entry_point
stringlengths 1
65
| original_triton_code
stringlengths 4.5k
619k
| python_code
stringlengths 208
60.9k
| triton_code
stringlengths 1.15k
275k
| repo_name
stringlengths 7
115
| module_name
stringlengths 1
65
| synthetic
bool 1
class | uuid
int64 0
18.5k
| licenses
sequencelengths 1
6
| stars
int64 0
19.8k
| sha
stringlengths 40
40
| repo_link
stringlengths 72
180
| pytorch_code
stringlengths 200
4.05k
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|
Features_2_to_1 | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ig/cig47vuhapdutooboct76wzrymtxcbuu7dlxiavnyagc5buk35ye.py
# Topologically Sorted Source Nodes: [max_4, max_5], Original ATen: [aten.max]
# Source node to ATen node mapping:
# max_4 => max_4
# max_5 => max_5
# Graph fragment:
# %max_4 : [num_users=1] = call_function[target=torch.ops.aten.max.dim](args = (%arg0_1, 1), kwargs = {})
# %max_5 : [num_users=1] = call_function[target=torch.ops.aten.max.dim](args = (%getitem_6, 1), kwargs = {})
triton_poi_fused_max_0 = async_compile.triton('triton_poi_fused_max_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 16, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_max_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (64*x1)), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + (64*x1)), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + (64*x1)), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + (64*x1)), xmask)
tmp7 = tl.load(in_ptr0 + (4 + x0 + (64*x1)), xmask)
tmp8 = tl.load(in_ptr0 + (20 + x0 + (64*x1)), xmask)
tmp10 = tl.load(in_ptr0 + (36 + x0 + (64*x1)), xmask)
tmp12 = tl.load(in_ptr0 + (52 + x0 + (64*x1)), xmask)
tmp15 = tl.load(in_ptr0 + (8 + x0 + (64*x1)), xmask)
tmp16 = tl.load(in_ptr0 + (24 + x0 + (64*x1)), xmask)
tmp18 = tl.load(in_ptr0 + (40 + x0 + (64*x1)), xmask)
tmp20 = tl.load(in_ptr0 + (56 + x0 + (64*x1)), xmask)
tmp23 = tl.load(in_ptr0 + (12 + x0 + (64*x1)), xmask)
tmp24 = tl.load(in_ptr0 + (28 + x0 + (64*x1)), xmask)
tmp26 = tl.load(in_ptr0 + (44 + x0 + (64*x1)), xmask)
tmp28 = tl.load(in_ptr0 + (60 + x0 + (64*x1)), xmask)
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp6 = triton_helpers.maximum(tmp4, tmp5)
tmp9 = triton_helpers.maximum(tmp7, tmp8)
tmp11 = triton_helpers.maximum(tmp9, tmp10)
tmp13 = triton_helpers.maximum(tmp11, tmp12)
tmp14 = triton_helpers.maximum(tmp6, tmp13)
tmp17 = triton_helpers.maximum(tmp15, tmp16)
tmp19 = triton_helpers.maximum(tmp17, tmp18)
tmp21 = triton_helpers.maximum(tmp19, tmp20)
tmp22 = triton_helpers.maximum(tmp14, tmp21)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp27 = triton_helpers.maximum(tmp25, tmp26)
tmp29 = triton_helpers.maximum(tmp27, tmp28)
tmp30 = triton_helpers.maximum(tmp22, tmp29)
tl.store(out_ptr0 + (x2), tmp30, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/kz/ckzd2zqltdogl55ubrceqzldx2tdkuzsxdgtqqc5o3e6elue77g2.py
# Topologically Sorted Source Nodes: [output], Original ATen: [aten.stack]
# Source node to ATen node mapping:
# output => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%permute, %expand, %getitem_2, %getitem_4, %expand_1], 2), kwargs = {})
triton_poi_fused_stack_1 = async_compile.triton('triton_poi_fused_stack_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_stack_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 14, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_stack_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 320
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 20
x1 = (xindex // 20) % 4
x2 = (xindex // 80)
x3 = (xindex // 20)
x4 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((20*x1) + (64*x2) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr0 + ((64*x2) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp11 = tl.load(in_ptr0 + (20 + (64*x2) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp12 = triton_helpers.maximum(tmp10, tmp11)
tmp13 = tl.load(in_ptr0 + (40 + (64*x2) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp14 = triton_helpers.maximum(tmp12, tmp13)
tmp15 = tl.load(in_ptr0 + (60 + (64*x2) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp16 = triton_helpers.maximum(tmp14, tmp15)
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp9, tmp16, tmp17)
tmp19 = tmp0 >= tmp7
tmp20 = tl.full([1], 12, tl.int64)
tmp21 = tmp0 < tmp20
tmp22 = tmp19 & tmp21
tmp23 = tl.load(in_ptr0 + ((16*x3) + ((-8) + x0)), tmp22 & xmask, eviction_policy='evict_last', other=0.0)
tmp24 = tl.load(in_ptr0 + (4 + (16*x3) + ((-8) + x0)), tmp22 & xmask, eviction_policy='evict_last', other=0.0)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp26 = tl.load(in_ptr0 + (8 + (16*x3) + ((-8) + x0)), tmp22 & xmask, eviction_policy='evict_last', other=0.0)
tmp27 = triton_helpers.maximum(tmp25, tmp26)
tmp28 = tl.load(in_ptr0 + (12 + (16*x3) + ((-8) + x0)), tmp22 & xmask, eviction_policy='evict_last', other=0.0)
tmp29 = triton_helpers.maximum(tmp27, tmp28)
tmp30 = tl.full(tmp29.shape, 0.0, tmp29.dtype)
tmp31 = tl.where(tmp22, tmp29, tmp30)
tmp32 = tmp0 >= tmp20
tmp33 = tl.full([1], 16, tl.int64)
tmp34 = tmp0 < tmp33
tmp35 = tmp32 & tmp34
tmp36 = tl.load(in_ptr0 + ((4*x1) + (64*x2) + ((-12) + x0)), tmp35 & xmask, eviction_policy='evict_last', other=0.0)
tmp37 = tl.load(in_ptr0 + (16 + (4*x1) + (64*x2) + ((-12) + x0)), tmp35 & xmask, eviction_policy='evict_last', other=0.0)
tmp38 = triton_helpers.maximum(tmp36, tmp37)
tmp39 = tl.load(in_ptr0 + (32 + (4*x1) + (64*x2) + ((-12) + x0)), tmp35 & xmask, eviction_policy='evict_last', other=0.0)
tmp40 = triton_helpers.maximum(tmp38, tmp39)
tmp41 = tl.load(in_ptr0 + (48 + (4*x1) + (64*x2) + ((-12) + x0)), tmp35 & xmask, eviction_policy='evict_last', other=0.0)
tmp42 = triton_helpers.maximum(tmp40, tmp41)
tmp43 = tl.full(tmp42.shape, 0.0, tmp42.dtype)
tmp44 = tl.where(tmp35, tmp42, tmp43)
tmp45 = tmp0 >= tmp33
tmp46 = tl.full([1], 20, tl.int64)
tmp47 = tmp0 < tmp46
tmp48 = tl.load(in_ptr1 + ((4*x2) + ((-16) + x0)), tmp45 & xmask, eviction_policy='evict_last', other=0.0)
tmp49 = tl.where(tmp35, tmp44, tmp48)
tmp50 = tl.where(tmp22, tmp31, tmp49)
tmp51 = tl.where(tmp9, tmp18, tmp50)
tmp52 = tl.where(tmp4, tmp5, tmp51)
tl.store(out_ptr0 + (x4), tmp52, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [max_4, max_5], Original ATen: [aten.max]
stream0 = get_raw_stream(0)
triton_poi_fused_max_0.run(arg0_1, buf0, 16, grid=grid(16), stream=stream0)
buf1 = empty_strided_cuda((4, 4, 20), (80, 20, 1), torch.float32)
# Topologically Sorted Source Nodes: [output], Original ATen: [aten.stack]
triton_poi_fused_stack_1.run(arg0_1, buf0, buf1, 320, grid=grid(320), stream=stream0)
del arg0_1
del buf0
return (buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.optim
import torch.nn as nn
class Features_2_to_1(nn.Module):
def __init__(self):
"""
take a batch (bs, n_vertices, n_vertices, in_features)
and returns (bs, n_vertices, basis * in_features)
where basis = 5
"""
super().__init__()
def forward(self, x):
b, n, _, in_features = x.size()
basis = 5
diag_part = torch.diagonal(x, dim1=1, dim2=2).permute(0, 2, 1)
max_diag_part = torch.max(diag_part, 1)[0].unsqueeze(1)
max_of_rows = torch.max(x, 2)[0]
max_of_cols = torch.max(x, 1)[0]
max_all = torch.max(torch.max(x, 1)[0], 1)[0].unsqueeze(1)
op1 = diag_part
op2 = max_diag_part.expand_as(op1)
op3 = max_of_rows
op4 = max_of_cols
op5 = max_all.expand_as(op1)
output = torch.stack([op1, op2, op3, op4, op5], dim=2)
assert output.size() == (b, n, basis, in_features), output.size()
return output.view(b, n, basis * in_features)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.optim
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_max_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)
tmp7 = tl.load(in_ptr0 + (4 + x0 + 64 * x1), xmask)
tmp8 = tl.load(in_ptr0 + (20 + x0 + 64 * x1), xmask)
tmp10 = tl.load(in_ptr0 + (36 + x0 + 64 * x1), xmask)
tmp12 = tl.load(in_ptr0 + (52 + x0 + 64 * x1), xmask)
tmp15 = tl.load(in_ptr0 + (8 + x0 + 64 * x1), xmask)
tmp16 = tl.load(in_ptr0 + (24 + x0 + 64 * x1), xmask)
tmp18 = tl.load(in_ptr0 + (40 + x0 + 64 * x1), xmask)
tmp20 = tl.load(in_ptr0 + (56 + x0 + 64 * x1), xmask)
tmp23 = tl.load(in_ptr0 + (12 + x0 + 64 * x1), xmask)
tmp24 = tl.load(in_ptr0 + (28 + x0 + 64 * x1), xmask)
tmp26 = tl.load(in_ptr0 + (44 + x0 + 64 * x1), xmask)
tmp28 = tl.load(in_ptr0 + (60 + x0 + 64 * x1), xmask)
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp6 = triton_helpers.maximum(tmp4, tmp5)
tmp9 = triton_helpers.maximum(tmp7, tmp8)
tmp11 = triton_helpers.maximum(tmp9, tmp10)
tmp13 = triton_helpers.maximum(tmp11, tmp12)
tmp14 = triton_helpers.maximum(tmp6, tmp13)
tmp17 = triton_helpers.maximum(tmp15, tmp16)
tmp19 = triton_helpers.maximum(tmp17, tmp18)
tmp21 = triton_helpers.maximum(tmp19, tmp20)
tmp22 = triton_helpers.maximum(tmp14, tmp21)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp27 = triton_helpers.maximum(tmp25, tmp26)
tmp29 = triton_helpers.maximum(tmp27, tmp28)
tmp30 = triton_helpers.maximum(tmp22, tmp29)
tl.store(out_ptr0 + x2, tmp30, xmask)
@triton.jit
def triton_poi_fused_stack_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 320
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 20
x1 = xindex // 20 % 4
x2 = xindex // 80
x3 = xindex // 20
x4 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (20 * x1 + 64 * x2 + x0), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr0 + (64 * x2 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tl.load(in_ptr0 + (20 + 64 * x2 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp12 = triton_helpers.maximum(tmp10, tmp11)
tmp13 = tl.load(in_ptr0 + (40 + 64 * x2 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp14 = triton_helpers.maximum(tmp12, tmp13)
tmp15 = tl.load(in_ptr0 + (60 + 64 * x2 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp16 = triton_helpers.maximum(tmp14, tmp15)
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp9, tmp16, tmp17)
tmp19 = tmp0 >= tmp7
tmp20 = tl.full([1], 12, tl.int64)
tmp21 = tmp0 < tmp20
tmp22 = tmp19 & tmp21
tmp23 = tl.load(in_ptr0 + (16 * x3 + (-8 + x0)), tmp22 & xmask,
eviction_policy='evict_last', other=0.0)
tmp24 = tl.load(in_ptr0 + (4 + 16 * x3 + (-8 + x0)), tmp22 & xmask,
eviction_policy='evict_last', other=0.0)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp26 = tl.load(in_ptr0 + (8 + 16 * x3 + (-8 + x0)), tmp22 & xmask,
eviction_policy='evict_last', other=0.0)
tmp27 = triton_helpers.maximum(tmp25, tmp26)
tmp28 = tl.load(in_ptr0 + (12 + 16 * x3 + (-8 + x0)), tmp22 & xmask,
eviction_policy='evict_last', other=0.0)
tmp29 = triton_helpers.maximum(tmp27, tmp28)
tmp30 = tl.full(tmp29.shape, 0.0, tmp29.dtype)
tmp31 = tl.where(tmp22, tmp29, tmp30)
tmp32 = tmp0 >= tmp20
tmp33 = tl.full([1], 16, tl.int64)
tmp34 = tmp0 < tmp33
tmp35 = tmp32 & tmp34
tmp36 = tl.load(in_ptr0 + (4 * x1 + 64 * x2 + (-12 + x0)), tmp35 &
xmask, eviction_policy='evict_last', other=0.0)
tmp37 = tl.load(in_ptr0 + (16 + 4 * x1 + 64 * x2 + (-12 + x0)), tmp35 &
xmask, eviction_policy='evict_last', other=0.0)
tmp38 = triton_helpers.maximum(tmp36, tmp37)
tmp39 = tl.load(in_ptr0 + (32 + 4 * x1 + 64 * x2 + (-12 + x0)), tmp35 &
xmask, eviction_policy='evict_last', other=0.0)
tmp40 = triton_helpers.maximum(tmp38, tmp39)
tmp41 = tl.load(in_ptr0 + (48 + 4 * x1 + 64 * x2 + (-12 + x0)), tmp35 &
xmask, eviction_policy='evict_last', other=0.0)
tmp42 = triton_helpers.maximum(tmp40, tmp41)
tmp43 = tl.full(tmp42.shape, 0.0, tmp42.dtype)
tmp44 = tl.where(tmp35, tmp42, tmp43)
tmp45 = tmp0 >= tmp33
tl.full([1], 20, tl.int64)
tmp48 = tl.load(in_ptr1 + (4 * x2 + (-16 + x0)), tmp45 & xmask,
eviction_policy='evict_last', other=0.0)
tmp49 = tl.where(tmp35, tmp44, tmp48)
tmp50 = tl.where(tmp22, tmp31, tmp49)
tmp51 = tl.where(tmp9, tmp18, tmp50)
tmp52 = tl.where(tmp4, tmp5, tmp51)
tl.store(out_ptr0 + x4, tmp52, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_max_0[grid(16)](arg0_1, buf0, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 20), (80, 20, 1), torch.float32)
triton_poi_fused_stack_1[grid(320)](arg0_1, buf0, buf1, 320, XBLOCK
=128, num_warps=4, num_stages=1)
del arg0_1
del buf0
return buf1,
class Features_2_to_1New(nn.Module):
def __init__(self):
"""
take a batch (bs, n_vertices, n_vertices, in_features)
and returns (bs, n_vertices, basis * in_features)
where basis = 5
"""
super().__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MauTrib/gnn-en-folie | Features_2_to_1 | false | 831 | [
"Apache-2.0"
] | 0 | 3ca639919a2b285a41641717f4131107c015b510 | https://github.com/MauTrib/gnn-en-folie/tree/3ca639919a2b285a41641717f4131107c015b510 | import torch
import torch.optim
import torch.nn as nn
class Model(nn.Module):
def __init__(self):
"""
take a batch (bs, n_vertices, n_vertices, in_features)
and returns (bs, n_vertices, basis * in_features)
where basis = 5
"""
super().__init__()
def forward(self, x):
b, n, _, in_features = x.size()
basis = 5
diag_part = torch.diagonal(x, dim1=1, dim2=2).permute(0, 2, 1)
max_diag_part = torch.max(diag_part, 1)[0].unsqueeze(1)
max_of_rows = torch.max(x, 2)[0]
max_of_cols = torch.max(x, 1)[0]
max_all = torch.max(torch.max(x, 1)[0], 1)[0].unsqueeze(1)
op1 = diag_part
op2 = max_diag_part.expand_as(op1)
op3 = max_of_rows
op4 = max_of_cols
op5 = max_all.expand_as(op1)
output = torch.stack([op1, op2, op3, op4, op5], dim=2)
assert output.size() == (b, n, basis, in_features), output.size()
return output.view(b, n, basis * in_features)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
Scale | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/vj/cvjysphv56i4vvevccww6n5duxxdl7sscct4ytyabkj3frgyogia.py
# Topologically Sorted Source Nodes: [mul], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# mul => mul
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg0_1, 30), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 30.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class Scale(nn.Module):
def __init__(self, scale=30):
super(Scale, self).__init__()
self.scale = scale
def forward(self, x):
return x * self.scale
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 30.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](arg0_1, buf0, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class ScaleNew(nn.Module):
def __init__(self, scale=30):
super(ScaleNew, self).__init__()
self.scale = scale
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MickeyZeng/Data-Visualization | Scale | false | 832 | [
"MIT"
] | 0 | c7005d1096545d7a5eb96dd0c9bc13e874d42fa4 | https://github.com/MickeyZeng/Data-Visualization/tree/c7005d1096545d7a5eb96dd0c9bc13e874d42fa4 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, scale=30):
super().__init__()
self.scale = scale
def forward(self, x):
return x * self.scale
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
MutualBiAffineAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/c4/cc4khg7fwbxxm2fufox7nnkf4gfybrmj5ir2tx3zuxfioc5b2dya.py
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2], -1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/hj/chjzotk5iydxvuetxetlv36s7car7cdb24whkuqihxwcy5kkr4o2.py
# Topologically Sorted Source Nodes: [tanh], Original ATen: [aten.tanh]
# Source node to ATen node mapping:
# tanh => tanh
# Graph fragment:
# %tanh : [num_users=2] = call_function[target=torch.ops.aten.tanh.default](args = (%view_1,), kwargs = {})
triton_poi_fused_tanh_1 = async_compile.triton('triton_poi_fused_tanh_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_tanh_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_tanh_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + (x2), tmp3, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/y3/cy3gemhwjvq3uqncl6n6peh7zmyo5ilbe7mjwdurj3vr3k2edujx.py
# Topologically Sorted Source Nodes: [softmax, att_1, sum_1, add, att_2, att_res], Original ATen: [aten._softmax, aten.mul, aten.sum, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# att_1 => mul
# att_2 => div_1
# att_res => mul_1
# softmax => amax, div, exp, sub, sum_1
# sum_1 => sum_2
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%view_3, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_3, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div, %unsqueeze), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%mul, [-1], True), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sum_2, 1e-20), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mul, %add), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div_1, %primals_2), kwargs = {})
triton_poi_fused__softmax_add_div_mul_sum_2 = async_compile.triton('triton_poi_fused__softmax_add_div_mul_sum_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_add_div_mul_sum_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_add_div_mul_sum_2(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4) % 64
x3 = (xindex // 4)
x4 = xindex % 256
x5 = xindex
tmp0 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + (x4), xmask, eviction_policy='evict_last')
tmp1 = tmp0 - tmp0
tmp2 = tl_math.exp(tmp1)
tmp3 = tmp2 / tmp2
tmp5 = tmp3 * tmp4
tmp6 = 1e-20
tmp7 = tmp5 + tmp6
tmp8 = tmp5 / tmp7
tmp10 = tmp8 * tmp9
tl.store(out_ptr0 + (x5), tmp10, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 8), (8, 1))
assert_size_stride(primals_4, (4, ), (1, ))
assert_size_stride(primals_5, (1, 4), (4, 1))
assert_size_stride(primals_6, (1, ), (1, ))
assert_size_stride(primals_7, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 8), (128, 32, 8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, primals_2, buf0, 512, grid=grid(512), stream=stream0)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf0, (64, 8), (8, 1), 0), reinterpret_tensor(primals_3, (8, 4), (1, 8), 0), out=buf1)
del primals_3
buf2 = reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [tanh], Original ATen: [aten.tanh]
triton_poi_fused_tanh_1.run(buf2, primals_4, 256, grid=grid(256), stream=stream0)
del primals_4
buf4 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [att], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_6, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 1), (1, 4), 0), alpha=1, beta=1, out=buf4)
del primals_6
buf5 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [softmax, att_1, sum_1, add, att_2, att_res], Original ATen: [aten._softmax, aten.mul, aten.sum, aten.add, aten.div]
triton_poi_fused__softmax_add_div_mul_sum_2.run(buf4, primals_7, primals_2, buf5, 1024, grid=grid(1024), stream=stream0)
return (buf5, primals_2, primals_7, reinterpret_tensor(buf0, (64, 8), (8, 1), 0), buf2, buf4, primals_5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((1, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
from torch.autograd import *
class MutualBiAffineAttention(nn.Module):
"""
Mutual BiAffine Attention between 2 kinds of features.
"""
def __init__(self, hidden_size):
super(MutualBiAffineAttention, self).__init__()
self.linear1 = nn.Linear(2 * hidden_size, hidden_size)
self.linear2 = nn.Linear(hidden_size, 1)
def forward(self, in1, att_feats, att_mask):
att = self.linear2(torch.tanh(self.linear1(torch.cat([in1,
att_feats], -1))))
att_mask = att_mask.unsqueeze(-1)
att = torch.softmax(att, -1) * att_mask
att = att / (att.sum(-1, keepdim=True) + 1e-20)
att_res = att * att_feats
return att_res
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4])]
def get_init_inputs():
return [[], {'hidden_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
from torch.autograd import *
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_tanh_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
@triton.jit
def triton_poi_fused__softmax_add_div_mul_sum_2(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 64
x3 = xindex // 4
x4 = xindex % 256
x5 = xindex
tmp0 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + x3, xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + x4, xmask, eviction_policy='evict_last')
tmp1 = tmp0 - tmp0
tmp2 = tl_math.exp(tmp1)
tmp3 = tmp2 / tmp2
tmp5 = tmp3 * tmp4
tmp6 = 1e-20
tmp7 = tmp5 + tmp6
tmp8 = tmp5 / tmp7
tmp10 = tmp8 * tmp9
tl.store(out_ptr0 + x5, tmp10, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 8), (8, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (1, 4), (4, 1))
assert_size_stride(primals_6, (1,), (1,))
assert_size_stride(primals_7, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 8), (128, 32, 8, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(512)](primals_1, primals_2, buf0, 512,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 8), (8, 1), 0),
reinterpret_tensor(primals_3, (8, 4), (1, 8), 0), out=buf1)
del primals_3
buf2 = reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf1
triton_poi_fused_tanh_1[grid(256)](buf2, primals_4, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_4
buf4 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.addmm(primals_6, reinterpret_tensor(buf2, (64, 4), (
4, 1), 0), reinterpret_tensor(primals_5, (4, 1), (1, 4), 0),
alpha=1, beta=1, out=buf4)
del primals_6
buf5 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1),
torch.float32)
triton_poi_fused__softmax_add_div_mul_sum_2[grid(1024)](buf4,
primals_7, primals_2, buf5, 1024, XBLOCK=256, num_warps=4,
num_stages=1)
return buf5, primals_2, primals_7, reinterpret_tensor(buf0, (64, 8), (8,
1), 0), buf2, buf4, primals_5
class MutualBiAffineAttentionNew(nn.Module):
"""
Mutual BiAffine Attention between 2 kinds of features.
"""
def __init__(self, hidden_size):
super(MutualBiAffineAttentionNew, self).__init__()
self.linear1 = nn.Linear(2 * hidden_size, hidden_size)
self.linear2 = nn.Linear(hidden_size, 1)
def forward(self, input_0, input_1, input_2):
primals_3 = self.linear1.weight
primals_4 = self.linear1.bias
primals_5 = self.linear2.weight
primals_6 = self.linear2.bias
primals_1 = input_0
primals_2 = input_1
primals_7 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
| Maxi-0902/DRAN | MutualBiAffineAttention | false | 833 | [
"MIT"
] | 0 | c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | https://github.com/Maxi-0902/DRAN/tree/c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | import torch
import torch.nn as nn
from torch.autograd import *
class Model(nn.Module):
"""
Mutual BiAffine Attention between 2 kinds of features.
"""
def __init__(self, hidden_size):
super().__init__()
self.linear1 = nn.Linear(2 * hidden_size, hidden_size)
self.linear2 = nn.Linear(hidden_size, 1)
def forward(self, in1, att_feats, att_mask):
att = self.linear2(torch.tanh(self.linear1(torch.cat([in1,
att_feats], -1))))
att_mask = att_mask.unsqueeze(-1)
att = torch.softmax(att, -1) * att_mask
att = att / (att.sum(-1, keepdim=True) + 1e-20)
att_res = att * att_feats
return att_res
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
DownBlock | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/6z/c6zaho43pxjqk5eqidz3d5ckh3z6vlh34gcbntokha6jjsr4all2.py
# Topologically Sorted Source Nodes: [out, l0], Original ATen: [aten.convolution, aten._prelu_kernel]
# Source node to ATen node mapping:
# l0 => gt, mul, where
# out => convolution
# Graph fragment:
# %convolution : [num_users=4] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [4, 4], [2, 2], [1, 1], False, [0, 0], 1), kwargs = {})
# %gt : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%convolution, 0), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view, %convolution), kwargs = {})
# %where : [num_users=3] = call_function[target=torch.ops.aten.where.self](args = (%gt, %convolution, %mul), kwargs = {})
triton_poi_fused__prelu_kernel_convolution_0 = async_compile.triton('triton_poi_fused__prelu_kernel_convolution_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__prelu_kernel_convolution_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_0(in_out_ptr0, in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (0))
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/v2/cv2gefdciotml3zwtkzv4ghtu2a4dbeoas7q3ue7dcfa4f2mizfk.py
# Topologically Sorted Source Nodes: [out_1, h0, sub], Original ATen: [aten.convolution, aten._prelu_kernel, aten.sub]
# Source node to ATen node mapping:
# h0 => gt_1, mul_1, where_1
# out_1 => convolution_1
# sub => sub
# Graph fragment:
# %convolution_1 : [num_users=4] = call_function[target=torch.ops.aten.convolution.default](args = (%where, %primals_5, %primals_6, [4, 4], [2, 2], [1, 1], True, [0, 0], 1), kwargs = {})
# %gt_1 : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%convolution_1, 0), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, %convolution_1), kwargs = {})
# %where_1 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%gt_1, %convolution_1, %mul_1), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%where_1, %primals_3), kwargs = {})
triton_poi_fused__prelu_kernel_convolution_sub_1 = async_compile.triton('triton_poi_fused__prelu_kernel_convolution_sub_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__prelu_kernel_convolution_sub_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_sub_1(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 16) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (0))
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + (x3), xmask)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 - tmp9
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
tl.store(out_ptr0 + (x3), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/p4/cp477hx6wscm5qo2wzce4hwmhoeb74h5hefxsw2hibtz4mcbgjag.py
# Topologically Sorted Source Nodes: [out_2, l1, add], Original ATen: [aten.convolution, aten._prelu_kernel, aten.add]
# Source node to ATen node mapping:
# add => add
# l1 => gt_2, mul_2, where_2
# out_2 => convolution_2
# Graph fragment:
# %convolution_2 : [num_users=4] = call_function[target=torch.ops.aten.convolution.default](args = (%sub, %primals_8, %primals_9, [4, 4], [2, 2], [1, 1], False, [0, 0], 1), kwargs = {})
# %gt_2 : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%convolution_2, 0), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_2, %convolution_2), kwargs = {})
# %where_2 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%gt_2, %convolution_2, %mul_2), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%where_2, %where), kwargs = {})
triton_poi_fused__prelu_kernel_add_convolution_2 = async_compile.triton('triton_poi_fused__prelu_kernel_add_convolution_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__prelu_kernel_add_convolution_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__prelu_kernel_add_convolution_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (0))
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + (x2), xmask)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 + tmp9
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (1, ), (1, ))
assert_size_stride(primals_5, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_6, (4, ), (1, ))
assert_size_stride(primals_7, (1, ), (1, ))
assert_size_stride(primals_8, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_9, (4, ), (1, ))
assert_size_stride(primals_10, (1, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(4, 4), padding=(2, 2), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 1, 1), (4, 1, 1, 1))
buf1 = buf0; del buf0 # reuse
buf2 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [out, l0], Original ATen: [aten.convolution, aten._prelu_kernel]
stream0 = get_raw_stream(0)
triton_poi_fused__prelu_kernel_convolution_0.run(buf1, primals_2, primals_4, buf2, 16, grid=grid(16), stream=stream0)
del primals_2
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.convolution]
buf3 = extern_kernels.convolution(buf2, primals_5, stride=(4, 4), padding=(2, 2), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 4, 4, 4), (64, 16, 4, 1))
buf4 = buf3; del buf3 # reuse
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_1, h0, sub], Original ATen: [aten.convolution, aten._prelu_kernel, aten.sub]
triton_poi_fused__prelu_kernel_convolution_sub_1.run(buf4, primals_6, primals_7, primals_3, buf5, 256, grid=grid(256), stream=stream0)
del primals_6
# Topologically Sorted Source Nodes: [out_2], Original ATen: [aten.convolution]
buf6 = extern_kernels.convolution(buf5, primals_8, stride=(4, 4), padding=(2, 2), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 4, 1, 1), (4, 1, 1, 1))
buf7 = buf6; del buf6 # reuse
buf8 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_2, l1, add], Original ATen: [aten.convolution, aten._prelu_kernel, aten.add]
triton_poi_fused__prelu_kernel_add_convolution_2.run(buf7, primals_9, primals_10, buf2, buf8, 16, grid=grid(16), stream=stream0)
del primals_9
return (buf8, primals_1, primals_3, primals_4, primals_5, primals_7, primals_8, primals_10, buf1, buf2, buf4, buf5, buf7, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 8, 8), (256, 64, 8, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, 4, 8, 8), (256, 64, 8, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 4, 8, 8), (256, 64, 8, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class ConvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=3, stride=1,
padding=1, bias=True, norm=None):
super(ConvBlock, self).__init__()
self.conv = nn.Conv2d(input_size, output_size, kernel_size, stride,
padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.conv(x))
else:
out = self.conv(x)
return self.act(out)
class DeconvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=4, stride=2,
padding=1, bias=True, activation='prelu', norm=None):
super(DeconvBlock, self).__init__()
self.deconv = nn.ConvTranspose2d(input_size, output_size,
kernel_size, stride, padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.activation = activation
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.deconv(x))
else:
out = self.deconv(x)
if self.activation is not None:
return self.act(out)
else:
return out
class DownBlock(nn.Module):
def __init__(self, num_filter, kernel_size=8, stride=4, padding=2, bias
=True, activation='prelu', norm=None):
super(DownBlock, self).__init__()
self.down_conv1 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
self.down_conv2 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
self.down_conv3 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
def forward(self, x):
l0 = self.down_conv1(x)
h0 = self.down_conv2(l0)
l1 = self.down_conv3(h0 - x)
return l1 + l0
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'num_filter': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_0(in_out_ptr0, in_ptr0,
in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tl.store(in_out_ptr0 + x2, tmp2, xmask)
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_sub_1(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + x3, xmask)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 - tmp9
tl.store(in_out_ptr0 + x3, tmp2, xmask)
tl.store(out_ptr0 + x3, tmp10, xmask)
@triton.jit
def triton_poi_fused__prelu_kernel_add_convolution_2(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 + tmp9
tl.store(in_out_ptr0 + x2, tmp2, xmask)
tl.store(out_ptr0 + x2, tmp10, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (1,), (1,))
assert_size_stride(primals_5, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (1,), (1,))
assert_size_stride(primals_8, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_9, (4,), (1,))
assert_size_stride(primals_10, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(4,
4), padding=(2, 2), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 1, 1), (4, 1, 1, 1))
buf1 = buf0
del buf0
buf2 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__prelu_kernel_convolution_0[grid(16)](buf1,
primals_2, primals_4, buf2, 16, XBLOCK=16, num_warps=1,
num_stages=1)
del primals_2
buf3 = extern_kernels.convolution(buf2, primals_5, stride=(4, 4),
padding=(2, 2), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 4, 4, 4), (64, 16, 4, 1))
buf4 = buf3
del buf3
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__prelu_kernel_convolution_sub_1[grid(256)](buf4,
primals_6, primals_7, primals_3, buf5, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_6
buf6 = extern_kernels.convolution(buf5, primals_8, stride=(4, 4),
padding=(2, 2), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 4, 1, 1), (4, 1, 1, 1))
buf7 = buf6
del buf6
buf8 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
triton_poi_fused__prelu_kernel_add_convolution_2[grid(16)](buf7,
primals_9, primals_10, buf2, buf8, 16, XBLOCK=16, num_warps=1,
num_stages=1)
del primals_9
return (buf8, primals_1, primals_3, primals_4, primals_5, primals_7,
primals_8, primals_10, buf1, buf2, buf4, buf5, buf7)
class ConvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=3, stride=1,
padding=1, bias=True, norm=None):
super(ConvBlock, self).__init__()
self.conv = nn.Conv2d(input_size, output_size, kernel_size, stride,
padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.conv(x))
else:
out = self.conv(x)
return self.act(out)
class DeconvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=4, stride=2,
padding=1, bias=True, activation='prelu', norm=None):
super(DeconvBlock, self).__init__()
self.deconv = nn.ConvTranspose2d(input_size, output_size,
kernel_size, stride, padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.activation = activation
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.deconv(x))
else:
out = self.deconv(x)
if self.activation is not None:
return self.act(out)
else:
return out
class DownBlockNew(nn.Module):
def __init__(self, num_filter, kernel_size=8, stride=4, padding=2, bias
=True, activation='prelu', norm=None):
super(DownBlockNew, self).__init__()
self.down_conv1 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
self.down_conv2 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
self.down_conv3 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
def forward(self, input_0):
primals_1 = self.down_conv1.conv.weight
primals_2 = self.down_conv1.conv.bias
primals_4 = self.down_conv1.act.weight
primals_5 = self.down_conv2.deconv.weight
primals_6 = self.down_conv2.deconv.bias
primals_7 = self.down_conv2.act.weight
primals_8 = self.down_conv3.conv.weight
primals_9 = self.down_conv3.conv.bias
primals_10 = self.down_conv3.act.weight
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return output[0]
| MatusBako/MakeFacesGreatAgain | DownBlock | false | 834 | [
"MIT"
] | 0 | e4941a8460db79dec566ed02d4b23eafb416a6db | https://github.com/MatusBako/MakeFacesGreatAgain/tree/e4941a8460db79dec566ed02d4b23eafb416a6db | import torch
import torch.nn as nn
class ConvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=3, stride=1,
padding=1, bias=True, norm=None):
super().__init__()
self.conv = nn.Conv2d(input_size, output_size, kernel_size, stride,
padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.conv(x))
else:
out = self.conv(x)
return self.act(out)
class DeconvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=4, stride=2,
padding=1, bias=True, activation='prelu', norm=None):
super().__init__()
self.deconv = nn.ConvTranspose2d(input_size, output_size,
kernel_size, stride, padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.activation = activation
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.deconv(x))
else:
out = self.deconv(x)
if self.activation is not None:
return self.act(out)
else:
return out
class Model(nn.Module):
def __init__(self, num_filter, kernel_size=8, stride=4, padding=2, bias
=True, activation='prelu', norm=None):
super().__init__()
self.down_conv1 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
self.down_conv2 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
self.down_conv3 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
def forward(self, x):
l0 = self.down_conv1(x)
h0 = self.down_conv2(l0)
l1 = self.down_conv3(h0 - x)
return l1 + l0
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
RewardCriterion | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/jk/cjk4a6hzo7lmkian2zuwtbzbsa4v76s5vf3coi67m65hich66au5.py
# Topologically Sorted Source Nodes: [neg, mul, output, sum_1, sum_2, output_1], Original ATen: [aten.neg, aten.mul, aten.sum, aten.div]
# Source node to ATen node mapping:
# mul => mul
# neg => neg
# output => mul_1
# output_1 => div
# sum_1 => sum_1
# sum_2 => sum_2
# Graph fragment:
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%view,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%neg, %view_1), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %view_2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%mul_1,), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%view_2,), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_1, %sum_2), kwargs = {})
triton_per_fused_div_mul_neg_sum_0 = async_compile.triton('triton_per_fused_div_mul_neg_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 16],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {4: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 5), equal_to_1=(4,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_div_mul_neg_sum_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 2, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_div_mul_neg_sum_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp2 = tl.load(in_ptr1 + (r0), None)
tmp1 = -tmp0
tmp3 = tmp1 * tmp2
tmp4 = r0 % 4
tmp5 = tl.full([1, 1], 0, tl.int64)
tmp6 = tmp4 >= tmp5
tmp7 = tl.full([1, 1], 1, tl.int64)
tmp8 = tmp4 < tmp7
tmp9 = 1.0
tmp10 = tl.full(tmp9.shape, 0.0, tmp9.dtype)
tmp11 = tl.where(tmp8, tmp9, tmp10)
tmp12 = tmp4 >= tmp7
tmp13 = tl.full([1, 1], 4, tl.int64)
tmp14 = tmp4 < tmp13
tmp15 = tl.load(in_ptr2 + (tl.broadcast_to((4*(r0 // 4)) + ((-1) + (r0 % 4)), [XBLOCK, RBLOCK])), tmp12, eviction_policy='evict_last', other=0.0)
tmp16 = 0.0
tmp17 = tmp15 > tmp16
tmp18 = tmp17.to(tl.float32)
tmp19 = tl.full(tmp18.shape, 0.0, tmp18.dtype)
tmp20 = tl.where(tmp12, tmp18, tmp19)
tmp21 = tl.where(tmp8, tmp11, tmp20)
tmp22 = tmp3 * tmp21
tmp23 = tl.broadcast_to(tmp22, [XBLOCK, RBLOCK])
tmp25 = tl.sum(tmp23, 1)[:, None]
tmp26 = tl.broadcast_to(tmp21, [XBLOCK, RBLOCK])
tmp28 = tl.sum(tmp26, 1)[:, None]
tmp29 = tmp25 / tmp28
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp29, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
assert_size_stride(arg2_1, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf2 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [neg, mul, output, sum_1, sum_2, output_1], Original ATen: [aten.neg, aten.mul, aten.sum, aten.div]
stream0 = get_raw_stream(0)
triton_per_fused_div_mul_neg_sum_0.run(buf2, arg0_1, arg1_1, arg2_1, 1, 16, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
del arg2_1
return (buf2, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
arg2_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1, arg2_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
from torch.autograd import *
def to_contiguous(tensor):
if tensor.is_contiguous():
return tensor
else:
return tensor.contiguous()
class RewardCriterion(nn.Module):
def __init__(self):
super(RewardCriterion, self).__init__()
def forward(self, input, seq, reward, gpn_loss=None):
input = to_contiguous(input).view(-1)
reward = to_contiguous(reward).view(-1)
mask = (seq > 0).float()
mask = to_contiguous(torch.cat([mask.new(mask.size(0), 1).fill_(1),
mask[:, :-1]], 1)).view(-1)
if gpn_loss is None:
output = -input * reward * mask
output = torch.sum(output) / torch.sum(mask)
else:
gpn_loss = gpn_loss.unsqueeze(1).expand(gpn_loss.size(0), seq.
size(1)).contiguous().view(-1)
output = (-input * reward + gpn_loss * torch.exp(reward)) * mask
output = torch.sum(output) / torch.sum(mask)
return output
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
from torch.autograd import *
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_div_mul_neg_sum_0(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp2 = tl.load(in_ptr1 + r0, None)
tmp1 = -tmp0
tmp3 = tmp1 * tmp2
tmp4 = r0 % 4
tl.full([1, 1], 0, tl.int64)
tmp7 = tl.full([1, 1], 1, tl.int64)
tmp8 = tmp4 < tmp7
tmp9 = 1.0
tmp10 = tl.full(tmp9.shape, 0.0, tmp9.dtype)
tmp11 = tl.where(tmp8, tmp9, tmp10)
tmp12 = tmp4 >= tmp7
tl.full([1, 1], 4, tl.int64)
tmp15 = tl.load(in_ptr2 + tl.broadcast_to(4 * (r0 // 4) + (-1 + r0 % 4),
[XBLOCK, RBLOCK]), tmp12, eviction_policy='evict_last', other=0.0)
tmp16 = 0.0
tmp17 = tmp15 > tmp16
tmp18 = tmp17.to(tl.float32)
tmp19 = tl.full(tmp18.shape, 0.0, tmp18.dtype)
tmp20 = tl.where(tmp12, tmp18, tmp19)
tmp21 = tl.where(tmp8, tmp11, tmp20)
tmp22 = tmp3 * tmp21
tmp23 = tl.broadcast_to(tmp22, [XBLOCK, RBLOCK])
tmp25 = tl.sum(tmp23, 1)[:, None]
tmp26 = tl.broadcast_to(tmp21, [XBLOCK, RBLOCK])
tmp28 = tl.sum(tmp26, 1)[:, None]
tmp29 = tmp25 / tmp28
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp29, None)
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
assert_size_stride(arg2_1, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf2 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_div_mul_neg_sum_0[grid(1)](buf2, arg0_1, arg1_1,
arg2_1, 1, 16, XBLOCK=1, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
del arg2_1
return buf2,
def to_contiguous(tensor):
if tensor.is_contiguous():
return tensor
else:
return tensor.contiguous()
class RewardCriterionNew(nn.Module):
def __init__(self):
super(RewardCriterionNew, self).__init__()
def forward(self, input_0, input_1, input_2):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
output = call([arg0_1, arg1_1, arg2_1])
return output[0]
| Maxi-0902/DRAN | RewardCriterion | false | 835 | [
"MIT"
] | 0 | c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | https://github.com/Maxi-0902/DRAN/tree/c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | import torch
import torch.nn as nn
from torch.autograd import *
def to_contiguous(tensor):
if tensor.is_contiguous():
return tensor
else:
return tensor.contiguous()
class Model(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input, seq, reward, gpn_loss=None):
input = to_contiguous(input).view(-1)
reward = to_contiguous(reward).view(-1)
mask = (seq > 0).float()
mask = to_contiguous(torch.cat([mask.new(mask.size(0), 1).fill_(1),
mask[:, :-1]], 1)).view(-1)
if gpn_loss is None:
output = -input * reward * mask
output = torch.sum(output) / torch.sum(mask)
else:
gpn_loss = gpn_loss.unsqueeze(1).expand(gpn_loss.size(0), seq.
size(1)).contiguous().view(-1)
output = (-input * reward + gpn_loss * torch.exp(reward)) * mask
output = torch.sum(output) / torch.sum(mask)
return output
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return []
|
SelfAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/gd/cgdyncc5fbhtbwk4m3qh4p3orh6ir2n4sbfvfheve5ebdbvm6btr.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_1,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x3 = xindex
y0 = yindex % 4
y1 = (yindex // 4) % 4
y2 = (yindex // 16)
y4 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x3) + (16*y2) + (64*y1)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x3 + (4*y4)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2a/c2aleeuqhoz4te7w3qwhl66rqffdzonerxbyu7iyqwrmvsg67b3e.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# x_1 => clone_2
# Graph fragment:
# %clone_2 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_4,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2x/c2xtedo5flggmbs3xi3ahonfvuhv7jopaiz4kzrgrsin73g5evcv.py
# Topologically Sorted Source Nodes: [x_1, attn], Original ATen: [aten.clone, aten._softmax]
# Source node to ATen node mapping:
# attn => amax, exp, sub
# x_1 => clone_3
# Graph fragment:
# %clone_3 : [num_users=2] = call_function[target=torch.ops.aten.clone.default](args = (%permute_6,), kwargs = {memory_format: torch.contiguous_format})
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%clone_3, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%clone_3, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
triton_poi_fused__softmax_clone_2 = async_compile.triton('triton_poi_fused__softmax_clone_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_clone_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_clone_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = (xindex // 16)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x3), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6j/c6jxcwk2zob7wa3y7plr7b6myq6bexpfj24ifkfsbtise4t4sq6g.py
# Topologically Sorted Source Nodes: [attn], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# attn => div, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_3 = async_compile.triton('triton_poi_fused__softmax_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = (yindex // 4)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x2) + (16*y1)), xmask & ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (4 + y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (8 + y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (12 + y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2 + (4*y3)), tmp8, xmask & ymask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(primals_2, buf0, 64, 4, grid=grid(64, 4), stream=stream0)
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf1)
del primals_1
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf1, buf2, 256, grid=grid(256), stream=stream0)
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_3, (4, 4), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 1, 4), torch.float32)
# Topologically Sorted Source Nodes: [x_1, attn], Original ATen: [aten.clone, aten._softmax]
triton_poi_fused__softmax_clone_2.run(buf3, buf4, 256, grid=grid(256), stream=stream0)
buf5 = reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf3 # reuse
# Topologically Sorted Source Nodes: [attn], Original ATen: [aten._softmax]
triton_poi_fused__softmax_3.run(buf4, buf5, 64, 4, grid=grid(64, 4), stream=stream0)
buf6 = reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0); del buf4 # reuse
# Topologically Sorted Source Nodes: [x_2], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf5, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(primals_2, (16, 4, 4), (16, 4, 1), 0), out=buf6)
return (reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0), buf5, reinterpret_tensor(buf0, (64, 4), (4, 1), 0), buf1, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), buf5, reinterpret_tensor(primals_2, (16, 4, 4), (16, 1, 4), 0), primals_3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn as nn
import torch.utils.data.distributed
class SelfAttention(nn.Module):
"""
Self SelfAttention Layer
Given $X\\in \\mathbb{R}^{n imes in_feature}$, the attention is calculated by: $a=Softmax(W_2tanh(W_1X))$, where
$W_1 \\in \\mathbb{R}^{hidden imes in_feature}$, $W_2 \\in \\mathbb{R}^{out_feature imes hidden}$.
The final output is: $out=aX$, which is unrelated with input $n$.
"""
def __init__(self, *, hidden, in_feature, out_feature):
"""
The init function.
:param hidden: the hidden dimension, can be viewed as the number of experts.
:param in_feature: the input feature dimension.
:param out_feature: the output feature dimension.
"""
super(SelfAttention, self).__init__()
self.w1 = torch.nn.Parameter(torch.FloatTensor(hidden, in_feature))
self.w2 = torch.nn.Parameter(torch.FloatTensor(out_feature, hidden))
self.reset_parameters()
def reset_parameters(self):
"""
Use xavier_normal method to initialize parameters.
"""
nn.init.xavier_normal_(self.w1)
nn.init.xavier_normal_(self.w2)
def forward(self, X):
"""
The forward function.
:param X: The input feature map. $X \\in \\mathbb{R}^{n imes in_feature}$.
:return: The final embeddings and attention matrix.
"""
x = torch.tanh(torch.matmul(self.w1, X.transpose(1, 0)))
x = torch.matmul(self.w2, x)
attn = torch.nn.functional.softmax(x, dim=-1)
x = torch.matmul(attn, X)
return x, attn
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'hidden': 4, 'in_feature': 4, 'out_feature': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch import nn as nn
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x3 = xindex
y0 = yindex % 4
y1 = yindex // 4 % 4
y2 = yindex // 16
y4 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x3 + 16 * y2 + 64 * y1), xmask &
ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x3 + 4 * y4), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_poi_fused__softmax_clone_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = xindex // 16
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (4 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (8 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (12 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2 + 4 * y3), tmp8, xmask & ymask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(64, 4)](primals_2, buf0, 64, 4,
XBLOCK=4, YBLOCK=32, num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf1)
del primals_1
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(256)](buf1, buf2, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 4), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 1, 4), torch.float32)
triton_poi_fused__softmax_clone_2[grid(256)](buf3, buf4, 256,
XBLOCK=128, num_warps=4, num_stages=1)
buf5 = reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf3
triton_poi_fused__softmax_3[grid(64, 4)](buf4, buf5, 64, 4, XBLOCK=
4, YBLOCK=32, num_warps=4, num_stages=1)
buf6 = reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0)
del buf4
extern_kernels.bmm(reinterpret_tensor(buf5, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(primals_2, (16, 4, 4), (16, 4, 1), 0),
out=buf6)
return reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0
), buf5, reinterpret_tensor(buf0, (64, 4), (4, 1), 0
), buf1, reinterpret_tensor(buf2, (64, 4), (4, 1), 0
), buf5, reinterpret_tensor(primals_2, (16, 4, 4), (16, 1, 4), 0
), primals_3
class SelfAttentionNew(nn.Module):
"""
Self SelfAttention Layer
Given $X\\in \\mathbb{R}^{n imes in_feature}$, the attention is calculated by: $a=Softmax(W_2tanh(W_1X))$, where
$W_1 \\in \\mathbb{R}^{hidden imes in_feature}$, $W_2 \\in \\mathbb{R}^{out_feature imes hidden}$.
The final output is: $out=aX$, which is unrelated with input $n$.
"""
def __init__(self, *, hidden, in_feature, out_feature):
"""
The init function.
:param hidden: the hidden dimension, can be viewed as the number of experts.
:param in_feature: the input feature dimension.
:param out_feature: the output feature dimension.
"""
super(SelfAttentionNew, self).__init__()
self.w1 = torch.nn.Parameter(torch.FloatTensor(hidden, in_feature))
self.w2 = torch.nn.Parameter(torch.FloatTensor(out_feature, hidden))
self.reset_parameters()
def reset_parameters(self):
"""
Use xavier_normal method to initialize parameters.
"""
nn.init.xavier_normal_(self.w1)
nn.init.xavier_normal_(self.w2)
def forward(self, input_0):
primals_1 = self.w1
primals_3 = self.w2
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0], output[1]
| MatthewMasters/grover | SelfAttention | false | 836 | [
"MIT"
] | 0 | 737a340754bc4c63134ef84019a0a84023fd69a3 | https://github.com/MatthewMasters/grover/tree/737a340754bc4c63134ef84019a0a84023fd69a3 | import torch
from torch import nn as nn
import torch.utils.data.distributed
class Model(nn.Module):
"""
Self SelfAttention Layer
Given $X\\in \\mathbb{R}^{n imes in_feature}$, the attention is calculated by: $a=Softmax(W_2tanh(W_1X))$, where
$W_1 \\in \\mathbb{R}^{hidden imes in_feature}$, $W_2 \\in \\mathbb{R}^{out_feature imes hidden}$.
The final output is: $out=aX$, which is unrelated with input $n$.
"""
def __init__(self, *, hidden, in_feature, out_feature):
"""
The init function.
:param hidden: the hidden dimension, can be viewed as the number of experts.
:param in_feature: the input feature dimension.
:param out_feature: the output feature dimension.
"""
super().__init__()
self.w1 = torch.nn.Parameter(torch.FloatTensor(hidden, in_feature))
self.w2 = torch.nn.Parameter(torch.FloatTensor(out_feature, hidden))
self.reset_parameters()
def reset_parameters(self):
"""
Use xavier_normal method to initialize parameters.
"""
nn.init.xavier_normal_(self.w1)
nn.init.xavier_normal_(self.w2)
def forward(self, X):
"""
The forward function.
:param X: The input feature map. $X \\in \\mathbb{R}^{n imes in_feature}$.
:return: The final embeddings and attention matrix.
"""
x = torch.tanh(torch.matmul(self.w1, X.transpose(1, 0)))
x = torch.matmul(self.w2, x)
attn = torch.nn.functional.softmax(x, dim=-1)
x = torch.matmul(attn, X)
return x, attn
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
DWConv | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nk/cnkqgsnerxuygjzfsg7bb62yz5o7s7mc4zdh3yqihlwokuyxwi6t.py
# Unsorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
triton_poi_fused_0 = async_compile.triton('triton_poi_fused_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4096, 4096], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 3072
xnumel = 4096
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 768
y1 = (yindex // 768)
tmp0 = tl.load(in_ptr0 + (x2 + (4096*y3)), None, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + (768*x2) + (3145728*y1)), tmp0, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/gd/cgdltxpi4qz3npc22hwzzjki3xywudrei6v5aadkwvmd56hvm73m.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# x => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [1, 1], [1, 1], False, [0, 0], 768), kwargs = {})
triton_poi_fused_convolution_1 = async_compile.triton('triton_poi_fused_convolution_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4096, 4096], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 3072
xnumel = 4096
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y0 = yindex % 768
y1 = (yindex // 768)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (768*x2) + (3145728*y1)), None, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (y0), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + (4096*y3)), tmp2, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (768, 1, 3, 3), (9, 9, 3, 1))
assert_size_stride(primals_2, (768, ), (1, ))
assert_size_stride(primals_3, (4, 768, 64, 64), (3145728, 4096, 64, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 768, 64, 64), (3145728, 1, 49152, 768), torch.float32)
# Unsorted Source Nodes: [], Original ATen: []
stream0 = get_raw_stream(0)
triton_poi_fused_0.run(primals_3, buf0, 3072, 4096, grid=grid(3072, 4096), stream=stream0)
del primals_3
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(buf0, primals_1, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=768, bias=None)
assert_size_stride(buf1, (4, 768, 64, 64), (3145728, 1, 49152, 768))
buf2 = empty_strided_cuda((4, 768, 64, 64), (3145728, 4096, 64, 1), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
triton_poi_fused_convolution_1.run(buf1, primals_2, buf2, 3072, 4096, grid=grid(3072, 4096), stream=stream0)
del buf1
del primals_2
return (buf2, primals_1, buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((768, 1, 3, 3), (9, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((768, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 768, 64, 64), (3145728, 4096, 64, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class DWConv(nn.Module):
def __init__(self, dim=768):
super(DWConv, self).__init__()
self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
def forward(self, x):
x = self.dwconv(x)
return x
def get_inputs():
return [torch.rand([4, 768, 64, 64])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 768
y1 = yindex // 768
tmp0 = tl.load(in_ptr0 + (x2 + 4096 * y3), None, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 768 * x2 + 3145728 * y1), tmp0, None)
@triton.jit
def triton_poi_fused_convolution_1(in_ptr0, in_ptr1, out_ptr0, ynumel,
xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y0 = yindex % 768
y1 = yindex // 768
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 768 * x2 + 3145728 * y1), None,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4096 * y3), tmp2, None)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (768, 1, 3, 3), (9, 9, 3, 1))
assert_size_stride(primals_2, (768,), (1,))
assert_size_stride(primals_3, (4, 768, 64, 64), (3145728, 4096, 64, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 768, 64, 64), (3145728, 1, 49152, 768
), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(3072, 4096)](primals_3, buf0, 3072, 4096,
XBLOCK=32, YBLOCK=32, num_warps=4, num_stages=1)
del primals_3
buf1 = extern_kernels.convolution(buf0, primals_1, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=768, bias=None)
assert_size_stride(buf1, (4, 768, 64, 64), (3145728, 1, 49152, 768))
buf2 = empty_strided_cuda((4, 768, 64, 64), (3145728, 4096, 64, 1),
torch.float32)
triton_poi_fused_convolution_1[grid(3072, 4096)](buf1, primals_2,
buf2, 3072, 4096, XBLOCK=64, YBLOCK=64, num_warps=8, num_stages=1)
del buf1
del primals_2
return buf2, primals_1, buf0
class DWConvNew(nn.Module):
def __init__(self, dim=768):
super(DWConvNew, self).__init__()
self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
def forward(self, input_0):
primals_1 = self.dwconv.weight
primals_2 = self.dwconv.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| LSH9832/MyPythonModules | DWConv | false | 837 | [
"MIT"
] | 0 | 442566a0fbd6ebe2bc20b6914686a1e2663d10c0 | https://github.com/LSH9832/MyPythonModules/tree/442566a0fbd6ebe2bc20b6914686a1e2663d10c0 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, dim=768):
super().__init__()
self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
def forward(self, x):
x = self.dwconv(x)
return x
def get_inputs():
return [torch.rand([4, 768, 64, 64])]
def get_init_inputs():
return []
|
Squareplus | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/s7/cs7rg7ti3jxwvhvqtykhnyfxbmihzp32355jskzvodnvizlhto2c.py
# Topologically Sorted Source Nodes: [square, add, sqrt, add_1, truediv], Original ATen: [aten.pow, aten.add, aten.sqrt, aten.div]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# sqrt => sqrt
# square => pow_1
# truediv => div
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%arg0_1, 2), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%pow_1, 4), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%add,), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%arg0_1, %sqrt), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%add_1, 2), kwargs = {})
triton_poi_fused_add_div_pow_sqrt_0 = async_compile.triton('triton_poi_fused_add_div_pow_sqrt_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_pow_sqrt_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_pow_sqrt_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tmp0 * tmp0
tmp2 = 4.0
tmp3 = tmp1 + tmp2
tmp4 = libdevice.sqrt(tmp3)
tmp5 = tmp0 + tmp4
tmp6 = 0.5
tmp7 = tmp5 * tmp6
tl.store(out_ptr0 + (x0), tmp7, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [square, add, sqrt, add_1, truediv], Original ATen: [aten.pow, aten.add, aten.sqrt, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_pow_sqrt_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch as t
import torch.nn as nn
class Squareplus(nn.Module):
def __init__(self, a=2):
super().__init__()
self.a = a
def forward(self, x):
"""The 'squareplus' activation function: has very similar properties to
softplus, but is far cheaper computationally.
- squareplus(0) = 1 (softplus(0) = ln 2)
- gradient diminishes more slowly for negative inputs.
- ReLU = (x + sqrt(x^2))/2
- 'squareplus' becomes smoother with higher 'a'
"""
return (x + t.sqrt(t.square(x) + self.a * self.a)) / 2
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_div_pow_sqrt_0(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tmp0 * tmp0
tmp2 = 4.0
tmp3 = tmp1 + tmp2
tmp4 = libdevice.sqrt(tmp3)
tmp5 = tmp0 + tmp4
tmp6 = 0.5
tmp7 = tmp5 * tmp6
tl.store(out_ptr0 + x0, tmp7, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_pow_sqrt_0[grid(256)](arg0_1, buf0, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class SquareplusNew(nn.Module):
def __init__(self, a=2):
super().__init__()
self.a = a
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MaximeRobeyns/BDRL | Squareplus | false | 838 | [
"Apache-2.0"
] | 0 | 55e295d5aaca6745d35525114b472ad118c14a6d | https://github.com/MaximeRobeyns/BDRL/tree/55e295d5aaca6745d35525114b472ad118c14a6d | import torch
import torch as t
import torch.nn as nn
class Model(nn.Module):
def __init__(self, a=2):
super().__init__()
self.a = a
def forward(self, x):
"""The 'squareplus' activation function: has very similar properties to
softplus, but is far cheaper computationally.
- squareplus(0) = 1 (softplus(0) = ln 2)
- gradient diminishes more slowly for negative inputs.
- ReLU = (x + sqrt(x^2))/2
- 'squareplus' becomes smoother with higher 'a'
"""
return (x + t.sqrt(t.square(x) + self.a * self.a)) / 2
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
DiceBCELoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xa/cxaldjb3lyrttpd6oyeccpxx6xt35jg65irbgseyklqochxbvrit.py
# Topologically Sorted Source Nodes: [BCE, mul, intersection, mul_1, add, sum_2, sum_3, add_1, add_2, truediv, dice_loss, Dice_BCE], Original ATen: [aten.binary_cross_entropy, aten.mul, aten.sum, aten.add, aten.div, aten.rsub]
# Source node to ATen node mapping:
# BCE => full_default, full_default_1, log, log1p, maximum, maximum_1, mean, mul_2, mul_3, neg, sub_1, sub_2
# Dice_BCE => add_3
# add => add
# add_1 => add_1
# add_2 => add_2
# dice_loss => sub
# intersection => sum_1
# mul => mul
# mul_1 => mul_1
# sum_2 => sum_2
# sum_3 => sum_3
# truediv => div
# Graph fragment:
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_1, 1), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%view,), kwargs = {})
# %log1p : [num_users=1] = call_function[target=torch.ops.aten.log1p.default](args = (%neg,), kwargs = {})
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], -100), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %maximum : [num_users=1] = call_function[target=torch.ops.aten.maximum.default](args = (%log1p, %full_default), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_1, %maximum), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%view,), kwargs = {})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], -100), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %maximum_1 : [num_users=1] = call_function[target=torch.ops.aten.maximum.default](args = (%log, %full_default_1), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, %maximum_1), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_2, %mul_3), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%sub_2,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view, %view_1), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%mul,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sum_1, 2.0), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, 1), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%view,), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%view_1,), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sum_2, %sum_3), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_1, 1), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%add, %add_2), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %div), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean, %sub), kwargs = {})
triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0 = async_compile.triton('triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 2, 'num_reduction': 4, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp3 = tl.load(in_ptr1 + (r0), None)
tmp1 = 1.0
tmp2 = tmp0 - tmp1
tmp4 = tl.sigmoid(tmp3)
tmp5 = -tmp4
tmp6 = libdevice.log1p(tmp5)
tmp7 = -100.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp2 * tmp8
tmp10 = tl_math.log(tmp4)
tmp11 = triton_helpers.maximum(tmp10, tmp7)
tmp12 = tmp0 * tmp11
tmp13 = tmp9 - tmp12
tmp14 = tl.broadcast_to(tmp13, [RBLOCK])
tmp16 = triton_helpers.promote_to_tensor(tl.sum(tmp14, 0))
tmp17 = tmp4 * tmp0
tmp18 = tl.broadcast_to(tmp17, [RBLOCK])
tmp20 = triton_helpers.promote_to_tensor(tl.sum(tmp18, 0))
tmp21 = tl.broadcast_to(tmp4, [RBLOCK])
tmp23 = triton_helpers.promote_to_tensor(tl.sum(tmp21, 0))
tmp24 = tl.broadcast_to(tmp0, [RBLOCK])
tmp26 = triton_helpers.promote_to_tensor(tl.sum(tmp24, 0))
tmp27 = 256.0
tmp28 = tmp16 / tmp27
tmp29 = 2.0
tmp30 = tmp20 * tmp29
tmp31 = tmp30 + tmp1
tmp32 = tmp23 + tmp26
tmp33 = tmp32 + tmp1
tmp34 = tmp31 / tmp33
tmp35 = tmp1 - tmp34
tmp36 = tmp28 + tmp35
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp36, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf4 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [BCE, mul, intersection, mul_1, add, sum_2, sum_3, add_1, add_2, truediv, dice_loss, Dice_BCE], Original ATen: [aten.binary_cross_entropy, aten.mul, aten.sum, aten.add, aten.div, aten.rsub]
stream0 = get_raw_stream(0)
triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0.run(buf4, arg1_1, arg0_1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
return (buf4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
from torch.nn import functional as F
class DiceBCELoss(nn.Module):
def __init__(self, weight=None, size_average=True):
super(DiceBCELoss, self).__init__()
def forward(self, inputs, targets, smooth=1):
inputs = torch.sigmoid(inputs)
inputs = inputs.view(-1)
targets = targets.view(-1)
intersection = (inputs * targets).sum()
dice_loss = 1 - (2.0 * intersection + smooth) / (inputs.sum() +
targets.sum() + smooth)
BCE = F.binary_cross_entropy(inputs, targets, reduction='mean')
Dice_BCE = BCE + dice_loss
return Dice_BCE
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0(in_out_ptr0,
in_ptr0, in_ptr1, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp3 = tl.load(in_ptr1 + r0, None)
tmp1 = 1.0
tmp2 = tmp0 - tmp1
tmp4 = tl.sigmoid(tmp3)
tmp5 = -tmp4
tmp6 = libdevice.log1p(tmp5)
tmp7 = -100.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp2 * tmp8
tmp10 = tl_math.log(tmp4)
tmp11 = triton_helpers.maximum(tmp10, tmp7)
tmp12 = tmp0 * tmp11
tmp13 = tmp9 - tmp12
tmp14 = tl.broadcast_to(tmp13, [RBLOCK])
tmp16 = triton_helpers.promote_to_tensor(tl.sum(tmp14, 0))
tmp17 = tmp4 * tmp0
tmp18 = tl.broadcast_to(tmp17, [RBLOCK])
tmp20 = triton_helpers.promote_to_tensor(tl.sum(tmp18, 0))
tmp21 = tl.broadcast_to(tmp4, [RBLOCK])
tmp23 = triton_helpers.promote_to_tensor(tl.sum(tmp21, 0))
tmp24 = tl.broadcast_to(tmp0, [RBLOCK])
tmp26 = triton_helpers.promote_to_tensor(tl.sum(tmp24, 0))
tmp27 = 256.0
tmp28 = tmp16 / tmp27
tmp29 = 2.0
tmp30 = tmp20 * tmp29
tmp31 = tmp30 + tmp1
tmp32 = tmp23 + tmp26
tmp33 = tmp32 + tmp1
tmp34 = tmp31 / tmp33
tmp35 = tmp1 - tmp34
tmp36 = tmp28 + tmp35
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp36, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf4 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_binary_cross_entropy_div_mul_rsub_sum_0[grid(1)](
buf4, arg1_1, arg0_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf4,
class DiceBCELossNew(nn.Module):
def __init__(self, weight=None, size_average=True):
super(DiceBCELossNew, self).__init__()
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| Michaelistaken/PathPretrain | DiceBCELoss | false | 839 | [
"MIT"
] | 0 | 650b7eb02e67f6d864d81808eb7230c48fe6946a | https://github.com/Michaelistaken/PathPretrain/tree/650b7eb02e67f6d864d81808eb7230c48fe6946a | import torch
from torch import nn
from torch.nn import functional as F
class Model(nn.Module):
def __init__(self, weight=None, size_average=True):
super().__init__()
def forward(self, inputs, targets, smooth=1):
inputs = torch.sigmoid(inputs)
inputs = inputs.view(-1)
targets = targets.view(-1)
intersection = (inputs * targets).sum()
dice_loss = 1 - (2.0 * intersection + smooth) / (inputs.sum() +
targets.sum() + smooth)
BCE = F.binary_cross_entropy(inputs, targets, reduction='mean')
Dice_BCE = BCE + dice_loss
return Dice_BCE
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
DiceLoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/sc/csc4e2vzzrvn5vasjwt33h5twu2lbmnykvnkrdwx4vjrfjftxot7.py
# Topologically Sorted Source Nodes: [mul, intersection, mul_1, add, sum_2, sum_3, add_1, add_2, dice, sub], Original ATen: [aten.mul, aten.sum, aten.add, aten.div, aten.rsub]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# add_2 => add_2
# dice => div
# intersection => sum_1
# mul => mul
# mul_1 => mul_1
# sub => sub
# sum_2 => sum_2
# sum_3 => sum_3
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg0_1, %arg1_1), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%mul,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sum_1, 2.0), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, 1e-05), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%arg0_1,), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%arg1_1,), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sum_2, %sum_3), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_1, 1e-05), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%add, %add_2), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %div), kwargs = {})
triton_per_fused_add_div_mul_rsub_sum_0 = async_compile.triton('triton_per_fused_add_div_mul_rsub_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_div_mul_rsub_sum_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 2, 'num_reduction': 3, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_div_mul_rsub_sum_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = tl.load(in_ptr1 + (r0), None)
tmp2 = tmp0 * tmp1
tmp3 = tl.broadcast_to(tmp2, [RBLOCK])
tmp5 = triton_helpers.promote_to_tensor(tl.sum(tmp3, 0))
tmp6 = tl.broadcast_to(tmp0, [RBLOCK])
tmp8 = triton_helpers.promote_to_tensor(tl.sum(tmp6, 0))
tmp9 = tl.broadcast_to(tmp1, [RBLOCK])
tmp11 = triton_helpers.promote_to_tensor(tl.sum(tmp9, 0))
tmp12 = 2.0
tmp13 = tmp5 * tmp12
tmp14 = 1e-05
tmp15 = tmp13 + tmp14
tmp16 = tmp8 + tmp11
tmp17 = tmp16 + tmp14
tmp18 = tmp15 / tmp17
tmp19 = 1.0
tmp20 = tmp19 - tmp18
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp20, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf3 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [mul, intersection, mul_1, add, sum_2, sum_3, add_1, add_2, dice, sub], Original ATen: [aten.mul, aten.sum, aten.add, aten.div, aten.rsub]
stream0 = get_raw_stream(0)
triton_per_fused_add_div_mul_rsub_sum_0.run(buf3, arg0_1, arg1_1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
return (buf3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class DiceLoss(nn.Module):
def __init__(self, weight=None, size_average=True):
super(DiceLoss, self).__init__()
def forward(self, inputs, targets):
intersection = (inputs * targets).sum()
dice = (2.0 * intersection + 1e-05) / (inputs.sum() + targets.sum() +
1e-05)
return 1 - dice
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_div_mul_rsub_sum_0(in_out_ptr0, in_ptr0, in_ptr1,
xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp2 = tmp0 * tmp1
tmp3 = tl.broadcast_to(tmp2, [RBLOCK])
tmp5 = triton_helpers.promote_to_tensor(tl.sum(tmp3, 0))
tmp6 = tl.broadcast_to(tmp0, [RBLOCK])
tmp8 = triton_helpers.promote_to_tensor(tl.sum(tmp6, 0))
tmp9 = tl.broadcast_to(tmp1, [RBLOCK])
tmp11 = triton_helpers.promote_to_tensor(tl.sum(tmp9, 0))
tmp12 = 2.0
tmp13 = tmp5 * tmp12
tmp14 = 1e-05
tmp15 = tmp13 + tmp14
tmp16 = tmp8 + tmp11
tmp17 = tmp16 + tmp14
tmp18 = tmp15 / tmp17
tmp19 = 1.0
tmp20 = tmp19 - tmp18
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp20, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf3 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_div_mul_rsub_sum_0[grid(1)](buf3, arg0_1,
arg1_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf3,
class DiceLossNew(nn.Module):
def __init__(self, weight=None, size_average=True):
super(DiceLossNew, self).__init__()
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| MohannadEhabBarakat/U-2-Net | DiceLoss | false | 840 | [
"Apache-2.0"
] | 0 | 89a4eba7a565e7afcd4ac04b11b55099ebef687c | https://github.com/MohannadEhabBarakat/U-2-Net/tree/89a4eba7a565e7afcd4ac04b11b55099ebef687c | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, weight=None, size_average=True):
super().__init__()
def forward(self, inputs, targets):
intersection = (inputs * targets).sum()
dice = (2.0 * intersection + 1e-05) / (inputs.sum() + targets.sum() +
1e-05)
return 1 - dice
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
MSELoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/fo/cfogynji4ibod7ws4da5mkh5phgwd7jmfcb3cuupcagtpbr6jgm3.py
# Topologically Sorted Source Nodes: [loss, loss_1, loss_2], Original ATen: [aten.mse_loss, aten.mean, aten.mul]
# Source node to ATen node mapping:
# loss => mean, pow_1, sub
# loss_1 => mean_1
# loss_2 => mul
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg0_1, %arg1_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%pow_1,), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%mean,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mean_1, 1.0), kwargs = {})
triton_per_fused_mean_mse_loss_mul_0 = async_compile.triton('triton_per_fused_mean_mse_loss_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_mean_mse_loss_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 2, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_mean_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = tl.load(in_ptr1 + (r0), None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp7 = 256.0
tmp8 = tmp6 / tmp7
tmp9 = 1.0
tmp10 = tmp8 / tmp9
tmp11 = tmp10 * tmp9
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp11, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [loss, loss_1, loss_2], Original ATen: [aten.mse_loss, aten.mean, aten.mul]
stream0 = get_raw_stream(0)
triton_per_fused_mean_mse_loss_mul_0.run(buf1, arg0_1, arg1_1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
return (buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import functools
import torch
import torch.nn as nn
import torch.nn.functional as F
def reduce_loss(loss, reduction):
"""Reduce loss as specified.
Args:
loss (Tensor): Elementwise loss tensor.
reduction (str): Options are "none", "mean" and "sum".
Return:
Tensor: Reduced loss tensor.
"""
reduction_enum = F._Reduction.get_enum(reduction)
if reduction_enum == 0:
return loss
elif reduction_enum == 1:
return loss.mean()
elif reduction_enum == 2:
return loss.sum()
def weight_reduce_loss(loss, weight=None, reduction='mean', avg_factor=None):
"""Apply element-wise weight and reduce loss.
Args:
loss (Tensor): Element-wise loss.
weight (Tensor): Element-wise weights.
reduction (str): Same as built-in losses of PyTorch.
avg_factor (float): Average factor when computing the mean of losses.
Returns:
Tensor: Processed loss values.
"""
if weight is not None:
loss = loss * weight
if avg_factor is None:
loss = reduce_loss(loss, reduction)
elif reduction == 'mean':
loss = loss.sum() / avg_factor
elif reduction != 'none':
raise ValueError('avg_factor can not be used with reduction="sum"')
return loss
def weighted_loss(loss_func):
"""Create a weighted version of a given loss function.
To use this decorator, the loss function must have the signature like
``loss_func(pred, target, **kwargs)``. The function only needs to compute
element-wise loss without any reduction. This decorator will add weight
and reduction arguments to the function. The decorated function will have
the signature like ``loss_func(pred, target, weight=None, reduction='mean',
avg_factor=None, **kwargs)``.
:Example:
>>> import torch
>>> @weighted_loss
>>> def l1_loss(pred, target):
>>> return (pred - target).abs()
>>> pred = torch.Tensor([0, 2, 3])
>>> target = torch.Tensor([1, 1, 1])
>>> weight = torch.Tensor([1, 0, 1])
>>> l1_loss(pred, target)
tensor(1.3333)
>>> l1_loss(pred, target, weight)
tensor(1.)
>>> l1_loss(pred, target, reduction='none')
tensor([1., 1., 2.])
>>> l1_loss(pred, target, weight, avg_factor=2)
tensor(1.5000)
"""
@functools.wraps(loss_func)
def wrapper(pred, target, weight=None, reduction='mean', avg_factor=
None, **kwargs):
loss = loss_func(pred, target, **kwargs)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
return wrapper
@weighted_loss
def mse_loss(pred, target, reduction='mean'):
"""Warpper of mse loss."""
return F.mse_loss(pred, target, reduction)
class MSELoss(nn.Module):
"""MSELoss.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, reduction='mean', loss_weight=1.0):
super().__init__()
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self, pred, target, weight=None):
"""Forward function of loss.
Args:
pred (torch.Tensor): The prediction.
target (torch.Tensor): The learning target of the prediction.
weight (torch.Tensor, optional): Weight of the loss for each
prediction. Defaults to None.
avg_factor (int, optional): Average factor that is used to average
the loss. Defaults to None.
Returns:
torch.Tensor: The calculated loss
"""
loss = self.loss_weight * mse_loss(pred, target, reduction=self.
reduction)
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import functools
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_mean_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1,
xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp7 = 256.0
tmp8 = tmp6 / tmp7
tmp9 = 1.0
tmp10 = tmp8 / tmp9
tmp11 = tmp10 * tmp9
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp11, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_mean_mse_loss_mul_0[grid(1)](buf1, arg0_1, arg1_1,
1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf1,
def reduce_loss(loss, reduction):
"""Reduce loss as specified.
Args:
loss (Tensor): Elementwise loss tensor.
reduction (str): Options are "none", "mean" and "sum".
Return:
Tensor: Reduced loss tensor.
"""
reduction_enum = F._Reduction.get_enum(reduction)
if reduction_enum == 0:
return loss
elif reduction_enum == 1:
return loss.mean()
elif reduction_enum == 2:
return loss.sum()
def weight_reduce_loss(loss, weight=None, reduction='mean', avg_factor=None):
"""Apply element-wise weight and reduce loss.
Args:
loss (Tensor): Element-wise loss.
weight (Tensor): Element-wise weights.
reduction (str): Same as built-in losses of PyTorch.
avg_factor (float): Average factor when computing the mean of losses.
Returns:
Tensor: Processed loss values.
"""
if weight is not None:
loss = loss * weight
if avg_factor is None:
loss = reduce_loss(loss, reduction)
elif reduction == 'mean':
loss = loss.sum() / avg_factor
elif reduction != 'none':
raise ValueError('avg_factor can not be used with reduction="sum"')
return loss
def weighted_loss(loss_func):
"""Create a weighted version of a given loss function.
To use this decorator, the loss function must have the signature like
``loss_func(pred, target, **kwargs)``. The function only needs to compute
element-wise loss without any reduction. This decorator will add weight
and reduction arguments to the function. The decorated function will have
the signature like ``loss_func(pred, target, weight=None, reduction='mean',
avg_factor=None, **kwargs)``.
:Example:
>>> import torch
>>> @weighted_loss
>>> def l1_loss(pred, target):
>>> return (pred - target).abs()
>>> pred = torch.Tensor([0, 2, 3])
>>> target = torch.Tensor([1, 1, 1])
>>> weight = torch.Tensor([1, 0, 1])
>>> l1_loss(pred, target)
tensor(1.3333)
>>> l1_loss(pred, target, weight)
tensor(1.)
>>> l1_loss(pred, target, reduction='none')
tensor([1., 1., 2.])
>>> l1_loss(pred, target, weight, avg_factor=2)
tensor(1.5000)
"""
@functools.wraps(loss_func)
def wrapper(pred, target, weight=None, reduction='mean', avg_factor=
None, **kwargs):
loss = loss_func(pred, target, **kwargs)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
return wrapper
@weighted_loss
def mse_loss(pred, target, reduction='mean'):
"""Warpper of mse loss."""
return F.mse_loss(pred, target, reduction)
class MSELossNew(nn.Module):
"""MSELoss.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, reduction='mean', loss_weight=1.0):
super().__init__()
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| Min-Sheng/mmregression | MSELoss | false | 841 | [
"Apache-2.0"
] | 0 | 6d70383d89ccb3dea7f425b665c2a184d014a99f | https://github.com/Min-Sheng/mmregression/tree/6d70383d89ccb3dea7f425b665c2a184d014a99f | import functools
import torch
import torch.nn as nn
import torch.nn.functional as F
def reduce_loss(loss, reduction):
"""Reduce loss as specified.
Args:
loss (Tensor): Elementwise loss tensor.
reduction (str): Options are "none", "mean" and "sum".
Return:
Tensor: Reduced loss tensor.
"""
reduction_enum = F._Reduction.get_enum(reduction)
if reduction_enum == 0:
return loss
elif reduction_enum == 1:
return loss.mean()
elif reduction_enum == 2:
return loss.sum()
def weight_reduce_loss(loss, weight=None, reduction='mean', avg_factor=None):
"""Apply element-wise weight and reduce loss.
Args:
loss (Tensor): Element-wise loss.
weight (Tensor): Element-wise weights.
reduction (str): Same as built-in losses of PyTorch.
avg_factor (float): Average factor when computing the mean of losses.
Returns:
Tensor: Processed loss values.
"""
if weight is not None:
loss = loss * weight
if avg_factor is None:
loss = reduce_loss(loss, reduction)
elif reduction == 'mean':
loss = loss.sum() / avg_factor
elif reduction != 'none':
raise ValueError('avg_factor can not be used with reduction="sum"')
return loss
def weighted_loss(loss_func):
"""Create a weighted version of a given loss function.
To use this decorator, the loss function must have the signature like
``loss_func(pred, target, **kwargs)``. The function only needs to compute
element-wise loss without any reduction. This decorator will add weight
and reduction arguments to the function. The decorated function will have
the signature like ``loss_func(pred, target, weight=None, reduction='mean',
avg_factor=None, **kwargs)``.
:Example:
>>> import torch
>>> @weighted_loss
>>> def l1_loss(pred, target):
>>> return (pred - target).abs()
>>> pred = torch.Tensor([0, 2, 3])
>>> target = torch.Tensor([1, 1, 1])
>>> weight = torch.Tensor([1, 0, 1])
>>> l1_loss(pred, target)
tensor(1.3333)
>>> l1_loss(pred, target, weight)
tensor(1.)
>>> l1_loss(pred, target, reduction='none')
tensor([1., 1., 2.])
>>> l1_loss(pred, target, weight, avg_factor=2)
tensor(1.5000)
"""
@functools.wraps(loss_func)
def wrapper(pred, target, weight=None, reduction='mean', avg_factor=
None, **kwargs):
loss = loss_func(pred, target, **kwargs)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
return wrapper
@weighted_loss
def mse_loss(pred, target, reduction='mean'):
"""Warpper of mse loss."""
return F.mse_loss(pred, target, reduction)
class Model(nn.Module):
"""MSELoss.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, reduction='mean', loss_weight=1.0):
super().__init__()
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self, pred, target, weight=None):
"""Forward function of loss.
Args:
pred (torch.Tensor): The prediction.
target (torch.Tensor): The learning target of the prediction.
weight (torch.Tensor, optional): Weight of the loss for each
prediction. Defaults to None.
avg_factor (int, optional): Average factor that is used to average
the loss. Defaults to None.
Returns:
torch.Tensor: The calculated loss
"""
loss = self.loss_weight * mse_loss(pred, target, reduction=self.
reduction)
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
UpBlock | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ta/cta3qfkas5srst7l4ccnpj7oz7hel5h53n7tjqwjnpk4uswjkepy.py
# Topologically Sorted Source Nodes: [out, h0], Original ATen: [aten.convolution, aten._prelu_kernel]
# Source node to ATen node mapping:
# h0 => gt, mul, where
# out => convolution
# Graph fragment:
# %convolution : [num_users=4] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [4, 4], [2, 2], [1, 1], True, [0, 0], 1), kwargs = {})
# %gt : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%convolution, 0), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view, %convolution), kwargs = {})
# %where : [num_users=3] = call_function[target=torch.ops.aten.where.self](args = (%gt, %convolution, %mul), kwargs = {})
triton_poi_fused__prelu_kernel_convolution_0 = async_compile.triton('triton_poi_fused__prelu_kernel_convolution_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4096],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__prelu_kernel_convolution_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_0(in_out_ptr0, in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = (xindex // 256) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), None)
tmp1 = tl.load(in_ptr0 + (x1), None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (0))
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tl.store(in_out_ptr0 + (x3), tmp2, None)
tl.store(out_ptr0 + (x3), tmp8, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/v2/cv2gefdciotml3zwtkzv4ghtu2a4dbeoas7q3ue7dcfa4f2mizfk.py
# Topologically Sorted Source Nodes: [out_1, l0, sub], Original ATen: [aten.convolution, aten._prelu_kernel, aten.sub]
# Source node to ATen node mapping:
# l0 => gt_1, mul_1, where_1
# out_1 => convolution_1
# sub => sub
# Graph fragment:
# %convolution_1 : [num_users=4] = call_function[target=torch.ops.aten.convolution.default](args = (%where, %primals_5, %primals_6, [4, 4], [2, 2], [1, 1], False, [0, 0], 1), kwargs = {})
# %gt_1 : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%convolution_1, 0), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, %convolution_1), kwargs = {})
# %where_1 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%gt_1, %convolution_1, %mul_1), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%where_1, %primals_3), kwargs = {})
triton_poi_fused__prelu_kernel_convolution_sub_1 = async_compile.triton('triton_poi_fused__prelu_kernel_convolution_sub_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__prelu_kernel_convolution_sub_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_sub_1(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 16) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (0))
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + (x3), xmask)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 - tmp9
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
tl.store(out_ptr0 + (x3), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/vz/cvznbzzyrmqjqwdtg4sdvku7chdpvxprj52fd6jpud4fhhmvo2pr.py
# Topologically Sorted Source Nodes: [out_2, h1, add], Original ATen: [aten.convolution, aten._prelu_kernel, aten.add]
# Source node to ATen node mapping:
# add => add
# h1 => gt_2, mul_2, where_2
# out_2 => convolution_2
# Graph fragment:
# %convolution_2 : [num_users=4] = call_function[target=torch.ops.aten.convolution.default](args = (%sub, %primals_8, %primals_9, [4, 4], [2, 2], [1, 1], True, [0, 0], 1), kwargs = {})
# %gt_2 : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%convolution_2, 0), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_2, %convolution_2), kwargs = {})
# %where_2 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%gt_2, %convolution_2, %mul_2), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%where_2, %where), kwargs = {})
triton_poi_fused__prelu_kernel_add_convolution_2 = async_compile.triton('triton_poi_fused__prelu_kernel_add_convolution_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4096],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__prelu_kernel_add_convolution_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__prelu_kernel_add_convolution_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = (xindex // 256) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), None)
tmp1 = tl.load(in_ptr0 + (x1), None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (0))
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + (x3), None)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 + tmp9
tl.store(in_out_ptr0 + (x3), tmp2, None)
tl.store(out_ptr0 + (x3), tmp10, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (1, ), (1, ))
assert_size_stride(primals_5, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_6, (4, ), (1, ))
assert_size_stride(primals_7, (1, ), (1, ))
assert_size_stride(primals_8, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_9, (4, ), (1, ))
assert_size_stride(primals_10, (1, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(4, 4), padding=(2, 2), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 16, 16), (1024, 256, 16, 1))
buf1 = buf0; del buf0 # reuse
buf2 = empty_strided_cuda((4, 4, 16, 16), (1024, 256, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [out, h0], Original ATen: [aten.convolution, aten._prelu_kernel]
stream0 = get_raw_stream(0)
triton_poi_fused__prelu_kernel_convolution_0.run(buf1, primals_2, primals_4, buf2, 4096, grid=grid(4096), stream=stream0)
del primals_2
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.convolution]
buf3 = extern_kernels.convolution(buf2, primals_5, stride=(4, 4), padding=(2, 2), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 4, 4, 4), (64, 16, 4, 1))
buf4 = buf3; del buf3 # reuse
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_1, l0, sub], Original ATen: [aten.convolution, aten._prelu_kernel, aten.sub]
triton_poi_fused__prelu_kernel_convolution_sub_1.run(buf4, primals_6, primals_7, primals_3, buf5, 256, grid=grid(256), stream=stream0)
del primals_6
# Topologically Sorted Source Nodes: [out_2], Original ATen: [aten.convolution]
buf6 = extern_kernels.convolution(buf5, primals_8, stride=(4, 4), padding=(2, 2), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 4, 16, 16), (1024, 256, 16, 1))
buf7 = buf6; del buf6 # reuse
buf8 = empty_strided_cuda((4, 4, 16, 16), (1024, 256, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_2, h1, add], Original ATen: [aten.convolution, aten._prelu_kernel, aten.add]
triton_poi_fused__prelu_kernel_add_convolution_2.run(buf7, primals_9, primals_10, buf2, buf8, 4096, grid=grid(4096), stream=stream0)
del primals_9
return (buf8, primals_1, primals_3, primals_4, primals_5, primals_7, primals_8, primals_10, buf1, buf2, buf4, buf5, buf7, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 8, 8), (256, 64, 8, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, 4, 8, 8), (256, 64, 8, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 4, 8, 8), (256, 64, 8, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class ConvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=3, stride=1,
padding=1, bias=True, norm=None):
super(ConvBlock, self).__init__()
self.conv = nn.Conv2d(input_size, output_size, kernel_size, stride,
padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.conv(x))
else:
out = self.conv(x)
return self.act(out)
class DeconvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=4, stride=2,
padding=1, bias=True, activation='prelu', norm=None):
super(DeconvBlock, self).__init__()
self.deconv = nn.ConvTranspose2d(input_size, output_size,
kernel_size, stride, padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.activation = activation
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.deconv(x))
else:
out = self.deconv(x)
if self.activation is not None:
return self.act(out)
else:
return out
class UpBlock(nn.Module):
def __init__(self, num_filter, kernel_size=8, stride=4, padding=2, bias
=True, activation='prelu', norm=None):
super(UpBlock, self).__init__()
self.up_conv1 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
self.up_conv2 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
self.up_conv3 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
def forward(self, x):
h0 = self.up_conv1(x)
l0 = self.up_conv2(h0)
h1 = self.up_conv3(l0 - x)
return h1 + h0
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'num_filter': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_0(in_out_ptr0, in_ptr0,
in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 256 % 4
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tl.store(in_out_ptr0 + x3, tmp2, None)
tl.store(out_ptr0 + x3, tmp8, None)
@triton.jit
def triton_poi_fused__prelu_kernel_convolution_sub_1(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + x3, xmask)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 - tmp9
tl.store(in_out_ptr0 + x3, tmp2, xmask)
tl.store(out_ptr0 + x3, tmp10, xmask)
@triton.jit
def triton_poi_fused__prelu_kernel_add_convolution_2(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 256 % 4
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp9 = tl.load(in_ptr2 + x3, None)
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp7 = tmp6 * tmp2
tmp8 = tl.where(tmp4, tmp2, tmp7)
tmp10 = tmp8 + tmp9
tl.store(in_out_ptr0 + x3, tmp2, None)
tl.store(out_ptr0 + x3, tmp10, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (1,), (1,))
assert_size_stride(primals_5, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (1,), (1,))
assert_size_stride(primals_8, (4, 4, 8, 8), (256, 64, 8, 1))
assert_size_stride(primals_9, (4,), (1,))
assert_size_stride(primals_10, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(4,
4), padding=(2, 2), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 16, 16), (1024, 256, 16, 1))
buf1 = buf0
del buf0
buf2 = empty_strided_cuda((4, 4, 16, 16), (1024, 256, 16, 1), torch
.float32)
get_raw_stream(0)
triton_poi_fused__prelu_kernel_convolution_0[grid(4096)](buf1,
primals_2, primals_4, buf2, 4096, XBLOCK=256, num_warps=4,
num_stages=1)
del primals_2
buf3 = extern_kernels.convolution(buf2, primals_5, stride=(4, 4),
padding=(2, 2), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 4, 4, 4), (64, 16, 4, 1))
buf4 = buf3
del buf3
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__prelu_kernel_convolution_sub_1[grid(256)](buf4,
primals_6, primals_7, primals_3, buf5, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_6
buf6 = extern_kernels.convolution(buf5, primals_8, stride=(4, 4),
padding=(2, 2), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 4, 16, 16), (1024, 256, 16, 1))
buf7 = buf6
del buf6
buf8 = empty_strided_cuda((4, 4, 16, 16), (1024, 256, 16, 1), torch
.float32)
triton_poi_fused__prelu_kernel_add_convolution_2[grid(4096)](buf7,
primals_9, primals_10, buf2, buf8, 4096, XBLOCK=256, num_warps=
4, num_stages=1)
del primals_9
return (buf8, primals_1, primals_3, primals_4, primals_5, primals_7,
primals_8, primals_10, buf1, buf2, buf4, buf5, buf7)
class ConvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=3, stride=1,
padding=1, bias=True, norm=None):
super(ConvBlock, self).__init__()
self.conv = nn.Conv2d(input_size, output_size, kernel_size, stride,
padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.conv(x))
else:
out = self.conv(x)
return self.act(out)
class DeconvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=4, stride=2,
padding=1, bias=True, activation='prelu', norm=None):
super(DeconvBlock, self).__init__()
self.deconv = nn.ConvTranspose2d(input_size, output_size,
kernel_size, stride, padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.activation = activation
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.deconv(x))
else:
out = self.deconv(x)
if self.activation is not None:
return self.act(out)
else:
return out
class UpBlockNew(nn.Module):
def __init__(self, num_filter, kernel_size=8, stride=4, padding=2, bias
=True, activation='prelu', norm=None):
super(UpBlockNew, self).__init__()
self.up_conv1 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
self.up_conv2 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
self.up_conv3 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
def forward(self, input_0):
primals_1 = self.up_conv1.deconv.weight
primals_2 = self.up_conv1.deconv.bias
primals_4 = self.up_conv1.act.weight
primals_5 = self.up_conv2.conv.weight
primals_6 = self.up_conv2.conv.bias
primals_7 = self.up_conv2.act.weight
primals_8 = self.up_conv3.deconv.weight
primals_9 = self.up_conv3.deconv.bias
primals_10 = self.up_conv3.act.weight
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return output[0]
| MatusBako/MakeFacesGreatAgain | UpBlock | false | 842 | [
"MIT"
] | 0 | e4941a8460db79dec566ed02d4b23eafb416a6db | https://github.com/MatusBako/MakeFacesGreatAgain/tree/e4941a8460db79dec566ed02d4b23eafb416a6db | import torch
import torch.nn as nn
class ConvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=3, stride=1,
padding=1, bias=True, norm=None):
super().__init__()
self.conv = nn.Conv2d(input_size, output_size, kernel_size, stride,
padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.conv(x))
else:
out = self.conv(x)
return self.act(out)
class DeconvBlock(nn.Module):
def __init__(self, input_size, output_size, kernel_size=4, stride=2,
padding=1, bias=True, activation='prelu', norm=None):
super().__init__()
self.deconv = nn.ConvTranspose2d(input_size, output_size,
kernel_size, stride, padding, bias=bias)
self.norm = norm
if self.norm == 'batch':
self.bn = nn.BatchNorm2d(output_size)
elif self.norm == 'instance':
self.bn = nn.InstanceNorm2d(output_size)
self.activation = activation
self.act = nn.PReLU()
def forward(self, x):
if self.norm is not None:
out = self.bn(self.deconv(x))
else:
out = self.deconv(x)
if self.activation is not None:
return self.act(out)
else:
return out
class Model(nn.Module):
def __init__(self, num_filter, kernel_size=8, stride=4, padding=2, bias
=True, activation='prelu', norm=None):
super().__init__()
self.up_conv1 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
self.up_conv2 = ConvBlock(num_filter, num_filter, kernel_size,
stride, padding, norm=None)
self.up_conv3 = DeconvBlock(num_filter, num_filter, kernel_size,
stride, padding, activation=activation, norm=None)
def forward(self, x):
h0 = self.up_conv1(x)
l0 = self.up_conv2(h0)
h1 = self.up_conv3(l0 - x)
return h1 + h0
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
AdaptiveAveragePooling | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/sq/csqnmu3e2m5az5dh7sty7sp2hq3r2hrzddgkwp2u7fhw4rm6niet.py
# Topologically Sorted Source Nodes: [outs], Original ATen: [aten._adaptive_avg_pool2d]
# Source node to ATen node mapping:
# outs => _adaptive_avg_pool2d
# Graph fragment:
# %_adaptive_avg_pool2d : [num_users=1] = call_function[target=torch.ops.aten._adaptive_avg_pool2d.default](args = (%arg0_1, [5, 4]), kwargs = {})
triton_poi_fused__adaptive_avg_pool2d_0 = async_compile.triton('triton_poi_fused__adaptive_avg_pool2d_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__adaptive_avg_pool2d_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__adaptive_avg_pool2d_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 320
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4) % 5
x0 = xindex % 4
x2 = (xindex // 20)
x4 = xindex
tmp0 = ((4*x1) // 5)
tmp1 = ((8 + (4*x1)) // 5)
tmp2 = tmp0 < tmp1
tmp3 = x0
tmp4 = 1 + x0
tmp5 = tmp3 < tmp4
tmp6 = tmp2 & tmp5
tmp7 = tl.load(in_ptr0 + (x0 + (4*((4*x1) // 5)) + (16*x2)), tmp6 & xmask, other=0.0)
tmp8 = tmp4 < tmp4
tmp9 = tmp2 & tmp8
tmp10 = tl.load(in_ptr0 + (1 + x0 + (4*((4*x1) // 5)) + (16*x2)), tmp9 & xmask, other=0.0)
tmp11 = tmp10 + tmp7
tmp12 = 1 + ((4*x1) // 5)
tmp13 = tmp12 < tmp1
tmp14 = tmp13 & tmp5
tmp15 = tl.load(in_ptr0 + (4 + x0 + (4*((4*x1) // 5)) + (16*x2)), tmp14 & xmask, other=0.0)
tmp16 = tmp15 + tmp11
tmp17 = tmp13 & tmp8
tmp18 = tl.load(in_ptr0 + (5 + x0 + (4*((4*x1) // 5)) + (16*x2)), tmp17 & xmask, other=0.0)
tmp19 = tmp18 + tmp16
tmp20 = 1.0
tmp21 = tl.full(tmp20.shape, 0.0, tmp20.dtype)
tmp22 = tl.where(tmp6, tmp20, tmp21)
tmp23 = tl.where(tmp9, tmp20, tmp21)
tmp24 = tmp23 + tmp22
tmp25 = tl.where(tmp14, tmp20, tmp21)
tmp26 = tmp25 + tmp24
tmp27 = tl.where(tmp17, tmp20, tmp21)
tmp28 = tmp27 + tmp26
tmp29 = tmp19 / tmp28
tl.store(out_ptr0 + (x4), tmp29, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 5, 4), (80, 20, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [outs], Original ATen: [aten._adaptive_avg_pool2d]
stream0 = get_raw_stream(0)
triton_poi_fused__adaptive_avg_pool2d_0.run(arg0_1, buf0, 320, grid=grid(320), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class AdaptiveAveragePooling(nn.Module):
"""Adaptive Pooling neck.
Args:
dim (int): Dimensions of each sample channel, can be one of {1, 2, 3}.
Default: 2
output_size (int | tuple): output size,
If dim equals to 1: output_size is a single integer.
Else, if output_size is a tuple of integers with the length of dim.
Default: (5, 4)
"""
def __init__(self, dim=2, output_size=(5, 4)):
super(AdaptiveAveragePooling, self).__init__()
assert dim in [1, 2, 3
], f'AdaptiveAveragePooling dim only support {1, 2, 3}, get {dim} instead.'
if dim == 1:
assert isinstance(output_size, int)
self.aap = nn.AdaptiveAvgPool1d(output_size)
elif dim == 2:
assert isinstance(output_size, tuple)
assert len(output_size) == 2
self.aap = nn.AdaptiveAvgPool2d(output_size)
else:
assert isinstance(output_size, tuple)
assert len(output_size) == 3
self.aap = nn.AdaptiveAvgPool3d(output_size)
self.output_size = output_size
def init_weights(self):
pass
def forward(self, inputs):
if isinstance(inputs, tuple):
outs = tuple([self.aap(x) for x in inputs])
elif isinstance(inputs, torch.Tensor):
outs = self.aap(inputs)
else:
raise TypeError('neck inputs should be tuple or torch.tensor')
return outs
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__adaptive_avg_pool2d_0(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 320
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 5
x0 = xindex % 4
x2 = xindex // 20
x4 = xindex
tmp0 = 4 * x1 // 5
tmp1 = (8 + 4 * x1) // 5
tmp2 = tmp0 < tmp1
tmp3 = x0
tmp4 = 1 + x0
tmp5 = tmp3 < tmp4
tmp6 = tmp2 & tmp5
tmp7 = tl.load(in_ptr0 + (x0 + 4 * (4 * x1 // 5) + 16 * x2), tmp6 &
xmask, other=0.0)
tmp8 = tmp4 < tmp4
tmp9 = tmp2 & tmp8
tmp10 = tl.load(in_ptr0 + (1 + x0 + 4 * (4 * x1 // 5) + 16 * x2), tmp9 &
xmask, other=0.0)
tmp11 = tmp10 + tmp7
tmp12 = 1 + 4 * x1 // 5
tmp13 = tmp12 < tmp1
tmp14 = tmp13 & tmp5
tmp15 = tl.load(in_ptr0 + (4 + x0 + 4 * (4 * x1 // 5) + 16 * x2), tmp14 &
xmask, other=0.0)
tmp16 = tmp15 + tmp11
tmp17 = tmp13 & tmp8
tmp18 = tl.load(in_ptr0 + (5 + x0 + 4 * (4 * x1 // 5) + 16 * x2), tmp17 &
xmask, other=0.0)
tmp19 = tmp18 + tmp16
tmp20 = 1.0
tmp21 = tl.full(tmp20.shape, 0.0, tmp20.dtype)
tmp22 = tl.where(tmp6, tmp20, tmp21)
tmp23 = tl.where(tmp9, tmp20, tmp21)
tmp24 = tmp23 + tmp22
tmp25 = tl.where(tmp14, tmp20, tmp21)
tmp26 = tmp25 + tmp24
tmp27 = tl.where(tmp17, tmp20, tmp21)
tmp28 = tmp27 + tmp26
tmp29 = tmp19 / tmp28
tl.store(out_ptr0 + x4, tmp29, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 5, 4), (80, 20, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__adaptive_avg_pool2d_0[grid(320)](arg0_1, buf0,
320, XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class AdaptiveAveragePoolingNew(nn.Module):
"""Adaptive Pooling neck.
Args:
dim (int): Dimensions of each sample channel, can be one of {1, 2, 3}.
Default: 2
output_size (int | tuple): output size,
If dim equals to 1: output_size is a single integer.
Else, if output_size is a tuple of integers with the length of dim.
Default: (5, 4)
"""
def __init__(self, dim=2, output_size=(5, 4)):
super(AdaptiveAveragePoolingNew, self).__init__()
assert dim in [1, 2, 3
], f'AdaptiveAveragePooling dim only support {1, 2, 3}, get {dim} instead.'
if dim == 1:
assert isinstance(output_size, int)
self.aap = nn.AdaptiveAvgPool1d(output_size)
elif dim == 2:
assert isinstance(output_size, tuple)
assert len(output_size) == 2
self.aap = nn.AdaptiveAvgPool2d(output_size)
else:
assert isinstance(output_size, tuple)
assert len(output_size) == 3
self.aap = nn.AdaptiveAvgPool3d(output_size)
self.output_size = output_size
def init_weights(self):
pass
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Min-Sheng/mmregression | AdaptiveAveragePooling | false | 843 | [
"Apache-2.0"
] | 0 | 6d70383d89ccb3dea7f425b665c2a184d014a99f | https://github.com/Min-Sheng/mmregression/tree/6d70383d89ccb3dea7f425b665c2a184d014a99f | import torch
import torch.nn as nn
class Model(nn.Module):
"""Adaptive Pooling neck.
Args:
dim (int): Dimensions of each sample channel, can be one of {1, 2, 3}.
Default: 2
output_size (int | tuple): output size,
If dim equals to 1: output_size is a single integer.
Else, if output_size is a tuple of integers with the length of dim.
Default: (5, 4)
"""
def __init__(self, dim=2, output_size=(5, 4)):
super().__init__()
assert dim in [1, 2, 3
], f'AdaptiveAveragePooling dim only support {1, 2, 3}, get {dim} instead.'
if dim == 1:
assert isinstance(output_size, int)
self.aap = nn.AdaptiveAvgPool1d(output_size)
elif dim == 2:
assert isinstance(output_size, tuple)
assert len(output_size) == 2
self.aap = nn.AdaptiveAvgPool2d(output_size)
else:
assert isinstance(output_size, tuple)
assert len(output_size) == 3
self.aap = nn.AdaptiveAvgPool3d(output_size)
self.output_size = output_size
def init_weights(self):
pass
def forward(self, inputs):
if isinstance(inputs, tuple):
outs = tuple([self.aap(x) for x in inputs])
elif isinstance(inputs, torch.Tensor):
outs = self.aap(inputs)
else:
raise TypeError('neck inputs should be tuple or torch.tensor')
return outs
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
DiceLoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/qz/cqza6p5fjiie2hfiu5dfjqqugrnzziwuwxzlhzy2aa7khopxjbym.py
# Topologically Sorted Source Nodes: [output], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# output => amax, exp, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%arg0_1, [1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg0_1, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
triton_poi_fused__softmax_0 = async_compile.triton('triton_poi_fused__softmax_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x3), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ow/cowcmfm2nngivhyq4bs2mbwucbdqfobg6iiyxzaiowjvardl53yg.py
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# contiguous => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x3), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/bh/cbh2bmdr7w3qvbew4pqw4afcg6vw2r54b5y4we3wgnpjykfrrvgk.py
# Topologically Sorted Source Nodes: [mul, intersect, add, denominator], Original ATen: [aten.mul, aten.sum, aten.add]
# Source node to ATen node mapping:
# add => add
# denominator => sum_3
# intersect => sum_2
# mul => mul
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view, %view_1), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%mul, [-1]), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view, %view_1), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%add, [-1]), kwargs = {})
triton_per_fused_add_mul_sum_2 = async_compile.triton('triton_per_fused_add_mul_sum_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[4, 64],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_mul_sum_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 2, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_mul_sum_2(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 4
rnumel = 64
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + ((16*x0) + (64*(r1 // 16)) + (r1 % 16)), xmask, other=0.0)
tmp1 = tl.load(in_ptr1 + ((16*x0) + (64*(r1 // 16)) + (r1 % 16)), xmask, other=0.0)
tmp2 = tmp0 * tmp1
tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp5 = tl.where(xmask, tmp3, 0)
tmp6 = tl.sum(tmp5, 1)[:, None]
tmp7 = tmp0 + tmp1
tmp8 = tl.broadcast_to(tmp7, [XBLOCK, RBLOCK])
tmp10 = tl.where(xmask, tmp8, 0)
tmp11 = tl.sum(tmp10, 1)[:, None]
tl.store(out_ptr0 + (x0), tmp6, xmask)
tl.store(out_ptr1 + (x0), tmp11, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/bk/cbk4ubll74yajr3doqwfgyyn546u572foe56segenriamhebmgpk.py
# Topologically Sorted Source Nodes: [dice, dice_1, sub], Original ATen: [aten.div, aten.mean, aten.rsub]
# Source node to ATen node mapping:
# dice => div_1
# dice_1 => mean
# sub => sub_1
# Graph fragment:
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_2, %sum_3), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%div_1,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %mean), kwargs = {})
triton_per_fused_div_mean_rsub_3 = async_compile.triton('triton_per_fused_div_mean_rsub_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 4],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_div_mean_rsub_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_div_mean_rsub_3(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 4
RBLOCK: tl.constexpr = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = tl.load(in_ptr1 + (r0), None)
tmp2 = tmp0 / tmp1
tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp5 = tl.sum(tmp3, 1)[:, None]
tmp6 = 4.0
tmp7 = tmp5 / tmp6
tmp8 = 1.0
tmp9 = tmp8 - tmp7
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp9, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [output], Original ATen: [aten._softmax]
stream0 = get_raw_stream(0)
triton_poi_fused__softmax_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
buf1 = empty_strided_cuda((4, 4, 4, 4), (16, 64, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf0, buf1, 256, grid=grid(256), stream=stream0)
del buf0
buf2 = empty_strided_cuda((4, ), (1, ), torch.float32)
buf3 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [mul, intersect, add, denominator], Original ATen: [aten.mul, aten.sum, aten.add]
triton_per_fused_add_mul_sum_2.run(buf1, arg1_1, buf2, buf3, 4, 64, grid=grid(4), stream=stream0)
del arg1_1
del buf1
buf4 = empty_strided_cuda((), (), torch.float32)
buf5 = buf4; del buf4 # reuse
# Topologically Sorted Source Nodes: [dice, dice_1, sub], Original ATen: [aten.div, aten.mean, aten.rsub]
triton_per_fused_div_mean_rsub_3.run(buf5, buf2, buf3, 1, 4, grid=grid(1), stream=stream0)
del buf2
del buf3
return (buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
def flatten(tensor):
"""Flattens a given tensor such that the channel axis is first.
The shapes are transformed as follows:
(N, C, D, H, W) -> (C, N * D * H * W)
"""
C = tensor.size(1)
axis_order = (1, 0) + tuple(range(2, tensor.dim()))
transposed = tensor.permute(axis_order)
return transposed.contiguous().view(C, -1)
class DiceLoss(nn.Module):
def __init__(self):
super().__init__()
self.epsilon = 1e-05
def forward(self, output, target):
assert output.size() == target.size(
), "'input' and 'target' must have the same shape"
output = F.softmax(output, dim=1)
output = flatten(output)
target = flatten(target)
intersect = (output * target).sum(-1)
denominator = (output + target).sum(-1)
dice = intersect / denominator
dice = torch.mean(dice)
return 1 - dice
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x3, tmp8, xmask)
@triton.jit
def triton_per_fused_add_mul_sum_2(in_ptr0, in_ptr1, out_ptr0, out_ptr1,
xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (16 * x0 + 64 * (r1 // 16) + r1 % 16), xmask,
other=0.0)
tmp1 = tl.load(in_ptr1 + (16 * x0 + 64 * (r1 // 16) + r1 % 16), xmask,
other=0.0)
tmp2 = tmp0 * tmp1
tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp5 = tl.where(xmask, tmp3, 0)
tmp6 = tl.sum(tmp5, 1)[:, None]
tmp7 = tmp0 + tmp1
tmp8 = tl.broadcast_to(tmp7, [XBLOCK, RBLOCK])
tmp10 = tl.where(xmask, tmp8, 0)
tmp11 = tl.sum(tmp10, 1)[:, None]
tl.store(out_ptr0 + x0, tmp6, xmask)
tl.store(out_ptr1 + x0, tmp11, xmask)
@triton.jit
def triton_per_fused_div_mean_rsub_3(in_out_ptr0, in_ptr0, in_ptr1, xnumel,
rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 4
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp2 = tmp0 / tmp1
tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp5 = tl.sum(tmp3, 1)[:, None]
tmp6 = 4.0
tmp7 = tmp5 / tmp6
tmp8 = 1.0
tmp9 = tmp8 - tmp7
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp9, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__softmax_0[grid(256)](arg0_1, buf0, 256, XBLOCK=
128, num_warps=4, num_stages=1)
del arg0_1
buf1 = empty_strided_cuda((4, 4, 4, 4), (16, 64, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(256)](buf0, buf1, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf0
buf2 = empty_strided_cuda((4,), (1,), torch.float32)
buf3 = empty_strided_cuda((4,), (1,), torch.float32)
triton_per_fused_add_mul_sum_2[grid(4)](buf1, arg1_1, buf2, buf3, 4,
64, XBLOCK=1, num_warps=2, num_stages=1)
del arg1_1
del buf1
buf4 = empty_strided_cuda((), (), torch.float32)
buf5 = buf4
del buf4
triton_per_fused_div_mean_rsub_3[grid(1)](buf5, buf2, buf3, 1, 4,
XBLOCK=1, num_warps=2, num_stages=1)
del buf2
del buf3
return buf5,
def flatten(tensor):
"""Flattens a given tensor such that the channel axis is first.
The shapes are transformed as follows:
(N, C, D, H, W) -> (C, N * D * H * W)
"""
C = tensor.size(1)
axis_order = (1, 0) + tuple(range(2, tensor.dim()))
transposed = tensor.permute(axis_order)
return transposed.contiguous().view(C, -1)
class DiceLossNew(nn.Module):
def __init__(self):
super().__init__()
self.epsilon = 1e-05
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| MikoyChinese/Learn | DiceLoss | false | 844 | [
"Apache-2.0"
] | 0 | c482b1e84496279935b5bb2cfc1e6d78e2868c63 | https://github.com/MikoyChinese/Learn/tree/c482b1e84496279935b5bb2cfc1e6d78e2868c63 | import torch
import torch.nn as nn
import torch.nn.functional as F
def flatten(tensor):
"""Flattens a given tensor such that the channel axis is first.
The shapes are transformed as follows:
(N, C, D, H, W) -> (C, N * D * H * W)
"""
C = tensor.size(1)
axis_order = (1, 0) + tuple(range(2, tensor.dim()))
transposed = tensor.permute(axis_order)
return transposed.contiguous().view(C, -1)
class Model(nn.Module):
def __init__(self):
super().__init__()
self.epsilon = 1e-05
def forward(self, output, target):
assert output.size() == target.size(
), "'input' and 'target' must have the same shape"
output = F.softmax(output, dim=1)
output = flatten(output)
target = flatten(target)
intersect = (output * target).sum(-1)
denominator = (output + target).sum(-1)
dice = intersect / denominator
dice = torch.mean(dice)
return 1 - dice
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
Residual_Block | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/tc/ctchh66224xunisoqylba7qzpsl47ose6e5qu7bcgrxigfwirvf4.py
# Topologically Sorted Source Nodes: [instance_norm, output], Original ATen: [aten.repeat, aten._native_batch_norm_legit, aten.leaky_relu]
# Source node to ATen node mapping:
# instance_norm => add, add_1, mul, mul_1, repeat, rsqrt, sub, var_mean
# output => gt, mul_2, where
# Graph fragment:
# %repeat : [num_users=2] = call_function[target=torch.ops.aten.repeat.default](args = (%primals_3, [4]), kwargs = {})
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%view, [0, 2, 3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=2] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view, %getitem_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %unsqueeze_1), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %unsqueeze_3), kwargs = {})
# %gt : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%view_1, 0), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 0.2), kwargs = {})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%gt, %view_1, %mul_2), kwargs = {})
triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0 = async_compile.triton('triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.reduction(
size_hints=[256, 4096],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32', 8: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7, 8), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 2, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, out_ptr3, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
xnumel = 256
rnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0 % 64), xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x0), tmp0, xmask)
tmp3_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp1 = tl.load(in_ptr1 + (r1 + (4096*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp3_mean_next, tmp3_m2_next, tmp3_weight_next = triton_helpers.welford_reduce(
tmp2, tmp3_mean, tmp3_m2, tmp3_weight, roffset == 0
)
tmp3_mean = tl.where(rmask & xmask, tmp3_mean_next, tmp3_mean)
tmp3_m2 = tl.where(rmask & xmask, tmp3_m2_next, tmp3_m2)
tmp3_weight = tl.where(rmask & xmask, tmp3_weight_next, tmp3_weight)
tmp3_tmp, tmp4_tmp, tmp5_tmp = triton_helpers.welford(
tmp3_mean, tmp3_m2, tmp3_weight, 1
)
tmp3 = tmp3_tmp[:, None]
tmp4 = tmp4_tmp[:, None]
tmp5 = tmp5_tmp[:, None]
tl.store(out_ptr1 + (x0), tmp3, xmask)
tmp15 = tl.load(in_ptr2 + (x0 % 64), xmask, eviction_policy='evict_last')
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp6 = tl.load(in_ptr1 + (r1 + (4096*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0)
tmp7 = tmp6 - tmp3
tmp8 = 4096.0
tmp9 = tmp4 / tmp8
tmp10 = 1e-05
tmp11 = tmp9 + tmp10
tmp12 = libdevice.rsqrt(tmp11)
tmp13 = tmp7 * tmp12
tmp14 = tmp13 * tmp0
tmp16 = tmp14 + tmp15
tmp17 = 0.0
tmp18 = tmp16 > tmp17
tmp19 = 0.2
tmp20 = tmp16 * tmp19
tmp21 = tl.where(tmp18, tmp16, tmp20)
tl.store(in_out_ptr0 + (r1 + (4096*x0)), tmp21, rmask & xmask)
tmp22 = 4096.0
tmp23 = tmp4 / tmp22
tmp24 = 1e-05
tmp25 = tmp23 + tmp24
tmp26 = libdevice.rsqrt(tmp25)
tl.store(out_ptr3 + (x0), tmp26, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/st/cstrjrus2xtc6elawqi42w5evvdehzcp4d5fjmxji4eznl5szfll.py
# Topologically Sorted Source Nodes: [output_1, output_2], Original ATen: [aten.repeat, aten._native_batch_norm_legit, aten.add]
# Source node to ATen node mapping:
# output_1 => add_2, repeat_2, rsqrt_1, var_mean_1
# output_2 => add_4
# Graph fragment:
# %repeat_2 : [num_users=2] = call_function[target=torch.ops.aten.repeat.default](args = (%primals_6, [4]), kwargs = {})
# %var_mean_1 : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%view_5, [0, 2, 3]), kwargs = {correction: 0, keepdim: True})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_2, 1e-05), kwargs = {})
# %rsqrt_1 : [num_users=2] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add_2,), kwargs = {})
# %add_4 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_6, %primals_1), kwargs = {})
triton_red_fused__native_batch_norm_legit_add_repeat_1 = async_compile.triton('triton_red_fused__native_batch_norm_legit_add_repeat_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.reduction(
size_hints=[256, 4096],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: '*fp32', 8: 'i32', 9: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7, 8, 9), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__native_batch_norm_legit_add_repeat_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 2, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_red_fused__native_batch_norm_legit_add_repeat_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, out_ptr1, out_ptr3, out_ptr4, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
xnumel = 256
rnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0 % 64), xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x0), tmp0, xmask)
tmp3_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp1 = tl.load(in_ptr1 + (r1 + (4096*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp3_mean_next, tmp3_m2_next, tmp3_weight_next = triton_helpers.welford_reduce(
tmp2, tmp3_mean, tmp3_m2, tmp3_weight, roffset == 0
)
tmp3_mean = tl.where(rmask & xmask, tmp3_mean_next, tmp3_mean)
tmp3_m2 = tl.where(rmask & xmask, tmp3_m2_next, tmp3_m2)
tmp3_weight = tl.where(rmask & xmask, tmp3_weight_next, tmp3_weight)
tmp3_tmp, tmp4_tmp, tmp5_tmp = triton_helpers.welford(
tmp3_mean, tmp3_m2, tmp3_weight, 1
)
tmp3 = tmp3_tmp[:, None]
tmp4 = tmp4_tmp[:, None]
tmp5 = tmp5_tmp[:, None]
tl.store(out_ptr1 + (x0), tmp3, xmask)
x2 = xindex % 64
tmp15 = tl.load(in_ptr2 + (x2), xmask, eviction_policy='evict_last')
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp6 = tl.load(in_ptr1 + (r1 + (4096*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0)
tmp17 = tl.load(in_ptr3 + (r1 + (4096*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0)
tmp7 = tmp6 - tmp3
tmp8 = 4096.0
tmp9 = tmp4 / tmp8
tmp10 = 1e-05
tmp11 = tmp9 + tmp10
tmp12 = libdevice.rsqrt(tmp11)
tmp13 = tmp7 * tmp12
tmp14 = tmp13 * tmp0
tmp16 = tmp14 + tmp15
tmp18 = tmp16 + tmp17
tl.store(out_ptr3 + (r1 + (4096*x0)), tmp18, rmask & xmask)
tmp19 = 4096.0
tmp20 = tmp4 / tmp19
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr4 + (x0), tmp23, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7 = args
args.clear()
assert_size_stride(primals_1, (4, 64, 64, 64), (262144, 4096, 64, 1))
assert_size_stride(primals_2, (64, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_3, (64, ), (1, ))
assert_size_stride(primals_4, (64, ), (1, ))
assert_size_stride(primals_5, (64, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_6, (64, ), (1, ))
assert_size_stride(primals_7, (64, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 64, 64, 64), (262144, 4096, 64, 1))
buf1 = empty_strided_cuda((256, ), (1, ), torch.float32)
buf2 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256), torch.float32)
buf6 = empty_strided_cuda((1, 256, 64, 64), (1048576, 4096, 64, 1), torch.float32)
buf7 = reinterpret_tensor(buf6, (4, 64, 64, 64), (262144, 4096, 64, 1), 0); del buf6 # reuse
buf5 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256), torch.float32)
# Topologically Sorted Source Nodes: [instance_norm, output], Original ATen: [aten.repeat, aten._native_batch_norm_legit, aten.leaky_relu]
stream0 = get_raw_stream(0)
triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0.run(buf7, primals_3, buf0, primals_4, buf1, buf2, buf5, 256, 4096, grid=grid(256), stream=stream0)
del primals_3
del primals_4
# Topologically Sorted Source Nodes: [conv2d_1], Original ATen: [aten.convolution]
buf8 = extern_kernels.convolution(buf7, primals_5, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 64, 64, 64), (262144, 4096, 64, 1))
buf9 = empty_strided_cuda((256, ), (1, ), torch.float32)
buf10 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256), torch.float32)
buf14 = empty_strided_cuda((4, 64, 64, 64), (262144, 4096, 64, 1), torch.float32)
buf13 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256), torch.float32)
# Topologically Sorted Source Nodes: [output_1, output_2], Original ATen: [aten.repeat, aten._native_batch_norm_legit, aten.add]
triton_red_fused__native_batch_norm_legit_add_repeat_1.run(primals_6, buf8, primals_7, primals_1, buf9, buf10, buf14, buf13, 256, 4096, grid=grid(256), stream=stream0)
del primals_6
del primals_7
return (buf14, primals_1, primals_2, primals_5, buf0, buf1, reinterpret_tensor(buf5, (256, ), (1, ), 0), buf7, buf8, buf9, reinterpret_tensor(buf13, (256, ), (1, ), 0), reinterpret_tensor(buf10, (1, 256, 1, 1), (256, 1, 1, 1), 0), reinterpret_tensor(buf2, (1, 256, 1, 1), (256, 1, 1, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 64, 64, 64), (262144, 4096, 64, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((64, 64, 3, 3), (576, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((64, 64, 3, 3), (576, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
from torch import add
class Residual_Block(nn.Module):
def __init__(self):
super(Residual_Block, self).__init__()
self.conv1 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size
=3, padding=1, bias=False)
self.in1 = nn.InstanceNorm2d(64, affine=True)
self.relu = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size
=3, padding=1, bias=False)
self.in2 = nn.InstanceNorm2d(64, affine=True)
def forward(self, x):
identity_data = x
output = self.relu(self.in1(self.conv1(x)))
output = self.in2(self.conv2(output))
output = add(output, identity_data)
return output
def get_inputs():
return [torch.rand([4, 64, 64, 64])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0(in_out_ptr0,
in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, out_ptr3, xnumel, rnumel,
XBLOCK: tl.constexpr, RBLOCK: tl.constexpr):
xnumel = 256
rnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0 % 64, xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + x0, tmp0, xmask)
tmp3_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp1 = tl.load(in_ptr1 + (r1 + 4096 * x0), rmask & xmask,
eviction_policy='evict_last', other=0.0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp3_mean_next, tmp3_m2_next, tmp3_weight_next = (triton_helpers.
welford_reduce(tmp2, tmp3_mean, tmp3_m2, tmp3_weight, roffset == 0)
)
tmp3_mean = tl.where(rmask & xmask, tmp3_mean_next, tmp3_mean)
tmp3_m2 = tl.where(rmask & xmask, tmp3_m2_next, tmp3_m2)
tmp3_weight = tl.where(rmask & xmask, tmp3_weight_next, tmp3_weight)
tmp3_tmp, tmp4_tmp, tmp5_tmp = triton_helpers.welford(tmp3_mean,
tmp3_m2, tmp3_weight, 1)
tmp3 = tmp3_tmp[:, None]
tmp4 = tmp4_tmp[:, None]
tmp5_tmp[:, None]
tl.store(out_ptr1 + x0, tmp3, xmask)
tmp15 = tl.load(in_ptr2 + x0 % 64, xmask, eviction_policy='evict_last')
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp6 = tl.load(in_ptr1 + (r1 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp7 = tmp6 - tmp3
tmp8 = 4096.0
tmp9 = tmp4 / tmp8
tmp10 = 1e-05
tmp11 = tmp9 + tmp10
tmp12 = libdevice.rsqrt(tmp11)
tmp13 = tmp7 * tmp12
tmp14 = tmp13 * tmp0
tmp16 = tmp14 + tmp15
tmp17 = 0.0
tmp18 = tmp16 > tmp17
tmp19 = 0.2
tmp20 = tmp16 * tmp19
tmp21 = tl.where(tmp18, tmp16, tmp20)
tl.store(in_out_ptr0 + (r1 + 4096 * x0), tmp21, rmask & xmask)
tmp22 = 4096.0
tmp23 = tmp4 / tmp22
tmp24 = 1e-05
tmp25 = tmp23 + tmp24
tmp26 = libdevice.rsqrt(tmp25)
tl.store(out_ptr3 + x0, tmp26, xmask)
@triton.jit
def triton_red_fused__native_batch_norm_legit_add_repeat_1(in_ptr0, in_ptr1,
in_ptr2, in_ptr3, out_ptr0, out_ptr1, out_ptr3, out_ptr4, xnumel,
rnumel, XBLOCK: tl.constexpr, RBLOCK: tl.constexpr):
xnumel = 256
rnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0 % 64, xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + x0, tmp0, xmask)
tmp3_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp3_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp1 = tl.load(in_ptr1 + (r1 + 4096 * x0), rmask & xmask,
eviction_policy='evict_last', other=0.0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp3_mean_next, tmp3_m2_next, tmp3_weight_next = (triton_helpers.
welford_reduce(tmp2, tmp3_mean, tmp3_m2, tmp3_weight, roffset == 0)
)
tmp3_mean = tl.where(rmask & xmask, tmp3_mean_next, tmp3_mean)
tmp3_m2 = tl.where(rmask & xmask, tmp3_m2_next, tmp3_m2)
tmp3_weight = tl.where(rmask & xmask, tmp3_weight_next, tmp3_weight)
tmp3_tmp, tmp4_tmp, tmp5_tmp = triton_helpers.welford(tmp3_mean,
tmp3_m2, tmp3_weight, 1)
tmp3 = tmp3_tmp[:, None]
tmp4 = tmp4_tmp[:, None]
tmp5_tmp[:, None]
tl.store(out_ptr1 + x0, tmp3, xmask)
x2 = xindex % 64
tmp15 = tl.load(in_ptr2 + x2, xmask, eviction_policy='evict_last')
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r1 = rindex
tmp6 = tl.load(in_ptr1 + (r1 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp17 = tl.load(in_ptr3 + (r1 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp7 = tmp6 - tmp3
tmp8 = 4096.0
tmp9 = tmp4 / tmp8
tmp10 = 1e-05
tmp11 = tmp9 + tmp10
tmp12 = libdevice.rsqrt(tmp11)
tmp13 = tmp7 * tmp12
tmp14 = tmp13 * tmp0
tmp16 = tmp14 + tmp15
tmp18 = tmp16 + tmp17
tl.store(out_ptr3 + (r1 + 4096 * x0), tmp18, rmask & xmask)
tmp19 = 4096.0
tmp20 = tmp4 / tmp19
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr4 + x0, tmp23, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 64, 64, 64), (262144, 4096, 64, 1))
assert_size_stride(primals_2, (64, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_3, (64,), (1,))
assert_size_stride(primals_4, (64,), (1,))
assert_size_stride(primals_5, (64, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_6, (64,), (1,))
assert_size_stride(primals_7, (64,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1,
1), padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 64, 64, 64), (262144, 4096, 64, 1))
buf1 = empty_strided_cuda((256,), (1,), torch.float32)
buf2 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256), torch
.float32)
buf6 = empty_strided_cuda((1, 256, 64, 64), (1048576, 4096, 64, 1),
torch.float32)
buf7 = reinterpret_tensor(buf6, (4, 64, 64, 64), (262144, 4096, 64,
1), 0)
del buf6
buf5 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256), torch
.float32)
get_raw_stream(0)
triton_red_fused__native_batch_norm_legit_leaky_relu_repeat_0[grid(256)
](buf7, primals_3, buf0, primals_4, buf1, buf2, buf5, 256, 4096,
XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1)
del primals_3
del primals_4
buf8 = extern_kernels.convolution(buf7, primals_5, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 64, 64, 64), (262144, 4096, 64, 1))
buf9 = empty_strided_cuda((256,), (1,), torch.float32)
buf10 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256),
torch.float32)
buf14 = empty_strided_cuda((4, 64, 64, 64), (262144, 4096, 64, 1),
torch.float32)
buf13 = empty_strided_cuda((1, 256, 1, 1), (256, 1, 256, 256),
torch.float32)
triton_red_fused__native_batch_norm_legit_add_repeat_1[grid(256)](
primals_6, buf8, primals_7, primals_1, buf9, buf10, buf14,
buf13, 256, 4096, XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1
)
del primals_6
del primals_7
return (buf14, primals_1, primals_2, primals_5, buf0, buf1,
reinterpret_tensor(buf5, (256,), (1,), 0), buf7, buf8, buf9,
reinterpret_tensor(buf13, (256,), (1,), 0), reinterpret_tensor(
buf10, (1, 256, 1, 1), (256, 1, 1, 1), 0), reinterpret_tensor(buf2,
(1, 256, 1, 1), (256, 1, 1, 1), 0))
class Residual_BlockNew(nn.Module):
def __init__(self):
super(Residual_BlockNew, self).__init__()
self.conv1 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size
=3, padding=1, bias=False)
self.in1 = nn.InstanceNorm2d(64, affine=True)
self.relu = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size
=3, padding=1, bias=False)
self.in2 = nn.InstanceNorm2d(64, affine=True)
def forward(self, input_0):
primals_2 = self.conv1.weight
primals_3 = self.in1.weight
primals_4 = self.in1.bias
primals_5 = self.conv2.weight
primals_6 = self.in2.weight
primals_7 = self.in2.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
| MatusBako/MakeFacesGreatAgain | Residual_Block | false | 845 | [
"MIT"
] | 0 | e4941a8460db79dec566ed02d4b23eafb416a6db | https://github.com/MatusBako/MakeFacesGreatAgain/tree/e4941a8460db79dec566ed02d4b23eafb416a6db | import torch
import torch.nn as nn
from torch import add
class Model(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size
=3, padding=1, bias=False)
self.in1 = nn.InstanceNorm2d(64, affine=True)
self.relu = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size
=3, padding=1, bias=False)
self.in2 = nn.InstanceNorm2d(64, affine=True)
def forward(self, x):
identity_data = x
output = self.relu(self.in1(self.conv1(x)))
output = self.in2(self.conv2(output))
output = add(output, identity_data)
return output
def get_inputs():
return [torch.rand([4, 64, 64, 64])]
def get_init_inputs():
return []
|
ScaledDotProductAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xe/cxeuttfzx4xq2jmzwzvkech4crjirky5wjckb34lnep5o6sog3uw.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + (4*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x4), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fn/cfnr6wn6wbusamhilcgctjberp7g5kksyakcze32k6ntswznc2de.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone_1
# Graph fragment:
# %clone_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand_1,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = (yindex // 16)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (16*x2) + (64*y1)), xmask & ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (y0), ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + (4*y3)), tmp2, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ka/ckaneo6wn23ipwgbubou64jdtwieswlrn7w7r7kqky4aagh3v6l3.py
# Topologically Sorted Source Nodes: [wrapped_sqrt, att_1], Original ATen: [aten.sqrt, aten._softmax]
# Source node to ATen node mapping:
# att_1 => exp
# wrapped_sqrt => full_default
# Graph fragment:
# %full_default : [num_users=2] = call_function[target=torch.ops.aten.full.default](args = ([], 2.0), kwargs = {dtype: torch.float64, layout: torch.strided, device: cpu, pin_memory: False})
# %scalar_tensor_default : [num_users=2] = call_function[target=torch.ops.aten.scalar_tensor.default](args = (1,), kwargs = {dtype: torch.float32, device: cuda:0, pin_memory: False})
# %ge_scalar : [num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%full_default, 0), kwargs = {})
# %neg_default : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%scalar_tensor_default,), kwargs = {})
# %where_self : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ge_scalar, %scalar_tensor_default, %neg_default), kwargs = {})
# %mul_tensor : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_11, %where_self), kwargs = {})
# %amax_default : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%mul_tensor, [-1], True), kwargs = {})
# %sub_tensor : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_tensor, %amax_default), kwargs = {})
# %mul_tensor_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where_self, %full_default), kwargs = {})
# %div_tensor : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sub_tensor, %mul_tensor_1), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%div_tensor,), kwargs = {})
triton_poi_fused__softmax_sqrt_2 = async_compile.triton('triton_poi_fused__softmax_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_sqrt_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_sqrt_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp8 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp1 = tl.full([1], 2.0, tl.float64)
tmp2 = tl.full([1], 0.0, tl.float64)
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp9 = tmp8 * tmp6
tmp11 = tmp10 * tmp6
tmp12 = triton_helpers.maximum(tmp9, tmp11)
tmp14 = tmp13 * tmp6
tmp15 = triton_helpers.maximum(tmp12, tmp14)
tmp17 = tmp16 * tmp6
tmp18 = triton_helpers.maximum(tmp15, tmp17)
tmp19 = tmp7 - tmp18
tmp20 = tmp6.to(tl.float64)
tmp21 = tmp20 * tmp1
tmp22 = tmp21.to(tl.float32)
tmp23 = tmp19 / tmp22
tmp24 = tl_math.exp(tmp23)
tl.store(out_ptr0 + (x2), tmp24, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ry/cryn7ntc2gpkbfzbre3xh7lffx7zkbskw6oihbzsekkgajmdbki6.py
# Topologically Sorted Source Nodes: [att_1], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# att_1 => div_1, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div_1 : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_3 = async_compile.triton('triton_poi_fused__softmax_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6b/c6busvilz5nn36jjet3bmw7cqddirh4sgalamjr3fsrp3sbsacfi.py
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# contiguous => clone_3
# Graph fragment:
# %clone_3 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_6,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_4 = async_compile.triton('triton_poi_fused_clone_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tl.store(out_ptr0 + (x4), tmp0, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16, ), (1, ))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16, ), (1, ))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16, ), (1, ))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 16), (16, 1))
assert_size_stride(primals_11, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(buf0, primals_4, buf3, 256, grid=grid(256), stream=stream0)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf1, primals_6, buf4, 64, 4, grid=grid(64, 4), stream=stream0)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [wrapped_sqrt, att_1], Original ATen: [aten.sqrt, aten._softmax]
triton_poi_fused__softmax_sqrt_2.run(buf5, buf6, 256, grid=grid(256), stream=stream0)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [att_1], Original ATen: [aten._softmax]
triton_poi_fused__softmax_3.run(buf6, buf7, 256, grid=grid(256), stream=stream0)
buf8 = buf6; del buf6 # reuse
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.clone]
triton_poi_fused_clone_0.run(buf2, primals_8, buf8, 256, grid=grid(256), stream=stream0)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0); del buf2 # reuse
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
triton_poi_fused_clone_4.run(buf9, buf10, 256, grid=grid(256), stream=stream0)
del buf9
buf11 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_11, reinterpret_tensor(buf10, (16, 16), (16, 1), 0), reinterpret_tensor(primals_10, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf11)
del primals_11
return (reinterpret_tensor(buf11, (4, 4, 4), (16, 4, 1), 0), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), buf7, reinterpret_tensor(buf10, (16, 16), (16, 1), 0), primals_10, reinterpret_tensor(buf8, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class ScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nv, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, n_k, d_v)
:return: out (b_s, nq, d_model)
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4])
]
def get_init_inputs():
return [[], {'d_model': 4, 'd_k': 4, 'd_v': 4, 'h': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
from torch.autograd import *
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = yindex // 16
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 16 * x2 + 64 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_sqrt_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp8 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp13 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = tl.full([1], 2.0, tl.float64)
tmp2 = tl.full([1], 0.0, tl.float64)
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp9 = tmp8 * tmp6
tmp11 = tmp10 * tmp6
tmp12 = triton_helpers.maximum(tmp9, tmp11)
tmp14 = tmp13 * tmp6
tmp15 = triton_helpers.maximum(tmp12, tmp14)
tmp17 = tmp16 * tmp6
tmp18 = triton_helpers.maximum(tmp15, tmp17)
tmp19 = tmp7 - tmp18
tmp20 = tmp6.to(tl.float64)
tmp21 = tmp20 * tmp1
tmp22 = tmp21.to(tl.float32)
tmp23 = tmp19 / tmp22
tmp24 = tl_math.exp(tmp23)
tl.store(out_ptr0 + x2, tmp24, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tl.store(out_ptr0 + x4, tmp0, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16,), (1,))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16,), (1,))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 16), (16, 1))
assert_size_stride(primals_11, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(256)](buf0, primals_4, buf3, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_clone_1[grid(64, 4)](buf1, primals_6, buf4, 64, 4,
XBLOCK=4, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_sqrt_2[grid(256)](buf5, buf6, 256, XBLOCK
=128, num_warps=4, num_stages=1)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused__softmax_3[grid(256)](buf6, buf7, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf8 = buf6
del buf6
triton_poi_fused_clone_0[grid(256)](buf2, primals_8, buf8, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_4[grid(256)](buf9, buf10, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf9
buf11 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_11, reinterpret_tensor(buf10, (16, 16),
(16, 1), 0), reinterpret_tensor(primals_10, (16, 4), (1, 16), 0
), alpha=1, beta=1, out=buf11)
del primals_11
return reinterpret_tensor(buf11, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_2, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf7, reinterpret_tensor(buf10, (16, 16), (16, 1), 0
), primals_10, reinterpret_tensor(buf8, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0)
class ScaledDotProductAttentionNew(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductAttentionNew, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, input_0, input_1, input_2):
primals_3 = self.fc_q.weight
primals_4 = self.fc_q.bias
primals_5 = self.fc_k.weight
primals_6 = self.fc_k.bias
primals_7 = self.fc_v.weight
primals_8 = self.fc_v.bias
primals_10 = self.fc_o.weight
primals_11 = self.fc_o.bias
primals_1 = input_0
primals_2 = input_1
primals_9 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11])
return output[0]
| Maxi-0902/DRAN | ScaledDotProductAttention | false | 846 | [
"MIT"
] | 0 | c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | https://github.com/Maxi-0902/DRAN/tree/c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class Model(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super().__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nv, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, n_k, d_v)
:return: out (b_s, nq, d_model)
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4])
]
def get_init_inputs():
return [4, 4, 4, 4]
|
activation_quantize_fn | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ey/ceyfjb5l4jwqlb5hnlsrrauzvcyyuprkp5vh3odjdm5fjdhyuxlp.py
# Topologically Sorted Source Nodes: [mul, round_1, out], Original ATen: [aten.mul, aten.round, aten.div]
# Source node to ATen node mapping:
# mul => mul
# out => div
# round_1 => round_1
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg0_1, 15.0), kwargs = {})
# %round_1 : [num_users=1] = call_function[target=torch.ops.aten.round.default](args = (%mul,), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%round_1, 15.0), kwargs = {})
triton_poi_fused_div_mul_round_0 = async_compile.triton('triton_poi_fused_div_mul_round_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_div_mul_round_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_div_mul_round_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 15.0
tmp2 = tmp0 * tmp1
tmp3 = libdevice.nearbyint(tmp2)
tmp4 = 0.06666666666666667
tmp5 = tmp3 * tmp4
tl.store(out_ptr0 + (x0), tmp5, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul, round_1, out], Original ATen: [aten.mul, aten.round, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_div_mul_round_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.utils.data
import torch.nn as nn
def uniform_quantize(k):
class qfn(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if k == 32:
out = input
elif k == 1:
out = torch.sign(input)
else:
n = float(2 ** k - 1)
out = torch.round(input * n) / n
return out
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
return qfn().apply
class activation_quantize_fn(nn.Module):
def __init__(self, a_bit):
super(activation_quantize_fn, self).__init__()
assert a_bit <= 8 or a_bit == 32
self.a_bit = a_bit
self.uniform_q = uniform_quantize(k=a_bit)
def forward(self, x):
if self.a_bit == 32:
activation_q = x
else:
activation_q = self.uniform_q(torch.clamp(x, 0, 1))
return activation_q
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'a_bit': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.utils.data
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_div_mul_round_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 15.0
tmp2 = tmp0 * tmp1
tmp3 = libdevice.nearbyint(tmp2)
tmp4 = 0.06666666666666667
tmp5 = tmp3 * tmp4
tl.store(out_ptr0 + x0, tmp5, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_div_mul_round_0[grid(256)](arg0_1, buf0, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del arg0_1
return buf0,
def uniform_quantize(k):
class qfn(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if k == 32:
out = input
elif k == 1:
out = torch.sign(input)
else:
n = float(2 ** k - 1)
out = torch.round(input * n) / n
return out
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
return qfn().apply
class activation_quantize_fnNew(nn.Module):
def __init__(self, a_bit):
super(activation_quantize_fnNew, self).__init__()
assert a_bit <= 8 or a_bit == 32
self.a_bit = a_bit
self.uniform_q = uniform_quantize(k=a_bit)
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MohammedHAlali/pytorch_DoReFaNet | activation_quantize_fn | false | 847 | [
"MIT"
] | 0 | d208089b9172f02c09cc6633158ed5b5d6cd7f1e | https://github.com/MohammedHAlali/pytorch_DoReFaNet/tree/d208089b9172f02c09cc6633158ed5b5d6cd7f1e | import torch
import torch.utils.data
import torch.nn as nn
def uniform_quantize(k):
class qfn(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if k == 32:
out = input
elif k == 1:
out = torch.sign(input)
else:
n = float(2 ** k - 1)
out = torch.round(input * n) / n
return out
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
return qfn().apply
class Model(nn.Module):
def __init__(self, a_bit):
super().__init__()
assert a_bit <= 8 or a_bit == 32
self.a_bit = a_bit
self.uniform_q = uniform_quantize(k=a_bit)
def forward(self, x):
if self.a_bit == 32:
activation_q = x
else:
activation_q = self.uniform_q(torch.clamp(x, 0, 1))
return activation_q
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
FeedForward | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/p7/cp7qy7k3id7baiclou4lkuuzxq3f2zjcjcnoamy3rdevi3rsgjjj.py
# Topologically Sorted Source Nodes: [gelu], Original ATen: [aten.gelu]
# Source node to ATen node mapping:
# gelu => add, erf, mul, mul_1, mul_2
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 0.5), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 0.7071067811865476), kwargs = {})
# %erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_1,), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %add), kwargs = {})
triton_poi_fused_gelu_0 = async_compile.triton('triton_poi_fused_gelu_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gelu_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_gelu_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + (x0), tmp8, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (16, 4), (4, 1))
assert_size_stride(primals_2, (16, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 16), (16, 1))
assert_size_stride(primals_5, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 16), (1, 4), 0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [gelu], Original ATen: [aten.gelu]
stream0 = get_raw_stream(0)
triton_poi_fused_gelu_0.run(buf0, buf1, 1024, grid=grid(1024), stream=stream0)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 16), (16, 1), 0), reinterpret_tensor(primals_4, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf2)
del primals_5
return (reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), buf0, reinterpret_tensor(buf1, (64, 16), (16, 1), 0), primals_4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'num_features': 4, 'expansion_factor': 4, 'dropout': 0.5}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_gelu_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (16, 4), (4, 1))
assert_size_stride(primals_2, (16,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 16), (16, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 16), (1, 4),
0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused_gelu_0[grid(1024)](buf0, buf1, 1024, XBLOCK=128,
num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 16),
(16, 1), 0), reinterpret_tensor(primals_4, (16, 4), (1, 16), 0),
alpha=1, beta=1, out=buf2)
del primals_5
return reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf0, reinterpret_tensor(buf1, (64, 16), (16, 1), 0), primals_4
class FeedForwardNew(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| Misuzu-Kurenai/mlp-singer | FeedForward | false | 848 | [
"MIT"
] | 0 | 416451045bb9b3965aaf496e84a8b45332a6ba59 | https://github.com/Misuzu-Kurenai/mlp-singer/tree/416451045bb9b3965aaf496e84a8b45332a6ba59 | import torch
import torch.nn.functional as F
from torch import nn
class Model(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 0.5]
|
ScaledDotProductWithBoxAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xe/cxeuttfzx4xq2jmzwzvkech4crjirky5wjckb34lnep5o6sog3uw.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + (4*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x4), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fn/cfnr6wn6wbusamhilcgctjberp7g5kksyakcze32k6ntswznc2de.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone_1
# Graph fragment:
# %clone_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand_1,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = (yindex // 16)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (16*x2) + (64*y1)), xmask & ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (y0), ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + (4*y3)), tmp2, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/an/cank72e7b53ox2aydemjjtk4g5r4l3ezhruq4nvlf7sdzagvxw7i.py
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
# Source node to ATen node mapping:
# att => div
# clamp => clamp_min
# log => log
# w_mn => add
# w_mn_1 => amax, exp, sub, sum_1
# wrapped_sqrt => full_default
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 2.0), kwargs = {dtype: torch.float64, layout: torch.strided, device: cpu, pin_memory: False})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%view_11, %full_default), kwargs = {})
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%primals_10, 1e-06), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_min,), kwargs = {})
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%log, %div), kwargs = {})
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%add, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
triton_poi_fused__softmax_add_clamp_div_log_sqrt_2 = async_compile.triton('triton_poi_fused__softmax_add_clamp_div_log_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_add_clamp_div_log_sqrt_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_2(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (4*x2), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr1 + (1 + (4*x2)), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr1 + (2 + (4*x2)), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr1 + (3 + (4*x2)), xmask, eviction_policy='evict_last')
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = triton_helpers.maximum(tmp8, tmp1)
tmp10 = tl_math.log(tmp9)
tmp12 = tmp11 / tmp5
tmp13 = tmp10 + tmp12
tmp14 = triton_helpers.maximum(tmp7, tmp13)
tmp16 = triton_helpers.maximum(tmp15, tmp1)
tmp17 = tl_math.log(tmp16)
tmp19 = tmp18 / tmp5
tmp20 = tmp17 + tmp19
tmp21 = triton_helpers.maximum(tmp14, tmp20)
tmp23 = triton_helpers.maximum(tmp22, tmp1)
tmp24 = tl_math.log(tmp23)
tmp26 = tmp25 / tmp5
tmp27 = tmp24 + tmp26
tmp28 = triton_helpers.maximum(tmp21, tmp27)
tmp29 = tmp7 - tmp28
tmp30 = tl_math.exp(tmp29)
tmp31 = tmp13 - tmp28
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp34 = tmp20 - tmp28
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tmp37 = tmp27 - tmp28
tmp38 = tl_math.exp(tmp37)
tmp39 = tmp36 + tmp38
tl.store(out_ptr0 + (x2), tmp28, xmask)
tl.store(out_ptr1 + (x2), tmp39, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/rp/crpmptrcembmuvzewgtdminpn6dx6emnjq6iahknzv7cxb67vchn.py
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
# Source node to ATen node mapping:
# att => div
# clamp => clamp_min
# log => log
# w_mn => add
# w_mn_1 => amax, div_1, exp, sub
# wrapped_sqrt => full_default
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 2.0), kwargs = {dtype: torch.float64, layout: torch.strided, device: cpu, pin_memory: False})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%view_11, %full_default), kwargs = {})
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%primals_10, 1e-06), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_min,), kwargs = {})
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%log, %div), kwargs = {})
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%add, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
# %div_1 : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_add_clamp_div_log_sqrt_3 = async_compile.triton('triton_poi_fused__softmax_add_clamp_div_log_sqrt_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_add_clamp_div_log_sqrt_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_3(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex % 64
x4 = xindex
x5 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x3), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_out_ptr0 + (x4), xmask)
tmp8 = tl.load(in_ptr1 + (x5), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr2 + (x5), xmask, eviction_policy='evict_last')
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = tmp7 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp12 = tmp10 / tmp11
tl.store(in_out_ptr0 + (x4), tmp12, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6b/c6busvilz5nn36jjet3bmw7cqddirh4sgalamjr3fsrp3sbsacfi.py
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# contiguous => clone_4
# Graph fragment:
# %clone_4 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_6,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_4 = async_compile.triton('triton_poi_fused_clone_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tl.store(out_ptr0 + (x4), tmp0, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16, ), (1, ))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16, ), (1, ))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16, ), (1, ))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_11, (4, 16), (16, 1))
assert_size_stride(primals_12, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(buf0, primals_4, buf3, 256, grid=grid(256), stream=stream0)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf1, primals_6, buf4, 64, 4, grid=grid(64, 4), stream=stream0)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
triton_poi_fused__softmax_add_clamp_div_log_sqrt_2.run(primals_10, buf5, buf6, buf7, 64, grid=grid(64), stream=stream0)
buf8 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
triton_poi_fused__softmax_add_clamp_div_log_sqrt_3.run(buf8, primals_10, buf6, buf7, 256, grid=grid(256), stream=stream0)
del buf6
del primals_10
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.clone]
triton_poi_fused_clone_0.run(buf2, primals_8, buf9, 256, grid=grid(256), stream=stream0)
del primals_8
buf10 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0); del buf2 # reuse
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1), 0), out=buf10)
buf11 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
triton_poi_fused_clone_4.run(buf10, buf11, 256, grid=grid(256), stream=stream0)
del buf10
buf12 = reinterpret_tensor(buf7, (16, 4), (4, 1), 0); del buf7 # reuse
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_12, reinterpret_tensor(buf11, (16, 16), (16, 1), 0), reinterpret_tensor(primals_11, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf12)
del primals_12
return (reinterpret_tensor(buf12, (4, 4, 4), (16, 4, 1), 0), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), buf8, reinterpret_tensor(buf11, (16, 16), (16, 1), 0), primals_11, reinterpret_tensor(buf9, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_12 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class ScaledDotProductWithBoxAttention(nn.Module):
"""
Scaled dot-product attention with box
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1, comment=None):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductWithBoxAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
self.comment = comment
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
w_g = box_relation_embed_matrix
w_a = att
w_mn = torch.log(torch.clamp(w_g, min=1e-06)) + w_a
w_mn = torch.softmax(w_mn, -1)
att = self.dropout(w_mn)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4,
4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'd_k': 4, 'd_v': 4, 'h': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
from torch.autograd import *
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = yindex // 16
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 16 * x2 + 64 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_2(in_ptr0, in_ptr1,
out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + 4 * x2, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr1 + (1 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp15 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr1 + (2 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp22 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = triton_helpers.maximum(tmp8, tmp1)
tmp10 = tl_math.log(tmp9)
tmp12 = tmp11 / tmp5
tmp13 = tmp10 + tmp12
tmp14 = triton_helpers.maximum(tmp7, tmp13)
tmp16 = triton_helpers.maximum(tmp15, tmp1)
tmp17 = tl_math.log(tmp16)
tmp19 = tmp18 / tmp5
tmp20 = tmp17 + tmp19
tmp21 = triton_helpers.maximum(tmp14, tmp20)
tmp23 = triton_helpers.maximum(tmp22, tmp1)
tmp24 = tl_math.log(tmp23)
tmp26 = tmp25 / tmp5
tmp27 = tmp24 + tmp26
tmp28 = triton_helpers.maximum(tmp21, tmp27)
tmp29 = tmp7 - tmp28
tmp30 = tl_math.exp(tmp29)
tmp31 = tmp13 - tmp28
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp34 = tmp20 - tmp28
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tmp37 = tmp27 - tmp28
tmp38 = tl_math.exp(tmp37)
tmp39 = tmp36 + tmp38
tl.store(out_ptr0 + x2, tmp28, xmask)
tl.store(out_ptr1 + x2, tmp39, xmask)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_3(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex % 64
x4 = xindex
x5 = xindex // 4
tmp0 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_out_ptr0 + x4, xmask)
tmp8 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr2 + x5, xmask, eviction_policy='evict_last')
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = tmp7 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp12 = tmp10 / tmp11
tl.store(in_out_ptr0 + x4, tmp12, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tl.store(out_ptr0 + x4, tmp0, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16,), (1,))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16,), (1,))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_11, (4, 16), (16, 1))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(256)](buf0, primals_4, buf3, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_clone_1[grid(64, 4)](buf1, primals_6, buf4, 64, 4,
XBLOCK=4, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_add_clamp_div_log_sqrt_2[grid(64)](primals_10
, buf5, buf6, buf7, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf8 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused__softmax_add_clamp_div_log_sqrt_3[grid(256)](buf8,
primals_10, buf6, buf7, 256, XBLOCK=256, num_warps=4, num_stages=1)
del buf6
del primals_10
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_0[grid(256)](buf2, primals_8, buf9, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_8
buf10 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1), 0), out=buf10)
buf11 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_4[grid(256)](buf10, buf11, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf10
buf12 = reinterpret_tensor(buf7, (16, 4), (4, 1), 0)
del buf7
extern_kernels.addmm(primals_12, reinterpret_tensor(buf11, (16, 16),
(16, 1), 0), reinterpret_tensor(primals_11, (16, 4), (1, 16), 0
), alpha=1, beta=1, out=buf12)
del primals_12
return reinterpret_tensor(buf12, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_2, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf8, reinterpret_tensor(buf11, (16, 16), (16, 1), 0
), primals_11, reinterpret_tensor(buf9, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0)
class ScaledDotProductWithBoxAttentionNew(nn.Module):
"""
Scaled dot-product attention with box
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1, comment=None):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductWithBoxAttentionNew, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
self.comment = comment
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, input_0, input_1, input_2, input_3):
primals_3 = self.fc_q.weight
primals_4 = self.fc_q.bias
primals_5 = self.fc_k.weight
primals_6 = self.fc_k.bias
primals_7 = self.fc_v.weight
primals_8 = self.fc_v.bias
primals_11 = self.fc_o.weight
primals_12 = self.fc_o.bias
primals_1 = input_0
primals_2 = input_1
primals_9 = input_2
primals_10 = input_3
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
| Maxi-0902/DRAN | ScaledDotProductWithBoxAttention | false | 849 | [
"MIT"
] | 0 | c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | https://github.com/Maxi-0902/DRAN/tree/c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class Model(nn.Module):
"""
Scaled dot-product attention with box
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1, comment=None):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super().__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
self.comment = comment
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
w_g = box_relation_embed_matrix
w_a = att
w_mn = torch.log(torch.clamp(w_g, min=1e-06)) + w_a
w_mn = torch.softmax(w_mn, -1)
att = self.dropout(w_mn)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4,
4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4, 4, 4]
|
GlobalAveragePooling | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/is/cispe7zbbl4nxt2jjus6h5iou2w7htohqj7z2oz6g7nqz6vbpbqr.py
# Topologically Sorted Source Nodes: [avg_pool2d], Original ATen: [aten.avg_pool2d]
# Source node to ATen node mapping:
# avg_pool2d => avg_pool2d
# Graph fragment:
# %avg_pool2d : [num_users=1] = call_function[target=torch.ops.aten.avg_pool2d.default](args = (%arg0_1, [4, 4]), kwargs = {})
triton_poi_fused_avg_pool2d_0 = async_compile.triton('triton_poi_fused_avg_pool2d_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_avg_pool2d_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 16, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_avg_pool2d_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (16*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (16*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (16*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (16*x0)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (4 + (16*x0)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (5 + (16*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (6 + (16*x0)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr0 + (7 + (16*x0)), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (8 + (16*x0)), xmask, eviction_policy='evict_last')
tmp17 = tl.load(in_ptr0 + (9 + (16*x0)), xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_ptr0 + (10 + (16*x0)), xmask, eviction_policy='evict_last')
tmp21 = tl.load(in_ptr0 + (11 + (16*x0)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr0 + (12 + (16*x0)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr0 + (13 + (16*x0)), xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr0 + (14 + (16*x0)), xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr0 + (15 + (16*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 + tmp0
tmp4 = tmp3 + tmp2
tmp6 = tmp5 + tmp4
tmp8 = tmp7 + tmp6
tmp10 = tmp9 + tmp8
tmp12 = tmp11 + tmp10
tmp14 = tmp13 + tmp12
tmp16 = tmp15 + tmp14
tmp18 = tmp17 + tmp16
tmp20 = tmp19 + tmp18
tmp22 = tmp21 + tmp20
tmp24 = tmp23 + tmp22
tmp26 = tmp25 + tmp24
tmp28 = tmp27 + tmp26
tmp30 = tmp29 + tmp28
tmp31 = 0.0625
tmp32 = tmp30 * tmp31
tl.store(out_ptr0 + (x0), tmp32, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [avg_pool2d], Original ATen: [aten.avg_pool2d]
stream0 = get_raw_stream(0)
triton_poi_fused_avg_pool2d_0.run(arg0_1, buf0, 16, grid=grid(16), stream=stream0)
del arg0_1
return (reinterpret_tensor(buf0, (4, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
class GlobalAveragePooling(nn.Module):
def __init__(self):
super(GlobalAveragePooling, self).__init__()
def forward(self, feat):
num_channels = feat.size(1)
return F.avg_pool2d(feat, (feat.size(2), feat.size(3))).view(-1,
num_channels)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_avg_pool2d_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 16 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp3 = tl.load(in_ptr0 + (2 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp5 = tl.load(in_ptr0 + (3 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp7 = tl.load(in_ptr0 + (4 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp9 = tl.load(in_ptr0 + (5 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp11 = tl.load(in_ptr0 + (6 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp13 = tl.load(in_ptr0 + (7 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp15 = tl.load(in_ptr0 + (8 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp17 = tl.load(in_ptr0 + (9 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp19 = tl.load(in_ptr0 + (10 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp21 = tl.load(in_ptr0 + (11 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr0 + (12 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp25 = tl.load(in_ptr0 + (13 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp27 = tl.load(in_ptr0 + (14 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp29 = tl.load(in_ptr0 + (15 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp2 = tmp1 + tmp0
tmp4 = tmp3 + tmp2
tmp6 = tmp5 + tmp4
tmp8 = tmp7 + tmp6
tmp10 = tmp9 + tmp8
tmp12 = tmp11 + tmp10
tmp14 = tmp13 + tmp12
tmp16 = tmp15 + tmp14
tmp18 = tmp17 + tmp16
tmp20 = tmp19 + tmp18
tmp22 = tmp21 + tmp20
tmp24 = tmp23 + tmp22
tmp26 = tmp25 + tmp24
tmp28 = tmp27 + tmp26
tmp30 = tmp29 + tmp28
tmp31 = 0.0625
tmp32 = tmp30 * tmp31
tl.store(out_ptr0 + x0, tmp32, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_avg_pool2d_0[grid(16)](arg0_1, buf0, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del arg0_1
return reinterpret_tensor(buf0, (4, 4), (4, 1), 0),
class GlobalAveragePoolingNew(nn.Module):
def __init__(self):
super(GlobalAveragePoolingNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MrChenFeng/Project-Template | GlobalAveragePooling | false | 850 | [
"MIT"
] | 0 | 42a335c6abb710bbae6407cbb0ca461533bc12f9 | https://github.com/MrChenFeng/Project-Template/tree/42a335c6abb710bbae6407cbb0ca461533bc12f9 | import torch
import torch.nn as nn
import torch.nn.functional as F
class Model(nn.Module):
def __init__(self):
super().__init__()
def forward(self, feat):
num_channels = feat.size(1)
return F.avg_pool2d(feat, (feat.size(2), feat.size(3))).view(-1,
num_channels)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
RadialPredictionLayer | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/fv/cfvgww2wivyppqi3fi2tgwfgcxo2txzj7xtl2zvn66oo6sg4rwba.py
# Topologically Sorted Source Nodes: [sub, pow_1, sum_1, sqrt, neg], Original ATen: [aten.sub, aten.pow, aten.sum, aten.sqrt, aten.neg]
# Source node to ATen node mapping:
# neg => neg
# pow_1 => pow_1
# sqrt => sqrt
# sub => sub
# sum_1 => sum_1
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%unsqueeze, %unsqueeze_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [2]), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%sum_1,), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%sqrt,), kwargs = {})
triton_poi_fused_neg_pow_sqrt_sub_sum_0 = async_compile.triton('triton_poi_fused_neg_pow_sqrt_sub_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_neg_pow_sqrt_sub_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_neg_pow_sqrt_sub_sum_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = (xindex // 16)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (64*x1)), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (16 + x0 + (64*x1)), xmask)
tmp8 = tl.load(in_ptr0 + (32 + x0 + (64*x1)), xmask)
tmp12 = tl.load(in_ptr0 + (48 + x0 + (64*x1)), xmask)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp5 = tmp4 - tmp1
tmp6 = tmp5 * tmp5
tmp7 = tmp3 + tmp6
tmp9 = tmp8 - tmp1
tmp10 = tmp9 * tmp9
tmp11 = tmp7 + tmp10
tmp13 = tmp12 - tmp1
tmp14 = tmp13 * tmp13
tmp15 = tmp11 + tmp14
tmp16 = libdevice.sqrt(tmp15)
tmp17 = -tmp16
tl.store(out_ptr0 + (x2), tmp17, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1, 4, 4), (16, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sub, pow_1, sum_1, sqrt, neg], Original ATen: [aten.sub, aten.pow, aten.sum, aten.sqrt, aten.neg]
stream0 = get_raw_stream(0)
triton_poi_fused_neg_pow_sqrt_sub_sum_0.run(arg0_1, arg1_1, buf0, 64, grid=grid(64), stream=stream0)
del arg0_1
del arg1_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class RadialPredictionLayer(torch.nn.Module):
""" The RPL classification layer with fixed prototypes
"""
def __init__(self, in_features, out_features):
super(RadialPredictionLayer, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.prototypes = nn.Parameter(torch.diag(torch.ones(self.
in_features)), requires_grad=False)
def forward(self, x):
return -((x[:, None, :] - self.prototypes[None, :, :]) ** 2).sum(dim=2
).sqrt()
def extra_repr(self):
return 'in_features={}, out_features={}'.format(self.in_features,
self.out_features)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_features': 4, 'out_features': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_neg_pow_sqrt_sub_sum_0(in_ptr0, in_ptr1, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)
tmp8 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp12 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp5 = tmp4 - tmp1
tmp6 = tmp5 * tmp5
tmp7 = tmp3 + tmp6
tmp9 = tmp8 - tmp1
tmp10 = tmp9 * tmp9
tmp11 = tmp7 + tmp10
tmp13 = tmp12 - tmp1
tmp14 = tmp13 * tmp13
tmp15 = tmp11 + tmp14
tmp16 = libdevice.sqrt(tmp15)
tmp17 = -tmp16
tl.store(out_ptr0 + x2, tmp17, xmask)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1, 4, 4), (16, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_neg_pow_sqrt_sub_sum_0[grid(64)](arg0_1, arg1_1,
buf0, 64, XBLOCK=64, num_warps=1, num_stages=1)
del arg0_1
del arg1_1
return buf0,
class RadialPredictionLayerNew(torch.nn.Module):
""" The RPL classification layer with fixed prototypes
"""
def __init__(self, in_features, out_features):
super(RadialPredictionLayerNew, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.prototypes = nn.Parameter(torch.diag(torch.ones(self.
in_features)), requires_grad=False)
def extra_repr(self):
return 'in_features={}, out_features={}'.format(self.in_features,
self.out_features)
def forward(self, input_0):
arg1_1 = self.prototypes
arg0_1 = input_0
output = call([arg0_1, arg1_1])
return output[0]
| Monkso/RPL-Softmax_RoadSigns | RadialPredictionLayer | false | 851 | [
"MIT"
] | 0 | 3df929d779ff02ec796e717659943bb46311ba0f | https://github.com/Monkso/RPL-Softmax_RoadSigns/tree/3df929d779ff02ec796e717659943bb46311ba0f | import torch
import torch.nn as nn
class Model(torch.nn.Module):
""" The RPL classification layer with fixed prototypes
"""
def __init__(self, in_features, out_features):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.prototypes = nn.Parameter(torch.diag(torch.ones(self.
in_features)), requires_grad=False)
def forward(self, x):
return -((x[:, None, :] - self.prototypes[None, :, :]) ** 2).sum(dim=2
).sqrt()
def extra_repr(self):
return 'in_features={}, out_features={}'.format(self.in_features,
self.out_features)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
BBoxTransform | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/6s/c6sbsxdzyxnyaoinhkjjxf63e3bablnuxukt3ntviotrgdiysvm5.py
# Topologically Sorted Source Nodes: [stack], Original ATen: [aten.stack]
# Source node to ATen node mapping:
# stack => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%sub_3, %sub_2, %add_5, %add_4], 2), kwargs = {})
triton_poi_fused_stack_0 = async_compile.triton('triton_poi_fused_stack_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_stack_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 16, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_stack_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = (xindex // 16)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (1 + (4*x0) + (16*x1)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tl.load(in_ptr1 + (3 + (4*x0) + (16*x1)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp7 = tl.load(in_ptr1 + (1 + (4*x0) + (16*x1)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp8 = tmp6 - tmp7
tmp9 = tmp5 * tmp8
tmp10 = tmp7 + tmp6
tmp11 = 0.5
tmp12 = tmp10 * tmp11
tmp13 = tmp9 + tmp12
tmp14 = tl.load(in_ptr0 + (3 + (4*x0) + (16*x1)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp15 * tmp8
tmp17 = tmp16 * tmp11
tmp18 = tmp13 - tmp17
tmp19 = tl.full(tmp18.shape, 0.0, tmp18.dtype)
tmp20 = tl.where(tmp4, tmp18, tmp19)
tmp21 = tmp0 >= tmp3
tmp22 = tl.full([1], 8, tl.int64)
tmp23 = tmp0 < tmp22
tmp24 = tmp21 & tmp23
tmp25 = tl.load(in_ptr0 + ((4*((-4) + x0)) + (16*x1)), tmp24 & xmask, eviction_policy='evict_last', other=0.0)
tmp26 = tl.load(in_ptr1 + (2 + (4*((-4) + x0)) + (16*x1)), tmp24 & xmask, eviction_policy='evict_last', other=0.0)
tmp27 = tl.load(in_ptr1 + ((4*((-4) + x0)) + (16*x1)), tmp24 & xmask, eviction_policy='evict_last', other=0.0)
tmp28 = tmp26 - tmp27
tmp29 = tmp25 * tmp28
tmp30 = tmp27 + tmp26
tmp31 = tmp30 * tmp11
tmp32 = tmp29 + tmp31
tmp33 = tl.load(in_ptr0 + (2 + (4*((-4) + x0)) + (16*x1)), tmp24 & xmask, eviction_policy='evict_last', other=0.0)
tmp34 = tl_math.exp(tmp33)
tmp35 = tmp34 * tmp28
tmp36 = tmp35 * tmp11
tmp37 = tmp32 - tmp36
tmp38 = tl.full(tmp37.shape, 0.0, tmp37.dtype)
tmp39 = tl.where(tmp24, tmp37, tmp38)
tmp40 = tmp0 >= tmp22
tmp41 = tl.full([1], 12, tl.int64)
tmp42 = tmp0 < tmp41
tmp43 = tmp40 & tmp42
tmp44 = tl.load(in_ptr0 + (1 + (4*((-8) + x0)) + (16*x1)), tmp43 & xmask, eviction_policy='evict_last', other=0.0)
tmp45 = tl.load(in_ptr1 + (3 + (4*((-8) + x0)) + (16*x1)), tmp43 & xmask, eviction_policy='evict_last', other=0.0)
tmp46 = tl.load(in_ptr1 + (1 + (4*((-8) + x0)) + (16*x1)), tmp43 & xmask, eviction_policy='evict_last', other=0.0)
tmp47 = tmp45 - tmp46
tmp48 = tmp44 * tmp47
tmp49 = tmp46 + tmp45
tmp50 = tmp49 * tmp11
tmp51 = tmp48 + tmp50
tmp52 = tl.load(in_ptr0 + (3 + (4*((-8) + x0)) + (16*x1)), tmp43 & xmask, eviction_policy='evict_last', other=0.0)
tmp53 = tl_math.exp(tmp52)
tmp54 = tmp53 * tmp47
tmp55 = tmp54 * tmp11
tmp56 = tmp51 + tmp55
tmp57 = tl.full(tmp56.shape, 0.0, tmp56.dtype)
tmp58 = tl.where(tmp43, tmp56, tmp57)
tmp59 = tmp0 >= tmp41
tmp60 = tl.full([1], 16, tl.int64)
tmp61 = tmp0 < tmp60
tmp62 = tl.load(in_ptr0 + ((4*((-12) + x0)) + (16*x1)), tmp59 & xmask, eviction_policy='evict_last', other=0.0)
tmp63 = tl.load(in_ptr1 + (2 + (4*((-12) + x0)) + (16*x1)), tmp59 & xmask, eviction_policy='evict_last', other=0.0)
tmp64 = tl.load(in_ptr1 + ((4*((-12) + x0)) + (16*x1)), tmp59 & xmask, eviction_policy='evict_last', other=0.0)
tmp65 = tmp63 - tmp64
tmp66 = tmp62 * tmp65
tmp67 = tmp64 + tmp63
tmp68 = tmp67 * tmp11
tmp69 = tmp66 + tmp68
tmp70 = tl.load(in_ptr0 + (2 + (4*((-12) + x0)) + (16*x1)), tmp59 & xmask, eviction_policy='evict_last', other=0.0)
tmp71 = tl_math.exp(tmp70)
tmp72 = tmp71 * tmp65
tmp73 = tmp72 * tmp11
tmp74 = tmp69 + tmp73
tmp75 = tl.full(tmp74.shape, 0.0, tmp74.dtype)
tmp76 = tl.where(tmp59, tmp74, tmp75)
tmp77 = tl.where(tmp43, tmp58, tmp76)
tmp78 = tl.where(tmp24, tmp39, tmp77)
tmp79 = tl.where(tmp4, tmp20, tmp78)
tl.store(out_ptr0 + (x2), tmp79, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 16), (64, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [stack], Original ATen: [aten.stack]
stream0 = get_raw_stream(0)
triton_poi_fused_stack_0.run(arg1_1, arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
del arg1_1
return (reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class BBoxTransform(nn.Module):
def forward(self, anchors, regression):
"""
Args:
anchors: [batch_size, boxes, (y1, x1, y2, x2)]
regression: [batch_size, boxes, (dy, dx, dh, dw)]
"""
y_centers_a = (anchors[..., 0] + anchors[..., 2]) / 2
x_centers_a = (anchors[..., 1] + anchors[..., 3]) / 2
ha = anchors[..., 2] - anchors[..., 0]
wa = anchors[..., 3] - anchors[..., 1]
w = regression[..., 3].exp() * wa
h = regression[..., 2].exp() * ha
y_centers = regression[..., 0] * ha + y_centers_a
x_centers = regression[..., 1] * wa + x_centers_a
ymin = y_centers - h / 2.0
xmin = x_centers - w / 2.0
ymax = y_centers + h / 2.0
xmax = x_centers + w / 2.0
return torch.stack([xmin, ymin, xmax, ymax], dim=2)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_stack_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (1 + 4 * x0 + 16 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp6 = tl.load(in_ptr1 + (3 + 4 * x0 + 16 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp7 = tl.load(in_ptr1 + (1 + 4 * x0 + 16 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp8 = tmp6 - tmp7
tmp9 = tmp5 * tmp8
tmp10 = tmp7 + tmp6
tmp11 = 0.5
tmp12 = tmp10 * tmp11
tmp13 = tmp9 + tmp12
tmp14 = tl.load(in_ptr0 + (3 + 4 * x0 + 16 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp15 * tmp8
tmp17 = tmp16 * tmp11
tmp18 = tmp13 - tmp17
tmp19 = tl.full(tmp18.shape, 0.0, tmp18.dtype)
tmp20 = tl.where(tmp4, tmp18, tmp19)
tmp21 = tmp0 >= tmp3
tmp22 = tl.full([1], 8, tl.int64)
tmp23 = tmp0 < tmp22
tmp24 = tmp21 & tmp23
tmp25 = tl.load(in_ptr0 + (4 * (-4 + x0) + 16 * x1), tmp24 & xmask,
eviction_policy='evict_last', other=0.0)
tmp26 = tl.load(in_ptr1 + (2 + 4 * (-4 + x0) + 16 * x1), tmp24 & xmask,
eviction_policy='evict_last', other=0.0)
tmp27 = tl.load(in_ptr1 + (4 * (-4 + x0) + 16 * x1), tmp24 & xmask,
eviction_policy='evict_last', other=0.0)
tmp28 = tmp26 - tmp27
tmp29 = tmp25 * tmp28
tmp30 = tmp27 + tmp26
tmp31 = tmp30 * tmp11
tmp32 = tmp29 + tmp31
tmp33 = tl.load(in_ptr0 + (2 + 4 * (-4 + x0) + 16 * x1), tmp24 & xmask,
eviction_policy='evict_last', other=0.0)
tmp34 = tl_math.exp(tmp33)
tmp35 = tmp34 * tmp28
tmp36 = tmp35 * tmp11
tmp37 = tmp32 - tmp36
tmp38 = tl.full(tmp37.shape, 0.0, tmp37.dtype)
tmp39 = tl.where(tmp24, tmp37, tmp38)
tmp40 = tmp0 >= tmp22
tmp41 = tl.full([1], 12, tl.int64)
tmp42 = tmp0 < tmp41
tmp43 = tmp40 & tmp42
tmp44 = tl.load(in_ptr0 + (1 + 4 * (-8 + x0) + 16 * x1), tmp43 & xmask,
eviction_policy='evict_last', other=0.0)
tmp45 = tl.load(in_ptr1 + (3 + 4 * (-8 + x0) + 16 * x1), tmp43 & xmask,
eviction_policy='evict_last', other=0.0)
tmp46 = tl.load(in_ptr1 + (1 + 4 * (-8 + x0) + 16 * x1), tmp43 & xmask,
eviction_policy='evict_last', other=0.0)
tmp47 = tmp45 - tmp46
tmp48 = tmp44 * tmp47
tmp49 = tmp46 + tmp45
tmp50 = tmp49 * tmp11
tmp51 = tmp48 + tmp50
tmp52 = tl.load(in_ptr0 + (3 + 4 * (-8 + x0) + 16 * x1), tmp43 & xmask,
eviction_policy='evict_last', other=0.0)
tmp53 = tl_math.exp(tmp52)
tmp54 = tmp53 * tmp47
tmp55 = tmp54 * tmp11
tmp56 = tmp51 + tmp55
tmp57 = tl.full(tmp56.shape, 0.0, tmp56.dtype)
tmp58 = tl.where(tmp43, tmp56, tmp57)
tmp59 = tmp0 >= tmp41
tl.full([1], 16, tl.int64)
tmp62 = tl.load(in_ptr0 + (4 * (-12 + x0) + 16 * x1), tmp59 & xmask,
eviction_policy='evict_last', other=0.0)
tmp63 = tl.load(in_ptr1 + (2 + 4 * (-12 + x0) + 16 * x1), tmp59 & xmask,
eviction_policy='evict_last', other=0.0)
tmp64 = tl.load(in_ptr1 + (4 * (-12 + x0) + 16 * x1), tmp59 & xmask,
eviction_policy='evict_last', other=0.0)
tmp65 = tmp63 - tmp64
tmp66 = tmp62 * tmp65
tmp67 = tmp64 + tmp63
tmp68 = tmp67 * tmp11
tmp69 = tmp66 + tmp68
tmp70 = tl.load(in_ptr0 + (2 + 4 * (-12 + x0) + 16 * x1), tmp59 & xmask,
eviction_policy='evict_last', other=0.0)
tmp71 = tl_math.exp(tmp70)
tmp72 = tmp71 * tmp65
tmp73 = tmp72 * tmp11
tmp74 = tmp69 + tmp73
tmp75 = tl.full(tmp74.shape, 0.0, tmp74.dtype)
tmp76 = tl.where(tmp59, tmp74, tmp75)
tmp77 = tl.where(tmp43, tmp58, tmp76)
tmp78 = tl.where(tmp24, tmp39, tmp77)
tmp79 = tl.where(tmp4, tmp20, tmp78)
tl.store(out_ptr0 + x2, tmp79, xmask)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 16), (64, 16, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_stack_0[grid(256)](arg1_1, arg0_1, buf0, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del arg0_1
del arg1_1
return reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0),
class BBoxTransformNew(nn.Module):
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| MikoyChinese/Learn | BBoxTransform | false | 852 | [
"Apache-2.0"
] | 0 | c482b1e84496279935b5bb2cfc1e6d78e2868c63 | https://github.com/MikoyChinese/Learn/tree/c482b1e84496279935b5bb2cfc1e6d78e2868c63 | import torch
import torch.nn as nn
class Model(nn.Module):
def forward(self, anchors, regression):
"""
Args:
anchors: [batch_size, boxes, (y1, x1, y2, x2)]
regression: [batch_size, boxes, (dy, dx, dh, dw)]
"""
y_centers_a = (anchors[..., 0] + anchors[..., 2]) / 2
x_centers_a = (anchors[..., 1] + anchors[..., 3]) / 2
ha = anchors[..., 2] - anchors[..., 0]
wa = anchors[..., 3] - anchors[..., 1]
w = regression[..., 3].exp() * wa
h = regression[..., 2].exp() * ha
y_centers = regression[..., 0] * ha + y_centers_a
x_centers = regression[..., 1] * wa + x_centers_a
ymin = y_centers - h / 2.0
xmin = x_centers - w / 2.0
ymax = y_centers + h / 2.0
xmax = x_centers + w / 2.0
return torch.stack([xmin, ymin, xmax, ymax], dim=2)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
weight_quantize_fn | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xy/cxy5n3qjrhqoemybkvmtvjinhiprajrrxnbbem4oeydlrjod5nzb.py
# Topologically Sorted Source Nodes: [weight, truediv, abs_1, max_1, truediv_1, weight_1], Original ATen: [aten.tanh, aten.div, aten.abs, aten.max, aten.add]
# Source node to ATen node mapping:
# abs_1 => abs_1
# max_1 => max_1
# truediv => div
# truediv_1 => div_1
# weight => tanh
# weight_1 => add
# Graph fragment:
# %tanh : [num_users=2] = call_function[target=torch.ops.aten.tanh.default](args = (%arg0_1,), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%tanh, 2), kwargs = {})
# %abs_1 : [num_users=1] = call_function[target=torch.ops.aten.abs.default](args = (%tanh,), kwargs = {})
# %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.default](args = (%abs_1,), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%div, %max_1), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div_1, 0.5), kwargs = {})
triton_per_fused_abs_add_div_max_tanh_0 = async_compile.triton('triton_per_fused_abs_add_div_max_tanh_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_abs_add_div_max_tanh_0', 'mutated_arg_names': [], 'no_x_dim': True, 'num_load': 1, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_abs_add_div_max_tanh_0(in_ptr0, out_ptr0, out_ptr1, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = libdevice.tanh(tmp0)
tmp2 = tl_math.abs(tmp1)
tmp3 = tl.broadcast_to(tmp2, [RBLOCK])
tmp5 = triton_helpers.promote_to_tensor(triton_helpers.max2(tmp3, 0))
tmp6 = 0.5
tmp7 = tmp1 * tmp6
tmp8 = tmp7 / tmp5
tmp9 = tmp8 + tmp6
tl.store(out_ptr1 + (tl.broadcast_to(r0, [RBLOCK])), tmp9, None)
tl.store(out_ptr0 + (tl.full([1], 0, tl.int32)), tmp5, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [weight, truediv, abs_1, max_1, truediv_1, weight_1], Original ATen: [aten.tanh, aten.div, aten.abs, aten.max, aten.add]
stream0 = get_raw_stream(0)
triton_per_fused_abs_add_div_max_tanh_0.run(arg0_1, buf0, buf1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
return (buf0, buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.utils.data
import torch.nn as nn
def uniform_quantize(k):
class qfn(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if k == 32:
out = input
elif k == 1:
out = torch.sign(input)
else:
n = float(2 ** k - 1)
out = torch.round(input * n) / n
return out
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
return qfn().apply
class weight_quantize_fn(nn.Module):
def __init__(self, w_bit):
super(weight_quantize_fn, self).__init__()
assert w_bit <= 8 or w_bit == 32
self.w_bit = w_bit
self.uniform_q = uniform_quantize(k=w_bit)
def forward(self, x):
if self.w_bit == 32:
weight_q = x
elif self.w_bit == 1:
E = torch.mean(torch.abs(x)).detach()
weight_q = self.uniform_q(x / E) * E
else:
weight = torch.tanh(x)
max_w = torch.max(torch.abs(weight)).detach()
weight = weight / 2 / max_w + 0.5
weight_q = max_w * (2 * self.uniform_q(weight) - 1)
return weight_q
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'w_bit': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.utils.data
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_abs_add_div_max_tanh_0(in_ptr0, out_ptr0, out_ptr1,
xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = libdevice.tanh(tmp0)
tmp2 = tl_math.abs(tmp1)
tmp3 = tl.broadcast_to(tmp2, [RBLOCK])
tmp5 = triton_helpers.promote_to_tensor(triton_helpers.max2(tmp3, 0))
tmp6 = 0.5
tmp7 = tmp1 * tmp6
tmp8 = tmp7 / tmp5
tmp9 = tmp8 + tmp6
tl.store(out_ptr1 + tl.broadcast_to(r0, [RBLOCK]), tmp9, None)
tl.store(out_ptr0 + tl.full([1], 0, tl.int32), tmp5, None)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_per_fused_abs_add_div_max_tanh_0[grid(1)](arg0_1, buf0, buf1,
1, 256, num_warps=2, num_stages=1)
del arg0_1
return buf0, buf1
def uniform_quantize(k):
class qfn(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if k == 32:
out = input
elif k == 1:
out = torch.sign(input)
else:
n = float(2 ** k - 1)
out = torch.round(input * n) / n
return out
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
return qfn().apply
class weight_quantize_fnNew(nn.Module):
def __init__(self, w_bit):
super(weight_quantize_fnNew, self).__init__()
assert w_bit <= 8 or w_bit == 32
self.w_bit = w_bit
self.uniform_q = uniform_quantize(k=w_bit)
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MohammedHAlali/pytorch_DoReFaNet | weight_quantize_fn | false | 853 | [
"MIT"
] | 0 | d208089b9172f02c09cc6633158ed5b5d6cd7f1e | https://github.com/MohammedHAlali/pytorch_DoReFaNet/tree/d208089b9172f02c09cc6633158ed5b5d6cd7f1e | import torch
import torch.utils.data
import torch.nn as nn
def uniform_quantize(k):
class qfn(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if k == 32:
out = input
elif k == 1:
out = torch.sign(input)
else:
n = float(2 ** k - 1)
out = torch.round(input * n) / n
return out
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input
return qfn().apply
class Model(nn.Module):
def __init__(self, w_bit):
super().__init__()
assert w_bit <= 8 or w_bit == 32
self.w_bit = w_bit
self.uniform_q = uniform_quantize(k=w_bit)
def forward(self, x):
if self.w_bit == 32:
weight_q = x
elif self.w_bit == 1:
E = torch.mean(torch.abs(x)).detach()
weight_q = self.uniform_q(x / E) * E
else:
weight = torch.tanh(x)
max_w = torch.max(torch.abs(weight)).detach()
weight = weight / 2 / max_w + 0.5
weight_q = max_w * (2 * self.uniform_q(weight) - 1)
return weight_q
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
RBFLayer | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/zk/czktkx3bod3v62xyjig5yabz7tjg5pp6cdzcusscv2i5rh24prte.py
# Topologically Sorted Source Nodes: [sub, pow_1, sum_1, pow_2, distances], Original ATen: [aten.sub, aten.pow, aten.sum, aten.div]
# Source node to ATen node mapping:
# distances => div
# pow_1 => pow_1
# pow_2 => pow_2
# sub => sub
# sum_1 => sum_1
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%expand, %expand_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [-1]), kwargs = {})
# %pow_2 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_1, 0.5), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%pow_2, %unsqueeze_2), kwargs = {})
triton_poi_fused_div_pow_sub_sum_0 = async_compile.triton('triton_poi_fused_div_pow_sub_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_div_pow_sub_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 9, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_div_pow_sub_sum_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x0 = xindex % 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4*x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr1 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr1 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp6 = tmp4 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tmp3 + tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp19 = libdevice.sqrt(tmp18)
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 / tmp21
tl.store(out_ptr0 + (x2), tmp22, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sub, pow_1, sum_1, pow_2, distances], Original ATen: [aten.sub, aten.pow, aten.sum, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_div_pow_sub_sum_0.run(primals_1, primals_2, primals_3, buf0, 16, grid=grid(16), stream=stream0)
return (buf0, primals_1, primals_2, primals_3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class RBFLayer(nn.Module):
"""
Transforms incoming data using a given radial basis function:
u_{i} = rbf(||x - c_{i}|| / s_{i})
Arguments:
in_features: size of each input sample
out_features: size of each output sample
Shape:
- Input: (N, in_features) where N is an arbitrary batch size
- Output: (N, out_features) where N is an arbitrary batch size
Attributes:
centres: the learnable centres of shape (out_features, in_features).
The values are initialised from a standard normal distribution.
Normalising inputs to have mean 0 and standard deviation 1 is
recommended.
log_sigmas: logarithm of the learnable scaling factors of shape (out_features).
basis_func: the radial basis function used to transform the scaled
distances.
"""
def __init__(self, in_features, out_features, basis_func):
super(RBFLayer, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.centres = nn.Parameter(torch.Tensor(out_features, in_features))
self.log_sigmas = nn.Parameter(torch.Tensor(out_features))
self.basis_func = basis_func
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.centres, 0, 1)
nn.init.constant_(self.log_sigmas, 0)
def forward(self, input):
size = input.size(0), self.out_features, self.in_features
x = input.unsqueeze(1).expand(size)
c = self.centres.unsqueeze(0).expand(size)
distances = (x - c).pow(2).sum(-1).pow(0.5) / torch.exp(self.log_sigmas
).unsqueeze(0)
return distances
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [[], {'in_features': 4, 'out_features': 4, 'basis_func': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_div_pow_sub_sum_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x0 = xindex % 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp14 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp15 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp6 = tmp4 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tmp3 + tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp19 = libdevice.sqrt(tmp18)
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 / tmp21
tl.store(out_ptr0 + x2, tmp22, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_div_pow_sub_sum_0[grid(16)](primals_1, primals_2,
primals_3, buf0, 16, XBLOCK=16, num_warps=1, num_stages=1)
return buf0, primals_1, primals_2, primals_3
class RBFLayerNew(nn.Module):
"""
Transforms incoming data using a given radial basis function:
u_{i} = rbf(||x - c_{i}|| / s_{i})
Arguments:
in_features: size of each input sample
out_features: size of each output sample
Shape:
- Input: (N, in_features) where N is an arbitrary batch size
- Output: (N, out_features) where N is an arbitrary batch size
Attributes:
centres: the learnable centres of shape (out_features, in_features).
The values are initialised from a standard normal distribution.
Normalising inputs to have mean 0 and standard deviation 1 is
recommended.
log_sigmas: logarithm of the learnable scaling factors of shape (out_features).
basis_func: the radial basis function used to transform the scaled
distances.
"""
def __init__(self, in_features, out_features, basis_func):
super(RBFLayerNew, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.centres = nn.Parameter(torch.Tensor(out_features, in_features))
self.log_sigmas = nn.Parameter(torch.Tensor(out_features))
self.basis_func = basis_func
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.centres, 0, 1)
nn.init.constant_(self.log_sigmas, 0)
def forward(self, input_0):
primals_1 = self.centres
primals_3 = self.log_sigmas
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| MorganeAyle/SNIP-it | RBFLayer | false | 854 | [
"MIT"
] | 0 | df2bf44d6d3f7e4ea7733242a79c916735a7b49e | https://github.com/MorganeAyle/SNIP-it/tree/df2bf44d6d3f7e4ea7733242a79c916735a7b49e | import torch
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class Model(nn.Module):
"""
Transforms incoming data using a given radial basis function:
u_{i} = rbf(||x - c_{i}|| / s_{i})
Arguments:
in_features: size of each input sample
out_features: size of each output sample
Shape:
- Input: (N, in_features) where N is an arbitrary batch size
- Output: (N, out_features) where N is an arbitrary batch size
Attributes:
centres: the learnable centres of shape (out_features, in_features).
The values are initialised from a standard normal distribution.
Normalising inputs to have mean 0 and standard deviation 1 is
recommended.
log_sigmas: logarithm of the learnable scaling factors of shape (out_features).
basis_func: the radial basis function used to transform the scaled
distances.
"""
def __init__(self, in_features, out_features, basis_func):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.centres = nn.Parameter(torch.Tensor(out_features, in_features))
self.log_sigmas = nn.Parameter(torch.Tensor(out_features))
self.basis_func = basis_func
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.centres, 0, 1)
nn.init.constant_(self.log_sigmas, 0)
def forward(self, input):
size = input.size(0), self.out_features, self.in_features
x = input.unsqueeze(1).expand(size)
c = self.centres.unsqueeze(0).expand(size)
distances = (x - c).pow(2).sum(-1).pow(0.5) / torch.exp(self.log_sigmas
).unsqueeze(0)
return distances
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
ChannelMixer | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/hp/chpdwpegv6lvistek2wqgimtufecqvfp6grp5rpblk5yjicjzqd2.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
# Source node to ATen node mapping:
# x => add, rsqrt, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%primals_1, [3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
triton_poi_fused_native_layer_norm_0 = async_compile.triton('triton_poi_fused_native_layer_norm_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + (x0), tmp8, xmask)
tl.store(out_ptr1 + (x0), tmp23, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/lh/clhh73owbiuj4adasmetdqsot2nlmw2ljupnw2q4yt3du76mikww.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
# Source node to ATen node mapping:
# x => add, add_1, mul, mul_1, rsqrt, sub, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%primals_1, [3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %getitem_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_2), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %primals_3), kwargs = {})
triton_poi_fused_native_layer_norm_1 = async_compile.triton('triton_poi_fused_native_layer_norm_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_native_layer_norm_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/w3/cw34ecblrjywdduwwge3aadqjo3z4k2b7rzmrjodmdmh6egawevs.py
# Topologically Sorted Source Nodes: [gelu], Original ATen: [aten.gelu]
# Source node to ATen node mapping:
# gelu => add_2, erf, mul_2, mul_3, mul_4
# Graph fragment:
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 0.5), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 0.7071067811865476), kwargs = {})
# %erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_3,), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_2, %add_2), kwargs = {})
triton_poi_fused_gelu_2 = async_compile.triton('triton_poi_fused_gelu_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gelu_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_gelu_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + (x0), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/5q/c5qmnkuxezgezseizmolw3mx24fyy6xp3cfoz3egpqwcprxgwjre.py
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.add]
# Source node to ATen node mapping:
# out => add_3
# Graph fragment:
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_3, %primals_1), kwargs = {})
triton_poi_fused_add_3 = async_compile.triton('triton_poi_fused_add_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_3(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x2), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (16, 4), (4, 1))
assert_size_stride(primals_5, (16, ), (1, ))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
stream0 = get_raw_stream(0)
triton_poi_fused_native_layer_norm_0.run(primals_1, buf0, buf1, 64, grid=grid(64), stream=stream0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
triton_poi_fused_native_layer_norm_1.run(primals_1, buf0, buf1, primals_2, primals_3, buf2, 256, grid=grid(256), stream=stream0)
del buf0
del buf1
del primals_2
del primals_3
buf3 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 16), (1, 4), 0), alpha=1, beta=1, out=buf3)
del primals_5
buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [gelu], Original ATen: [aten.gelu]
triton_poi_fused_gelu_2.run(buf3, buf4, 1024, grid=grid(1024), stream=stream0)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf4, (64, 16), (16, 1), 0), reinterpret_tensor(primals_6, (16, 4), (1, 16), 0), out=buf5)
buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.add]
triton_poi_fused_add_3.run(buf6, primals_7, primals_1, 256, grid=grid(256), stream=stream0)
del primals_7
return (buf6, primals_1, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), buf3, reinterpret_tensor(buf4, (64, 16), (16, 1), 0), primals_6, primals_4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class ChannelMixer(nn.Module):
def __init__(self, d_model, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(d_model, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = self.mlp(x)
out = x + residual
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'expansion_factor': 4, 'dropout': 0.5}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn.functional as F
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_gelu_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
@triton.jit
def triton_poi_fused_add_3(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (16, 4), (4, 1))
assert_size_stride(primals_5, (16,), (1,))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
get_raw_stream(0)
triton_poi_fused_native_layer_norm_0[grid(64)](primals_1, buf0,
buf1, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_native_layer_norm_1[grid(256)](primals_1, buf0,
buf1, primals_2, primals_3, buf2, 256, XBLOCK=128, num_warps=4,
num_stages=1)
del buf0
del buf1
del primals_2
del primals_3
buf3 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf2, (64, 4), (
4, 1), 0), reinterpret_tensor(primals_4, (4, 16), (1, 4), 0),
alpha=1, beta=1, out=buf3)
del primals_5
buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
triton_poi_fused_gelu_2[grid(1024)](buf3, buf4, 1024, XBLOCK=128,
num_warps=4, num_stages=1)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf4, (64, 16), (16, 1), 0),
reinterpret_tensor(primals_6, (16, 4), (1, 16), 0), out=buf5)
buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused_add_3[grid(256)](buf6, primals_7, primals_1, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_7
return buf6, primals_1, reinterpret_tensor(buf2, (64, 4), (4, 1), 0
), buf3, reinterpret_tensor(buf4, (64, 16), (16, 1), 0
), primals_6, primals_4
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class ChannelMixerNew(nn.Module):
def __init__(self, d_model, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(d_model, expansion_factor, dropout)
def forward(self, input_0):
primals_2 = self.norm.weight
primals_3 = self.norm.bias
primals_4 = self.mlp.fc1.weight
primals_5 = self.mlp.fc1.bias
primals_6 = self.mlp.fc2.weight
primals_7 = self.mlp.fc2.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
| Misuzu-Kurenai/mlp-singer | ChannelMixer | false | 855 | [
"MIT"
] | 0 | 416451045bb9b3965aaf496e84a8b45332a6ba59 | https://github.com/Misuzu-Kurenai/mlp-singer/tree/416451045bb9b3965aaf496e84a8b45332a6ba59 | import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class Model(nn.Module):
def __init__(self, d_model, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(d_model, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = self.mlp(x)
out = x + residual
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 0.5]
|
TokenMixer | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/hp/chpdwpegv6lvistek2wqgimtufecqvfp6grp5rpblk5yjicjzqd2.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
# Source node to ATen node mapping:
# x => add, rsqrt, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%primals_1, [3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
triton_poi_fused_native_layer_norm_0 = async_compile.triton('triton_poi_fused_native_layer_norm_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + (x0), tmp8, xmask)
tl.store(out_ptr1 + (x0), tmp23, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/r4/cr4j7hzjw2quuyvesstjj4anrvqw5igyzqrj3sm3fqqycoyilvv7.py
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# linear => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x5 = (xindex // 4)
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr1 + (x5), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + (x5), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/z5/cz5ftw3jjdcl7kf3rjnx3ukf3ssp6cwhcagtkza76pvoehzvmo6f.py
# Topologically Sorted Source Nodes: [linear, gelu], Original ATen: [aten.add, aten.gelu]
# Source node to ATen node mapping:
# gelu => add_3, erf, mul_2, mul_3, mul_4
# linear => add_2
# Graph fragment:
# %add_2 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_1, %primals_5), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%add_2, 0.5), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%add_2, 0.7071067811865476), kwargs = {})
# %erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_3,), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_2, %add_3), kwargs = {})
triton_poi_fused_add_gelu_2 = async_compile.triton('triton_poi_fused_add_gelu_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_gelu_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_gelu_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.5
tmp4 = tmp2 * tmp3
tmp5 = 0.7071067811865476
tmp6 = tmp2 * tmp5
tmp7 = libdevice.erf(tmp6)
tmp8 = 1.0
tmp9 = tmp7 + tmp8
tmp10 = tmp4 * tmp9
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/e5/ce5bdqzsyer25as6qwj6lscnvnsdinzp4i4e6yl4gdcwu7bl4lft.py
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.add]
# Source node to ATen node mapping:
# out => add_4
# Graph fragment:
# %add_4 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%permute_3, %primals_1), kwargs = {})
triton_poi_fused_add_3 = async_compile.triton('triton_poi_fused_add_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_3(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
tmp0 = tl.load(in_out_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tl.store(in_out_ptr0 + (x4), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (16, 4), (4, 1))
assert_size_stride(primals_5, (16, ), (1, ))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
stream0 = get_raw_stream(0)
triton_poi_fused_native_layer_norm_0.run(primals_1, buf0, buf1, 64, grid=grid(64), stream=stream0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(primals_1, buf0, buf1, primals_2, primals_3, buf2, 256, grid=grid(256), stream=stream0)
del buf0
del buf1
del primals_2
del primals_3
buf3 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 16), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear, gelu], Original ATen: [aten.add, aten.gelu]
triton_poi_fused_add_gelu_2.run(buf3, primals_5, buf4, 1024, grid=grid(1024), stream=stream0)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf4, (64, 16), (16, 1), 0), reinterpret_tensor(primals_6, (16, 4), (1, 16), 0), out=buf5)
buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 4, 16, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.add]
triton_poi_fused_add_3.run(buf6, primals_7, primals_1, 256, grid=grid(256), stream=stream0)
del primals_7
return (buf6, primals_1, primals_5, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), buf3, reinterpret_tensor(buf4, (64, 16), (16, 1), 0), primals_6, primals_4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class TokenMixer(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(seq_len, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = x.transpose(1, 2)
x = self.mlp(x)
x = x.transpose(1, 2)
out = x + residual
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'seq_len': 4, 'expansion_factor': 4,
'dropout': 0.5}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn.functional as F
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x5 = xindex // 4
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x5, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), tmp8, xmask)
@triton.jit
def triton_poi_fused_add_gelu_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.5
tmp4 = tmp2 * tmp3
tmp5 = 0.7071067811865476
tmp6 = tmp2 * tmp5
tmp7 = libdevice.erf(tmp6)
tmp8 = 1.0
tmp9 = tmp7 + tmp8
tmp10 = tmp4 * tmp9
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_3(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
tmp0 = tl.load(in_out_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tl.store(in_out_ptr0 + x4, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (16, 4), (4, 1))
assert_size_stride(primals_5, (16,), (1,))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
get_raw_stream(0)
triton_poi_fused_native_layer_norm_0[grid(64)](primals_1, buf0,
buf1, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(256)](primals_1, buf0, buf1,
primals_2, primals_3, buf2, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del buf0
del buf1
del primals_2
del primals_3
buf3 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 16), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
triton_poi_fused_add_gelu_2[grid(1024)](buf3, primals_5, buf4, 1024,
XBLOCK=256, num_warps=4, num_stages=1)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf4, (64, 16), (16, 1), 0),
reinterpret_tensor(primals_6, (16, 4), (1, 16), 0), out=buf5)
buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 4, 16, 1), 0)
del buf5
triton_poi_fused_add_3[grid(256)](buf6, primals_7, primals_1, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_7
return buf6, primals_1, primals_5, reinterpret_tensor(buf2, (64, 4), (4,
1), 0), buf3, reinterpret_tensor(buf4, (64, 16), (16, 1), 0
), primals_6, primals_4
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class TokenMixerNew(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(seq_len, expansion_factor, dropout)
def forward(self, input_0):
primals_2 = self.norm.weight
primals_3 = self.norm.bias
primals_4 = self.mlp.fc1.weight
primals_5 = self.mlp.fc1.bias
primals_6 = self.mlp.fc2.weight
primals_7 = self.mlp.fc2.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
| Misuzu-Kurenai/mlp-singer | TokenMixer | false | 856 | [
"MIT"
] | 0 | 416451045bb9b3965aaf496e84a8b45332a6ba59 | https://github.com/Misuzu-Kurenai/mlp-singer/tree/416451045bb9b3965aaf496e84a8b45332a6ba59 | import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class Model(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(seq_len, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = x.transpose(1, 2)
x = self.mlp(x)
x = x.transpose(1, 2)
out = x + residual
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'seq_len': 4, 'expansion_factor': 4,
'dropout': 0.5}]
|
MultiHeadBoxAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xe/cxeuttfzx4xq2jmzwzvkech4crjirky5wjckb34lnep5o6sog3uw.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + (4*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x4), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fn/cfnr6wn6wbusamhilcgctjberp7g5kksyakcze32k6ntswznc2de.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone_1
# Graph fragment:
# %clone_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand_1,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = (yindex // 16)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (16*x2) + (64*y1)), xmask & ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (y0), ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + (4*y3)), tmp2, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/an/cank72e7b53ox2aydemjjtk4g5r4l3ezhruq4nvlf7sdzagvxw7i.py
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
# Source node to ATen node mapping:
# att => div
# clamp => clamp_min
# log => log
# w_mn => add
# w_mn_1 => amax, exp, sub, sum_1
# wrapped_sqrt => full_default
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 2.0), kwargs = {dtype: torch.float64, layout: torch.strided, device: cpu, pin_memory: False})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%view_11, %full_default), kwargs = {})
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%primals_10, 1e-06), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_min,), kwargs = {})
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%log, %div), kwargs = {})
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%add, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
triton_poi_fused__softmax_add_clamp_div_log_sqrt_2 = async_compile.triton('triton_poi_fused__softmax_add_clamp_div_log_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_add_clamp_div_log_sqrt_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_2(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (4*x2), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr1 + (1 + (4*x2)), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr1 + (2 + (4*x2)), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr1 + (3 + (4*x2)), xmask, eviction_policy='evict_last')
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = triton_helpers.maximum(tmp8, tmp1)
tmp10 = tl_math.log(tmp9)
tmp12 = tmp11 / tmp5
tmp13 = tmp10 + tmp12
tmp14 = triton_helpers.maximum(tmp7, tmp13)
tmp16 = triton_helpers.maximum(tmp15, tmp1)
tmp17 = tl_math.log(tmp16)
tmp19 = tmp18 / tmp5
tmp20 = tmp17 + tmp19
tmp21 = triton_helpers.maximum(tmp14, tmp20)
tmp23 = triton_helpers.maximum(tmp22, tmp1)
tmp24 = tl_math.log(tmp23)
tmp26 = tmp25 / tmp5
tmp27 = tmp24 + tmp26
tmp28 = triton_helpers.maximum(tmp21, tmp27)
tmp29 = tmp7 - tmp28
tmp30 = tl_math.exp(tmp29)
tmp31 = tmp13 - tmp28
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp34 = tmp20 - tmp28
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tmp37 = tmp27 - tmp28
tmp38 = tl_math.exp(tmp37)
tmp39 = tmp36 + tmp38
tl.store(out_ptr0 + (x2), tmp28, xmask)
tl.store(out_ptr1 + (x2), tmp39, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/rp/crpmptrcembmuvzewgtdminpn6dx6emnjq6iahknzv7cxb67vchn.py
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
# Source node to ATen node mapping:
# att => div
# clamp => clamp_min
# log => log
# w_mn => add
# w_mn_1 => amax, div_1, exp, sub
# wrapped_sqrt => full_default
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 2.0), kwargs = {dtype: torch.float64, layout: torch.strided, device: cpu, pin_memory: False})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%view_11, %full_default), kwargs = {})
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%primals_10, 1e-06), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_min,), kwargs = {})
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%log, %div), kwargs = {})
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%add, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
# %div_1 : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_add_clamp_div_log_sqrt_3 = async_compile.triton('triton_poi_fused__softmax_add_clamp_div_log_sqrt_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_add_clamp_div_log_sqrt_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_3(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex % 64
x4 = xindex
x5 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x3), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_out_ptr0 + (x4), xmask)
tmp8 = tl.load(in_ptr1 + (x5), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr2 + (x5), xmask, eviction_policy='evict_last')
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = tmp7 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp12 = tmp10 / tmp11
tl.store(in_out_ptr0 + (x4), tmp12, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6b/c6busvilz5nn36jjet3bmw7cqddirh4sgalamjr3fsrp3sbsacfi.py
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# contiguous => clone_4
# Graph fragment:
# %clone_4 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_6,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_4 = async_compile.triton('triton_poi_fused_clone_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tl.store(out_ptr0 + (x4), tmp0, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6m/c6mhj5zwirfhy5e4o45uaeov72uwfby4udubpm2fcz42iqvs2g57.py
# Topologically Sorted Source Nodes: [add_1, out_3], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# add_1 => add_1
# out_3 => var_mean
# Graph fragment:
# %add_1 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_1, %view_17), kwargs = {})
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%add_1, [2]), kwargs = {correction: 0, keepdim: True})
triton_poi_fused_add_native_layer_norm_5 = async_compile.triton('triton_poi_fused_add_native_layer_norm_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4*x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + (x0), tmp16, xmask)
tl.store(out_ptr1 + (x0), tmp28, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/iz/cizh7p23zwsiqbrt6dvrlvjzpyujwvyyaolptfk5xtby6foymiaz.py
# Topologically Sorted Source Nodes: [add_1, out_3], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# add_1 => add_1
# out_3 => add_2, add_3, mul, mul_1, rsqrt, sub_1
# Graph fragment:
# %add_1 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_1, %view_17), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add_2,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add_1, %getitem_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_1, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_13), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %primals_14), kwargs = {})
triton_poi_fused_add_native_layer_norm_6 = async_compile.triton('triton_poi_fused_add_native_layer_norm_6', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_6', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp3 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + (x2), tmp13, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13, primals_14 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16, ), (1, ))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16, ), (1, ))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16, ), (1, ))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_11, (4, 16), (16, 1))
assert_size_stride(primals_12, (4, ), (1, ))
assert_size_stride(primals_13, (4, ), (1, ))
assert_size_stride(primals_14, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(buf0, primals_4, buf3, 256, grid=grid(256), stream=stream0)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf1, primals_6, buf4, 64, 4, grid=grid(64, 4), stream=stream0)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
triton_poi_fused__softmax_add_clamp_div_log_sqrt_2.run(primals_10, buf5, buf6, buf7, 64, grid=grid(64), stream=stream0)
buf8 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [wrapped_sqrt, att, clamp, log, w_mn, w_mn_1], Original ATen: [aten.sqrt, aten.div, aten.clamp, aten.log, aten.add, aten._softmax]
triton_poi_fused__softmax_add_clamp_div_log_sqrt_3.run(buf8, primals_10, buf6, buf7, 256, grid=grid(256), stream=stream0)
del primals_10
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.clone]
triton_poi_fused_clone_0.run(buf2, primals_8, buf9, 256, grid=grid(256), stream=stream0)
del primals_8
buf10 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0); del buf2 # reuse
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1), 0), out=buf10)
buf11 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
triton_poi_fused_clone_4.run(buf10, buf11, 256, grid=grid(256), stream=stream0)
del buf10
buf12 = reinterpret_tensor(buf7, (16, 4), (4, 1), 0); del buf7 # reuse
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_12, reinterpret_tensor(buf11, (16, 16), (16, 1), 0), reinterpret_tensor(primals_11, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf12)
del primals_12
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf14 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
# Topologically Sorted Source Nodes: [add_1, out_3], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_5.run(primals_1, buf12, buf13, buf14, 16, grid=grid(16), stream=stream0)
buf15 = reinterpret_tensor(buf6, (4, 4, 4), (16, 4, 1), 0); del buf6 # reuse
# Topologically Sorted Source Nodes: [add_1, out_3], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_6.run(primals_1, buf12, buf13, buf14, primals_13, primals_14, buf15, 64, grid=grid(64), stream=stream0)
del buf13
del buf14
del primals_14
return (buf15, primals_1, primals_13, reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), buf8, reinterpret_tensor(buf11, (16, 16), (16, 1), 0), buf12, primals_11, reinterpret_tensor(buf9, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_12 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_13 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_14 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13, primals_14])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| from torch.nn import Module
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class ScaledDotProductWithBoxAttention(nn.Module):
"""
Scaled dot-product attention with box
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1, comment=None):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductWithBoxAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
self.comment = comment
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
w_g = box_relation_embed_matrix
w_a = att
w_mn = torch.log(torch.clamp(w_g, min=1e-06)) + w_a
w_mn = torch.softmax(w_mn, -1)
att = self.dropout(w_mn)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class MultiHeadBoxAttention(Module):
"""
Multi-head attention layer with Dropout and Layer Normalization and box attention.
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1,
identity_map_reordering=False, can_be_stateful=False,
attention_module=None, attention_module_kwargs=None, comment=None):
super(MultiHeadBoxAttention, self).__init__()
self.identity_map_reordering = identity_map_reordering
self.attention = ScaledDotProductWithBoxAttention(d_model=d_model,
d_k=d_k, d_v=d_v, h=h, comment=comment)
self.dropout = nn.Dropout(p=dropout)
self.layer_norm = nn.LayerNorm(d_model)
self.can_be_stateful = can_be_stateful
if self.can_be_stateful:
self.register_state('running_keys', torch.zeros((0, d_model)))
self.register_state('running_values', torch.zeros((0, d_model)))
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
if self.can_be_stateful and self._is_stateful:
self.running_keys = torch.cat([self.running_keys, keys], 1)
keys = self.running_keys
self.running_values = torch.cat([self.running_values, values], 1)
values = self.running_values
if self.identity_map_reordering:
q_norm = self.layer_norm(queries)
k_norm = self.layer_norm(keys)
v_norm = self.layer_norm(values)
out = self.attention(q_norm, k_norm, v_norm,
box_relation_embed_matrix, attention_mask, attention_weights)
out = queries + self.dropout(torch.relu(out))
else:
out = self.attention(queries, keys, values,
box_relation_embed_matrix, attention_mask, attention_weights)
out = self.dropout(out)
out = self.layer_norm(queries + out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4,
4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'd_k': 4, 'd_v': 4, 'h': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch.nn import Module
import numpy as np
import torch.nn as nn
from torch.autograd import *
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = yindex // 16
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 16 * x2 + 64 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_2(in_ptr0, in_ptr1,
out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + 4 * x2, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr1 + (1 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp15 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr1 + (2 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp22 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = triton_helpers.maximum(tmp8, tmp1)
tmp10 = tl_math.log(tmp9)
tmp12 = tmp11 / tmp5
tmp13 = tmp10 + tmp12
tmp14 = triton_helpers.maximum(tmp7, tmp13)
tmp16 = triton_helpers.maximum(tmp15, tmp1)
tmp17 = tl_math.log(tmp16)
tmp19 = tmp18 / tmp5
tmp20 = tmp17 + tmp19
tmp21 = triton_helpers.maximum(tmp14, tmp20)
tmp23 = triton_helpers.maximum(tmp22, tmp1)
tmp24 = tl_math.log(tmp23)
tmp26 = tmp25 / tmp5
tmp27 = tmp24 + tmp26
tmp28 = triton_helpers.maximum(tmp21, tmp27)
tmp29 = tmp7 - tmp28
tmp30 = tl_math.exp(tmp29)
tmp31 = tmp13 - tmp28
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp34 = tmp20 - tmp28
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tmp37 = tmp27 - tmp28
tmp38 = tl_math.exp(tmp37)
tmp39 = tmp36 + tmp38
tl.store(out_ptr0 + x2, tmp28, xmask)
tl.store(out_ptr1 + x2, tmp39, xmask)
@triton.jit
def triton_poi_fused__softmax_add_clamp_div_log_sqrt_3(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex % 64
x4 = xindex
x5 = xindex // 4
tmp0 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_out_ptr0 + x4, xmask)
tmp8 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr2 + x5, xmask, eviction_policy='evict_last')
tmp1 = 1e-06
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = tl_math.log(tmp2)
tmp5 = 2.0
tmp6 = tmp4 / tmp5
tmp7 = tmp3 + tmp6
tmp9 = tmp7 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp12 = tmp10 / tmp11
tl.store(in_out_ptr0 + x4, tmp12, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tl.store(out_ptr0 + x4, tmp0, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16,), (1,))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16,), (1,))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_11, (4, 16), (16, 1))
assert_size_stride(primals_12, (4,), (1,))
assert_size_stride(primals_13, (4,), (1,))
assert_size_stride(primals_14, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(256)](buf0, primals_4, buf3, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_clone_1[grid(64, 4)](buf1, primals_6, buf4, 64, 4,
XBLOCK=4, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_add_clamp_div_log_sqrt_2[grid(64)](primals_10
, buf5, buf6, buf7, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf8 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused__softmax_add_clamp_div_log_sqrt_3[grid(256)](buf8,
primals_10, buf6, buf7, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_10
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_0[grid(256)](buf2, primals_8, buf9, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_8
buf10 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1), 0), out=buf10)
buf11 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_4[grid(256)](buf10, buf11, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf10
buf12 = reinterpret_tensor(buf7, (16, 4), (4, 1), 0)
del buf7
extern_kernels.addmm(primals_12, reinterpret_tensor(buf11, (16, 16),
(16, 1), 0), reinterpret_tensor(primals_11, (16, 4), (1, 16), 0
), alpha=1, beta=1, out=buf12)
del primals_12
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf14 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](primals_1, buf12,
buf13, buf14, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf15 = reinterpret_tensor(buf6, (4, 4, 4), (16, 4, 1), 0)
del buf6
triton_poi_fused_add_native_layer_norm_6[grid(64)](primals_1, buf12,
buf13, buf14, primals_13, primals_14, buf15, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf13
del buf14
del primals_14
return buf15, primals_1, primals_13, reinterpret_tensor(primals_2, (16,
4), (4, 1), 0), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf8, reinterpret_tensor(buf11, (16, 16), (16, 1), 0
), buf12, primals_11, reinterpret_tensor(buf9, (16, 4, 4), (16, 1,
4), 0), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0)
class ScaledDotProductWithBoxAttention(nn.Module):
"""
Scaled dot-product attention with box
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1, comment=None):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductWithBoxAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
self.comment = comment
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
w_g = box_relation_embed_matrix
w_a = att
w_mn = torch.log(torch.clamp(w_g, min=1e-06)) + w_a
w_mn = torch.softmax(w_mn, -1)
att = self.dropout(w_mn)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class MultiHeadBoxAttentionNew(Module):
"""
Multi-head attention layer with Dropout and Layer Normalization and box attention.
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1,
identity_map_reordering=False, can_be_stateful=False,
attention_module=None, attention_module_kwargs=None, comment=None):
super(MultiHeadBoxAttentionNew, self).__init__()
self.identity_map_reordering = identity_map_reordering
self.attention = ScaledDotProductWithBoxAttention(d_model=d_model,
d_k=d_k, d_v=d_v, h=h, comment=comment)
self.dropout = nn.Dropout(p=dropout)
self.layer_norm = nn.LayerNorm(d_model)
self.can_be_stateful = can_be_stateful
if self.can_be_stateful:
self.register_state('running_keys', torch.zeros((0, d_model)))
self.register_state('running_values', torch.zeros((0, d_model)))
def forward(self, input_0, input_1, input_2, input_3):
primals_3 = self.attention.fc_q.weight
primals_4 = self.attention.fc_q.bias
primals_5 = self.attention.fc_k.weight
primals_6 = self.attention.fc_k.bias
primals_7 = self.attention.fc_v.weight
primals_8 = self.attention.fc_v.bias
primals_11 = self.attention.fc_o.weight
primals_12 = self.attention.fc_o.bias
primals_13 = self.layer_norm.weight
primals_14 = self.layer_norm.bias
primals_1 = input_0
primals_2 = input_1
primals_9 = input_2
primals_10 = input_3
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14])
return output[0]
| Maxi-0902/DRAN | MultiHeadBoxAttention | false | 857 | [
"MIT"
] | 0 | c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | https://github.com/Maxi-0902/DRAN/tree/c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | from torch.nn import Module
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class ScaledDotProductWithBoxAttention(nn.Module):
"""
Scaled dot-product attention with box
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1, comment=None):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super().__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
self.comment = comment
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
w_g = box_relation_embed_matrix
w_a = att
w_mn = torch.log(torch.clamp(w_g, min=1e-06)) + w_a
w_mn = torch.softmax(w_mn, -1)
att = self.dropout(w_mn)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class Model(Module):
"""
Multi-head attention layer with Dropout and Layer Normalization and box attention.
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1,
identity_map_reordering=False, can_be_stateful=False,
attention_module=None, attention_module_kwargs=None, comment=None):
super().__init__()
self.identity_map_reordering = identity_map_reordering
self.attention = ScaledDotProductWithBoxAttention(d_model=d_model,
d_k=d_k, d_v=d_v, h=h, comment=comment)
self.dropout = nn.Dropout(p=dropout)
self.layer_norm = nn.LayerNorm(d_model)
self.can_be_stateful = can_be_stateful
if self.can_be_stateful:
self.register_state('running_keys', torch.zeros((0, d_model)))
self.register_state('running_values', torch.zeros((0, d_model)))
def forward(self, queries, keys, values, box_relation_embed_matrix,
attention_mask=None, attention_weights=None):
if self.can_be_stateful and self._is_stateful:
self.running_keys = torch.cat([self.running_keys, keys], 1)
keys = self.running_keys
self.running_values = torch.cat([self.running_val
# ... truncated (>4000 chars) for memory efficiency |
FixedSubnetConv | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/p3/cp3qleddjiuuytozrtebx5pzf2ycpwtw4mkq2jsx7qqswymv2bm6.py
# Topologically Sorted Source Nodes: [w], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# w => mul
# Graph fragment:
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_1, %primals_2), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask)
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/tc/ctcagp37ljugm52zu6ckorigrppqo67voefe2f2odg5r6hyllhyu.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# x => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_4, %mul, %primals_3, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_1 = async_compile.triton('triton_poi_fused_convolution_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [w], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(primals_1, primals_2, buf0, 256, grid=grid(256), stream=stream0)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(primals_4, buf0, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 1, 1), (4, 1, 1, 1))
buf2 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
triton_poi_fused_convolution_1.run(buf2, primals_3, 16, grid=grid(16), stream=stream0)
del primals_3
return (buf2, primals_1, primals_2, primals_4, buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
from torch import nn
from torch.nn import functional as F
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class FixedSubnetConv(nn.Conv2d):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.scores = nn.Parameter(torch.Tensor(self.weight.size()))
nn.init.kaiming_uniform_(self.scores, a=math.sqrt(5))
def set_prune_rate(self, prune_rate):
self.prune_rate = prune_rate
None
def set_subnet(self):
output = self.clamped_scores().clone()
_, idx = self.clamped_scores().flatten().abs().sort()
p = int(self.prune_rate * self.clamped_scores().numel())
flat_oup = output.flatten()
flat_oup[idx[:p]] = 0
flat_oup[idx[p:]] = 1
self.scores = torch.nn.Parameter(output)
self.scores.requires_grad = False
def clamped_scores(self):
return self.scores.abs()
def get_subnet(self):
return self.weight * self.scores
def forward(self, x):
w = self.get_subnet()
x = F.conv2d(x, w, self.bias, self.stride, self.padding, self.
dilation, self.groups)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'kernel_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import math
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask)
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](primals_1, primals_2, buf0, 256,
XBLOCK=128, num_warps=4, num_stages=1)
buf1 = extern_kernels.convolution(primals_4, buf0, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 1, 1), (4, 1, 1, 1))
buf2 = buf1
del buf1
triton_poi_fused_convolution_1[grid(16)](buf2, primals_3, 16,
XBLOCK=16, num_warps=1, num_stages=1)
del primals_3
return buf2, primals_1, primals_2, primals_4, buf0
class FixedSubnetConvNew(nn.Conv2d):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.scores = nn.Parameter(torch.Tensor(self.weight.size()))
nn.init.kaiming_uniform_(self.scores, a=math.sqrt(5))
def set_prune_rate(self, prune_rate):
self.prune_rate = prune_rate
None
def set_subnet(self):
output = self.clamped_scores().clone()
_, idx = self.clamped_scores().flatten().abs().sort()
p = int(self.prune_rate * self.clamped_scores().numel())
flat_oup = output.flatten()
flat_oup[idx[:p]] = 0
flat_oup[idx[p:]] = 1
self.scores = torch.nn.Parameter(output)
self.scores.requires_grad = False
def clamped_scores(self):
return self.scores.abs()
def get_subnet(self):
return self.weight * self.scores
def forward(self, input_0):
primals_1 = self.weight
primals_3 = self.bias
primals_2 = self.scores
primals_4 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
| MorganeAyle/SNIP-it | FixedSubnetConv | false | 858 | [
"MIT"
] | 0 | df2bf44d6d3f7e4ea7733242a79c916735a7b49e | https://github.com/MorganeAyle/SNIP-it/tree/df2bf44d6d3f7e4ea7733242a79c916735a7b49e | import math
import torch
from torch import nn
from torch.nn import functional as F
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class Model(nn.Conv2d):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.scores = nn.Parameter(torch.Tensor(self.weight.size()))
nn.init.kaiming_uniform_(self.scores, a=math.sqrt(5))
def set_prune_rate(self, prune_rate):
self.prune_rate = prune_rate
None
def set_subnet(self):
output = self.clamped_scores().clone()
_, idx = self.clamped_scores().flatten().abs().sort()
p = int(self.prune_rate * self.clamped_scores().numel())
flat_oup = output.flatten()
flat_oup[idx[:p]] = 0
flat_oup[idx[p:]] = 1
self.scores = torch.nn.Parameter(output)
self.scores.requires_grad = False
def clamped_scores(self):
return self.scores.abs()
def get_subnet(self):
return self.weight * self.scores
def forward(self, x):
w = self.get_subnet()
x = F.conv2d(x, w, self.bias, self.stride, self.padding, self.
dilation, self.groups)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
SplitChannels | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/u5/cu56dhpcth43gy4shrd7mcexf4nfa6qetnnhwe4mno4v6ug76h6j.py
# Topologically Sorted Source Nodes: [a_1], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# a_1 => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%arg0_1,), kwargs = {})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tl.store(out_ptr0 + (x0), tmp0, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [a_1], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
buf1 = empty_strided_cuda((4, 0, 4, 4), (0, 16, 4, 1), torch.float32)
return (buf0, buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
class SplitChannels(torch.nn.Module):
def __init__(self, split_location):
super(SplitChannels, self).__init__()
self.split_location = split_location
def forward(self, x):
a, b = x[:, :self.split_location], x[:, self.split_location:]
a, b = a.clone(), b.clone()
del x
return a, b
def inverse(self, x, y):
return torch.cat([x, y], dim=1)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'split_location': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tl.store(out_ptr0 + x0, tmp0, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
buf1 = empty_strided_cuda((4, 0, 4, 4), (0, 16, 4, 1), torch.float32)
return buf0, buf1
class SplitChannelsNew(torch.nn.Module):
def __init__(self, split_location):
super(SplitChannelsNew, self).__init__()
self.split_location = split_location
def inverse(self, x, y):
return torch.cat([x, y], dim=1)
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0], output[1]
| Myyyr/segmentation | SplitChannels | false | 859 | [
"MIT"
] | 0 | 6b9423e327cff1eb23599404031b7fb8e9ecf75d | https://github.com/Myyyr/segmentation/tree/6b9423e327cff1eb23599404031b7fb8e9ecf75d | import torch
class Model(torch.nn.Module):
def __init__(self, split_location):
super().__init__()
self.split_location = split_location
def forward(self, x):
a, b = x[:, :self.split_location], x[:, self.split_location:]
a, b = a.clone(), b.clone()
del x
return a, b
def inverse(self, x, y):
return torch.cat([x, y], dim=1)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
PoseMap | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xz/cxza2g5csa2qi5maf6i3bhrrrbzuuxyn2xm5sovg7ycfagm5zps5.py
# Topologically Sorted Source Nodes: [res, sum_2, div, res_1], Original ATen: [aten.sum, aten.repeat, aten.div]
# Source node to ATen node mapping:
# div => repeat
# res => sum_1
# res_1 => div
# sum_2 => sum_2
# Graph fragment:
# %sum_1 : [num_users=2] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%arg0_1, [1], True), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%sum_1, [2, 3], True), kwargs = {})
# %repeat : [num_users=1] = call_function[target=torch.ops.aten.repeat.default](args = (%sum_2, [1, 1, 4, 4]), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_1, %repeat), kwargs = {})
triton_per_fused_div_repeat_sum_0 = async_compile.triton('triton_per_fused_div_repeat_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[4, 16],
reduction_hint=ReductionHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_div_repeat_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_div_repeat_sum_0(in_ptr0, out_ptr1, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 4
rnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + (64*x0)), xmask, other=0.0)
tmp1 = tl.load(in_ptr0 + (16 + r1 + (64*x0)), xmask, other=0.0)
tmp3 = tl.load(in_ptr0 + (32 + r1 + (64*x0)), xmask, other=0.0)
tmp5 = tl.load(in_ptr0 + (48 + r1 + (64*x0)), xmask, other=0.0)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp9 = tl.where(xmask, tmp7, 0)
tmp10 = tl.sum(tmp9, 1)[:, None]
tmp11 = tmp6 / tmp10
tl.store(out_ptr1 + (r1 + (16*x0)), tmp11, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf1 = empty_strided_cuda((4, 1, 4, 4), (16, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [res, sum_2, div, res_1], Original ATen: [aten.sum, aten.repeat, aten.div]
stream0 = get_raw_stream(0)
triton_per_fused_div_repeat_sum_0.run(arg0_1, buf1, 4, 16, grid=grid(4), stream=stream0)
del arg0_1
return (buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class PoseMap(nn.Module):
def __init__(self):
super(PoseMap, self).__init__()
pass
def forward(self, x):
assert len(x.shape) == 4, 'The HeatMap shape should be BxCxHxW'
res = x.sum(dim=1, keepdim=True)
H = x.shape[2]
W = x.shape[3]
div = res.sum(dim=[2, 3], keepdim=True).repeat(1, 1, H, W)
res = res / div
return res
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_div_repeat_sum_0(in_ptr0, out_ptr1, xnumel, rnumel,
XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0)
tmp1 = tl.load(in_ptr0 + (16 + r1 + 64 * x0), xmask, other=0.0)
tmp3 = tl.load(in_ptr0 + (32 + r1 + 64 * x0), xmask, other=0.0)
tmp5 = tl.load(in_ptr0 + (48 + r1 + 64 * x0), xmask, other=0.0)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp9 = tl.where(xmask, tmp7, 0)
tmp10 = tl.sum(tmp9, 1)[:, None]
tmp11 = tmp6 / tmp10
tl.store(out_ptr1 + (r1 + 16 * x0), tmp11, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf1 = empty_strided_cuda((4, 1, 4, 4), (16, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_per_fused_div_repeat_sum_0[grid(4)](arg0_1, buf1, 4, 16,
XBLOCK=1, num_warps=2, num_stages=1)
del arg0_1
return buf1,
class PoseMapNew(nn.Module):
def __init__(self):
super(PoseMapNew, self).__init__()
pass
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| MrChenFeng/Project-Template | PoseMap | false | 860 | [
"MIT"
] | 0 | 42a335c6abb710bbae6407cbb0ca461533bc12f9 | https://github.com/MrChenFeng/Project-Template/tree/42a335c6abb710bbae6407cbb0ca461533bc12f9 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self):
super().__init__()
pass
def forward(self, x):
assert len(x.shape) == 4, 'The HeatMap shape should be BxCxHxW'
res = x.sum(dim=1, keepdim=True)
H = x.shape[2]
W = x.shape[3]
div = res.sum(dim=[2, 3], keepdim=True).repeat(1, 1, H, W)
res = res / div
return res
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
ScalarAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nc/cncwsucylpsg2zmlivjfxu6vbd64ztxjndlsix2ysjtby3xohgk4.py
# Topologically Sorted Source Nodes: [tanh], Original ATen: [aten.tanh]
# Source node to ATen node mapping:
# tanh => tanh
# Graph fragment:
# %tanh : [num_users=2] = call_function[target=torch.ops.aten.tanh.default](args = (%view_1,), kwargs = {})
triton_poi_fused_tanh_0 = async_compile.triton('triton_poi_fused_tanh_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_tanh_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_tanh_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + (x2), tmp3, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/dm/cdmkcxuzpnailvibeivaikqdr4zvashgzwju7qijhq5aizlo3aor.py
# Topologically Sorted Source Nodes: [softmax], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# softmax => amax, exp, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%view_3, [1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_3, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
triton_poi_fused__softmax_1 = async_compile.triton('triton_poi_fused__softmax_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = (xindex // 16)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x3), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/kt/cktghousutx6xui2sl2rvevzmb7gkacvfhntjq5n2xzeu7v57oz6.py
# Topologically Sorted Source Nodes: [softmax], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# softmax => div, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [1], True), kwargs = {})
# %div : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_2 = async_compile.triton('triton_poi_fused__softmax_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = (xindex // 16)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + (16*x2)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x3), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fj/cfjxsjft32uejuiclvkwroildivviknko2qfji2c5dxkqu3gx26a.py
# Topologically Sorted Source Nodes: [mul, x], Original ATen: [aten.mul, aten.sum]
# Source node to ATen node mapping:
# mul => mul
# x => sum_2
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%expand, %primals_3), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%mul, [1]), kwargs = {})
triton_poi_fused_mul_sum_3 = async_compile.triton('triton_poi_fused_mul_sum_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_sum_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_sum_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4) % 4
x2 = (xindex // 16)
x3 = xindex % 16
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x1 + (16*x2)), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (x3 + (64*x2)), xmask)
tmp3 = tl.load(in_ptr0 + (4 + x1 + (16*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (16 + x3 + (64*x2)), xmask)
tmp7 = tl.load(in_ptr0 + (8 + x1 + (16*x2)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (32 + x3 + (64*x2)), xmask)
tmp11 = tl.load(in_ptr0 + (12 + x1 + (16*x2)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (48 + x3 + (64*x2)), xmask)
tmp2 = tmp0 * tmp1
tmp5 = tmp3 * tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 * tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 * tmp12
tmp14 = tmp10 + tmp13
tl.store(out_ptr0 + (x4), tmp14, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (1, 4), (4, 1))
assert_size_stride(primals_5, (1, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [tanh], Original ATen: [aten.tanh]
stream0 = get_raw_stream(0)
triton_poi_fused_tanh_0.run(buf1, primals_2, 256, grid=grid(256), stream=stream0)
del primals_2
buf3 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 1), (1, 4), 0), alpha=1, beta=1, out=buf3)
del primals_5
buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [softmax], Original ATen: [aten._softmax]
triton_poi_fused__softmax_1.run(buf3, buf4, 64, grid=grid(64), stream=stream0)
buf5 = reinterpret_tensor(buf3, (4, 4, 4, 1), (16, 4, 1, 1), 0); del buf3 # reuse
# Topologically Sorted Source Nodes: [softmax], Original ATen: [aten._softmax]
triton_poi_fused__softmax_2.run(buf4, buf5, 64, grid=grid(64), stream=stream0)
buf6 = reinterpret_tensor(buf4, (4, 4, 4), (16, 4, 1), 0); del buf4 # reuse
# Topologically Sorted Source Nodes: [mul, x], Original ATen: [aten.mul, aten.sum]
triton_poi_fused_mul_sum_3.run(buf5, primals_3, buf6, 64, grid=grid(64), stream=stream0)
return (buf6, buf5, primals_3, buf1, buf5, primals_4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((1, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.utils.data
import torch.utils.checkpoint
class ScalarAttention(nn.Module):
def __init__(self, in_size, hidden_size):
super(ScalarAttention, self).__init__()
self.hidden = nn.Linear(in_size, hidden_size)
nn.init.orthogonal_(self.hidden.weight.data)
self.out = nn.Linear(hidden_size, 1)
nn.init.orthogonal_(self.hidden.weight.data)
self.softmax = nn.Softmax(dim=1)
def forward(self, input):
self.alpha = self.softmax(self.out(torch.tanh(self.hidden(input))))
x = (self.alpha.expand_as(input) * input).sum(dim=1)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_size': 4, 'hidden_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
import torch.utils.data
import torch.utils.checkpoint
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_tanh_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = xindex // 16
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = xindex // 16
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x3, tmp8, xmask)
@triton.jit
def triton_poi_fused_mul_sum_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 4
x2 = xindex // 16
x3 = xindex % 16
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x1 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr1 + (x3 + 64 * x2), xmask)
tmp3 = tl.load(in_ptr0 + (4 + x1 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr1 + (16 + x3 + 64 * x2), xmask)
tmp7 = tl.load(in_ptr0 + (8 + x1 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp8 = tl.load(in_ptr1 + (32 + x3 + 64 * x2), xmask)
tmp11 = tl.load(in_ptr0 + (12 + x1 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp12 = tl.load(in_ptr1 + (48 + x3 + 64 * x2), xmask)
tmp2 = tmp0 * tmp1
tmp5 = tmp3 * tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 * tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 * tmp12
tmp14 = tmp10 + tmp13
tl.store(out_ptr0 + x4, tmp14, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (1, 4), (4, 1))
assert_size_stride(primals_5, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_tanh_0[grid(256)](buf1, primals_2, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_2
buf3 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 4), (
4, 1), 0), reinterpret_tensor(primals_4, (4, 1), (1, 4), 0),
alpha=1, beta=1, out=buf3)
del primals_5
buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_1[grid(64)](buf3, buf4, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf5 = reinterpret_tensor(buf3, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf3
triton_poi_fused__softmax_2[grid(64)](buf4, buf5, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf6 = reinterpret_tensor(buf4, (4, 4, 4), (16, 4, 1), 0)
del buf4
triton_poi_fused_mul_sum_3[grid(64)](buf5, primals_3, buf6, 64,
XBLOCK=64, num_warps=1, num_stages=1)
return buf6, buf5, primals_3, buf1, buf5, primals_4
class ScalarAttentionNew(nn.Module):
def __init__(self, in_size, hidden_size):
super(ScalarAttentionNew, self).__init__()
self.hidden = nn.Linear(in_size, hidden_size)
nn.init.orthogonal_(self.hidden.weight.data)
self.out = nn.Linear(hidden_size, 1)
nn.init.orthogonal_(self.hidden.weight.data)
self.softmax = nn.Softmax(dim=1)
def forward(self, input_0):
primals_1 = self.hidden.weight
primals_2 = self.hidden.bias
primals_4 = self.out.weight
primals_5 = self.out.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| MarvinLvn/platalea | ScalarAttention | false | 861 | [
"Apache-2.0"
] | 0 | 31def0813c90a3259f86f7d86cb576cd66dca3fe | https://github.com/MarvinLvn/platalea/tree/31def0813c90a3259f86f7d86cb576cd66dca3fe | import torch
import torch.nn as nn
import torch.utils.data
import torch.utils.checkpoint
class Model(nn.Module):
def __init__(self, in_size, hidden_size):
super().__init__()
self.hidden = nn.Linear(in_size, hidden_size)
nn.init.orthogonal_(self.hidden.weight.data)
self.out = nn.Linear(hidden_size, 1)
nn.init.orthogonal_(self.hidden.weight.data)
self.softmax = nn.Softmax(dim=1)
def forward(self, input):
self.alpha = self.softmax(self.out(torch.tanh(self.hidden(input))))
x = (self.alpha.expand_as(input) * input).sum(dim=1)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
MultiHeadAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xe/cxeuttfzx4xq2jmzwzvkech4crjirky5wjckb34lnep5o6sog3uw.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + (4*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x4), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fn/cfnr6wn6wbusamhilcgctjberp7g5kksyakcze32k6ntswznc2de.py
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# matmul => clone_1
# Graph fragment:
# %clone_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%expand_1,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = (yindex // 16)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (16*x2) + (64*y1)), xmask & ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (y0), ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + (4*y3)), tmp2, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ka/ckaneo6wn23ipwgbubou64jdtwieswlrn7w7r7kqky4aagh3v6l3.py
# Topologically Sorted Source Nodes: [wrapped_sqrt, att_1], Original ATen: [aten.sqrt, aten._softmax]
# Source node to ATen node mapping:
# att_1 => exp
# wrapped_sqrt => full_default
# Graph fragment:
# %full_default : [num_users=2] = call_function[target=torch.ops.aten.full.default](args = ([], 2.0), kwargs = {dtype: torch.float64, layout: torch.strided, device: cpu, pin_memory: False})
# %scalar_tensor_default : [num_users=2] = call_function[target=torch.ops.aten.scalar_tensor.default](args = (1,), kwargs = {dtype: torch.float32, device: cuda:0, pin_memory: False})
# %ge_scalar : [num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%full_default, 0), kwargs = {})
# %neg_default : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%scalar_tensor_default,), kwargs = {})
# %where_self : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ge_scalar, %scalar_tensor_default, %neg_default), kwargs = {})
# %mul_tensor : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_11, %where_self), kwargs = {})
# %amax_default : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%mul_tensor, [-1], True), kwargs = {})
# %sub_tensor : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_tensor, %amax_default), kwargs = {})
# %mul_tensor_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where_self, %full_default), kwargs = {})
# %div_tensor : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sub_tensor, %mul_tensor_1), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%div_tensor,), kwargs = {})
triton_poi_fused__softmax_sqrt_2 = async_compile.triton('triton_poi_fused__softmax_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_sqrt_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_sqrt_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp8 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp1 = tl.full([1], 2.0, tl.float64)
tmp2 = tl.full([1], 0.0, tl.float64)
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp9 = tmp8 * tmp6
tmp11 = tmp10 * tmp6
tmp12 = triton_helpers.maximum(tmp9, tmp11)
tmp14 = tmp13 * tmp6
tmp15 = triton_helpers.maximum(tmp12, tmp14)
tmp17 = tmp16 * tmp6
tmp18 = triton_helpers.maximum(tmp15, tmp17)
tmp19 = tmp7 - tmp18
tmp20 = tmp6.to(tl.float64)
tmp21 = tmp20 * tmp1
tmp22 = tmp21.to(tl.float32)
tmp23 = tmp19 / tmp22
tmp24 = tl_math.exp(tmp23)
tl.store(out_ptr0 + (x2), tmp24, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ry/cryn7ntc2gpkbfzbre3xh7lffx7zkbskw6oihbzsekkgajmdbki6.py
# Topologically Sorted Source Nodes: [att_1], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# att_1 => div_1, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div_1 : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_3 = async_compile.triton('triton_poi_fused__softmax_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6b/c6busvilz5nn36jjet3bmw7cqddirh4sgalamjr3fsrp3sbsacfi.py
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# contiguous => clone_3
# Graph fragment:
# %clone_3 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_6,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_4 = async_compile.triton('triton_poi_fused_clone_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tl.store(out_ptr0 + (x4), tmp0, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6m/c6mhj5zwirfhy5e4o45uaeov72uwfby4udubpm2fcz42iqvs2g57.py
# Topologically Sorted Source Nodes: [add, out_3], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# add => add
# out_3 => var_mean
# Graph fragment:
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_1, %view_17), kwargs = {})
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%add, [2]), kwargs = {correction: 0, keepdim: True})
triton_poi_fused_add_native_layer_norm_5 = async_compile.triton('triton_poi_fused_add_native_layer_norm_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4*x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + (x0), tmp16, xmask)
tl.store(out_ptr1 + (x0), tmp28, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/iz/cizh7p23zwsiqbrt6dvrlvjzpyujwvyyaolptfk5xtby6foymiaz.py
# Topologically Sorted Source Nodes: [add, out_3], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# add => add
# out_3 => add_1, add_2, mul, mul_1, rsqrt, sub_1
# Graph fragment:
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_1, %view_17), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add_1,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add, %getitem_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_1, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_12), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %primals_13), kwargs = {})
triton_poi_fused_add_native_layer_norm_6 = async_compile.triton('triton_poi_fused_add_native_layer_norm_6', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_6', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp3 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + (x2), tmp13, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16, ), (1, ))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16, ), (1, ))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16, ), (1, ))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 16), (16, 1))
assert_size_stride(primals_11, (4, ), (1, ))
assert_size_stride(primals_12, (4, ), (1, ))
assert_size_stride(primals_13, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(buf0, primals_4, buf3, 256, grid=grid(256), stream=stream0)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf1, primals_6, buf4, 64, 4, grid=grid(64, 4), stream=stream0)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [wrapped_sqrt, att_1], Original ATen: [aten.sqrt, aten._softmax]
triton_poi_fused__softmax_sqrt_2.run(buf5, buf6, 256, grid=grid(256), stream=stream0)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [att_1], Original ATen: [aten._softmax]
triton_poi_fused__softmax_3.run(buf6, buf7, 256, grid=grid(256), stream=stream0)
buf8 = buf6; del buf6 # reuse
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.clone]
triton_poi_fused_clone_0.run(buf2, primals_8, buf8, 256, grid=grid(256), stream=stream0)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0); del buf2 # reuse
# Topologically Sorted Source Nodes: [matmul_1], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [contiguous], Original ATen: [aten.clone]
triton_poi_fused_clone_4.run(buf9, buf10, 256, grid=grid(256), stream=stream0)
del buf9
buf11 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_11, reinterpret_tensor(buf10, (16, 16), (16, 1), 0), reinterpret_tensor(primals_10, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf11)
del primals_11
buf12 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
# Topologically Sorted Source Nodes: [add, out_3], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_5.run(primals_1, buf11, buf12, buf13, 16, grid=grid(16), stream=stream0)
buf14 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [add, out_3], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_6.run(primals_1, buf11, buf12, buf13, primals_12, primals_13, buf14, 64, grid=grid(64), stream=stream0)
del buf12
del buf13
del primals_13
return (buf14, primals_1, primals_12, reinterpret_tensor(primals_2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0), buf7, reinterpret_tensor(buf10, (16, 16), (16, 1), 0), buf11, primals_10, reinterpret_tensor(buf8, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_12 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_13 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| from torch.nn import Module
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class ScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nv, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, n_k, d_v)
:return: out (b_s, nq, d_model)
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class MultiHeadAttention(Module):
"""
Multi-head attention layer with Dropout and Layer Normalization.
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1,
identity_map_reordering=False, can_be_stateful=False,
attention_module=None, attention_module_kwargs=None):
super(MultiHeadAttention, self).__init__()
self.identity_map_reordering = identity_map_reordering
if attention_module is not None:
if attention_module_kwargs is not None:
self.attention = attention_module(d_model=d_model, d_k=d_k,
d_v=d_v, h=h, **attention_module_kwargs)
else:
self.attention = attention_module(d_model=d_model, d_k=d_k,
d_v=d_v, h=h)
else:
self.attention = ScaledDotProductAttention(d_model=d_model, d_k
=d_k, d_v=d_v, h=h)
self.dropout = nn.Dropout(p=dropout)
self.layer_norm = nn.LayerNorm(d_model)
self.can_be_stateful = can_be_stateful
if self.can_be_stateful:
self.register_state('running_keys', torch.zeros((0, d_model)))
self.register_state('running_values', torch.zeros((0, d_model)))
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
if self.can_be_stateful and self._is_stateful:
self.running_keys = torch.cat([self.running_keys, keys], 1)
keys = self.running_keys
self.running_values = torch.cat([self.running_values, values], 1)
values = self.running_values
if self.identity_map_reordering:
q_norm = self.layer_norm(queries)
k_norm = self.layer_norm(keys)
v_norm = self.layer_norm(values)
out = self.attention(q_norm, k_norm, v_norm, attention_mask,
attention_weights)
out = queries + self.dropout(torch.relu(out))
else:
out = self.attention(queries, keys, values, attention_mask,
attention_weights)
out = self.dropout(out)
out = self.layer_norm(queries + out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4])
]
def get_init_inputs():
return [[], {'d_model': 4, 'd_k': 4, 'd_v': 4, 'h': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch.nn import Module
import numpy as np
import torch.nn as nn
from torch.autograd import *
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = yindex // 16
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 16 * x2 + 64 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_sqrt_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp8 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp13 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = tl.full([1], 2.0, tl.float64)
tmp2 = tl.full([1], 0.0, tl.float64)
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp9 = tmp8 * tmp6
tmp11 = tmp10 * tmp6
tmp12 = triton_helpers.maximum(tmp9, tmp11)
tmp14 = tmp13 * tmp6
tmp15 = triton_helpers.maximum(tmp12, tmp14)
tmp17 = tmp16 * tmp6
tmp18 = triton_helpers.maximum(tmp15, tmp17)
tmp19 = tmp7 - tmp18
tmp20 = tmp6.to(tl.float64)
tmp21 = tmp20 * tmp1
tmp22 = tmp21.to(tl.float32)
tmp23 = tmp19 / tmp22
tmp24 = tl_math.exp(tmp23)
tl.store(out_ptr0 + x2, tmp24, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tl.store(out_ptr0 + x4, tmp0, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16,), (1,))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16,), (1,))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 16), (16, 1))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
assert_size_stride(primals_13, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(256)](buf0, primals_4, buf3, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_clone_1[grid(64, 4)](buf1, primals_6, buf4, 64, 4,
XBLOCK=4, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_sqrt_2[grid(256)](buf5, buf6, 256, XBLOCK
=128, num_warps=4, num_stages=1)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused__softmax_3[grid(256)](buf6, buf7, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf8 = buf6
del buf6
triton_poi_fused_clone_0[grid(256)](buf2, primals_8, buf8, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_4[grid(256)](buf9, buf10, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf9
buf11 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_11, reinterpret_tensor(buf10, (16, 16),
(16, 1), 0), reinterpret_tensor(primals_10, (16, 4), (1, 16), 0
), alpha=1, beta=1, out=buf11)
del primals_11
buf12 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](primals_1, buf11,
buf12, buf13, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf14 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(64)](primals_1, buf11,
buf12, buf13, primals_12, primals_13, buf14, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf12
del buf13
del primals_13
return buf14, primals_1, primals_12, reinterpret_tensor(primals_2, (16,
4), (4, 1), 0), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf7, reinterpret_tensor(buf10, (16, 16), (16, 1), 0
), buf11, primals_10, reinterpret_tensor(buf8, (16, 4, 4), (16, 1,
4), 0), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0)
class ScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nv, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, n_k, d_v)
:return: out (b_s, nq, d_model)
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class MultiHeadAttentionNew(Module):
"""
Multi-head attention layer with Dropout and Layer Normalization.
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1,
identity_map_reordering=False, can_be_stateful=False,
attention_module=None, attention_module_kwargs=None):
super(MultiHeadAttentionNew, self).__init__()
self.identity_map_reordering = identity_map_reordering
if attention_module is not None:
if attention_module_kwargs is not None:
self.attention = attention_module(d_model=d_model, d_k=d_k,
d_v=d_v, h=h, **attention_module_kwargs)
else:
self.attention = attention_module(d_model=d_model, d_k=d_k,
d_v=d_v, h=h)
else:
self.attention = ScaledDotProductAttention(d_model=d_model, d_k
=d_k, d_v=d_v, h=h)
self.dropout = nn.Dropout(p=dropout)
self.layer_norm = nn.LayerNorm(d_model)
self.can_be_stateful = can_be_stateful
if self.can_be_stateful:
self.register_state('running_keys', torch.zeros((0, d_model)))
self.register_state('running_values', torch.zeros((0, d_model)))
def forward(self, input_0, input_1, input_2):
primals_3 = self.attention.fc_q.weight
primals_4 = self.attention.fc_q.bias
primals_5 = self.attention.fc_k.weight
primals_6 = self.attention.fc_k.bias
primals_7 = self.attention.fc_v.weight
primals_8 = self.attention.fc_v.bias
primals_10 = self.attention.fc_o.weight
primals_11 = self.attention.fc_o.bias
primals_12 = self.layer_norm.weight
primals_13 = self.layer_norm.bias
primals_1 = input_0
primals_2 = input_1
primals_9 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13])
return output[0]
| Maxi-0902/DRAN | MultiHeadAttention | false | 862 | [
"MIT"
] | 0 | c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | https://github.com/Maxi-0902/DRAN/tree/c3dbfcbc018446544150dc4e151442d6a9fcd4d9 | from torch.nn import Module
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import *
class ScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super().__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
nn.init.xavier_uniform_(self.fc_q.weight)
nn.init.xavier_uniform_(self.fc_k.weight)
nn.init.xavier_uniform_(self.fc_v.weight)
nn.init.xavier_uniform_(self.fc_o.weight)
nn.init.constant_(self.fc_q.bias, 0)
nn.init.constant_(self.fc_k.bias, 0)
nn.init.constant_(self.fc_v.bias, 0)
nn.init.constant_(self.fc_o.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nv, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, n_k, d_v)
:return: out (b_s, nq, d_model)
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class Model(Module):
"""
Multi-head attention layer with Dropout and Layer Normalization.
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1,
identity_map_reordering=False, can_be_stateful=False,
attention_module=None, attention_module_kwargs=None):
super().__init__()
self.identity_map_reordering = identity_map_reordering
if attention_module is not None:
if attention_module_kwargs is not None:
self.attention = attention_module(d_model=d_model, d_k=d_k,
d_v=d_v, h=h, **attention_module_kwargs)
else:
self.attention = attention_module(d_model=d_model, d_k=d_k,
d_v=d_v, h=h)
else:
self.attention = ScaledDotProductAttention(d_model=d_model, d_k
=d_k, d_v=d_v, h=h)
self.dropout = nn.Dropout(p=dropout)
self.layer_norm = nn.LayerNorm(d_model)
self.can_be_stateful = can_be_stateful
if self.can_be_stateful:
self.register_state('running_keys', torch.zeros((0, d_model)))
self.register_state('running_values', torch.zeros((0, d_model)))
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
if self.can_be_stateful and self._is_stateful:
self.running_keys = torch.cat([self.running_keys, keys], 1)
keys = self.running_keys
self.running_values = to
# ... truncated (>4000 chars) for memory efficiency |
MNIST_classifier | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/qp/cqpf3ql535lobs6mc6l5mep65wu2cj446ohcw5e75cpwggiavfgv.py
# Topologically Sorted Source Nodes: [conv2d, x], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# conv2d => convolution
# x => relu
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [2, 2], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution,), kwargs = {})
triton_poi_fused_convolution_relu_0 = async_compile.triton('triton_poi_fused_convolution_relu_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[131072],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 115200
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 900) % 32
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x3), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/xk/cxkvyfvspf73ksgokcqqswzdbo3mm6zcwn3oysmga7t3dttjw46i.py
# Topologically Sorted Source Nodes: [conv2d_1, x_1], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# conv2d_1 => convolution_1
# x_1 => relu_1
# Graph fragment:
# %convolution_1 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%relu, %primals_4, %primals_5, [2, 2], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu_1 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution_1,), kwargs = {})
triton_poi_fused_convolution_relu_1 = async_compile.triton('triton_poi_fused_convolution_relu_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[65536],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 50176
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 196) % 64
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x3), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/x2/cx2gv3rubc3d6i6vje7ofz7rc6geevvsbewnta34aqec3wib5n4a.py
# Topologically Sorted Source Nodes: [conv2d_3, x_3], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# conv2d_3 => convolution_3
# x_3 => relu_3
# Graph fragment:
# %convolution_3 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%relu_2, %primals_8, %primals_9, [2, 2], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu_3 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution_3,), kwargs = {})
triton_poi_fused_convolution_relu_2 = async_compile.triton('triton_poi_fused_convolution_relu_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32768],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 18432
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = (xindex // 36) % 128
tmp0 = tl.load(in_out_ptr0 + (x3), None)
tmp1 = tl.load(in_ptr0 + (x1), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x3), tmp4, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/t6/ct66yluuxdsnlrfjtwku3llhylssn2uww2aqi2puk6bplzrkfiof.py
# Topologically Sorted Source Nodes: [conv2d_4], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# conv2d_4 => convolution_4
# Graph fragment:
# %convolution_4 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%relu_3, %primals_10, %primals_11, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_3 = async_compile.triton('triton_poi_fused_convolution_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1000
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 25) % 10
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11 = args
args.clear()
assert_size_stride(primals_1, (32, 1, 5, 5), (25, 25, 5, 1))
assert_size_stride(primals_2, (32, ), (1, ))
assert_size_stride(primals_3, (4, 1, 64, 64), (4096, 4096, 64, 1))
assert_size_stride(primals_4, (64, 32, 3, 3), (288, 9, 3, 1))
assert_size_stride(primals_5, (64, ), (1, ))
assert_size_stride(primals_6, (64, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_7, (64, ), (1, ))
assert_size_stride(primals_8, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_9, (128, ), (1, ))
assert_size_stride(primals_10, (10, 128, 2, 2), (512, 4, 2, 1))
assert_size_stride(primals_11, (10, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(2, 2), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 32, 30, 30), (28800, 900, 30, 1))
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [conv2d, x], Original ATen: [aten.convolution, aten.relu]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_relu_0.run(buf1, primals_2, 115200, grid=grid(115200), stream=stream0)
del primals_2
# Topologically Sorted Source Nodes: [conv2d_1], Original ATen: [aten.convolution]
buf2 = extern_kernels.convolution(buf1, primals_4, stride=(2, 2), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 64, 14, 14), (12544, 196, 14, 1))
buf3 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [conv2d_1, x_1], Original ATen: [aten.convolution, aten.relu]
triton_poi_fused_convolution_relu_1.run(buf3, primals_5, 50176, grid=grid(50176), stream=stream0)
del primals_5
# Topologically Sorted Source Nodes: [conv2d_2], Original ATen: [aten.convolution]
buf4 = extern_kernels.convolution(buf3, primals_6, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (4, 64, 14, 14), (12544, 196, 14, 1))
buf5 = buf4; del buf4 # reuse
# Topologically Sorted Source Nodes: [conv2d_2, x_2], Original ATen: [aten.convolution, aten.relu]
triton_poi_fused_convolution_relu_1.run(buf5, primals_7, 50176, grid=grid(50176), stream=stream0)
del primals_7
# Topologically Sorted Source Nodes: [conv2d_3], Original ATen: [aten.convolution]
buf6 = extern_kernels.convolution(buf5, primals_8, stride=(2, 2), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 128, 6, 6), (4608, 36, 6, 1))
buf7 = buf6; del buf6 # reuse
# Topologically Sorted Source Nodes: [conv2d_3, x_3], Original ATen: [aten.convolution, aten.relu]
triton_poi_fused_convolution_relu_2.run(buf7, primals_9, 18432, grid=grid(18432), stream=stream0)
del primals_9
# Topologically Sorted Source Nodes: [conv2d_4], Original ATen: [aten.convolution]
buf8 = extern_kernels.convolution(buf7, primals_10, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 10, 5, 5), (250, 25, 5, 1))
buf9 = buf8; del buf8 # reuse
# Topologically Sorted Source Nodes: [conv2d_4], Original ATen: [aten.convolution]
triton_poi_fused_convolution_3.run(buf9, primals_11, 1000, grid=grid(1000), stream=stream0)
del primals_11
return (reinterpret_tensor(buf9, (100, 10), (10, 1), 0), primals_1, primals_3, primals_4, primals_6, primals_8, primals_10, buf1, buf3, buf5, buf7, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((32, 1, 5, 5), (25, 25, 5, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((32, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 1, 64, 64), (4096, 4096, 64, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((64, 32, 3, 3), (288, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((64, 64, 3, 3), (576, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((128, 64, 3, 3), (576, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((128, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((10, 128, 2, 2), (512, 4, 2, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((10, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class MNIST_classifier(nn.Module):
def __init__(self):
super(MNIST_classifier, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 5, stride=2)
self.conv2 = nn.Conv2d(32, 64, 3, stride=2)
self.conv3 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
self.conv4 = nn.Conv2d(64, 128, 3, stride=2)
self.conv5 = nn.Conv2d(128, 10, 2)
self.act = nn.ReLU(inplace=True)
def forward(self, x):
x = self.act(self.conv1(x))
x = self.act(self.conv2(x))
x = self.act(self.conv3(x))
x = self.act(self.conv4(x))
x = self.conv5(x).view(-1, 10)
return x
def get_inputs():
return [torch.rand([4, 1, 64, 64])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 115200
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 900 % 32
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_1(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 50176
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 196 % 64
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 36 % 128
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 1000
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 25 % 10
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11) = args
args.clear()
assert_size_stride(primals_1, (32, 1, 5, 5), (25, 25, 5, 1))
assert_size_stride(primals_2, (32,), (1,))
assert_size_stride(primals_3, (4, 1, 64, 64), (4096, 4096, 64, 1))
assert_size_stride(primals_4, (64, 32, 3, 3), (288, 9, 3, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (64, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_7, (64,), (1,))
assert_size_stride(primals_8, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_9, (128,), (1,))
assert_size_stride(primals_10, (10, 128, 2, 2), (512, 4, 2, 1))
assert_size_stride(primals_11, (10,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(2,
2), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 32, 30, 30), (28800, 900, 30, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_relu_0[grid(115200)](buf1, primals_2,
115200, XBLOCK=512, num_warps=8, num_stages=1)
del primals_2
buf2 = extern_kernels.convolution(buf1, primals_4, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 64, 14, 14), (12544, 196, 14, 1))
buf3 = buf2
del buf2
triton_poi_fused_convolution_relu_1[grid(50176)](buf3, primals_5,
50176, XBLOCK=512, num_warps=4, num_stages=1)
del primals_5
buf4 = extern_kernels.convolution(buf3, primals_6, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (4, 64, 14, 14), (12544, 196, 14, 1))
buf5 = buf4
del buf4
triton_poi_fused_convolution_relu_1[grid(50176)](buf5, primals_7,
50176, XBLOCK=512, num_warps=4, num_stages=1)
del primals_7
buf6 = extern_kernels.convolution(buf5, primals_8, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 128, 6, 6), (4608, 36, 6, 1))
buf7 = buf6
del buf6
triton_poi_fused_convolution_relu_2[grid(18432)](buf7, primals_9,
18432, XBLOCK=256, num_warps=4, num_stages=1)
del primals_9
buf8 = extern_kernels.convolution(buf7, primals_10, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 10, 5, 5), (250, 25, 5, 1))
buf9 = buf8
del buf8
triton_poi_fused_convolution_3[grid(1000)](buf9, primals_11, 1000,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_11
return (reinterpret_tensor(buf9, (100, 10), (10, 1), 0), primals_1,
primals_3, primals_4, primals_6, primals_8, primals_10, buf1, buf3,
buf5, buf7)
class MNIST_classifierNew(nn.Module):
def __init__(self):
super(MNIST_classifierNew, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 5, stride=2)
self.conv2 = nn.Conv2d(32, 64, 3, stride=2)
self.conv3 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
self.conv4 = nn.Conv2d(64, 128, 3, stride=2)
self.conv5 = nn.Conv2d(128, 10, 2)
self.act = nn.ReLU(inplace=True)
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_6 = self.conv3.weight
primals_7 = self.conv3.bias
primals_8 = self.conv4.weight
primals_9 = self.conv4.bias
primals_10 = self.conv5.weight
primals_11 = self.conv5.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11])
return output[0]
| MorganeAyle/SNIP-it | MNIST_classifier | false | 863 | [
"MIT"
] | 0 | df2bf44d6d3f7e4ea7733242a79c916735a7b49e | https://github.com/MorganeAyle/SNIP-it/tree/df2bf44d6d3f7e4ea7733242a79c916735a7b49e | import torch
from torch import nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class Model(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(1, 32, 5, stride=2)
self.conv2 = nn.Conv2d(32, 64, 3, stride=2)
self.conv3 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
self.conv4 = nn.Conv2d(64, 128, 3, stride=2)
self.conv5 = nn.Conv2d(128, 10, 2)
self.act = nn.ReLU(inplace=True)
def forward(self, x):
x = self.act(self.conv1(x))
x = self.act(self.conv2(x))
x = self.act(self.conv3(x))
x = self.act(self.conv4(x))
x = self.conv5(x).view(-1, 10)
return x
def get_inputs():
return [torch.rand([4, 1, 64, 64])]
def get_init_inputs():
return []
|
QuickGELU | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/vb/cvbzgl2kevabwrh6e5ay2dasnq6gi3a7zeveknxks7jkbuq64ime.py
# Topologically Sorted Source Nodes: [mul, sigmoid, mul_1], Original ATen: [aten.mul, aten.sigmoid]
# Source node to ATen node mapping:
# mul => mul
# mul_1 => mul_1
# sigmoid => sigmoid
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg0_1, 1.702), kwargs = {})
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%mul,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg0_1, %sigmoid), kwargs = {})
triton_poi_fused_mul_sigmoid_0 = async_compile.triton('triton_poi_fused_mul_sigmoid_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_sigmoid_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_sigmoid_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 1.702
tmp2 = tmp0 * tmp1
tmp3 = tl.sigmoid(tmp2)
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + (x0), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul, sigmoid, mul_1], Original ATen: [aten.mul, aten.sigmoid]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_sigmoid_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
import torch.distributed.nn
class QuickGELU(nn.Module):
def forward(self, x: 'torch.Tensor'):
return x * torch.sigmoid(1.702 * x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch import nn
import torch.distributed.nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_sigmoid_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 1.702
tmp2 = tmp0 * tmp1
tmp3 = tl.sigmoid(tmp2)
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + x0, tmp4, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_sigmoid_0[grid(256)](arg0_1, buf0, 256, XBLOCK
=256, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class QuickGELUNew(nn.Module):
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| NYU-DICE-Lab/open_clip | QuickGELU | false | 864 | [
"MIT"
] | 0 | fd71804b503135fb1c7cc8de3a0d6599741c8ed9 | https://github.com/NYU-DICE-Lab/open_clip/tree/fd71804b503135fb1c7cc8de3a0d6599741c8ed9 | import torch
from torch import nn
import torch.distributed.nn
class Model(nn.Module):
def forward(self, x: 'torch.Tensor'):
return x * torch.sigmoid(1.702 * x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
QREmbeddingBag | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/iy/ciyon33lu2o3f2lebmdecxtlu42syxcwsaphharmspenqyfixtlk.py
# Topologically Sorted Source Nodes: [embed_q], Original ATen: [aten.arange]
# Source node to ATen node mapping:
# embed_q => iota
# Graph fragment:
# %iota : [num_users=3] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 4, dtype: torch.int64, device: cuda:0, requires_grad: False})
triton_poi_fused_arange_0 = async_compile.triton('triton_poi_fused_arange_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*i64', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0,), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_arange_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_arange_0(out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = 4*x0
tl.store(out_ptr0 + (x0), tmp0, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/zf/czfbkirglecm7wywddzz66rq764tpg3xvmbcbwoorawgjupe2rmx.py
# Topologically Sorted Source Nodes: [truediv, input_q, remainder, input_r], Original ATen: [aten.div, aten._to_copy, aten.remainder]
# Source node to ATen node mapping:
# input_q => convert_element_type
# input_r => convert_element_type_1
# remainder => remainder
# truediv => div
# Graph fragment:
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, 4), kwargs = {})
# %convert_element_type : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%div, torch.int64), kwargs = {})
# %remainder : [num_users=1] = call_function[target=torch.ops.aten.remainder.Scalar](args = (%primals_1, 4), kwargs = {})
# %convert_element_type_1 : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%remainder, torch.int64), kwargs = {})
triton_poi_fused__to_copy_div_remainder_1 = async_compile.triton('triton_poi_fused__to_copy_div_remainder_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*i64', 2: '*i64', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_div_remainder_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__to_copy_div_remainder_1(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.25
tmp2 = tmp0 * tmp1
tmp3 = tmp2.to(tl.int64)
tmp4 = 4.0
tmp5 = tmp0 % tmp4
tmp6 = tl.full([1], 0, tl.int32)
tmp7 = tmp5 != tmp6
tmp8 = libdevice.signbit(tmp5) if (tmp5).dtype is tl.float32 else tmp5 < 0
tmp9 = libdevice.signbit(tmp4) if (tmp4).dtype is tl.float32 else tmp4 < 0
tmp10 = tmp8 != tmp9
tmp11 = tmp7 & tmp10
tmp12 = tmp5 + tmp4
tmp13 = tl.where(tmp11, tmp12, tmp5)
tmp14 = tmp13.to(tl.int64)
tl.store(out_ptr0 + (x0), tmp3, xmask)
tl.store(out_ptr1 + (x0), tmp14, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/s6/cs6amwv6tcyb72n2inb7juqh7pzigbagrrevtktdjfchpnz3tffn.py
# Topologically Sorted Source Nodes: [embed], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# embed => mul
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%getitem, %getitem_4), kwargs = {})
triton_poi_fused_mul_2 = async_compile.triton('triton_poi_fused_mul_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask)
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (1, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, ), (1, ), torch.int64)
# Topologically Sorted Source Nodes: [embed_q], Original ATen: [aten.arange]
stream0 = get_raw_stream(0)
triton_poi_fused_arange_0.run(buf0, 4, grid=grid(4), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.int64)
buf7 = empty_strided_cuda((4, 4), (4, 1), torch.int64)
# Topologically Sorted Source Nodes: [truediv, input_q, remainder, input_r], Original ATen: [aten.div, aten._to_copy, aten.remainder]
triton_poi_fused__to_copy_div_remainder_1.run(primals_1, buf1, buf7, 16, grid=grid(16), stream=stream0)
del primals_1
# Topologically Sorted Source Nodes: [embed_q], Original ATen: [aten._embedding_bag]
buf2 = torch.ops.aten._embedding_bag.default(primals_2, reinterpret_tensor(buf1, (16, ), (1, ), 0), buf0, False, 1)
del primals_2
buf3 = buf2[0]
buf4 = buf2[1]
buf5 = buf2[2]
buf6 = buf2[3]
del buf2
# Topologically Sorted Source Nodes: [embed_r], Original ATen: [aten._embedding_bag]
buf8 = torch.ops.aten._embedding_bag.default(primals_3, reinterpret_tensor(buf7, (16, ), (1, ), 0), buf0, False, 1)
del primals_3
buf9 = buf8[0]
buf10 = buf8[1]
buf11 = buf8[2]
buf12 = buf8[3]
del buf8
buf13 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [embed], Original ATen: [aten.mul]
triton_poi_fused_mul_2.run(buf3, buf9, buf13, 16, grid=grid(16), stream=stream0)
return (buf13, buf0, reinterpret_tensor(buf1, (16, ), (1, ), 0), buf3, buf4, buf5, buf6, reinterpret_tensor(buf7, (16, ), (1, ), 0), buf9, buf10, buf11, buf12, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((1, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import numpy as np
import torch.nn as nn
from torch.nn.parameter import Parameter
import torch.nn.functional as F
class QREmbeddingBag(nn.Module):
"""Computes sums or means over two 'bags' of embeddings, one using the quotient
of the indices and the other using the remainder of the indices, without
instantiating the intermediate embeddings, then performs an operation to combine these.
For bags of constant length and no :attr:`per_sample_weights`, this class
* with ``mode="sum"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.sum(dim=0)``,
* with ``mode="mean"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.mean(dim=0)``,
* with ``mode="max"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.max(dim=0)``.
However, :class:`~torch.nn.EmbeddingBag` is much more time and memory efficient than using a chain of these
operations.
QREmbeddingBag also supports per-sample weights as an argument to the forward
pass. This scales the output of the Embedding before performing a weighted
reduction as specified by ``mode``. If :attr:`per_sample_weights`` is passed, the
only supported ``mode`` is ``"sum"``, which computes a weighted sum according to
:attr:`per_sample_weights`.
Known Issues:
Autograd breaks with multiple GPUs. It breaks only with multiple embeddings.
Args:
num_categories (int): total number of unique categories. The input indices must be in
0, 1, ..., num_categories - 1.
embedding_dim (list): list of sizes for each embedding vector in each table. If ``"add"``
or ``"mult"`` operation are used, these embedding dimensions must be
the same. If a single embedding_dim is used, then it will use this
embedding_dim for both embedding tables.
num_collisions (int): number of collisions to enforce.
operation (string, optional): ``"concat"``, ``"add"``, or ``"mult". Specifies the operation
to compose embeddings. ``"concat"`` concatenates the embeddings,
``"add"`` sums the embeddings, and ``"mult"`` multiplies
(component-wise) the embeddings.
Default: ``"mult"``
max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
is renormalized to have norm :attr:`max_norm`.
norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
scale_grad_by_freq (boolean, optional): if given, this will scale gradients by the inverse of frequency of
the words in the mini-batch. Default ``False``.
Note: this option is not supported when ``mode="max"``.
mode (string, optional): ``"sum"``, ``"mean"`` or ``"max"``. Specifies the way to reduce the bag.
``"sum"`` computes the weighted sum, taking :attr:`per_sample_weights`
into consideration. ``"mean"`` computes the average of the values
in the bag, ``"max"`` computes the max value over each bag.
Default: ``"mean"``
sparse (bool, optional): if ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor. See
Notes for more details regarding sparse gradients. Note: this option is not
supported when ``mode="max"``.
Attributes:
weight (Tensor): the learnable weights of each embedding table is the module of shape
`(num_embeddings, embedding_dim)` initialized using a uniform distribution
with sqrt(1 / num_categories).
Inputs: :attr:`input` (LongTensor), :attr:`offsets` (LongTensor, optional), and
:attr:`per_index_weights` (Tensor, optional)
- If :attr:`input` is 2D of shape `(B, N)`,
it will be treated as ``B`` bags (sequences) each of fixed length ``N``, and
this will return ``B`` values aggregated in a way depending on the :attr:`mode`.
:attr:`offsets` is ignored and required to be ``None`` in this case.
- If :attr:`input` is 1D of shape `(N)`,
it will be treated as a concatenation of multiple bags (sequences).
:attr:`offsets` is required to be a 1D tensor containing the
starting index positions of each bag in :attr:`input`. Therefore,
for :attr:`offsets` of shape `(B)`, :attr:`input` will be viewed as
having ``B`` bags. Empty bags (i.e., having 0-length) will have
returned vectors filled by zeros.
per_sample_weights (Tensor, optional): a tensor of float / double weights, or None
to indicate all weights should be taken to be ``1``. If specified, :attr:`per_sample_weights`
must have exactly the same shape as input and is treated as having the same
:attr:`offsets`, if those are not ``None``. Only supported for ``mode='sum'``.
Output shape: `(B, embedding_dim)`
"""
__constants__ = ['num_categories', 'embedding_dim', 'num_collisions',
'operation', 'max_norm', 'norm_type', 'scale_grad_by_freq', 'mode',
'sparse']
def __init__(self, num_categories, embedding_dim, num_collisions,
operation='mult', max_norm=None, norm_type=2.0, scale_grad_by_freq=
False, mode='mean', sparse=False, _weight=None):
super(QREmbeddingBag, self).__init__()
assert operation in ['concat', 'mult', 'add'], 'Not valid operation!'
self.num_categories = num_categories
if isinstance(embedding_dim, int) or len(embedding_dim) == 1:
self.embedding_dim = [embedding_dim, embedding_dim]
else:
self.embedding_dim = embedding_dim
self.num_collisions = num_collisions
self.operation = operation
self.max_norm = max_norm
self.norm_type = norm_type
self.scale_grad_by_freq = scale_grad_by_freq
if self.operation == 'add' or self.operation == 'mult':
assert self.embedding_dim[0] == self.embedding_dim[1
], 'Embedding dimensions do not match!'
self.num_embeddings = [int(np.ceil(num_categories / num_collisions)
), num_collisions]
if _weight is None:
self.weight_q = Parameter(torch.Tensor(self.num_embeddings[0],
self.embedding_dim[0]))
self.weight_r = Parameter(torch.Tensor(self.num_embeddings[1],
self.embedding_dim[1]))
self.reset_parameters()
else:
assert list(_weight[0].shape) == [self.num_embeddings[0], self.
embedding_dim[0]
], 'Shape of weight for quotient table does not match num_embeddings and embedding_dim'
assert list(_weight[1].shape) == [self.num_embeddings[1], self.
embedding_dim[1]
], 'Shape of weight for remainder table does not match num_embeddings and embedding_dim'
self.weight_q = Parameter(_weight[0])
self.weight_r = Parameter(_weight[1])
self.mode = mode
self.sparse = sparse
def reset_parameters(self):
nn.init.uniform_(self.weight_q, np.sqrt(1 / self.num_categories))
nn.init.uniform_(self.weight_r, np.sqrt(1 / self.num_categories))
def forward(self, input, offsets=None, per_sample_weights=None):
input_q = (input / self.num_collisions).long()
input_r = torch.remainder(input, self.num_collisions).long()
embed_q = F.embedding_bag(input_q, self.weight_q, offsets, self.
max_norm, self.norm_type, self.scale_grad_by_freq, self.mode,
self.sparse, per_sample_weights)
embed_r = F.embedding_bag(input_r, self.weight_r, offsets, self.
max_norm, self.norm_type, self.scale_grad_by_freq, self.mode,
self.sparse, per_sample_weights)
if self.operation == 'concat':
embed = torch.cat((embed_q, embed_r), dim=1)
elif self.operation == 'add':
embed = embed_q + embed_r
elif self.operation == 'mult':
embed = embed_q * embed_r
return embed
def extra_repr(self):
s = '{num_embeddings}, {embedding_dim}'
if self.max_norm is not None:
s += ', max_norm={max_norm}'
if self.norm_type != 2:
s += ', norm_type={norm_type}'
if self.scale_grad_by_freq is not False:
s += ', scale_grad_by_freq={scale_grad_by_freq}'
s += ', mode={mode}'
return s.format(**self.__dict__)
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [[], {'num_categories': 4, 'embedding_dim': 4, 'num_collisions': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import numpy as np
import torch.nn as nn
from torch.nn.parameter import Parameter
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_arange_0(out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = 4 * x0
tl.store(out_ptr0 + x0, tmp0, xmask)
@triton.jit
def triton_poi_fused__to_copy_div_remainder_1(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.25
tmp2 = tmp0 * tmp1
tmp3 = tmp2.to(tl.int64)
tmp4 = 4.0
tmp5 = tmp0 % tmp4
tmp6 = tl.full([1], 0, tl.int32)
tmp7 = tmp5 != tmp6
tmp8 = libdevice.signbit(tmp5) if tmp5.dtype is tl.float32 else tmp5 < 0
tmp9 = libdevice.signbit(tmp4) if tmp4.dtype is tl.float32 else tmp4 < 0
tmp10 = tmp8 != tmp9
tmp11 = tmp7 & tmp10
tmp12 = tmp5 + tmp4
tmp13 = tl.where(tmp11, tmp12, tmp5)
tmp14 = tmp13.to(tl.int64)
tl.store(out_ptr0 + x0, tmp3, xmask)
tl.store(out_ptr1 + x0, tmp14, xmask)
@triton.jit
def triton_poi_fused_mul_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask)
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (1, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4,), (1,), torch.int64)
get_raw_stream(0)
triton_poi_fused_arange_0[grid(4)](buf0, 4, XBLOCK=4, num_warps=1,
num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.int64)
buf7 = empty_strided_cuda((4, 4), (4, 1), torch.int64)
triton_poi_fused__to_copy_div_remainder_1[grid(16)](primals_1, buf1,
buf7, 16, XBLOCK=16, num_warps=1, num_stages=1)
del primals_1
buf2 = torch.ops.aten._embedding_bag.default(primals_2,
reinterpret_tensor(buf1, (16,), (1,), 0), buf0, False, 1)
del primals_2
buf3 = buf2[0]
buf4 = buf2[1]
buf5 = buf2[2]
buf6 = buf2[3]
del buf2
buf8 = torch.ops.aten._embedding_bag.default(primals_3,
reinterpret_tensor(buf7, (16,), (1,), 0), buf0, False, 1)
del primals_3
buf9 = buf8[0]
buf10 = buf8[1]
buf11 = buf8[2]
buf12 = buf8[3]
del buf8
buf13 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_mul_2[grid(16)](buf3, buf9, buf13, 16, XBLOCK=16,
num_warps=1, num_stages=1)
return buf13, buf0, reinterpret_tensor(buf1, (16,), (1,), 0
), buf3, buf4, buf5, buf6, reinterpret_tensor(buf7, (16,), (1,), 0
), buf9, buf10, buf11, buf12
class QREmbeddingBagNew(nn.Module):
"""Computes sums or means over two 'bags' of embeddings, one using the quotient
of the indices and the other using the remainder of the indices, without
instantiating the intermediate embeddings, then performs an operation to combine these.
For bags of constant length and no :attr:`per_sample_weights`, this class
* with ``mode="sum"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.sum(dim=0)``,
* with ``mode="mean"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.mean(dim=0)``,
* with ``mode="max"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.max(dim=0)``.
However, :class:`~torch.nn.EmbeddingBag` is much more time and memory efficient than using a chain of these
operations.
QREmbeddingBag also supports per-sample weights as an argument to the forward
pass. This scales the output of the Embedding before performing a weighted
reduction as specified by ``mode``. If :attr:`per_sample_weights`` is passed, the
only supported ``mode`` is ``"sum"``, which computes a weighted sum according to
:attr:`per_sample_weights`.
Known Issues:
Autograd breaks with multiple GPUs. It breaks only with multiple embeddings.
Args:
num_categories (int): total number of unique categories. The input indices must be in
0, 1, ..., num_categories - 1.
embedding_dim (list): list of sizes for each embedding vector in each table. If ``"add"``
or ``"mult"`` operation are used, these embedding dimensions must be
the same. If a single embedding_dim is used, then it will use this
embedding_dim for both embedding tables.
num_collisions (int): number of collisions to enforce.
operation (string, optional): ``"concat"``, ``"add"``, or ``"mult". Specifies the operation
to compose embeddings. ``"concat"`` concatenates the embeddings,
``"add"`` sums the embeddings, and ``"mult"`` multiplies
(component-wise) the embeddings.
Default: ``"mult"``
max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
is renormalized to have norm :attr:`max_norm`.
norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
scale_grad_by_freq (boolean, optional): if given, this will scale gradients by the inverse of frequency of
the words in the mini-batch. Default ``False``.
Note: this option is not supported when ``mode="max"``.
mode (string, optional): ``"sum"``, ``"mean"`` or ``"max"``. Specifies the way to reduce the bag.
``"sum"`` computes the weighted sum, taking :attr:`per_sample_weights`
into consideration. ``"mean"`` computes the average of the values
in the bag, ``"max"`` computes the max value over each bag.
Default: ``"mean"``
sparse (bool, optional): if ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor. See
Notes for more details regarding sparse gradients. Note: this option is not
supported when ``mode="max"``.
Attributes:
weight (Tensor): the learnable weights of each embedding table is the module of shape
`(num_embeddings, embedding_dim)` initialized using a uniform distribution
with sqrt(1 / num_categories).
Inputs: :attr:`input` (LongTensor), :attr:`offsets` (LongTensor, optional), and
:attr:`per_index_weights` (Tensor, optional)
- If :attr:`input` is 2D of shape `(B, N)`,
it will be treated as ``B`` bags (sequences) each of fixed length ``N``, and
this will return ``B`` values aggregated in a way depending on the :attr:`mode`.
:attr:`offsets` is ignored and required to be ``None`` in this case.
- If :attr:`input` is 1D of shape `(N)`,
it will be treated as a concatenation of multiple bags (sequences).
:attr:`offsets` is required to be a 1D tensor containing the
starting index positions of each bag in :attr:`input`. Therefore,
for :attr:`offsets` of shape `(B)`, :attr:`input` will be viewed as
having ``B`` bags. Empty bags (i.e., having 0-length) will have
returned vectors filled by zeros.
per_sample_weights (Tensor, optional): a tensor of float / double weights, or None
to indicate all weights should be taken to be ``1``. If specified, :attr:`per_sample_weights`
must have exactly the same shape as input and is treated as having the same
:attr:`offsets`, if those are not ``None``. Only supported for ``mode='sum'``.
Output shape: `(B, embedding_dim)`
"""
__constants__ = ['num_categories', 'embedding_dim', 'num_collisions',
'operation', 'max_norm', 'norm_type', 'scale_grad_by_freq', 'mode',
'sparse']
def __init__(self, num_categories, embedding_dim, num_collisions,
operation='mult', max_norm=None, norm_type=2.0, scale_grad_by_freq=
False, mode='mean', sparse=False, _weight=None):
super(QREmbeddingBagNew, self).__init__()
assert operation in ['concat', 'mult', 'add'], 'Not valid operation!'
self.num_categories = num_categories
if isinstance(embedding_dim, int) or len(embedding_dim) == 1:
self.embedding_dim = [embedding_dim, embedding_dim]
else:
self.embedding_dim = embedding_dim
self.num_collisions = num_collisions
self.operation = operation
self.max_norm = max_norm
self.norm_type = norm_type
self.scale_grad_by_freq = scale_grad_by_freq
if self.operation == 'add' or self.operation == 'mult':
assert self.embedding_dim[0] == self.embedding_dim[1
], 'Embedding dimensions do not match!'
self.num_embeddings = [int(np.ceil(num_categories / num_collisions)
), num_collisions]
if _weight is None:
self.weight_q = Parameter(torch.Tensor(self.num_embeddings[0],
self.embedding_dim[0]))
self.weight_r = Parameter(torch.Tensor(self.num_embeddings[1],
self.embedding_dim[1]))
self.reset_parameters()
else:
assert list(_weight[0].shape) == [self.num_embeddings[0], self.
embedding_dim[0]
], 'Shape of weight for quotient table does not match num_embeddings and embedding_dim'
assert list(_weight[1].shape) == [self.num_embeddings[1], self.
embedding_dim[1]
], 'Shape of weight for remainder table does not match num_embeddings and embedding_dim'
self.weight_q = Parameter(_weight[0])
self.weight_r = Parameter(_weight[1])
self.mode = mode
self.sparse = sparse
def reset_parameters(self):
nn.init.uniform_(self.weight_q, np.sqrt(1 / self.num_categories))
nn.init.uniform_(self.weight_r, np.sqrt(1 / self.num_categories))
def extra_repr(self):
s = '{num_embeddings}, {embedding_dim}'
if self.max_norm is not None:
s += ', max_norm={max_norm}'
if self.norm_type != 2:
s += ', norm_type={norm_type}'
if self.scale_grad_by_freq is not False:
s += ', scale_grad_by_freq={scale_grad_by_freq}'
s += ', mode={mode}'
return s.format(**self.__dict__)
def forward(self, input_0):
primals_2 = self.weight_q
primals_1 = self.weight_r
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| MrDoghead/dlrm | QREmbeddingBag | false | 865 | [
"MIT"
] | 0 | 9b0d8ea992daa515104c7967f30110684283ebb1 | https://github.com/MrDoghead/dlrm/tree/9b0d8ea992daa515104c7967f30110684283ebb1 | import torch
import numpy as np
import torch.nn as nn
from torch.nn.parameter import Parameter
import torch.nn.functional as F
class Model(nn.Module):
"""Computes sums or means over two 'bags' of embeddings, one using the quotient
of the indices and the other using the remainder of the indices, without
instantiating the intermediate embeddings, then performs an operation to combine these.
For bags of constant length and no :attr:`per_sample_weights`, this class
* with ``mode="sum"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.sum(dim=0)``,
* with ``mode="mean"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.mean(dim=0)``,
* with ``mode="max"`` is equivalent to :class:`~torch.nn.Embedding` followed by ``torch.max(dim=0)``.
However, :class:`~torch.nn.EmbeddingBag` is much more time and memory efficient than using a chain of these
operations.
QREmbeddingBag also supports per-sample weights as an argument to the forward
pass. This scales the output of the Embedding before performing a weighted
reduction as specified by ``mode``. If :attr:`per_sample_weights`` is passed, the
only supported ``mode`` is ``"sum"``, which computes a weighted sum according to
:attr:`per_sample_weights`.
Known Issues:
Autograd breaks with multiple GPUs. It breaks only with multiple embeddings.
Args:
num_categories (int): total number of unique categories. The input indices must be in
0, 1, ..., num_categories - 1.
embedding_dim (list): list of sizes for each embedding vector in each table. If ``"add"``
or ``"mult"`` operation are used, these embedding dimensions must be
the same. If a single embedding_dim is used, then it will use this
embedding_dim for both embedding tables.
num_collisions (int): number of collisions to enforce.
operation (string, optional): ``"concat"``, ``"add"``, or ``"mult". Specifies the operation
to compose embeddings. ``"concat"`` concatenates the embeddings,
``"add"`` sums the embeddings, and ``"mult"`` multiplies
(component-wise) the embeddings.
Default: ``"mult"``
max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
is renormalized to have norm :attr:`max_norm`.
norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
scale_grad_by_freq (boolean, optional): if given, this will scale gradients by the inverse of frequency of
the words in the mini-batch. Default ``False``.
Note: this option is not supported when ``mode="max"``.
mode (string, optional): ``"sum"``, ``"mean"`` or ``"max"``. Specifies the way to reduce the bag.
``"sum"`` computes the weighted sum, taking :attr:`per_sample_weights`
into consideration. ``"mean"`` computes the average of the values
in the bag, ``"max"`` computes the max value over each bag.
Default: ``"mean"``
sparse (bool, optional): if ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor. See
Notes for more details regarding sparse gradients. Note: this option is not
supported when ``mode="max"``.
Attributes:
weight (Tensor): the learnable weights of each embedding table is the module of shape
`(num_embeddings, embedding_dim)` initialized using a uniform distrib
# ... truncated (>4000 chars) for memory efficiency |
MixerBlock | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/hp/chpdwpegv6lvistek2wqgimtufecqvfp6grp5rpblk5yjicjzqd2.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
# Source node to ATen node mapping:
# x => add, rsqrt, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%primals_1, [3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
triton_poi_fused_native_layer_norm_0 = async_compile.triton('triton_poi_fused_native_layer_norm_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + (x0), tmp8, xmask)
tl.store(out_ptr1 + (x0), tmp23, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/r4/cr4j7hzjw2quuyvesstjj4anrvqw5igyzqrj3sm3fqqycoyilvv7.py
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# linear => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x5 = (xindex // 4)
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr1 + (x5), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + (x5), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/z5/cz5ftw3jjdcl7kf3rjnx3ukf3ssp6cwhcagtkza76pvoehzvmo6f.py
# Topologically Sorted Source Nodes: [linear, gelu], Original ATen: [aten.add, aten.gelu]
# Source node to ATen node mapping:
# gelu => add_3, erf, mul_2, mul_3, mul_4
# linear => add_2
# Graph fragment:
# %add_2 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_1, %primals_5), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%add_2, 0.5), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%add_2, 0.7071067811865476), kwargs = {})
# %erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_3,), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_2, %add_3), kwargs = {})
triton_poi_fused_add_gelu_2 = async_compile.triton('triton_poi_fused_add_gelu_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_gelu_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_gelu_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.5
tmp4 = tmp2 * tmp3
tmp5 = 0.7071067811865476
tmp6 = tmp2 * tmp5
tmp7 = libdevice.erf(tmp6)
tmp8 = 1.0
tmp9 = tmp7 + tmp8
tmp10 = tmp4 * tmp9
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/x2/cx2zp56snyrw27gd3kepcb5sb7pjfocirge6if2uvy2hu2bpjaxk.py
# Topologically Sorted Source Nodes: [out, x_5], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# out => add_4
# x_5 => clone_3, var_mean_1
# Graph fragment:
# %add_4 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%permute_3, %primals_1), kwargs = {})
# %clone_3 : [num_users=2] = call_function[target=torch.ops.aten.clone.default](args = (%add_4,), kwargs = {memory_format: torch.contiguous_format})
# %var_mean_1 : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%clone_3, [3]), kwargs = {correction: 0, keepdim: True})
triton_poi_fused_add_native_layer_norm_3 = async_compile.triton('triton_poi_fused_add_native_layer_norm_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_3(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16)
x3 = xindex
tmp0 = tl.load(in_ptr0 + ((4*x1) + (16*x0) + (64*x2)), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4*x3), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + (4*x1) + (16*x0) + (64*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + (4*x3)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + (4*x1) + (16*x0) + (64*x2)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + (4*x3)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + (4*x1) + (16*x0) + (64*x2)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (3 + (4*x3)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + (x3), tmp16, xmask)
tl.store(out_ptr1 + (x3), tmp28, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/p4/cp4poqr3q5r235xhpttemdweyjndyygisvksz7yupbqdk3m4obp6.py
# Topologically Sorted Source Nodes: [out, x_5], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# out => add_4
# x_5 => add_5, add_6, clone_3, mul_5, mul_6, rsqrt_1, sub_1
# Graph fragment:
# %add_4 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%permute_3, %primals_1), kwargs = {})
# %clone_3 : [num_users=2] = call_function[target=torch.ops.aten.clone.default](args = (%add_4,), kwargs = {memory_format: torch.contiguous_format})
# %add_5 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_2, 1e-05), kwargs = {})
# %rsqrt_1 : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add_5,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%clone_3, %getitem_3), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_1, %rsqrt_1), kwargs = {})
# %mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_5, %primals_8), kwargs = {})
# %add_6 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_6, %primals_9), kwargs = {})
triton_poi_fused_add_native_layer_norm_4 = async_compile.triton('triton_poi_fused_add_native_layer_norm_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_4(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x4 = xindex
x5 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp1 = tl.load(in_ptr1 + (x4), xmask)
tmp3 = tl.load(in_ptr2 + (x5), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x5), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + (x4), tmp13, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/qi/cqiodoqhh76dofgtvqgxsw37fmapfat76jykoxdahek5ub776gmr.py
# Topologically Sorted Source Nodes: [gelu_1], Original ATen: [aten.gelu]
# Source node to ATen node mapping:
# gelu_1 => add_7, erf_1, mul_7, mul_8, mul_9
# Graph fragment:
# %mul_7 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_5, 0.5), kwargs = {})
# %mul_8 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_5, 0.7071067811865476), kwargs = {})
# %erf_1 : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_8,), kwargs = {})
# %add_7 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf_1, 1), kwargs = {})
# %mul_9 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_7, %add_7), kwargs = {})
triton_poi_fused_gelu_5 = async_compile.triton('triton_poi_fused_gelu_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gelu_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_gelu_5(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + (x0), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2c/c2cw224zqp7xs4x6cuohaw22y32i2w6jyi5ugr25q6udpxxzd4pc.py
# Topologically Sorted Source Nodes: [out, out_1], Original ATen: [aten.add]
# Source node to ATen node mapping:
# out => add_4
# out_1 => add_8
# Graph fragment:
# %add_4 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%permute_3, %primals_1), kwargs = {})
# %add_8 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_7, %add_4), kwargs = {})
triton_poi_fused_add_6 = async_compile.triton('triton_poi_fused_add_6', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_6', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_6(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
tmp0 = tl.load(in_out_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x0 + (4*x2) + (16*x1) + (64*x3)), xmask)
tmp4 = tl.load(in_ptr2 + (x4), xmask)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + (x4), tmp6, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (16, 4), (4, 1))
assert_size_stride(primals_5, (16, ), (1, ))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4, ), (1, ))
assert_size_stride(primals_8, (4, ), (1, ))
assert_size_stride(primals_9, (4, ), (1, ))
assert_size_stride(primals_10, (16, 4), (4, 1))
assert_size_stride(primals_11, (16, ), (1, ))
assert_size_stride(primals_12, (4, 16), (16, 1))
assert_size_stride(primals_13, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
stream0 = get_raw_stream(0)
triton_poi_fused_native_layer_norm_0.run(primals_1, buf0, buf1, 64, grid=grid(64), stream=stream0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(primals_1, buf0, buf1, primals_2, primals_3, buf2, 256, grid=grid(256), stream=stream0)
del primals_2
del primals_3
buf3 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 16), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear, gelu], Original ATen: [aten.add, aten.gelu]
triton_poi_fused_add_gelu_2.run(buf3, primals_5, buf4, 1024, grid=grid(1024), stream=stream0)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_7, reinterpret_tensor(buf4, (64, 16), (16, 1), 0), reinterpret_tensor(primals_6, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf5)
del primals_7
buf6 = buf1; del buf1 # reuse
buf7 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [out, x_5], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_3.run(buf5, primals_1, buf6, buf7, 64, grid=grid(64), stream=stream0)
buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out, x_5], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_4.run(buf5, primals_1, buf6, buf7, primals_8, primals_9, buf8, 256, grid=grid(256), stream=stream0)
del buf6
del buf7
del primals_9
buf9 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear_2], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_11, reinterpret_tensor(buf8, (64, 4), (4, 1), 0), reinterpret_tensor(primals_10, (4, 16), (1, 4), 0), alpha=1, beta=1, out=buf9)
del primals_11
buf10 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.float32)
# Topologically Sorted Source Nodes: [gelu_1], Original ATen: [aten.gelu]
triton_poi_fused_gelu_5.run(buf9, buf10, 1024, grid=grid(1024), stream=stream0)
buf11 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf10, (64, 16), (16, 1), 0), reinterpret_tensor(primals_12, (16, 4), (1, 16), 0), out=buf11)
buf12 = reinterpret_tensor(buf11, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf11 # reuse
# Topologically Sorted Source Nodes: [out, out_1], Original ATen: [aten.add]
triton_poi_fused_add_6.run(buf12, primals_13, buf5, primals_1, 256, grid=grid(256), stream=stream0)
del primals_13
return (buf12, primals_1, primals_5, primals_8, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), buf3, reinterpret_tensor(buf4, (64, 16), (16, 1), 0), buf5, reinterpret_tensor(buf8, (64, 4), (4, 1), 0), buf9, reinterpret_tensor(buf10, (64, 16), (16, 1), 0), primals_12, primals_10, primals_6, primals_4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_12 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_13 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class ChannelMixer(nn.Module):
def __init__(self, d_model, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(d_model, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = self.mlp(x)
out = x + residual
return out
class TokenMixer(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(seq_len, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = x.transpose(1, 2)
x = self.mlp(x)
x = x.transpose(1, 2)
out = x + residual
return out
class MixerBlock(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.token_mixer = TokenMixer(d_model, seq_len, expansion_factor,
dropout)
self.channel_mixer = ChannelMixer(d_model, expansion_factor, dropout)
def forward(self, x):
x = self.token_mixer(x)
x = self.channel_mixer(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'seq_len': 4, 'expansion_factor': 4,
'dropout': 0.5}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn.functional as F
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x5 = xindex // 4
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x5, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), tmp8, xmask)
@triton.jit
def triton_poi_fused_add_gelu_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.5
tmp4 = tmp2 * tmp3
tmp5 = 0.7071067811865476
tmp6 = tmp2 * tmp5
tmp7 = libdevice.erf(tmp6)
tmp8 = 1.0
tmp9 = tmp7 + tmp8
tmp10 = tmp4 * tmp9
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_3(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16
x3 = xindex
tmp0 = tl.load(in_ptr0 + (4 * x1 + 16 * x0 + 64 * x2), xmask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x3, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x1 + 16 * x0 + 64 * x2), xmask,
eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x3), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x1 + 16 * x0 + 64 * x2), xmask,
eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x3), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x1 + 16 * x0 + 64 * x2), xmask,
eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (3 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x3, tmp16, xmask)
tl.store(out_ptr1 + x3, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_4(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
x5 = xindex // 4
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp1 = tl.load(in_ptr1 + x4, xmask)
tmp3 = tl.load(in_ptr2 + x5, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x5, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x4, tmp13, xmask)
@triton.jit
def triton_poi_fused_gelu_5(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
@triton.jit
def triton_poi_fused_add_6(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
tmp0 = tl.load(in_out_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp4 = tl.load(in_ptr2 + x4, xmask)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + x4, tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (16, 4), (4, 1))
assert_size_stride(primals_5, (16,), (1,))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4,), (1,))
assert_size_stride(primals_10, (16, 4), (4, 1))
assert_size_stride(primals_11, (16,), (1,))
assert_size_stride(primals_12, (4, 16), (16, 1))
assert_size_stride(primals_13, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
get_raw_stream(0)
triton_poi_fused_native_layer_norm_0[grid(64)](primals_1, buf0,
buf1, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(256)](primals_1, buf0, buf1,
primals_2, primals_3, buf2, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del primals_2
del primals_3
buf3 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 16), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
triton_poi_fused_add_gelu_2[grid(1024)](buf3, primals_5, buf4, 1024,
XBLOCK=256, num_warps=4, num_stages=1)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, reinterpret_tensor(buf4, (64, 16),
(16, 1), 0), reinterpret_tensor(primals_6, (16, 4), (1, 16), 0),
alpha=1, beta=1, out=buf5)
del primals_7
buf6 = buf1
del buf1
buf7 = buf0
del buf0
triton_poi_fused_add_native_layer_norm_3[grid(64)](buf5, primals_1,
buf6, buf7, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_4[grid(256)](buf5, primals_1,
buf6, buf7, primals_8, primals_9, buf8, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del buf6
del buf7
del primals_9
buf9 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
extern_kernels.addmm(primals_11, reinterpret_tensor(buf8, (64, 4),
(4, 1), 0), reinterpret_tensor(primals_10, (4, 16), (1, 4), 0),
alpha=1, beta=1, out=buf9)
del primals_11
buf10 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
triton_poi_fused_gelu_5[grid(1024)](buf9, buf10, 1024, XBLOCK=128,
num_warps=4, num_stages=1)
buf11 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf10, (64, 16), (16, 1), 0),
reinterpret_tensor(primals_12, (16, 4), (1, 16), 0), out=buf11)
buf12 = reinterpret_tensor(buf11, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf11
triton_poi_fused_add_6[grid(256)](buf12, primals_13, buf5,
primals_1, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_13
return buf12, primals_1, primals_5, primals_8, reinterpret_tensor(buf2,
(64, 4), (4, 1), 0), buf3, reinterpret_tensor(buf4, (64, 16), (16,
1), 0), buf5, reinterpret_tensor(buf8, (64, 4), (4, 1), 0
), buf9, reinterpret_tensor(buf10, (64, 16), (16, 1), 0
), primals_12, primals_10, primals_6, primals_4
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class ChannelMixer(nn.Module):
def __init__(self, d_model, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(d_model, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = self.mlp(x)
out = x + residual
return out
class TokenMixer(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(seq_len, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = x.transpose(1, 2)
x = self.mlp(x)
x = x.transpose(1, 2)
out = x + residual
return out
class MixerBlockNew(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.token_mixer = TokenMixer(d_model, seq_len, expansion_factor,
dropout)
self.channel_mixer = ChannelMixer(d_model, expansion_factor, dropout)
def forward(self, input_0):
primals_2 = self.token_mixer.norm.weight
primals_3 = self.token_mixer.norm.bias
primals_4 = self.token_mixer.mlp.fc1.weight
primals_5 = self.token_mixer.mlp.fc1.bias
primals_6 = self.token_mixer.mlp.fc2.weight
primals_7 = self.token_mixer.mlp.fc2.bias
primals_8 = self.channel_mixer.norm.weight
primals_9 = self.channel_mixer.norm.bias
primals_10 = self.channel_mixer.mlp.fc1.weight
primals_11 = self.channel_mixer.mlp.fc1.bias
primals_12 = self.channel_mixer.mlp.fc2.weight
primals_13 = self.channel_mixer.mlp.fc2.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13])
return output[0]
| Misuzu-Kurenai/mlp-singer | MixerBlock | false | 866 | [
"MIT"
] | 0 | 416451045bb9b3965aaf496e84a8b45332a6ba59 | https://github.com/Misuzu-Kurenai/mlp-singer/tree/416451045bb9b3965aaf496e84a8b45332a6ba59 | import torch
import torch.nn.functional as F
from torch import nn
class FeedForward(nn.Module):
def __init__(self, num_features, expansion_factor, dropout):
super().__init__()
num_hidden = expansion_factor * num_features
self.fc1 = nn.Linear(num_features, num_hidden)
self.fc2 = nn.Linear(num_hidden, num_features)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
x = self.dropout1(F.gelu(self.fc1(x)))
x = self.dropout2(self.fc2(x))
return x
class ChannelMixer(nn.Module):
def __init__(self, d_model, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(d_model, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = self.mlp(x)
out = x + residual
return out
class TokenMixer(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.norm = nn.LayerNorm(d_model)
self.mlp = FeedForward(seq_len, expansion_factor, dropout)
def forward(self, x):
residual = x
x = self.norm(x)
x = x.transpose(1, 2)
x = self.mlp(x)
x = x.transpose(1, 2)
out = x + residual
return out
class Model(nn.Module):
def __init__(self, d_model, seq_len, expansion_factor, dropout):
super().__init__()
self.token_mixer = TokenMixer(d_model, seq_len, expansion_factor,
dropout)
self.channel_mixer = ChannelMixer(d_model, expansion_factor, dropout)
def forward(self, x):
x = self.token_mixer(x)
x = self.channel_mixer(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'seq_len': 4, 'expansion_factor': 4,
'dropout': 0.5}]
|
linformerAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ez/cezmv74yrhrunjwqrletcmzzbnanma4ylsle3v7w345t7kxp622s.py
# Topologically Sorted Source Nodes: [k_1], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# k_1 => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_0 = async_compile.triton('triton_poi_fused_clone_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64, 4], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = (yindex // 4)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x2) + (16*y1)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + (4*y3)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2m/c2m2l4qnodgs3ot5g3cwne522qhk3gkewrbzygpibw7yj5zhwduc.py
# Topologically Sorted Source Nodes: [k_1, v_1], Original ATen: [aten.add]
# Source node to ATen node mapping:
# k_1 => add
# v_1 => add_1
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_1, %primals_3), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_7, %primals_3), kwargs = {})
triton_poi_fused_add_1 = async_compile.triton('triton_poi_fused_add_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[2048],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_1', 'mutated_arg_names': ['in_out_ptr0', 'in_out_ptr1'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_1(in_out_ptr0, in_out_ptr1, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1280
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 20
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_out_ptr1 + (x2), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp3 + tmp1
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
tl.store(in_out_ptr1 + (x2), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/3w/c3wpg3vhtw2uvjzlz5viwm3dnhvkbjudvttnxifnse5cypcefklu.py
# Topologically Sorted Source Nodes: [sqrt, P_bar_1], Original ATen: [aten.sqrt, aten._softmax]
# Source node to ATen node mapping:
# P_bar_1 => div_1, exp, sum_1
# sqrt => full_default
# Graph fragment:
# %full_default : [num_users=2] = call_function[target=torch.ops.aten.full.default](args = ([], 4.4721360206604), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %scalar_tensor_default : [num_users=2] = call_function[target=torch.ops.aten.scalar_tensor.default](args = (1,), kwargs = {dtype: torch.float32, device: cuda:0, pin_memory: False})
# %ge_scalar : [num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%full_default, 0), kwargs = {})
# %neg_default : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%scalar_tensor_default,), kwargs = {})
# %where_self : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ge_scalar, %scalar_tensor_default, %neg_default), kwargs = {})
# %mul_tensor : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_5, %where_self), kwargs = {})
# %amax_default : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%mul_tensor, [-1], True), kwargs = {})
# %sub_tensor : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_tensor, %amax_default), kwargs = {})
# %mul_tensor_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where_self, %full_default), kwargs = {})
# %div_tensor : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sub_tensor, %mul_tensor_1), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%div_tensor,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div_1 : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_per_fused__softmax_sqrt_2 = async_compile.triton('triton_per_fused__softmax_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[64, 32],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused__softmax_sqrt_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 2, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused__softmax_sqrt_2(in_ptr0, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 64
rnumel = 20
RBLOCK: tl.constexpr = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = rindex < rnumel
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + (20*x0)), rmask & xmask, other=0.0)
tmp1 = 4.4721360206604
tmp2 = 0.0
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp8 = tl.broadcast_to(tmp7, [XBLOCK, RBLOCK])
tmp10 = tl.where(rmask & xmask, tmp8, float("-inf"))
tmp11 = triton_helpers.max2(tmp10, 1)[:, None]
tmp12 = tmp7 - tmp11
tmp13 = tmp6 * tmp1
tmp14 = tmp12 / tmp13
tmp15 = tl_math.exp(tmp14)
tmp16 = tl.broadcast_to(tmp15, [XBLOCK, RBLOCK])
tmp18 = tl.where(rmask & xmask, tmp16, 0)
tmp19 = tl.sum(tmp18, 1)[:, None]
tmp20 = tmp15 / tmp19
tl.store(out_ptr2 + (r1 + (20*x0)), tmp20, rmask & xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (20, 4), (4, 1))
assert_size_stride(primals_3, (20, ), (1, ))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [k_1], Original ATen: [aten.clone]
stream0 = get_raw_stream(0)
triton_poi_fused_clone_0.run(primals_1, buf0, 64, 4, grid=grid(64, 4), stream=stream0)
del primals_1
buf1 = empty_strided_cuda((64, 20), (20, 1), torch.float32)
# Topologically Sorted Source Nodes: [k_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 20), (1, 4), 0), out=buf1)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [v_1], Original ATen: [aten.clone]
triton_poi_fused_clone_0.run(primals_5, buf7, 64, 4, grid=grid(64, 4), stream=stream0)
del primals_5
buf8 = empty_strided_cuda((64, 20), (20, 1), torch.float32)
# Topologically Sorted Source Nodes: [v_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf7, (64, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 20), (1, 4), 0), out=buf8)
del primals_2
buf2 = reinterpret_tensor(buf1, (4, 4, 4, 20), (320, 80, 20, 1), 0); del buf1 # reuse
buf9 = reinterpret_tensor(buf8, (4, 4, 4, 20), (320, 80, 20, 1), 0); del buf8 # reuse
# Topologically Sorted Source Nodes: [k_1, v_1], Original ATen: [aten.add]
triton_poi_fused_add_1.run(buf2, buf9, primals_3, 1280, grid=grid(1280), stream=stream0)
del primals_3
buf3 = empty_strided_cuda((16, 4, 20), (80, 20, 1), torch.float32)
# Topologically Sorted Source Nodes: [q], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(primals_4, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf2, (16, 4, 20), (80, 20, 1), 0), out=buf3)
buf6 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [sqrt, P_bar_1], Original ATen: [aten.sqrt, aten._softmax]
triton_per_fused__softmax_sqrt_2.run(buf3, buf6, 64, 20, grid=grid(64), stream=stream0)
del buf3
buf10 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 20), (80, 20, 1), 0), reinterpret_tensor(buf9, (16, 20, 4), (80, 1, 20), 0), out=buf10)
return (reinterpret_tensor(buf10, (4, 4, 4, 4), (64, 16, 4, 1), 0), reinterpret_tensor(buf0, (64, 4), (4, 1), 0), buf6, reinterpret_tensor(buf7, (64, 4), (4, 1), 0), reinterpret_tensor(buf9, (16, 4, 20), (80, 20, 1), 0), reinterpret_tensor(primals_4, (16, 4, 4), (16, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((20, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((20, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
def get_EF(input_size, dim, method='learnable', head_dim=None, bias=True):
"""
Retuns the E or F matrix, initialized via xavier initialization.
This is the recommended way to do it according to the authors of the paper.
Includes a method for convolution, as well as a method for no additional params.
"""
assert method == 'learnable' or method == 'convolution' or method == 'no_params', "The method flag needs to be either 'learnable', 'convolution', or 'no_params'!"
if method == 'convolution':
conv = nn.Conv1d(head_dim, head_dim, kernel_size=int(input_size /
dim), stride=int(input_size / dim))
return conv
if method == 'no_params':
mat = torch.zeros((input_size, dim))
torch.nn.init.normal_(mat, mean=0.0, std=1 / dim)
return mat
lin = nn.Linear(input_size, dim, bias)
torch.nn.init.xavier_normal_(lin.weight)
return lin
class linformerAttention(nn.Module):
def __init__(self, dim, dropout, input_size, dim_k=20, full_attention=
False, parameter_sharing=None):
super().__init__()
self.dim = dim
self.dropout = nn.Dropout(dropout)
self.dim_k = dim_k
self.full_attention = full_attention
self.input_size = input_size
self.print_dim = False
self.E = get_EF(input_size, dim=self.dim_k, method='learnable',
head_dim=self.dim)
self.F = (get_EF(input_size, dim=self.dim_k, method='learnable',
head_dim=self.dim) if parameter_sharing == 'none' or
parameter_sharing == 'headwise' else self.E)
self.is_proj_tensor = isinstance(self.E, torch.Tensor)
def forward(self, q, k, v):
if self.print_dim:
None
None
None
if not self.full_attention:
if self.is_proj_tensor:
self.E = self.E
b, h, *_ = q.shape
projection_E = repeat(self.E, 'j d -> b h j d', b=b, h=h)
k = torch.einsum('...di,...dj->...ij', k, projection_E)
else:
k = torch.einsum('...ij->...ji', k)
k = self.E(k)
if self.print_dim:
None
None
None
q = torch.einsum('...id,...dj->...ij', q, k)
P_bar = q / torch.sqrt(torch.tensor(self.dim_k).type(q.type()))
P_bar = P_bar.softmax(dim=-1)
P_bar = self.dropout(P_bar)
if not self.full_attention:
if self.is_proj_tensor:
self.F = self.F
v = torch.matmul(v, self.F)
else:
v = torch.einsum('...ij->...ji', v)
v = self.F(v)
out = torch.einsum('...id,...jd->...ij', P_bar, v)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4])]
def get_init_inputs():
return [[], {'dim': 4, 'dropout': 0.5, 'input_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_1(in_out_ptr0, in_out_ptr1, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 1280
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 20
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_out_ptr1 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp3 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
tl.store(in_out_ptr1 + x2, tmp4, xmask)
@triton.jit
def triton_per_fused__softmax_sqrt_2(in_ptr0, out_ptr2, xnumel, rnumel,
XBLOCK: tl.constexpr):
xnumel = 64
rnumel = 20
RBLOCK: tl.constexpr = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
rmask = rindex < rnumel
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 20 * x0), rmask & xmask, other=0.0)
tmp1 = 4.4721360206604
tmp2 = 0.0
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp8 = tl.broadcast_to(tmp7, [XBLOCK, RBLOCK])
tmp10 = tl.where(rmask & xmask, tmp8, float('-inf'))
tmp11 = triton_helpers.max2(tmp10, 1)[:, None]
tmp12 = tmp7 - tmp11
tmp13 = tmp6 * tmp1
tmp14 = tmp12 / tmp13
tmp15 = tl_math.exp(tmp14)
tmp16 = tl.broadcast_to(tmp15, [XBLOCK, RBLOCK])
tmp18 = tl.where(rmask & xmask, tmp16, 0)
tmp19 = tl.sum(tmp18, 1)[:, None]
tmp20 = tmp15 / tmp19
tl.store(out_ptr2 + (r1 + 20 * x0), tmp20, rmask & xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (20, 4), (4, 1))
assert_size_stride(primals_3, (20,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(64, 4)](primals_1, buf0, 64, 4,
XBLOCK=4, YBLOCK=32, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((64, 20), (20, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 20), (1, 4), 0), out=buf1)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_0[grid(64, 4)](primals_5, buf7, 64, 4,
XBLOCK=4, YBLOCK=32, num_warps=4, num_stages=1)
del primals_5
buf8 = empty_strided_cuda((64, 20), (20, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf7, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 20), (1, 4), 0), out=buf8)
del primals_2
buf2 = reinterpret_tensor(buf1, (4, 4, 4, 20), (320, 80, 20, 1), 0)
del buf1
buf9 = reinterpret_tensor(buf8, (4, 4, 4, 20), (320, 80, 20, 1), 0)
del buf8
triton_poi_fused_add_1[grid(1280)](buf2, buf9, primals_3, 1280,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_3
buf3 = empty_strided_cuda((16, 4, 20), (80, 20, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(primals_4, (16, 4, 4), (16, 4,
1), 0), reinterpret_tensor(buf2, (16, 4, 20), (80, 20, 1), 0),
out=buf3)
buf6 = buf2
del buf2
triton_per_fused__softmax_sqrt_2[grid(64)](buf3, buf6, 64, 20,
XBLOCK=8, num_warps=2, num_stages=1)
del buf3
buf10 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 20), (80, 20, 1
), 0), reinterpret_tensor(buf9, (16, 20, 4), (80, 1, 20), 0),
out=buf10)
return reinterpret_tensor(buf10, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(buf0, (64, 4), (4, 1), 0
), buf6, reinterpret_tensor(buf7, (64, 4), (4, 1), 0
), reinterpret_tensor(buf9, (16, 4, 20), (80, 20, 1), 0
), reinterpret_tensor(primals_4, (16, 4, 4), (16, 1, 4), 0)
def get_EF(input_size, dim, method='learnable', head_dim=None, bias=True):
"""
Retuns the E or F matrix, initialized via xavier initialization.
This is the recommended way to do it according to the authors of the paper.
Includes a method for convolution, as well as a method for no additional params.
"""
assert method == 'learnable' or method == 'convolution' or method == 'no_params', "The method flag needs to be either 'learnable', 'convolution', or 'no_params'!"
if method == 'convolution':
conv = nn.Conv1d(head_dim, head_dim, kernel_size=int(input_size /
dim), stride=int(input_size / dim))
return conv
if method == 'no_params':
mat = torch.zeros((input_size, dim))
torch.nn.init.normal_(mat, mean=0.0, std=1 / dim)
return mat
lin = nn.Linear(input_size, dim, bias)
torch.nn.init.xavier_normal_(lin.weight)
return lin
class linformerAttentionNew(nn.Module):
def __init__(self, dim, dropout, input_size, dim_k=20, full_attention=
False, parameter_sharing=None):
super().__init__()
self.dim = dim
self.dropout = nn.Dropout(dropout)
self.dim_k = dim_k
self.full_attention = full_attention
self.input_size = input_size
self.print_dim = False
self.E = get_EF(input_size, dim=self.dim_k, method='learnable',
head_dim=self.dim)
self.F = (get_EF(input_size, dim=self.dim_k, method='learnable',
head_dim=self.dim) if parameter_sharing == 'none' or
parameter_sharing == 'headwise' else self.E)
self.is_proj_tensor = isinstance(self.E, torch.Tensor)
def forward(self, input_0, input_1, input_2):
primals_2 = self.E.weight
primals_3 = self.E.bias
primals_1 = input_0
primals_4 = input_1
primals_5 = input_2
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| MohammadrezaRezvani/performer-pytorch | linformerAttention | false | 867 | [
"MIT"
] | 0 | 347dd58111f4f79b8991f7609552203609856b4b | https://github.com/MohammadrezaRezvani/performer-pytorch/tree/347dd58111f4f79b8991f7609552203609856b4b | import torch
from torch import nn
def get_EF(input_size, dim, method='learnable', head_dim=None, bias=True):
"""
Retuns the E or F matrix, initialized via xavier initialization.
This is the recommended way to do it according to the authors of the paper.
Includes a method for convolution, as well as a method for no additional params.
"""
assert method == 'learnable' or method == 'convolution' or method == 'no_params', "The method flag needs to be either 'learnable', 'convolution', or 'no_params'!"
if method == 'convolution':
conv = nn.Conv1d(head_dim, head_dim, kernel_size=int(input_size /
dim), stride=int(input_size / dim))
return conv
if method == 'no_params':
mat = torch.zeros((input_size, dim))
torch.nn.init.normal_(mat, mean=0.0, std=1 / dim)
return mat
lin = nn.Linear(input_size, dim, bias)
torch.nn.init.xavier_normal_(lin.weight)
return lin
class Model(nn.Module):
def __init__(self, dim, dropout, input_size, dim_k=20, full_attention=
False, parameter_sharing=None):
super().__init__()
self.dim = dim
self.dropout = nn.Dropout(dropout)
self.dim_k = dim_k
self.full_attention = full_attention
self.input_size = input_size
self.print_dim = False
self.E = get_EF(input_size, dim=self.dim_k, method='learnable',
head_dim=self.dim)
self.F = (get_EF(input_size, dim=self.dim_k, method='learnable',
head_dim=self.dim) if parameter_sharing == 'none' or
parameter_sharing == 'headwise' else self.E)
self.is_proj_tensor = isinstance(self.E, torch.Tensor)
def forward(self, q, k, v):
if self.print_dim:
None
None
None
if not self.full_attention:
if self.is_proj_tensor:
self.E = self.E
b, h, *_ = q.shape
projection_E = repeat(self.E, 'j d -> b h j d', b=b, h=h)
k = torch.einsum('...di,...dj->...ij', k, projection_E)
else:
k = torch.einsum('...ij->...ji', k)
k = self.E(k)
if self.print_dim:
None
None
None
q = torch.einsum('...id,...dj->...ij', q, k)
P_bar = q / torch.sqrt(torch.tensor(self.dim_k).type(q.type()))
P_bar = P_bar.softmax(dim=-1)
P_bar = self.dropout(P_bar)
if not self.full_attention:
if self.is_proj_tensor:
self.F = self.F
v = torch.matmul(v, self.F)
else:
v = torch.einsum('...ij->...ji', v)
v = self.F(v)
out = torch.einsum('...id,...jd->...ij', P_bar, v)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4])]
def get_init_inputs():
return [4, 0.5, 4]
|
FGFunction | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/et/cetbvewldbe3nltlmooux3gvif26azg2nu4k4k7ld7hdklxsohux.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_group_norm]
# Source node to ATen node mapping:
# x => add, add_1, mul_2, rsqrt, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%view, [2, 3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 0.001), kwargs = {})
# %rsqrt : [num_users=2] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, %unsqueeze_7), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_2, %unsqueeze_4), kwargs = {})
triton_per_fused_native_group_norm_0 = async_compile.triton('triton_per_fused_native_group_norm_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[4, 64],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_native_group_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 4, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_native_group_norm_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr2, out_ptr3, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 4
rnumel = 64
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
r3 = (rindex // 16)
tmp0 = tl.load(in_ptr0 + (r1 + (64*x0)), xmask, other=0.0)
tmp29 = tl.load(in_ptr1 + (r3), None, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr2 + (r3), None, eviction_policy='evict_last')
tmp1 = 0.0
tmp2 = tmp0 > tmp1
tmp3 = 0.01
tmp4 = tmp0 * tmp3
tmp5 = tl.where(tmp2, tmp0, tmp4)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK, RBLOCK])
tmp8 = tl.where(xmask, tmp6, 0)
tmp9 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp11 = tl.where(xmask, tmp9, 0)
tmp12 = tl.sum(tmp11, 1)[:, None]
tmp13 = tl.full([XBLOCK, 1], 64, tl.int32)
tmp14 = tmp13.to(tl.float32)
tmp15 = tmp12 / tmp14
tmp16 = tmp6 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tl.broadcast_to(tmp17, [XBLOCK, RBLOCK])
tmp20 = tl.where(xmask, tmp18, 0)
tmp21 = tl.sum(tmp20, 1)[:, None]
tmp22 = tmp5 - tmp15
tmp23 = 64.0
tmp24 = tmp21 / tmp23
tmp25 = 0.001
tmp26 = tmp24 + tmp25
tmp27 = libdevice.rsqrt(tmp26)
tmp28 = tmp22 * tmp27
tmp30 = tmp28 * tmp29
tmp32 = tmp30 + tmp31
tl.store(out_ptr2 + (r1 + (64*x0)), tmp32, xmask)
tl.store(out_ptr3 + (x0), tmp27, xmask)
tl.store(out_ptr0 + (x0), tmp15, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3, 3), (108, 27, 9, 3, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (4, ), (1, ))
assert_size_stride(primals_5, (4, 4, 3, 3, 3), (108, 27, 9, 3, 1))
assert_size_stride(primals_6, (4, ), (1, ))
assert_size_stride(primals_7, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [conv3d], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), primals_1, stride=(1, 1, 1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False, output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf0, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1))
buf1 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf4 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_group_norm]
stream0 = get_raw_stream(0)
triton_per_fused_native_group_norm_0.run(buf0, primals_3, primals_4, buf1, buf5, buf4, 4, 64, grid=grid(4), stream=stream0)
del primals_4
# Topologically Sorted Source Nodes: [conv3d_1], Original ATen: [aten.convolution]
buf6 = extern_kernels.convolution(reinterpret_tensor(buf5, (1, 4, 4, 4, 4), (0, 64, 16, 4, 1), 0), primals_5, stride=(1, 1, 1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False, output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf6, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1))
buf7 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf11 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
triton_per_fused_native_group_norm_0.run(buf6, primals_6, primals_7, buf7, buf10, buf11, 4, 64, grid=grid(4), stream=stream0)
del primals_7
return (buf10, primals_1, primals_3, primals_5, primals_6, reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), buf0, reinterpret_tensor(buf1, (4, 1), (1, 1), 0), reinterpret_tensor(buf4, (4, 1), (1, 1), 0), reinterpret_tensor(buf5, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), buf6, reinterpret_tensor(buf7, (4, 1), (1, 1), 0), reinterpret_tensor(buf11, (4, 1), (1, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 3, 3, 3), (108, 27, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, 4, 3, 3, 3), (108, 27, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
from torch.nn import functional as F
class FGFunction(nn.Module):
"""Module used for F and G
Archi :
conv -> BN -> ReLu -> conv -> BN -> ReLu
"""
def __init__(self, channels):
super(FGFunction, self).__init__()
self.gn1 = nn.GroupNorm(1, channels, eps=0.001)
self.conv1 = nn.Conv3d(channels, channels, 3, padding=1, bias=False)
self.gn2 = nn.GroupNorm(1, channels, eps=0.001)
self.conv2 = nn.Conv3d(channels, channels, 3, padding=1, bias=False)
def forward(self, x):
x = self.gn1(F.leaky_relu(self.conv1(x), inplace=True))
x = self.gn2(F.leaky_relu(self.conv2(x), inplace=True))
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'channels': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_native_group_norm_0(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, out_ptr2, out_ptr3, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
r3 = rindex // 16
tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0)
tmp29 = tl.load(in_ptr1 + r3, None, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr2 + r3, None, eviction_policy='evict_last')
tmp1 = 0.0
tmp2 = tmp0 > tmp1
tmp3 = 0.01
tmp4 = tmp0 * tmp3
tmp5 = tl.where(tmp2, tmp0, tmp4)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK, RBLOCK])
tl.where(xmask, tmp6, 0)
tmp9 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp11 = tl.where(xmask, tmp9, 0)
tmp12 = tl.sum(tmp11, 1)[:, None]
tmp13 = tl.full([XBLOCK, 1], 64, tl.int32)
tmp14 = tmp13.to(tl.float32)
tmp15 = tmp12 / tmp14
tmp16 = tmp6 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tl.broadcast_to(tmp17, [XBLOCK, RBLOCK])
tmp20 = tl.where(xmask, tmp18, 0)
tmp21 = tl.sum(tmp20, 1)[:, None]
tmp22 = tmp5 - tmp15
tmp23 = 64.0
tmp24 = tmp21 / tmp23
tmp25 = 0.001
tmp26 = tmp24 + tmp25
tmp27 = libdevice.rsqrt(tmp26)
tmp28 = tmp22 * tmp27
tmp30 = tmp28 * tmp29
tmp32 = tmp30 + tmp31
tl.store(out_ptr2 + (r1 + 64 * x0), tmp32, xmask)
tl.store(out_ptr3 + x0, tmp27, xmask)
tl.store(out_ptr0 + x0, tmp15, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3, 3), (108, 27, 9, 3, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4, 4, 3, 3, 3), (108, 27, 9, 3, 1))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1,
4, 4, 4, 4), (256, 64, 16, 4, 1), 0), primals_1, stride=(1, 1,
1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf0, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1))
buf1 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf4 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
get_raw_stream(0)
triton_per_fused_native_group_norm_0[grid(4)](buf0, primals_3,
primals_4, buf1, buf5, buf4, 4, 64, XBLOCK=1, num_warps=2,
num_stages=1)
del primals_4
buf6 = extern_kernels.convolution(reinterpret_tensor(buf5, (1, 4, 4,
4, 4), (0, 64, 16, 4, 1), 0), primals_5, stride=(1, 1, 1),
padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf6, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1))
buf7 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf11 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_per_fused_native_group_norm_0[grid(4)](buf6, primals_6,
primals_7, buf7, buf10, buf11, 4, 64, XBLOCK=1, num_warps=2,
num_stages=1)
del primals_7
return (buf10, primals_1, primals_3, primals_5, primals_6,
reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1),
0), buf0, reinterpret_tensor(buf1, (4, 1), (1, 1), 0),
reinterpret_tensor(buf4, (4, 1), (1, 1), 0), reinterpret_tensor(
buf5, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), buf6,
reinterpret_tensor(buf7, (4, 1), (1, 1), 0), reinterpret_tensor(
buf11, (4, 1), (1, 1), 0))
class FGFunctionNew(nn.Module):
"""Module used for F and G
Archi :
conv -> BN -> ReLu -> conv -> BN -> ReLu
"""
def __init__(self, channels):
super(FGFunctionNew, self).__init__()
self.gn1 = nn.GroupNorm(1, channels, eps=0.001)
self.conv1 = nn.Conv3d(channels, channels, 3, padding=1, bias=False)
self.gn2 = nn.GroupNorm(1, channels, eps=0.001)
self.conv2 = nn.Conv3d(channels, channels, 3, padding=1, bias=False)
def forward(self, input_0):
primals_3 = self.gn1.weight
primals_4 = self.gn1.bias
primals_1 = self.conv1.weight
primals_6 = self.gn2.weight
primals_7 = self.gn2.bias
primals_5 = self.conv2.weight
primals_2 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
| Myyyr/segmentation | FGFunction | false | 868 | [
"MIT"
] | 0 | 6b9423e327cff1eb23599404031b7fb8e9ecf75d | https://github.com/Myyyr/segmentation/tree/6b9423e327cff1eb23599404031b7fb8e9ecf75d | import torch
from torch import nn
from torch.nn import functional as F
class Model(nn.Module):
"""Module used for F and G
Archi :
conv -> BN -> ReLu -> conv -> BN -> ReLu
"""
def __init__(self, channels):
super().__init__()
self.gn1 = nn.GroupNorm(1, channels, eps=0.001)
self.conv1 = nn.Conv3d(channels, channels, 3, padding=1, bias=False)
self.gn2 = nn.GroupNorm(1, channels, eps=0.001)
self.conv2 = nn.Conv3d(channels, channels, 3, padding=1, bias=False)
def forward(self, x):
x = self.gn1(F.leaky_relu(self.conv1(x), inplace=True))
x = self.gn2(F.leaky_relu(self.conv2(x), inplace=True))
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
GLU | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/si/csi3q2lvu3kn5mel4tdzl2vtzommba4fd5qu3gyxcskc5f7bdxxu.py
# Topologically Sorted Source Nodes: [sigmoid, mul], Original ATen: [aten.sigmoid, aten.mul]
# Source node to ATen node mapping:
# mul => mul
# sigmoid => sigmoid
# Graph fragment:
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%slice_4,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%slice_2, %sigmoid), kwargs = {})
triton_poi_fused_mul_sigmoid_0 = async_compile.triton('triton_poi_fused_mul_sigmoid_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_sigmoid_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_sigmoid_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 32
x1 = (xindex // 32)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (64*x1)), xmask)
tmp1 = tl.load(in_ptr0 + (32 + x0 + (64*x1)), xmask)
tmp2 = tl.sigmoid(tmp1)
tmp3 = tmp0 * tmp2
tl.store(out_ptr0 + (x2), tmp3, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sigmoid, mul], Original ATen: [aten.sigmoid, aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_sigmoid_0.run(arg0_1, buf0, 128, grid=grid(128), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.parallel
import torch.onnx
class GLU(nn.Module):
def __init__(self):
super(GLU, self).__init__()
def forward(self, x):
nc = x.size(1)
assert nc % 2 == 0, 'channels dont divide 2!'
nc = int(nc / 2)
return x[:, :nc] * torch.sigmoid(x[:, nc:])
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
import torch.nn.parallel
import torch.onnx
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_sigmoid_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 32
x1 = xindex // 32
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp2 = tl.sigmoid(tmp1)
tmp3 = tmp0 * tmp2
tl.store(out_ptr0 + x2, tmp3, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_sigmoid_0[grid(128)](arg0_1, buf0, 128, XBLOCK
=128, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class GLUNew(nn.Module):
def __init__(self):
super(GLUNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Nakachi-S/AttnGAN | GLU | false | 869 | [
"MIT"
] | 0 | 2dfd1e38f78f2a58895d81131cd8c17e74dbacb2 | https://github.com/Nakachi-S/AttnGAN/tree/2dfd1e38f78f2a58895d81131cd8c17e74dbacb2 | import torch
import torch.nn as nn
import torch.nn.parallel
import torch.onnx
class Model(nn.Module):
def __init__(self):
super().__init__()
def forward(self, x):
nc = x.size(1)
assert nc % 2 == 0, 'channels dont divide 2!'
nc = int(nc / 2)
return x[:, :nc] * torch.sigmoid(x[:, nc:])
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
LinearBlock | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/67/c67nwtvejry7jovauj4gfwnoxaeorcxngrt3hs2o3y3i4niugvpw.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten._native_batch_norm_legit]
# Source node to ATen node mapping:
# x_1 => add, rsqrt, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%view_1, [0, 2]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 0.001), kwargs = {})
# %rsqrt : [num_users=2] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
triton_per_fused__native_batch_norm_legit_0 = async_compile.triton('triton_per_fused__native_batch_norm_legit_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[4, 16],
reduction_hint=ReductionHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused__native_batch_norm_legit_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 4, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused__native_batch_norm_legit_0(in_ptr0, out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 4
rnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex % 4
r2 = (rindex // 4)
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + (4*x0) + (16*r2)), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = 16.0
tmp18 = tmp16 / tmp17
tmp19 = 0.001
tmp20 = tmp18 + tmp19
tmp21 = libdevice.rsqrt(tmp20)
tl.store(out_ptr2 + (x0), tmp21, xmask)
tl.store(out_ptr0 + (x0), tmp10, xmask)
tl.store(out_ptr1 + (x0), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/dy/cdyw65rioafwiakcljlghmqa24g7jnrjibhdgma4b5qyneqwjh77.py
# Topologically Sorted Source Nodes: [x_1, x_2], Original ATen: [aten._native_batch_norm_legit, aten.relu, aten.threshold_backward]
# Source node to ATen node mapping:
# x_1 => add, add_1, mul, mul_1, rsqrt, sub, var_mean
# x_2 => relu
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%view_1, [0, 2]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 0.001), kwargs = {})
# %rsqrt : [num_users=2] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_1, %getitem_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %unsqueeze), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %unsqueeze_1), kwargs = {})
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%add_1,), kwargs = {})
# %le : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu, 0), kwargs = {})
triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1 = async_compile.triton('triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*i1', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 4) % 4
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr3 + (x1), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr4 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = 16.0
tmp5 = tmp3 / tmp4
tmp6 = 0.001
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp2 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tmp14 = tl.full([1], 0, tl.int32)
tmp15 = triton_helpers.maximum(tmp14, tmp13)
tmp16 = 0.0
tmp17 = tmp15 <= tmp16
tl.store(out_ptr0 + (x3), tmp15, xmask)
tl.store(out_ptr1 + (x3), tmp17, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, ), (1, ))
assert_size_stride(primals_5, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (16, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((1, 4, 1), (4, 1, 4), torch.float32)
buf2 = empty_strided_cuda((1, 4, 1), (4, 1, 4), torch.float32)
buf4 = empty_strided_cuda((1, 4, 1), (4, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten._native_batch_norm_legit]
stream0 = get_raw_stream(0)
triton_per_fused__native_batch_norm_legit_0.run(buf0, buf1, buf2, buf4, 4, 16, grid=grid(4), stream=stream0)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.bool)
# Topologically Sorted Source Nodes: [x_1, x_2], Original ATen: [aten._native_batch_norm_legit, aten.relu, aten.threshold_backward]
triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1.run(buf0, buf1, buf2, primals_4, primals_5, buf5, buf6, 64, grid=grid(64), stream=stream0)
del buf2
del primals_5
return (buf5, primals_4, reinterpret_tensor(primals_3, (16, 4), (4, 1), 0), buf0, reinterpret_tensor(buf4, (4, ), (1, ), 0), buf6, reinterpret_tensor(buf1, (1, 4, 1), (4, 1, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
from scipy.stats import truncnorm
def truncated_normal_(tensor, mean=0.0, std=1.0):
values = truncnorm.rvs(-2, 2, size=tensor.shape)
values = mean + std * values
tensor.copy_(torch.from_numpy(values))
return tensor
def fc_init_(module):
if hasattr(module, 'weight') and module.weight is not None:
truncated_normal_(module.weight.data, mean=0.0, std=0.01)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.constant_(module.bias.data, 0.0)
return module
class LinearBlock(nn.Module):
def __init__(self, input_size, output_size):
super(LinearBlock, self).__init__()
self.relu = nn.ReLU()
self.normalize = nn.BatchNorm1d(output_size, affine=True, momentum=
0.999, eps=0.001, track_running_stats=False)
self.linear = nn.Linear(input_size, output_size)
fc_init_(self.linear)
def forward(self, x):
x = self.linear(x)
x = self.normalize(x)
x = self.relu(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'input_size': 4, 'output_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
from scipy.stats import truncnorm
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused__native_batch_norm_legit_0(in_ptr0, out_ptr0, out_ptr1,
out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex % 4
r2 = rindex // 4
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 4 * x0 + 16 * r2), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = 16.0
tmp18 = tmp16 / tmp17
tmp19 = 0.001
tmp20 = tmp18 + tmp19
tmp21 = libdevice.rsqrt(tmp20)
tl.store(out_ptr2 + x0, tmp21, xmask)
tl.store(out_ptr0 + x0, tmp10, xmask)
tl.store(out_ptr1 + x0, tmp16, xmask)
@triton.jit
def triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1(in_ptr0
, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, out_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr4 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = 16.0
tmp5 = tmp3 / tmp4
tmp6 = 0.001
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp2 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tmp14 = tl.full([1], 0, tl.int32)
tmp15 = triton_helpers.maximum(tmp14, tmp13)
tmp16 = 0.0
tmp17 = tmp15 <= tmp16
tl.store(out_ptr0 + x3, tmp15, xmask)
tl.store(out_ptr1 + x3, tmp17, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (16,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((1, 4, 1), (4, 1, 4), torch.float32)
buf2 = empty_strided_cuda((1, 4, 1), (4, 1, 4), torch.float32)
buf4 = empty_strided_cuda((1, 4, 1), (4, 1, 1), torch.float32)
get_raw_stream(0)
triton_per_fused__native_batch_norm_legit_0[grid(4)](buf0, buf1,
buf2, buf4, 4, 16, XBLOCK=1, num_warps=2, num_stages=1)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.bool)
triton_poi_fused__native_batch_norm_legit_relu_threshold_backward_1[
grid(64)](buf0, buf1, buf2, primals_4, primals_5, buf5, buf6,
64, XBLOCK=64, num_warps=1, num_stages=1)
del buf2
del primals_5
return buf5, primals_4, reinterpret_tensor(primals_3, (16, 4), (4, 1), 0
), buf0, reinterpret_tensor(buf4, (4,), (1,), 0
), buf6, reinterpret_tensor(buf1, (1, 4, 1), (4, 1, 1), 0)
def truncated_normal_(tensor, mean=0.0, std=1.0):
values = truncnorm.rvs(-2, 2, size=tensor.shape)
values = mean + std * values
tensor.copy_(torch.from_numpy(values))
return tensor
def fc_init_(module):
if hasattr(module, 'weight') and module.weight is not None:
truncated_normal_(module.weight.data, mean=0.0, std=0.01)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.constant_(module.bias.data, 0.0)
return module
class LinearBlockNew(nn.Module):
def __init__(self, input_size, output_size):
super(LinearBlockNew, self).__init__()
self.relu = nn.ReLU()
self.normalize = nn.BatchNorm1d(output_size, affine=True, momentum=
0.999, eps=0.001, track_running_stats=False)
self.linear = nn.Linear(input_size, output_size)
fc_init_(self.linear)
def forward(self, input_0):
primals_2 = self.normalize.weight
primals_4 = self.normalize.bias
primals_1 = self.linear.weight
primals_5 = self.linear.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| JasonMa2016/learn2learn | LinearBlock | false | 870 | [
"MIT"
] | 0 | 502e1ea6db64481d7464fdda4d4d0be9b0f1089a | https://github.com/JasonMa2016/learn2learn/tree/502e1ea6db64481d7464fdda4d4d0be9b0f1089a | import torch
from torch import nn
from scipy.stats import truncnorm
def truncated_normal_(tensor, mean=0.0, std=1.0):
values = truncnorm.rvs(-2, 2, size=tensor.shape)
values = mean + std * values
tensor.copy_(torch.from_numpy(values))
return tensor
def fc_init_(module):
if hasattr(module, 'weight') and module.weight is not None:
truncated_normal_(module.weight.data, mean=0.0, std=0.01)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.constant_(module.bias.data, 0.0)
return module
class Model(nn.Module):
def __init__(self, input_size, output_size):
super().__init__()
self.relu = nn.ReLU()
self.normalize = nn.BatchNorm1d(output_size, affine=True, momentum=
0.999, eps=0.001, track_running_stats=False)
self.linear = nn.Linear(input_size, output_size)
fc_init_(self.linear)
def forward(self, x):
x = self.linear(x)
x = self.normalize(x)
x = self.relu(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
ClipLoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/zq/czqmztr2vwx7nz3nkq3ph7yqafp67nxh2us7hz55luebu7lh6ogw.py
# Topologically Sorted Source Nodes: [mul, mul_1], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# mul => mul
# mul_1 => mul_1
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg1_1, %arg0_1), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg1_1, %arg2_1), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask)
tmp3 = tl.load(in_ptr2 + (x0), xmask)
tmp2 = tmp0 * tmp1
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + (x0), tmp2, xmask)
tl.store(out_ptr1 + (x0), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ul/culvxc5xcnacfjypzxghwcyc2445sqsz25ci4rib6axjxs3fv3so.py
# Topologically Sorted Source Nodes: [cross_entropy], Original ATen: [aten._log_softmax]
# Source node to ATen node mapping:
# cross_entropy => amax, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%mm, [1], True), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mm, %amax), kwargs = {})
triton_poi_fused__log_softmax_1 = async_compile.triton('triton_poi_fused__log_softmax_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__log_softmax_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__log_softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/uk/cuk4x7whqaqekks743sb2fwjzsgutcays75mxn552glq6cnc7cwy.py
# Topologically Sorted Source Nodes: [labels, cross_entropy, cross_entropy_1, add, total_loss], Original ATen: [aten.arange, aten.nll_loss_forward, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# cross_entropy => convert_element_type, div, full_default_1, ne_1, ne_2, neg, sum_2, sum_3, where_1
# cross_entropy_1 => convert_element_type_1, div_1, full_default_3, ne_4, ne_5, neg_1, sum_5, sum_6, where_3
# labels => iota
# total_loss => div_2
# Graph fragment:
# %iota : [num_users=8] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %ne_1 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%iota, -100), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%squeeze,), kwargs = {})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_1 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%ne_1, %neg, %full_default_1), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%where_1,), kwargs = {})
# %ne_2 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%iota, -100), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%ne_2,), kwargs = {})
# %convert_element_type : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%sum_2, torch.float32), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_3, %convert_element_type), kwargs = {})
# %ne_4 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%iota, -100), kwargs = {})
# %neg_1 : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%squeeze_1,), kwargs = {})
# %full_default_3 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_3 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%ne_4, %neg_1, %full_default_3), kwargs = {})
# %sum_6 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%where_3,), kwargs = {})
# %ne_5 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%iota, -100), kwargs = {})
# %sum_5 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%ne_5,), kwargs = {})
# %convert_element_type_1 : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%sum_5, torch.float32), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_6, %convert_element_type_1), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %div_1), kwargs = {})
# %div_2 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%add, 2), kwargs = {})
triton_per_fused_add_arange_div_nll_loss_forward_2 = async_compile.triton('triton_per_fused_add_arange_div_nll_loss_forward_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 4],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_arange_div_nll_loss_forward_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 4, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_arange_div_nll_loss_forward_2(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 4
RBLOCK: tl.constexpr = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp6 = tl.load(in_ptr0 + (4*r0), None, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + (4*r0)), None, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (2 + (4*r0)), None, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr0 + (3 + (4*r0)), None, eviction_policy='evict_last')
tmp26 = tl.load(in_ptr1 + (4*r0), None, eviction_policy='evict_last')
tmp28 = tl.load(in_ptr1 + (1 + (4*r0)), None, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr1 + (2 + (4*r0)), None, eviction_policy='evict_last')
tmp34 = tl.load(in_ptr1 + (3 + (4*r0)), None, eviction_policy='evict_last')
tmp0 = r0
tmp1 = tl.full([1, 1], -100, tl.int64)
tmp2 = tmp0 != tmp1
tmp3 = tl.full([1, 1], 0, tl.int64)
tmp4 = tl.where(tmp2, tmp0, tmp3)
tmp5 = tl.load(in_ptr0 + (tmp4 + (4*r0)), None, eviction_policy='evict_last')
tmp7 = tl_math.exp(tmp6)
tmp9 = tl_math.exp(tmp8)
tmp10 = tmp7 + tmp9
tmp12 = tl_math.exp(tmp11)
tmp13 = tmp10 + tmp12
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp13 + tmp15
tmp17 = tl_math.log(tmp16)
tmp18 = tmp5 - tmp17
tmp19 = -tmp18
tmp20 = 0.0
tmp21 = tl.where(tmp2, tmp19, tmp20)
tmp22 = tl.broadcast_to(tmp21, [XBLOCK, RBLOCK])
tmp24 = tl.sum(tmp22, 1)[:, None]
tmp25 = tl.load(in_ptr1 + (tmp4 + (4*r0)), None, eviction_policy='evict_last')
tmp27 = tl_math.exp(tmp26)
tmp29 = tl_math.exp(tmp28)
tmp30 = tmp27 + tmp29
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tmp37 = tl_math.log(tmp36)
tmp38 = tmp25 - tmp37
tmp39 = -tmp38
tmp40 = tl.where(tmp2, tmp39, tmp20)
tmp41 = tl.broadcast_to(tmp40, [XBLOCK, RBLOCK])
tmp43 = tl.sum(tmp41, 1)[:, None]
tmp44 = tmp2.to(tl.int64)
tmp45 = tl.broadcast_to(tmp44, [XBLOCK, RBLOCK])
tmp47 = tl.sum(tmp45, 1)[:, None]
tmp48 = tmp47.to(tl.float32)
tmp49 = tmp24 / tmp48
tmp50 = tmp43 / tmp48
tmp51 = tmp49 + tmp50
tmp52 = 0.5
tmp53 = tmp51 * tmp52
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp53, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
assert_size_stride(arg2_1, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul, mul_1], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(arg1_1, arg0_1, arg2_1, buf0, buf5, 16, grid=grid(16), stream=stream0)
del arg1_1
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul, logits_per_image], Original ATen: [aten.mul, aten.mm]
extern_kernels.mm(buf0, reinterpret_tensor(arg2_1, (4, 4), (1, 4), 0), out=buf1)
del arg2_1
buf2 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [cross_entropy], Original ATen: [aten._log_softmax]
triton_poi_fused__log_softmax_1.run(buf1, buf2, 16, grid=grid(16), stream=stream0)
buf6 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [mul_1, logits_per_text], Original ATen: [aten.mul, aten.mm]
extern_kernels.mm(buf5, reinterpret_tensor(arg0_1, (4, 4), (1, 4), 0), out=buf6)
del arg0_1
buf7 = buf5; del buf5 # reuse
# Topologically Sorted Source Nodes: [cross_entropy_1], Original ATen: [aten._log_softmax]
triton_poi_fused__log_softmax_1.run(buf6, buf7, 16, grid=grid(16), stream=stream0)
del buf6
buf3 = empty_strided_cuda((), (), torch.float32)
buf10 = buf3; del buf3 # reuse
# Topologically Sorted Source Nodes: [labels, cross_entropy, cross_entropy_1, add, total_loss], Original ATen: [aten.arange, aten.nll_loss_forward, aten.add, aten.div]
triton_per_fused_add_arange_div_nll_loss_forward_2.run(buf10, buf2, buf7, 1, 4, grid=grid(1), stream=stream0)
del buf2
del buf7
return (buf10, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
arg2_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1, arg2_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
from torch.nn import functional as F
from torch import distributed as dist
import torch.distributed.nn
def gather_features(image_features, text_features, local_loss=False,
gather_with_grad=False, rank=0, world_size=1, use_horovod=False):
if use_horovod:
assert hvd is not None, 'Please install horovod'
if gather_with_grad:
all_image_features = hvd.allgather(image_features)
all_text_features = hvd.allgather(text_features)
else:
with torch.no_grad():
all_image_features = hvd.allgather(image_features)
all_text_features = hvd.allgather(text_features)
if not local_loss:
gathered_image_features = list(all_image_features.chunk(
world_size, dim=0))
gathered_text_features = list(all_text_features.chunk(
world_size, dim=0))
gathered_image_features[rank] = image_features
gathered_text_features[rank] = text_features
all_image_features = torch.cat(gathered_image_features, dim=0)
all_text_features = torch.cat(gathered_text_features, dim=0)
elif gather_with_grad:
all_image_features = torch.cat(torch.distributed.nn.all_gather(
image_features), dim=0)
all_text_features = torch.cat(torch.distributed.nn.all_gather(
text_features), dim=0)
else:
gathered_image_features = [torch.zeros_like(image_features) for _ in
range(world_size)]
gathered_text_features = [torch.zeros_like(text_features) for _ in
range(world_size)]
dist.all_gather(gathered_image_features, image_features)
dist.all_gather(gathered_text_features, text_features)
if not local_loss:
gathered_image_features[rank] = image_features
gathered_text_features[rank] = text_features
all_image_features = torch.cat(gathered_image_features, dim=0)
all_text_features = torch.cat(gathered_text_features, dim=0)
return all_image_features, all_text_features
class ClipLoss(nn.Module):
def __init__(self, local_loss=False, gather_with_grad=False,
cache_labels=False, rank=0, world_size=1, use_horovod=False):
super().__init__()
self.local_loss = local_loss
self.gather_with_grad = gather_with_grad
self.cache_labels = cache_labels
self.rank = rank
self.world_size = world_size
self.use_horovod = use_horovod
self.prev_num_logits = 0
self.labels = {}
def forward(self, image_features, text_features, logit_scale):
device = image_features.device
if self.world_size > 1:
all_image_features, all_text_features = gather_features(
image_features, text_features, self.local_loss, self.
gather_with_grad, self.rank, self.world_size, self.use_horovod)
if self.local_loss:
logits_per_image = (logit_scale * image_features @
all_text_features.T)
logits_per_text = (logit_scale * text_features @
all_image_features.T)
else:
logits_per_image = (logit_scale * all_image_features @
all_text_features.T)
logits_per_text = logits_per_image.T
else:
logits_per_image = logit_scale * image_features @ text_features.T
logits_per_text = logit_scale * text_features @ image_features.T
num_logits = logits_per_image.shape[0]
if self.prev_num_logits != num_logits or device not in self.labels:
labels = torch.arange(num_logits, device=device, dtype=torch.long)
if self.world_size > 1 and self.local_loss:
labels = labels + num_logits * self.rank
if self.cache_labels:
self.labels[device] = labels
self.prev_num_logits = num_logits
else:
labels = self.labels[device]
total_loss = (F.cross_entropy(logits_per_image, labels) + F.
cross_entropy(logits_per_text, labels)) / 2
return total_loss
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
from torch import distributed as dist
import torch.distributed.nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask)
tmp3 = tl.load(in_ptr2 + x0, xmask)
tmp2 = tmp0 * tmp1
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + x0, tmp2, xmask)
tl.store(out_ptr1 + x0, tmp4, xmask)
@triton.jit
def triton_poi_fused__log_softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_per_fused_add_arange_div_nll_loss_forward_2(in_out_ptr0, in_ptr0,
in_ptr1, xnumel, rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 4
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp6 = tl.load(in_ptr0 + 4 * r0, None, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + 4 * r0), None, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (2 + 4 * r0), None, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr0 + (3 + 4 * r0), None, eviction_policy='evict_last')
tmp26 = tl.load(in_ptr1 + 4 * r0, None, eviction_policy='evict_last')
tmp28 = tl.load(in_ptr1 + (1 + 4 * r0), None, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr1 + (2 + 4 * r0), None, eviction_policy='evict_last')
tmp34 = tl.load(in_ptr1 + (3 + 4 * r0), None, eviction_policy='evict_last')
tmp0 = r0
tmp1 = tl.full([1, 1], -100, tl.int64)
tmp2 = tmp0 != tmp1
tmp3 = tl.full([1, 1], 0, tl.int64)
tmp4 = tl.where(tmp2, tmp0, tmp3)
tmp5 = tl.load(in_ptr0 + (tmp4 + 4 * r0), None, eviction_policy=
'evict_last')
tmp7 = tl_math.exp(tmp6)
tmp9 = tl_math.exp(tmp8)
tmp10 = tmp7 + tmp9
tmp12 = tl_math.exp(tmp11)
tmp13 = tmp10 + tmp12
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp13 + tmp15
tmp17 = tl_math.log(tmp16)
tmp18 = tmp5 - tmp17
tmp19 = -tmp18
tmp20 = 0.0
tmp21 = tl.where(tmp2, tmp19, tmp20)
tmp22 = tl.broadcast_to(tmp21, [XBLOCK, RBLOCK])
tmp24 = tl.sum(tmp22, 1)[:, None]
tmp25 = tl.load(in_ptr1 + (tmp4 + 4 * r0), None, eviction_policy=
'evict_last')
tmp27 = tl_math.exp(tmp26)
tmp29 = tl_math.exp(tmp28)
tmp30 = tmp27 + tmp29
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tmp37 = tl_math.log(tmp36)
tmp38 = tmp25 - tmp37
tmp39 = -tmp38
tmp40 = tl.where(tmp2, tmp39, tmp20)
tmp41 = tl.broadcast_to(tmp40, [XBLOCK, RBLOCK])
tmp43 = tl.sum(tmp41, 1)[:, None]
tmp44 = tmp2.to(tl.int64)
tmp45 = tl.broadcast_to(tmp44, [XBLOCK, RBLOCK])
tmp47 = tl.sum(tmp45, 1)[:, None]
tmp48 = tmp47.to(tl.float32)
tmp49 = tmp24 / tmp48
tmp50 = tmp43 / tmp48
tmp51 = tmp49 + tmp50
tmp52 = 0.5
tmp53 = tmp51 * tmp52
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp53, None)
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
assert_size_stride(arg2_1, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(16)](arg1_1, arg0_1, arg2_1, buf0, buf5,
16, XBLOCK=16, num_warps=1, num_stages=1)
del arg1_1
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf0, reinterpret_tensor(arg2_1, (4, 4), (1, 4),
0), out=buf1)
del arg2_1
buf2 = buf0
del buf0
triton_poi_fused__log_softmax_1[grid(16)](buf1, buf2, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf6 = buf1
del buf1
extern_kernels.mm(buf5, reinterpret_tensor(arg0_1, (4, 4), (1, 4),
0), out=buf6)
del arg0_1
buf7 = buf5
del buf5
triton_poi_fused__log_softmax_1[grid(16)](buf6, buf7, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del buf6
buf3 = empty_strided_cuda((), (), torch.float32)
buf10 = buf3
del buf3
triton_per_fused_add_arange_div_nll_loss_forward_2[grid(1)](buf10,
buf2, buf7, 1, 4, XBLOCK=1, num_warps=2, num_stages=1)
del buf2
del buf7
return buf10,
def gather_features(image_features, text_features, local_loss=False,
gather_with_grad=False, rank=0, world_size=1, use_horovod=False):
if use_horovod:
assert hvd is not None, 'Please install horovod'
if gather_with_grad:
all_image_features = hvd.allgather(image_features)
all_text_features = hvd.allgather(text_features)
else:
with torch.no_grad():
all_image_features = hvd.allgather(image_features)
all_text_features = hvd.allgather(text_features)
if not local_loss:
gathered_image_features = list(all_image_features.chunk(
world_size, dim=0))
gathered_text_features = list(all_text_features.chunk(
world_size, dim=0))
gathered_image_features[rank] = image_features
gathered_text_features[rank] = text_features
all_image_features = torch.cat(gathered_image_features, dim=0)
all_text_features = torch.cat(gathered_text_features, dim=0)
elif gather_with_grad:
all_image_features = torch.cat(torch.distributed.nn.all_gather(
image_features), dim=0)
all_text_features = torch.cat(torch.distributed.nn.all_gather(
text_features), dim=0)
else:
gathered_image_features = [torch.zeros_like(image_features) for _ in
range(world_size)]
gathered_text_features = [torch.zeros_like(text_features) for _ in
range(world_size)]
dist.all_gather(gathered_image_features, image_features)
dist.all_gather(gathered_text_features, text_features)
if not local_loss:
gathered_image_features[rank] = image_features
gathered_text_features[rank] = text_features
all_image_features = torch.cat(gathered_image_features, dim=0)
all_text_features = torch.cat(gathered_text_features, dim=0)
return all_image_features, all_text_features
class ClipLossNew(nn.Module):
def __init__(self, local_loss=False, gather_with_grad=False,
cache_labels=False, rank=0, world_size=1, use_horovod=False):
super().__init__()
self.local_loss = local_loss
self.gather_with_grad = gather_with_grad
self.cache_labels = cache_labels
self.rank = rank
self.world_size = world_size
self.use_horovod = use_horovod
self.prev_num_logits = 0
self.labels = {}
def forward(self, input_0, input_1, input_2):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
output = call([arg0_1, arg1_1, arg2_1])
return output[0]
| NYU-DICE-Lab/open_clip | ClipLoss | false | 871 | [
"MIT"
] | 0 | fd71804b503135fb1c7cc8de3a0d6599741c8ed9 | https://github.com/NYU-DICE-Lab/open_clip/tree/fd71804b503135fb1c7cc8de3a0d6599741c8ed9 | import torch
from torch import nn
from torch.nn import functional as F
from torch import distributed as dist
import torch.distributed.nn
def gather_features(image_features, text_features, local_loss=False,
gather_with_grad=False, rank=0, world_size=1, use_horovod=False):
if use_horovod:
assert hvd is not None, 'Please install horovod'
if gather_with_grad:
all_image_features = hvd.allgather(image_features)
all_text_features = hvd.allgather(text_features)
else:
with torch.no_grad():
all_image_features = hvd.allgather(image_features)
all_text_features = hvd.allgather(text_features)
if not local_loss:
gathered_image_features = list(all_image_features.chunk(
world_size, dim=0))
gathered_text_features = list(all_text_features.chunk(
world_size, dim=0))
gathered_image_features[rank] = image_features
gathered_text_features[rank] = text_features
all_image_features = torch.cat(gathered_image_features, dim=0)
all_text_features = torch.cat(gathered_text_features, dim=0)
elif gather_with_grad:
all_image_features = torch.cat(torch.distributed.nn.all_gather(
image_features), dim=0)
all_text_features = torch.cat(torch.distributed.nn.all_gather(
text_features), dim=0)
else:
gathered_image_features = [torch.zeros_like(image_features) for _ in
range(world_size)]
gathered_text_features = [torch.zeros_like(text_features) for _ in
range(world_size)]
dist.all_gather(gathered_image_features, image_features)
dist.all_gather(gathered_text_features, text_features)
if not local_loss:
gathered_image_features[rank] = image_features
gathered_text_features[rank] = text_features
all_image_features = torch.cat(gathered_image_features, dim=0)
all_text_features = torch.cat(gathered_text_features, dim=0)
return all_image_features, all_text_features
class Model(nn.Module):
def __init__(self, local_loss=False, gather_with_grad=False,
cache_labels=False, rank=0, world_size=1, use_horovod=False):
super().__init__()
self.local_loss = local_loss
self.gather_with_grad = gather_with_grad
self.cache_labels = cache_labels
self.rank = rank
self.world_size = world_size
self.use_horovod = use_horovod
self.prev_num_logits = 0
self.labels = {}
def forward(self, image_features, text_features, logit_scale):
device = image_features.device
if self.world_size > 1:
all_image_features, all_text_features = gather_features(
image_features, text_features, self.local_loss, self.
gather_with_grad, self.rank, self.world_size, self.use_horovod)
if self.local_loss:
logits_per_image = (logit_scale * image_features @
all_text_features.T)
logits_per_text = (logit_scale * text_features @
all_image_features.T)
else:
logits_per_image = (logit_scale * all_image_features @
all_text_features.T)
logits_per_text = logits_per_image.T
else:
logits_per_image = logit_scale * image_features @ text_features.T
logits_per_text = logit_scale * text_features @ image_features.T
num_logits = logits_per_image.shape[0]
if self.prev_num_logits != num_logits or device not in self.labels:
labels = torch.arange(num_logits, device=device, dtype=torch.long)
if self.world_size > 1 and self.local_loss:
labels = labels + num_logits * self.rank
if self.cache_labels:
self.labels[device] = labels
self.prev_
# ... truncated (>4000 chars) for memory efficiency |
DenseBlock | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/pz/cpzxz6lkuih5yejx2jgx2k2tt7datj3a4ghmi4d5empkhgoz5dif.py
# Topologically Sorted Source Nodes: [conv1d], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# conv1d => convolution
# Graph fragment:
# %convolution : [num_users=2] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1], [1], [1], False, [0], 1), kwargs = {})
triton_poi_fused_convolution_0 = async_compile.triton('triton_poi_fused_convolution_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 5) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/u6/cu6p4t7q5wtwgs5cpdhj7mpibqv5emb5qdj4v6x2yo45mtrk3dke.py
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# out => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_3, %mul], 1), kwargs = {})
triton_poi_fused_cat_1 = async_compile.triton('triton_poi_fused_cat_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4) % 8
x0 = xindex % 4
x2 = (xindex // 32)
x3 = xindex
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + (4*x1) + (16*x2)), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + (x0 + (5*((-4) + x1)) + (20*x2)), tmp6 & xmask, other=0.0)
tmp10 = libdevice.tanh(tmp9)
tmp11 = tl.load(in_ptr2 + (x0 + (5*((-4) + x1)) + (20*x2)), tmp6 & xmask, other=0.0)
tmp12 = tl.sigmoid(tmp11)
tmp13 = tmp10 * tmp12
tmp14 = tl.full(tmp13.shape, 0.0, tmp13.dtype)
tmp15 = tl.where(tmp6, tmp13, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + (x3), tmp16, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 2), (8, 2, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4, 2), (8, 2, 1))
assert_size_stride(primals_5, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [conv1d], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,), padding=(1,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 5), (20, 5, 1))
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [conv1d], Original ATen: [aten.convolution]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_0.run(buf1, primals_2, 80, grid=grid(80), stream=stream0)
del primals_2
# Topologically Sorted Source Nodes: [conv1d_1], Original ATen: [aten.convolution]
buf2 = extern_kernels.convolution(primals_3, primals_4, stride=(1,), padding=(1,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf2, (4, 4, 5), (20, 5, 1))
buf3 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [conv1d_1], Original ATen: [aten.convolution]
triton_poi_fused_convolution_0.run(buf3, primals_5, 80, grid=grid(80), stream=stream0)
del primals_5
buf4 = empty_strided_cuda((4, 8, 4), (32, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.cat]
triton_poi_fused_cat_1.run(primals_3, buf1, buf3, buf4, 128, grid=grid(128), stream=stream0)
return (buf4, primals_1, primals_3, primals_4, buf1, buf3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 2), (8, 2, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 4, 2), (8, 2, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
class CausalConv1d(nn.Module):
"""A 1D causal convolution layer.
Input: (B, D_in, T), where B is the minibatch size, D_in is the number of
dimensions per step, and T is the number of steps.
Output: (B, D_out, T), where B is the minibatch size, D_out is the number
of dimensions in the output, and T is the number of steps.
Arguments:
in_channels (int): number of input channels
out_channels (int): number of output channels
"""
def __init__(self, in_channels, out_channels, dilation=1):
super(CausalConv1d, self).__init__()
self.padding = dilation
self.causal_conv = nn.Conv1d(in_channels, out_channels, 2, padding=
self.padding, dilation=dilation)
def forward(self, minibatch):
return self.causal_conv(minibatch)[:, :, :-self.padding]
class DenseBlock(nn.Module):
"""Two parallel 1D causal convolution layers w/tanh and sigmoid activations
Input: (B, D_in, T), where B is the minibatch size, D_in is the number of
dimensions of the input, and T is the number of steps.
Output: (B, D_in+F, T), where where `B` is the minibatch size, `D_in` is the
number of dimensions of the input, `F` is the number of filters, and `T`
is the length of the input sequence.
Arguments:
in_channels (int): number of input channels
filters (int): number of filters per channel
"""
def __init__(self, in_channels, filters, dilation=1):
super(DenseBlock, self).__init__()
self.causal_conv1 = CausalConv1d(in_channels, filters, dilation=
dilation)
self.causal_conv2 = CausalConv1d(in_channels, filters, dilation=
dilation)
def forward(self, minibatch):
tanh = F.tanh(self.causal_conv1(minibatch))
sig = F.sigmoid(self.causal_conv2(minibatch))
out = torch.cat([minibatch, tanh * sig], dim=1)
return out
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'filters': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 5 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 8
x0 = xindex % 4
x2 = xindex // 32
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 4 * x1 + 16 * x2), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (x0 + 5 * (-4 + x1) + 20 * x2), tmp6 & xmask,
other=0.0)
tmp10 = libdevice.tanh(tmp9)
tmp11 = tl.load(in_ptr2 + (x0 + 5 * (-4 + x1) + 20 * x2), tmp6 & xmask,
other=0.0)
tmp12 = tl.sigmoid(tmp11)
tmp13 = tmp10 * tmp12
tmp14 = tl.full(tmp13.shape, 0.0, tmp13.dtype)
tmp15 = tl.where(tmp6, tmp13, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + x3, tmp16, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 2), (8, 2, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4, 2), (8, 2, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,),
padding=(1,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 5), (20, 5, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(80)](buf1, primals_2, 80,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf2 = extern_kernels.convolution(primals_3, primals_4, stride=(1,),
padding=(1,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf2, (4, 4, 5), (20, 5, 1))
buf3 = buf2
del buf2
triton_poi_fused_convolution_0[grid(80)](buf3, primals_5, 80,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_5
buf4 = empty_strided_cuda((4, 8, 4), (32, 4, 1), torch.float32)
triton_poi_fused_cat_1[grid(128)](primals_3, buf1, buf3, buf4, 128,
XBLOCK=128, num_warps=4, num_stages=1)
return buf4, primals_1, primals_3, primals_4, buf1, buf3
class CausalConv1d(nn.Module):
"""A 1D causal convolution layer.
Input: (B, D_in, T), where B is the minibatch size, D_in is the number of
dimensions per step, and T is the number of steps.
Output: (B, D_out, T), where B is the minibatch size, D_out is the number
of dimensions in the output, and T is the number of steps.
Arguments:
in_channels (int): number of input channels
out_channels (int): number of output channels
"""
def __init__(self, in_channels, out_channels, dilation=1):
super(CausalConv1d, self).__init__()
self.padding = dilation
self.causal_conv = nn.Conv1d(in_channels, out_channels, 2, padding=
self.padding, dilation=dilation)
def forward(self, minibatch):
return self.causal_conv(minibatch)[:, :, :-self.padding]
class DenseBlockNew(nn.Module):
"""Two parallel 1D causal convolution layers w/tanh and sigmoid activations
Input: (B, D_in, T), where B is the minibatch size, D_in is the number of
dimensions of the input, and T is the number of steps.
Output: (B, D_in+F, T), where where `B` is the minibatch size, `D_in` is the
number of dimensions of the input, `F` is the number of filters, and `T`
is the length of the input sequence.
Arguments:
in_channels (int): number of input channels
filters (int): number of filters per channel
"""
def __init__(self, in_channels, filters, dilation=1):
super(DenseBlockNew, self).__init__()
self.causal_conv1 = CausalConv1d(in_channels, filters, dilation=
dilation)
self.causal_conv2 = CausalConv1d(in_channels, filters, dilation=
dilation)
def forward(self, input_0):
primals_1 = self.causal_conv1.causal_conv.weight
primals_2 = self.causal_conv1.causal_conv.bias
primals_4 = self.causal_conv2.causal_conv.weight
primals_5 = self.causal_conv2.causal_conv.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| NagisaZj/oyster | DenseBlock | false | 872 | [
"MIT"
] | 0 | 069a510fe63bb29ecd9871e0e189e58b03c8cad9 | https://github.com/NagisaZj/oyster/tree/069a510fe63bb29ecd9871e0e189e58b03c8cad9 | import torch
import torch.nn as nn
import torch.nn.functional as F
class CausalConv1d(nn.Module):
"""A 1D causal convolution layer.
Input: (B, D_in, T), where B is the minibatch size, D_in is the number of
dimensions per step, and T is the number of steps.
Output: (B, D_out, T), where B is the minibatch size, D_out is the number
of dimensions in the output, and T is the number of steps.
Arguments:
in_channels (int): number of input channels
out_channels (int): number of output channels
"""
def __init__(self, in_channels, out_channels, dilation=1):
super().__init__()
self.padding = dilation
self.causal_conv = nn.Conv1d(in_channels, out_channels, 2, padding=
self.padding, dilation=dilation)
def forward(self, minibatch):
return self.causal_conv(minibatch)[:, :, :-self.padding]
class Model(nn.Module):
"""Two parallel 1D causal convolution layers w/tanh and sigmoid activations
Input: (B, D_in, T), where B is the minibatch size, D_in is the number of
dimensions of the input, and T is the number of steps.
Output: (B, D_in+F, T), where where `B` is the minibatch size, `D_in` is the
number of dimensions of the input, `F` is the number of filters, and `T`
is the length of the input sequence.
Arguments:
in_channels (int): number of input channels
filters (int): number of filters per channel
"""
def __init__(self, in_channels, filters, dilation=1):
super().__init__()
self.causal_conv1 = CausalConv1d(in_channels, filters, dilation=
dilation)
self.causal_conv2 = CausalConv1d(in_channels, filters, dilation=
dilation)
def forward(self, minibatch):
tanh = F.tanh(self.causal_conv1(minibatch))
sig = F.sigmoid(self.causal_conv2(minibatch))
out = torch.cat([minibatch, tanh * sig], dim=1)
return out
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
distLinear | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/m6/cm645lheesrjji6wgkstt4nu675ugbbjruised3fke4juyuyosol.py
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
# Source node to ATen node mapping:
# _weight_norm => pow_3, pow_4, sum_2
# Graph fragment:
# %pow_3 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_3, 2), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_3, [1], True), kwargs = {})
# %pow_4 : [num_users=2] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_2, 0.5), kwargs = {})
triton_poi_fused__weight_norm_interface_0 = async_compile.triton('triton_poi_fused__weight_norm_interface_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__weight_norm_interface_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__weight_norm_interface_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp1 = tmp0 * tmp0
tmp3 = tmp2 * tmp2
tmp4 = tmp1 + tmp3
tmp6 = tmp5 * tmp5
tmp7 = tmp4 + tmp6
tmp9 = tmp8 * tmp8
tmp10 = tmp7 + tmp9
tmp11 = libdevice.sqrt(tmp10)
tl.store(out_ptr0 + (x0), tmp11, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/dp/cdpmihjazxc2dpfye4tlkemiovtq5jgmt3cquzgrtbm3gn32us7u.py
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
# Source node to ATen node mapping:
# _weight_norm => div_1, mul
# Graph fragment:
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_2, %pow_4), kwargs = {})
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_3, %div_1), kwargs = {})
triton_poi_fused__weight_norm_interface_1 = async_compile.triton('triton_poi_fused__weight_norm_interface_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__weight_norm_interface_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__weight_norm_interface_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 / tmp2
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/v2/cv2bongkb2dgiiqtf7a3gfmvfar5me3uzhr5tfvzijs3i7yi2oub.py
# Topologically Sorted Source Nodes: [add, x_normalized], Original ATen: [aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# x_normalized => div
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%expand, 1e-05), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, %add), kwargs = {})
triton_poi_fused_add_div_2 = async_compile.triton('triton_poi_fused_add_div_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tl.store(out_ptr0 + (x3), tmp15, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/nq/cnqverg3qpl5ed2t36fkymquzuwlsmkpmbft4egnvgh24e2ct6wu.py
# Topologically Sorted Source Nodes: [scores], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# scores => mul_1
# Graph fragment:
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 2), kwargs = {})
triton_poi_fused_mul_3 = async_compile.triton('triton_poi_fused_mul_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_3(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
tmp1 = 2.0
tmp2 = tmp0 * tmp1
tl.store(in_out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 1), (1, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
stream0 = get_raw_stream(0)
triton_poi_fused__weight_norm_interface_0.run(primals_3, buf0, 4, grid=grid(4), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
triton_poi_fused__weight_norm_interface_1.run(primals_3, primals_2, buf0, buf1, 16, grid=grid(16), stream=stream0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [add, x_normalized], Original ATen: [aten.add, aten.div]
triton_poi_fused_add_div_2.run(primals_1, buf2, 256, grid=grid(256), stream=stream0)
del primals_1
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [cos_dist], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0), reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf3)
buf4 = reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf3 # reuse
# Topologically Sorted Source Nodes: [scores], Original ATen: [aten.mul]
triton_poi_fused_mul_3.run(buf4, 256, grid=grid(256), stream=stream0)
return (buf4, buf1, primals_2, primals_3, buf0, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 1), (1, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.utils.data
from torch.nn.utils.weight_norm import WeightNorm
import torch.nn.parallel
import torch.optim
class distLinear(nn.Module):
def __init__(self, indim, outdim):
super(distLinear, self).__init__()
self.L = nn.Linear(indim, outdim, bias=False)
self.class_wise_learnable_norm = True
if self.class_wise_learnable_norm:
WeightNorm.apply(self.L, 'weight', dim=0)
if outdim <= 200:
self.scale_factor = 2
else:
self.scale_factor = 10
def forward(self, x):
x_norm = torch.norm(x, p=2, dim=1).unsqueeze(1).expand_as(x)
x_normalized = x.div(x_norm + 1e-05)
if not self.class_wise_learnable_norm:
L_norm = torch.norm(self.L.weight.data, p=2, dim=1).unsqueeze(1
).expand_as(self.L.weight.data)
self.L.weight.data = self.L.weight.data.div(L_norm + 1e-05)
cos_dist = self.L(x_normalized)
scores = self.scale_factor * cos_dist
return scores
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'indim': 4, 'outdim': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.utils.data
from torch.nn.utils.weight_norm import WeightNorm
import torch.nn.parallel
import torch.optim
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__weight_norm_interface_0(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp1 = tmp0 * tmp0
tmp3 = tmp2 * tmp2
tmp4 = tmp1 + tmp3
tmp6 = tmp5 * tmp5
tmp7 = tmp4 + tmp6
tmp9 = tmp8 * tmp8
tmp10 = tmp7 + tmp9
tmp11 = libdevice.sqrt(tmp10)
tl.store(out_ptr0 + x0, tmp11, xmask)
@triton.jit
def triton_poi_fused__weight_norm_interface_1(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp3 = tmp1 / tmp2
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_add_div_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tl.store(out_ptr0 + x3, tmp15, xmask)
@triton.jit
def triton_poi_fused_mul_3(in_out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = 2.0
tmp2 = tmp0 * tmp1
tl.store(in_out_ptr0 + x0, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 1), (1, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__weight_norm_interface_0[grid(4)](primals_3, buf0,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused__weight_norm_interface_1[grid(16)](primals_3,
primals_2, buf0, buf1, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_div_2[grid(256)](primals_1, buf2, 256, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_1
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0),
reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf3)
buf4 = reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf3
triton_poi_fused_mul_3[grid(256)](buf4, 256, XBLOCK=256, num_warps=
4, num_stages=1)
return buf4, buf1, primals_2, primals_3, buf0, reinterpret_tensor(buf2,
(64, 4), (4, 1), 0)
class distLinearNew(nn.Module):
def __init__(self, indim, outdim):
super(distLinearNew, self).__init__()
self.L = nn.Linear(indim, outdim, bias=False)
self.class_wise_learnable_norm = True
if self.class_wise_learnable_norm:
WeightNorm.apply(self.L, 'weight', dim=0)
if outdim <= 200:
self.scale_factor = 2
else:
self.scale_factor = 10
def forward(self, input_0):
primals_2 = self.L.weight_g
primals_3 = self.L.weight_v
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| MuawizChaudhary/STARTUP | distLinear | false | 873 | [
"MIT"
] | 0 | 03f39b34a4ec232f132173b4a1e67ea04165e52b | https://github.com/MuawizChaudhary/STARTUP/tree/03f39b34a4ec232f132173b4a1e67ea04165e52b | import torch
import torch.nn as nn
import torch.utils.data
from torch.nn.utils.weight_norm import WeightNorm
import torch.nn.parallel
import torch.optim
class Model(nn.Module):
def __init__(self, indim, outdim):
super().__init__()
self.L = nn.Linear(indim, outdim, bias=False)
self.class_wise_learnable_norm = True
if self.class_wise_learnable_norm:
WeightNorm.apply(self.L, 'weight', dim=0)
if outdim <= 200:
self.scale_factor = 2
else:
self.scale_factor = 10
def forward(self, x):
x_norm = torch.norm(x, p=2, dim=1).unsqueeze(1).expand_as(x)
x_normalized = x.div(x_norm + 1e-05)
if not self.class_wise_learnable_norm:
L_norm = torch.norm(self.L.weight.data, p=2, dim=1).unsqueeze(1
).expand_as(self.L.weight.data)
self.L.weight.data = self.L.weight.data.div(L_norm + 1e-05)
cos_dist = self.L(x_normalized)
scores = self.scale_factor * cos_dist
return scores
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
BothContextGate | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nc/cnc6a3vkphurm472zdavmn3qnff4lmaezxs63jlllw2kks2e62a4.py
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# input_tensor => cat
# Graph fragment:
# %cat : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2, %primals_3], 1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = (xindex // 12)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tmp12 = tl.full([1], 12, tl.int64)
tmp13 = tmp0 < tmp12
tmp14 = tl.load(in_ptr2 + ((4*x1) + ((-8) + x0)), tmp11 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + (x2), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/as/castj5utepsqaf7nmxz7ydwohxvblspgu7zqqkf5geoow2xnjdtg.py
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat_1 => cat_1
# Graph fragment:
# %cat_1 : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2], 1), kwargs = {})
triton_poi_fused_cat_1 = async_compile.triton('triton_poi_fused_cat_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/42/c42q6xmrtx4mk6cvsm764oigvfmjedisa43isepa27ioqcfzfgtm.py
# Topologically Sorted Source Nodes: [z, sub, mul, mul_1, add, tanh], Original ATen: [aten.sigmoid, aten.rsub, aten.mul, aten.add, aten.tanh]
# Source node to ATen node mapping:
# add => add
# mul => mul
# mul_1 => mul_1
# sub => sub
# tanh => tanh
# z => sigmoid
# Graph fragment:
# %sigmoid : [num_users=2] = call_function[target=torch.ops.aten.sigmoid.default](args = (%addmm,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1.0, %sigmoid), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %addmm_2), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %addmm_1), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %mul_1), kwargs = {})
# %tanh : [num_users=1] = call_function[target=torch.ops.aten.tanh.default](args = (%add,), kwargs = {})
triton_poi_fused_add_mul_rsub_sigmoid_tanh_2 = async_compile.triton('triton_poi_fused_add_mul_rsub_sigmoid_tanh_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_rsub_sigmoid_tanh_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_rsub_sigmoid_tanh_2(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp4 = tl.load(in_ptr1 + (x0), xmask)
tmp6 = tl.load(in_ptr2 + (x0), xmask)
tmp1 = tl.sigmoid(tmp0)
tmp2 = 1.0
tmp3 = tmp2 - tmp1
tmp5 = tmp3 * tmp4
tmp7 = tmp1 * tmp6
tmp8 = tmp5 + tmp7
tmp9 = libdevice.tanh(tmp8)
tl.store(out_ptr0 + (x0), tmp9, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4, ), (1, ))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, ), (1, ))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, primals_2, primals_3, buf0, 48, grid=grid(48), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, buf0, reinterpret_tensor(primals_4, (12, 4), (1, 12), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [proj_source], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_7, primals_3, reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf2)
del primals_6
del primals_7
buf3 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
triton_poi_fused_cat_1.run(primals_1, primals_2, buf3, 32, grid=grid(32), stream=stream0)
del primals_1
del primals_2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [proj_target], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_9, buf3, reinterpret_tensor(primals_8, (8, 4), (1, 8), 0), alpha=1, beta=1, out=buf4)
del primals_8
del primals_9
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [z, sub, mul, mul_1, add, tanh], Original ATen: [aten.sigmoid, aten.rsub, aten.mul, aten.add, aten.tanh]
triton_poi_fused_add_mul_rsub_sigmoid_tanh_2.run(buf1, buf4, buf2, buf5, 16, grid=grid(16), stream=stream0)
return (buf5, primals_3, buf0, buf1, buf2, buf3, buf4, buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 12), (12, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class BothContextGate(nn.Module):
"""Apply the context gate to both contexts"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(BothContextGate, self).__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, prev_emb, dec_state, attn_state):
z, source, target = self.context_gate(prev_emb, dec_state, attn_state)
return self.tanh((1.0 - z) * target + z * source)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = xindex // 12
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tl.full([1], 12, tl.int64)
tmp14 = tl.load(in_ptr2 + (4 * x1 + (-8 + x0)), tmp11 & xmask,
eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + x2, tmp16, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_mul_rsub_sigmoid_tanh_2(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp4 = tl.load(in_ptr1 + x0, xmask)
tmp6 = tl.load(in_ptr2 + x0, xmask)
tmp1 = tl.sigmoid(tmp0)
tmp2 = 1.0
tmp3 = tmp2 - tmp1
tmp5 = tmp3 * tmp4
tmp7 = tmp1 * tmp6
tmp8 = tmp5 + tmp7
tmp9 = libdevice.tanh(tmp8)
tl.store(out_ptr0 + x0, tmp9, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(48)](primals_1, primals_2, primals_3,
buf0, 48, XBLOCK=64, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, buf0, reinterpret_tensor(primals_4,
(12, 4), (1, 12), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, primals_3, reinterpret_tensor(
primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf2)
del primals_6
del primals_7
buf3 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
triton_poi_fused_cat_1[grid(32)](primals_1, primals_2, buf3, 32,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_1
del primals_2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, buf3, reinterpret_tensor(primals_8,
(8, 4), (1, 8), 0), alpha=1, beta=1, out=buf4)
del primals_8
del primals_9
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_add_mul_rsub_sigmoid_tanh_2[grid(16)](buf1, buf4,
buf2, buf5, 16, XBLOCK=16, num_warps=1, num_stages=1)
return buf5, primals_3, buf0, buf1, buf2, buf3, buf4, buf5
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class BothContextGateNew(nn.Module):
"""Apply the context gate to both contexts"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(BothContextGateNew, self).__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, input_0, input_1, input_2):
primals_4 = self.context_gate.gate.weight
primals_5 = self.context_gate.gate.bias
primals_1 = self.context_gate.source_proj.weight
primals_7 = self.context_gate.source_proj.bias
primals_8 = self.context_gate.target_proj.weight
primals_9 = self.context_gate.target_proj.bias
primals_2 = input_0
primals_3 = input_1
primals_6 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
| NaomiatLibrary/OpenNMT-kpg-release | BothContextGate | false | 874 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class Model(nn.Module):
"""Apply the context gate to both contexts"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, prev_emb, dec_state, attn_state):
z, source, target = self.context_gate(prev_emb, dec_state, attn_state)
return self.tanh((1.0 - z) * target + z * source)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
|
SourceContextGate | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nc/cnc6a3vkphurm472zdavmn3qnff4lmaezxs63jlllw2kks2e62a4.py
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# input_tensor => cat
# Graph fragment:
# %cat : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2, %primals_3], 1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = (xindex // 12)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tmp12 = tl.full([1], 12, tl.int64)
tmp13 = tmp0 < tmp12
tmp14 = tl.load(in_ptr2 + ((4*x1) + ((-8) + x0)), tmp11 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + (x2), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/as/castj5utepsqaf7nmxz7ydwohxvblspgu7zqqkf5geoow2xnjdtg.py
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat_1 => cat_1
# Graph fragment:
# %cat_1 : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2], 1), kwargs = {})
triton_poi_fused_cat_1 = async_compile.triton('triton_poi_fused_cat_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6s/c6s456zdtwmlik5mrlyh27gyo7fuehy7vzu5kdq6caawwt6hjyvr.py
# Topologically Sorted Source Nodes: [z, mul, add, tanh], Original ATen: [aten.sigmoid, aten.mul, aten.add, aten.tanh]
# Source node to ATen node mapping:
# add => add
# mul => mul
# tanh => tanh
# z => sigmoid
# Graph fragment:
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%addmm,), kwargs = {})
# %add_tensor : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default, %primals_9), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %addmm_1), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_tensor, %mul), kwargs = {})
# %tanh : [num_users=1] = call_function[target=torch.ops.aten.tanh.default](args = (%add,), kwargs = {})
triton_poi_fused_add_mul_sigmoid_tanh_2 = async_compile.triton('triton_poi_fused_add_mul_sigmoid_tanh_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_sigmoid_tanh_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_tanh_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x2), xmask)
tmp5 = tl.load(in_ptr2 + (x2), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tl.sigmoid(tmp3)
tmp6 = tmp4 * tmp5
tmp7 = tmp2 + tmp6
tmp8 = libdevice.tanh(tmp7)
tl.store(in_out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4, ), (1, ))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, ), (1, ))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, primals_2, primals_3, buf0, 48, grid=grid(48), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, buf0, reinterpret_tensor(primals_4, (12, 4), (1, 12), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [proj_source], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_7, primals_3, reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf2)
del primals_6
del primals_7
buf3 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
triton_poi_fused_cat_1.run(primals_1, primals_2, buf3, 32, grid=grid(32), stream=stream0)
del primals_1
del primals_2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(buf3, reinterpret_tensor(primals_8, (8, 4), (1, 8), 0), out=buf4)
del primals_8
buf5 = buf4; del buf4 # reuse
# Topologically Sorted Source Nodes: [z, mul, add, tanh], Original ATen: [aten.sigmoid, aten.mul, aten.add, aten.tanh]
triton_poi_fused_add_mul_sigmoid_tanh_2.run(buf5, primals_9, buf1, buf2, 16, grid=grid(16), stream=stream0)
del primals_9
return (buf5, primals_3, buf0, buf1, buf2, buf3, buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 12), (12, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class SourceContextGate(nn.Module):
"""Apply the context gate only to the source context"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(SourceContextGate, self).__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, prev_emb, dec_state, attn_state):
z, source, target = self.context_gate(prev_emb, dec_state, attn_state)
return self.tanh(target + z * source)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = xindex // 12
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tl.full([1], 12, tl.int64)
tmp14 = tl.load(in_ptr2 + (4 * x1 + (-8 + x0)), tmp11 & xmask,
eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + x2, tmp16, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_tanh_2(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + x2, xmask)
tmp5 = tl.load(in_ptr2 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp4 = tl.sigmoid(tmp3)
tmp6 = tmp4 * tmp5
tmp7 = tmp2 + tmp6
tmp8 = libdevice.tanh(tmp7)
tl.store(in_out_ptr0 + x2, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(48)](primals_1, primals_2, primals_3,
buf0, 48, XBLOCK=64, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, buf0, reinterpret_tensor(primals_4,
(12, 4), (1, 12), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, primals_3, reinterpret_tensor(
primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf2)
del primals_6
del primals_7
buf3 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
triton_poi_fused_cat_1[grid(32)](primals_1, primals_2, buf3, 32,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_1
del primals_2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf3, reinterpret_tensor(primals_8, (8, 4), (1, 8
), 0), out=buf4)
del primals_8
buf5 = buf4
del buf4
triton_poi_fused_add_mul_sigmoid_tanh_2[grid(16)](buf5, primals_9,
buf1, buf2, 16, XBLOCK=16, num_warps=1, num_stages=1)
del primals_9
return buf5, primals_3, buf0, buf1, buf2, buf3, buf5
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class SourceContextGateNew(nn.Module):
"""Apply the context gate only to the source context"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(SourceContextGateNew, self).__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, input_0, input_1, input_2):
primals_4 = self.context_gate.gate.weight
primals_5 = self.context_gate.gate.bias
primals_1 = self.context_gate.source_proj.weight
primals_7 = self.context_gate.source_proj.bias
primals_8 = self.context_gate.target_proj.weight
primals_9 = self.context_gate.target_proj.bias
primals_2 = input_0
primals_3 = input_1
primals_6 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
| NaomiatLibrary/OpenNMT-kpg-release | SourceContextGate | false | 875 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class Model(nn.Module):
"""Apply the context gate only to the source context"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, prev_emb, dec_state, attn_state):
z, source, target = self.context_gate(prev_emb, dec_state, attn_state)
return self.tanh(target + z * source)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
|
IIDTransform | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/m4/cm4fnxaqni4hnszc6wcn2ebn3g33kx3ecngfkcujjyvto2zaoqdr.py
# Topologically Sorted Source Nodes: [masked_tensor, albedo_refined, shading_tensor, shading_tensor_1, masked_tensor_1, shading_refined, albedo_refined_1, albedo_refined_2, masked_tensor_2, albedo_refined_3, masked_tensor_3, albedo_tensor, masked_tensor_4, shading_tensor_2, rgb_recon, rgb_recon_1], Original ATen: [aten.le, aten.masked_fill, aten.div, aten.clamp, aten.ge, aten.mul]
# Source node to ATen node mapping:
# albedo_refined => full_default, where
# albedo_refined_1 => div_1
# albedo_refined_2 => clamp_max_1, clamp_min_1
# albedo_refined_3 => full_default_2, where_2
# albedo_tensor => full_default_3, where_3
# masked_tensor => le
# masked_tensor_1 => le_1
# masked_tensor_2 => ge
# masked_tensor_3 => le_2
# masked_tensor_4 => le_3
# rgb_recon => mul
# rgb_recon_1 => clamp_max_2, clamp_min_2
# shading_refined => full_default_1, where_1
# shading_tensor => div
# shading_tensor_1 => clamp_max, clamp_min
# shading_tensor_2 => full_default_4, where_4
# Graph fragment:
# %le : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%arg0_1, 0.0), kwargs = {})
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 1.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%le, %full_default, %arg0_1), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%arg1_1, %where), kwargs = {})
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%div, 0.0), kwargs = {})
# %clamp_max : [num_users=3] = call_function[target=torch.ops.aten.clamp_max.default](args = (%clamp_min, 1.0), kwargs = {})
# %le_1 : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%clamp_max, 0.0), kwargs = {})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 1.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_1 : [num_users=3] = call_function[target=torch.ops.aten.where.self](args = (%le_1, %full_default_1, %clamp_max), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%arg1_1, %where_1), kwargs = {})
# %clamp_min_1 : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%div_1, 0.0), kwargs = {})
# %clamp_max_1 : [num_users=2] = call_function[target=torch.ops.aten.clamp_max.default](args = (%clamp_min_1, 1.0), kwargs = {})
# %ge : [num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%clamp_max_1, 1.0), kwargs = {})
# %full_default_2 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_2 : [num_users=3] = call_function[target=torch.ops.aten.where.self](args = (%ge, %full_default_2, %clamp_max_1), kwargs = {})
# %le_2 : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%where_2, 0.0), kwargs = {})
# %full_default_3 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 1.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_3 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%le_2, %full_default_3, %where_2), kwargs = {})
# %le_3 : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%where_1, 0.0), kwargs = {})
# %full_default_4 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 1.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_4 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%le_3, %full_default_4, %where_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where_3, %where_4), kwargs = {})
# %clamp_min_2 : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%mul, 0.0), kwargs = {})
# %clamp_max_2 : [num_users=1] = call_function[target=torch.ops.aten.clamp_max.default](args = (%clamp_min_2, 1.0), kwargs = {})
triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0 = async_compile.triton('triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask)
tmp2 = 0.0
tmp3 = tmp1 <= tmp2
tmp4 = 1.0
tmp5 = tl.where(tmp3, tmp4, tmp1)
tmp6 = tmp0 / tmp5
tmp7 = triton_helpers.maximum(tmp6, tmp2)
tmp8 = triton_helpers.minimum(tmp7, tmp4)
tmp9 = tmp8 <= tmp2
tmp10 = tl.where(tmp9, tmp4, tmp8)
tmp11 = tmp0 / tmp10
tmp12 = triton_helpers.maximum(tmp11, tmp2)
tmp13 = triton_helpers.minimum(tmp12, tmp4)
tmp14 = tmp13 >= tmp4
tmp15 = tl.where(tmp14, tmp2, tmp13)
tmp16 = tmp15 <= tmp2
tmp17 = tl.where(tmp16, tmp4, tmp15)
tmp18 = tmp10 <= tmp2
tmp19 = tl.where(tmp18, tmp4, tmp10)
tmp20 = tmp17 * tmp19
tmp21 = triton_helpers.maximum(tmp20, tmp2)
tmp22 = triton_helpers.minimum(tmp21, tmp4)
tl.store(out_ptr0 + (x0), tmp8, xmask)
tl.store(out_ptr1 + (x0), tmp15, xmask)
tl.store(out_ptr2 + (x0), tmp22, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [masked_tensor, albedo_refined, shading_tensor, shading_tensor_1, masked_tensor_1, shading_refined, albedo_refined_1, albedo_refined_2, masked_tensor_2, albedo_refined_3, masked_tensor_3, albedo_tensor, masked_tensor_4, shading_tensor_2, rgb_recon, rgb_recon_1], Original ATen: [aten.le, aten.masked_fill, aten.div, aten.clamp, aten.ge, aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0.run(arg1_1, arg0_1, buf0, buf1, buf2, 256, grid=grid(256), stream=stream0)
del arg0_1
del arg1_1
return (buf2, buf0, buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.parallel
import torch.utils.data
from torchvision import transforms
import torch.nn as nn
import torch.cuda
class IIDTransform(nn.Module):
def __init__(self):
super(IIDTransform, self).__init__()
self.transform_op = transforms.Normalize((0.5,), (0.5,))
def mask_fill_nonzeros(self, input_tensor):
output_tensor = torch.clone(input_tensor)
masked_tensor = input_tensor <= 0.0
return output_tensor.masked_fill(masked_tensor, 1.0)
def revert_mask_fill_nonzeros(self, input_tensor):
output_tensor = torch.clone(input_tensor)
masked_tensor = input_tensor >= 1.0
return output_tensor.masked_fill_(masked_tensor, 0.0)
def forward(self, rgb_tensor, albedo_tensor):
min = 0.0
max = 1.0
shading_tensor = self.extract_shading(rgb_tensor, albedo_tensor, False)
shading_refined = self.mask_fill_nonzeros(shading_tensor)
albedo_refined = rgb_tensor / shading_refined
albedo_refined = torch.clip(albedo_refined, min, max)
albedo_refined = self.revert_mask_fill_nonzeros(albedo_refined)
rgb_recon = self.produce_rgb(albedo_refined, shading_refined, False)
rgb_recon = self.transform_op(rgb_recon)
albedo_refined = self.transform_op(albedo_refined)
shading_tensor = self.transform_op(shading_tensor)
return rgb_recon, albedo_refined, shading_tensor
def extract_shading(self, rgb_tensor, albedo_tensor, one_channel=False):
min = 0.0
max = 1.0
albedo_refined = self.mask_fill_nonzeros(albedo_tensor)
shading_tensor = rgb_tensor / albedo_refined
if one_channel is True:
shading_tensor = kornia.color.rgb_to_grayscale(shading_tensor)
shading_tensor = torch.clip(shading_tensor, min, max)
return shading_tensor
def produce_rgb(self, albedo_tensor, shading_tensor, tozeroone=True):
if tozeroone:
albedo_tensor = albedo_tensor * 0.5 + 0.5
shading_tensor = shading_tensor * 0.5 + 0.5
albedo_tensor = self.mask_fill_nonzeros(albedo_tensor)
shading_tensor = self.mask_fill_nonzeros(shading_tensor)
rgb_recon = albedo_tensor * shading_tensor
rgb_recon = torch.clip(rgb_recon, 0.0, 1.0)
return rgb_recon
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn.parallel
import torch.utils.data
from torchvision import transforms
import torch.nn as nn
import torch.cuda
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0(in_ptr0, in_ptr1,
out_ptr0, out_ptr1, out_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask)
tmp2 = 0.0
tmp3 = tmp1 <= tmp2
tmp4 = 1.0
tmp5 = tl.where(tmp3, tmp4, tmp1)
tmp6 = tmp0 / tmp5
tmp7 = triton_helpers.maximum(tmp6, tmp2)
tmp8 = triton_helpers.minimum(tmp7, tmp4)
tmp9 = tmp8 <= tmp2
tmp10 = tl.where(tmp9, tmp4, tmp8)
tmp11 = tmp0 / tmp10
tmp12 = triton_helpers.maximum(tmp11, tmp2)
tmp13 = triton_helpers.minimum(tmp12, tmp4)
tmp14 = tmp13 >= tmp4
tmp15 = tl.where(tmp14, tmp2, tmp13)
tmp16 = tmp15 <= tmp2
tmp17 = tl.where(tmp16, tmp4, tmp15)
tmp18 = tmp10 <= tmp2
tmp19 = tl.where(tmp18, tmp4, tmp10)
tmp20 = tmp17 * tmp19
tmp21 = triton_helpers.maximum(tmp20, tmp2)
tmp22 = triton_helpers.minimum(tmp21, tmp4)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp15, xmask)
tl.store(out_ptr2 + x0, tmp22, xmask)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clamp_div_ge_le_masked_fill_mul_0[grid(256)](arg1_1,
arg0_1, buf0, buf1, buf2, 256, XBLOCK=128, num_warps=4,
num_stages=1)
del arg0_1
del arg1_1
return buf2, buf0, buf1
class IIDTransformNew(nn.Module):
def __init__(self):
super(IIDTransformNew, self).__init__()
self.transform_op = transforms.Normalize((0.5,), (0.5,))
def mask_fill_nonzeros(self, input_tensor):
output_tensor = torch.clone(input_tensor)
masked_tensor = input_tensor <= 0.0
return output_tensor.masked_fill(masked_tensor, 1.0)
def revert_mask_fill_nonzeros(self, input_tensor):
output_tensor = torch.clone(input_tensor)
masked_tensor = input_tensor >= 1.0
return output_tensor.masked_fill_(masked_tensor, 0.0)
def extract_shading(self, rgb_tensor, albedo_tensor, one_channel=False):
min = 0.0
max = 1.0
albedo_refined = self.mask_fill_nonzeros(albedo_tensor)
shading_tensor = rgb_tensor / albedo_refined
if one_channel is True:
shading_tensor = kornia.color.rgb_to_grayscale(shading_tensor)
shading_tensor = torch.clip(shading_tensor, min, max)
return shading_tensor
def produce_rgb(self, albedo_tensor, shading_tensor, tozeroone=True):
if tozeroone:
albedo_tensor = albedo_tensor * 0.5 + 0.5
shading_tensor = shading_tensor * 0.5 + 0.5
albedo_tensor = self.mask_fill_nonzeros(albedo_tensor)
shading_tensor = self.mask_fill_nonzeros(shading_tensor)
rgb_recon = albedo_tensor * shading_tensor
rgb_recon = torch.clip(rgb_recon, 0.0, 1.0)
return rgb_recon
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0], output[1], output[2]
| NeilDG/NeuralNets-Experiment3 | IIDTransform | false | 876 | [
"MIT"
] | 0 | f0d2f788eeca49f803f65810c155491ce687cf9e | https://github.com/NeilDG/NeuralNets-Experiment3/tree/f0d2f788eeca49f803f65810c155491ce687cf9e | import torch
import torch.nn.parallel
import torch.utils.data
from torchvision import transforms
import torch.nn as nn
import torch.cuda
class Model(nn.Module):
def __init__(self):
super().__init__()
self.transform_op = transforms.Normalize((0.5,), (0.5,))
def mask_fill_nonzeros(self, input_tensor):
output_tensor = torch.clone(input_tensor)
masked_tensor = input_tensor <= 0.0
return output_tensor.masked_fill(masked_tensor, 1.0)
def revert_mask_fill_nonzeros(self, input_tensor):
output_tensor = torch.clone(input_tensor)
masked_tensor = input_tensor >= 1.0
return output_tensor.masked_fill_(masked_tensor, 0.0)
def forward(self, rgb_tensor, albedo_tensor):
min = 0.0
max = 1.0
shading_tensor = self.extract_shading(rgb_tensor, albedo_tensor, False)
shading_refined = self.mask_fill_nonzeros(shading_tensor)
albedo_refined = rgb_tensor / shading_refined
albedo_refined = torch.clip(albedo_refined, min, max)
albedo_refined = self.revert_mask_fill_nonzeros(albedo_refined)
rgb_recon = self.produce_rgb(albedo_refined, shading_refined, False)
rgb_recon = self.transform_op(rgb_recon)
albedo_refined = self.transform_op(albedo_refined)
shading_tensor = self.transform_op(shading_tensor)
return rgb_recon, albedo_refined, shading_tensor
def extract_shading(self, rgb_tensor, albedo_tensor, one_channel=False):
min = 0.0
max = 1.0
albedo_refined = self.mask_fill_nonzeros(albedo_tensor)
shading_tensor = rgb_tensor / albedo_refined
if one_channel is True:
shading_tensor = kornia.color.rgb_to_grayscale(shading_tensor)
shading_tensor = torch.clip(shading_tensor, min, max)
return shading_tensor
def produce_rgb(self, albedo_tensor, shading_tensor, tozeroone=True):
if tozeroone:
albedo_tensor = albedo_tensor * 0.5 + 0.5
shading_tensor = shading_tensor * 0.5 + 0.5
albedo_tensor = self.mask_fill_nonzeros(albedo_tensor)
shading_tensor = self.mask_fill_nonzeros(shading_tensor)
rgb_recon = albedo_tensor * shading_tensor
rgb_recon = torch.clip(rgb_recon, 0.0, 1.0)
return rgb_recon
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
TargetContextGate | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nc/cnc6a3vkphurm472zdavmn3qnff4lmaezxs63jlllw2kks2e62a4.py
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# input_tensor => cat
# Graph fragment:
# %cat : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2, %primals_3], 1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = (xindex // 12)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tmp12 = tl.full([1], 12, tl.int64)
tmp13 = tmp0 < tmp12
tmp14 = tl.load(in_ptr2 + ((4*x1) + ((-8) + x0)), tmp11 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + (x2), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/as/castj5utepsqaf7nmxz7ydwohxvblspgu7zqqkf5geoow2xnjdtg.py
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat_1 => cat_1
# Graph fragment:
# %cat_1 : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2], 1), kwargs = {})
triton_poi_fused_cat_1 = async_compile.triton('triton_poi_fused_cat_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/7l/c7lf5woemwxcoeo376uvq4tswpw24vydykzmrhtxemqtbvcbg3gw.py
# Topologically Sorted Source Nodes: [z, mul, add, tanh], Original ATen: [aten.sigmoid, aten.mul, aten.add, aten.tanh]
# Source node to ATen node mapping:
# add => add
# mul => mul
# tanh => tanh
# z => sigmoid
# Graph fragment:
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%addmm,), kwargs = {})
# %add_tensor : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default, %primals_7), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %addmm_2), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %add_tensor), kwargs = {})
# %tanh : [num_users=1] = call_function[target=torch.ops.aten.tanh.default](args = (%add,), kwargs = {})
triton_poi_fused_add_mul_sigmoid_tanh_2 = async_compile.triton('triton_poi_fused_add_mul_sigmoid_tanh_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_sigmoid_tanh_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_tanh_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp2 = tl.load(in_ptr1 + (x2), xmask)
tmp4 = tl.load(in_out_ptr0 + (x2), xmask)
tmp5 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp1 = tl.sigmoid(tmp0)
tmp3 = tmp1 * tmp2
tmp6 = tmp4 + tmp5
tmp7 = tmp3 + tmp6
tmp8 = libdevice.tanh(tmp7)
tl.store(in_out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4, ), (1, ))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, ), (1, ))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, primals_2, primals_3, buf0, 48, grid=grid(48), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, buf0, reinterpret_tensor(primals_4, (12, 4), (1, 12), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(primals_3, reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf2)
del primals_6
buf3 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
triton_poi_fused_cat_1.run(primals_1, primals_2, buf3, 32, grid=grid(32), stream=stream0)
del primals_1
del primals_2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [proj_target], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_9, buf3, reinterpret_tensor(primals_8, (8, 4), (1, 8), 0), alpha=1, beta=1, out=buf4)
del primals_8
del primals_9
buf5 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [z, mul, add, tanh], Original ATen: [aten.sigmoid, aten.mul, aten.add, aten.tanh]
triton_poi_fused_add_mul_sigmoid_tanh_2.run(buf5, buf1, buf4, primals_7, 16, grid=grid(16), stream=stream0)
del primals_7
return (buf5, primals_3, buf0, buf1, buf3, buf4, buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 12), (12, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class TargetContextGate(nn.Module):
"""Apply the context gate only to the target context"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(TargetContextGate, self).__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, prev_emb, dec_state, attn_state):
z, source, target = self.context_gate(prev_emb, dec_state, attn_state)
return self.tanh(z * target + source)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = xindex // 12
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tl.full([1], 12, tl.int64)
tmp14 = tl.load(in_ptr2 + (4 * x1 + (-8 + x0)), tmp11 & xmask,
eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + x2, tmp16, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_tanh_2(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp2 = tl.load(in_ptr1 + x2, xmask)
tmp4 = tl.load(in_out_ptr0 + x2, xmask)
tmp5 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.sigmoid(tmp0)
tmp3 = tmp1 * tmp2
tmp6 = tmp4 + tmp5
tmp7 = tmp3 + tmp6
tmp8 = libdevice.tanh(tmp7)
tl.store(in_out_ptr0 + x2, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(48)](primals_1, primals_2, primals_3,
buf0, 48, XBLOCK=64, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, buf0, reinterpret_tensor(primals_4,
(12, 4), (1, 12), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(primals_3, reinterpret_tensor(primals_6, (4, 4),
(1, 4), 0), out=buf2)
del primals_6
buf3 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
triton_poi_fused_cat_1[grid(32)](primals_1, primals_2, buf3, 32,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_1
del primals_2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, buf3, reinterpret_tensor(primals_8,
(8, 4), (1, 8), 0), alpha=1, beta=1, out=buf4)
del primals_8
del primals_9
buf5 = buf2
del buf2
triton_poi_fused_add_mul_sigmoid_tanh_2[grid(16)](buf5, buf1, buf4,
primals_7, 16, XBLOCK=16, num_warps=1, num_stages=1)
del primals_7
return buf5, primals_3, buf0, buf1, buf3, buf4, buf5
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class TargetContextGateNew(nn.Module):
"""Apply the context gate only to the target context"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(TargetContextGateNew, self).__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, input_0, input_1, input_2):
primals_4 = self.context_gate.gate.weight
primals_5 = self.context_gate.gate.bias
primals_1 = self.context_gate.source_proj.weight
primals_7 = self.context_gate.source_proj.bias
primals_8 = self.context_gate.target_proj.weight
primals_9 = self.context_gate.target_proj.bias
primals_2 = input_0
primals_3 = input_1
primals_6 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
| NaomiatLibrary/OpenNMT-kpg-release | TargetContextGate | false | 877 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
class Model(nn.Module):
"""Apply the context gate only to the target context"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
self.context_gate = ContextGate(embeddings_size, decoder_size,
attention_size, output_size)
self.tanh = nn.Tanh()
def forward(self, prev_emb, dec_state, attn_state):
z, source, target = self.context_gate(prev_emb, dec_state, attn_state)
return self.tanh(z * target + source)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
|
UpSampler | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/yg/cygmdn2auc3tes3j4f4vkac3dflpmrpbzuwybdwsm7q2cg5n4bbd.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten._to_copy]
# Source node to ATen node mapping:
# x_1 => convert_element_type_1
# Graph fragment:
# %convert_element_type_1 : [num_users=7] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%view, torch.int64), kwargs = {})
triton_poi_fused__to_copy_0 = async_compile.triton('triton_poi_fused__to_copy_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*i64', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0,), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__to_copy_0(out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tmp1 = tmp0.to(tl.float32)
tmp2 = 0.5
tmp3 = tmp1 + tmp2
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = tmp5 - tmp2
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp8.to(tl.int32)
tl.store(out_ptr0 + (x0), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/7n/c7nzf22cnpbzyd2j2wkts5chwjan7tdk7qdk4mpjozf4w4f6h2mb.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.add, aten.clamp]
# Source node to ATen node mapping:
# x_1 => add_1, clamp_max
# Graph fragment:
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%convert_element_type_1, 1), kwargs = {})
# %clamp_max : [num_users=5] = call_function[target=torch.ops.aten.clamp_max.default](args = (%add_1, 3), kwargs = {})
triton_poi_fused_add_clamp_1 = async_compile.triton('triton_poi_fused_add_clamp_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*i64', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0,), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_clamp_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_clamp_1(out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tmp1 = tmp0.to(tl.float32)
tmp2 = 0.5
tmp3 = tmp1 + tmp2
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = tmp5 - tmp2
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp8.to(tl.int32)
tmp10 = tl.full([1], 1, tl.int64)
tmp11 = tmp9 + tmp10
tmp12 = tl.full([1], 3, tl.int64)
tmp13 = triton_helpers.minimum(tmp11, tmp12)
tl.store(out_ptr0 + (x0), tmp13, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ix/cixcyqvwbk6rxsv5l4regwkl4tvvkxddkkv7q5vh5eqteaen7y4v.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.arange, aten._to_copy, aten.add, aten.mul, aten.sub, aten.clamp]
# Source node to ATen node mapping:
# x_1 => add, clamp_max_3, clamp_min, clamp_min_3, convert_element_type, iota, mul, sub, sub_3
# Graph fragment:
# %iota : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %convert_element_type : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%iota, torch.float32), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%convert_element_type, 0.5), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%add, 1.0), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul, 0.5), kwargs = {})
# %clamp_min : [num_users=4] = call_function[target=torch.ops.aten.clamp_min.default](args = (%sub, 0.0), kwargs = {})
# %sub_3 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%clamp_min, %convert_element_type_5), kwargs = {})
# %clamp_min_3 : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%sub_3, 0.0), kwargs = {})
# %clamp_max_3 : [num_users=5] = call_function[target=torch.ops.aten.clamp_max.default](args = (%clamp_min_3, 1.0), kwargs = {})
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2 = async_compile.triton('triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0,), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2(out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tmp1 = tmp0.to(tl.float32)
tmp2 = 0.5
tmp3 = tmp1 + tmp2
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = tmp5 - tmp2
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp8.to(tl.int32)
tmp10 = tmp9.to(tl.float32)
tmp11 = tmp8 - tmp10
tmp12 = triton_helpers.maximum(tmp11, tmp7)
tmp13 = triton_helpers.minimum(tmp12, tmp4)
tl.store(out_ptr0 + (x0), tmp13, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/e2/ce2jh5wwu6hmidjna5uagh6ukx2dyd4fceacrm2n3c54e5dujrjx.py
# Topologically Sorted Source Nodes: [x, x_1], Original ATen: [aten.convolution, aten._unsafe_index, aten.sub, aten.mul, aten.add]
# Source node to ATen node mapping:
# x => convolution
# x_1 => _unsafe_index, _unsafe_index_1, _unsafe_index_2, _unsafe_index_3, _unsafe_index_4, _unsafe_index_5, _unsafe_index_6, _unsafe_index_7, add_10, add_11, add_12, add_6, add_7, add_8, add_9, mul_3, mul_4, mul_5, mul_6, mul_7, mul_8, mul_9, sub_10, sub_12, sub_4, sub_5, sub_6, sub_7, sub_9
# Graph fragment:
# %convolution : [num_users=8] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1, 1], [0, 0, 0], [1, 1, 1], False, [0, 0, 0], 1), kwargs = {})
# %_unsafe_index : [num_users=2] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %convert_element_type_1, %convert_element_type_3, %convert_element_type_5]), kwargs = {})
# %_unsafe_index_1 : [num_users=1] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %convert_element_type_1, %convert_element_type_3, %clamp_max_2]), kwargs = {})
# %_unsafe_index_2 : [num_users=2] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %convert_element_type_1, %clamp_max_1, %convert_element_type_5]), kwargs = {})
# %_unsafe_index_3 : [num_users=1] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %convert_element_type_1, %clamp_max_1, %clamp_max_2]), kwargs = {})
# %_unsafe_index_4 : [num_users=2] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %clamp_max, %convert_element_type_3, %convert_element_type_5]), kwargs = {})
# %_unsafe_index_5 : [num_users=1] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %clamp_max, %convert_element_type_3, %clamp_max_2]), kwargs = {})
# %_unsafe_index_6 : [num_users=2] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %clamp_max, %clamp_max_1, %convert_element_type_5]), kwargs = {})
# %_unsafe_index_7 : [num_users=1] = call_function[target=torch.ops.aten._unsafe_index.Tensor](args = (%convolution, [None, None, %clamp_max, %clamp_max_1, %clamp_max_2]), kwargs = {})
# %sub_4 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%_unsafe_index_1, %_unsafe_index), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_4, %clamp_max_3), kwargs = {})
# %add_6 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%_unsafe_index, %mul_3), kwargs = {})
# %sub_5 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%_unsafe_index_3, %_unsafe_index_2), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_5, %clamp_max_3), kwargs = {})
# %add_7 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%_unsafe_index_2, %mul_4), kwargs = {})
# %sub_6 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%_unsafe_index_5, %_unsafe_index_4), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_6, %clamp_max_3), kwargs = {})
# %add_8 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%_unsafe_index_4, %mul_5), kwargs = {})
# %sub_7 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%_unsafe_index_7, %_unsafe_index_6), kwargs = {})
# %mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_7, %clamp_max_3), kwargs = {})
# %add_9 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%_unsafe_index_6, %mul_6), kwargs = {})
# %sub_9 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add_7, %add_6), kwargs = {})
# %mul_7 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_9, %clamp_max_4), kwargs = {})
# %add_10 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_6, %mul_7), kwargs = {})
# %sub_10 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add_9, %add_8), kwargs = {})
# %mul_8 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_10, %clamp_max_4), kwargs = {})
# %add_11 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_8, %mul_8), kwargs = {})
# %sub_12 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add_11, %add_10), kwargs = {})
# %mul_9 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_12, %clamp_max_5), kwargs = {})
# %add_12 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_10, %mul_9), kwargs = {})
triton_poi_fused__unsafe_index_add_convolution_mul_sub_3 = async_compile.triton('triton_poi_fused__unsafe_index_add_convolution_mul_sub_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*i64', 2: '*i64', 3: '*i64', 4: '*fp32', 5: '*fp32', 6: '*i64', 7: '*i64', 8: '*fp32', 9: '*fp32', 10: '*i64', 11: '*fp32', 12: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__unsafe_index_add_convolution_mul_sub_3', 'mutated_arg_names': ['in_out_ptr1'], 'no_x_dim': False, 'num_load': 10, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__unsafe_index_add_convolution_mul_sub_3(in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7, in_ptr8, in_ptr9, in_ptr10, xnumel, XBLOCK : tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = (xindex // 16) % 4
x1 = (xindex // 4) % 4
x0 = xindex % 4
x6 = (xindex // 64)
x3 = (xindex // 64) % 4
x8 = xindex
tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr4 + (x3), xmask, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr5 + (x1), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr6 + (x0), xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr7 + (x0), xmask, eviction_policy='evict_last')
tmp38 = tl.load(in_ptr8 + (x1), xmask, eviction_policy='evict_last')
tmp40 = tl.load(in_ptr9 + (x2), xmask, eviction_policy='evict_last')
tmp63 = tl.load(in_ptr10 + (x2), xmask, eviction_policy='evict_last')
tmp1 = tl.full([XBLOCK], 4, tl.int32)
tmp2 = tmp0 + tmp1
tmp3 = tmp0 < 0
tmp4 = tl.where(tmp3, tmp2, tmp0)
tmp6 = tmp5 + tmp1
tmp7 = tmp5 < 0
tmp8 = tl.where(tmp7, tmp6, tmp5)
tmp10 = tmp9 + tmp1
tmp11 = tmp9 < 0
tmp12 = tl.where(tmp11, tmp10, tmp9)
tmp13 = tl.load(in_ptr3 + (tmp12 + (4*tmp8) + (16*tmp4) + (64*x6)), xmask, eviction_policy='evict_last')
tmp15 = tmp13 + tmp14
tmp17 = tmp16 + tmp1
tmp18 = tmp16 < 0
tmp19 = tl.where(tmp18, tmp17, tmp16)
tmp20 = tl.load(in_ptr3 + (tmp12 + (4*tmp19) + (16*tmp4) + (64*x6)), xmask, eviction_policy='evict_last')
tmp21 = tmp20 + tmp14
tmp23 = tmp22 + tmp1
tmp24 = tmp22 < 0
tmp25 = tl.where(tmp24, tmp23, tmp22)
tmp26 = tl.load(in_ptr3 + (tmp25 + (4*tmp19) + (16*tmp4) + (64*x6)), xmask, eviction_policy='evict_last')
tmp27 = tmp26 + tmp14
tmp28 = tmp27 - tmp21
tmp30 = tmp28 * tmp29
tmp31 = tmp21 + tmp30
tmp32 = tl.load(in_ptr3 + (tmp25 + (4*tmp8) + (16*tmp4) + (64*x6)), xmask, eviction_policy='evict_last')
tmp33 = tmp32 + tmp14
tmp34 = tmp33 - tmp15
tmp35 = tmp34 * tmp29
tmp36 = tmp15 + tmp35
tmp37 = tmp36 - tmp31
tmp39 = tmp37 * tmp38
tmp41 = tmp40 + tmp1
tmp42 = tmp40 < 0
tmp43 = tl.where(tmp42, tmp41, tmp40)
tmp44 = tl.load(in_ptr3 + (tmp12 + (4*tmp8) + (16*tmp43) + (64*x6)), xmask, eviction_policy='evict_last')
tmp45 = tmp44 + tmp14
tmp46 = tl.load(in_ptr3 + (tmp12 + (4*tmp19) + (16*tmp43) + (64*x6)), xmask, eviction_policy='evict_last')
tmp47 = tmp46 + tmp14
tmp48 = tl.load(in_ptr3 + (tmp25 + (4*tmp19) + (16*tmp43) + (64*x6)), xmask, eviction_policy='evict_last')
tmp49 = tmp48 + tmp14
tmp50 = tmp49 - tmp47
tmp51 = tmp50 * tmp29
tmp52 = tmp47 + tmp51
tmp53 = tl.load(in_ptr3 + (tmp25 + (4*tmp8) + (16*tmp43) + (64*x6)), xmask, eviction_policy='evict_last')
tmp54 = tmp53 + tmp14
tmp55 = tmp54 - tmp45
tmp56 = tmp55 * tmp29
tmp57 = tmp45 + tmp56
tmp58 = tmp57 - tmp52
tmp59 = tmp58 * tmp38
tmp60 = tmp31 + tmp39
tmp61 = tmp52 + tmp59
tmp62 = tmp61 - tmp60
tmp64 = tmp62 * tmp63
tmp65 = tmp60 + tmp64
tl.store(in_out_ptr1 + (x8), tmp65, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 1, 1, 1), (4, 1, 1, 1, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4, 4), (256, 64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), transposed=False, output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4, 4), (256, 64, 16, 4, 1))
buf1 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.int64)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten._to_copy]
stream0 = get_raw_stream(0)
triton_poi_fused__to_copy_0.run(buf1, 4, grid=grid(4), stream=stream0)
buf2 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.int64)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.add, aten.clamp]
triton_poi_fused_add_clamp_1.run(buf2, 4, grid=grid(4), stream=stream0)
buf3 = empty_strided_cuda((4, 1), (1, 1), torch.int64)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten._to_copy]
triton_poi_fused__to_copy_0.run(buf3, 4, grid=grid(4), stream=stream0)
buf4 = empty_strided_cuda((4, 1), (1, 1), torch.int64)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.add, aten.clamp]
triton_poi_fused_add_clamp_1.run(buf4, 4, grid=grid(4), stream=stream0)
buf5 = empty_strided_cuda((4, ), (1, ), torch.int64)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.arange, aten._to_copy, aten.add, aten.mul, aten.sub, aten.clamp]
triton_poi_fused__to_copy_0.run(buf5, 4, grid=grid(4), stream=stream0)
buf6 = empty_strided_cuda((4, ), (1, ), torch.int64)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.add, aten.clamp]
triton_poi_fused_add_clamp_1.run(buf6, 4, grid=grid(4), stream=stream0)
buf11 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.arange, aten._to_copy, aten.add, aten.mul, aten.sub, aten.clamp]
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2.run(buf11, 4, grid=grid(4), stream=stream0)
buf14 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.sub, aten.clamp]
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2.run(buf14, 4, grid=grid(4), stream=stream0)
buf17 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.sub, aten.clamp]
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2.run(buf17, 4, grid=grid(4), stream=stream0)
buf8 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1), torch.float32)
buf15 = buf8; del buf8 # reuse
buf18 = buf15; del buf15 # reuse
# Topologically Sorted Source Nodes: [x, x_1], Original ATen: [aten.convolution, aten._unsafe_index, aten.sub, aten.mul, aten.add]
triton_poi_fused__unsafe_index_add_convolution_mul_sub_3.run(buf18, buf1, buf4, buf5, buf0, primals_2, buf3, buf6, buf11, buf14, buf2, buf17, 1024, grid=grid(1024), stream=stream0)
del buf0
del primals_2
return (buf18, primals_1, primals_3, buf1, buf2, buf3, buf4, buf5, buf6, buf11, buf14, buf17, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 1, 1, 1), (4, 1, 1, 1, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4, 4), (256, 64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
from torch.nn import functional as F
class UpSampler(nn.Module):
"""Up Sample module
Decrease the channels size and increase the spatial size of tensor
Extends:
nn.Module
"""
def __init__(self, inChannels, outChannels, spatial_size):
"""
Arguments:
inChannels {int} -- Number of in channels
outChannels {int} -- Number of out channels
spatial_size {tuple} -- Spatial size to get
"""
super(UpSampler, self).__init__()
self.spatial_size = spatial_size
self.conv = nn.Conv3d(inChannels, outChannels, 1)
def forward(self, x):
x = self.conv(x)
x = F.interpolate(x, self.spatial_size, mode='trilinear',
align_corners=False)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4, 4])]
def get_init_inputs():
return [[], {'inChannels': 4, 'outChannels': 4, 'spatial_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__to_copy_0(out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tmp1 = tmp0.to(tl.float32)
tmp2 = 0.5
tmp3 = tmp1 + tmp2
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = tmp5 - tmp2
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp8.to(tl.int32)
tl.store(out_ptr0 + x0, tmp9, xmask)
@triton.jit
def triton_poi_fused_add_clamp_1(out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tmp1 = tmp0.to(tl.float32)
tmp2 = 0.5
tmp3 = tmp1 + tmp2
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = tmp5 - tmp2
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp8.to(tl.int32)
tmp10 = tl.full([1], 1, tl.int64)
tmp11 = tmp9 + tmp10
tmp12 = tl.full([1], 3, tl.int64)
tmp13 = triton_helpers.minimum(tmp11, tmp12)
tl.store(out_ptr0 + x0, tmp13, xmask)
@triton.jit
def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2(out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tmp1 = tmp0.to(tl.float32)
tmp2 = 0.5
tmp3 = tmp1 + tmp2
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = tmp5 - tmp2
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp8.to(tl.int32)
tmp10 = tmp9.to(tl.float32)
tmp11 = tmp8 - tmp10
tmp12 = triton_helpers.maximum(tmp11, tmp7)
tmp13 = triton_helpers.minimum(tmp12, tmp4)
tl.store(out_ptr0 + x0, tmp13, xmask)
@triton.jit
def triton_poi_fused__unsafe_index_add_convolution_mul_sub_3(in_out_ptr1,
in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7,
in_ptr8, in_ptr9, in_ptr10, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16 % 4
x1 = xindex // 4 % 4
x0 = xindex % 4
x6 = xindex // 64
x3 = xindex // 64 % 4
x8 = xindex
tmp0 = tl.load(in_ptr0 + x2, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr4 + x3, xmask, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr5 + x1, xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr6 + x0, xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr7 + x0, xmask, eviction_policy='evict_last')
tmp38 = tl.load(in_ptr8 + x1, xmask, eviction_policy='evict_last')
tmp40 = tl.load(in_ptr9 + x2, xmask, eviction_policy='evict_last')
tmp63 = tl.load(in_ptr10 + x2, xmask, eviction_policy='evict_last')
tmp1 = tl.full([XBLOCK], 4, tl.int32)
tmp2 = tmp0 + tmp1
tmp3 = tmp0 < 0
tmp4 = tl.where(tmp3, tmp2, tmp0)
tmp6 = tmp5 + tmp1
tmp7 = tmp5 < 0
tmp8 = tl.where(tmp7, tmp6, tmp5)
tmp10 = tmp9 + tmp1
tmp11 = tmp9 < 0
tmp12 = tl.where(tmp11, tmp10, tmp9)
tmp13 = tl.load(in_ptr3 + (tmp12 + 4 * tmp8 + 16 * tmp4 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp15 = tmp13 + tmp14
tmp17 = tmp16 + tmp1
tmp18 = tmp16 < 0
tmp19 = tl.where(tmp18, tmp17, tmp16)
tmp20 = tl.load(in_ptr3 + (tmp12 + 4 * tmp19 + 16 * tmp4 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp21 = tmp20 + tmp14
tmp23 = tmp22 + tmp1
tmp24 = tmp22 < 0
tmp25 = tl.where(tmp24, tmp23, tmp22)
tmp26 = tl.load(in_ptr3 + (tmp25 + 4 * tmp19 + 16 * tmp4 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp27 = tmp26 + tmp14
tmp28 = tmp27 - tmp21
tmp30 = tmp28 * tmp29
tmp31 = tmp21 + tmp30
tmp32 = tl.load(in_ptr3 + (tmp25 + 4 * tmp8 + 16 * tmp4 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp33 = tmp32 + tmp14
tmp34 = tmp33 - tmp15
tmp35 = tmp34 * tmp29
tmp36 = tmp15 + tmp35
tmp37 = tmp36 - tmp31
tmp39 = tmp37 * tmp38
tmp41 = tmp40 + tmp1
tmp42 = tmp40 < 0
tmp43 = tl.where(tmp42, tmp41, tmp40)
tmp44 = tl.load(in_ptr3 + (tmp12 + 4 * tmp8 + 16 * tmp43 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp45 = tmp44 + tmp14
tmp46 = tl.load(in_ptr3 + (tmp12 + 4 * tmp19 + 16 * tmp43 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp47 = tmp46 + tmp14
tmp48 = tl.load(in_ptr3 + (tmp25 + 4 * tmp19 + 16 * tmp43 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp49 = tmp48 + tmp14
tmp50 = tmp49 - tmp47
tmp51 = tmp50 * tmp29
tmp52 = tmp47 + tmp51
tmp53 = tl.load(in_ptr3 + (tmp25 + 4 * tmp8 + 16 * tmp43 + 64 * x6),
xmask, eviction_policy='evict_last')
tmp54 = tmp53 + tmp14
tmp55 = tmp54 - tmp45
tmp56 = tmp55 * tmp29
tmp57 = tmp45 + tmp56
tmp58 = tmp57 - tmp52
tmp59 = tmp58 * tmp38
tmp60 = tmp31 + tmp39
tmp61 = tmp52 + tmp59
tmp62 = tmp61 - tmp60
tmp64 = tmp62 * tmp63
tmp65 = tmp60 + tmp64
tl.store(in_out_ptr1 + x8, tmp65, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 1, 1, 1), (4, 1, 1, 1, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4, 4), (256, 64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4, 4), (256, 64, 16, 4, 1))
buf1 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.int64)
get_raw_stream(0)
triton_poi_fused__to_copy_0[grid(4)](buf1, 4, XBLOCK=4, num_warps=1,
num_stages=1)
buf2 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.int64)
triton_poi_fused_add_clamp_1[grid(4)](buf2, 4, XBLOCK=4, num_warps=
1, num_stages=1)
buf3 = empty_strided_cuda((4, 1), (1, 1), torch.int64)
triton_poi_fused__to_copy_0[grid(4)](buf3, 4, XBLOCK=4, num_warps=1,
num_stages=1)
buf4 = empty_strided_cuda((4, 1), (1, 1), torch.int64)
triton_poi_fused_add_clamp_1[grid(4)](buf4, 4, XBLOCK=4, num_warps=
1, num_stages=1)
buf5 = empty_strided_cuda((4,), (1,), torch.int64)
triton_poi_fused__to_copy_0[grid(4)](buf5, 4, XBLOCK=4, num_warps=1,
num_stages=1)
buf6 = empty_strided_cuda((4,), (1,), torch.int64)
triton_poi_fused_add_clamp_1[grid(4)](buf6, 4, XBLOCK=4, num_warps=
1, num_stages=1)
buf11 = empty_strided_cuda((4,), (1,), torch.float32)
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2[grid(4)](buf11,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf14 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2[grid(4)](buf14,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf17 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.float32)
triton_poi_fused__to_copy_add_arange_clamp_mul_sub_2[grid(4)](buf17,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf8 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1),
torch.float32)
buf15 = buf8
del buf8
buf18 = buf15
del buf15
triton_poi_fused__unsafe_index_add_convolution_mul_sub_3[grid(1024)](
buf18, buf1, buf4, buf5, buf0, primals_2, buf3, buf6, buf11,
buf14, buf2, buf17, 1024, XBLOCK=128, num_warps=4, num_stages=1)
del buf0
del primals_2
return (buf18, primals_1, primals_3, buf1, buf2, buf3, buf4, buf5, buf6,
buf11, buf14, buf17)
class UpSamplerNew(nn.Module):
"""Up Sample module
Decrease the channels size and increase the spatial size of tensor
Extends:
nn.Module
"""
def __init__(self, inChannels, outChannels, spatial_size):
"""
Arguments:
inChannels {int} -- Number of in channels
outChannels {int} -- Number of out channels
spatial_size {tuple} -- Spatial size to get
"""
super(UpSamplerNew, self).__init__()
self.spatial_size = spatial_size
self.conv = nn.Conv3d(inChannels, outChannels, 1)
def forward(self, input_0):
primals_1 = self.conv.weight
primals_2 = self.conv.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Myyyr/segmentation | UpSampler | false | 878 | [
"MIT"
] | 0 | 6b9423e327cff1eb23599404031b7fb8e9ecf75d | https://github.com/Myyyr/segmentation/tree/6b9423e327cff1eb23599404031b7fb8e9ecf75d | import torch
from torch import nn
from torch.nn import functional as F
class Model(nn.Module):
"""Up Sample module
Decrease the channels size and increase the spatial size of tensor
Extends:
nn.Module
"""
def __init__(self, inChannels, outChannels, spatial_size):
"""
Arguments:
inChannels {int} -- Number of in channels
outChannels {int} -- Number of out channels
spatial_size {tuple} -- Spatial size to get
"""
super().__init__()
self.spatial_size = spatial_size
self.conv = nn.Conv3d(inChannels, outChannels, 1)
def forward(self, x):
x = self.conv(x)
x = F.interpolate(x, self.spatial_size, mode='trilinear',
align_corners=False)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
Conv | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/dn/cdn4hbqepg3y55cba52oq2rbsg46dywnx7zzzodovhv5qj6yewkg.py
# Topologically Sorted Source Nodes: [weight], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# weight => mul
# Graph fragment:
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_1, 0.23570226039551584), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 144
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.23570226039551584
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/32/c32v7egt4mupqssam3gmac2qgv3ujprjybthsgweflmot256qqw7.py
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# conv2d => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %mul, %primals_2, [1, 1], [1, 1], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_1 = async_compile.triton('triton_poi_fused_convolution_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 16) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 3, 3), (36, 9, 3, 1), torch.float32)
# Topologically Sorted Source Nodes: [weight], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(primals_1, buf0, 144, grid=grid(144), stream=stream0)
del primals_1
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(primals_3, buf0, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 4, 4), (64, 16, 4, 1))
buf2 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
triton_poi_fused_convolution_1.run(buf2, primals_2, 256, grid=grid(256), stream=stream0)
del primals_2
return (buf2, buf0, primals_3, buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 3, 3), (36, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
from math import sqrt
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Conv(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1,
padding=1):
super().__init__()
conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
kernel_size=kernel_size, stride=stride, padding=padding)
conv.weight.data.normal_()
conv.bias.data.zero_()
self.conv = equal_lr(conv)
def forward(self, input):
return self.conv(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
from math import sqrt
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 144
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.23570226039551584
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 3, 3), (36, 9, 3, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(144)](primals_1, buf0, 144, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_1
buf1 = extern_kernels.convolution(primals_3, buf0, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 4, 4), (64, 16, 4, 1))
buf2 = buf1
del buf1
triton_poi_fused_convolution_1[grid(256)](buf2, primals_2, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
return buf2, buf0, primals_3, buf0
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class ConvNew(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1,
padding=1):
super().__init__()
conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
kernel_size=kernel_size, stride=stride, padding=padding)
conv.weight.data.normal_()
conv.bias.data.zero_()
self.conv = equal_lr(conv)
def forward(self, input_0):
primals_2 = self.conv.bias
primals_1 = self.conv.weight_orig
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio | Conv | false | 879 | [
"MIT"
] | 0 | 231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | https://github.com/NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio/tree/231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | import torch
import torch.nn as nn
from math import sqrt
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Model(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1,
padding=1):
super().__init__()
conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
kernel_size=kernel_size, stride=stride, padding=padding)
conv.weight.data.normal_()
conv.bias.data.zero_()
self.conv = equal_lr(conv)
def forward(self, input):
return self.conv(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
AverageAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/75/c75w3rgnfmm4c7hp5div65urlkb5kzh2656pt75swmio7vzn3vp3.py
# Topologically Sorted Source Nodes: [ones, triangle, mask], Original ATen: [aten.ones, aten.tril, aten.mul]
# Source node to ATen node mapping:
# mask => mul_1
# ones => full_default
# triangle => full_default_1, le, sub, where
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([4, 4], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%unsqueeze, %unsqueeze_1), kwargs = {})
# %le : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%sub, 0), kwargs = {})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%le, %full_default, %full_default_1), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where, %permute), kwargs = {})
triton_poi_fused_mul_ones_tril_0 = async_compile.triton('triton_poi_fused_mul_ones_tril_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_ones_tril_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_ones_tril_0(out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = x0 + ((-1)*x1)
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 <= tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = 1 + x1
tmp7 = tmp6.to(tl.float32)
tmp8 = tmp3 / tmp7
tmp9 = tmp5 * tmp8
tl.store(out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ha/chavpwdtejkyqus2olvrr56v6fhdolpm5dx6l26ahmwfvz664fnv.py
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %bmm], -1), kwargs = {})
triton_poi_fused_cat_1 = async_compile.triton('triton_poi_fused_cat_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/bj/cbjkk5x2yiy67l3q4l7ooe5u7plvwkualpweocfe25rsydr62zek.py
# Topologically Sorted Source Nodes: [sigmoid, mul_1, sigmoid_1, mul_2, gating_outputs_1], Original ATen: [aten.sigmoid, aten.mul, aten.add, aten.sigmoid_backward]
# Source node to ATen node mapping:
# gating_outputs_1 => add_1
# mul_1 => mul_2
# mul_2 => mul_3
# sigmoid => sigmoid
# sigmoid_1 => sigmoid_1
# Graph fragment:
# %sigmoid : [num_users=3] = call_function[target=torch.ops.aten.sigmoid.default](args = (%getitem,), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %primals_1), kwargs = {})
# %sigmoid_1 : [num_users=3] = call_function[target=torch.ops.aten.sigmoid.default](args = (%getitem_1,), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid_1, %bmm), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_2, %mul_3), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %sigmoid_1), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid_1, %sub_1), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %sigmoid), kwargs = {})
# %mul_8 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %sub_2), kwargs = {})
triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2 = async_compile.triton('triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, out_ptr1, out_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (8*x1)), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + (x2), xmask)
tmp6 = tl.load(in_ptr0 + (4 + x0 + (8*x1)), xmask)
tmp7 = tl.load(in_ptr1 + (4 + x0), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr3 + (x2), xmask)
tmp2 = tmp0 + tmp1
tmp3 = tl.sigmoid(tmp2)
tmp5 = tmp3 * tmp4
tmp8 = tmp6 + tmp7
tmp9 = tl.sigmoid(tmp8)
tmp11 = tmp9 * tmp10
tmp12 = tmp5 + tmp11
tmp13 = 1.0
tmp14 = tmp13 - tmp9
tmp15 = tmp9 * tmp14
tmp16 = tmp13 - tmp3
tmp17 = tmp3 * tmp16
tl.store(out_ptr0 + (x2), tmp12, xmask)
tl.store(out_ptr1 + (x2), tmp15, xmask)
tl.store(out_ptr2 + (x2), tmp17, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (8, 8), (8, 1))
assert_size_stride(primals_3, (8, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [ones, triangle, mask], Original ATen: [aten.ones, aten.tril, aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_ones_tril_0.run(buf0, 16, grid=grid(16), stream=stream0)
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [average_outputs], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf0, (4, 4, 4), (0, 4, 1), 0), primals_1, out=buf1)
del buf0
buf2 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
triton_poi_fused_cat_1.run(primals_1, buf1, buf2, 128, grid=grid(128), stream=stream0)
buf3 = empty_strided_cuda((16, 8), (8, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf2, (16, 8), (8, 1), 0), reinterpret_tensor(primals_2, (8, 8), (1, 8), 0), out=buf3)
del primals_2
buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sigmoid, mul_1, sigmoid_1, mul_2, gating_outputs_1], Original ATen: [aten.sigmoid, aten.mul, aten.add, aten.sigmoid_backward]
triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2.run(buf3, primals_3, primals_1, buf1, buf4, buf5, buf6, 64, grid=grid(64), stream=stream0)
del buf3
del primals_3
return (buf4, buf1, primals_1, buf1, reinterpret_tensor(buf2, (16, 8), (8, 1), 0), buf5, buf6, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((8, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((8, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class PositionwiseFeedForward(nn.Module):
""" A two-layer Feed-Forward-Network with residual layer norm.
Args:
d_model (int): the size of input for the first-layer of the FFN.
d_ff (int): the hidden layer size of the second-layer
of the FNN.
dropout (float): dropout probability in :math:`[0, 1)`.
"""
def __init__(self, d_model, d_ff, dropout=0.1):
super(PositionwiseFeedForward, self).__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-06)
self.dropout_1 = nn.Dropout(dropout)
self.relu = nn.ReLU()
self.dropout_2 = nn.Dropout(dropout)
def forward(self, x):
"""Layer definition.
Args:
x: ``(batch_size, input_len, model_dim)``
Returns:
(FloatTensor): Output ``(batch_size, input_len, model_dim)``.
"""
inter = self.dropout_1(self.relu(self.w_1(self.layer_norm(x))))
output = self.dropout_2(self.w_2(inter))
return output + x
def update_dropout(self, dropout):
self.dropout_1.p = dropout
self.dropout_2.p = dropout
class AverageAttention(nn.Module):
"""
Average Attention module from
"Accelerating Neural Transformer via an Average Attention Network"
:cite:`DBLP:journals/corr/abs-1805-00631`.
Args:
model_dim (int): the dimension of keys/values/queries,
must be divisible by head_count
dropout (float): dropout parameter
"""
def __init__(self, model_dim, dropout=0.1, aan_useffn=False):
self.model_dim = model_dim
self.aan_useffn = aan_useffn
super(AverageAttention, self).__init__()
if aan_useffn:
self.average_layer = PositionwiseFeedForward(model_dim,
model_dim, dropout)
self.gating_layer = nn.Linear(model_dim * 2, model_dim * 2)
def cumulative_average_mask(self, batch_size, inputs_len, device):
"""
Builds the mask to compute the cumulative average as described in
:cite:`DBLP:journals/corr/abs-1805-00631` -- Figure 3
Args:
batch_size (int): batch size
inputs_len (int): length of the inputs
Returns:
(FloatTensor):
* A Tensor of shape ``(batch_size, input_len, input_len)``
"""
triangle = torch.tril(torch.ones(inputs_len, inputs_len, dtype=
torch.float, device=device))
weights = torch.ones(1, inputs_len, dtype=torch.float, device=device
) / torch.arange(1, inputs_len + 1, dtype=torch.float, device=
device)
mask = triangle * weights.transpose(0, 1)
return mask.unsqueeze(0).expand(batch_size, inputs_len, inputs_len)
def cumulative_average(self, inputs, mask_or_step, layer_cache=None,
step=None):
"""
Computes the cumulative average as described in
:cite:`DBLP:journals/corr/abs-1805-00631` -- Equations (1) (5) (6)
Args:
inputs (FloatTensor): sequence to average
``(batch_size, input_len, dimension)``
mask_or_step: if cache is set, this is assumed
to be the current step of the
dynamic decoding. Otherwise, it is the mask matrix
used to compute the cumulative average.
layer_cache: a dictionary containing the cumulative average
of the previous step.
Returns:
a tensor of the same shape and type as ``inputs``.
"""
if layer_cache is not None:
step = mask_or_step
average_attention = (inputs + step * layer_cache['prev_g']) / (step
+ 1)
layer_cache['prev_g'] = average_attention
return average_attention
else:
mask = mask_or_step
return torch.matmul(mask, inputs)
def forward(self, inputs, mask=None, layer_cache=None, step=None):
"""
Args:
inputs (FloatTensor): ``(batch_size, input_len, model_dim)``
Returns:
(FloatTensor, FloatTensor):
* gating_outputs ``(batch_size, input_len, model_dim)``
* average_outputs average attention
``(batch_size, input_len, model_dim)``
"""
batch_size = inputs.size(0)
inputs_len = inputs.size(1)
average_outputs = self.cumulative_average(inputs, self.
cumulative_average_mask(batch_size, inputs_len, inputs.device) if
layer_cache is None else step, layer_cache=layer_cache)
if self.aan_useffn:
average_outputs = self.average_layer(average_outputs)
gating_outputs = self.gating_layer(torch.cat((inputs,
average_outputs), -1))
input_gate, forget_gate = torch.chunk(gating_outputs, 2, dim=2)
gating_outputs = torch.sigmoid(input_gate) * inputs + torch.sigmoid(
forget_gate) * average_outputs
return gating_outputs, average_outputs
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'model_dim': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_ones_tril_0(out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = x0 + -1 * x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 <= tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = 1 + x1
tmp7 = tmp6.to(tl.float32)
tmp8 = tmp3 / tmp7
tmp9 = tmp5 * tmp8
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2(in_ptr0, in_ptr1,
in_ptr2, in_ptr3, out_ptr0, out_ptr1, out_ptr2, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 8 * x1), xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + x2, xmask)
tmp6 = tl.load(in_ptr0 + (4 + x0 + 8 * x1), xmask)
tmp7 = tl.load(in_ptr1 + (4 + x0), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr3 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp3 = tl.sigmoid(tmp2)
tmp5 = tmp3 * tmp4
tmp8 = tmp6 + tmp7
tmp9 = tl.sigmoid(tmp8)
tmp11 = tmp9 * tmp10
tmp12 = tmp5 + tmp11
tmp13 = 1.0
tmp14 = tmp13 - tmp9
tmp15 = tmp9 * tmp14
tmp16 = tmp13 - tmp3
tmp17 = tmp3 * tmp16
tl.store(out_ptr0 + x2, tmp12, xmask)
tl.store(out_ptr1 + x2, tmp15, xmask)
tl.store(out_ptr2 + x2, tmp17, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (8, 8), (8, 1))
assert_size_stride(primals_3, (8,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_ones_tril_0[grid(16)](buf0, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf0, (4, 4, 4), (0, 4, 1), 0
), primals_1, out=buf1)
del buf0
buf2 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
triton_poi_fused_cat_1[grid(128)](primals_1, buf1, buf2, 128,
XBLOCK=128, num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((16, 8), (8, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (16, 8), (8, 1), 0),
reinterpret_tensor(primals_2, (8, 8), (1, 8), 0), out=buf3)
del primals_2
buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_mul_sigmoid_sigmoid_backward_2[grid(64)](buf3,
primals_3, primals_1, buf1, buf4, buf5, buf6, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf3
del primals_3
return buf4, buf1, primals_1, buf1, reinterpret_tensor(buf2, (16, 8), (
8, 1), 0), buf5, buf6
class PositionwiseFeedForward(nn.Module):
""" A two-layer Feed-Forward-Network with residual layer norm.
Args:
d_model (int): the size of input for the first-layer of the FFN.
d_ff (int): the hidden layer size of the second-layer
of the FNN.
dropout (float): dropout probability in :math:`[0, 1)`.
"""
def __init__(self, d_model, d_ff, dropout=0.1):
super(PositionwiseFeedForward, self).__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-06)
self.dropout_1 = nn.Dropout(dropout)
self.relu = nn.ReLU()
self.dropout_2 = nn.Dropout(dropout)
def forward(self, x):
"""Layer definition.
Args:
x: ``(batch_size, input_len, model_dim)``
Returns:
(FloatTensor): Output ``(batch_size, input_len, model_dim)``.
"""
inter = self.dropout_1(self.relu(self.w_1(self.layer_norm(x))))
output = self.dropout_2(self.w_2(inter))
return output + x
def update_dropout(self, dropout):
self.dropout_1.p = dropout
self.dropout_2.p = dropout
class AverageAttentionNew(nn.Module):
"""
Average Attention module from
"Accelerating Neural Transformer via an Average Attention Network"
:cite:`DBLP:journals/corr/abs-1805-00631`.
Args:
model_dim (int): the dimension of keys/values/queries,
must be divisible by head_count
dropout (float): dropout parameter
"""
def __init__(self, model_dim, dropout=0.1, aan_useffn=False):
self.model_dim = model_dim
self.aan_useffn = aan_useffn
super(AverageAttentionNew, self).__init__()
if aan_useffn:
self.average_layer = PositionwiseFeedForward(model_dim,
model_dim, dropout)
self.gating_layer = nn.Linear(model_dim * 2, model_dim * 2)
def cumulative_average_mask(self, batch_size, inputs_len, device):
"""
Builds the mask to compute the cumulative average as described in
:cite:`DBLP:journals/corr/abs-1805-00631` -- Figure 3
Args:
batch_size (int): batch size
inputs_len (int): length of the inputs
Returns:
(FloatTensor):
* A Tensor of shape ``(batch_size, input_len, input_len)``
"""
triangle = torch.tril(torch.ones(inputs_len, inputs_len, dtype=
torch.float, device=device))
weights = torch.ones(1, inputs_len, dtype=torch.float, device=device
) / torch.arange(1, inputs_len + 1, dtype=torch.float, device=
device)
mask = triangle * weights.transpose(0, 1)
return mask.unsqueeze(0).expand(batch_size, inputs_len, inputs_len)
def cumulative_average(self, inputs, mask_or_step, layer_cache=None,
step=None):
"""
Computes the cumulative average as described in
:cite:`DBLP:journals/corr/abs-1805-00631` -- Equations (1) (5) (6)
Args:
inputs (FloatTensor): sequence to average
``(batch_size, input_len, dimension)``
mask_or_step: if cache is set, this is assumed
to be the current step of the
dynamic decoding. Otherwise, it is the mask matrix
used to compute the cumulative average.
layer_cache: a dictionary containing the cumulative average
of the previous step.
Returns:
a tensor of the same shape and type as ``inputs``.
"""
if layer_cache is not None:
step = mask_or_step
average_attention = (inputs + step * layer_cache['prev_g']) / (step
+ 1)
layer_cache['prev_g'] = average_attention
return average_attention
else:
mask = mask_or_step
return torch.matmul(mask, inputs)
def forward(self, input_0):
primals_2 = self.gating_layer.weight
primals_3 = self.gating_layer.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0], output[1]
| NaomiatLibrary/OpenNMT-kpg-release | AverageAttention | false | 880 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class PositionwiseFeedForward(nn.Module):
""" A two-layer Feed-Forward-Network with residual layer norm.
Args:
d_model (int): the size of input for the first-layer of the FFN.
d_ff (int): the hidden layer size of the second-layer
of the FNN.
dropout (float): dropout probability in :math:`[0, 1)`.
"""
def __init__(self, d_model, d_ff, dropout=0.1):
super().__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-06)
self.dropout_1 = nn.Dropout(dropout)
self.relu = nn.ReLU()
self.dropout_2 = nn.Dropout(dropout)
def forward(self, x):
"""Layer definition.
Args:
x: ``(batch_size, input_len, model_dim)``
Returns:
(FloatTensor): Output ``(batch_size, input_len, model_dim)``.
"""
inter = self.dropout_1(self.relu(self.w_1(self.layer_norm(x))))
output = self.dropout_2(self.w_2(inter))
return output + x
def update_dropout(self, dropout):
self.dropout_1.p = dropout
self.dropout_2.p = dropout
class Model(nn.Module):
"""
Average Attention module from
"Accelerating Neural Transformer via an Average Attention Network"
:cite:`DBLP:journals/corr/abs-1805-00631`.
Args:
model_dim (int): the dimension of keys/values/queries,
must be divisible by head_count
dropout (float): dropout parameter
"""
def __init__(self, model_dim, dropout=0.1, aan_useffn=False):
self.model_dim = model_dim
self.aan_useffn = aan_useffn
super().__init__()
if aan_useffn:
self.average_layer = PositionwiseFeedForward(model_dim,
model_dim, dropout)
self.gating_layer = nn.Linear(model_dim * 2, model_dim * 2)
def cumulative_average_mask(self, batch_size, inputs_len, device):
"""
Builds the mask to compute the cumulative average as described in
:cite:`DBLP:journals/corr/abs-1805-00631` -- Figure 3
Args:
batch_size (int): batch size
inputs_len (int): length of the inputs
Returns:
(FloatTensor):
* A Tensor of shape ``(batch_size, input_len, input_len)``
"""
triangle = torch.tril(torch.ones(inputs_len, inputs_len, dtype=
torch.float, device=device))
weights = torch.ones(1, inputs_len, dtype=torch.float, device=device
) / torch.arange(1, inputs_len + 1, dtype=torch.float, device=
device)
mask = triangle * weights.transpose(0, 1)
return mask.unsqueeze(0).expand(batch_size, inputs_len, inputs_len)
def cumulative_average(self, inputs, mask_or_step, layer_cache=None,
step=None):
"""
Computes the cumulative average as described in
:cite:`DBLP:journals/corr/abs-1805-00631` -- Equations (1) (5) (6)
Args:
inputs (FloatTensor): sequence to average
``(batch_size, input_len, dimension)``
mask_or_step: if cache is set, this is assumed
to be the current step of the
dynamic decoding. Otherwise, it is the mask matrix
used to compute the cumulative average.
layer_cache: a dictionary containing the cumulative average
of the previous step.
Returns:
a tensor of the same shape and type as ``inputs``.
"""
if layer_cache is not None:
step = mask_or_step
average_attention = (inputs + step * layer_cache['prev_g']) / (step
+ 1)
layer_cache['prev_g'] = average_attention
return average_attention
else:
mask = mask_or_step
return torch.matmul(mask, inputs)
def fo
# ... truncated (>4000 chars) for memory efficiency |
ContextGate | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nc/cnc6a3vkphurm472zdavmn3qnff4lmaezxs63jlllw2kks2e62a4.py
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# input_tensor => cat
# Graph fragment:
# %cat : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2, %primals_3], 1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = (xindex // 12)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp9 & xmask, eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tmp12 = tl.full([1], 12, tl.int64)
tmp13 = tmp0 < tmp12
tmp14 = tl.load(in_ptr2 + ((4*x1) + ((-8) + x0)), tmp11 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + (x2), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/x7/cx7xziu4lpr42gzh3hblzhyhhr2agimvsluvyrub77hqbwauajw5.py
# Topologically Sorted Source Nodes: [z], Original ATen: [aten.sigmoid]
# Source node to ATen node mapping:
# z => sigmoid
# Graph fragment:
# %add_tensor : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default, %primals_5), kwargs = {})
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%add_tensor,), kwargs = {})
triton_poi_fused_sigmoid_1 = async_compile.triton('triton_poi_fused_sigmoid_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sigmoid_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_sigmoid_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.sigmoid(tmp2)
tl.store(in_out_ptr0 + (x2), tmp3, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/zd/czdeq2ohbgubcyeps2ukquvfhigxtyega57i24ketclusfgmyedi.py
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat_1 => cat_1
# Graph fragment:
# %cat_1 : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%primals_1, %primals_2], 1), kwargs = {})
triton_poi_fused_cat_2 = async_compile.triton('triton_poi_fused_cat_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4, ), (1, ))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, ), (1, ))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
# Topologically Sorted Source Nodes: [input_tensor], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, primals_2, primals_3, buf0, 48, grid=grid(48), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(buf0, reinterpret_tensor(primals_4, (12, 4), (1, 12), 0), out=buf1)
del primals_4
buf2 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [z], Original ATen: [aten.sigmoid]
triton_poi_fused_sigmoid_1.run(buf2, primals_5, 16, grid=grid(16), stream=stream0)
del primals_5
buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [proj_source], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_7, primals_3, reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf3)
del primals_6
del primals_7
buf4 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat_1], Original ATen: [aten.cat]
triton_poi_fused_cat_2.run(primals_1, primals_2, buf4, 32, grid=grid(32), stream=stream0)
del primals_1
del primals_2
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [proj_target], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_9, buf4, reinterpret_tensor(primals_8, (8, 4), (1, 8), 0), alpha=1, beta=1, out=buf5)
del primals_8
del primals_9
return (buf2, buf3, buf5, primals_3, buf0, buf2, buf4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 12), (12, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class ContextGate(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGate, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 48
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x1 = xindex // 12
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tl.full([1], 12, tl.int64)
tmp14 = tl.load(in_ptr2 + (4 * x1 + (-8 + x0)), tmp11 & xmask,
eviction_policy='evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + x2, tmp16, xmask)
@triton.jit
def triton_poi_fused_sigmoid_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.sigmoid(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
@triton.jit
def triton_poi_fused_cat_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 12), (12, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 12), (12, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(48)](primals_1, primals_2, primals_3,
buf0, 48, XBLOCK=64, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf0, reinterpret_tensor(primals_4, (12, 4), (1,
12), 0), out=buf1)
del primals_4
buf2 = buf1
del buf1
triton_poi_fused_sigmoid_1[grid(16)](buf2, primals_5, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del primals_5
buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, primals_3, reinterpret_tensor(
primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf3)
del primals_6
del primals_7
buf4 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
triton_poi_fused_cat_2[grid(32)](primals_1, primals_2, buf4, 32,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_1
del primals_2
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, buf4, reinterpret_tensor(primals_8,
(8, 4), (1, 8), 0), alpha=1, beta=1, out=buf5)
del primals_8
del primals_9
return buf2, buf3, buf5, primals_3, buf0, buf2, buf4
class ContextGateNew(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super(ContextGateNew, self).__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, input_0, input_1, input_2):
primals_4 = self.gate.weight
primals_5 = self.gate.bias
primals_1 = self.source_proj.weight
primals_7 = self.source_proj.bias
primals_8 = self.target_proj.weight
primals_9 = self.target_proj.bias
primals_2 = input_0
primals_3 = input_1
primals_6 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0], output[1], output[2]
| NaomiatLibrary/OpenNMT-kpg-release | ContextGate | false | 881 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class Model(nn.Module):
"""
Context gate is a decoder module that takes as input the previous word
embedding, the current decoder state and the attention state, and
produces a gate.
The gate can be used to select the input from the target side context
(decoder state), from the source context (attention state) or both.
"""
def __init__(self, embeddings_size, decoder_size, attention_size,
output_size):
super().__init__()
input_size = embeddings_size + decoder_size + attention_size
self.gate = nn.Linear(input_size, output_size, bias=True)
self.sig = nn.Sigmoid()
self.source_proj = nn.Linear(attention_size, output_size)
self.target_proj = nn.Linear(embeddings_size + decoder_size,
output_size)
def forward(self, prev_emb, dec_state, attn_state):
input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1)
z = self.sig(self.gate(input_tensor))
proj_source = self.source_proj(attn_state)
proj_target = self.target_proj(torch.cat((prev_emb, dec_state), dim=1))
return z, proj_source, proj_target
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embeddings_size': 4, 'decoder_size': 4, 'attention_size':
4, 'output_size': 4}]
|
MultiHeadAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/fg/cfg742icmosiwp5ugziye26din5ueqx3v7ntptkkpyackudldrxs.py
# Topologically Sorted Source Nodes: [eq], Original ATen: [aten.eq]
# Source node to ATen node mapping:
# eq => eq
# Graph fragment:
# %eq : [num_users=2] = call_function[target=torch.ops.aten.eq.Scalar](args = (%primals_7, 0), kwargs = {})
triton_poi_fused_eq_0 = async_compile.triton('triton_poi_fused_eq_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*i1', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_eq_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_eq_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.0
tmp2 = tmp0 == tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/qj/cqjih4pwu4dm7bujxf22gnjibtqmub34ehk2ak4q4x2axdts4nnl.py
# Topologically Sorted Source Nodes: [attention_1, attention_3], Original ATen: [aten.masked_fill, aten._softmax]
# Source node to ATen node mapping:
# attention_1 => full_default, where
# attention_3 => exp, sum_1
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], -1.0000000200408773e+20), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq, %full_default, %view_9), kwargs = {})
# %mul_tensor : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where, 1), kwargs = {})
# %amax_default : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%mul_tensor, [3], True), kwargs = {})
# %sub_tensor : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_tensor, %amax_default), kwargs = {})
# %div_tensor : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sub_tensor, 2.0), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%div_tensor,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [3], True), kwargs = {})
triton_poi_fused__softmax_masked_fill_1 = async_compile.triton('triton_poi_fused__softmax_masked_fill_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16, 4], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*i1', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_masked_fill_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 9, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_masked_fill_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 4
y1 = (yindex // 4)
tmp0 = tl.load(in_ptr0 + ((4*x2) + (16*y3)), xmask & ymask, eviction_policy='evict_last').to(tl.int1)
tmp1 = tl.load(in_ptr1 + (y0 + (4*x2) + (16*y1)), xmask & ymask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + (y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + (4*x2) + (16*y3)), xmask & ymask, eviction_policy='evict_last').to(tl.int1)
tmp9 = tl.load(in_ptr2 + (4 + y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr0 + (2 + (4*x2) + (16*y3)), xmask & ymask, eviction_policy='evict_last').to(tl.int1)
tmp15 = tl.load(in_ptr2 + (8 + y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr0 + (3 + (4*x2) + (16*y3)), xmask & ymask, eviction_policy='evict_last').to(tl.int1)
tmp21 = tl.load(in_ptr2 + (12 + y0 + (16*y1)), ymask, eviction_policy='evict_last')
tmp3 = tmp1 * tmp2
tmp4 = -1.0000000200408773e+20
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp6 = 1.0
tmp7 = tmp5 * tmp6
tmp10 = tmp1 * tmp9
tmp11 = tl.where(tmp8, tmp4, tmp10)
tmp12 = tmp11 * tmp6
tmp13 = triton_helpers.maximum(tmp7, tmp12)
tmp16 = tmp1 * tmp15
tmp17 = tl.where(tmp14, tmp4, tmp16)
tmp18 = tmp17 * tmp6
tmp19 = triton_helpers.maximum(tmp13, tmp18)
tmp22 = tmp1 * tmp21
tmp23 = tl.where(tmp20, tmp4, tmp22)
tmp24 = tmp23 * tmp6
tmp25 = triton_helpers.maximum(tmp19, tmp24)
tmp26 = tmp7 - tmp25
tmp27 = 0.5
tmp28 = tmp26 * tmp27
tmp29 = tl_math.exp(tmp28)
tmp30 = tmp12 - tmp25
tmp31 = tmp30 * tmp27
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp29 + tmp32
tmp34 = tmp18 - tmp25
tmp35 = tmp34 * tmp27
tmp36 = tl_math.exp(tmp35)
tmp37 = tmp33 + tmp36
tmp38 = tmp24 - tmp25
tmp39 = tmp38 * tmp27
tmp40 = tl_math.exp(tmp39)
tmp41 = tmp37 + tmp40
tl.store(out_ptr0 + (x2 + (4*y3)), tmp25, xmask & ymask)
tl.store(out_ptr1 + (x2 + (4*y3)), tmp41, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/jm/cjmepcz4xj5qri3tmkprchzy3sga6hejrzqq6r5xnqpjctta5tca.py
# Topologically Sorted Source Nodes: [attention_1, attention_3], Original ATen: [aten.masked_fill, aten._softmax]
# Source node to ATen node mapping:
# attention_1 => full_default, where
# attention_3 => div_1, exp
# Graph fragment:
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], -1.0000000200408773e+20), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq, %full_default, %view_9), kwargs = {})
# %mul_tensor : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%where, 1), kwargs = {})
# %sub_tensor : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_tensor, %amax_default), kwargs = {})
# %div_tensor : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sub_tensor, 2.0), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%div_tensor,), kwargs = {})
# %div_1 : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_masked_fill_2 = async_compile.triton('triton_poi_fused__softmax_masked_fill_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*i1', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_masked_fill_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_masked_fill_2(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x1 = (xindex // 4) % 4
x2 = (xindex // 16) % 4
x3 = (xindex // 64)
x0 = xindex % 4
x5 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x4), xmask).to(tl.int1)
tmp1 = tl.load(in_ptr1 + (x2 + (4*x1) + (16*x3)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + (x2 + (4*x0) + (16*x3)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr3 + (x5), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr4 + (x5), xmask, eviction_policy='evict_last')
tmp3 = tmp1 * tmp2
tmp4 = -1.0000000200408773e+20
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp6 = 1.0
tmp7 = tmp5 * tmp6
tmp9 = tmp7 - tmp8
tmp10 = 0.5
tmp11 = tmp9 * tmp10
tmp12 = tl_math.exp(tmp11)
tmp14 = tmp12 / tmp13
tl.store(out_ptr0 + (x4), tmp14, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/xt/cxtkkmujo4ytg6ycpz5lk5livtstr63pg5nsf5ijewjbtrfrqx6k.py
# Topologically Sorted Source Nodes: [einsum_1], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# einsum_1 => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_8,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_3 = async_compile.triton('triton_poi_fused_clone_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16, 4], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = (yindex // 4)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x2) + (16*y1)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + (4*y3)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/q4/cq4lrbjfvbivmpg2zkxhkatw7yc2rqarfj625cpqjlxqgfutfyet.py
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.add]
# Source node to ATen node mapping:
# out_1 => add
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_16, %primals_9), kwargs = {})
triton_poi_fused_add_4 = async_compile.triton('triton_poi_fused_add_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_4', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_4(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_2, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_3, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_4, (1, 1), (1, 1))
assert_size_stride(primals_5, (1, 1), (1, 1))
assert_size_stride(primals_6, (1, 1), (1, 1))
assert_size_stride(primals_7, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_8, (4, 4), (4, 1))
assert_size_stride(primals_9, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [values_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(primals_2, (64, 1), (1, 1), 0), primals_4, out=buf0)
del primals_4
buf1 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [keys_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 1), (1, 1), 0), primals_5, out=buf1)
del primals_5
buf2 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [queries_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(primals_1, (64, 1), (1, 1), 0), primals_6, out=buf2)
del primals_6
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
# Topologically Sorted Source Nodes: [eq], Original ATen: [aten.eq]
stream0 = get_raw_stream(0)
triton_poi_fused_eq_0.run(primals_7, buf3, 256, grid=grid(256), stream=stream0)
del primals_7
buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
# Topologically Sorted Source Nodes: [attention_1, attention_3], Original ATen: [aten.masked_fill, aten._softmax]
triton_poi_fused__softmax_masked_fill_1.run(buf3, buf2, buf1, buf4, buf5, 16, 4, grid=grid(16, 4), stream=stream0)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [attention_1, attention_3], Original ATen: [aten.masked_fill, aten._softmax]
triton_poi_fused__softmax_masked_fill_2.run(buf3, buf2, buf1, buf4, buf5, buf6, 256, grid=grid(256), stream=stream0)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 1, 1), (16, 4, 1, 1, 1), 0); del buf5 # reuse
# Topologically Sorted Source Nodes: [einsum_1], Original ATen: [aten.clone]
triton_poi_fused_clone_3.run(buf0, buf7, 16, 4, grid=grid(16, 4), stream=stream0)
buf8 = reinterpret_tensor(buf0, (16, 4, 1), (4, 1, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [einsum_1], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf7, (16, 4, 1), (4, 1, 0), 0), out=buf8)
buf9 = reinterpret_tensor(buf4, (4, 4, 4), (16, 4, 1), 0); del buf4 # reuse
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.clone]
triton_poi_fused_clone_3.run(buf8, buf9, 16, 4, grid=grid(16, 4), stream=stream0)
buf10 = reinterpret_tensor(buf8, (16, 4), (4, 1), 0); del buf8 # reuse
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf9, (16, 4), (4, 1), 0), reinterpret_tensor(primals_8, (4, 4), (1, 4), 0), out=buf10)
buf11 = reinterpret_tensor(buf10, (4, 4, 4), (16, 4, 1), 0); del buf10 # reuse
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten.add]
triton_poi_fused_add_4.run(buf11, primals_9, 64, grid=grid(64), stream=stream0)
del primals_9
return (buf11, reinterpret_tensor(primals_2, (64, 1), (1, 1), 0), reinterpret_tensor(primals_3, (64, 1), (1, 1), 0), buf1, reinterpret_tensor(primals_1, (64, 1), (1, 1), 0), buf2, buf3, buf6, reinterpret_tensor(buf9, (16, 4), (4, 1), 0), primals_8, reinterpret_tensor(buf7, (16, 1, 4), (4, 1, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 1), (16, 4, 1, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 1), (16, 4, 1, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 1), (16, 4, 1, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, 1), (1, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((1, 1), (1, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((1, 1), (1, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class MultiHeadAttention(nn.Module):
def __init__(self, embedding_size, number_of_heads):
super(MultiHeadAttention, self).__init__()
self.embedding_size = embedding_size
self.number_of_heads = number_of_heads
self.head_dimension = embedding_size // number_of_heads
assert self.head_dimension * number_of_heads == embedding_size, 'Embedding size needs to be divisible by the number of heads'
self.value = nn.Linear(self.head_dimension, self.head_dimension,
bias=False)
self.key = nn.Linear(self.head_dimension, self.head_dimension, bias
=False)
self.query = nn.Linear(self.head_dimension, self.head_dimension,
bias=False)
self.full_connection = nn.Linear(number_of_heads * self.
head_dimension, embedding_size)
def forward(self, query, value, key, mask):
batch_size = query.shape[0]
value_len, key_len, query_len = value.shape[1], key.shape[1
], query.shape[1]
values = value.reshape(batch_size, value_len, self.number_of_heads,
self.head_dimension)
keys = key.reshape(batch_size, key_len, self.number_of_heads, self.
head_dimension)
queries = query.reshape(batch_size, query_len, self.number_of_heads,
self.head_dimension)
values = self.value(values)
keys = self.key(keys)
queries = self.query(queries)
attention = torch.einsum('bqhd,bkhd->bhqk', [queries, keys])
if mask is not None:
attention = attention.masked_fill(mask == 0, float('-1e20'))
attention /= self.embedding_size ** 0.5
attention = torch.softmax(attention, dim=3)
out = torch.einsum('bhql,blhd->bqhd', [attention, values]).reshape(
batch_size, query_len, self.number_of_heads * self.head_dimension)
out = self.full_connection(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 1]), torch.rand([4, 4, 4, 1]), torch.rand(
[4, 4, 4, 1]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'embedding_size': 4, 'number_of_heads': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_eq_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.0
tmp2 = tmp0 == tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused__softmax_masked_fill_1(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.
constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 4
y1 = yindex // 4
tmp0 = tl.load(in_ptr0 + (4 * x2 + 16 * y3), xmask & ymask,
eviction_policy='evict_last').to(tl.int1)
tmp1 = tl.load(in_ptr1 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + (y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp8 = tl.load(in_ptr0 + (1 + 4 * x2 + 16 * y3), xmask & ymask,
eviction_policy='evict_last').to(tl.int1)
tmp9 = tl.load(in_ptr2 + (4 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp14 = tl.load(in_ptr0 + (2 + 4 * x2 + 16 * y3), xmask & ymask,
eviction_policy='evict_last').to(tl.int1)
tmp15 = tl.load(in_ptr2 + (8 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp20 = tl.load(in_ptr0 + (3 + 4 * x2 + 16 * y3), xmask & ymask,
eviction_policy='evict_last').to(tl.int1)
tmp21 = tl.load(in_ptr2 + (12 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp3 = tmp1 * tmp2
tmp4 = -1.0000000200408773e+20
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp6 = 1.0
tmp7 = tmp5 * tmp6
tmp10 = tmp1 * tmp9
tmp11 = tl.where(tmp8, tmp4, tmp10)
tmp12 = tmp11 * tmp6
tmp13 = triton_helpers.maximum(tmp7, tmp12)
tmp16 = tmp1 * tmp15
tmp17 = tl.where(tmp14, tmp4, tmp16)
tmp18 = tmp17 * tmp6
tmp19 = triton_helpers.maximum(tmp13, tmp18)
tmp22 = tmp1 * tmp21
tmp23 = tl.where(tmp20, tmp4, tmp22)
tmp24 = tmp23 * tmp6
tmp25 = triton_helpers.maximum(tmp19, tmp24)
tmp26 = tmp7 - tmp25
tmp27 = 0.5
tmp28 = tmp26 * tmp27
tmp29 = tl_math.exp(tmp28)
tmp30 = tmp12 - tmp25
tmp31 = tmp30 * tmp27
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp29 + tmp32
tmp34 = tmp18 - tmp25
tmp35 = tmp34 * tmp27
tmp36 = tl_math.exp(tmp35)
tmp37 = tmp33 + tmp36
tmp38 = tmp24 - tmp25
tmp39 = tmp38 * tmp27
tmp40 = tl_math.exp(tmp39)
tmp41 = tmp37 + tmp40
tl.store(out_ptr0 + (x2 + 4 * y3), tmp25, xmask & ymask)
tl.store(out_ptr1 + (x2 + 4 * y3), tmp41, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_masked_fill_2(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x0 = xindex % 4
x5 = xindex // 4
tmp0 = tl.load(in_ptr0 + x4, xmask).to(tl.int1)
tmp1 = tl.load(in_ptr1 + (x2 + 4 * x1 + 16 * x3), xmask,
eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + (x2 + 4 * x0 + 16 * x3), xmask,
eviction_policy='evict_last')
tmp8 = tl.load(in_ptr3 + x5, xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr4 + x5, xmask, eviction_policy='evict_last')
tmp3 = tmp1 * tmp2
tmp4 = -1.0000000200408773e+20
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp6 = 1.0
tmp7 = tmp5 * tmp6
tmp9 = tmp7 - tmp8
tmp10 = 0.5
tmp11 = tmp9 * tmp10
tmp12 = tl_math.exp(tmp11)
tmp14 = tmp12 / tmp13
tl.store(out_ptr0 + x4, tmp14, xmask)
@triton.jit
def triton_poi_fused_clone_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_4(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_2, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_3, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_4, (1, 1), (1, 1))
assert_size_stride(primals_5, (1, 1), (1, 1))
assert_size_stride(primals_6, (1, 1), (1, 1))
assert_size_stride(primals_7, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_8, (4, 4), (4, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (64, 1), (1, 1), 0),
primals_4, out=buf0)
del primals_4
buf1 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 1), (1, 1), 0),
primals_5, out=buf1)
del primals_5
buf2 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (64, 1), (1, 1), 0),
primals_6, out=buf2)
del primals_6
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
get_raw_stream(0)
triton_poi_fused_eq_0[grid(256)](primals_7, buf3, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_7
buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_masked_fill_1[grid(16, 4)](buf3, buf2,
buf1, buf4, buf5, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1,
num_stages=1)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_masked_fill_2[grid(256)](buf3, buf2, buf1,
buf4, buf5, buf6, 256, XBLOCK=256, num_warps=4, num_stages=1)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 1, 1), (16, 4, 1, 1, 1), 0)
del buf5
triton_poi_fused_clone_3[grid(16, 4)](buf0, buf7, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf8 = reinterpret_tensor(buf0, (16, 4, 1), (4, 1, 1), 0)
del buf0
extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf7, (16, 4, 1), (4, 1, 0), 0), out=buf8)
buf9 = reinterpret_tensor(buf4, (4, 4, 4), (16, 4, 1), 0)
del buf4
triton_poi_fused_clone_3[grid(16, 4)](buf8, buf9, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf10 = reinterpret_tensor(buf8, (16, 4), (4, 1), 0)
del buf8
extern_kernels.mm(reinterpret_tensor(buf9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_8, (4, 4), (1, 4), 0), out=buf10)
buf11 = reinterpret_tensor(buf10, (4, 4, 4), (16, 4, 1), 0)
del buf10
triton_poi_fused_add_4[grid(64)](buf11, primals_9, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_9
return buf11, reinterpret_tensor(primals_2, (64, 1), (1, 1), 0
), reinterpret_tensor(primals_3, (64, 1), (1, 1), 0
), buf1, reinterpret_tensor(primals_1, (64, 1), (1, 1), 0
), buf2, buf3, buf6, reinterpret_tensor(buf9, (16, 4), (4, 1), 0
), primals_8, reinterpret_tensor(buf7, (16, 1, 4), (4, 1, 1), 0)
class MultiHeadAttentionNew(nn.Module):
def __init__(self, embedding_size, number_of_heads):
super(MultiHeadAttentionNew, self).__init__()
self.embedding_size = embedding_size
self.number_of_heads = number_of_heads
self.head_dimension = embedding_size // number_of_heads
assert self.head_dimension * number_of_heads == embedding_size, 'Embedding size needs to be divisible by the number of heads'
self.value = nn.Linear(self.head_dimension, self.head_dimension,
bias=False)
self.key = nn.Linear(self.head_dimension, self.head_dimension, bias
=False)
self.query = nn.Linear(self.head_dimension, self.head_dimension,
bias=False)
self.full_connection = nn.Linear(number_of_heads * self.
head_dimension, embedding_size)
def forward(self, input_0, input_1, input_2, input_3):
primals_4 = self.value.weight
primals_5 = self.key.weight
primals_6 = self.query.weight
primals_8 = self.full_connection.weight
primals_9 = self.full_connection.bias
primals_1 = input_0
primals_2 = input_1
primals_3 = input_2
primals_7 = input_3
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
| NMT-hub/transformer | MultiHeadAttention | false | 882 | [
"MIT"
] | 0 | e5b332da6a322e8025c30ee7e31fe34a323e7388 | https://github.com/NMT-hub/transformer/tree/e5b332da6a322e8025c30ee7e31fe34a323e7388 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, embedding_size, number_of_heads):
super().__init__()
self.embedding_size = embedding_size
self.number_of_heads = number_of_heads
self.head_dimension = embedding_size // number_of_heads
assert self.head_dimension * number_of_heads == embedding_size, 'Embedding size needs to be divisible by the number of heads'
self.value = nn.Linear(self.head_dimension, self.head_dimension,
bias=False)
self.key = nn.Linear(self.head_dimension, self.head_dimension, bias
=False)
self.query = nn.Linear(self.head_dimension, self.head_dimension,
bias=False)
self.full_connection = nn.Linear(number_of_heads * self.
head_dimension, embedding_size)
def forward(self, query, value, key, mask):
batch_size = query.shape[0]
value_len, key_len, query_len = value.shape[1], key.shape[1
], query.shape[1]
values = value.reshape(batch_size, value_len, self.number_of_heads,
self.head_dimension)
keys = key.reshape(batch_size, key_len, self.number_of_heads, self.
head_dimension)
queries = query.reshape(batch_size, query_len, self.number_of_heads,
self.head_dimension)
values = self.value(values)
keys = self.key(keys)
queries = self.query(queries)
attention = torch.einsum('bqhd,bkhd->bhqk', [queries, keys])
if mask is not None:
attention = attention.masked_fill(mask == 0, float('-1e20'))
attention /= self.embedding_size ** 0.5
attention = torch.softmax(attention, dim=3)
out = torch.einsum('bhql,blhd->bqhd', [attention, values]).reshape(
batch_size, query_len, self.number_of_heads * self.head_dimension)
out = self.full_connection(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 1]), torch.rand([4, 4, 4, 1]), torch.rand(
[4, 4, 4, 1]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
AdaIN | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/52/c526p7iwll7vx7gobeuv6q3lym4ek7lbhopuykpcibc57bou263i.py
# Topologically Sorted Source Nodes: [weight], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# weight => mul
# Graph fragment:
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_1, 0.7071067811865476), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.7071067811865476
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/jo/cjo3wxmtawsvu7opemz2xwvsknw4nxv74xivifhgb7csue6qqjbi.py
# Topologically Sorted Source Nodes: [out, mul_1, out_1], Original ATen: [aten._native_batch_norm_legit, aten.mul, aten.add]
# Source node to ATen node mapping:
# mul_1 => mul_2
# out => add, rsqrt, var_mean
# out_1 => add_1
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%view, [0, 2, 3]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_2, 1e-05), kwargs = {})
# %rsqrt : [num_users=2] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%getitem, %view_1), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_2, %getitem_1), kwargs = {})
triton_per_fused__native_batch_norm_legit_add_mul_1 = async_compile.triton('triton_per_fused__native_batch_norm_legit_add_mul_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[16, 16],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused__native_batch_norm_legit_add_mul_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 4, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused__native_batch_norm_legit_add_mul_1(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 16
rnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
x2 = xindex % 4
x3 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (r1 + (16*x0)), xmask, other=0.0)
tmp22 = tl.load(in_ptr1 + (x2 + (8*x3)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr2 + (x2), xmask, eviction_policy='evict_last')
tmp28 = tl.load(in_ptr1 + (4 + x2 + (8*x3)), xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr2 + (4 + x2), xmask, eviction_policy='evict_last')
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = 16.0
tmp18 = tmp16 / tmp17
tmp19 = 1e-05
tmp20 = tmp18 + tmp19
tmp21 = libdevice.rsqrt(tmp20)
tmp24 = tmp22 + tmp23
tmp25 = tmp0 - tmp10
tmp26 = tmp25 * tmp21
tmp27 = tmp24 * tmp26
tmp30 = tmp28 + tmp29
tmp31 = tmp27 + tmp30
tl.debug_barrier()
tl.store(in_out_ptr0 + (x0), tmp21, xmask)
tl.store(out_ptr1 + (r1 + (16*x0)), tmp31, xmask)
tl.store(out_ptr0 + (x0), tmp10, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (8, 4), (4, 1))
assert_size_stride(primals_2, (8, ), (1, ))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((8, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [weight], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(primals_1, buf0, 32, grid=grid(32), stream=stream0)
del primals_1
buf1 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(primals_3, reinterpret_tensor(buf0, (4, 8), (1, 4), 0), out=buf1)
buf2 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 1, 1), torch.float32)
buf3 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32)
buf5 = reinterpret_tensor(buf3, (1, 16, 1, 1), (16, 1, 1, 1), 0); del buf3 # reuse
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out, mul_1, out_1], Original ATen: [aten._native_batch_norm_legit, aten.mul, aten.add]
triton_per_fused__native_batch_norm_legit_add_mul_1.run(buf5, primals_4, buf1, primals_2, buf2, buf6, 16, 16, grid=grid(16), stream=stream0)
del buf1
del primals_2
return (buf6, buf0, primals_3, primals_4, buf2, buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((8, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((8, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
from math import sqrt
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Linear(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
linear = nn.Linear(in_dim, out_dim)
linear.weight.data.normal_()
linear.bias.data.zero_()
self.linear = equal_lr(linear)
def forward(self, input):
return self.linear(input)
class AdaIN(nn.Module):
def __init__(self, in_channels, style_dim):
super().__init__()
self.norm = nn.InstanceNorm2d(in_channels)
self.style = Linear(style_dim, in_channels * 2)
self.style.linear.bias.data[:in_channels] = 1
self.style.linear.bias.data[in_channels:] = 0
def forward(self, input, style):
style = self.style(style).unsqueeze(2).unsqueeze(3)
gamma, beta = style.chunk(2, 1)
out = self.norm(input)
out = gamma * out + beta
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'style_dim': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
from math import sqrt
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.7071067811865476
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_per_fused__native_batch_norm_legit_add_mul_1(in_out_ptr0,
in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, xnumel, rnumel, XBLOCK:
tl.constexpr):
xnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
x2 = xindex % 4
x3 = xindex // 4
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp22 = tl.load(in_ptr1 + (x2 + 8 * x3), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr2 + x2, xmask, eviction_policy='evict_last')
tmp28 = tl.load(in_ptr1 + (4 + x2 + 8 * x3), xmask, eviction_policy=
'evict_last')
tmp29 = tl.load(in_ptr2 + (4 + x2), xmask, eviction_policy='evict_last')
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = 16.0
tmp18 = tmp16 / tmp17
tmp19 = 1e-05
tmp20 = tmp18 + tmp19
tmp21 = libdevice.rsqrt(tmp20)
tmp24 = tmp22 + tmp23
tmp25 = tmp0 - tmp10
tmp26 = tmp25 * tmp21
tmp27 = tmp24 * tmp26
tmp30 = tmp28 + tmp29
tmp31 = tmp27 + tmp30
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp21, xmask)
tl.store(out_ptr1 + (r1 + 16 * x0), tmp31, xmask)
tl.store(out_ptr0 + x0, tmp10, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (8, 4), (4, 1))
assert_size_stride(primals_2, (8,), (1,))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((8, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(32)](primals_1, buf0, 32, XBLOCK=32,
num_warps=1, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
extern_kernels.mm(primals_3, reinterpret_tensor(buf0, (4, 8), (1, 4
), 0), out=buf1)
buf2 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 1, 1), torch.float32)
buf3 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32
)
buf5 = reinterpret_tensor(buf3, (1, 16, 1, 1), (16, 1, 1, 1), 0)
del buf3
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_per_fused__native_batch_norm_legit_add_mul_1[grid(16)](buf5,
primals_4, buf1, primals_2, buf2, buf6, 16, 16, XBLOCK=8,
num_warps=2, num_stages=1)
del buf1
del primals_2
return buf6, buf0, primals_3, primals_4, buf2, buf5
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Linear(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
linear = nn.Linear(in_dim, out_dim)
linear.weight.data.normal_()
linear.bias.data.zero_()
self.linear = equal_lr(linear)
def forward(self, input):
return self.linear(input)
class AdaINNew(nn.Module):
def __init__(self, in_channels, style_dim):
super().__init__()
self.norm = nn.InstanceNorm2d(in_channels)
self.style = Linear(style_dim, in_channels * 2)
self.style.linear.bias.data[:in_channels] = 1
self.style.linear.bias.data[in_channels:] = 0
def forward(self, input_0, input_1):
primals_2 = self.style.linear.bias
primals_1 = self.style.linear.weight_orig
primals_4 = input_0
primals_3 = input_1
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
| NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio | AdaIN | false | 883 | [
"MIT"
] | 0 | 231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | https://github.com/NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio/tree/231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | import torch
import torch.nn as nn
from math import sqrt
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Linear(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
linear = nn.Linear(in_dim, out_dim)
linear.weight.data.normal_()
linear.bias.data.zero_()
self.linear = equal_lr(linear)
def forward(self, input):
return self.linear(input)
class Model(nn.Module):
def __init__(self, in_channels, style_dim):
super().__init__()
self.norm = nn.InstanceNorm2d(in_channels)
self.style = Linear(style_dim, in_channels * 2)
self.style.linear.bias.data[:in_channels] = 1
self.style.linear.bias.data[in_channels:] = 0
def forward(self, input, style):
style = self.style(style).unsqueeze(2).unsqueeze(3)
gamma, beta = style.chunk(2, 1)
out = self.norm(input)
out = gamma * out + beta
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [4, 4]
|
GlobalAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/r6/cr6neze6yovkog6kjrk5k2db63h47ozkojywfys6karxe7dlumrz.py
# Topologically Sorted Source Nodes: [align_vectors], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# align_vectors => amax, exp, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%view, [-1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
triton_poi_fused__softmax_0 = async_compile.triton('triton_poi_fused__softmax_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/kj/ckjtlefzavjukjsytvkak6ek26zmzexpcbnlwelx4k5kascjxlf3.py
# Topologically Sorted Source Nodes: [align_vectors], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# align_vectors => div, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_1 = async_compile.triton('triton_poi_fused__softmax_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ip/cip3p4ibqio6uu76ccsemd7wjusq5ptlow3dt2zxzouyuz2sqywf.py
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%bmm_1, %primals_1], 2), kwargs = {})
triton_poi_fused_cat_2 = async_compile.triton('triton_poi_fused_cat_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = (xindex // 8)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x1) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/f5/cf5pnuv5il7avsmzck3quom7r6zvcfuulsdwpzlv2epzfmcgqgwb.py
# Topologically Sorted Source Nodes: [attn_h_2], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# attn_h_2 => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_2,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_3 = async_compile.triton('triton_poi_fused_clone_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_3(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16)
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1)), xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(out_ptr0 + (x3), tmp1, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/u4/cu4fypgfipklcxtitafatnyqdaatx5tws6qfndqotcy4qivcph6d.py
# Topologically Sorted Source Nodes: [align_vectors_2], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# align_vectors_2 => clone_1
# Graph fragment:
# %clone_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_3,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_4 = async_compile.triton('triton_poi_fused_clone_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4) % 4
x2 = (xindex // 16)
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x2) + (16*x1)), xmask)
tl.store(out_ptr0 + (x3), tmp0, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (4, 8), (8, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [align], Original ATen: [aten.bmm]
extern_kernels.bmm(primals_1, reinterpret_tensor(primals_2, (4, 4, 4), (16, 1, 4), 0), out=buf0)
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [align_vectors], Original ATen: [aten._softmax]
stream0 = get_raw_stream(0)
triton_poi_fused__softmax_0.run(buf0, buf1, 64, grid=grid(64), stream=stream0)
buf2 = reinterpret_tensor(buf0, (16, 4), (4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [align_vectors], Original ATen: [aten._softmax]
triton_poi_fused__softmax_1.run(buf1, buf2, 64, grid=grid(64), stream=stream0)
buf3 = reinterpret_tensor(buf1, (4, 4, 4), (16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [c], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf2, (4, 4, 4), (16, 4, 1), 0), primals_2, out=buf3)
del primals_2
buf4 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
triton_poi_fused_cat_2.run(buf3, primals_1, buf4, 128, grid=grid(128), stream=stream0)
del primals_1
buf5 = reinterpret_tensor(buf3, (16, 4), (4, 1), 0); del buf3 # reuse
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf4, (16, 8), (8, 1), 0), reinterpret_tensor(primals_3, (8, 4), (1, 8), 0), out=buf5)
del primals_3
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [attn_h_2], Original ATen: [aten.clone]
triton_poi_fused_clone_3.run(buf5, buf6, 64, grid=grid(64), stream=stream0)
buf7 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [align_vectors_2], Original ATen: [aten.clone]
triton_poi_fused_clone_4.run(buf2, buf7, 64, grid=grid(64), stream=stream0)
del buf2
return (buf6, buf7, reinterpret_tensor(buf4, (16, 8), (8, 1), 0), buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.cuda
import torch.distributed
def aeq(*args):
"""
Assert all arguments have the same value
"""
arguments = (arg for arg in args)
first = next(arguments)
assert all(arg == first for arg in arguments
), 'Not all arguments have the same value: ' + str(args)
def sequence_mask(lengths, max_len=None):
"""
Creates a boolean mask from sequence lengths.
"""
batch_size = lengths.numel()
max_len = max_len or lengths.max()
return torch.arange(0, max_len, device=lengths.device).type_as(lengths
).repeat(batch_size, 1).lt(lengths.unsqueeze(1))
class GlobalAttention(nn.Module):
"""
Global attention takes a matrix and a query vector. It
then computes a parameterized convex combination of the matrix
based on the input query.
Constructs a unit mapping a query `q` of size `dim`
and a source matrix `H` of size `n x dim`, to an output
of size `dim`.
.. mermaid::
graph BT
A[Query]
subgraph RNN
C[H 1]
D[H 2]
E[H N]
end
F[Attn]
G[Output]
A --> F
C --> F
D --> F
E --> F
C -.-> G
D -.-> G
E -.-> G
F --> G
All models compute the output as
:math:`c = \\sum_{j=1}^{\\text{SeqLength}} a_j H_j` where
:math:`a_j` is the softmax of a score function.
Then then apply a projection layer to [q, c].
However they
differ on how they compute the attention score.
* Luong Attention (dot, general):
* dot: :math:`\\text{score}(H_j,q) = H_j^T q`
* general: :math:`\\text{score}(H_j, q) = H_j^T W_a q`
* Bahdanau Attention (mlp):
* :math:`\\text{score}(H_j, q) = v_a^T \\text{tanh}(W_a q + U_a h_j)`
Args:
dim (int): dimensionality of query and key
coverage (bool): use coverage term
attn_type (str): type of attention to use, options [dot,general,mlp]
attn_func (str): attention function to use, options [softmax,sparsemax]
"""
def __init__(self, dim, coverage=False, attn_type='dot', attn_func=
'softmax'):
super(GlobalAttention, self).__init__()
self.dim = dim
assert attn_type in ['dot', 'general', 'mlp'
], 'Please select a valid attention type (got {:s}).'.format(
attn_type)
self.attn_type = attn_type
assert attn_func in ['softmax', 'sparsemax'
], 'Please select a valid attention function.'
self.attn_func = attn_func
if self.attn_type == 'general':
self.linear_in = nn.Linear(dim, dim, bias=False)
elif self.attn_type == 'mlp':
self.linear_context = nn.Linear(dim, dim, bias=False)
self.linear_query = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, 1, bias=False)
out_bias = self.attn_type == 'mlp'
self.linear_out = nn.Linear(dim * 2, dim, bias=out_bias)
if coverage:
self.linear_cover = nn.Linear(1, dim, bias=False)
def score(self, h_t, h_s):
"""
Args:
h_t (FloatTensor): sequence of queries ``(batch, tgt_len, dim)``
h_s (FloatTensor): sequence of sources ``(batch, src_len, dim``
Returns:
FloatTensor: raw attention scores (unnormalized) for each src index
``(batch, tgt_len, src_len)``
"""
src_batch, src_len, src_dim = h_s.size()
tgt_batch, tgt_len, tgt_dim = h_t.size()
aeq(src_batch, tgt_batch)
aeq(src_dim, tgt_dim)
aeq(self.dim, src_dim)
if self.attn_type in ['general', 'dot']:
if self.attn_type == 'general':
h_t_ = h_t.view(tgt_batch * tgt_len, tgt_dim)
h_t_ = self.linear_in(h_t_)
h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim)
h_s_ = h_s.transpose(1, 2)
return torch.bmm(h_t, h_s_)
else:
dim = self.dim
wq = self.linear_query(h_t.view(-1, dim))
wq = wq.view(tgt_batch, tgt_len, 1, dim)
wq = wq.expand(tgt_batch, tgt_len, src_len, dim)
uh = self.linear_context(h_s.contiguous().view(-1, dim))
uh = uh.view(src_batch, 1, src_len, dim)
uh = uh.expand(src_batch, tgt_len, src_len, dim)
wquh = torch.tanh(wq + uh)
return self.v(wquh.view(-1, dim)).view(tgt_batch, tgt_len, src_len)
def forward(self, source, memory_bank, memory_lengths=None, coverage=None):
"""
Args:
source (FloatTensor): query vectors ``(batch, tgt_len, dim)``
memory_bank (FloatTensor): source vectors ``(batch, src_len, dim)``
memory_lengths (LongTensor): the source context lengths ``(batch,)``
coverage (FloatTensor): None (not supported yet)
Returns:
(FloatTensor, FloatTensor):
* Computed vector ``(tgt_len, batch, dim)``
* Attention distribtutions for each query
``(tgt_len, batch, src_len)``
"""
if source.dim() == 2:
one_step = True
source = source.unsqueeze(1)
else:
one_step = False
batch, source_l, dim = memory_bank.size()
batch_, target_l, dim_ = source.size()
aeq(batch, batch_)
aeq(dim, dim_)
aeq(self.dim, dim)
if coverage is not None:
batch_, source_l_ = coverage.size()
aeq(batch, batch_)
aeq(source_l, source_l_)
if coverage is not None:
cover = coverage.view(-1).unsqueeze(1)
memory_bank += self.linear_cover(cover).view_as(memory_bank)
memory_bank = torch.tanh(memory_bank)
align = self.score(source, memory_bank)
if memory_lengths is not None:
mask = sequence_mask(memory_lengths, max_len=align.size(-1))
mask = mask.unsqueeze(1)
align.masked_fill_(~mask, -float('inf'))
if self.attn_func == 'softmax':
align_vectors = F.softmax(align.view(batch * target_l, source_l
), -1)
else:
align_vectors = sparsemax(align.view(batch * target_l, source_l
), -1)
align_vectors = align_vectors.view(batch, target_l, source_l)
c = torch.bmm(align_vectors, memory_bank)
concat_c = torch.cat([c, source], 2).view(batch * target_l, dim * 2)
attn_h = self.linear_out(concat_c).view(batch, target_l, dim)
if self.attn_type in ['general', 'dot']:
attn_h = torch.tanh(attn_h)
if one_step:
attn_h = attn_h.squeeze(1)
align_vectors = align_vectors.squeeze(1)
batch_, dim_ = attn_h.size()
aeq(batch, batch_)
aeq(dim, dim_)
batch_, source_l_ = align_vectors.size()
aeq(batch, batch_)
aeq(source_l, source_l_)
else:
attn_h = attn_h.transpose(0, 1).contiguous()
align_vectors = align_vectors.transpose(0, 1).contiguous()
target_l_, batch_, dim_ = attn_h.size()
aeq(target_l, target_l_)
aeq(batch, batch_)
aeq(dim, dim_)
target_l_, batch_, source_l_ = align_vectors.size()
aeq(target_l, target_l_)
aeq(batch, batch_)
aeq(source_l, source_l_)
return attn_h, align_vectors
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'dim': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_cat_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_clone_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1), xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(out_ptr0 + x3, tmp1, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1), xmask)
tl.store(out_ptr0 + x3, tmp0, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (4, 8), (8, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(primals_1, reinterpret_tensor(primals_2, (4, 4,
4), (16, 1, 4), 0), out=buf0)
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__softmax_0[grid(64)](buf0, buf1, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf2 = reinterpret_tensor(buf0, (16, 4), (4, 1), 0)
del buf0
triton_poi_fused__softmax_1[grid(64)](buf1, buf2, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf3 = reinterpret_tensor(buf1, (4, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf2, (4, 4, 4), (16, 4, 1),
0), primals_2, out=buf3)
del primals_2
buf4 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
triton_poi_fused_cat_2[grid(128)](buf3, primals_1, buf4, 128,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_1
buf5 = reinterpret_tensor(buf3, (16, 4), (4, 1), 0)
del buf3
extern_kernels.mm(reinterpret_tensor(buf4, (16, 8), (8, 1), 0),
reinterpret_tensor(primals_3, (8, 4), (1, 8), 0), out=buf5)
del primals_3
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_clone_3[grid(64)](buf5, buf6, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf7 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_clone_4[grid(64)](buf2, buf7, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf2
return buf6, buf7, reinterpret_tensor(buf4, (16, 8), (8, 1), 0), buf5
def aeq(*args):
"""
Assert all arguments have the same value
"""
arguments = (arg for arg in args)
first = next(arguments)
assert all(arg == first for arg in arguments
), 'Not all arguments have the same value: ' + str(args)
def sequence_mask(lengths, max_len=None):
"""
Creates a boolean mask from sequence lengths.
"""
batch_size = lengths.numel()
max_len = max_len or lengths.max()
return torch.arange(0, max_len, device=lengths.device).type_as(lengths
).repeat(batch_size, 1).lt(lengths.unsqueeze(1))
class GlobalAttentionNew(nn.Module):
"""
Global attention takes a matrix and a query vector. It
then computes a parameterized convex combination of the matrix
based on the input query.
Constructs a unit mapping a query `q` of size `dim`
and a source matrix `H` of size `n x dim`, to an output
of size `dim`.
.. mermaid::
graph BT
A[Query]
subgraph RNN
C[H 1]
D[H 2]
E[H N]
end
F[Attn]
G[Output]
A --> F
C --> F
D --> F
E --> F
C -.-> G
D -.-> G
E -.-> G
F --> G
All models compute the output as
:math:`c = \\sum_{j=1}^{\\text{SeqLength}} a_j H_j` where
:math:`a_j` is the softmax of a score function.
Then then apply a projection layer to [q, c].
However they
differ on how they compute the attention score.
* Luong Attention (dot, general):
* dot: :math:`\\text{score}(H_j,q) = H_j^T q`
* general: :math:`\\text{score}(H_j, q) = H_j^T W_a q`
* Bahdanau Attention (mlp):
* :math:`\\text{score}(H_j, q) = v_a^T \\text{tanh}(W_a q + U_a h_j)`
Args:
dim (int): dimensionality of query and key
coverage (bool): use coverage term
attn_type (str): type of attention to use, options [dot,general,mlp]
attn_func (str): attention function to use, options [softmax,sparsemax]
"""
def __init__(self, dim, coverage=False, attn_type='dot', attn_func=
'softmax'):
super(GlobalAttentionNew, self).__init__()
self.dim = dim
assert attn_type in ['dot', 'general', 'mlp'
], 'Please select a valid attention type (got {:s}).'.format(
attn_type)
self.attn_type = attn_type
assert attn_func in ['softmax', 'sparsemax'
], 'Please select a valid attention function.'
self.attn_func = attn_func
if self.attn_type == 'general':
self.linear_in = nn.Linear(dim, dim, bias=False)
elif self.attn_type == 'mlp':
self.linear_context = nn.Linear(dim, dim, bias=False)
self.linear_query = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, 1, bias=False)
out_bias = self.attn_type == 'mlp'
self.linear_out = nn.Linear(dim * 2, dim, bias=out_bias)
if coverage:
self.linear_cover = nn.Linear(1, dim, bias=False)
def score(self, h_t, h_s):
"""
Args:
h_t (FloatTensor): sequence of queries ``(batch, tgt_len, dim)``
h_s (FloatTensor): sequence of sources ``(batch, src_len, dim``
Returns:
FloatTensor: raw attention scores (unnormalized) for each src index
``(batch, tgt_len, src_len)``
"""
src_batch, src_len, src_dim = h_s.size()
tgt_batch, tgt_len, tgt_dim = h_t.size()
aeq(src_batch, tgt_batch)
aeq(src_dim, tgt_dim)
aeq(self.dim, src_dim)
if self.attn_type in ['general', 'dot']:
if self.attn_type == 'general':
h_t_ = h_t.view(tgt_batch * tgt_len, tgt_dim)
h_t_ = self.linear_in(h_t_)
h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim)
h_s_ = h_s.transpose(1, 2)
return torch.bmm(h_t, h_s_)
else:
dim = self.dim
wq = self.linear_query(h_t.view(-1, dim))
wq = wq.view(tgt_batch, tgt_len, 1, dim)
wq = wq.expand(tgt_batch, tgt_len, src_len, dim)
uh = self.linear_context(h_s.contiguous().view(-1, dim))
uh = uh.view(src_batch, 1, src_len, dim)
uh = uh.expand(src_batch, tgt_len, src_len, dim)
wquh = torch.tanh(wq + uh)
return self.v(wquh.view(-1, dim)).view(tgt_batch, tgt_len, src_len)
def forward(self, input_0, input_1):
primals_3 = self.linear_out.weight
primals_1 = input_0
primals_2 = input_1
output = call([primals_1, primals_2, primals_3])
return output[0], output[1]
| NaomiatLibrary/OpenNMT-kpg-release | GlobalAttention | false | 884 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.cuda
import torch.distributed
def aeq(*args):
"""
Assert all arguments have the same value
"""
arguments = (arg for arg in args)
first = next(arguments)
assert all(arg == first for arg in arguments
), 'Not all arguments have the same value: ' + str(args)
def sequence_mask(lengths, max_len=None):
"""
Creates a boolean mask from sequence lengths.
"""
batch_size = lengths.numel()
max_len = max_len or lengths.max()
return torch.arange(0, max_len, device=lengths.device).type_as(lengths
).repeat(batch_size, 1).lt(lengths.unsqueeze(1))
class Model(nn.Module):
"""
Global attention takes a matrix and a query vector. It
then computes a parameterized convex combination of the matrix
based on the input query.
Constructs a unit mapping a query `q` of size `dim`
and a source matrix `H` of size `n x dim`, to an output
of size `dim`.
.. mermaid::
graph BT
A[Query]
subgraph RNN
C[H 1]
D[H 2]
E[H N]
end
F[Attn]
G[Output]
A --> F
C --> F
D --> F
E --> F
C -.-> G
D -.-> G
E -.-> G
F --> G
All models compute the output as
:math:`c = \\sum_{j=1}^{\\text{SeqLength}} a_j H_j` where
:math:`a_j` is the softmax of a score function.
Then then apply a projection layer to [q, c].
However they
differ on how they compute the attention score.
* Luong Attention (dot, general):
* dot: :math:`\\text{score}(H_j,q) = H_j^T q`
* general: :math:`\\text{score}(H_j, q) = H_j^T W_a q`
* Bahdanau Attention (mlp):
* :math:`\\text{score}(H_j, q) = v_a^T \\text{tanh}(W_a q + U_a h_j)`
Args:
dim (int): dimensionality of query and key
coverage (bool): use coverage term
attn_type (str): type of attention to use, options [dot,general,mlp]
attn_func (str): attention function to use, options [softmax,sparsemax]
"""
def __init__(self, dim, coverage=False, attn_type='dot', attn_func=
'softmax'):
super().__init__()
self.dim = dim
assert attn_type in ['dot', 'general', 'mlp'
], 'Please select a valid attention type (got {:s}).'.format(
attn_type)
self.attn_type = attn_type
assert attn_func in ['softmax', 'sparsemax'
], 'Please select a valid attention function.'
self.attn_func = attn_func
if self.attn_type == 'general':
self.linear_in = nn.Linear(dim, dim, bias=False)
elif self.attn_type == 'mlp':
self.linear_context = nn.Linear(dim, dim, bias=False)
self.linear_query = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, 1, bias=False)
out_bias = self.attn_type == 'mlp'
self.linear_out = nn.Linear(dim * 2, dim, bias=out_bias)
if coverage:
self.linear_cover = nn.Linear(1, dim, bias=False)
def score(self, h_t, h_s):
"""
Args:
h_t (FloatTensor): sequence of queries ``(batch, tgt_len, dim)``
h_s (FloatTensor): sequence of sources ``(batch, src_len, dim``
Returns:
FloatTensor: raw attention scores (unnormalized) for each src index
``(batch, tgt_len, src_len)``
"""
src_batch, src_len, src_dim = h_s.size()
tgt_batch, tgt_len, tgt_dim = h_t.size()
aeq(src_batch, tgt_batch)
aeq(src_dim, tgt_dim)
aeq(self.dim, src_dim)
if self.attn_type in ['general', 'dot']:
if self.attn_type == 'general':
h_t_ = h_t.view(tgt_batch * tgt_len, tgt_dim)
h_t_ = self.linear_in(h_t_)
h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim)
h_s_ = h_s.transpose(1, 2)
return
# ... truncated (>4000 chars) for memory efficiency |
PolicyNetwork | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/mf/cmfr5g3giykxhf4xdxjhcr3mepjprumgz7flrc7pbhcb6obxihdl.py
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.relu]
# Source node to ATen node mapping:
# out => relu
# Graph fragment:
# %add_tensor : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default, %primals_3), kwargs = {})
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%add_tensor,), kwargs = {})
triton_poi_fused_relu_0 = async_compile.triton('triton_poi_fused_relu_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[2048],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 2048
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + (x2), None)
tmp1 = tl.load(in_ptr0 + (x0), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x2), tmp4, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/hz/chz2sqsqk26mwhf2dxhgh44jfpu2er5yqjftwkzfav5ctqtx5e7f.py
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# out_1 => amax, exp, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%addmm_1, [1], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%addmm_1, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
triton_poi_fused__softmax_1 = async_compile.triton('triton_poi_fused__softmax_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/3f/c3fx6bzkalkw7u7askqdnz4rzlcoyqiec4r434sjc5x3axxgkrmr.py
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# out_1 => div, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [1], True), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_2 = async_compile.triton('triton_poi_fused__softmax_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (32, 4), (4, 1))
assert_size_stride(primals_3, (32, ), (1, ))
assert_size_stride(primals_4, (4, 32), (32, 1))
assert_size_stride(primals_5, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 32), (32, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_1, (64, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 32), (1, 4), 0), out=buf0)
del primals_2
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [out], Original ATen: [aten.relu]
stream0 = get_raw_stream(0)
triton_poi_fused_relu_0.run(buf1, primals_3, 2048, grid=grid(2048), stream=stream0)
del primals_3
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, buf1, reinterpret_tensor(primals_4, (32, 4), (1, 32), 0), alpha=1, beta=1, out=buf2)
del primals_5
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten._softmax]
triton_poi_fused__softmax_1.run(buf2, buf3, 256, grid=grid(256), stream=stream0)
buf4 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [out_1], Original ATen: [aten._softmax]
triton_poi_fused__softmax_2.run(buf3, buf4, 256, grid=grid(256), stream=stream0)
del buf3
return (buf4, reinterpret_tensor(primals_1, (64, 4), (4, 1), 0), buf1, buf4, primals_4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((32, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((32, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 32), (32, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
class PolicyNetwork(torch.nn.Module):
def __init__(self, input_size, output_size):
super().__init__()
self.input_size = input_size
self.linear1 = nn.Linear(input_size, 32)
self.linear2 = nn.Linear(32, output_size)
def forward(self, x):
x = x.view((-1, self.input_size))
out = F.relu(self.linear1(x))
out = F.softmax(self.linear2(out), dim=1)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_size': 4, 'output_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (32, 4), (4, 1))
assert_size_stride(primals_3, (32,), (1,))
assert_size_stride(primals_4, (4, 32), (32, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 32), (32, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 32), (1, 4), 0), out=buf0)
del primals_2
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_relu_0[grid(2048)](buf1, primals_3, 2048, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_3
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, buf1, reinterpret_tensor(primals_4,
(32, 4), (1, 32), 0), alpha=1, beta=1, out=buf2)
del primals_5
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(256)](buf2, buf3, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf4 = buf2
del buf2
triton_poi_fused__softmax_2[grid(256)](buf3, buf4, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf3
return buf4, reinterpret_tensor(primals_1, (64, 4), (4, 1), 0
), buf1, buf4, primals_4
class PolicyNetworkNew(torch.nn.Module):
def __init__(self, input_size, output_size):
super().__init__()
self.input_size = input_size
self.linear1 = nn.Linear(input_size, 32)
self.linear2 = nn.Linear(32, output_size)
def forward(self, input_0):
primals_2 = self.linear1.weight
primals_3 = self.linear1.bias
primals_4 = self.linear2.weight
primals_5 = self.linear2.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| NiccoloSacchi/rlcard | PolicyNetwork | false | 885 | [
"MIT"
] | 0 | 046129e8616b12e25652957869a94ab5fd838ae1 | https://github.com/NiccoloSacchi/rlcard/tree/046129e8616b12e25652957869a94ab5fd838ae1 | import torch
import torch.nn as nn
import torch.nn.functional as F
class Model(torch.nn.Module):
def __init__(self, input_size, output_size):
super().__init__()
self.input_size = input_size
self.linear1 = nn.Linear(input_size, 32)
self.linear2 = nn.Linear(32, output_size)
def forward(self, x):
x = x.view((-1, self.input_size))
out = F.relu(self.linear1(x))
out = F.softmax(self.linear2(out), dim=1)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
VdConv1D | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/sb/csbvvvjfzxf5es7wjns765b33dzd6xpxzrgv3hhj5xca64vmwvsf.py
# Topologically Sorted Source Nodes: [exp, mul, sigma], Original ATen: [aten.exp, aten.mul]
# Source node to ATen node mapping:
# exp => exp
# mul => mul
# sigma => mul_1
# Graph fragment:
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%primals_4,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp, %primals_1), kwargs = {})
# %mul_1 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_1), kwargs = {})
triton_poi_fused_exp_mul_0 = async_compile.triton('triton_poi_fused_exp_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_exp_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_exp_mul_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (0))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp3 = tl.load(in_ptr1 + (x0), xmask)
tmp2 = tl_math.exp(tmp1)
tmp4 = tmp2 * tmp3
tmp5 = tmp4 * tmp3
tl.store(out_ptr0 + (x0), tmp5, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/in/cin3u2ol5konbraa6fo43xltih7heuofqc5jodxk3n25bhvs7vtj.py
# Topologically Sorted Source Nodes: [mul_2], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# mul_2 => mul_2
# Graph fragment:
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %primals_2), kwargs = {})
triton_poi_fused_mul_1 = async_compile.triton('triton_poi_fused_mul_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tmp0 * tmp0
tl.store(out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ty/ctypukwtlyaep3dpxr6nglgyxao7kyz3ouk4pzxoyieqoeh54i5e.py
# Topologically Sorted Source Nodes: [add, std, mul_3, out], Original ATen: [aten.add, aten.sqrt, aten.mul]
# Source node to ATen node mapping:
# add => add
# mul_3 => mul_3
# out => add_1
# std => sqrt
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%squeeze_1, 1e-16), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%add,), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sqrt, 0.0), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%squeeze, %mul_3), kwargs = {})
triton_poi_fused_add_mul_sqrt_2 = async_compile.triton('triton_poi_fused_add_mul_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_sqrt_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_sqrt_2(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask)
tmp3 = tl.load(in_ptr1 + (x0), xmask)
tmp2 = tmp0 + tmp1
tmp4 = 1e-16
tmp5 = tmp3 + tmp4
tmp6 = libdevice.sqrt(tmp5)
tmp7 = 0.0
tmp8 = tmp6 * tmp7
tmp9 = tmp2 + tmp8
tl.store(in_out_ptr0 + (x0), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ew/cew35xf34xxwboqmlqqsfqrobstwpzou7opfqptunulkezckblnx.py
# Topologically Sorted Source Nodes: [neg, exp_1, log1p, sum_1, mul_4, kl], Original ATen: [aten.neg, aten.exp, aten.log1p, aten.sum, aten.mul]
# Source node to ATen node mapping:
# exp_1 => exp_1
# kl => mul_5
# log1p => log1p
# mul_4 => mul_4
# neg => neg
# sum_1 => sum_1
# Graph fragment:
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%primals_4,), kwargs = {})
# %exp_1 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%neg,), kwargs = {})
# %log1p : [num_users=1] = call_function[target=torch.ops.aten.log1p.default](args = (%exp_1,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%log1p,), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sum_1, 0.5), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_4, 64.0), kwargs = {})
triton_poi_fused_exp_log1p_mul_neg_sum_3 = async_compile.triton('triton_poi_fused_exp_log1p_mul_neg_sum_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {2: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=(2,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_exp_log1p_mul_neg_sum_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_exp_log1p_mul_neg_sum_3(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + (0))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = -tmp1
tmp3 = tl_math.exp(tmp2)
tmp4 = libdevice.log1p(tmp3)
tmp5 = 0.5
tmp6 = tmp4 * tmp5
tmp7 = 64.0
tmp8 = tmp6 * tmp7
tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32)), tmp8, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (1, 1), (1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [mean], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1, 4, 4), (16, 4, 1), 0), primals_1, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf0, (1, 4, 1), (4, 1, 1))
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [exp, mul, sigma], Original ATen: [aten.exp, aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_exp_mul_0.run(primals_4, primals_1, buf1, 64, grid=grid(64), stream=stream0)
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul_2], Original ATen: [aten.mul]
triton_poi_fused_mul_1.run(primals_2, buf2, 16, grid=grid(16), stream=stream0)
# Topologically Sorted Source Nodes: [conv1d_1], Original ATen: [aten.convolution]
buf3 = extern_kernels.convolution(reinterpret_tensor(buf2, (1, 4, 4), (0, 4, 1), 0), buf1, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf3, (1, 4, 1), (4, 1, 1))
buf4 = reinterpret_tensor(buf0, (4, 1), (1, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [add, std, mul_3, out], Original ATen: [aten.add, aten.sqrt, aten.mul]
triton_poi_fused_add_mul_sqrt_2.run(buf4, primals_3, buf3, 4, grid=grid(4), stream=stream0)
del primals_3
buf5 = empty_strided_cuda((), (), torch.float32)
# Topologically Sorted Source Nodes: [neg, exp_1, log1p, sum_1, mul_4, kl], Original ATen: [aten.neg, aten.exp, aten.log1p, aten.sum, aten.mul]
triton_poi_fused_exp_log1p_mul_neg_sum_3.run(primals_4, buf5, 1, grid=grid(1), stream=stream0)
return (buf4, buf5, primals_1, primals_4, reinterpret_tensor(primals_2, (1, 4, 4), (16, 4, 1), 0), buf1, reinterpret_tensor(buf2, (1, 4, 4), (16, 4, 1), 0), buf3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, 1), (1, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def calculate_kl(log_alpha):
return 0.5 * torch.sum(torch.log1p(torch.exp(-log_alpha)))
class VdConv1D(nn.Module):
"""
Conv1D Layer variational dropout
"""
def __init__(self, in_channels, out_channels, kernel_size, alpha_shape=
(1, 1), bias=True, stride=1, padding=0, dilation=1):
super(VdConv1D, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.dilation = dilation
self.alpha_shape = alpha_shape
self.groups = 1
self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels,
self.kernel_size))
self.log_alpha = nn.Parameter(torch.Tensor(*alpha_shape))
if bias:
self.bias = nn.Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.out_bias = lambda input, kernel: F.conv1d(input, kernel, self.
bias, self.stride, self.padding, self.dilation, self.groups)
self.out_nobias = lambda input, kernel: F.conv1d(input, kernel,
None, self.stride, self.padding, self.dilation, self.groups)
self.reset_parameters()
self.kl_value = calculate_kl
def reset_parameters(self):
n = self.in_channels
n *= self.kernel_size
stdv = 1.0 / math.sqrt(n)
self.weight.data.uniform_(-stdv, stdv)
if self.bias is not None:
self.bias.data.uniform_(-stdv, stdv)
self.log_alpha.data.fill_(-5.0)
def forward(self, x, sample=False):
mean = self.out_bias(x, self.weight)
sigma = torch.exp(self.log_alpha) * self.weight * self.weight
std = torch.sqrt(1e-16 + self.out_nobias(x * x, sigma))
if self.training or sample:
epsilon = std.data.new(std.size()).normal_()
else:
epsilon = 0.0
out = mean + std * epsilon
kl = self.kl_loss()
return out, kl
def kl_loss(self):
return self.weight.nelement() / self.log_alpha.nelement(
) * self.kl_value(self.log_alpha)
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'kernel_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_exp_mul_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp3 = tl.load(in_ptr1 + x0, xmask)
tmp2 = tl_math.exp(tmp1)
tmp4 = tmp2 * tmp3
tmp5 = tmp4 * tmp3
tl.store(out_ptr0 + x0, tmp5, xmask)
@triton.jit
def triton_poi_fused_mul_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tmp0 * tmp0
tl.store(out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_poi_fused_add_mul_sqrt_2(in_out_ptr0, in_ptr0, in_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask)
tmp3 = tl.load(in_ptr1 + x0, xmask)
tmp2 = tmp0 + tmp1
tmp4 = 1e-16
tmp5 = tmp3 + tmp4
tmp6 = libdevice.sqrt(tmp5)
tmp7 = 0.0
tmp8 = tmp6 * tmp7
tmp9 = tmp2 + tmp8
tl.store(in_out_ptr0 + x0, tmp9, xmask)
@triton.jit
def triton_poi_fused_exp_log1p_mul_neg_sum_3(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = -tmp1
tmp3 = tl_math.exp(tmp2)
tmp4 = libdevice.log1p(tmp3)
tmp5 = 0.5
tmp6 = tmp4 * tmp5
tmp7 = 64.0
tmp8 = tmp6 * tmp7
tl.store(out_ptr0 + tl.full([XBLOCK], 0, tl.int32), tmp8, None)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (1, 1), (1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1,
4, 4), (16, 4, 1), 0), primals_1, stride=(1,), padding=(0,),
dilation=(1,), transposed=False, output_padding=(0,), groups=1,
bias=None)
assert_size_stride(buf0, (1, 4, 1), (4, 1, 1))
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_exp_mul_0[grid(64)](primals_4, primals_1, buf1, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_mul_1[grid(16)](primals_2, buf2, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf3 = extern_kernels.convolution(reinterpret_tensor(buf2, (1, 4, 4
), (0, 4, 1), 0), buf1, stride=(1,), padding=(0,), dilation=(1,
), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf3, (1, 4, 1), (4, 1, 1))
buf4 = reinterpret_tensor(buf0, (4, 1), (1, 1), 0)
del buf0
triton_poi_fused_add_mul_sqrt_2[grid(4)](buf4, primals_3, buf3, 4,
XBLOCK=4, num_warps=1, num_stages=1)
del primals_3
buf5 = empty_strided_cuda((), (), torch.float32)
triton_poi_fused_exp_log1p_mul_neg_sum_3[grid(1)](primals_4, buf5,
1, XBLOCK=1, num_warps=1, num_stages=1)
return buf4, buf5, primals_1, primals_4, reinterpret_tensor(primals_2,
(1, 4, 4), (16, 4, 1), 0), buf1, reinterpret_tensor(buf2, (1, 4, 4),
(16, 4, 1), 0), buf3
def calculate_kl(log_alpha):
return 0.5 * torch.sum(torch.log1p(torch.exp(-log_alpha)))
class VdConv1DNew(nn.Module):
"""
Conv1D Layer variational dropout
"""
def __init__(self, in_channels, out_channels, kernel_size, alpha_shape=
(1, 1), bias=True, stride=1, padding=0, dilation=1):
super(VdConv1DNew, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.dilation = dilation
self.alpha_shape = alpha_shape
self.groups = 1
self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels,
self.kernel_size))
self.log_alpha = nn.Parameter(torch.Tensor(*alpha_shape))
if bias:
self.bias = nn.Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.out_bias = lambda input, kernel: F.conv1d(input, kernel, self.
bias, self.stride, self.padding, self.dilation, self.groups)
self.out_nobias = lambda input, kernel: F.conv1d(input, kernel,
None, self.stride, self.padding, self.dilation, self.groups)
self.reset_parameters()
self.kl_value = calculate_kl
def reset_parameters(self):
n = self.in_channels
n *= self.kernel_size
stdv = 1.0 / math.sqrt(n)
self.weight.data.uniform_(-stdv, stdv)
if self.bias is not None:
self.bias.data.uniform_(-stdv, stdv)
self.log_alpha.data.fill_(-5.0)
def kl_loss(self):
return self.weight.nelement() / self.log_alpha.nelement(
) * self.kl_value(self.log_alpha)
def forward(self, input_0):
primals_1 = self.weight
primals_4 = self.log_alpha
primals_3 = self.bias
primals_2 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0], output[1]
| Neronjust2017/pytorch-classification-project | VdConv1D | false | 886 | [
"MIT"
] | 0 | fc5f4d7c46d071765f682ce20e6580646d4e5c76 | https://github.com/Neronjust2017/pytorch-classification-project/tree/fc5f4d7c46d071765f682ce20e6580646d4e5c76 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def calculate_kl(log_alpha):
return 0.5 * torch.sum(torch.log1p(torch.exp(-log_alpha)))
class Model(nn.Module):
"""
Conv1D Layer variational dropout
"""
def __init__(self, in_channels, out_channels, kernel_size, alpha_shape=
(1, 1), bias=True, stride=1, padding=0, dilation=1):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.dilation = dilation
self.alpha_shape = alpha_shape
self.groups = 1
self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels,
self.kernel_size))
self.log_alpha = nn.Parameter(torch.Tensor(*alpha_shape))
if bias:
self.bias = nn.Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.out_bias = lambda input, kernel: F.conv1d(input, kernel, self.
bias, self.stride, self.padding, self.dilation, self.groups)
self.out_nobias = lambda input, kernel: F.conv1d(input, kernel,
None, self.stride, self.padding, self.dilation, self.groups)
self.reset_parameters()
self.kl_value = calculate_kl
def reset_parameters(self):
n = self.in_channels
n *= self.kernel_size
stdv = 1.0 / math.sqrt(n)
self.weight.data.uniform_(-stdv, stdv)
if self.bias is not None:
self.bias.data.uniform_(-stdv, stdv)
self.log_alpha.data.fill_(-5.0)
def forward(self, x, sample=False):
mean = self.out_bias(x, self.weight)
sigma = torch.exp(self.log_alpha) * self.weight * self.weight
std = torch.sqrt(1e-16 + self.out_nobias(x * x, sigma))
if self.training or sample:
epsilon = std.data.new(std.size()).normal_()
else:
epsilon = 0.0
out = mean + std * epsilon
kl = self.kl_loss()
return out, kl
def kl_loss(self):
return self.weight.nelement() / self.log_alpha.nelement(
) * self.kl_value(self.log_alpha)
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
SpatialAttention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/46/c46mg7rvdztu6n5oosf5c4if7ziag6obrxhwbn43lcdfibfuom7w.py
# Topologically Sorted Source Nodes: [feature], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# feature => cat
# Graph fragment:
# %cat : [num_users=2] = call_function[target=torch.ops.aten.cat.default](args = ([%mean, %getitem], 1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 16) % 2
x0 = xindex % 16
x2 = (xindex // 32)
x3 = xindex
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + (64*x2)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp7 = tmp5 + tmp6
tmp8 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp9 = tmp7 + tmp8
tmp10 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp11 = tmp9 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = tl.full(tmp13.shape, 0.0, tmp13.dtype)
tmp15 = tl.where(tmp4, tmp13, tmp14)
tmp16 = tmp0 >= tmp3
tmp17 = tl.full([1], 2, tl.int64)
tmp18 = tmp0 < tmp17
tmp19 = tl.load(in_ptr0 + (x0 + (64*x2)), tmp16 & xmask, eviction_policy='evict_last', other=0.0)
tmp20 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), tmp16 & xmask, eviction_policy='evict_last', other=0.0)
tmp21 = triton_helpers.maximum(tmp19, tmp20)
tmp22 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), tmp16 & xmask, eviction_policy='evict_last', other=0.0)
tmp23 = triton_helpers.maximum(tmp21, tmp22)
tmp24 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), tmp16 & xmask, eviction_policy='evict_last', other=0.0)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp26 = tl.full(tmp25.shape, 0.0, tmp25.dtype)
tmp27 = tl.where(tmp16, tmp25, tmp26)
tmp28 = tl.where(tmp4, tmp15, tmp27)
tl.store(out_ptr0 + (x3), tmp28, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2s/c2s2pec3vduo4xn2kfu53hypzbhir2ql56lmvazfmymhwv2ehhv5.py
# Topologically Sorted Source Nodes: [sigmoid], Original ATen: [aten.sigmoid]
# Source node to ATen node mapping:
# sigmoid => sigmoid
# Graph fragment:
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%convolution,), kwargs = {})
triton_poi_fused_sigmoid_1 = async_compile.triton('triton_poi_fused_sigmoid_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sigmoid_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_sigmoid_1(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
tmp1 = tl.sigmoid(tmp0)
tl.store(in_out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 2, 7, 7), (98, 49, 7, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [feature], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, buf0, 128, grid=grid(128), stream=stream0)
del primals_1
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1, 1), padding=(3, 3), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 1, 4, 4), (16, 16, 4, 1))
buf2 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [sigmoid], Original ATen: [aten.sigmoid]
triton_poi_fused_sigmoid_1.run(buf2, 64, grid=grid(64), stream=stream0)
return (buf2, primals_2, buf0, buf2, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((1, 2, 7, 7), (98, 49, 7, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.parallel
import torch.utils.data
import torch.nn as nn
import torch.cuda
class SpatialAttention(nn.Module):
def __init__(self):
super(SpatialAttention, self).__init__()
self.conv = nn.Conv2d(2, 1, 7, padding=3, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, x: 'torch.Tensor') ->torch.Tensor:
feat_avg = torch.mean(x, dim=1, keepdim=True)
feat_max = torch.max(x, dim=1, keepdim=True)[0]
feature = torch.cat((feat_avg, feat_max), dim=1)
return self.sigmoid(self.conv(feature))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn.parallel
import torch.utils.data
import torch.nn as nn
import torch.cuda
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 16 % 2
x0 = xindex % 16
x2 = xindex // 32
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 64 * x2), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp7 = tmp5 + tmp6
tmp8 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp9 = tmp7 + tmp8
tmp10 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp9 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = tl.full(tmp13.shape, 0.0, tmp13.dtype)
tmp15 = tl.where(tmp4, tmp13, tmp14)
tmp16 = tmp0 >= tmp3
tl.full([1], 2, tl.int64)
tmp19 = tl.load(in_ptr0 + (x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp20 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp21 = triton_helpers.maximum(tmp19, tmp20)
tmp22 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp23 = triton_helpers.maximum(tmp21, tmp22)
tmp24 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp26 = tl.full(tmp25.shape, 0.0, tmp25.dtype)
tmp27 = tl.where(tmp16, tmp25, tmp26)
tmp28 = tl.where(tmp4, tmp15, tmp27)
tl.store(out_ptr0 + x3, tmp28, xmask)
@triton.jit
def triton_poi_fused_sigmoid_1(in_out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.sigmoid(tmp0)
tl.store(in_out_ptr0 + x0, tmp1, xmask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 2, 7, 7), (98, 49, 7, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(128)](primals_1, buf0, 128, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_1
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1, 1),
padding=(3, 3), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 1, 4, 4), (16, 16, 4, 1))
buf2 = buf1
del buf1
triton_poi_fused_sigmoid_1[grid(64)](buf2, 64, XBLOCK=64, num_warps
=1, num_stages=1)
return buf2, primals_2, buf0, buf2
class SpatialAttentionNew(nn.Module):
def __init__(self):
super(SpatialAttentionNew, self).__init__()
self.conv = nn.Conv2d(2, 1, 7, padding=3, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, input_0):
primals_2 = self.conv.weight
primals_1 = input_0
output = call([primals_1, primals_2])
return output[0]
| NeilDG/NeuralNets-Experiment3 | SpatialAttention | false | 887 | [
"MIT"
] | 0 | f0d2f788eeca49f803f65810c155491ce687cf9e | https://github.com/NeilDG/NeuralNets-Experiment3/tree/f0d2f788eeca49f803f65810c155491ce687cf9e | import torch
import torch.nn.parallel
import torch.utils.data
import torch.nn as nn
import torch.cuda
class Model(nn.Module):
def __init__(self):
super().__init__()
self.conv = nn.Conv2d(2, 1, 7, padding=3, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, x: 'torch.Tensor') ->torch.Tensor:
feat_avg = torch.mean(x, dim=1, keepdim=True)
feat_max = torch.max(x, dim=1, keepdim=True)[0]
feature = torch.cat((feat_avg, feat_max), dim=1)
return self.sigmoid(self.conv(feature))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
Loss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/dy/cdye32zdr3cicbbaex3vhj3rludh22pvzjjm43zjdv7rkt2htpyv.py
# Topologically Sorted Source Nodes: [sub_5, abs_1, zeros, sub_3, sign, gamma_U_hard, sub_4, sign_1, gamma_L_hard, gamma_hard, mul_4, mean_2, mean_3, add, qd_lhs_hard, zero, sub_1, mul, gamma_U, sub_2, mul_1, gamma_L, gamma_, PICP_soft, sub_7, max_3, pow_1, qd_rhs_soft, qd_loss_soft, PICP, sub, MPIW, tensor_1, y_var_limited, exp, add_4, log, tensor_2, y_var], Original ATen: [aten.sub, aten.abs, aten.zeros_like, aten.sign, aten.maximum, aten.mul, aten.mean, aten.add, aten.div, aten.lift_fresh, aten.sigmoid, aten.rsub, aten.pow, aten.minimum, aten.exp, aten.log]
# Source node to ATen node mapping:
# MPIW => mean_1
# PICP => mean_11
# PICP_soft => mean_6
# abs_1 => abs_1
# add => add
# add_4 => add_4
# exp => exp
# gamma_ => mul_2
# gamma_L => sigmoid_1
# gamma_L_hard => maximum_1
# gamma_U => sigmoid
# gamma_U_hard => maximum
# gamma_hard => mul_3
# log => log
# max_3 => maximum_2
# mean_2 => mean_2
# mean_3 => mean_3
# mul => mul
# mul_1 => mul_1
# mul_4 => mul_4
# pow_1 => pow_1
# qd_lhs_hard => div
# qd_loss_soft => add_2
# qd_rhs_soft => mul_6
# sign => sign
# sign_1 => sign_1
# sub => sub
# sub_1 => sub_1
# sub_2 => sub_2
# sub_3 => sub_3
# sub_4 => sub_4
# sub_5 => sub_5
# sub_7 => sub_7
# tensor_1 => full_default_2
# tensor_2 => full_default_3
# y_var => maximum_4
# y_var_limited => minimum
# zero => full_default_1
# zeros => full_default
# Graph fragment:
# %sub_5 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select, %select_1), kwargs = {})
# %abs_1 : [num_users=1] = call_function[target=torch.ops.aten.abs.default](args = (%sub_5,), kwargs = {})
# %full_default : [num_users=2] = call_function[target=torch.ops.aten.full.default](args = ([4, 4, 4], 0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %sub_3 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select, %select_2), kwargs = {})
# %sign : [num_users=1] = call_function[target=torch.ops.aten.sign.default](args = (%sub_3,), kwargs = {})
# %maximum : [num_users=1] = call_function[target=torch.ops.aten.maximum.default](args = (%full_default, %sign), kwargs = {})
# %sub_4 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select_2, %select_1), kwargs = {})
# %sign_1 : [num_users=1] = call_function[target=torch.ops.aten.sign.default](args = (%sub_4,), kwargs = {})
# %maximum_1 : [num_users=1] = call_function[target=torch.ops.aten.maximum.default](args = (%full_default, %sign_1), kwargs = {})
# %mul_3 : [num_users=3] = call_function[target=torch.ops.aten.mul.Tensor](args = (%maximum, %maximum_1), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%abs_1, %mul_3), kwargs = {})
# %mean_2 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%mul_4,), kwargs = {})
# %mean_3 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%mul_3,), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean_3, 0.001), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mean_2, %add), kwargs = {})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cpu, pin_memory: False})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select, %select_2), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_1, 100.0), kwargs = {})
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%mul,), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select_2, %select_1), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_2, 100.0), kwargs = {})
# %sigmoid_1 : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%mul_1,), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %sigmoid_1), kwargs = {})
# %mean_6 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%mul_2,), kwargs = {})
# %sub_7 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (0.9, %mean_6), kwargs = {})
# %maximum_2 : [num_users=1] = call_function[target=torch.ops.aten.maximum.default](args = (%full_default_1, %sub_7), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%maximum_2, 2), kwargs = {})
# %mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%pow_1, 20.0), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %mul_6), kwargs = {})
# %mean_11 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%mul_3,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select, %select_1), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%sub,), kwargs = {})
# %full_default_2 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 10.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cpu, pin_memory: False})
# %minimum : [num_users=1] = call_function[target=torch.ops.aten.minimum.default](args = (%select_1, %full_default_2), kwargs = {})
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%minimum,), kwargs = {})
# %add_4 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%exp, 1.0), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%add_4,), kwargs = {})
# %full_default_3 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 9.999999747378752e-06), kwargs = {dtype: torch.float32, layout: torch.strided, device: cpu, pin_memory: False})
# %maximum_4 : [num_users=1] = call_function[target=torch.ops.aten.maximum.default](args = (%log, %full_default_3), kwargs = {})
triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0 = async_compile.triton('triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 64],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {6: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 7), equal_to_1=(6,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0', 'mutated_arg_names': ['in_out_ptr0', 'in_out_ptr1', 'in_out_ptr2'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 5, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0(in_out_ptr0, in_out_ptr1, in_out_ptr2, in_ptr0, in_ptr1, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 64
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = (rindex // 16)
r2 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + (64*r1)), None)
tmp1 = tl.load(in_ptr0 + (16 + r0 + (64*r1)), None)
tmp4 = tl.load(in_ptr1 + (r0 + (64*r1)), None)
tmp2 = tmp0 - tmp1
tmp3 = tl_math.abs(tmp2)
tmp5 = tmp0 - tmp4
tmp6 = tl.full([1, 1], 0, tl.int32)
tmp7 = tmp6 < tmp5
tmp8 = tmp7.to(tl.int8)
tmp9 = tmp5 < tmp6
tmp10 = tmp9.to(tl.int8)
tmp11 = tmp8 - tmp10
tmp12 = tmp11.to(tmp5.dtype)
tmp13 = 0.0
tmp14 = triton_helpers.maximum(tmp13, tmp12)
tmp15 = tmp4 - tmp1
tmp16 = tmp6 < tmp15
tmp17 = tmp16.to(tl.int8)
tmp18 = tmp15 < tmp6
tmp19 = tmp18.to(tl.int8)
tmp20 = tmp17 - tmp19
tmp21 = tmp20.to(tmp15.dtype)
tmp22 = triton_helpers.maximum(tmp13, tmp21)
tmp23 = tmp14 * tmp22
tmp24 = tmp3 * tmp23
tmp25 = tl.broadcast_to(tmp24, [XBLOCK, RBLOCK])
tmp27 = tl.sum(tmp25, 1)[:, None]
tmp28 = tl.broadcast_to(tmp23, [XBLOCK, RBLOCK])
tmp30 = tl.sum(tmp28, 1)[:, None]
tmp31 = 100.0
tmp32 = tmp5 * tmp31
tmp33 = tl.sigmoid(tmp32)
tmp34 = tmp15 * tmp31
tmp35 = tl.sigmoid(tmp34)
tmp36 = tmp33 * tmp35
tmp37 = tl.broadcast_to(tmp36, [XBLOCK, RBLOCK])
tmp39 = tl.sum(tmp37, 1)[:, None]
tmp40 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp42 = tl.sum(tmp40, 1)[:, None]
tmp43 = 10.0
tmp44 = triton_helpers.minimum(tmp1, tmp43)
tmp45 = tl_math.exp(tmp44)
tmp46 = 1.0
tmp47 = tmp45 + tmp46
tmp48 = tl_math.log(tmp47)
tmp49 = 9.999999747378752e-06
tmp50 = triton_helpers.maximum(tmp48, tmp49)
tmp51 = 64.0
tmp52 = tmp27 / tmp51
tmp53 = tmp30 / tmp51
tmp54 = 0.001
tmp55 = tmp53 + tmp54
tmp56 = tmp52 / tmp55
tmp57 = tmp39 / tmp51
tmp58 = 0.9
tmp59 = tmp58 - tmp57
tmp60 = triton_helpers.maximum(tmp13, tmp59)
tmp61 = tmp60 * tmp60
tmp62 = 20.0
tmp63 = tmp61 * tmp62
tmp64 = tmp56 + tmp63
tmp65 = tmp42 / tmp51
tl.store(out_ptr2 + (tl.broadcast_to(r2, [XBLOCK, RBLOCK])), tmp50, None)
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp64, None)
tl.debug_barrier()
tl.store(in_out_ptr1 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp53, None)
tl.debug_barrier()
tl.store(in_out_ptr2 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp65, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf3 = empty_strided_cuda((), (), torch.float32)
buf4 = empty_strided_cuda((), (), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf6 = buf0; del buf0 # reuse
buf7 = buf3; del buf3 # reuse
buf8 = buf4; del buf4 # reuse
# Topologically Sorted Source Nodes: [sub_5, abs_1, zeros, sub_3, sign, gamma_U_hard, sub_4, sign_1, gamma_L_hard, gamma_hard, mul_4, mean_2, mean_3, add, qd_lhs_hard, zero, sub_1, mul, gamma_U, sub_2, mul_1, gamma_L, gamma_, PICP_soft, sub_7, max_3, pow_1, qd_rhs_soft, qd_loss_soft, PICP, sub, MPIW, tensor_1, y_var_limited, exp, add_4, log, tensor_2, y_var], Original ATen: [aten.sub, aten.abs, aten.zeros_like, aten.sign, aten.maximum, aten.mul, aten.mean, aten.add, aten.div, aten.lift_fresh, aten.sigmoid, aten.rsub, aten.pow, aten.minimum, aten.exp, aten.log]
stream0 = get_raw_stream(0)
triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0.run(buf6, buf7, buf8, arg0_1, arg1_1, buf5, 1, 64, grid=grid(1), stream=stream0)
del arg1_1
return (buf6, buf7, buf8, buf5, reinterpret_tensor(arg0_1, (4, 4, 4), (64, 4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn as nn
class Loss(nn.Module):
def __init__(self, device, type_in='pred_intervals', alpha=0.1,
loss_type='qd_soft', censor_R=False, soften=100.0, lambda_in=10.0,
sigma_in=0.5):
super().__init__()
self.alpha = alpha
self.lambda_in = lambda_in
self.soften = soften
self.loss_type = loss_type
self.type_in = type_in
self.censor_R = censor_R
self.sigma_in = sigma_in
self.device = device
def forward(self, y_pred, y_true):
if self.type_in == 'pred_intervals':
metric = []
metric_name = []
y_U = y_pred[:, 0]
y_L = y_pred[:, 1]
y_T = y_true[:, 0]
N_ = y_T.shape[0]
alpha_ = self.alpha
lambda_ = self.lambda_in
torch.mean(y_pred, dim=1)
MPIW = torch.mean(y_U - y_L)
gamma_U = torch.sigmoid((y_U - y_T) * self.soften)
gamma_L = torch.sigmoid((y_T - y_L) * self.soften)
gamma_ = torch.mul(gamma_U, gamma_L)
torch.ones_like(gamma_)
zeros = torch.zeros_like(y_U)
gamma_U_hard = torch.max(zeros, torch.sign(y_U - y_T))
gamma_L_hard = torch.max(zeros, torch.sign(y_T - y_L))
gamma_hard = torch.mul(gamma_U_hard, gamma_L_hard)
qd_lhs_hard = torch.div(torch.mean(torch.abs(y_U - y_L) *
gamma_hard), torch.mean(gamma_hard) + 0.001)
torch.div(torch.mean(torch.abs(y_U - y_L) * gamma_), torch.mean
(gamma_) + 0.001)
PICP_soft = torch.mean(gamma_)
PICP_hard = torch.mean(gamma_hard)
zero = torch.tensor(0.0)
qd_rhs_soft = lambda_ * math.sqrt(N_) * torch.pow(torch.max(
zero, 1.0 - alpha_ - PICP_soft), 2)
qd_rhs_hard = lambda_ * math.sqrt(N_) * torch.pow(torch.max(
zero, 1.0 - alpha_ - PICP_hard), 2)
qd_loss_soft = qd_lhs_hard + qd_rhs_soft
qd_loss_hard = qd_lhs_hard + qd_rhs_hard
y_mean = y_U
y_var_limited = torch.min(y_L, torch.tensor(10.0))
y_var = torch.max(torch.log(1.0 + torch.exp(y_var_limited)),
torch.tensor(1e-05))
self.y_mean = y_mean
self.y_var = y_var
gauss_loss = torch.log(y_var) / 2.0 + torch.div(torch.pow(y_T -
y_mean, 2), 2.0 * y_var)
gauss_loss = torch.mean(gauss_loss)
if self.loss_type == 'qd_soft':
loss = qd_loss_soft
elif self.loss_type == 'qd_hard':
loss = qd_loss_hard
elif self.loss_type == 'gauss_like':
loss = gauss_loss
elif self.loss_type == 'picp':
loss = PICP_hard
elif self.loss_type == 'mse':
loss = torch.mean(torch.pow(y_U - y_T, 2))
torch.mean(gamma_U_hard)
torch.mean(gamma_L_hard)
PICP = torch.mean(gamma_hard)
metric.append(PICP)
metric_name.append('PICP')
metric.append(MPIW)
metric_name.append('MPIW')
return loss, PICP, MPIW
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'device': 0}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0(
in_out_ptr0, in_out_ptr1, in_out_ptr2, in_ptr0, in_ptr1, out_ptr2,
xnumel, rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = rindex // 16
r2 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + 64 * r1), None)
tmp1 = tl.load(in_ptr0 + (16 + r0 + 64 * r1), None)
tmp4 = tl.load(in_ptr1 + (r0 + 64 * r1), None)
tmp2 = tmp0 - tmp1
tmp3 = tl_math.abs(tmp2)
tmp5 = tmp0 - tmp4
tmp6 = tl.full([1, 1], 0, tl.int32)
tmp7 = tmp6 < tmp5
tmp8 = tmp7.to(tl.int8)
tmp9 = tmp5 < tmp6
tmp10 = tmp9.to(tl.int8)
tmp11 = tmp8 - tmp10
tmp12 = tmp11.to(tmp5.dtype)
tmp13 = 0.0
tmp14 = triton_helpers.maximum(tmp13, tmp12)
tmp15 = tmp4 - tmp1
tmp16 = tmp6 < tmp15
tmp17 = tmp16.to(tl.int8)
tmp18 = tmp15 < tmp6
tmp19 = tmp18.to(tl.int8)
tmp20 = tmp17 - tmp19
tmp21 = tmp20.to(tmp15.dtype)
tmp22 = triton_helpers.maximum(tmp13, tmp21)
tmp23 = tmp14 * tmp22
tmp24 = tmp3 * tmp23
tmp25 = tl.broadcast_to(tmp24, [XBLOCK, RBLOCK])
tmp27 = tl.sum(tmp25, 1)[:, None]
tmp28 = tl.broadcast_to(tmp23, [XBLOCK, RBLOCK])
tmp30 = tl.sum(tmp28, 1)[:, None]
tmp31 = 100.0
tmp32 = tmp5 * tmp31
tmp33 = tl.sigmoid(tmp32)
tmp34 = tmp15 * tmp31
tmp35 = tl.sigmoid(tmp34)
tmp36 = tmp33 * tmp35
tmp37 = tl.broadcast_to(tmp36, [XBLOCK, RBLOCK])
tmp39 = tl.sum(tmp37, 1)[:, None]
tmp40 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp42 = tl.sum(tmp40, 1)[:, None]
tmp43 = 10.0
tmp44 = triton_helpers.minimum(tmp1, tmp43)
tmp45 = tl_math.exp(tmp44)
tmp46 = 1.0
tmp47 = tmp45 + tmp46
tmp48 = tl_math.log(tmp47)
tmp49 = 9.999999747378752e-06
tmp50 = triton_helpers.maximum(tmp48, tmp49)
tmp51 = 64.0
tmp52 = tmp27 / tmp51
tmp53 = tmp30 / tmp51
tmp54 = 0.001
tmp55 = tmp53 + tmp54
tmp56 = tmp52 / tmp55
tmp57 = tmp39 / tmp51
tmp58 = 0.9
tmp59 = tmp58 - tmp57
tmp60 = triton_helpers.maximum(tmp13, tmp59)
tmp61 = tmp60 * tmp60
tmp62 = 20.0
tmp63 = tmp61 * tmp62
tmp64 = tmp56 + tmp63
tmp65 = tmp42 / tmp51
tl.store(out_ptr2 + tl.broadcast_to(r2, [XBLOCK, RBLOCK]), tmp50, None)
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp64, None)
tl.debug_barrier()
tl.store(in_out_ptr1 + tl.full([XBLOCK, 1], 0, tl.int32), tmp53, None)
tl.debug_barrier()
tl.store(in_out_ptr2 + tl.full([XBLOCK, 1], 0, tl.int32), tmp65, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf3 = empty_strided_cuda((), (), torch.float32)
buf4 = empty_strided_cuda((), (), torch.float32)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf6 = buf0
del buf0
buf7 = buf3
del buf3
buf8 = buf4
del buf4
get_raw_stream(0)
triton_per_fused_abs_add_div_exp_lift_fresh_log_maximum_mean_minimum_mul_pow_rsub_sigmoid_sign_sub_zeros_like_0[
grid(1)](buf6, buf7, buf8, arg0_1, arg1_1, buf5, 1, 64, XBLOCK=
1, num_warps=2, num_stages=1)
del arg1_1
return buf6, buf7, buf8, buf5, reinterpret_tensor(arg0_1, (4, 4, 4), (
64, 4, 1), 0)
class LossNew(nn.Module):
def __init__(self, device, type_in='pred_intervals', alpha=0.1,
loss_type='qd_soft', censor_R=False, soften=100.0, lambda_in=10.0,
sigma_in=0.5):
super().__init__()
self.alpha = alpha
self.lambda_in = lambda_in
self.soften = soften
self.loss_type = loss_type
self.type_in = type_in
self.censor_R = censor_R
self.sigma_in = sigma_in
self.device = device
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0], output[1], output[2]
| Neronjust2017/pytorch-classification-project | Loss | false | 888 | [
"MIT"
] | 0 | fc5f4d7c46d071765f682ce20e6580646d4e5c76 | https://github.com/Neronjust2017/pytorch-classification-project/tree/fc5f4d7c46d071765f682ce20e6580646d4e5c76 | import math
import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, device, type_in='pred_intervals', alpha=0.1,
loss_type='qd_soft', censor_R=False, soften=100.0, lambda_in=10.0,
sigma_in=0.5):
super().__init__()
self.alpha = alpha
self.lambda_in = lambda_in
self.soften = soften
self.loss_type = loss_type
self.type_in = type_in
self.censor_R = censor_R
self.sigma_in = sigma_in
self.device = device
def forward(self, y_pred, y_true):
if self.type_in == 'pred_intervals':
metric = []
metric_name = []
y_U = y_pred[:, 0]
y_L = y_pred[:, 1]
y_T = y_true[:, 0]
N_ = y_T.shape[0]
alpha_ = self.alpha
lambda_ = self.lambda_in
torch.mean(y_pred, dim=1)
MPIW = torch.mean(y_U - y_L)
gamma_U = torch.sigmoid((y_U - y_T) * self.soften)
gamma_L = torch.sigmoid((y_T - y_L) * self.soften)
gamma_ = torch.mul(gamma_U, gamma_L)
torch.ones_like(gamma_)
zeros = torch.zeros_like(y_U)
gamma_U_hard = torch.max(zeros, torch.sign(y_U - y_T))
gamma_L_hard = torch.max(zeros, torch.sign(y_T - y_L))
gamma_hard = torch.mul(gamma_U_hard, gamma_L_hard)
qd_lhs_hard = torch.div(torch.mean(torch.abs(y_U - y_L) *
gamma_hard), torch.mean(gamma_hard) + 0.001)
torch.div(torch.mean(torch.abs(y_U - y_L) * gamma_), torch.mean
(gamma_) + 0.001)
PICP_soft = torch.mean(gamma_)
PICP_hard = torch.mean(gamma_hard)
zero = torch.tensor(0.0)
qd_rhs_soft = lambda_ * math.sqrt(N_) * torch.pow(torch.max(
zero, 1.0 - alpha_ - PICP_soft), 2)
qd_rhs_hard = lambda_ * math.sqrt(N_) * torch.pow(torch.max(
zero, 1.0 - alpha_ - PICP_hard), 2)
qd_loss_soft = qd_lhs_hard + qd_rhs_soft
qd_loss_hard = qd_lhs_hard + qd_rhs_hard
y_mean = y_U
y_var_limited = torch.min(y_L, torch.tensor(10.0))
y_var = torch.max(torch.log(1.0 + torch.exp(y_var_limited)),
torch.tensor(1e-05))
self.y_mean = y_mean
self.y_var = y_var
gauss_loss = torch.log(y_var) / 2.0 + torch.div(torch.pow(y_T -
y_mean, 2), 2.0 * y_var)
gauss_loss = torch.mean(gauss_loss)
if self.loss_type == 'qd_soft':
loss = qd_loss_soft
elif self.loss_type == 'qd_hard':
loss = qd_loss_hard
elif self.loss_type == 'gauss_like':
loss = gauss_loss
elif self.loss_type == 'picp':
loss = PICP_hard
elif self.loss_type == 'mse':
loss = torch.mean(torch.pow(y_U - y_T, 2))
torch.mean(gamma_U_hard)
torch.mean(gamma_L_hard)
PICP = torch.mean(gamma_hard)
metric.append(PICP)
metric_name.append('PICP')
metric.append(MPIW)
metric_name.append('MPIW')
return loss, PICP, MPIW
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [0]
|
Linear | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/oy/coy4v6ev22tv33nc6asaz3obrskaw2f3vho4q3aj4yqpth7c2y2m.py
# Topologically Sorted Source Nodes: [weight], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# weight => mul
# Graph fragment:
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_1, 0.7071067811865476), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.7071067811865476
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [weight], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(primals_1, buf0, 16, grid=grid(16), stream=stream0)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(buf0, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf1)
del primals_2
return (reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0), buf0, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
from math import sqrt
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Linear(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
linear = nn.Linear(in_dim, out_dim)
linear.weight.data.normal_()
linear.bias.data.zero_()
self.linear = equal_lr(linear)
def forward(self, input):
return self.linear(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_dim': 4, 'out_dim': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
from math import sqrt
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.7071067811865476
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(16)](primals_1, buf0, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(buf0, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf1)
del primals_2
return reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0
), buf0, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0)
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class LinearNew(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
linear = nn.Linear(in_dim, out_dim)
linear.weight.data.normal_()
linear.bias.data.zero_()
self.linear = equal_lr(linear)
def forward(self, input_0):
primals_2 = self.linear.bias
primals_1 = self.linear.weight_orig
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio | Linear | false | 889 | [
"MIT"
] | 0 | 231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | https://github.com/NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio/tree/231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | import torch
import torch.nn as nn
from math import sqrt
def equal_lr(module, name='weight'):
EqualLR.apply(module, name)
return module
class EqualLR:
def __init__(self, name):
self.name = name
def compute_weight(self, module):
weight = getattr(module, self.name + '_orig')
fan_in = weight.data.size(1) * weight.data[0][0].numel()
return weight * sqrt(2 / fan_in)
@staticmethod
def apply(module, name):
fn = EqualLR(name)
weight = getattr(module, name)
del module._parameters[name]
module.register_parameter(name + '_orig', nn.Parameter(weight.data))
module.register_forward_pre_hook(fn)
return fn
def __call__(self, module, input):
weight = self.compute_weight(module)
setattr(module, self.name, weight)
class Model(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
linear = nn.Linear(in_dim, out_dim)
linear.weight.data.normal_()
linear.bias.data.zero_()
self.linear = equal_lr(linear)
def forward(self, input):
return self.linear(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
PixelNorm | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ra/craeaitbyt7vnntesgul7eammhmieohm5a3bcuvnw64zksvbwmqs.py
# Topologically Sorted Source Nodes: [pow_1, mean, add, sqrt, truediv], Original ATen: [aten.pow, aten.mean, aten.add, aten.sqrt, aten.div]
# Source node to ATen node mapping:
# add => add
# mean => mean
# pow_1 => pow_1
# sqrt => sqrt
# truediv => div
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%arg0_1, 2), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%pow_1, [0], True), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean, 1e-08), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%add,), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%arg0_1, %sqrt), kwargs = {})
triton_poi_fused_add_div_mean_pow_sqrt_0 = async_compile.triton('triton_poi_fused_add_div_mean_pow_sqrt_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mean_pow_sqrt_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_mean_pow_sqrt_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (64 + x0), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (128 + x0), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (192 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = 1e-08
tmp15 = tmp13 + tmp14
tmp16 = libdevice.sqrt(tmp15)
tmp17 = tmp0 / tmp16
tl.store(out_ptr0 + (x2), tmp17, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [pow_1, mean, add, sqrt, truediv], Original ATen: [aten.pow, aten.mean, aten.add, aten.sqrt, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_mean_pow_sqrt_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class PixelNorm(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input):
return input / torch.sqrt(torch.mean(input ** 2, dim=0, keepdim=
True) + 1e-08)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_div_mean_pow_sqrt_0(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (64 + x0), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (128 + x0), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (192 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = 1e-08
tmp15 = tmp13 + tmp14
tmp16 = libdevice.sqrt(tmp15)
tmp17 = tmp0 / tmp16
tl.store(out_ptr0 + x2, tmp17, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_mean_pow_sqrt_0[grid(256)](arg0_1, buf0,
256, XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class PixelNormNew(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio | PixelNorm | false | 890 | [
"MIT"
] | 0 | 231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | https://github.com/NethraGunti/Woven-Artificial-Profile-WARP-Face-Video-Synthesis-from-Profile-and-Audio/tree/231d8daa8dddfd5eda8a092eb99c5d0e59d8b3f7 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input):
return input / torch.sqrt(torch.mean(input ** 2, dim=0, keepdim=
True) + 1e-08)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
LayerNorm | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/dg/cdgw6x7nju4bzp2wyuwgeanbco7zcjis6yiusovvnpz6zw3yjd3l.py
# Topologically Sorted Source Nodes: [u, sub], Original ATen: [aten.mean, aten.sub]
# Source node to ATen node mapping:
# sub => sub
# u => mean
# Graph fragment:
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})
triton_poi_fused_mean_sub_0 = async_compile.triton('triton_poi_fused_mean_sub_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mean_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = 4.0
tmp9 = tmp7 / tmp8
tmp10 = tmp0 - tmp9
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/qa/cqanrp6ysxh6sybzulc3onfaha6cuqejs54bwpkhct7ohd5rdj6b.py
# Topologically Sorted Source Nodes: [pow_1, s, add, sqrt, x, mul, add_1], Original ATen: [aten.pow, aten.mean, aten.add, aten.sqrt, aten.div, aten.mul]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# mul => mul
# pow_1 => pow_1
# s => mean_1
# sqrt => sqrt
# x => div
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%pow_1, [-1], True), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean_1, 1e-05), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%add,), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sub, %sqrt), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %div), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %primals_3), kwargs = {})
triton_poi_fused_add_div_mean_mul_pow_sqrt_1 = async_compile.triton('triton_poi_fused_add_div_mean_mul_pow_sqrt_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mean_mul_pow_sqrt_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_mean_mul_pow_sqrt_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp2 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tmp2 * tmp2
tmp5 = tmp4 * tmp4
tmp6 = tmp3 + tmp5
tmp8 = tmp7 * tmp7
tmp9 = tmp6 + tmp8
tmp11 = tmp10 * tmp10
tmp12 = tmp9 + tmp11
tmp13 = 4.0
tmp14 = tmp12 / tmp13
tmp15 = 1e-05
tmp16 = tmp14 + tmp15
tmp17 = libdevice.sqrt(tmp16)
tmp18 = tmp1 / tmp17
tmp19 = tmp0 * tmp18
tmp21 = tmp19 + tmp20
tl.store(out_ptr0 + (x2), tmp21, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [u, sub], Original ATen: [aten.mean, aten.sub]
stream0 = get_raw_stream(0)
triton_poi_fused_mean_sub_0.run(primals_1, buf0, 256, grid=grid(256), stream=stream0)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [pow_1, s, add, sqrt, x, mul, add_1], Original ATen: [aten.pow, aten.mean, aten.add, aten.sqrt, aten.div, aten.mul]
triton_poi_fused_add_div_mean_mul_pow_sqrt_1.run(primals_2, buf0, primals_3, buf1, 256, grid=grid(256), stream=stream0)
del buf0
del primals_2
del primals_3
return (buf1, primals_1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class LayerNorm(nn.Module):
"""Construct a layernorm module in the OpenAI style (epsilon inside the square root)."""
def __init__(self, n_state, e=1e-05):
super(LayerNorm, self).__init__()
self.g = nn.Parameter(torch.ones(n_state))
self.b = nn.Parameter(torch.zeros(n_state))
self.e = e
"""
Input:
x: n_state-dim
Output:
o: n_state-dim
"""
def forward(self, x):
u = x.mean(-1, keepdim=True)
s = (x - u).pow(2).mean(-1, keepdim=True)
x = (x - u) / torch.sqrt(s + self.e)
return self.g * x + self.b
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'n_state': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = 4.0
tmp9 = tmp7 / tmp8
tmp10 = tmp0 - tmp9
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_div_mean_mul_pow_sqrt_1(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp2 * tmp2
tmp5 = tmp4 * tmp4
tmp6 = tmp3 + tmp5
tmp8 = tmp7 * tmp7
tmp9 = tmp6 + tmp8
tmp11 = tmp10 * tmp10
tmp12 = tmp9 + tmp11
tmp13 = 4.0
tmp14 = tmp12 / tmp13
tmp15 = 1e-05
tmp16 = tmp14 + tmp15
tmp17 = libdevice.sqrt(tmp16)
tmp18 = tmp1 / tmp17
tmp19 = tmp0 * tmp18
tmp21 = tmp19 + tmp20
tl.store(out_ptr0 + x2, tmp21, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mean_sub_0[grid(256)](primals_1, buf0, 256, XBLOCK
=128, num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_div_mean_mul_pow_sqrt_1[grid(256)](primals_2,
buf0, primals_3, buf1, 256, XBLOCK=256, num_warps=4, num_stages=1)
del buf0
del primals_2
del primals_3
return buf1, primals_1
class LayerNormNew(nn.Module):
"""Construct a layernorm module in the OpenAI style (epsilon inside the square root)."""
def __init__(self, n_state, e=1e-05):
super(LayerNormNew, self).__init__()
self.g = nn.Parameter(torch.ones(n_state))
self.b = nn.Parameter(torch.zeros(n_state))
self.e = e
"""
Input:
x: n_state-dim
Output:
o: n_state-dim
"""
def forward(self, input_0):
primals_2 = self.g
primals_3 = self.b
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| NickSchoelkopf/SummerTime | LayerNorm | false | 891 | [
"Apache-2.0"
] | 0 | 9a89aab8e1544e3c52c043b9c47ab325e665e11e | https://github.com/NickSchoelkopf/SummerTime/tree/9a89aab8e1544e3c52c043b9c47ab325e665e11e | import torch
import torch.nn as nn
class Model(nn.Module):
"""Construct a layernorm module in the OpenAI style (epsilon inside the square root)."""
def __init__(self, n_state, e=1e-05):
super().__init__()
self.g = nn.Parameter(torch.ones(n_state))
self.b = nn.Parameter(torch.zeros(n_state))
self.e = e
"""
Input:
x: n_state-dim
Output:
o: n_state-dim
"""
def forward(self, x):
u = x.mean(-1, keepdim=True)
s = (x - u).pow(2).mean(-1, keepdim=True)
x = (x - u) / torch.sqrt(s + self.e)
return self.g * x + self.b
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
VdLinear | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xw/cxw7ds6vzofd5wrsg64nof7fgm4mm2ikktzhwb255ex5w5gucy4o.py
# Topologically Sorted Source Nodes: [mul_2], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# mul_2 => mul_2
# Graph fragment:
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %primals_2), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tmp0 * tmp0
tl.store(out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ly/cly2pej4mgekycfavj7bswvsbztsdzh6bt4ygykas43ajfzetcv2.py
# Topologically Sorted Source Nodes: [exp, mul, sigma], Original ATen: [aten.exp, aten.mul]
# Source node to ATen node mapping:
# exp => exp
# mul => mul
# sigma => mul_1
# Graph fragment:
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%primals_4,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp, %primals_1), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_1), kwargs = {})
triton_poi_fused_exp_mul_1 = async_compile.triton('triton_poi_fused_exp_mul_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_exp_mul_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_exp_mul_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (0))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp3 = tl.load(in_ptr1 + (x0), xmask)
tmp2 = tl_math.exp(tmp1)
tmp4 = tmp2 * tmp3
tmp5 = tmp4 * tmp3
tl.store(out_ptr0 + (x0), tmp5, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/3i/c3iwgo35oba6hfwttwuu53leevvzqcxhozcfra4umbek3rgdulha.py
# Topologically Sorted Source Nodes: [mean_1, add_1, std, mul_3, out], Original ATen: [aten.add, aten.sqrt, aten.mul]
# Source node to ATen node mapping:
# add_1 => add_1
# mean_1 => add
# mul_3 => mul_3
# out => add_2
# std => sqrt
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_1, %primals_3), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%view_3, 1e-16), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%add_1,), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sqrt, 0.0), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add, %mul_3), kwargs = {})
triton_poi_fused_add_mul_sqrt_2 = async_compile.triton('triton_poi_fused_add_mul_sqrt_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_sqrt_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_sqrt_2(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x2), xmask)
tmp2 = tmp0 + tmp1
tmp4 = 1e-16
tmp5 = tmp3 + tmp4
tmp6 = libdevice.sqrt(tmp5)
tmp7 = 0.0
tmp8 = tmp6 * tmp7
tmp9 = tmp2 + tmp8
tl.store(in_out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fn/cfnudspbq3r6goygu27k7y3howtldaho3mz37otp7h5drijht3hx.py
# Topologically Sorted Source Nodes: [neg, exp_1, log1p, sum_1, mul_4, mul_5, kl], Original ATen: [aten.neg, aten.exp, aten.log1p, aten.sum, aten.mul, aten.div]
# Source node to ATen node mapping:
# exp_1 => exp_1
# kl => div
# log1p => log1p
# mul_4 => mul_4
# mul_5 => mul_5
# neg => neg
# sum_1 => sum_1
# Graph fragment:
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%primals_4,), kwargs = {})
# %exp_1 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%neg,), kwargs = {})
# %log1p : [num_users=1] = call_function[target=torch.ops.aten.log1p.default](args = (%exp_1,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%log1p,), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sum_1, 0.5), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_4, 16), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mul_5, 1), kwargs = {})
triton_poi_fused_div_exp_log1p_mul_neg_sum_3 = async_compile.triton('triton_poi_fused_div_exp_log1p_mul_neg_sum_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {2: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=(2,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_div_exp_log1p_mul_neg_sum_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_div_exp_log1p_mul_neg_sum_3(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + (0))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = -tmp1
tmp3 = tl_math.exp(tmp2)
tmp4 = libdevice.log1p(tmp3)
tmp5 = 0.5
tmp6 = tmp4 * tmp5
tmp7 = 16.0
tmp8 = tmp6 * tmp7
tmp9 = 1.0
tmp10 = tmp8 * tmp9
tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32)), tmp10, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (1, 4), (4, 1))
assert_size_stride(primals_4, (1, 1), (1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mean], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(primals_2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mul_2], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(primals_2, buf1, 256, grid=grid(256), stream=stream0)
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [exp, mul, sigma], Original ATen: [aten.exp, aten.mul]
triton_poi_fused_exp_mul_1.run(primals_4, primals_1, buf2, 16, grid=grid(16), stream=stream0)
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear_1], Original ATen: [aten.mm]
extern_kernels.mm(reinterpret_tensor(buf1, (64, 4), (4, 1), 0), reinterpret_tensor(buf2, (4, 4), (1, 4), 0), out=buf3)
del buf2
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [mean_1, add_1, std, mul_3, out], Original ATen: [aten.add, aten.sqrt, aten.mul]
triton_poi_fused_add_mul_sqrt_2.run(buf4, primals_3, buf3, 256, grid=grid(256), stream=stream0)
del primals_3
buf5 = empty_strided_cuda((), (), torch.float32)
# Topologically Sorted Source Nodes: [neg, exp_1, log1p, sum_1, mul_4, mul_5, kl], Original ATen: [aten.neg, aten.exp, aten.log1p, aten.sum, aten.mul, aten.div]
triton_poi_fused_div_exp_log1p_mul_neg_sum_3.run(primals_4, buf5, 1, grid=grid(1), stream=stream0)
return (buf4, buf5, primals_1, primals_4, reinterpret_tensor(primals_2, (64, 4), (4, 1), 0), reinterpret_tensor(buf1, (64, 4), (4, 1), 0), buf3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((1, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, 1), (1, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def calculate_kl(log_alpha):
return 0.5 * torch.sum(torch.log1p(torch.exp(-log_alpha)))
class VdLinear(nn.Module):
"""
Linear Layer variational dropout
"""
def __init__(self, n_in, n_out, alpha_shape=(1, 1), bias=True):
super(VdLinear, self).__init__()
self.n_in = n_in
self.n_out = n_out
self.alpha_shape = alpha_shape
self.bias = bias
self.W = nn.Parameter(torch.Tensor(self.n_out, self.n_in))
self.log_alpha = nn.Parameter(torch.Tensor(*self.alpha_shape))
if bias:
self.bias = nn.Parameter(torch.Tensor(1, self.n_out))
else:
self.register_parameter('bias', None)
self.reset_parameters()
self.kl_value = calculate_kl
def reset_parameters(self):
stdv = 1.0 / math.sqrt(self.W.size(1))
self.W.data.uniform_(-stdv, stdv)
self.log_alpha.data.fill_(-5.0)
if self.bias is not None:
self.bias.data.zero_()
def forward(self, X, sample=False):
mean = F.linear(X, self.W)
if self.bias is not None:
mean = mean + self.bias
sigma = torch.exp(self.log_alpha) * self.W * self.W
std = torch.sqrt(1e-16 + F.linear(X * X, sigma))
if self.training or sample:
epsilon = std.data.new(std.size()).normal_()
else:
epsilon = 0.0
out = mean + std * epsilon
kl = self.kl_loss()
return out, kl
def kl_loss(self):
return self.W.nelement() * self.kl_value(self.log_alpha
) / self.log_alpha.nelement()
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'n_in': 4, 'n_out': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tmp0 * tmp0
tl.store(out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_poi_fused_exp_mul_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp3 = tl.load(in_ptr1 + x0, xmask)
tmp2 = tl_math.exp(tmp1)
tmp4 = tmp2 * tmp3
tmp5 = tmp4 * tmp3
tl.store(out_ptr0 + x0, tmp5, xmask)
@triton.jit
def triton_poi_fused_add_mul_sqrt_2(in_out_ptr0, in_ptr0, in_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp4 = 1e-16
tmp5 = tmp3 + tmp4
tmp6 = libdevice.sqrt(tmp5)
tmp7 = 0.0
tmp8 = tmp6 * tmp7
tmp9 = tmp2 + tmp8
tl.store(in_out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused_div_exp_log1p_mul_neg_sum_3(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = -tmp1
tmp3 = tl_math.exp(tmp2)
tmp4 = libdevice.log1p(tmp3)
tmp5 = 0.5
tmp6 = tmp4 * tmp5
tmp7 = 16.0
tmp8 = tmp6 * tmp7
tmp9 = 1.0
tmp10 = tmp8 * tmp9
tl.store(out_ptr0 + tl.full([XBLOCK], 0, tl.int32), tmp10, None)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (1, 4), (4, 1))
assert_size_stride(primals_4, (1, 1), (1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](primals_2, buf1, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_exp_mul_1[grid(16)](primals_4, primals_1, buf2, 16,
XBLOCK=16, num_warps=1, num_stages=1)
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 4), (4, 1), 0),
reinterpret_tensor(buf2, (4, 4), (1, 4), 0), out=buf3)
del buf2
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_add_mul_sqrt_2[grid(256)](buf4, primals_3, buf3,
256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_3
buf5 = empty_strided_cuda((), (), torch.float32)
triton_poi_fused_div_exp_log1p_mul_neg_sum_3[grid(1)](primals_4,
buf5, 1, XBLOCK=1, num_warps=1, num_stages=1)
return buf4, buf5, primals_1, primals_4, reinterpret_tensor(primals_2,
(64, 4), (4, 1), 0), reinterpret_tensor(buf1, (64, 4), (4, 1), 0), buf3
def calculate_kl(log_alpha):
return 0.5 * torch.sum(torch.log1p(torch.exp(-log_alpha)))
class VdLinearNew(nn.Module):
"""
Linear Layer variational dropout
"""
def __init__(self, n_in, n_out, alpha_shape=(1, 1), bias=True):
super(VdLinearNew, self).__init__()
self.n_in = n_in
self.n_out = n_out
self.alpha_shape = alpha_shape
self.bias = bias
self.W = nn.Parameter(torch.Tensor(self.n_out, self.n_in))
self.log_alpha = nn.Parameter(torch.Tensor(*self.alpha_shape))
if bias:
self.bias = nn.Parameter(torch.Tensor(1, self.n_out))
else:
self.register_parameter('bias', None)
self.reset_parameters()
self.kl_value = calculate_kl
def reset_parameters(self):
stdv = 1.0 / math.sqrt(self.W.size(1))
self.W.data.uniform_(-stdv, stdv)
self.log_alpha.data.fill_(-5.0)
if self.bias is not None:
self.bias.data.zero_()
def kl_loss(self):
return self.W.nelement() * self.kl_value(self.log_alpha
) / self.log_alpha.nelement()
def forward(self, input_0):
primals_1 = self.W
primals_4 = self.log_alpha
primals_3 = self.bias
primals_2 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0], output[1]
| Neronjust2017/pytorch-classification-project | VdLinear | false | 892 | [
"MIT"
] | 0 | fc5f4d7c46d071765f682ce20e6580646d4e5c76 | https://github.com/Neronjust2017/pytorch-classification-project/tree/fc5f4d7c46d071765f682ce20e6580646d4e5c76 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def calculate_kl(log_alpha):
return 0.5 * torch.sum(torch.log1p(torch.exp(-log_alpha)))
class Model(nn.Module):
"""
Linear Layer variational dropout
"""
def __init__(self, n_in, n_out, alpha_shape=(1, 1), bias=True):
super().__init__()
self.n_in = n_in
self.n_out = n_out
self.alpha_shape = alpha_shape
self.bias = bias
self.W = nn.Parameter(torch.Tensor(self.n_out, self.n_in))
self.log_alpha = nn.Parameter(torch.Tensor(*self.alpha_shape))
if bias:
self.bias = nn.Parameter(torch.Tensor(1, self.n_out))
else:
self.register_parameter('bias', None)
self.reset_parameters()
self.kl_value = calculate_kl
def reset_parameters(self):
stdv = 1.0 / math.sqrt(self.W.size(1))
self.W.data.uniform_(-stdv, stdv)
self.log_alpha.data.fill_(-5.0)
if self.bias is not None:
self.bias.data.zero_()
def forward(self, X, sample=False):
mean = F.linear(X, self.W)
if self.bias is not None:
mean = mean + self.bias
sigma = torch.exp(self.log_alpha) * self.W * self.W
std = torch.sqrt(1e-16 + F.linear(X * X, sigma))
if self.training or sample:
epsilon = std.data.new(std.size()).normal_()
else:
epsilon = 0.0
out = mean + std * epsilon
kl = self.kl_loss()
return out, kl
def kl_loss(self):
return self.W.nelement() * self.kl_value(self.log_alpha
) / self.log_alpha.nelement()
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
C | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/cc/ccc3jyqbz5rjanlhmaez5ahwrmffxjcm5mibxaw5efspqo45e4re.py
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# conv2d => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [4, 4], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_0 = async_compile.triton('triton_poi_fused_convolution_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[2048],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 1296
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 81) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(4, 4), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 9, 9), (324, 81, 9, 1))
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_0.run(buf1, primals_2, 1296, grid=grid(1296), stream=stream0)
del primals_2
return (buf1, primals_1, primals_3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class C(nn.Module):
def __init__(self, input_channel, output_channel, kernel_size, stride,
padding, activation=None):
"""
At the final layer, a 3x3 convolution is used to map each 64-component feature vector to the desired
number of classes.
:param input_channel: input channel size
:param output_channel: output channel size
"""
super(C, self).__init__()
if activation == 'sigmoid':
self.layer = nn.Sequential([nn.Conv2d(input_channel,
output_channel, kernel_size=kernel_size, stride=stride,
padding=padding), nn.Sigmoid()])
elif activation == 'tanh':
self.layer = nn.Sequential([nn.Conv2d(input_channel,
output_channel, kernel_size=kernel_size, stride=stride,
padding=padding), nn.Tanh()])
else:
self.layer = nn.Conv2d(input_channel, output_channel,
kernel_size=kernel_size, stride=stride, padding=padding)
def forward(self, x):
return self.layer(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_channel': 4, 'output_channel': 4, 'kernel_size': 4,
'stride': 1, 'padding': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 1296
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 81 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(4, 4), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 9, 9), (324, 81, 9, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(1296)](buf1, primals_2, 1296,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
return buf1, primals_1, primals_3
class CNew(nn.Module):
def __init__(self, input_channel, output_channel, kernel_size, stride,
padding, activation=None):
"""
At the final layer, a 3x3 convolution is used to map each 64-component feature vector to the desired
number of classes.
:param input_channel: input channel size
:param output_channel: output channel size
"""
super(CNew, self).__init__()
if activation == 'sigmoid':
self.layer = nn.Sequential([nn.Conv2d(input_channel,
output_channel, kernel_size=kernel_size, stride=stride,
padding=padding), nn.Sigmoid()])
elif activation == 'tanh':
self.layer = nn.Sequential([nn.Conv2d(input_channel,
output_channel, kernel_size=kernel_size, stride=stride,
padding=padding), nn.Tanh()])
else:
self.layer = nn.Conv2d(input_channel, output_channel,
kernel_size=kernel_size, stride=stride, padding=padding)
def forward(self, input_0):
primals_1 = self.layer.weight
primals_2 = self.layer.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Nikronic/Deep-Halftoning | C | false | 893 | [
"MIT"
] | 0 | 9564c592abf139ccab2791c1dbb354505edab5f9 | https://github.com/Nikronic/Deep-Halftoning/tree/9564c592abf139ccab2791c1dbb354505edab5f9 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, input_channel, output_channel, kernel_size, stride,
padding, activation=None):
"""
At the final layer, a 3x3 convolution is used to map each 64-component feature vector to the desired
number of classes.
:param input_channel: input channel size
:param output_channel: output channel size
"""
super().__init__()
if activation == 'sigmoid':
self.layer = nn.Sequential([nn.Conv2d(input_channel,
output_channel, kernel_size=kernel_size, stride=stride,
padding=padding), nn.Sigmoid()])
elif activation == 'tanh':
self.layer = nn.Sequential([nn.Conv2d(input_channel,
output_channel, kernel_size=kernel_size, stride=stride,
padding=padding), nn.Tanh()])
else:
self.layer = nn.Conv2d(input_channel, output_channel,
kernel_size=kernel_size, stride=stride, padding=padding)
def forward(self, x):
return self.layer(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_channel': 4, 'output_channel': 4, 'kernel_size': 4,
'stride': 1, 'padding': 4}]
|
ConcatReLU | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xs/cxswp3kc2i244jtoczawuzhlykppvkideitsqdgeyptjm7b3iubt.py
# Topologically Sorted Source Nodes: [cat, relu], Original ATen: [aten.cat, aten.relu]
# Source node to ATen node mapping:
# cat => cat
# relu => relu
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%arg0_1, %neg], 1), kwargs = {})
# %relu : [num_users=1] = call_function[target=torch.ops.aten.relu.default](args = (%cat,), kwargs = {})
triton_poi_fused_cat_relu_0 = async_compile.triton('triton_poi_fused_cat_relu_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_relu_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_relu_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 16) % 8
x0 = xindex % 16
x2 = (xindex // 128)
x3 = xindex
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + (16*x1) + (64*x2)), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr0 + (x0 + (16*((-4) + x1)) + (64*x2)), tmp6 & xmask, other=0.0)
tmp10 = -tmp9
tmp11 = tl.full(tmp10.shape, 0.0, tmp10.dtype)
tmp12 = tl.where(tmp6, tmp10, tmp11)
tmp13 = tl.where(tmp4, tmp5, tmp12)
tmp14 = tl.full([1], 0, tl.int32)
tmp15 = triton_helpers.maximum(tmp14, tmp13)
tl.store(out_ptr0 + (x3), tmp15, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat, relu], Original ATen: [aten.cat, aten.relu]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_relu_0.run(arg0_1, buf0, 512, grid=grid(512), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
def concat_relu(x):
"""Concatenated ReLU (http://arxiv.org/abs/1603.05201)."""
return F.relu(torch.cat([x, -x], dim=1))
class ConcatReLU(nn.Module):
"""Concatenated ReLU (http://arxiv.org/abs/1603.05201)."""
def forward(self, input):
return concat_relu(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_relu_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 16 % 8
x0 = xindex % 16
x2 = xindex // 128
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 16 * x1 + 64 * x2), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr0 + (x0 + 16 * (-4 + x1) + 64 * x2), tmp6 & xmask,
other=0.0)
tmp10 = -tmp9
tmp11 = tl.full(tmp10.shape, 0.0, tmp10.dtype)
tmp12 = tl.where(tmp6, tmp10, tmp11)
tmp13 = tl.where(tmp4, tmp5, tmp12)
tmp14 = tl.full([1], 0, tl.int32)
tmp15 = triton_helpers.maximum(tmp14, tmp13)
tl.store(out_ptr0 + x3, tmp15, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_relu_0[grid(512)](arg0_1, buf0, 512, XBLOCK=
128, num_warps=4, num_stages=1)
del arg0_1
return buf0,
def concat_relu(x):
"""Concatenated ReLU (http://arxiv.org/abs/1603.05201)."""
return F.relu(torch.cat([x, -x], dim=1))
class ConcatReLUNew(nn.Module):
"""Concatenated ReLU (http://arxiv.org/abs/1603.05201)."""
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Nintorac/survae_experiments | ConcatReLU | false | 894 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import torch
import torch.nn as nn
import torch.nn.functional as F
def concat_relu(x):
"""Concatenated ReLU (http://arxiv.org/abs/1603.05201)."""
return F.relu(torch.cat([x, -x], dim=1))
class Model(nn.Module):
"""Concatenated ReLU (http://arxiv.org/abs/1603.05201)."""
def forward(self, input):
return concat_relu(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
CrossEntropyLossOneHot | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/g7/cg7ria7sppao445b77eeypymkonyosw7h6ft4c2uydidxu3qyegi.py
# Topologically Sorted Source Nodes: [input_1, log, crossentropy, sum_1, crossentropy_1, sum_2], Original ATen: [aten.clamp, aten.log, aten.mul, aten.sum, aten.neg]
# Source node to ATen node mapping:
# crossentropy => mul
# crossentropy_1 => neg
# input_1 => clamp_max, clamp_min
# log => log
# sum_1 => sum_1
# sum_2 => sum_2
# Graph fragment:
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%arg0_1, 1e-07), kwargs = {})
# %clamp_max : [num_users=1] = call_function[target=torch.ops.aten.clamp_max.default](args = (%clamp_min, 0.9999999), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_max,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg1_1, %log), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%mul, [-1]), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%sum_1,), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%neg,), kwargs = {})
triton_per_fused_clamp_log_mul_neg_sum_0 = async_compile.triton('triton_per_fused_clamp_log_mul_neg_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 64],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_clamp_log_mul_neg_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_clamp_log_mul_neg_sum_0(in_ptr0, in_ptr1, out_ptr0, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 64
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (4*r0), None, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4*r0), None, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + (4*r0)), None, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr1 + (1 + (4*r0)), None, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (2 + (4*r0)), None, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr1 + (2 + (4*r0)), None, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr0 + (3 + (4*r0)), None, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr1 + (3 + (4*r0)), None, eviction_policy='evict_last')
tmp2 = 1e-07
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp4 = 0.9999999
tmp5 = triton_helpers.minimum(tmp3, tmp4)
tmp6 = tl_math.log(tmp5)
tmp7 = tmp0 * tmp6
tmp10 = triton_helpers.maximum(tmp9, tmp2)
tmp11 = triton_helpers.minimum(tmp10, tmp4)
tmp12 = tl_math.log(tmp11)
tmp13 = tmp8 * tmp12
tmp14 = tmp7 + tmp13
tmp17 = triton_helpers.maximum(tmp16, tmp2)
tmp18 = triton_helpers.minimum(tmp17, tmp4)
tmp19 = tl_math.log(tmp18)
tmp20 = tmp15 * tmp19
tmp21 = tmp14 + tmp20
tmp24 = triton_helpers.maximum(tmp23, tmp2)
tmp25 = triton_helpers.minimum(tmp24, tmp4)
tmp26 = tl_math.log(tmp25)
tmp27 = tmp22 * tmp26
tmp28 = tmp21 + tmp27
tmp29 = -tmp28
tmp30 = tl.broadcast_to(tmp29, [XBLOCK, RBLOCK])
tmp32 = tl.sum(tmp30, 1)[:, None]
tl.store(out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp32, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
# Topologically Sorted Source Nodes: [input_1, log, crossentropy, sum_1, crossentropy_1, sum_2], Original ATen: [aten.clamp, aten.log, aten.mul, aten.sum, aten.neg]
stream0 = get_raw_stream(0)
triton_per_fused_clamp_log_mul_neg_sum_0.run(arg1_1, arg0_1, buf0, 1, 64, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import Tensor
from torch.nn.modules.loss import CrossEntropyLoss
class CrossEntropyLossOneHot(CrossEntropyLoss):
EPS: 'int' = 1e-07
def forward(self, input: 'Tensor', target: 'Tensor') ->Tensor:
assert self.weight is None or isinstance(self.weight, Tensor)
input = torch.clip(input, self.EPS, 1 - self.EPS)
crossentropy = target * torch.log(input)
crossentropy = -torch.sum(crossentropy, -1)
return torch.sum(crossentropy)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch.nn.modules.loss import CrossEntropyLoss
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_clamp_log_mul_neg_sum_0(in_ptr0, in_ptr1, out_ptr0,
xnumel, rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + 4 * r0, None, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * r0, None, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + 4 * r0), None, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr1 + (1 + 4 * r0), None, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (2 + 4 * r0), None, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr1 + (2 + 4 * r0), None, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr0 + (3 + 4 * r0), None, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr1 + (3 + 4 * r0), None, eviction_policy='evict_last')
tmp2 = 1e-07
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp4 = 0.9999999
tmp5 = triton_helpers.minimum(tmp3, tmp4)
tmp6 = tl_math.log(tmp5)
tmp7 = tmp0 * tmp6
tmp10 = triton_helpers.maximum(tmp9, tmp2)
tmp11 = triton_helpers.minimum(tmp10, tmp4)
tmp12 = tl_math.log(tmp11)
tmp13 = tmp8 * tmp12
tmp14 = tmp7 + tmp13
tmp17 = triton_helpers.maximum(tmp16, tmp2)
tmp18 = triton_helpers.minimum(tmp17, tmp4)
tmp19 = tl_math.log(tmp18)
tmp20 = tmp15 * tmp19
tmp21 = tmp14 + tmp20
tmp24 = triton_helpers.maximum(tmp23, tmp2)
tmp25 = triton_helpers.minimum(tmp24, tmp4)
tmp26 = tl_math.log(tmp25)
tmp27 = tmp22 * tmp26
tmp28 = tmp21 + tmp27
tmp29 = -tmp28
tmp30 = tl.broadcast_to(tmp29, [XBLOCK, RBLOCK])
tmp32 = tl.sum(tmp30, 1)[:, None]
tl.store(out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp32, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
get_raw_stream(0)
triton_per_fused_clamp_log_mul_neg_sum_0[grid(1)](arg1_1, arg0_1,
buf0, 1, 64, XBLOCK=1, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf0,
class CrossEntropyLossOneHotNew(CrossEntropyLoss):
EPS: 'int' = 1e-07
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| NikolayZakharevich/music-processing | CrossEntropyLossOneHot | false | 895 | [
"MIT"
] | 0 | 516a3bca585f211d232cac7ede6cc417fb8878fe | https://github.com/NikolayZakharevich/music-processing/tree/516a3bca585f211d232cac7ede6cc417fb8878fe | import torch
from torch import Tensor
from torch.nn.modules.loss import CrossEntropyLoss
class Model(CrossEntropyLoss):
EPS: 'int' = 1e-07
def forward(self, input: 'Tensor', target: 'Tensor') ->Tensor:
assert self.weight is None or isinstance(self.weight, Tensor)
input = torch.clip(input, self.EPS, 1 - self.EPS)
crossentropy = target * torch.log(input)
crossentropy = -torch.sum(crossentropy, -1)
return torch.sum(crossentropy)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
Lookahead | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xf/cxfjorpj7nzakemkplfiue4lxn3wssubujqanweih6rojbuoolb7.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.constant_pad_nd]
# Source node to ATen node mapping:
# x_1 => constant_pad_nd
# Graph fragment:
# %constant_pad_nd : [num_users=2] = call_function[target=torch.ops.aten.constant_pad_nd.default](args = (%permute_1, [0, 3], 0.0), kwargs = {})
triton_poi_fused_constant_pad_nd_0 = async_compile.triton('triton_poi_fused_constant_pad_nd_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16, 8], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_constant_pad_nd_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_constant_pad_nd_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 16
xnumel = 7
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = x1
tmp1 = tl.full([1, 1], 4, tl.int64)
tmp2 = tmp0 < tmp1
tmp3 = tl.load(in_ptr0 + (y0 + (16*x1)), tmp2 & xmask & ymask, eviction_policy='evict_last', other=0.0)
tl.store(out_ptr0 + (x1 + (7*y0)), tmp3, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/r7/cr7evfztwmdzfk7dfomwxcc4g457gqiibpdthzqr2y7vyuw4ptpg.py
# Topologically Sorted Source Nodes: [x_3], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# x_3 => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_3,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_1 = async_compile.triton('triton_poi_fused_clone_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4, 16], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x1)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x1 + (16*y0)), tmp0, xmask & ymask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 1, 4), (4, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 7), (28, 7, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.constant_pad_nd]
stream0 = get_raw_stream(0)
triton_poi_fused_constant_pad_nd_0.run(primals_1, buf0, 16, 7, grid=grid(16, 7), stream=stream0)
del primals_1
# Topologically Sorted Source Nodes: [x_2], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=4, bias=None)
assert_size_stride(buf1, (4, 4, 4), (16, 4, 1))
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_3], Original ATen: [aten.clone]
triton_poi_fused_clone_1.run(buf1, buf2, 4, 16, grid=grid(4, 16), stream=stream0)
del buf1
return (buf2, primals_2, buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 1, 4), (4, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.utils.data.distributed
from torch import nn
import torch.nn.functional as F
class Lookahead(nn.Module):
def __init__(self, n_features, context):
super(Lookahead, self).__init__()
assert context > 0
self.context = context
self.n_features = n_features
self.pad = 0, self.context - 1
self.conv = nn.Conv1d(self.n_features, self.n_features, kernel_size
=self.context, stride=1, groups=self.n_features, padding=0,
bias=False)
def forward(self, x):
x = x.transpose(0, 1).transpose(1, 2)
x = F.pad(x, pad=self.pad, value=0)
x = self.conv(x)
x = x.transpose(1, 2).transpose(0, 1).contiguous()
return x
def __repr__(self):
return self.__class__.__name__ + '(' + 'n_features=' + str(self.
n_features) + ', context=' + str(self.context) + ')'
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'n_features': 4, 'context': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.utils.data.distributed
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_constant_pad_nd_0(in_ptr0, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 7
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = x1
tmp1 = tl.full([1, 1], 4, tl.int64)
tmp2 = tmp0 < tmp1
tmp3 = tl.load(in_ptr0 + (y0 + 16 * x1), tmp2 & xmask & ymask,
eviction_policy='evict_last', other=0.0)
tl.store(out_ptr0 + (x1 + 7 * y0), tmp3, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (x1 + 16 * y0), tmp0, xmask & ymask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 1, 4), (4, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 7), (28, 7, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_constant_pad_nd_0[grid(16, 7)](primals_1, buf0, 16,
7, XBLOCK=8, YBLOCK=16, num_warps=4, num_stages=1)
del primals_1
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=4, bias=None)
assert_size_stride(buf1, (4, 4, 4), (16, 4, 1))
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(4, 16)](buf1, buf2, 4, 16, XBLOCK=16,
YBLOCK=4, num_warps=1, num_stages=1)
del buf1
return buf2, primals_2, buf0
class LookaheadNew(nn.Module):
def __init__(self, n_features, context):
super(LookaheadNew, self).__init__()
assert context > 0
self.context = context
self.n_features = n_features
self.pad = 0, self.context - 1
self.conv = nn.Conv1d(self.n_features, self.n_features, kernel_size
=self.context, stride=1, groups=self.n_features, padding=0,
bias=False)
def __repr__(self):
return self.__class__.__name__ + '(' + 'n_features=' + str(self.
n_features) + ', context=' + str(self.context) + ')'
def forward(self, input_0):
primals_2 = self.conv.weight
primals_1 = input_0
output = call([primals_1, primals_2])
return output[0]
| NikolaiBabkin/deepspeech.pytorch | Lookahead | false | 896 | [
"MIT"
] | 0 | 2b120c6b735cc46200e10f81e169c8d7b75e8495 | https://github.com/NikolaiBabkin/deepspeech.pytorch/tree/2b120c6b735cc46200e10f81e169c8d7b75e8495 | import torch
import torch.utils.data.distributed
from torch import nn
import torch.nn.functional as F
class Model(nn.Module):
def __init__(self, n_features, context):
super().__init__()
assert context > 0
self.context = context
self.n_features = n_features
self.pad = 0, self.context - 1
self.conv = nn.Conv1d(self.n_features, self.n_features, kernel_size
=self.context, stride=1, groups=self.n_features, padding=0,
bias=False)
def forward(self, x):
x = x.transpose(0, 1).transpose(1, 2)
x = F.pad(x, pad=self.pad, value=0)
x = self.conv(x)
x = x.transpose(1, 2).transpose(0, 1).contiguous()
return x
def __repr__(self):
return self.__class__.__name__ + '(' + 'n_features=' + str(self.
n_features) + ', context=' + str(self.context) + ')'
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
ResidualAttentionBlock | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/6s/c6sstbvcita246hkfqwdeatnmsh3e6vlcncrzcwlsoqg7dmxvabp.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
# Source node to ATen node mapping:
# x => add, rsqrt, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%primals_3, [1]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
triton_poi_fused_native_layer_norm_0 = async_compile.triton('triton_poi_fused_native_layer_norm_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + (x0), tmp8, xmask)
tl.store(out_ptr1 + (x0), tmp23, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/zv/czv3tzezwxkylzsgkrivaldxprnr7tvjr5iihe4mbc7bzdev5lsj.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
# Source node to ATen node mapping:
# x => add, add_1, mul, mul_1, rsqrt, sub, var_mean
# Graph fragment:
# %var_mean : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%primals_3, [1]), kwargs = {correction: 0, keepdim: True})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_3, %getitem_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_1), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %primals_2), kwargs = {})
triton_poi_fused_native_layer_norm_1 = async_compile.triton('triton_poi_fused_native_layer_norm_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_native_layer_norm_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ah/cahpqo3o7hv3q647n5lretlqvfljlubj4ic7gscxws4yvkm5jzff.py
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# multi_head_attention_forward => mul_2
# Graph fragment:
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Scalar](args = (%view_9, 1.0), kwargs = {})
triton_poi_fused_mul_2 = async_compile.triton('triton_poi_fused_mul_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/id/cidhx2qv2kci37tvy6z4ewaqfmq7whhaejd6kgfwwkpiinbays7z.py
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# multi_head_attention_forward => mul_3
# Graph fragment:
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Scalar](args = (%permute_6, 1.0), kwargs = {})
triton_poi_fused_mul_3 = async_compile.triton('triton_poi_fused_mul_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/5l/c5lr3fbju3rzhzjnfdhv5evmxwg6sfeq26etlwygdbmjwzziqmty.py
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten._safe_softmax]
# Source node to ATen node mapping:
# multi_head_attention_forward => amax, exp, sub_1
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%view_14, [-1], True), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_14, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub_1,), kwargs = {})
triton_poi_fused__safe_softmax_4 = async_compile.triton('triton_poi_fused__safe_softmax_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__safe_softmax_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__safe_softmax_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ue/cue46mvjwzenos4q7lzsgaczgcooiukadngwwlhdoblv6ugei3uv.py
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten._safe_softmax]
# Source node to ATen node mapping:
# multi_head_attention_forward => any_1, div, eq, full_default, logical_not, logical_not_1, sum_1, where
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [-1], True), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
# %eq : [num_users=1] = call_function[target=torch.ops.aten.eq.Scalar](args = (%view_14, -inf), kwargs = {})
# %logical_not : [num_users=1] = call_function[target=torch.ops.aten.logical_not.default](args = (%eq,), kwargs = {})
# %any_1 : [num_users=1] = call_function[target=torch.ops.aten.any.dim](args = (%logical_not, -1, True), kwargs = {})
# %logical_not_1 : [num_users=1] = call_function[target=torch.ops.aten.logical_not.default](args = (%any_1,), kwargs = {})
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1, 4, 4, 4], 0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%logical_not_1, %full_default, %div), kwargs = {})
triton_poi_fused__safe_softmax_5 = async_compile.triton('triton_poi_fused__safe_softmax_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__safe_softmax_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 9, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__safe_softmax_5(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr1 + (x2), xmask)
tmp26 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp1 = float("-inf")
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = (tmp4 != 0)
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = (tmp9 != 0)
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = (tmp15 != 0)
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = (tmp21 != 0)
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + (x2), tmp35, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/bu/cbuwmu2mfp5lph35cxxxg6q2iiyw5l7wwbgjvpc2uau2jqndueez.py
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.clone]
# Source node to ATen node mapping:
# multi_head_attention_forward => clone
# Graph fragment:
# %clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%permute_7,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_clone_6 = async_compile.triton('triton_poi_fused_clone_6', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4, 4], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_6', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_clone_6(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 4
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x1)), xmask & ymask)
tl.store(out_ptr0 + (x1 + (4*y0)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/vd/cvdxk4eaksba3zsprpzfb3vx7t7nnrtdy3faz6dzcvumwkab3o3e.py
# Topologically Sorted Source Nodes: [x_1, x_2], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# x_1 => add_2
# x_2 => var_mean_1
# Graph fragment:
# %add_2 : [num_users=3] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_3, %squeeze), kwargs = {})
# %var_mean_1 : [num_users=2] = call_function[target=torch.ops.aten.var_mean.correction](args = (%add_2, [1]), kwargs = {correction: 0, keepdim: True})
triton_poi_fused_add_native_layer_norm_7 = async_compile.triton('triton_poi_fused_add_native_layer_norm_7', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_7', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_7(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4*x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + (x0), tmp16, xmask)
tl.store(out_ptr1 + (x0), tmp28, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ee/ceevx2e7wi6j4h2hf3scvp3at64mcchaq7w4yqwjgj6s77ewvx2h.py
# Topologically Sorted Source Nodes: [x_1, x_2], Original ATen: [aten.add, aten.native_layer_norm]
# Source node to ATen node mapping:
# x_1 => add_2
# x_2 => add_3, add_4, mul_4, mul_5, rsqrt_1, sub_2
# Graph fragment:
# %add_2 : [num_users=3] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_3, %squeeze), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_8, 1e-05), kwargs = {})
# %rsqrt_1 : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add_3,), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%add_2, %getitem_9), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_2, %rsqrt_1), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_4, %primals_8), kwargs = {})
# %add_4 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_5, %primals_9), kwargs = {})
triton_poi_fused_add_native_layer_norm_8 = async_compile.triton('triton_poi_fused_add_native_layer_norm_8', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_native_layer_norm_8', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_native_layer_norm_8(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp3 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + (x2), tmp13, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/dp/cdplbbjhtn7wjfs5zbdr7dqzrhv6sxravwmmbhqyrtfejnoccqhe.py
# Topologically Sorted Source Nodes: [input_2], Original ATen: [aten.gelu]
# Source node to ATen node mapping:
# input_2 => add_5, erf, mul_6, mul_7, mul_8
# Graph fragment:
# %mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%addmm_4, 0.5), kwargs = {})
# %mul_7 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%addmm_4, 0.7071067811865476), kwargs = {})
# %erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_7,), kwargs = {})
# %add_5 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
# %mul_8 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_6, %add_5), kwargs = {})
triton_poi_fused_gelu_9 = async_compile.triton('triton_poi_fused_gelu_9', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gelu_9', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_gelu_9(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + (x0), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/g3/cg3qezdtknb7gkpzhrsk3kwygoxnfplxs7kvhkhixedr43ssvdqx.py
# Topologically Sorted Source Nodes: [x_1, x_3], Original ATen: [aten.add]
# Source node to ATen node mapping:
# x_1 => add_2
# x_3 => add_6
# Graph fragment:
# %add_2 : [num_users=3] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_3, %squeeze), kwargs = {})
# %add_tensor : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default, %primals_13), kwargs = {})
# %add_6 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_2, %add_tensor), kwargs = {})
triton_poi_fused_add_10 = async_compile.triton('triton_poi_fused_add_10', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_10', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_10(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp3 = tl.load(in_out_ptr0 + (x2), xmask)
tmp4 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + (x2), tmp6, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13 = args
args.clear()
assert_size_stride(primals_1, (4, ), (1, ))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (12, 4), (4, 1))
assert_size_stride(primals_5, (12, ), (1, ))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, ), (1, ))
assert_size_stride(primals_8, (4, ), (1, ))
assert_size_stride(primals_9, (4, ), (1, ))
assert_size_stride(primals_10, (16, 4), (4, 1))
assert_size_stride(primals_11, (16, ), (1, ))
assert_size_stride(primals_12, (4, 16), (16, 1))
assert_size_stride(primals_13, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1), (1, 4), torch.float32)
buf1 = empty_strided_cuda((4, 1), (1, 4), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
stream0 = get_raw_stream(0)
triton_poi_fused_native_layer_norm_0.run(primals_3, buf0, buf1, 4, grid=grid(4), stream=stream0)
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.native_layer_norm]
triton_poi_fused_native_layer_norm_1.run(primals_3, buf0, buf1, primals_1, primals_2, buf2, 16, grid=grid(16), stream=stream0)
del primals_1
del primals_2
buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(buf2, reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(buf2, reinterpret_tensor(primals_4, (4, 4), (1, 4), 16), out=buf4)
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.addmm]
extern_kernels.addmm(reinterpret_tensor(primals_5, (4, ), (1, ), 8), buf2, reinterpret_tensor(primals_4, (4, 4), (1, 4), 32), alpha=1, beta=1, out=buf5)
buf6 = reinterpret_tensor(buf3, (1, 4, 4, 1), (16, 1, 4, 16), 0); del buf3 # reuse
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.mul]
triton_poi_fused_mul_2.run(buf6, primals_5, 16, grid=grid(16), stream=stream0)
buf7 = reinterpret_tensor(buf4, (1, 4, 1, 4), (16, 1, 16, 4), 0); del buf4 # reuse
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.mul]
triton_poi_fused_mul_3.run(buf7, primals_5, 16, grid=grid(16), stream=stream0)
del primals_5
buf8 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf6, (4, 4, 1), (1, 4, 0), 0), reinterpret_tensor(buf7, (4, 1, 4), (1, 0, 4), 0), out=buf8)
buf9 = empty_strided_cuda((1, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten._safe_softmax]
triton_poi_fused__safe_softmax_4.run(buf8, buf9, 64, grid=grid(64), stream=stream0)
buf10 = empty_strided_cuda((1, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten._safe_softmax]
triton_poi_fused__safe_softmax_5.run(buf8, buf9, buf10, 64, grid=grid(64), stream=stream0)
buf11 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf10, (4, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf5, (4, 4, 1), (1, 4, 0), 0), out=buf11)
buf12 = empty_strided_cuda((4, 1, 4, 1), (4, 1, 1, 4), torch.float32)
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.clone]
triton_poi_fused_clone_6.run(buf11, buf12, 4, 4, grid=grid(4, 4), stream=stream0)
buf13 = reinterpret_tensor(buf11, (4, 4), (4, 1), 0); del buf11 # reuse
# Topologically Sorted Source Nodes: [multi_head_attention_forward], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_7, reinterpret_tensor(buf12, (4, 4), (4, 1), 0), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf13)
del primals_7
buf14 = buf1; del buf1 # reuse
buf15 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [x_1, x_2], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_7.run(primals_3, buf13, buf14, buf15, 4, grid=grid(4), stream=stream0)
buf16 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1, x_2], Original ATen: [aten.add, aten.native_layer_norm]
triton_poi_fused_add_native_layer_norm_8.run(primals_3, buf13, buf14, buf15, primals_8, primals_9, buf16, 16, grid=grid(16), stream=stream0)
del buf14
del buf15
del primals_9
buf17 = reinterpret_tensor(buf9, (4, 16), (16, 1), 0); del buf9 # reuse
# Topologically Sorted Source Nodes: [input_1], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_11, buf16, reinterpret_tensor(primals_10, (4, 16), (1, 4), 0), alpha=1, beta=1, out=buf17)
del primals_11
buf18 = reinterpret_tensor(buf8, (4, 16), (16, 1), 0); del buf8 # reuse
# Topologically Sorted Source Nodes: [input_2], Original ATen: [aten.gelu]
triton_poi_fused_gelu_9.run(buf17, buf18, 64, grid=grid(64), stream=stream0)
buf19 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(buf18, reinterpret_tensor(primals_12, (16, 4), (1, 16), 0), out=buf19)
buf20 = buf19; del buf19 # reuse
# Topologically Sorted Source Nodes: [x_1, x_3], Original ATen: [aten.add]
triton_poi_fused_add_10.run(buf20, primals_3, buf13, primals_13, 16, grid=grid(16), stream=stream0)
del primals_13
return (buf20, primals_3, primals_8, buf2, buf10, reinterpret_tensor(buf12, (4, 4), (4, 1), 0), buf13, buf16, buf17, buf18, primals_12, primals_10, primals_6, reinterpret_tensor(buf5, (4, 1, 4), (1, 1, 4), 0), reinterpret_tensor(buf6, (4, 1, 4), (1, 4, 4), 0), reinterpret_tensor(buf7, (4, 4, 1), (1, 4, 16), 0), reinterpret_tensor(primals_4, (4, 4), (4, 1), 32), reinterpret_tensor(primals_4, (4, 4), (4, 1), 16), reinterpret_tensor(primals_4, (4, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((12, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((12, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((16, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_12 = rand_strided((4, 16), (16, 1), device='cuda:0', dtype=torch.float32)
primals_13 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12, primals_13])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from typing import Callable
from torch import nn
from torch.nn import functional as F
import torch.distributed.nn
from collections import OrderedDict
from typing import Optional
class LayerNorm(nn.LayerNorm):
"""Subclass torch's LayerNorm to handle fp16."""
def forward(self, x: 'torch.Tensor'):
x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias,
self.eps)
return x
class ResidualAttentionBlock(nn.Module):
def __init__(self, d_model: 'int', n_head: 'int', act_layer: 'Callable'
=nn.GELU):
super().__init__()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = LayerNorm(d_model)
self.mlp = nn.Sequential(OrderedDict([('c_fc', nn.Linear(d_model,
d_model * 4)), ('gelu', act_layer()), ('c_proj', nn.Linear(
d_model * 4, d_model))]))
self.ln_2 = LayerNorm(d_model)
def attention(self, x: 'torch.Tensor', attn_mask:
'Optional[torch.Tensor]'=None):
return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
def forward(self, x: 'torch.Tensor', attn_mask:
'Optional[torch.Tensor]'=None):
x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
x = x + self.mlp(self.ln_2(x))
return x
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'n_head': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from typing import Callable
from torch import nn
from torch.nn import functional as F
import torch.distributed.nn
from collections import OrderedDict
from typing import Optional
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_native_layer_norm_0(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_mul_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_mul_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + (4 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused__safe_softmax_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__safe_softmax_5(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + x2, xmask)
tmp26 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp29 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp31 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = float('-inf')
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = tmp4 != 0
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = tmp9 != 0
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = tmp15 != 0
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = tmp21 != 0
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + x2, tmp35, xmask)
@triton.jit
def triton_poi_fused_clone_6(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask)
tl.store(out_ptr0 + (x1 + 4 * y0), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_7(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_8(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
@triton.jit
def triton_poi_fused_gelu_9(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
@triton.jit
def triton_poi_fused_add_10(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_out_ptr0 + x2, xmask)
tmp4 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + x2, tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13) = args
args.clear()
assert_size_stride(primals_1, (4,), (1,))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (12, 4), (4, 1))
assert_size_stride(primals_5, (12,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4,), (1,))
assert_size_stride(primals_10, (16, 4), (4, 1))
assert_size_stride(primals_11, (16,), (1,))
assert_size_stride(primals_12, (4, 16), (16, 1))
assert_size_stride(primals_13, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1), (1, 4), torch.float32)
buf1 = empty_strided_cuda((4, 1), (1, 4), torch.float32)
get_raw_stream(0)
triton_poi_fused_native_layer_norm_0[grid(4)](primals_3, buf0, buf1,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_native_layer_norm_1[grid(16)](primals_3, buf0,
buf1, primals_1, primals_2, buf2, 16, XBLOCK=16, num_warps=1,
num_stages=1)
del primals_1
del primals_2
buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf2, reinterpret_tensor(primals_4, (4, 4), (1, 4
), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf2, reinterpret_tensor(primals_4, (4, 4), (1, 4
), 16), out=buf4)
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(reinterpret_tensor(primals_5, (4,), (1,), 8),
buf2, reinterpret_tensor(primals_4, (4, 4), (1, 4), 32), alpha=
1, beta=1, out=buf5)
buf6 = reinterpret_tensor(buf3, (1, 4, 4, 1), (16, 1, 4, 16), 0)
del buf3
triton_poi_fused_mul_2[grid(16)](buf6, primals_5, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf7 = reinterpret_tensor(buf4, (1, 4, 1, 4), (16, 1, 16, 4), 0)
del buf4
triton_poi_fused_mul_3[grid(16)](buf7, primals_5, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del primals_5
buf8 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf6, (4, 4, 1), (1, 4, 0), 0
), reinterpret_tensor(buf7, (4, 1, 4), (1, 0, 4), 0), out=buf8)
buf9 = empty_strided_cuda((1, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__safe_softmax_4[grid(64)](buf8, buf9, 64, XBLOCK=
64, num_warps=1, num_stages=1)
buf10 = empty_strided_cuda((1, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__safe_softmax_5[grid(64)](buf8, buf9, buf10, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf11 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf10, (4, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf5, (4, 4, 1), (1, 4, 0), 0), out=buf11)
buf12 = empty_strided_cuda((4, 1, 4, 1), (4, 1, 1, 4), torch.float32)
triton_poi_fused_clone_6[grid(4, 4)](buf11, buf12, 4, 4, XBLOCK=4,
YBLOCK=4, num_warps=1, num_stages=1)
buf13 = reinterpret_tensor(buf11, (4, 4), (4, 1), 0)
del buf11
extern_kernels.addmm(primals_7, reinterpret_tensor(buf12, (4, 4), (
4, 1), 0), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf13)
del primals_7
buf14 = buf1
del buf1
buf15 = buf0
del buf0
triton_poi_fused_add_native_layer_norm_7[grid(4)](primals_3, buf13,
buf14, buf15, 4, XBLOCK=4, num_warps=1, num_stages=1)
buf16 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_8[grid(16)](primals_3, buf13,
buf14, buf15, primals_8, primals_9, buf16, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del buf14
del buf15
del primals_9
buf17 = reinterpret_tensor(buf9, (4, 16), (16, 1), 0)
del buf9
extern_kernels.addmm(primals_11, buf16, reinterpret_tensor(
primals_10, (4, 16), (1, 4), 0), alpha=1, beta=1, out=buf17)
del primals_11
buf18 = reinterpret_tensor(buf8, (4, 16), (16, 1), 0)
del buf8
triton_poi_fused_gelu_9[grid(64)](buf17, buf18, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf19 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf18, reinterpret_tensor(primals_12, (16, 4), (1,
16), 0), out=buf19)
buf20 = buf19
del buf19
triton_poi_fused_add_10[grid(16)](buf20, primals_3, buf13,
primals_13, 16, XBLOCK=16, num_warps=1, num_stages=1)
del primals_13
return (buf20, primals_3, primals_8, buf2, buf10, reinterpret_tensor(
buf12, (4, 4), (4, 1), 0), buf13, buf16, buf17, buf18, primals_12,
primals_10, primals_6, reinterpret_tensor(buf5, (4, 1, 4), (1, 1, 4
), 0), reinterpret_tensor(buf6, (4, 1, 4), (1, 4, 4), 0),
reinterpret_tensor(buf7, (4, 4, 1), (1, 4, 16), 0),
reinterpret_tensor(primals_4, (4, 4), (4, 1), 32),
reinterpret_tensor(primals_4, (4, 4), (4, 1), 16),
reinterpret_tensor(primals_4, (4, 4), (4, 1), 0))
class LayerNorm(nn.LayerNorm):
"""Subclass torch's LayerNorm to handle fp16."""
def forward(self, x: 'torch.Tensor'):
x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias,
self.eps)
return x
class ResidualAttentionBlockNew(nn.Module):
def __init__(self, d_model: 'int', n_head: 'int', act_layer: 'Callable'
=nn.GELU):
super().__init__()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = LayerNorm(d_model)
self.mlp = nn.Sequential(OrderedDict([('c_fc', nn.Linear(d_model,
d_model * 4)), ('gelu', act_layer()), ('c_proj', nn.Linear(
d_model * 4, d_model))]))
self.ln_2 = LayerNorm(d_model)
def attention(self, x: 'torch.Tensor', attn_mask:
'Optional[torch.Tensor]'=None):
return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
def forward(self, input_0):
primals_4 = self.attn.in_proj_weight
primals_5 = self.attn.in_proj_bias
primals_3 = self.attn.out_proj.weight
primals_1 = self.attn.out_proj.bias
primals_2 = self.ln_1.weight
primals_7 = self.ln_1.bias
primals_10 = self.mlp.c_fc.weight
primals_11 = self.mlp.c_fc.bias
primals_12 = self.mlp.c_proj.weight
primals_8 = self.mlp.c_proj.bias
primals_9 = self.ln_2.weight
primals_13 = self.ln_2.bias
primals_6 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13])
return output[0]
| NYU-DICE-Lab/open_clip | ResidualAttentionBlock | false | 897 | [
"MIT"
] | 0 | fd71804b503135fb1c7cc8de3a0d6599741c8ed9 | https://github.com/NYU-DICE-Lab/open_clip/tree/fd71804b503135fb1c7cc8de3a0d6599741c8ed9 | import torch
from typing import Callable
from torch import nn
from torch.nn import functional as F
import torch.distributed.nn
from collections import OrderedDict
from typing import Optional
class LayerNorm(nn.LayerNorm):
"""Subclass torch's LayerNorm to handle fp16."""
def forward(self, x: 'torch.Tensor'):
x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias,
self.eps)
return x
class Model(nn.Module):
def __init__(self, d_model: 'int', n_head: 'int', act_layer: 'Callable'
=nn.GELU):
super().__init__()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = LayerNorm(d_model)
self.mlp = nn.Sequential(OrderedDict([('c_fc', nn.Linear(d_model,
d_model * 4)), ('gelu', act_layer()), ('c_proj', nn.Linear(
d_model * 4, d_model))]))
self.ln_2 = LayerNorm(d_model)
def attention(self, x: 'torch.Tensor', attn_mask:
'Optional[torch.Tensor]'=None):
return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
def forward(self, x: 'torch.Tensor', attn_mask:
'Optional[torch.Tensor]'=None):
x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
x = x + self.mlp(self.ln_2(x))
return x
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [4, 4]
|
SoftCrossEntropyLoss2d | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/td/ctdj5kazgiki6gdaadhqtp2x7tq2ee5ey5hqqdcoqmp54jyhf74f.py
# Topologically Sorted Source Nodes: [log_softmax], Original ATen: [aten._log_softmax]
# Source node to ATen node mapping:
# log_softmax => amax, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%arg0_1, [1], True), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg0_1, %amax), kwargs = {})
triton_poi_fused__log_softmax_0 = async_compile.triton('triton_poi_fused__log_softmax_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__log_softmax_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__log_softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tl.store(out_ptr0 + (x3), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ne/cneuoe5ed43ex5ojv524lcm6efihmzp4tx5sn3qedrcityjvt6pz.py
# Topologically Sorted Source Nodes: [log_softmax, inputs], Original ATen: [aten._log_softmax, aten.neg]
# Source node to ATen node mapping:
# inputs => neg
# log_softmax => exp, log, sub_1, sum_1
# Graph fragment:
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [1], True), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%sum_1,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%sub, %log), kwargs = {})
# %neg : [num_users=4] = call_function[target=torch.ops.aten.neg.default](args = (%sub_1,), kwargs = {})
triton_poi_fused__log_softmax_neg_1 = async_compile.triton('triton_poi_fused__log_softmax_neg_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__log_softmax_neg_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__log_softmax_neg_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl_math.exp(tmp1)
tmp4 = tl_math.exp(tmp3)
tmp5 = tmp2 + tmp4
tmp7 = tl_math.exp(tmp6)
tmp8 = tmp5 + tmp7
tmp10 = tl_math.exp(tmp9)
tmp11 = tmp8 + tmp10
tmp12 = tl_math.log(tmp11)
tmp13 = tmp0 - tmp12
tmp14 = -tmp13
tl.store(out_ptr0 + (x3), tmp14, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/hm/chmjikblxn7fogqton2wplz4p5ubizfnslw2y2jskup5crqiglsi.py
# Topologically Sorted Source Nodes: [getitem], Original ATen: [aten.index]
# Source node to ATen node mapping:
# getitem => index
# Graph fragment:
# %index : [num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%neg, [%full_default]), kwargs = {})
triton_poi_fused_index_2 = async_compile.triton('triton_poi_fused_index_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4, 16], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_index_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_index_2(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (x1 + (16*y0)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + (4*x1)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/xg/cxgyo7qw6gisgut6u4bocabj7jkiaw3bydso4hzgts7qjaia74xe.py
# Topologically Sorted Source Nodes: [getitem_2], Original ATen: [aten.index]
# Source node to ATen node mapping:
# getitem_2 => index_2
# Graph fragment:
# %index_2 : [num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%neg, [%full_default_2]), kwargs = {})
triton_poi_fused_index_3 = async_compile.triton('triton_poi_fused_index_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4, 16], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_index_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_index_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (64 + x1 + (16*y0)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + (4*x1)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/iq/ciq2g5midlhlq7225k24quvfyfki7oejp5vufoszhy2i3ktzwl66.py
# Topologically Sorted Source Nodes: [getitem_4], Original ATen: [aten.index]
# Source node to ATen node mapping:
# getitem_4 => index_4
# Graph fragment:
# %index_4 : [num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%neg, [%full_default_4]), kwargs = {})
triton_poi_fused_index_4 = async_compile.triton('triton_poi_fused_index_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4, 16], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_index_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_index_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (128 + x1 + (16*y0)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + (4*x1)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/uw/cuwcbkvcc6zyei2rc3jzebdq4rl7s3nyyw3nqvslxd7f2o7txngx.py
# Topologically Sorted Source Nodes: [getitem_6], Original ATen: [aten.index]
# Source node to ATen node mapping:
# getitem_6 => index_6
# Graph fragment:
# %index_6 : [num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%neg, [%full_default_6]), kwargs = {})
triton_poi_fused_index_5 = async_compile.triton('triton_poi_fused_index_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4, 16], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_index_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_index_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (192 + x1 + (16*y0)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + (4*x1)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/zz/czzlytqgjwzprkdc572dfyzk5vpiocqqctt74eadeo2aq54ezkih.py
# Topologically Sorted Source Nodes: [truediv, loss, truediv_1, loss_1, truediv_2, loss_2, truediv_3, loss_3], Original ATen: [aten.div, aten.add]
# Source node to ATen node mapping:
# loss => add
# loss_1 => add_1
# loss_2 => add_2
# loss_3 => add_3
# truediv => div
# truediv_1 => div_1
# truediv_2 => div_2
# truediv_3 => div_3
# Graph fragment:
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%convolution, 16), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, 0), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%convolution_1, 16), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add, %div_1), kwargs = {})
# %div_2 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%convolution_2, 16), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_1, %div_2), kwargs = {})
# %div_3 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%convolution_3, 16), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_2, %div_3), kwargs = {})
triton_poi_fused_add_div_6 = async_compile.triton('triton_poi_fused_add_div_6', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {4: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=(4,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_6', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_6(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 1
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + (0))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp6 = tl.load(in_ptr1 + (0))
tmp7 = tl.broadcast_to(tmp6, [XBLOCK])
tmp10 = tl.load(in_out_ptr0 + (0))
tmp11 = tl.broadcast_to(tmp10, [XBLOCK])
tmp14 = tl.load(in_ptr2 + (0))
tmp15 = tl.broadcast_to(tmp14, [XBLOCK])
tmp2 = 0.0625
tmp3 = tmp1 * tmp2
tmp4 = 0.0
tmp5 = tmp3 + tmp4
tmp8 = tmp7 * tmp2
tmp9 = tmp5 + tmp8
tmp12 = tmp11 * tmp2
tmp13 = tmp9 + tmp12
tmp16 = tmp15 * tmp2
tmp17 = tmp13 + tmp16
tl.store(in_out_ptr0 + (tl.full([XBLOCK], 0, tl.int32)), tmp17, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [log_softmax], Original ATen: [aten._log_softmax]
stream0 = get_raw_stream(0)
triton_poi_fused__log_softmax_0.run(arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [log_softmax, inputs], Original ATen: [aten._log_softmax, aten.neg]
triton_poi_fused__log_softmax_neg_1.run(buf0, buf1, 256, grid=grid(256), stream=stream0)
del buf0
buf2 = empty_strided_cuda((1, 4, 4, 4), (64, 1, 16, 4), torch.float32)
# Topologically Sorted Source Nodes: [getitem], Original ATen: [aten.index]
triton_poi_fused_index_2.run(buf1, buf2, 4, 16, grid=grid(4, 16), stream=stream0)
buf3 = empty_strided_cuda((1, 4, 4, 4), (64, 1, 16, 4), torch.float32)
# Topologically Sorted Source Nodes: [getitem_1], Original ATen: [aten.index]
triton_poi_fused_index_2.run(arg1_1, buf3, 4, 16, grid=grid(4, 16), stream=stream0)
# Topologically Sorted Source Nodes: [getitem, getitem_1, conv2d], Original ATen: [aten.index, aten.convolution]
buf4 = extern_kernels.convolution(buf2, buf3, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (1, 1, 1, 1), (1, 1, 1, 1))
buf5 = buf3; del buf3 # reuse
# Topologically Sorted Source Nodes: [getitem_2], Original ATen: [aten.index]
triton_poi_fused_index_3.run(buf1, buf5, 4, 16, grid=grid(4, 16), stream=stream0)
buf6 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [getitem_3], Original ATen: [aten.index]
triton_poi_fused_index_3.run(arg1_1, buf6, 4, 16, grid=grid(4, 16), stream=stream0)
# Topologically Sorted Source Nodes: [getitem_2, getitem_3, conv2d_1], Original ATen: [aten.index, aten.convolution]
buf7 = extern_kernels.convolution(buf5, buf6, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf7, (1, 1, 1, 1), (1, 1, 1, 1))
buf8 = buf6; del buf6 # reuse
# Topologically Sorted Source Nodes: [getitem_4], Original ATen: [aten.index]
triton_poi_fused_index_4.run(buf1, buf8, 4, 16, grid=grid(4, 16), stream=stream0)
buf9 = buf5; del buf5 # reuse
# Topologically Sorted Source Nodes: [getitem_5], Original ATen: [aten.index]
triton_poi_fused_index_4.run(arg1_1, buf9, 4, 16, grid=grid(4, 16), stream=stream0)
# Topologically Sorted Source Nodes: [getitem_4, getitem_5, conv2d_2], Original ATen: [aten.index, aten.convolution]
buf10 = extern_kernels.convolution(buf8, buf9, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf10, (1, 1, 1, 1), (1, 1, 1, 1))
buf11 = buf9; del buf9 # reuse
# Topologically Sorted Source Nodes: [getitem_6], Original ATen: [aten.index]
triton_poi_fused_index_5.run(buf1, buf11, 4, 16, grid=grid(4, 16), stream=stream0)
del buf1
buf12 = buf8; del buf8 # reuse
# Topologically Sorted Source Nodes: [getitem_7], Original ATen: [aten.index]
triton_poi_fused_index_5.run(arg1_1, buf12, 4, 16, grid=grid(4, 16), stream=stream0)
del arg1_1
# Topologically Sorted Source Nodes: [getitem_6, getitem_7, conv2d_3], Original ATen: [aten.index, aten.convolution]
buf13 = extern_kernels.convolution(buf11, buf12, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf13, (1, 1, 1, 1), (1, 1, 1, 1))
del buf11
del buf12
buf14 = buf10; del buf10 # reuse
# Topologically Sorted Source Nodes: [truediv, loss, truediv_1, loss_1, truediv_2, loss_2, truediv_3, loss_3], Original ATen: [aten.div, aten.add]
triton_poi_fused_add_div_6.run(buf14, buf4, buf7, buf13, 1, grid=grid(1), stream=stream0)
del buf13
del buf4
del buf7
return (buf14, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.utils.data.distributed
import torch
import torch.nn as nn
from numpy import int64 as int64
from torchvision.transforms import functional as F
import torch.nn.functional as F
import torch.utils
class SoftCrossEntropyLoss2d(nn.Module):
def __init__(self):
super(SoftCrossEntropyLoss2d, self).__init__()
def forward(self, inputs, targets):
loss = 0
inputs = -F.log_softmax(inputs, dim=1)
for index in range(inputs.size()[0]):
loss += F.conv2d(inputs[range(index, index + 1)], targets[range
(index, index + 1)]) / (targets.size()[2] * targets.size()[3])
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.utils.data.distributed
import torch
import torch.nn as nn
from numpy import int64 as int64
import torch.utils
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__log_softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tl.store(out_ptr0 + x3, tmp8, xmask)
@triton.jit
def triton_poi_fused__log_softmax_neg_1(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl_math.exp(tmp1)
tmp4 = tl_math.exp(tmp3)
tmp5 = tmp2 + tmp4
tmp7 = tl_math.exp(tmp6)
tmp8 = tmp5 + tmp7
tmp10 = tl_math.exp(tmp9)
tmp11 = tmp8 + tmp10
tmp12 = tl_math.log(tmp11)
tmp13 = tmp0 - tmp12
tmp14 = -tmp13
tl.store(out_ptr0 + x3, tmp14, xmask)
@triton.jit
def triton_poi_fused_index_2(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (x1 + 16 * y0), xmask & ymask, eviction_policy
='evict_last')
tl.store(out_ptr0 + (y0 + 4 * x1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_index_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (64 + x1 + 16 * y0), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + 4 * x1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_index_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (128 + x1 + 16 * y0), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + 4 * x1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_index_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (192 + x1 + 16 * y0), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + 4 * x1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_div_6(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp6 = tl.load(in_ptr1 + 0)
tmp7 = tl.broadcast_to(tmp6, [XBLOCK])
tmp10 = tl.load(in_out_ptr0 + 0)
tmp11 = tl.broadcast_to(tmp10, [XBLOCK])
tmp14 = tl.load(in_ptr2 + 0)
tmp15 = tl.broadcast_to(tmp14, [XBLOCK])
tmp2 = 0.0625
tmp3 = tmp1 * tmp2
tmp4 = 0.0
tmp5 = tmp3 + tmp4
tmp8 = tmp7 * tmp2
tmp9 = tmp5 + tmp8
tmp12 = tmp11 * tmp2
tmp13 = tmp9 + tmp12
tmp16 = tmp15 * tmp2
tmp17 = tmp13 + tmp16
tl.store(in_out_ptr0 + tl.full([XBLOCK], 0, tl.int32), tmp17, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__log_softmax_0[grid(256)](arg0_1, buf0, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__log_softmax_neg_1[grid(256)](buf0, buf1, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del buf0
buf2 = empty_strided_cuda((1, 4, 4, 4), (64, 1, 16, 4), torch.float32)
triton_poi_fused_index_2[grid(4, 16)](buf1, buf2, 4, 16, XBLOCK=16,
YBLOCK=4, num_warps=1, num_stages=1)
buf3 = empty_strided_cuda((1, 4, 4, 4), (64, 1, 16, 4), torch.float32)
triton_poi_fused_index_2[grid(4, 16)](arg1_1, buf3, 4, 16, XBLOCK=
16, YBLOCK=4, num_warps=1, num_stages=1)
buf4 = extern_kernels.convolution(buf2, buf3, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (1, 1, 1, 1), (1, 1, 1, 1))
buf5 = buf3
del buf3
triton_poi_fused_index_3[grid(4, 16)](buf1, buf5, 4, 16, XBLOCK=16,
YBLOCK=4, num_warps=1, num_stages=1)
buf6 = buf2
del buf2
triton_poi_fused_index_3[grid(4, 16)](arg1_1, buf6, 4, 16, XBLOCK=
16, YBLOCK=4, num_warps=1, num_stages=1)
buf7 = extern_kernels.convolution(buf5, buf6, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf7, (1, 1, 1, 1), (1, 1, 1, 1))
buf8 = buf6
del buf6
triton_poi_fused_index_4[grid(4, 16)](buf1, buf8, 4, 16, XBLOCK=16,
YBLOCK=4, num_warps=1, num_stages=1)
buf9 = buf5
del buf5
triton_poi_fused_index_4[grid(4, 16)](arg1_1, buf9, 4, 16, XBLOCK=
16, YBLOCK=4, num_warps=1, num_stages=1)
buf10 = extern_kernels.convolution(buf8, buf9, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf10, (1, 1, 1, 1), (1, 1, 1, 1))
buf11 = buf9
del buf9
triton_poi_fused_index_5[grid(4, 16)](buf1, buf11, 4, 16, XBLOCK=16,
YBLOCK=4, num_warps=1, num_stages=1)
del buf1
buf12 = buf8
del buf8
triton_poi_fused_index_5[grid(4, 16)](arg1_1, buf12, 4, 16, XBLOCK=
16, YBLOCK=4, num_warps=1, num_stages=1)
del arg1_1
buf13 = extern_kernels.convolution(buf11, buf12, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf13, (1, 1, 1, 1), (1, 1, 1, 1))
del buf11
del buf12
buf14 = buf10
del buf10
triton_poi_fused_add_div_6[grid(1)](buf14, buf4, buf7, buf13, 1,
XBLOCK=1, num_warps=1, num_stages=1)
del buf13
del buf4
del buf7
return buf14,
class SoftCrossEntropyLoss2dNew(nn.Module):
def __init__(self):
super(SoftCrossEntropyLoss2dNew, self).__init__()
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| HRTNet/HRTNet | SoftCrossEntropyLoss2d | false | 898 | [
"MIT"
] | 0 | 6a51c9c34568988ea6125a1638794c63d8fadbea | https://github.com/HRTNet/HRTNet/tree/6a51c9c34568988ea6125a1638794c63d8fadbea | import torch
import torch.utils.data.distributed
import torch
import torch.nn as nn
from numpy import int64 as int64
from torchvision.transforms import functional as F
import torch.nn.functional as F
import torch.utils
class Model(nn.Module):
def __init__(self):
super().__init__()
def forward(self, inputs, targets):
loss = 0
inputs = -F.log_softmax(inputs, dim=1)
for index in range(inputs.size()[0]):
loss += F.conv2d(inputs[range(index, index + 1)], targets[range
(index, index + 1)]) / (targets.size()[2] * targets.size()[3])
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
GatedTanhUnit | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ic/cicum3b5lw57fdrli5dh4vobpwyztfzzh7svgy63gkajhb2lh6p7.py
# Topologically Sorted Source Nodes: [tanh, sigmoid, mul], Original ATen: [aten.tanh, aten.sigmoid, aten.mul]
# Source node to ATen node mapping:
# mul => mul
# sigmoid => sigmoid
# tanh => tanh
# Graph fragment:
# %tanh : [num_users=1] = call_function[target=torch.ops.aten.tanh.default](args = (%getitem,), kwargs = {})
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%getitem_1,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%tanh, %sigmoid), kwargs = {})
triton_poi_fused_mul_sigmoid_tanh_0 = async_compile.triton('triton_poi_fused_mul_sigmoid_tanh_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_sigmoid_tanh_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_sigmoid_tanh_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 2
x1 = (xindex // 2)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (4*x1)), xmask)
tmp2 = tl.load(in_ptr0 + (2 + x0 + (4*x1)), xmask)
tmp1 = libdevice.tanh(tmp0)
tmp3 = tl.sigmoid(tmp2)
tmp4 = tmp1 * tmp3
tl.store(out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 2), (32, 8, 2, 1), torch.float32)
# Topologically Sorted Source Nodes: [tanh, sigmoid, mul], Original ATen: [aten.tanh, aten.sigmoid, aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_sigmoid_tanh_0.run(arg0_1, buf0, 128, grid=grid(128), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
def gated_tanh(x, dim):
"""Gated Tanh activation."""
x_tanh, x_sigmoid = torch.chunk(x, 2, dim=dim)
return torch.tanh(x_tanh) * torch.sigmoid(x_sigmoid)
class GatedTanhUnit(nn.Module):
"""Gated Tanh activation."""
def __init__(self, dim=-1):
super(GatedTanhUnit, self).__init__()
self.dim = dim
def forward(self, x):
return gated_tanh(x, dim=self.dim)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_sigmoid_tanh_0(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 2
x1 = xindex // 2
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x1), xmask)
tmp2 = tl.load(in_ptr0 + (2 + x0 + 4 * x1), xmask)
tmp1 = libdevice.tanh(tmp0)
tmp3 = tl.sigmoid(tmp2)
tmp4 = tmp1 * tmp3
tl.store(out_ptr0 + x2, tmp4, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 2), (32, 8, 2, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_sigmoid_tanh_0[grid(128)](arg0_1, buf0, 128,
XBLOCK=128, num_warps=4, num_stages=1)
del arg0_1
return buf0,
def gated_tanh(x, dim):
"""Gated Tanh activation."""
x_tanh, x_sigmoid = torch.chunk(x, 2, dim=dim)
return torch.tanh(x_tanh) * torch.sigmoid(x_sigmoid)
class GatedTanhUnitNew(nn.Module):
"""Gated Tanh activation."""
def __init__(self, dim=-1):
super(GatedTanhUnitNew, self).__init__()
self.dim = dim
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Nintorac/survae_experiments | GatedTanhUnit | false | 899 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import torch
import torch.nn as nn
def gated_tanh(x, dim):
"""Gated Tanh activation."""
x_tanh, x_sigmoid = torch.chunk(x, 2, dim=dim)
return torch.tanh(x_tanh) * torch.sigmoid(x_sigmoid)
class Model(nn.Module):
"""Gated Tanh activation."""
def __init__(self, dim=-1):
super().__init__()
self.dim = dim
def forward(self, x):
return gated_tanh(x, dim=self.dim)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
ConcatELU | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ik/cik7ixpiwi2uiiqxn4hnifuuboljjrthi4sf66od2vu5ec6zrapy.py
# Topologically Sorted Source Nodes: [cat, elu], Original ATen: [aten.cat, aten.elu]
# Source node to ATen node mapping:
# cat => cat
# elu => expm1, gt, mul, mul_1, mul_2, where
# Graph fragment:
# %cat : [num_users=3] = call_function[target=torch.ops.aten.cat.default](args = ([%arg0_1, %neg], 1), kwargs = {})
# %gt : [num_users=1] = call_function[target=torch.ops.aten.gt.Scalar](args = (%cat, 0), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%cat, 1.0), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%cat, 1.0), kwargs = {})
# %expm1 : [num_users=1] = call_function[target=torch.ops.aten.expm1.default](args = (%mul_1,), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%expm1, 1.0), kwargs = {})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%gt, %mul, %mul_2), kwargs = {})
triton_poi_fused_cat_elu_0 = async_compile.triton('triton_poi_fused_cat_elu_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_elu_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_elu_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 16) % 8
x0 = xindex % 16
x2 = (xindex // 128)
x3 = xindex
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + (16*x1) + (64*x2)), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr0 + (x0 + (16*((-4) + x1)) + (64*x2)), tmp6 & xmask, other=0.0)
tmp10 = -tmp9
tmp11 = tl.full(tmp10.shape, 0.0, tmp10.dtype)
tmp12 = tl.where(tmp6, tmp10, tmp11)
tmp13 = tl.where(tmp4, tmp5, tmp12)
tmp14 = 0.0
tmp15 = tmp13 > tmp14
tmp16 = 1.0
tmp17 = tmp13 * tmp16
tmp18 = libdevice.expm1(tmp17)
tmp19 = tmp18 * tmp16
tmp20 = tl.where(tmp15, tmp17, tmp19)
tl.store(out_ptr0 + (x3), tmp20, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat, elu], Original ATen: [aten.cat, aten.elu]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_elu_0.run(arg0_1, buf0, 512, grid=grid(512), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
def concat_elu(x):
"""Like concatenated ReLU (http://arxiv.org/abs/1603.05201), but with ELU instead."""
return F.elu(torch.cat([x, -x], dim=1))
class ConcatELU(nn.Module):
"""Like concatenated ReLU (http://arxiv.org/abs/1603.05201), but with ELU instead."""
def forward(self, input):
return concat_elu(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_elu_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 16 % 8
x0 = xindex % 16
x2 = xindex // 128
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 16 * x1 + 64 * x2), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr0 + (x0 + 16 * (-4 + x1) + 64 * x2), tmp6 & xmask,
other=0.0)
tmp10 = -tmp9
tmp11 = tl.full(tmp10.shape, 0.0, tmp10.dtype)
tmp12 = tl.where(tmp6, tmp10, tmp11)
tmp13 = tl.where(tmp4, tmp5, tmp12)
tmp14 = 0.0
tmp15 = tmp13 > tmp14
tmp16 = 1.0
tmp17 = tmp13 * tmp16
tmp18 = libdevice.expm1(tmp17)
tmp19 = tmp18 * tmp16
tmp20 = tl.where(tmp15, tmp17, tmp19)
tl.store(out_ptr0 + x3, tmp20, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_elu_0[grid(512)](arg0_1, buf0, 512, XBLOCK=128,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
def concat_elu(x):
"""Like concatenated ReLU (http://arxiv.org/abs/1603.05201), but with ELU instead."""
return F.elu(torch.cat([x, -x], dim=1))
class ConcatELUNew(nn.Module):
"""Like concatenated ReLU (http://arxiv.org/abs/1603.05201), but with ELU instead."""
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Nintorac/survae_experiments | ConcatELU | false | 900 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import torch
import torch.nn as nn
import torch.nn.functional as F
def concat_elu(x):
"""Like concatenated ReLU (http://arxiv.org/abs/1603.05201), but with ELU instead."""
return F.elu(torch.cat([x, -x], dim=1))
class Model(nn.Module):
"""Like concatenated ReLU (http://arxiv.org/abs/1603.05201), but with ELU instead."""
def forward(self, input):
return concat_elu(input)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
PositionalEncoding1d | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/cm/ccm7y3tyrccqsaknetzzscd6vebg5vwn32zhm3o3kao5h5m3e4de.py
# Topologically Sorted Source Nodes: [add], Original ATen: [aten.add]
# Source node to ATen node mapping:
# add => add
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_2, %primals_1), kwargs = {})
triton_poi_fused_add_0 = async_compile.triton('triton_poi_fused_add_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x2 = (xindex // 16) % 4
tmp0 = tl.load(in_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + (4*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x4), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 1, 4), (4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [add], Original ATen: [aten.add]
stream0 = get_raw_stream(0)
triton_poi_fused_add_0.run(primals_2, primals_1, buf0, 256, grid=grid(256), stream=stream0)
del primals_1
del primals_2
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 1, 4), (4, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn as nn
class PositionalEncoding1d(nn.Module):
"""
Learning positional embeddings.
Args:
shape: Iterable, the shape of the input.
embedding_dim: int, the size of each embedding vector.
"""
def __init__(self, size, embedding_dim):
super(PositionalEncoding1d, self).__init__()
self.size = size
self.embedding_dim = embedding_dim
self.encode_l = nn.Parameter(torch.Tensor(size, 1, embedding_dim))
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.encode_l, std=0.125 / math.sqrt(self.
embedding_dim))
def forward(self, x):
return x + self.encode_l
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'size': 4, 'embedding_dim': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x2 = xindex // 16 % 4
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 1, 4), (4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_0[grid(256)](primals_2, primals_1, buf0, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_1
del primals_2
return buf0,
class PositionalEncoding1dNew(nn.Module):
"""
Learning positional embeddings.
Args:
shape: Iterable, the shape of the input.
embedding_dim: int, the size of each embedding vector.
"""
def __init__(self, size, embedding_dim):
super(PositionalEncoding1dNew, self).__init__()
self.size = size
self.embedding_dim = embedding_dim
self.encode_l = nn.Parameter(torch.Tensor(size, 1, embedding_dim))
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.encode_l, std=0.125 / math.sqrt(self.
embedding_dim))
def forward(self, input_0):
primals_1 = self.encode_l
primals_2 = input_0
output = call([primals_1, primals_2])
return output[0]
| Nintorac/survae_experiments | PositionalEncoding1d | false | 901 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import math
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Learning positional embeddings.
Args:
shape: Iterable, the shape of the input.
embedding_dim: int, the size of each embedding vector.
"""
def __init__(self, size, embedding_dim):
super().__init__()
self.size = size
self.embedding_dim = embedding_dim
self.encode_l = nn.Parameter(torch.Tensor(size, 1, embedding_dim))
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.encode_l, std=0.125 / math.sqrt(self.
embedding_dim))
def forward(self, x):
return x + self.encode_l
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
AutoregressiveShift | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/6t/c6tcgwx2rbm4gnoatd2jl443wwzzxsgjzexlqawnvkx7opkzocib.py
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%expand, %slice_1],), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = (xindex // 16)
x0 = xindex % 4
x3 = xindex % 16
x4 = xindex
tmp0 = x2
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 4, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + (x3 + (16*((-1) + x2))), tmp6 & xmask, other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x4), tmp10, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (1, 1, 4), (4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_1, primals_2, buf0, 64, grid=grid(64), stream=stream0)
del primals_1
del primals_2
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((1, 1, 4), (4, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class AutoregressiveShift(nn.Module):
"""Shifts input right to make model autoregressive."""
def __init__(self, embed_dim):
super(AutoregressiveShift, self).__init__()
self.embed_dim = embed_dim
self.first_token = nn.Parameter(torch.Tensor(1, 1, embed_dim))
self._reset_parameters()
def _reset_parameters(self):
nn.init.xavier_uniform_(self.first_token)
def forward(self, x):
first_token = self.first_token.expand(1, x.shape[1], self.embed_dim)
return torch.cat([first_token, x[:-1]], dim=0)
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'embed_dim': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16
x0 = xindex % 4
x3 = xindex % 16
x4 = xindex
tmp0 = x2
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + x0, tmp4 & xmask, eviction_policy='evict_last',
other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 4, tl.int64)
tmp9 = tl.load(in_ptr1 + (x3 + 16 * (-1 + x2)), tmp6 & xmask, other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x4, tmp10, xmask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (1, 1, 4), (4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(64)](primals_1, primals_2, buf0, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_1
del primals_2
return buf0,
class AutoregressiveShiftNew(nn.Module):
"""Shifts input right to make model autoregressive."""
def __init__(self, embed_dim):
super(AutoregressiveShiftNew, self).__init__()
self.embed_dim = embed_dim
self.first_token = nn.Parameter(torch.Tensor(1, 1, embed_dim))
self._reset_parameters()
def _reset_parameters(self):
nn.init.xavier_uniform_(self.first_token)
def forward(self, input_0):
primals_1 = self.first_token
primals_2 = input_0
output = call([primals_1, primals_2])
return output[0]
| Nintorac/survae_experiments | AutoregressiveShift | false | 902 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import torch
import torch.nn as nn
class Model(nn.Module):
"""Shifts input right to make model autoregressive."""
def __init__(self, embed_dim):
super().__init__()
self.embed_dim = embed_dim
self.first_token = nn.Parameter(torch.Tensor(1, 1, embed_dim))
self._reset_parameters()
def _reset_parameters(self):
nn.init.xavier_uniform_(self.first_token)
def forward(self, x):
first_token = self.first_token.expand(1, x.shape[1], self.embed_dim)
return torch.cat([first_token, x[:-1]], dim=0)
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [4]
|
GatedConv2d | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/57/c57ohskjse4k3tnvco77uckol5cma3wz6yjje4oj5puip3zrtjnd.py
# Topologically Sorted Source Nodes: [h], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# h => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [4, 4], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_0 = async_compile.triton('triton_poi_fused_convolution_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4096],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 3888
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 81) % 12
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/7r/c7rmyohqzoyh35lg73th6dpwy4h4clifhsgaiccc6rzjwmabryuf.py
# Topologically Sorted Source Nodes: [sigmoid, mul, add], Original ATen: [aten.sigmoid, aten.mul, aten.add]
# Source node to ATen node mapping:
# add => add
# mul => mul
# sigmoid => sigmoid
# Graph fragment:
# %sigmoid : [num_users=2] = call_function[target=torch.ops.aten.sigmoid.default](args = (%getitem_2,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%getitem_1, %sigmoid), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, %mul), kwargs = {})
triton_poi_fused_add_mul_sigmoid_1 = async_compile.triton('triton_poi_fused_add_mul_sigmoid_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[2048],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_sigmoid_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_1(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 1296
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 324
x1 = (xindex // 324)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (648 + x0 + (972*x1)), xmask)
tmp2 = tl.load(in_ptr0 + (x0 + (972*x1)), xmask)
tmp3 = tl.load(in_ptr0 + (324 + x0 + (972*x1)), xmask)
tmp1 = tl.sigmoid(tmp0)
tmp4 = tmp3 * tmp1
tmp5 = tmp2 + tmp4
tl.store(out_ptr0 + (x2), tmp1, xmask)
tl.store(out_ptr1 + (x2), tmp5, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (12, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (12, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [h], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(4, 4), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 12, 9, 9), (972, 81, 9, 1))
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [h], Original ATen: [aten.convolution]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_0.run(buf1, primals_2, 3888, grid=grid(3888), stream=stream0)
del primals_2
buf2 = empty_strided_cuda((4, 4, 9, 9), (324, 81, 9, 1), torch.float32)
buf3 = empty_strided_cuda((4, 4, 9, 9), (324, 81, 9, 1), torch.float32)
# Topologically Sorted Source Nodes: [sigmoid, mul, add], Original ATen: [aten.sigmoid, aten.mul, aten.add]
triton_poi_fused_add_mul_sigmoid_1.run(buf1, buf2, buf3, 1296, grid=grid(1296), stream=stream0)
return (buf3, primals_1, primals_3, reinterpret_tensor(buf1, (4, 4, 9, 9), (972, 81, 9, 1), 324), buf2, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((12, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((12, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class GatedConv2d(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, padding):
super(GatedConv2d, self).__init__()
self.in_channels = in_channels
self.conv = nn.Conv2d(in_channels, out_channels * 3, kernel_size=
kernel_size, padding=padding)
def forward(self, x):
h = self.conv(x)
a, b, c = torch.chunk(h, chunks=3, dim=1)
return a + b * torch.sigmoid(c)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'kernel_size': 4,
'padding': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 3888
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 81 % 12
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_1(in_ptr0, out_ptr0, out_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 1296
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 324
x1 = xindex // 324
x2 = xindex
tmp0 = tl.load(in_ptr0 + (648 + x0 + 972 * x1), xmask)
tmp2 = tl.load(in_ptr0 + (x0 + 972 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (324 + x0 + 972 * x1), xmask)
tmp1 = tl.sigmoid(tmp0)
tmp4 = tmp3 * tmp1
tmp5 = tmp2 + tmp4
tl.store(out_ptr0 + x2, tmp1, xmask)
tl.store(out_ptr1 + x2, tmp5, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (12, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (12,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(4, 4), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 12, 9, 9), (972, 81, 9, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(3888)](buf1, primals_2, 3888,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((4, 4, 9, 9), (324, 81, 9, 1), torch.float32)
buf3 = empty_strided_cuda((4, 4, 9, 9), (324, 81, 9, 1), torch.float32)
triton_poi_fused_add_mul_sigmoid_1[grid(1296)](buf1, buf2, buf3,
1296, XBLOCK=128, num_warps=4, num_stages=1)
return buf3, primals_1, primals_3, reinterpret_tensor(buf1, (4, 4, 9, 9
), (972, 81, 9, 1), 324), buf2
class GatedConv2dNew(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, padding):
super(GatedConv2dNew, self).__init__()
self.in_channels = in_channels
self.conv = nn.Conv2d(in_channels, out_channels * 3, kernel_size=
kernel_size, padding=padding)
def forward(self, input_0):
primals_1 = self.conv.weight
primals_2 = self.conv.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Nintorac/survae_experiments | GatedConv2d | false | 903 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, padding):
super().__init__()
self.in_channels = in_channels
self.conv = nn.Conv2d(in_channels, out_channels * 3, kernel_size=
kernel_size, padding=padding)
def forward(self, x):
h = self.conv(x)
a, b, c = torch.chunk(h, chunks=3, dim=1)
return a + b * torch.sigmoid(c)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'kernel_size': 4,
'padding': 4}]
|
HeatmapLoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xf/cxf2mveni2qzuqw5oi7mwzm5vuj4a63unfx2dikmtlxpnubv3xi5.py
# Topologically Sorted Source Nodes: [sub, l, mean, mean_1], Original ATen: [aten.sub, aten.pow, aten.mean]
# Source node to ATen node mapping:
# l => pow_1
# mean => mean
# mean_1 => mean_1
# sub => sub
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg0_1, %arg1_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%pow_1, [3]), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%mean, [2]), kwargs = {})
triton_poi_fused_mean_pow_sub_0 = async_compile.triton('triton_poi_fused_mean_pow_sub_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mean_pow_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 32, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mean_pow_sub_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (16*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (16*x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (1 + (16*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (1 + (16*x0)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (2 + (16*x0)), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr1 + (2 + (16*x0)), xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr0 + (3 + (16*x0)), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr1 + (3 + (16*x0)), xmask, eviction_policy='evict_last')
tmp21 = tl.load(in_ptr0 + (4 + (16*x0)), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr1 + (4 + (16*x0)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr0 + (5 + (16*x0)), xmask, eviction_policy='evict_last')
tmp26 = tl.load(in_ptr1 + (5 + (16*x0)), xmask, eviction_policy='evict_last')
tmp30 = tl.load(in_ptr0 + (6 + (16*x0)), xmask, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr1 + (6 + (16*x0)), xmask, eviction_policy='evict_last')
tmp35 = tl.load(in_ptr0 + (7 + (16*x0)), xmask, eviction_policy='evict_last')
tmp36 = tl.load(in_ptr1 + (7 + (16*x0)), xmask, eviction_policy='evict_last')
tmp42 = tl.load(in_ptr0 + (8 + (16*x0)), xmask, eviction_policy='evict_last')
tmp43 = tl.load(in_ptr1 + (8 + (16*x0)), xmask, eviction_policy='evict_last')
tmp46 = tl.load(in_ptr0 + (9 + (16*x0)), xmask, eviction_policy='evict_last')
tmp47 = tl.load(in_ptr1 + (9 + (16*x0)), xmask, eviction_policy='evict_last')
tmp51 = tl.load(in_ptr0 + (10 + (16*x0)), xmask, eviction_policy='evict_last')
tmp52 = tl.load(in_ptr1 + (10 + (16*x0)), xmask, eviction_policy='evict_last')
tmp56 = tl.load(in_ptr0 + (11 + (16*x0)), xmask, eviction_policy='evict_last')
tmp57 = tl.load(in_ptr1 + (11 + (16*x0)), xmask, eviction_policy='evict_last')
tmp63 = tl.load(in_ptr0 + (12 + (16*x0)), xmask, eviction_policy='evict_last')
tmp64 = tl.load(in_ptr1 + (12 + (16*x0)), xmask, eviction_policy='evict_last')
tmp67 = tl.load(in_ptr0 + (13 + (16*x0)), xmask, eviction_policy='evict_last')
tmp68 = tl.load(in_ptr1 + (13 + (16*x0)), xmask, eviction_policy='evict_last')
tmp72 = tl.load(in_ptr0 + (14 + (16*x0)), xmask, eviction_policy='evict_last')
tmp73 = tl.load(in_ptr1 + (14 + (16*x0)), xmask, eviction_policy='evict_last')
tmp77 = tl.load(in_ptr0 + (15 + (16*x0)), xmask, eviction_policy='evict_last')
tmp78 = tl.load(in_ptr1 + (15 + (16*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp6 = tmp4 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tmp3 + tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp19 = 4.0
tmp20 = tmp18 / tmp19
tmp23 = tmp21 - tmp22
tmp24 = tmp23 * tmp23
tmp27 = tmp25 - tmp26
tmp28 = tmp27 * tmp27
tmp29 = tmp24 + tmp28
tmp32 = tmp30 - tmp31
tmp33 = tmp32 * tmp32
tmp34 = tmp29 + tmp33
tmp37 = tmp35 - tmp36
tmp38 = tmp37 * tmp37
tmp39 = tmp34 + tmp38
tmp40 = tmp39 / tmp19
tmp41 = tmp20 + tmp40
tmp44 = tmp42 - tmp43
tmp45 = tmp44 * tmp44
tmp48 = tmp46 - tmp47
tmp49 = tmp48 * tmp48
tmp50 = tmp45 + tmp49
tmp53 = tmp51 - tmp52
tmp54 = tmp53 * tmp53
tmp55 = tmp50 + tmp54
tmp58 = tmp56 - tmp57
tmp59 = tmp58 * tmp58
tmp60 = tmp55 + tmp59
tmp61 = tmp60 / tmp19
tmp62 = tmp41 + tmp61
tmp65 = tmp63 - tmp64
tmp66 = tmp65 * tmp65
tmp69 = tmp67 - tmp68
tmp70 = tmp69 * tmp69
tmp71 = tmp66 + tmp70
tmp74 = tmp72 - tmp73
tmp75 = tmp74 * tmp74
tmp76 = tmp71 + tmp75
tmp79 = tmp77 - tmp78
tmp80 = tmp79 * tmp79
tmp81 = tmp76 + tmp80
tmp82 = tmp81 / tmp19
tmp83 = tmp62 + tmp82
tmp84 = tmp83 / tmp19
tl.store(out_ptr0 + (x0), tmp84, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/35/c35v4ojil7vwot3cvbaaxwst2htsi5dh2y4o4d7fr4ma3zmlulqt.py
# Topologically Sorted Source Nodes: [l_1], Original ATen: [aten.mean]
# Source node to ATen node mapping:
# l_1 => mean_2
# Graph fragment:
# %mean_2 : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%mean_1, [1]), kwargs = {})
triton_poi_fused_mean_1 = async_compile.triton('triton_poi_fused_mean_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mean_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mean_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tl.store(out_ptr0 + (x0), tmp8, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sub, l, mean, mean_1], Original ATen: [aten.sub, aten.pow, aten.mean]
stream0 = get_raw_stream(0)
triton_poi_fused_mean_pow_sub_0.run(arg0_1, arg1_1, buf0, 16, grid=grid(16), stream=stream0)
del arg0_1
del arg1_1
buf1 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [l_1], Original ATen: [aten.mean]
triton_poi_fused_mean_1.run(buf0, buf1, 4, grid=grid(4), stream=stream0)
del buf0
return (buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.utils.data
class HeatmapLoss(torch.nn.Module):
"""
loss for detection heatmap
"""
def __init__(self):
super(HeatmapLoss, self).__init__()
def forward(self, pred, gt):
l = (pred - gt) ** 2
l = l.mean(dim=3).mean(dim=2).mean(dim=1)
return l
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mean_pow_sub_0(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 16 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 16 * x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (1 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp5 = tl.load(in_ptr1 + (1 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp9 = tl.load(in_ptr0 + (2 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp10 = tl.load(in_ptr1 + (2 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp14 = tl.load(in_ptr0 + (3 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp15 = tl.load(in_ptr1 + (3 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp21 = tl.load(in_ptr0 + (4 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp22 = tl.load(in_ptr1 + (4 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp25 = tl.load(in_ptr0 + (5 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp26 = tl.load(in_ptr1 + (5 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp30 = tl.load(in_ptr0 + (6 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp31 = tl.load(in_ptr1 + (6 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp35 = tl.load(in_ptr0 + (7 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp36 = tl.load(in_ptr1 + (7 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp42 = tl.load(in_ptr0 + (8 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp43 = tl.load(in_ptr1 + (8 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp46 = tl.load(in_ptr0 + (9 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp47 = tl.load(in_ptr1 + (9 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp51 = tl.load(in_ptr0 + (10 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp52 = tl.load(in_ptr1 + (10 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp56 = tl.load(in_ptr0 + (11 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp57 = tl.load(in_ptr1 + (11 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp63 = tl.load(in_ptr0 + (12 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp64 = tl.load(in_ptr1 + (12 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp67 = tl.load(in_ptr0 + (13 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp68 = tl.load(in_ptr1 + (13 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp72 = tl.load(in_ptr0 + (14 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp73 = tl.load(in_ptr1 + (14 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp77 = tl.load(in_ptr0 + (15 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp78 = tl.load(in_ptr1 + (15 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp6 = tmp4 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tmp3 + tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp19 = 4.0
tmp20 = tmp18 / tmp19
tmp23 = tmp21 - tmp22
tmp24 = tmp23 * tmp23
tmp27 = tmp25 - tmp26
tmp28 = tmp27 * tmp27
tmp29 = tmp24 + tmp28
tmp32 = tmp30 - tmp31
tmp33 = tmp32 * tmp32
tmp34 = tmp29 + tmp33
tmp37 = tmp35 - tmp36
tmp38 = tmp37 * tmp37
tmp39 = tmp34 + tmp38
tmp40 = tmp39 / tmp19
tmp41 = tmp20 + tmp40
tmp44 = tmp42 - tmp43
tmp45 = tmp44 * tmp44
tmp48 = tmp46 - tmp47
tmp49 = tmp48 * tmp48
tmp50 = tmp45 + tmp49
tmp53 = tmp51 - tmp52
tmp54 = tmp53 * tmp53
tmp55 = tmp50 + tmp54
tmp58 = tmp56 - tmp57
tmp59 = tmp58 * tmp58
tmp60 = tmp55 + tmp59
tmp61 = tmp60 / tmp19
tmp62 = tmp41 + tmp61
tmp65 = tmp63 - tmp64
tmp66 = tmp65 * tmp65
tmp69 = tmp67 - tmp68
tmp70 = tmp69 * tmp69
tmp71 = tmp66 + tmp70
tmp74 = tmp72 - tmp73
tmp75 = tmp74 * tmp74
tmp76 = tmp71 + tmp75
tmp79 = tmp77 - tmp78
tmp80 = tmp79 * tmp79
tmp81 = tmp76 + tmp80
tmp82 = tmp81 / tmp19
tmp83 = tmp62 + tmp82
tmp84 = tmp83 / tmp19
tl.store(out_ptr0 + x0, tmp84, xmask)
@triton.jit
def triton_poi_fused_mean_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mean_pow_sub_0[grid(16)](arg0_1, arg1_1, buf0, 16,
XBLOCK=16, num_warps=1, num_stages=1)
del arg0_1
del arg1_1
buf1 = empty_strided_cuda((4,), (1,), torch.float32)
triton_poi_fused_mean_1[grid(4)](buf0, buf1, 4, XBLOCK=4, num_warps
=1, num_stages=1)
del buf0
return buf1,
class HeatmapLossNew(torch.nn.Module):
"""
loss for detection heatmap
"""
def __init__(self):
super(HeatmapLossNew, self).__init__()
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| NiranthS/pytorch_stacked_hourglass | HeatmapLoss | false | 904 | [
"BSD-3-Clause"
] | 0 | db9838eb13f6848ba3b9db844c1e023eb8688c3c | https://github.com/NiranthS/pytorch_stacked_hourglass/tree/db9838eb13f6848ba3b9db844c1e023eb8688c3c | import torch
import torch.utils.data
class Model(torch.nn.Module):
"""
loss for detection heatmap
"""
def __init__(self):
super().__init__()
def forward(self, pred, gt):
l = (pred - gt) ** 2
l = l.mean(dim=3).mean(dim=2).mean(dim=1)
return l
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
AvgPoolHead | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/is/cispe7zbbl4nxt2jjus6h5iou2w7htohqj7z2oz6g7nqz6vbpbqr.py
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.avg_pool2d]
# Source node to ATen node mapping:
# x => avg_pool2d
# Graph fragment:
# %avg_pool2d : [num_users=1] = call_function[target=torch.ops.aten.avg_pool2d.default](args = (%primals_1, [4, 4], [1, 1]), kwargs = {})
triton_poi_fused_avg_pool2d_0 = async_compile.triton('triton_poi_fused_avg_pool2d_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_avg_pool2d_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 16, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_avg_pool2d_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (16*x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (16*x0)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + (16*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + (16*x0)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (4 + (16*x0)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (5 + (16*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (6 + (16*x0)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr0 + (7 + (16*x0)), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (8 + (16*x0)), xmask, eviction_policy='evict_last')
tmp17 = tl.load(in_ptr0 + (9 + (16*x0)), xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_ptr0 + (10 + (16*x0)), xmask, eviction_policy='evict_last')
tmp21 = tl.load(in_ptr0 + (11 + (16*x0)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr0 + (12 + (16*x0)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr0 + (13 + (16*x0)), xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr0 + (14 + (16*x0)), xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr0 + (15 + (16*x0)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 + tmp0
tmp4 = tmp3 + tmp2
tmp6 = tmp5 + tmp4
tmp8 = tmp7 + tmp6
tmp10 = tmp9 + tmp8
tmp12 = tmp11 + tmp10
tmp14 = tmp13 + tmp12
tmp16 = tmp15 + tmp14
tmp18 = tmp17 + tmp16
tmp20 = tmp19 + tmp18
tmp22 = tmp21 + tmp20
tmp24 = tmp23 + tmp22
tmp26 = tmp25 + tmp24
tmp28 = tmp27 + tmp26
tmp30 = tmp29 + tmp28
tmp31 = 0.0625
tmp32 = tmp30 * tmp31
tl.store(out_ptr0 + (x0), tmp32, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.avg_pool2d]
stream0 = get_raw_stream(0)
triton_poi_fused_avg_pool2d_0.run(primals_1, buf0, 16, grid=grid(16), stream=stream0)
del primals_1
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_2], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_3, reinterpret_tensor(buf0, (4, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf1)
del primals_2
del primals_3
return (buf1, reinterpret_tensor(buf0, (4, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.optim
class AvgPoolHead(nn.Module):
def __init__(self, in_channels, out_channels, fea_map_size):
super(AvgPoolHead, self).__init__()
self.avgpool = nn.AvgPool2d(fea_map_size, stride=1)
self.fc = nn.Linear(in_channels, out_channels)
def forward(self, x):
x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'fea_map_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
import torch.optim
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_avg_pool2d_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 16 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp3 = tl.load(in_ptr0 + (2 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp5 = tl.load(in_ptr0 + (3 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp7 = tl.load(in_ptr0 + (4 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp9 = tl.load(in_ptr0 + (5 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp11 = tl.load(in_ptr0 + (6 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp13 = tl.load(in_ptr0 + (7 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp15 = tl.load(in_ptr0 + (8 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp17 = tl.load(in_ptr0 + (9 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp19 = tl.load(in_ptr0 + (10 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp21 = tl.load(in_ptr0 + (11 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr0 + (12 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp25 = tl.load(in_ptr0 + (13 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp27 = tl.load(in_ptr0 + (14 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp29 = tl.load(in_ptr0 + (15 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp2 = tmp1 + tmp0
tmp4 = tmp3 + tmp2
tmp6 = tmp5 + tmp4
tmp8 = tmp7 + tmp6
tmp10 = tmp9 + tmp8
tmp12 = tmp11 + tmp10
tmp14 = tmp13 + tmp12
tmp16 = tmp15 + tmp14
tmp18 = tmp17 + tmp16
tmp20 = tmp19 + tmp18
tmp22 = tmp21 + tmp20
tmp24 = tmp23 + tmp22
tmp26 = tmp25 + tmp24
tmp28 = tmp27 + tmp26
tmp30 = tmp29 + tmp28
tmp31 = 0.0625
tmp32 = tmp30 * tmp31
tl.store(out_ptr0 + x0, tmp32, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_avg_pool2d_0[grid(16)](primals_1, buf0, 16, XBLOCK
=16, num_warps=1, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_3, reinterpret_tensor(buf0, (4, 4), (4,
1), 0), reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), alpha
=1, beta=1, out=buf1)
del primals_2
del primals_3
return buf1, reinterpret_tensor(buf0, (4, 4), (4, 1), 0)
class AvgPoolHeadNew(nn.Module):
def __init__(self, in_channels, out_channels, fea_map_size):
super(AvgPoolHeadNew, self).__init__()
self.avgpool = nn.AvgPool2d(fea_map_size, stride=1)
self.fc = nn.Linear(in_channels, out_channels)
def forward(self, input_0):
primals_2 = self.fc.weight
primals_3 = self.fc.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| NiteshBharadwaj/structured_aleatoric_uncertainty_for_human_pose | AvgPoolHead | false | 905 | [
"MIT"
] | 0 | c74fb7384be562f0a0f1966b3fadf19e13a235f2 | https://github.com/NiteshBharadwaj/structured_aleatoric_uncertainty_for_human_pose/tree/c74fb7384be562f0a0f1966b3fadf19e13a235f2 | import torch
import torch.nn as nn
import torch.optim
class Model(nn.Module):
def __init__(self, in_channels, out_channels, fea_map_size):
super().__init__()
self.avgpool = nn.AvgPool2d(fea_map_size, stride=1)
self.fc = nn.Linear(in_channels, out_channels)
def forward(self, x):
x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
PositionalEncodingImage | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/m3/cm32zeebfqk27kpgoaoaod7i4nafyctqhhjnawbrgrkg2de3udg7.py
# Topologically Sorted Source Nodes: [add, add_1, add_2], Original ATen: [aten.add]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# add_2 => add_2
# Graph fragment:
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%primals_2, %primals_1), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add, %primals_3), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add_1, %primals_4), kwargs = {})
triton_poi_fused_add_0 = async_compile.triton('triton_poi_fused_add_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_0(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x3 = (xindex // 64)
x2 = (xindex // 16) % 4
x7 = xindex % 16
tmp0 = tl.load(in_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + (4*x3)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + (x0 + (4*x2)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + (x7), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tl.store(out_ptr0 + (x4), tmp6, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (1, 4, 1, 1, 4), (16, 4, 4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (1, 1, 4, 1, 4), (16, 16, 4, 4, 1))
assert_size_stride(primals_4, (1, 1, 1, 4, 4), (16, 16, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((1, 4, 4, 4, 4), (256, 64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [add, add_1, add_2], Original ATen: [aten.add]
stream0 = get_raw_stream(0)
triton_poi_fused_add_0.run(primals_2, primals_1, primals_3, primals_4, buf0, 256, grid=grid(256), stream=stream0)
del primals_1
del primals_2
del primals_3
del primals_4
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((1, 4, 1, 1, 4), (16, 4, 4, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((1, 1, 4, 1, 4), (16, 16, 4, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((1, 1, 1, 4, 4), (16, 16, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn as nn
class PositionalEncodingImage(nn.Module):
"""
Learning positional embeddings for images.
Embeddings for channel, height and width are added to form the full positional embedding.
These encodings correspond to the ones from Sparse Transformers (https://arxiv.org/abs/1904.10509).
Args:
image_shape: Iterable, the shape of the image.
embedding_dim: int, the size of each embedding vector.
"""
def __init__(self, image_shape, embedding_dim):
super(PositionalEncodingImage, self).__init__()
assert len(image_shape
) == 3, 'image_shape should have length 3: (C,H,W)'
self.image_shape = image_shape
self.embedding_dim = embedding_dim
c, h, w = image_shape
self.encode_c = nn.Parameter(torch.Tensor(1, c, 1, 1, embedding_dim))
self.encode_h = nn.Parameter(torch.Tensor(1, 1, h, 1, embedding_dim))
self.encode_w = nn.Parameter(torch.Tensor(1, 1, 1, w, embedding_dim))
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.encode_c, std=0.125 / math.sqrt(3 * self.
embedding_dim))
nn.init.normal_(self.encode_h, std=0.125 / math.sqrt(3 * self.
embedding_dim))
nn.init.normal_(self.encode_w, std=0.125 / math.sqrt(3 * self.
embedding_dim))
def forward(self, x):
return x + self.encode_c + self.encode_h + self.encode_w
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'image_shape': [4, 4, 4], 'embedding_dim': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_0(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
x3 = xindex // 64
x2 = xindex // 16 % 4
x7 = xindex % 16
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp3 = tl.load(in_ptr2 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp5 = tl.load(in_ptr3 + x7, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tl.store(out_ptr0 + x4, tmp6, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (1, 4, 1, 1, 4), (16, 4, 4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (1, 1, 4, 1, 4), (16, 16, 4, 4, 1))
assert_size_stride(primals_4, (1, 1, 1, 4, 4), (16, 16, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((1, 4, 4, 4, 4), (256, 64, 16, 4, 1),
torch.float32)
get_raw_stream(0)
triton_poi_fused_add_0[grid(256)](primals_2, primals_1, primals_3,
primals_4, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1)
del primals_1
del primals_2
del primals_3
del primals_4
return buf0,
class PositionalEncodingImageNew(nn.Module):
"""
Learning positional embeddings for images.
Embeddings for channel, height and width are added to form the full positional embedding.
These encodings correspond to the ones from Sparse Transformers (https://arxiv.org/abs/1904.10509).
Args:
image_shape: Iterable, the shape of the image.
embedding_dim: int, the size of each embedding vector.
"""
def __init__(self, image_shape, embedding_dim):
super(PositionalEncodingImageNew, self).__init__()
assert len(image_shape
) == 3, 'image_shape should have length 3: (C,H,W)'
self.image_shape = image_shape
self.embedding_dim = embedding_dim
c, h, w = image_shape
self.encode_c = nn.Parameter(torch.Tensor(1, c, 1, 1, embedding_dim))
self.encode_h = nn.Parameter(torch.Tensor(1, 1, h, 1, embedding_dim))
self.encode_w = nn.Parameter(torch.Tensor(1, 1, 1, w, embedding_dim))
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.encode_c, std=0.125 / math.sqrt(3 * self.
embedding_dim))
nn.init.normal_(self.encode_h, std=0.125 / math.sqrt(3 * self.
embedding_dim))
nn.init.normal_(self.encode_w, std=0.125 / math.sqrt(3 * self.
embedding_dim))
def forward(self, input_0):
primals_1 = self.encode_c
primals_3 = self.encode_h
primals_4 = self.encode_w
primals_2 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
| Nintorac/survae_experiments | PositionalEncodingImage | false | 906 | [
"MIT"
] | 0 | d68cc25e2604aab08b53617c1f3ffe4716f166c4 | https://github.com/Nintorac/survae_experiments/tree/d68cc25e2604aab08b53617c1f3ffe4716f166c4 | import math
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Learning positional embeddings for images.
Embeddings for channel, height and width are added to form the full positional embedding.
These encodings correspond to the ones from Sparse Transformers (https://arxiv.org/abs/1904.10509).
Args:
image_shape: Iterable, the shape of the image.
embedding_dim: int, the size of each embedding vector.
"""
def __init__(self, image_shape, embedding_dim):
super().__init__()
assert len(image_shape
) == 3, 'image_shape should have length 3: (C,H,W)'
self.image_shape = image_shape
self.embedding_dim = embedding_dim
c, h, w = image_shape
self.encode_c = nn.Parameter(torch.Tensor(1, c, 1, 1, embedding_dim))
self.encode_h = nn.Parameter(torch.Tensor(1, 1, h, 1, embedding_dim))
self.encode_w = nn.Parameter(torch.Tensor(1, 1, 1, w, embedding_dim))
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.encode_c, std=0.125 / math.sqrt(3 * self.
embedding_dim))
nn.init.normal_(self.encode_h, std=0.125 / math.sqrt(3 * self.
embedding_dim))
nn.init.normal_(self.encode_w, std=0.125 / math.sqrt(3 * self.
embedding_dim))
def forward(self, x):
return x + self.encode_c + self.encode_h + self.encode_w
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
LinearZeros | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/n3/cn3xefa3yajkd7wdkmubvtixzl5wmbx427xsxd2dcdu4gxfgqriv.py
# Topologically Sorted Source Nodes: [mul, exp, output_1], Original ATen: [aten.mul, aten.exp, aten.view]
# Source node to ATen node mapping:
# exp => exp
# mul => mul
# output_1 => mul_1, view_3
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_4, 3.0), kwargs = {})
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%mul,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, %exp), kwargs = {})
# %view_3 : [num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%view_2, [4, 4, 4, 4]), kwargs = {})
triton_poi_fused_exp_mul_view_0 = async_compile.triton('triton_poi_fused_exp_mul_view_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_exp_mul_view_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_exp_mul_view_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp2 = 3.0
tmp3 = tmp1 * tmp2
tmp4 = tl_math.exp(tmp3)
tmp5 = tmp0 * tmp4
tl.store(in_out_ptr0 + (x4), tmp5, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [output], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf2 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [mul, exp, output_1], Original ATen: [aten.mul, aten.exp, aten.view]
stream0 = get_raw_stream(0)
triton_poi_fused_exp_mul_view_0.run(buf2, buf0, primals_4, 256, grid=grid(256), stream=stream0)
return (buf2, primals_4, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class LinearZeros(nn.Linear):
def __init__(self, in_features, out_features, bias=True,
logscale_factor=3.0):
"""
Linear layer with zero initialization
:param in_features: size of each input sample
:type in_features: int
:param out_features: size of each output sample
:type out_features: int
:param bias: whether to learn an additive bias.
:type bias: bool
:param logscale_factor: factor of logscale
:type logscale_factor: float
"""
super().__init__(in_features, out_features, bias)
self.logscale_factor = logscale_factor
self.weight.data.zero_()
self.bias.data.zero_()
self.register_parameter('logs', nn.Parameter(torch.zeros(out_features))
)
def forward(self, x):
"""
Forward linear zero layer
:param x: input
:type x: torch.Tensor
:return: output
:rtype: torch.Tensor
"""
output = super().forward(x)
output *= torch.exp(self.logs * self.logscale_factor)
return output
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_features': 4, 'out_features': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_exp_mul_view_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = 3.0
tmp3 = tmp1 * tmp2
tmp4 = tl_math.exp(tmp3)
tmp5 = tmp0 * tmp4
tl.store(in_out_ptr0 + x4, tmp5, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf2 = buf1
del buf1
get_raw_stream(0)
triton_poi_fused_exp_mul_view_0[grid(256)](buf2, buf0, primals_4,
256, XBLOCK=128, num_warps=4, num_stages=1)
return buf2, primals_4, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf0
class LinearZerosNew(nn.Linear):
def __init__(self, in_features, out_features, bias=True,
logscale_factor=3.0):
"""
Linear layer with zero initialization
:param in_features: size of each input sample
:type in_features: int
:param out_features: size of each output sample
:type out_features: int
:param bias: whether to learn an additive bias.
:type bias: bool
:param logscale_factor: factor of logscale
:type logscale_factor: float
"""
super().__init__(in_features, out_features, bias)
self.logscale_factor = logscale_factor
self.weight.data.zero_()
self.bias.data.zero_()
self.register_parameter('logs', nn.Parameter(torch.zeros(out_features))
)
def forward(self, input_0):
primals_1 = self.weight
primals_2 = self.bias
primals_4 = self.logs
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
| NirDiamant/pytorch-glow | LinearZeros | false | 907 | [
"MIT"
] | 0 | 2ab11f3a8486b86a279fe4fa64f25aa91226ee8a | https://github.com/NirDiamant/pytorch-glow/tree/2ab11f3a8486b86a279fe4fa64f25aa91226ee8a | import torch
import torch.nn as nn
class Model(nn.Linear):
def __init__(self, in_features, out_features, bias=True,
logscale_factor=3.0):
"""
Linear layer with zero initialization
:param in_features: size of each input sample
:type in_features: int
:param out_features: size of each output sample
:type out_features: int
:param bias: whether to learn an additive bias.
:type bias: bool
:param logscale_factor: factor of logscale
:type logscale_factor: float
"""
super().__init__(in_features, out_features, bias)
self.logscale_factor = logscale_factor
self.weight.data.zero_()
self.bias.data.zero_()
self.register_parameter('logs', nn.Parameter(torch.zeros(out_features))
)
def forward(self, x):
"""
Forward linear zero layer
:param x: input
:type x: torch.Tensor
:return: output
:rtype: torch.Tensor
"""
output = super().forward(x)
output *= torch.exp(self.logs * self.logscale_factor)
return output
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
Conv2dZeros | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/zv/czvhqogqibwnm23q44xn6gzvpm2ac5f4wseb2e7zgfnmpvynwgoy.py
# Topologically Sorted Source Nodes: [x, mul, exp, x_1], Original ATen: [aten.convolution, aten.mul, aten.exp]
# Source node to ATen node mapping:
# exp => exp
# mul => mul
# x => convolution
# x_1 => mul_1
# Graph fragment:
# %convolution : [num_users=2] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [1, 1], [1, 1], False, [0, 0], 1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_4, 3), kwargs = {})
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%mul,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%convolution, %exp), kwargs = {})
triton_poi_fused_convolution_exp_mul_0 = async_compile.triton('triton_poi_fused_convolution_exp_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_exp_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_exp_mul_0(in_out_ptr0, in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 16) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = 3.0
tmp5 = tmp3 * tmp4
tmp6 = tl_math.exp(tmp5)
tmp7 = tmp2 * tmp6
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
tl.store(out_ptr0 + (x3), tmp7, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 1, 1), (1, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4), (64, 16, 4, 1))
buf1 = buf0; del buf0 # reuse
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x, mul, exp, x_1], Original ATen: [aten.convolution, aten.mul, aten.exp]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_exp_mul_0.run(buf1, primals_2, primals_4, buf2, 256, grid=grid(256), stream=stream0)
del primals_2
return (buf2, primals_1, primals_3, primals_4, buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 3, 3), (36, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 1, 1), (1, 1, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class ActNorm(nn.Module):
def __init__(self, num_channels, scale=1.0, logscale_factor=3.0,
batch_variance=False):
"""
Activation normalization layer
:param num_channels: number of channels
:type num_channels: int
:param scale: scale
:type scale: float
:param logscale_factor: factor for logscale
:type logscale_factor: float
:param batch_variance: use batch variance
:type batch_variance: bool
"""
super().__init__()
self.num_channels = num_channels
self.scale = scale
self.logscale_factor = logscale_factor
self.batch_variance = batch_variance
self.bias_inited = False
self.logs_inited = False
self.register_parameter('bias', nn.Parameter(torch.zeros(1, self.
num_channels, 1, 1)))
self.register_parameter('logs', nn.Parameter(torch.zeros(1, self.
num_channels, 1, 1)))
def actnorm_center(self, x, reverse=False):
"""
center operation of activation normalization
:param x: input
:type x: torch.Tensor
:param reverse: whether to reverse bias
:type reverse: bool
:return: centered input
:rtype: torch.Tensor
"""
if not self.bias_inited:
self.initialize_bias(x)
if not reverse:
return x + self.bias
else:
return x - self.bias
def actnorm_scale(self, x, logdet, reverse=False):
"""
scale operation of activation normalization
:param x: input
:type x: torch.Tensor
:param logdet: log determinant
:type logdet:
:param reverse: whether to reverse bias
:type reverse: bool
:return: centered input and logdet
:rtype: tuple(torch.Tensor, torch.Tensor)
"""
if not self.logs_inited:
self.initialize_logs(x)
logs = self.logs * self.logscale_factor
if not reverse:
x *= torch.exp(logs)
else:
x *= torch.exp(-logs)
if logdet is not None:
logdet_factor = ops.count_pixels(x)
dlogdet = torch.sum(logs) * logdet_factor
if reverse:
dlogdet *= -1
logdet += dlogdet
return x, logdet
def initialize_bias(self, x):
"""
Initialize bias
:param x: input
:type x: torch.Tensor
"""
if not self.training:
return
with torch.no_grad():
x_mean = -1.0 * ops.reduce_mean(x, dim=[0, 2, 3], keepdim=True)
self.bias.data.copy_(x_mean.data)
self.bias_inited = True
def initialize_logs(self, x):
"""
Initialize logs
:param x: input
:type x: torch.Tensor
"""
if not self.training:
return
with torch.no_grad():
if self.batch_variance:
x_var = ops.reduce_mean(x ** 2, keepdim=True)
else:
x_var = ops.reduce_mean(x ** 2, dim=[0, 2, 3], keepdim=True)
logs = torch.log(self.scale / (torch.sqrt(x_var) + 1e-06)
) / self.logscale_factor
self.logs.data.copy_(logs.data)
self.logs_inited = True
def forward(self, x, logdet=None, reverse=False):
"""
Forward activation normalization layer
:param x: input
:type x: torch.Tensor
:param logdet: log determinant
:type logdet:
:param reverse: whether to reverse bias
:type reverse: bool
:return: normalized input and logdet
:rtype: tuple(torch.Tensor, torch.Tensor)
"""
assert len(x.shape) == 4
assert x.shape[1
] == self.num_channels, 'Input shape should be NxCxHxW, however channels are {} instead of {}'.format(
x.shape[1], self.num_channels)
assert x.device == self.bias.device and x.device == self.logs.device, 'Expect input device {} instead of {}'.format(
self.bias.device, x.device)
if not reverse:
x = self.actnorm_center(x, reverse=False)
x, logdet = self.actnorm_scale(x, logdet, reverse=False)
else:
x, logdet = self.actnorm_scale(x, logdet, reverse=True)
x = self.actnorm_center(x, reverse=True)
return x, logdet
class Conv2d(nn.Conv2d):
@staticmethod
def get_padding(padding_type, kernel_size, stride):
"""
Get padding size.
mentioned in https://github.com/pytorch/pytorch/issues/3867#issuecomment-361775080
behaves as 'SAME' padding in TensorFlow
independent on input size when stride is 1
:param padding_type: type of padding in ['SAME', 'VALID']
:type padding_type: str
:param kernel_size: kernel size
:type kernel_size: tuple(int) or int
:param stride: stride
:type stride: int
:return: padding size
:rtype: tuple(int)
"""
assert padding_type in ['SAME', 'VALID'
], 'Unsupported padding type: {}'.format(padding_type)
if isinstance(kernel_size, int):
kernel_size = [kernel_size, kernel_size]
if padding_type == 'SAME':
assert stride == 1, "'SAME' padding only supports stride=1"
return tuple((k - 1) // 2 for k in kernel_size)
return tuple(0 for _ in kernel_size)
def __init__(self, in_channels, out_channels, kernel_size=(3, 3),
stride=1, padding_type='SAME', do_weightnorm=False, do_actnorm=True,
dilation=1, groups=1):
"""
Wrapper of nn.Conv2d with weight normalization and activation normalization
:param padding_type: type of padding in ['SAME', 'VALID']
:type padding_type: str
:param do_weightnorm: whether to do weight normalization after convolution
:type do_weightnorm: bool
:param do_actnorm: whether to do activation normalization after convolution
:type do_actnorm: bool
"""
padding = self.get_padding(padding_type, kernel_size, stride)
super().__init__(in_channels, out_channels, kernel_size, stride,
padding, dilation, groups, bias=not do_actnorm)
self.do_weight_norm = do_weightnorm
self.do_actnorm = do_actnorm
self.weight.data.normal_(mean=0.0, std=0.05)
if self.do_actnorm:
self.actnorm = ActNorm(out_channels)
else:
self.bias.data.zero_()
def forward(self, x):
"""
Forward wrapped Conv2d layer
:param x: input
:type x: torch.Tensor
:return: output
:rtype: torch.Tensor
"""
x = super().forward(x)
if self.do_actnorm:
x, _ = self.actnorm(x)
return x
class Conv2dZeros(nn.Conv2d):
def __init__(self, in_channels, out_channels, kernel_size=(3, 3),
stride=1, padding_type='SAME', logscale_factor=3, dilation=1,
groups=1, bias=True):
"""
Wrapper of nn.Conv2d with zero initialization and logs
:param padding_type: type of padding in ['SAME', 'VALID']
:type padding_type: str
:param logscale_factor: factor for logscale
:type logscale_factor: float
"""
padding = Conv2d.get_padding(padding_type, kernel_size, stride)
super().__init__(in_channels, out_channels, kernel_size, stride,
padding, dilation, groups, bias)
self.logscale_factor = logscale_factor
self.bias.data.zero_()
self.weight.data.zero_()
self.register_parameter('logs', nn.Parameter(torch.zeros(
out_channels, 1, 1)))
def forward(self, x):
"""
Forward wrapped Conv2d layer
:param x: input
:type x: torch.Tensor
:return: output
:rtype: torch.Tensor
"""
x = super().forward(x)
x *= torch.exp(self.logs * self.logscale_factor)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_convolution_exp_mul_0(in_out_ptr0, in_ptr0, in_ptr1,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = 3.0
tmp5 = tmp3 * tmp4
tmp6 = tl_math.exp(tmp5)
tmp7 = tmp2 * tmp6
tl.store(in_out_ptr0 + x3, tmp2, xmask)
tl.store(out_ptr0 + x3, tmp7, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 1, 1), (1, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4), (64, 16, 4, 1))
buf1 = buf0
del buf0
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_convolution_exp_mul_0[grid(256)](buf1, primals_2,
primals_4, buf2, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
return buf2, primals_1, primals_3, primals_4, buf1
class ActNorm(nn.Module):
def __init__(self, num_channels, scale=1.0, logscale_factor=3.0,
batch_variance=False):
"""
Activation normalization layer
:param num_channels: number of channels
:type num_channels: int
:param scale: scale
:type scale: float
:param logscale_factor: factor for logscale
:type logscale_factor: float
:param batch_variance: use batch variance
:type batch_variance: bool
"""
super().__init__()
self.num_channels = num_channels
self.scale = scale
self.logscale_factor = logscale_factor
self.batch_variance = batch_variance
self.bias_inited = False
self.logs_inited = False
self.register_parameter('bias', nn.Parameter(torch.zeros(1, self.
num_channels, 1, 1)))
self.register_parameter('logs', nn.Parameter(torch.zeros(1, self.
num_channels, 1, 1)))
def actnorm_center(self, x, reverse=False):
"""
center operation of activation normalization
:param x: input
:type x: torch.Tensor
:param reverse: whether to reverse bias
:type reverse: bool
:return: centered input
:rtype: torch.Tensor
"""
if not self.bias_inited:
self.initialize_bias(x)
if not reverse:
return x + self.bias
else:
return x - self.bias
def actnorm_scale(self, x, logdet, reverse=False):
"""
scale operation of activation normalization
:param x: input
:type x: torch.Tensor
:param logdet: log determinant
:type logdet:
:param reverse: whether to reverse bias
:type reverse: bool
:return: centered input and logdet
:rtype: tuple(torch.Tensor, torch.Tensor)
"""
if not self.logs_inited:
self.initialize_logs(x)
logs = self.logs * self.logscale_factor
if not reverse:
x *= torch.exp(logs)
else:
x *= torch.exp(-logs)
if logdet is not None:
logdet_factor = ops.count_pixels(x)
dlogdet = torch.sum(logs) * logdet_factor
if reverse:
dlogdet *= -1
logdet += dlogdet
return x, logdet
def initialize_bias(self, x):
"""
Initialize bias
:param x: input
:type x: torch.Tensor
"""
if not self.training:
return
with torch.no_grad():
x_mean = -1.0 * ops.reduce_mean(x, dim=[0, 2, 3], keepdim=True)
self.bias.data.copy_(x_mean.data)
self.bias_inited = True
def initialize_logs(self, x):
"""
Initialize logs
:param x: input
:type x: torch.Tensor
"""
if not self.training:
return
with torch.no_grad():
if self.batch_variance:
x_var = ops.reduce_mean(x ** 2, keepdim=True)
else:
x_var = ops.reduce_mean(x ** 2, dim=[0, 2, 3], keepdim=True)
logs = torch.log(self.scale / (torch.sqrt(x_var) + 1e-06)
) / self.logscale_factor
self.logs.data.copy_(logs.data)
self.logs_inited = True
def forward(self, x, logdet=None, reverse=False):
"""
Forward activation normalization layer
:param x: input
:type x: torch.Tensor
:param logdet: log determinant
:type logdet:
:param reverse: whether to reverse bias
:type reverse: bool
:return: normalized input and logdet
:rtype: tuple(torch.Tensor, torch.Tensor)
"""
assert len(x.shape) == 4
assert x.shape[1
] == self.num_channels, 'Input shape should be NxCxHxW, however channels are {} instead of {}'.format(
x.shape[1], self.num_channels)
assert x.device == self.bias.device and x.device == self.logs.device, 'Expect input device {} instead of {}'.format(
self.bias.device, x.device)
if not reverse:
x = self.actnorm_center(x, reverse=False)
x, logdet = self.actnorm_scale(x, logdet, reverse=False)
else:
x, logdet = self.actnorm_scale(x, logdet, reverse=True)
x = self.actnorm_center(x, reverse=True)
return x, logdet
class Conv2d(nn.Conv2d):
@staticmethod
def get_padding(padding_type, kernel_size, stride):
"""
Get padding size.
mentioned in https://github.com/pytorch/pytorch/issues/3867#issuecomment-361775080
behaves as 'SAME' padding in TensorFlow
independent on input size when stride is 1
:param padding_type: type of padding in ['SAME', 'VALID']
:type padding_type: str
:param kernel_size: kernel size
:type kernel_size: tuple(int) or int
:param stride: stride
:type stride: int
:return: padding size
:rtype: tuple(int)
"""
assert padding_type in ['SAME', 'VALID'
], 'Unsupported padding type: {}'.format(padding_type)
if isinstance(kernel_size, int):
kernel_size = [kernel_size, kernel_size]
if padding_type == 'SAME':
assert stride == 1, "'SAME' padding only supports stride=1"
return tuple((k - 1) // 2 for k in kernel_size)
return tuple(0 for _ in kernel_size)
def __init__(self, in_channels, out_channels, kernel_size=(3, 3),
stride=1, padding_type='SAME', do_weightnorm=False, do_actnorm=True,
dilation=1, groups=1):
"""
Wrapper of nn.Conv2d with weight normalization and activation normalization
:param padding_type: type of padding in ['SAME', 'VALID']
:type padding_type: str
:param do_weightnorm: whether to do weight normalization after convolution
:type do_weightnorm: bool
:param do_actnorm: whether to do activation normalization after convolution
:type do_actnorm: bool
"""
padding = self.get_padding(padding_type, kernel_size, stride)
super().__init__(in_channels, out_channels, kernel_size, stride,
padding, dilation, groups, bias=not do_actnorm)
self.do_weight_norm = do_weightnorm
self.do_actnorm = do_actnorm
self.weight.data.normal_(mean=0.0, std=0.05)
if self.do_actnorm:
self.actnorm = ActNorm(out_channels)
else:
self.bias.data.zero_()
def forward(self, x):
"""
Forward wrapped Conv2d layer
:param x: input
:type x: torch.Tensor
:return: output
:rtype: torch.Tensor
"""
x = super().forward(x)
if self.do_actnorm:
x, _ = self.actnorm(x)
return x
class Conv2dZerosNew(nn.Conv2d):
def __init__(self, in_channels, out_channels, kernel_size=(3, 3),
stride=1, padding_type='SAME', logscale_factor=3, dilation=1,
groups=1, bias=True):
"""
Wrapper of nn.Conv2d with zero initialization and logs
:param padding_type: type of padding in ['SAME', 'VALID']
:type padding_type: str
:param logscale_factor: factor for logscale
:type logscale_factor: float
"""
padding = Conv2d.get_padding(padding_type, kernel_size, stride)
super().__init__(in_channels, out_channels, kernel_size, stride,
padding, dilation, groups, bias)
self.logscale_factor = logscale_factor
self.bias.data.zero_()
self.weight.data.zero_()
self.register_parameter('logs', nn.Parameter(torch.zeros(
out_channels, 1, 1)))
def forward(self, input_0):
primals_1 = self.weight
primals_2 = self.bias
primals_4 = self.logs
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
| NirDiamant/pytorch-glow | Conv2dZeros | false | 908 | [
"MIT"
] | 0 | 2ab11f3a8486b86a279fe4fa64f25aa91226ee8a | https://github.com/NirDiamant/pytorch-glow/tree/2ab11f3a8486b86a279fe4fa64f25aa91226ee8a | import torch
import torch.nn as nn
class ActNorm(nn.Module):
def __init__(self, num_channels, scale=1.0, logscale_factor=3.0,
batch_variance=False):
"""
Activation normalization layer
:param num_channels: number of channels
:type num_channels: int
:param scale: scale
:type scale: float
:param logscale_factor: factor for logscale
:type logscale_factor: float
:param batch_variance: use batch variance
:type batch_variance: bool
"""
super().__init__()
self.num_channels = num_channels
self.scale = scale
self.logscale_factor = logscale_factor
self.batch_variance = batch_variance
self.bias_inited = False
self.logs_inited = False
self.register_parameter('bias', nn.Parameter(torch.zeros(1, self.
num_channels, 1, 1)))
self.register_parameter('logs', nn.Parameter(torch.zeros(1, self.
num_channels, 1, 1)))
def actnorm_center(self, x, reverse=False):
"""
center operation of activation normalization
:param x: input
:type x: torch.Tensor
:param reverse: whether to reverse bias
:type reverse: bool
:return: centered input
:rtype: torch.Tensor
"""
if not self.bias_inited:
self.initialize_bias(x)
if not reverse:
return x + self.bias
else:
return x - self.bias
def actnorm_scale(self, x, logdet, reverse=False):
"""
scale operation of activation normalization
:param x: input
:type x: torch.Tensor
:param logdet: log determinant
:type logdet:
:param reverse: whether to reverse bias
:type reverse: bool
:return: centered input and logdet
:rtype: tuple(torch.Tensor, torch.Tensor)
"""
if not self.logs_inited:
self.initialize_logs(x)
logs = self.logs * self.logscale_factor
if not reverse:
x *= torch.exp(logs)
else:
x *= torch.exp(-logs)
if logdet is not None:
logdet_factor = ops.count_pixels(x)
dlogdet = torch.sum(logs) * logdet_factor
if reverse:
dlogdet *= -1
logdet += dlogdet
return x, logdet
def initialize_bias(self, x):
"""
Initialize bias
:param x: input
:type x: torch.Tensor
"""
if not self.training:
return
with torch.no_grad():
x_mean = -1.0 * ops.reduce_mean(x, dim=[0, 2, 3], keepdim=True)
self.bias.data.copy_(x_mean.data)
self.bias_inited = True
def initialize_logs(self, x):
"""
Initialize logs
:param x: input
:type x: torch.Tensor
"""
if not self.training:
return
with torch.no_grad():
if self.batch_variance:
x_var = ops.reduce_mean(x ** 2, keepdim=True)
else:
x_var = ops.reduce_mean(x ** 2, dim=[0, 2, 3], keepdim=True)
logs = torch.log(self.scale / (torch.sqrt(x_var) + 1e-06)
) / self.logscale_factor
self.logs.data.copy_(logs.data)
self.logs_inited = True
def forward(self, x, logdet=None, reverse=False):
"""
Forward activation normalization layer
:param x: input
:type x: torch.Tensor
:param logdet: log determinant
:type logdet:
:param reverse: whether to reverse bias
:type reverse: bool
:return: normalized input and logdet
:rtype: tuple(torch.Tensor, torch.Tensor)
"""
assert len(x.shape) == 4
assert x.shape[1
] == self.num_channels, 'Input shape should be NxCxHxW, however channels are {} instead of {}'.format(
x.shape[1], self.num_channels)
assert x.
# ... truncated (>4000 chars) for memory efficiency |
GaussianKernel | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xe/cxedqk5kxfk4utnfrixbnyqhrlbgj4i6sv6keatvftozhw66gz74.py
# Topologically Sorted Source Nodes: [sub, pow_1, l2_distance_square, neg, mean, mul, mul_1, truediv, exp], Original ATen: [aten.sub, aten.pow, aten.sum, aten.neg, aten.mean, aten.mul, aten.div, aten.exp]
# Source node to ATen node mapping:
# exp => exp
# l2_distance_square => sum_1
# mean => mean
# mul => mul
# mul_1 => mul_1
# neg => neg
# pow_1 => pow_1
# sub => sub
# truediv => div
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%unsqueeze, %unsqueeze_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %sum_1 : [num_users=2] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [2]), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%sum_1,), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%sum_1,), kwargs = {})
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mean, 1.0), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, 2), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%neg, %mul_1), kwargs = {})
# %exp : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%div,), kwargs = {})
triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0 = async_compile.triton('triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 8, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0(in_out_ptr0, in_ptr0, out_ptr1, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = (rindex // 16) % 4
r2 = (rindex // 64)
r3 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + (64*r1)), None, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (r0 + (64*r2)), None, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (16 + r0 + (64*r1)), None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (16 + r0 + (64*r2)), None, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (32 + r0 + (64*r1)), None, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (32 + r0 + (64*r2)), None, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr0 + (48 + r0 + (64*r1)), None, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr0 + (48 + r0 + (64*r2)), None, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp6 = tmp4 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tmp3 + tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp19 = tl.broadcast_to(tmp18, [RBLOCK])
tmp21 = triton_helpers.promote_to_tensor(tl.sum(tmp19, 0))
tmp22 = 256.0
tmp23 = tmp21 / tmp22
tmp24 = 1.0
tmp25 = tmp23 * tmp24
tmp26 = -tmp18
tmp27 = 2.0
tmp28 = tmp25 * tmp27
tmp29 = tmp26 / tmp28
tmp30 = tl_math.exp(tmp29)
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp25, None)
tl.store(out_ptr1 + (tl.broadcast_to(r3, [RBLOCK])), tmp30, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf1 = empty_strided_cuda((), (), torch.float32)
buf2 = buf1; del buf1 # reuse
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sub, pow_1, l2_distance_square, neg, mean, mul, mul_1, truediv, exp], Original ATen: [aten.sub, aten.pow, aten.sum, aten.neg, aten.mean, aten.mul, aten.div, aten.exp]
stream0 = get_raw_stream(0)
triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0.run(buf2, arg0_1, buf3, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
return (buf3, buf2, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from typing import Optional
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class GaussianKernel(nn.Module):
"""Gaussian Kernel Matrix
Gaussian Kernel k is defined by
.. math::
k(x_1, x_2) = \\exp \\left( - \\dfrac{\\| x_1 - x_2 \\|^2}{2\\sigma^2} \\right)
where :math:`x_1, x_2 \\in R^d` are 1-d tensors.
Gaussian Kernel Matrix K is defined on input group :math:`X=(x_1, x_2, ..., x_m),`
.. math::
K(X)_{i,j} = k(x_i, x_j)
Also by default, during training this layer keeps running estimates of the
mean of L2 distances, which are then used to set hyperparameter :math:`\\sigma`.
Mathematically, the estimation is :math:`\\sigma^2 = \\dfrac{\\alpha}{n^2}\\sum_{i,j} \\| x_i - x_j \\|^2`.
If :attr:`track_running_stats` is set to ``False``, this layer then does not
keep running estimates, and use a fixed :math:`\\sigma` instead.
Parameters:
- sigma (float, optional): bandwidth :math:`\\sigma`. Default: None
- track_running_stats (bool, optional): If ``True``, this module tracks the running mean of :math:`\\sigma^2`.
Otherwise, it won't track such statistics and always uses fix :math:`\\sigma^2`. Default: ``True``
- alpha (float, optional): :math:`\\alpha` which decides the magnitude of :math:`\\sigma^2` when track_running_stats is set to ``True``
Inputs:
- X (tensor): input group :math:`X`
Shape:
- Inputs: :math:`(minibatch, F)` where F means the dimension of input features.
- Outputs: :math:`(minibatch, minibatch)`
"""
def __init__(self, sigma: 'Optional[float]'=None, track_running_stats:
'Optional[bool]'=True, alpha: 'Optional[float]'=1.0):
super(GaussianKernel, self).__init__()
assert track_running_stats or sigma is not None
self.sigma_square = torch.tensor(sigma * sigma
) if sigma is not None else None
self.track_running_stats = track_running_stats
self.alpha = alpha
def forward(self, X: 'torch.Tensor') ->torch.Tensor:
l2_distance_square = ((X.unsqueeze(0) - X.unsqueeze(1)) ** 2).sum(2)
if self.track_running_stats:
self.sigma_square = self.alpha * torch.mean(l2_distance_square.
detach())
return torch.exp(-l2_distance_square / (2 * self.sigma_square))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from typing import Optional
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0(in_out_ptr0,
in_ptr0, out_ptr1, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = rindex // 16 % 4
r2 = rindex // 64
r3 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + 64 * r1), None, eviction_policy='evict_last'
)
tmp1 = tl.load(in_ptr0 + (r0 + 64 * r2), None, eviction_policy='evict_last'
)
tmp4 = tl.load(in_ptr0 + (16 + r0 + 64 * r1), None, eviction_policy=
'evict_last')
tmp5 = tl.load(in_ptr0 + (16 + r0 + 64 * r2), None, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (32 + r0 + 64 * r1), None, eviction_policy=
'evict_last')
tmp10 = tl.load(in_ptr0 + (32 + r0 + 64 * r2), None, eviction_policy=
'evict_last')
tmp14 = tl.load(in_ptr0 + (48 + r0 + 64 * r1), None, eviction_policy=
'evict_last')
tmp15 = tl.load(in_ptr0 + (48 + r0 + 64 * r2), None, eviction_policy=
'evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp6 = tmp4 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tmp3 + tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp19 = tl.broadcast_to(tmp18, [RBLOCK])
tmp21 = triton_helpers.promote_to_tensor(tl.sum(tmp19, 0))
tmp22 = 256.0
tmp23 = tmp21 / tmp22
tmp24 = 1.0
tmp25 = tmp23 * tmp24
tmp26 = -tmp18
tmp27 = 2.0
tmp28 = tmp25 * tmp27
tmp29 = tmp26 / tmp28
tmp30 = tl_math.exp(tmp29)
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp25, None)
tl.store(out_ptr1 + tl.broadcast_to(r3, [RBLOCK]), tmp30, None)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf1 = empty_strided_cuda((), (), torch.float32)
buf2 = buf1
del buf1
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_per_fused_div_exp_mean_mul_neg_pow_sub_sum_0[grid(1)](buf2,
arg0_1, buf3, 1, 256, num_warps=2, num_stages=1)
del arg0_1
return buf3, buf2
class GaussianKernelNew(nn.Module):
"""Gaussian Kernel Matrix
Gaussian Kernel k is defined by
.. math::
k(x_1, x_2) = \\exp \\left( - \\dfrac{\\| x_1 - x_2 \\|^2}{2\\sigma^2} \\right)
where :math:`x_1, x_2 \\in R^d` are 1-d tensors.
Gaussian Kernel Matrix K is defined on input group :math:`X=(x_1, x_2, ..., x_m),`
.. math::
K(X)_{i,j} = k(x_i, x_j)
Also by default, during training this layer keeps running estimates of the
mean of L2 distances, which are then used to set hyperparameter :math:`\\sigma`.
Mathematically, the estimation is :math:`\\sigma^2 = \\dfrac{\\alpha}{n^2}\\sum_{i,j} \\| x_i - x_j \\|^2`.
If :attr:`track_running_stats` is set to ``False``, this layer then does not
keep running estimates, and use a fixed :math:`\\sigma` instead.
Parameters:
- sigma (float, optional): bandwidth :math:`\\sigma`. Default: None
- track_running_stats (bool, optional): If ``True``, this module tracks the running mean of :math:`\\sigma^2`.
Otherwise, it won't track such statistics and always uses fix :math:`\\sigma^2`. Default: ``True``
- alpha (float, optional): :math:`\\alpha` which decides the magnitude of :math:`\\sigma^2` when track_running_stats is set to ``True``
Inputs:
- X (tensor): input group :math:`X`
Shape:
- Inputs: :math:`(minibatch, F)` where F means the dimension of input features.
- Outputs: :math:`(minibatch, minibatch)`
"""
def __init__(self, sigma: 'Optional[float]'=None, track_running_stats:
'Optional[bool]'=True, alpha: 'Optional[float]'=1.0):
super(GaussianKernelNew, self).__init__()
assert track_running_stats or sigma is not None
self.sigma_square = torch.tensor(sigma * sigma
) if sigma is not None else None
self.track_running_stats = track_running_stats
self.alpha = alpha
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| NiteshBharadwaj/ignoringhumanpose | GaussianKernel | false | 909 | [
"MIT"
] | 0 | 1fb7a063fded9cff18f7de4e1d71845983077256 | https://github.com/NiteshBharadwaj/ignoringhumanpose/tree/1fb7a063fded9cff18f7de4e1d71845983077256 | import torch
from typing import Optional
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class Model(nn.Module):
"""Gaussian Kernel Matrix
Gaussian Kernel k is defined by
.. math::
k(x_1, x_2) = \\exp \\left( - \\dfrac{\\| x_1 - x_2 \\|^2}{2\\sigma^2} \\right)
where :math:`x_1, x_2 \\in R^d` are 1-d tensors.
Gaussian Kernel Matrix K is defined on input group :math:`X=(x_1, x_2, ..., x_m),`
.. math::
K(X)_{i,j} = k(x_i, x_j)
Also by default, during training this layer keeps running estimates of the
mean of L2 distances, which are then used to set hyperparameter :math:`\\sigma`.
Mathematically, the estimation is :math:`\\sigma^2 = \\dfrac{\\alpha}{n^2}\\sum_{i,j} \\| x_i - x_j \\|^2`.
If :attr:`track_running_stats` is set to ``False``, this layer then does not
keep running estimates, and use a fixed :math:`\\sigma` instead.
Parameters:
- sigma (float, optional): bandwidth :math:`\\sigma`. Default: None
- track_running_stats (bool, optional): If ``True``, this module tracks the running mean of :math:`\\sigma^2`.
Otherwise, it won't track such statistics and always uses fix :math:`\\sigma^2`. Default: ``True``
- alpha (float, optional): :math:`\\alpha` which decides the magnitude of :math:`\\sigma^2` when track_running_stats is set to ``True``
Inputs:
- X (tensor): input group :math:`X`
Shape:
- Inputs: :math:`(minibatch, F)` where F means the dimension of input features.
- Outputs: :math:`(minibatch, minibatch)`
"""
def __init__(self, sigma: 'Optional[float]'=None, track_running_stats:
'Optional[bool]'=True, alpha: 'Optional[float]'=1.0):
super().__init__()
assert track_running_stats or sigma is not None
self.sigma_square = torch.tensor(sigma * sigma
) if sigma is not None else None
self.track_running_stats = track_running_stats
self.alpha = alpha
def forward(self, X: 'torch.Tensor') ->torch.Tensor:
l2_distance_square = ((X.unsqueeze(0) - X.unsqueeze(1)) ** 2).sum(2)
if self.track_running_stats:
self.sigma_square = self.alpha * torch.mean(l2_distance_square.
detach())
return torch.exp(-l2_distance_square / (2 * self.sigma_square))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
Multihead_Attention_Layer | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/sq/csqwykl3c2yrclpjal3rybjtw5arriact6hjsdztyo22onsct7qj.py
# Topologically Sorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
# Graph fragment:
# %clone_default : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%getitem,), kwargs = {memory_format: torch.contiguous_format})
# %mul_scalar : [num_users=1] = call_function[target=torch.ops.aten.mul.Scalar](args = (%clone_default, 1.0), kwargs = {})
triton_poi_fused_0 = async_compile.triton('triton_poi_fused_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (12*x1)), xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/dl/cdlt6h63nkfd6ocug72tr3k6f2jfxeyryh2jzumkigktbnbhikx2.py
# Topologically Sorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
# Graph fragment:
# %mul_scalar_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Scalar](args = (%permute_default, 1.0), kwargs = {})
triton_poi_fused_1 = async_compile.triton('triton_poi_fused_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4 + x0 + (12*x1)), xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/3o/c3oaxjlvvp4afd7ykuqeeshinhwcjxyjj3abqkwr3yqqinwm4e4f.py
# Topologically Sorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
# Graph fragment:
# %amax_default : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%view_default_2, [-1], True), kwargs = {})
# %sub_tensor : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_default_2, %amax_default), kwargs = {})
# %exp_default : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub_tensor,), kwargs = {})
triton_poi_fused_2 = async_compile.triton('triton_poi_fused_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2w/c2wy7nnmjg5t737ww7jovvpbthc2goznalidqgo6rnve5cme7k3g.py
# Topologically Sorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
# Graph fragment:
# %sum_dim_int_list : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp_default, [-1], True), kwargs = {})
# %div_tensor : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp_default, %sum_dim_int_list), kwargs = {})
# %eq_scalar : [num_users=1] = call_function[target=torch.ops.aten.eq.Scalar](args = (%view_default_2, -inf), kwargs = {})
# %logical_not_default : [num_users=1] = call_function[target=torch.ops.aten.logical_not.default](args = (%eq_scalar,), kwargs = {})
# %any_dim : [num_users=1] = call_function[target=torch.ops.aten.any.dim](args = (%logical_not_default, -1, True), kwargs = {})
# %logical_not_default_1 : [num_users=1] = call_function[target=torch.ops.aten.logical_not.default](args = (%any_dim,), kwargs = {})
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([4, 4, 4, 4], 0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_self : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%logical_not_default_1, %full_default, %div_tensor), kwargs = {})
triton_poi_fused_3 = async_compile.triton('triton_poi_fused_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 9, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr1 + (x2), xmask)
tmp26 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp31 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp1 = float("-inf")
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = (tmp4 != 0)
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = (tmp9 != 0)
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = (tmp15 != 0)
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = (tmp21 != 0)
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + (x2), tmp35, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/5z/c5zfhxx74swnamnhqv47vziqfsfs5gjynqz3d6dvcra34sp4pdpf.py
# Topologically Sorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
# Graph fragment:
# %clone_default_2 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%getitem_2,), kwargs = {memory_format: torch.contiguous_format})
triton_poi_fused_4 = async_compile.triton('triton_poi_fused_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_4(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (8 + x0 + (12*x1)), xmask)
tl.store(out_ptr0 + (x2), tmp0, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (12, 4, 1), (4, 1, 1))
assert_size_stride(primals_3, (4, 4, 1), (4, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [qkv], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf0, (4, 12, 4), (48, 4, 1))
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
stream0 = get_raw_stream(0)
triton_poi_fused_0.run(buf0, buf1, 64, grid=grid(64), stream=stream0)
buf2 = empty_strided_cuda((4, 4, 1, 4), (16, 4, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
triton_poi_fused_1.run(buf0, buf2, 64, grid=grid(64), stream=stream0)
buf3 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.bmm(reinterpret_tensor(buf1, (16, 4, 1), (4, 1, 0), 0), reinterpret_tensor(buf2, (16, 1, 4), (4, 0, 1), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
triton_poi_fused_2.run(buf3, buf4, 256, grid=grid(256), stream=stream0)
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
triton_poi_fused_3.run(buf3, buf4, buf5, 256, grid=grid(256), stream=stream0)
del buf3
del buf4
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
triton_poi_fused_4.run(buf0, buf6, 64, grid=grid(64), stream=stream0)
del buf0
buf7 = empty_strided_cuda((16, 4, 1), (4, 1, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.bmm(reinterpret_tensor(buf5, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf6, (16, 4, 1), (4, 1, 0), 0), out=buf7)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.convolution]
buf8 = extern_kernels.convolution(reinterpret_tensor(buf7, (4, 4, 4), (16, 4, 1), 0), primals_3, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf8, (4, 4, 4), (16, 4, 1))
return (buf8, primals_1, primals_2, primals_3, buf5, reinterpret_tensor(buf6, (16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf1, (16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 4), 0), reinterpret_tensor(buf7, (4, 4, 4), (16, 4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((12, 4, 1), (4, 1, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 1), (4, 1, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def scaled_self_attention(q, k, v, key_size):
weight = torch.matmul(q, k)
weight = F.softmax(weight / math.sqrt(key_size), dim=-1)
attention = torch.matmul(weight, v)
return attention
class Multihead_Attention_Layer(nn.Module):
def __init__(self, embedding, heads):
super(Multihead_Attention_Layer, self).__init__()
self.embedding = embedding
self.heads = heads
assert self.embedding % self.heads == 0, 'The number of embedding channels must be divisible by the number of heads'
self.head_dim = self.embedding // self.heads
self.embedding_layer = nn.Conv1d(self.embedding, self.embedding * 3,
kernel_size=1, bias=False)
self.out = nn.Conv1d(self.embedding, self.embedding, kernel_size=1,
bias=False)
def forward(self, x):
batch_size, channels, num_points = x.size()
qkv = self.embedding_layer(x)
qkv = qkv.permute(0, 2, 1)
qkv = qkv.reshape(batch_size, num_points, self.heads, 3 * self.head_dim
)
qkv = qkv.permute(0, 2, 1, 3)
q, k, v = torch.chunk(qkv, 3, dim=-1)
k = k.permute(0, 1, 3, 2)
values = scaled_self_attention(q, k, v, self.embedding / self.heads)
values = values.permute(0, 2, 1, 3)
values = values.reshape(batch_size, num_points, channels)
x = values.permute(0, 2, 1)
x = self.out(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'embedding': 4, 'heads': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 12 * x1), xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x2, tmp2, xmask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4 + x0 + 12 * x1), xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x2, tmp2, xmask)
@triton.jit
def triton_poi_fused_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + x2, xmask)
tmp26 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp29 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp31 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = float('-inf')
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = tmp4 != 0
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = tmp9 != 0
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = tmp15 != 0
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = tmp21 != 0
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + x2, tmp35, xmask)
@triton.jit
def triton_poi_fused_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (8 + x0 + 12 * x1), xmask)
tl.store(out_ptr0 + x2, tmp0, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (12, 4, 1), (4, 1, 1))
assert_size_stride(primals_3, (4, 4, 1), (4, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf0, (4, 12, 4), (48, 4, 1))
buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(64)](buf0, buf1, 64, XBLOCK=64, num_warps=1,
num_stages=1)
buf2 = empty_strided_cuda((4, 4, 1, 4), (16, 4, 4, 1), torch.float32)
triton_poi_fused_1[grid(64)](buf0, buf2, 64, XBLOCK=64, num_warps=1,
num_stages=1)
buf3 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf1, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf2, (16, 1, 4), (4, 0, 1), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_2[grid(256)](buf3, buf4, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_3[grid(256)](buf3, buf4, buf5, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del buf3
del buf4
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_4[grid(64)](buf0, buf6, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del buf0
buf7 = empty_strided_cuda((16, 4, 1), (4, 1, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf5, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf6, (16, 4, 1), (4, 1, 0), 0), out=buf7)
buf8 = extern_kernels.convolution(reinterpret_tensor(buf7, (4, 4, 4
), (16, 4, 1), 0), primals_3, stride=(1,), padding=(0,),
dilation=(1,), transposed=False, output_padding=(0,), groups=1,
bias=None)
assert_size_stride(buf8, (4, 4, 4), (16, 4, 1))
return buf8, primals_1, primals_2, primals_3, buf5, reinterpret_tensor(buf6
, (16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf1, (16, 1, 4), (
4, 1, 1), 0), reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 4), 0
), reinterpret_tensor(buf7, (4, 4, 4), (16, 4, 1), 0)
def scaled_self_attention(q, k, v, key_size):
weight = torch.matmul(q, k)
weight = F.softmax(weight / math.sqrt(key_size), dim=-1)
attention = torch.matmul(weight, v)
return attention
class Multihead_Attention_LayerNew(nn.Module):
def __init__(self, embedding, heads):
super(Multihead_Attention_LayerNew, self).__init__()
self.embedding = embedding
self.heads = heads
assert self.embedding % self.heads == 0, 'The number of embedding channels must be divisible by the number of heads'
self.head_dim = self.embedding // self.heads
self.embedding_layer = nn.Conv1d(self.embedding, self.embedding * 3,
kernel_size=1, bias=False)
self.out = nn.Conv1d(self.embedding, self.embedding, kernel_size=1,
bias=False)
def forward(self, input_0):
primals_2 = self.embedding_layer.weight
primals_3 = self.out.weight
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| NilsLusch/Point-Cloud-Transformer | Multihead_Attention_Layer | false | 910 | [
"MIT"
] | 0 | 84a16b45b8949bbf8e7730b10bd5835e2ab4e642 | https://github.com/NilsLusch/Point-Cloud-Transformer/tree/84a16b45b8949bbf8e7730b10bd5835e2ab4e642 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def scaled_self_attention(q, k, v, key_size):
weight = torch.matmul(q, k)
weight = F.softmax(weight / math.sqrt(key_size), dim=-1)
attention = torch.matmul(weight, v)
return attention
class Model(nn.Module):
def __init__(self, embedding, heads):
super().__init__()
self.embedding = embedding
self.heads = heads
assert self.embedding % self.heads == 0, 'The number of embedding channels must be divisible by the number of heads'
self.head_dim = self.embedding // self.heads
self.embedding_layer = nn.Conv1d(self.embedding, self.embedding * 3,
kernel_size=1, bias=False)
self.out = nn.Conv1d(self.embedding, self.embedding, kernel_size=1,
bias=False)
def forward(self, x):
batch_size, channels, num_points = x.size()
qkv = self.embedding_layer(x)
qkv = qkv.permute(0, 2, 1)
qkv = qkv.reshape(batch_size, num_points, self.heads, 3 * self.head_dim
)
qkv = qkv.permute(0, 2, 1, 3)
q, k, v = torch.chunk(qkv, 3, dim=-1)
k = k.permute(0, 1, 3, 2)
values = scaled_self_attention(q, k, v, self.embedding / self.heads)
values = values.permute(0, 2, 1, 3)
values = values.reshape(batch_size, num_points, channels)
x = values.permute(0, 2, 1)
x = self.out(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
Conv | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/iu/ciuxern2omgit5ovksuiwlddxkww6e3pkid4q2h3sauzn5rbd35z.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# x_1 => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%permute, %primals_2, %primals_3, [1], [0], [1], False, [0], 1), kwargs = {})
triton_poi_fused_convolution_0 = async_compile.triton('triton_poi_fused_convolution_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16, 4], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = (yindex // 4)
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + (4*x2) + (16*y1)), xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + (4*y3)), tmp0, xmask & ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/au/cau4pihcaptiev5y2ewn2o2nvrwhk7hogc72cofmmtbyv4rxc2oy.py
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# x_1 => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%permute, %primals_2, %primals_3, [1], [0], [1], False, [0], 1), kwargs = {})
triton_poi_fused_convolution_1 = async_compile.triton('triton_poi_fused_convolution_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 4) % 4
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 1), (4, 1, 1))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.convolution]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_0.run(primals_1, buf0, 16, 4, grid=grid(16, 4), stream=stream0)
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 4), (16, 4, 1))
del buf0
buf2 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [x_1], Original ATen: [aten.convolution]
triton_poi_fused_convolution_1.run(buf2, primals_3, 64, grid=grid(64), stream=stream0)
del primals_3
return (reinterpret_tensor(buf2, (4, 4, 4), (16, 1, 4), 0), primals_2, reinterpret_tensor(primals_1, (4, 4, 4), (16, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 1), (4, 1, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class Conv(nn.Module):
"""
Convolution Module
"""
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
padding=0, dilation=1, bias=True, w_init='linear'):
"""
:param in_channels: dimension of input
:param out_channels: dimension of output
:param kernel_size: size of kernel
:param stride: size of stride
:param padding: size of padding
:param dilation: dilation rate
:param bias: boolean. if True, bias is included.
:param w_init: str. weight inits with xavier initialization.
"""
super(Conv, self).__init__()
self.conv = nn.Conv1d(in_channels, out_channels, kernel_size=
kernel_size, stride=stride, padding=padding, dilation=dilation,
bias=bias)
def forward(self, x):
x = x.contiguous().transpose(1, 2)
x = self.conv(x)
x = x.contiguous().transpose(1, 2)
return x
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_0(in_ptr0, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 1), (4, 1, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(16, 4)](primals_1, buf0, 16, 4,
XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1)
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 4), (16, 4, 1))
del buf0
buf2 = buf1
del buf1
triton_poi_fused_convolution_1[grid(64)](buf2, primals_3, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_3
return reinterpret_tensor(buf2, (4, 4, 4), (16, 1, 4), 0
), primals_2, reinterpret_tensor(primals_1, (4, 4, 4), (16, 1, 4), 0)
class ConvNew(nn.Module):
"""
Convolution Module
"""
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
padding=0, dilation=1, bias=True, w_init='linear'):
"""
:param in_channels: dimension of input
:param out_channels: dimension of output
:param kernel_size: size of kernel
:param stride: size of stride
:param padding: size of padding
:param dilation: dilation rate
:param bias: boolean. if True, bias is included.
:param w_init: str. weight inits with xavier initialization.
"""
super(ConvNew, self).__init__()
self.conv = nn.Conv1d(in_channels, out_channels, kernel_size=
kernel_size, stride=stride, padding=padding, dilation=dilation,
bias=bias)
def forward(self, input_0):
primals_2 = self.conv.weight
primals_3 = self.conv.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Numb523/FastSpeech2_emotion | Conv | false | 911 | [
"MIT"
] | 0 | a541ce89ddf66625ee57c0a294d0bec1ae701f0c | https://github.com/Numb523/FastSpeech2_emotion/tree/a541ce89ddf66625ee57c0a294d0bec1ae701f0c | import torch
import torch.nn as nn
class Model(nn.Module):
"""
Convolution Module
"""
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
padding=0, dilation=1, bias=True, w_init='linear'):
"""
:param in_channels: dimension of input
:param out_channels: dimension of output
:param kernel_size: size of kernel
:param stride: size of stride
:param padding: size of padding
:param dilation: dilation rate
:param bias: boolean. if True, bias is included.
:param w_init: str. weight inits with xavier initialization.
"""
super().__init__()
self.conv = nn.Conv1d(in_channels, out_channels, kernel_size=
kernel_size, stride=stride, padding=padding, dilation=dilation,
bias=bias)
def forward(self, x):
x = x.contiguous().transpose(1, 2)
x = self.conv(x)
x = x.contiguous().transpose(1, 2)
return x
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
Theta | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/cv/ccvh5pcjc57oskvep5v2okylauufvnesitt2j6dbyvr6fh3vwuer.py
# Topologically Sorted Source Nodes: [output], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# output => mul
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_1, 1.0), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/7k/c7kfhnpi6ogpuakmzhvlayoy3mb73znoruxx7bdkt5awlqvsi5js.py
# Topologically Sorted Source Nodes: [autograd_function_apply], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# autograd_function_apply => mul_1
# Graph fragment:
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 1.0), kwargs = {})
triton_poi_fused_mul_1 = async_compile.triton('triton_poi_fused_mul_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [output], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(primals_1, buf0, 256, grid=grid(256), stream=stream0)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf1)
del primals_2
buf2 = reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [autograd_function_apply], Original ATen: [aten.mul]
triton_poi_fused_mul_1.run(buf2, primals_3, 256, grid=grid(256), stream=stream0)
del primals_3
return (buf2, reinterpret_tensor(buf0, (64, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| from torch.autograd import Function
import torch
from typing import Optional
from typing import Tuple
import torch.nn as nn
from typing import Any
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class GradientReverseFunction(Function):
@staticmethod
def forward(ctx: 'Any', input: 'torch.Tensor', coeff: 'Optional[float]'=1.0
) ->torch.Tensor:
ctx.coeff = coeff
output = input * 1.0
return output
@staticmethod
def backward(ctx: 'Any', grad_output: 'torch.Tensor') ->Tuple[torch.
Tensor, Any]:
return grad_output.neg() * ctx.coeff, None
class GradientReverseLayer(nn.Module):
def __init__(self):
super(GradientReverseLayer, self).__init__()
def forward(self, *input):
return GradientReverseFunction.apply(*input)
class Theta(nn.Module):
"""
maximize loss respect to :math:` heta`
minimize loss respect to features
"""
def __init__(self, dim: 'int'):
super(Theta, self).__init__()
self.grl1 = GradientReverseLayer()
self.grl2 = GradientReverseLayer()
self.layer1 = nn.Linear(dim, dim)
nn.init.eye_(self.layer1.weight)
nn.init.zeros_(self.layer1.bias)
def forward(self, features: 'torch.Tensor') ->torch.Tensor:
features = self.grl1(features)
return self.grl2(self.layer1(features))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'dim': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch.autograd import Function
from typing import Optional
from typing import Tuple
import torch.nn as nn
from typing import Any
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused_mul_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](primals_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf1)
del primals_2
buf2 = reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf1
triton_poi_fused_mul_1[grid(256)](buf2, primals_3, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_3
return buf2, reinterpret_tensor(buf0, (64, 4), (4, 1), 0)
class GradientReverseFunction(Function):
@staticmethod
def forward(ctx: 'Any', input: 'torch.Tensor', coeff: 'Optional[float]'=1.0
) ->torch.Tensor:
ctx.coeff = coeff
output = input * 1.0
return output
@staticmethod
def backward(ctx: 'Any', grad_output: 'torch.Tensor') ->Tuple[torch.
Tensor, Any]:
return grad_output.neg() * ctx.coeff, None
class GradientReverseLayer(nn.Module):
def __init__(self):
super(GradientReverseLayer, self).__init__()
def forward(self, *input):
return GradientReverseFunction.apply(*input)
class ThetaNew(nn.Module):
"""
maximize loss respect to :math:` heta`
minimize loss respect to features
"""
def __init__(self, dim: 'int'):
super(ThetaNew, self).__init__()
self.grl1 = GradientReverseLayer()
self.grl2 = GradientReverseLayer()
self.layer1 = nn.Linear(dim, dim)
nn.init.eye_(self.layer1.weight)
nn.init.zeros_(self.layer1.bias)
def forward(self, input_0):
primals_2 = self.layer1.weight
primals_3 = self.layer1.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| NiteshBharadwaj/ignoringhumanpose | Theta | false | 912 | [
"MIT"
] | 0 | 1fb7a063fded9cff18f7de4e1d71845983077256 | https://github.com/NiteshBharadwaj/ignoringhumanpose/tree/1fb7a063fded9cff18f7de4e1d71845983077256 | from torch.autograd import Function
import torch
from typing import Optional
from typing import Tuple
import torch.nn as nn
from typing import Any
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class GradientReverseFunction(Function):
@staticmethod
def forward(ctx: 'Any', input: 'torch.Tensor', coeff: 'Optional[float]'=1.0
) ->torch.Tensor:
ctx.coeff = coeff
output = input * 1.0
return output
@staticmethod
def backward(ctx: 'Any', grad_output: 'torch.Tensor') ->Tuple[torch.
Tensor, Any]:
return grad_output.neg() * ctx.coeff, None
class GradientReverseLayer(nn.Module):
def __init__(self):
super().__init__()
def forward(self, *input):
return GradientReverseFunction.apply(*input)
class Model(nn.Module):
"""
maximize loss respect to :math:` heta`
minimize loss respect to features
"""
def __init__(self, dim: 'int'):
super().__init__()
self.grl1 = GradientReverseLayer()
self.grl2 = GradientReverseLayer()
self.layer1 = nn.Linear(dim, dim)
nn.init.eye_(self.layer1.weight)
nn.init.zeros_(self.layer1.bias)
def forward(self, features: 'torch.Tensor') ->torch.Tensor:
features = self.grl1(features)
return self.grl2(self.layer1(features))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
LastTimeStep | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/go/cgonf334lfnfwaep64nfomh4fnfi4b7mwswh3a6zm5yy2cczmhl3.py
# Topologically Sorted Source Nodes: [last_step_2], Original ATen: [aten.mean]
# Source node to ATen node mapping:
# last_step_2 => mean
# Graph fragment:
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%view, [2]), kwargs = {})
triton_poi_fused_mean_0 = async_compile.triton('triton_poi_fused_mean_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mean_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mean_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = 1.0
tmp2 = tmp0 / tmp1
tl.store(out_ptr0 + (x0), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [last_step_2], Original ATen: [aten.mean]
stream0 = get_raw_stream(0)
triton_poi_fused_mean_0.run(arg0_1, buf0, 64, grid=grid(64), stream=stream0)
del arg0_1
return (reinterpret_tensor(buf0, (4, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
import torch.utils.data
from typing import Tuple
class LastTimeStep(nn.Module):
"""
A class for extracting the hidden activations of the last time step following
the output of a PyTorch RNN module.
"""
def __init__(self, bidirectional=False):
super(LastTimeStep, self).__init__()
if bidirectional:
self.num_driections = 2
else:
self.num_driections = 1
def forward(self, x: 'Tuple'):
last_step = x[0]
batch_size = last_step.shape[1]
seq_len = last_step.shape[0]
last_step = last_step.view(seq_len, batch_size, self.num_driections, -1
)
last_step = torch.mean(last_step, 2)
last_step = last_step[0]
return last_step.reshape(batch_size, -1)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch import nn
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mean_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 1.0
tmp2 = tmp0 / tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mean_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del arg0_1
return reinterpret_tensor(buf0, (4, 4), (4, 1), 0),
class LastTimeStepNew(nn.Module):
"""
A class for extracting the hidden activations of the last time step following
the output of a PyTorch RNN module.
"""
def __init__(self, bidirectional=False):
super(LastTimeStepNew, self).__init__()
if bidirectional:
self.num_driections = 2
else:
self.num_driections = 1
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Onion-Team-VN/skilledlab | LastTimeStep | false | 913 | [
"Apache-2.0"
] | 0 | ac5cd7b5aee52da98aee8a32e5d161fd8b7dddab | https://github.com/Onion-Team-VN/skilledlab/tree/ac5cd7b5aee52da98aee8a32e5d161fd8b7dddab | import torch
from torch import nn
import torch.utils.data
from typing import Tuple
class Model(nn.Module):
"""
A class for extracting the hidden activations of the last time step following
the output of a PyTorch RNN module.
"""
def __init__(self, bidirectional=False):
super().__init__()
if bidirectional:
self.num_driections = 2
else:
self.num_driections = 1
def forward(self, x: 'Tuple'):
last_step = x[0]
batch_size = last_step.shape[1]
seq_len = last_step.shape[0]
last_step = last_step.view(seq_len, batch_size, self.num_driections, -1
)
last_step = torch.mean(last_step, 2)
last_step = last_step[0]
return last_step.reshape(batch_size, -1)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
myLoss2 | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/vv/cvvphp5tp5tokvtbdg3vl63hwc6ncuk6vnqgbneg6chzoeg2vkl4.py
# Topologically Sorted Source Nodes: [loss_1, loss_2, mul, add, loss], Original ATen: [aten.mse_loss, aten.mul, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# loss => div
# loss_1 => mean, pow_1, sub
# loss_2 => mean_1, pow_2, sub_1
# mul => mul
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg1_1, %arg0_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%pow_1,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg3_1, %arg2_1), kwargs = {})
# %pow_2 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub_1, 2), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%pow_2,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mean_1, 1.0), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean, %mul), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%add, 2.0), kwargs = {})
triton_per_fused_add_div_mse_loss_mul_0 = async_compile.triton('triton_per_fused_add_div_mse_loss_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {5: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 6), equal_to_1=(5,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_div_mse_loss_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 4, 'num_reduction': 2, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_div_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, in_ptr3, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = tl.load(in_ptr1 + (r0), None)
tmp7 = tl.load(in_ptr2 + (r0), None)
tmp8 = tl.load(in_ptr3 + (r0), None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp9 = tmp7 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tl.broadcast_to(tmp10, [RBLOCK])
tmp13 = triton_helpers.promote_to_tensor(tl.sum(tmp11, 0))
tmp14 = 256.0
tmp15 = tmp6 / tmp14
tmp16 = tmp13 / tmp14
tmp17 = 1.0
tmp18 = tmp16 * tmp17
tmp19 = tmp15 + tmp18
tmp20 = 0.5
tmp21 = tmp19 * tmp20
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp21, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg3_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf2 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [loss_1, loss_2, mul, add, loss], Original ATen: [aten.mse_loss, aten.mul, aten.add, aten.div]
stream0 = get_raw_stream(0)
triton_per_fused_add_div_mse_loss_mul_0.run(buf2, arg1_1, arg0_1, arg3_1, arg2_1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
del arg2_1
del arg3_1
return (buf2, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg2_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg3_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1, arg2_1, arg3_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
import torch.nn as nn
class myLoss2(nn.Module):
def __init__(self, alpha=1.0):
super(myLoss2, self).__init__()
self.alpha = alpha
def forward(self, sent_probs, doc_probs, sent_targets, doc_targets):
loss_1 = F.mse_loss(sent_probs, sent_targets)
loss_2 = F.mse_loss(doc_probs, doc_targets)
norm = 1.0 + self.alpha
loss = (loss_1 + self.alpha * loss_2) / norm
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_div_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, in_ptr3, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp7 = tl.load(in_ptr2 + r0, None)
tmp8 = tl.load(in_ptr3 + r0, None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp9 = tmp7 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tl.broadcast_to(tmp10, [RBLOCK])
tmp13 = triton_helpers.promote_to_tensor(tl.sum(tmp11, 0))
tmp14 = 256.0
tmp15 = tmp6 / tmp14
tmp16 = tmp13 / tmp14
tmp17 = 1.0
tmp18 = tmp16 * tmp17
tmp19 = tmp15 + tmp18
tmp20 = 0.5
tmp21 = tmp19 * tmp20
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp21, None)
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg3_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf2 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_div_mse_loss_mul_0[grid(1)](buf2, arg1_1,
arg0_1, arg3_1, arg2_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
del arg2_1
del arg3_1
return buf2,
class myLoss2New(nn.Module):
def __init__(self, alpha=1.0):
super(myLoss2New, self).__init__()
self.alpha = alpha
def forward(self, input_0, input_1, input_2, input_3):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
arg3_1 = input_3
output = call([arg0_1, arg1_1, arg2_1, arg3_1])
return output[0]
| PKULiuHui/LiveBlogSum | myLoss2 | false | 914 | [
"MIT"
] | 0 | b6a22521ee454e649981d70ddca6c89a1bac5a4c | https://github.com/PKULiuHui/LiveBlogSum/tree/b6a22521ee454e649981d70ddca6c89a1bac5a4c | import torch
import torch.nn.functional as F
import torch.nn as nn
class Model(nn.Module):
def __init__(self, alpha=1.0):
super().__init__()
self.alpha = alpha
def forward(self, sent_probs, doc_probs, sent_targets, doc_targets):
loss_1 = F.mse_loss(sent_probs, sent_targets)
loss_2 = F.mse_loss(doc_probs, doc_targets)
norm = 1.0 + self.alpha
loss = (loss_1 + self.alpha * loss_2) / norm
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
relu_constant_fraction | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/mv/cmvbbjeg7t4qnq4ujckrl5vqd7vxv7z4mkklfzyrzbxjlxbrdnqf.py
# Topologically Sorted Source Nodes: [sub, relu], Original ATen: [aten.sub, aten.relu]
# Source node to ATen node mapping:
# relu => relu
# sub => sub
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg1_1, %view), kwargs = {})
# %relu : [num_users=1] = call_function[target=torch.ops.aten.relu.default](args = (%sub,), kwargs = {})
triton_poi_fused_relu_sub_0 = async_compile.triton('triton_poi_fused_relu_sub_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_sub_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 16) % 4
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(out_ptr0 + (x3), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, ), (1, ))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sub, relu], Original ATen: [aten.sub, aten.relu]
stream0 = get_raw_stream(0)
triton_poi_fused_relu_sub_0.run(arg1_1, arg0_1, buf0, 256, grid=grid(256), stream=stream0)
del arg0_1
del arg1_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import numpy as np
from torch import nn
from torch.nn.functional import relu
def regula_falsi(func, a, b, iterations):
f_a = func(a, -1)
f_b = func(b, -1)
if torch.any(f_a * f_b >= 0):
None
raise Exception(
'You have not assumed right initial values in regula falsi')
c = a
break_indices = torch.zeros_like(a).bool()
for i in range(iterations):
c = (a * f_b - b * f_a) / (f_b - f_a)
f_c = func(c, i)
break_indices[f_c == 0] = True
b_eq_c_indices = (f_c * f_a < 0) & ~break_indices
b[b_eq_c_indices] = c[b_eq_c_indices]
a_eq_c_indices = ~(b_eq_c_indices | break_indices)
a[a_eq_c_indices] = c[a_eq_c_indices]
return c
class relu_constant_fraction(nn.Module):
def __init__(self, nb_channels):
super(relu_constant_fraction, self).__init__()
self.biases = nn.Parameter(torch.zeros(nb_channels))
self.biases.requires_grad = False
self.bias_buffer = None
def forward(self, x):
return relu(x - self.biases.view(1, -1, 1, 1))
def adjust_bias(self, desired_fraction, prev_layer_outputs):
if desired_fraction > 1 - 0.001:
self.biases.data = -10 * torch.ones_like(self.biases)
return
def get_fraction_deviation(biases, j):
activations = relu(prev_layer_outputs - biases.view(1, -1, 1, 1))
ratios = (activations > 0.001).float().mean(dim=(0, 2, 3))
return ratios - desired_fraction
with torch.no_grad():
solutions = regula_falsi(get_fraction_deviation, -3 * torch.
ones_like(self.biases), 3 * torch.ones_like(self.biases), 20)
momentum = 0.75
dampening = 0.0
lr = 0.5
delta = solutions - self.biases
buf = self.bias_buffer
if buf is None:
buf = torch.clone(delta).detach()
self.bias_buffer = buf
else:
buf.mul_(momentum).add_(delta, alpha=1 - dampening)
delta = buf
self.biases.add_(delta, alpha=lr)
def get_activation_fractions(self, prev_layer_outputs):
activations = relu(prev_layer_outputs - self.biases.view(1, -1, 1, 1))
ratios = (activations > 0.001).float().mean(dim=(0, 2, 3))
return ratios
def show_trajectory(self, prev_layer_outputs):
import matplotlib.pyplot as plt
bias_values = np.linspace(-10, 10, 1000)
fractions = np.zeros((1000, self.biases.shape[0]))
for j, bias in enumerate(bias_values):
cumulative_ratios = torch.zeros_like(self.biases)
batch_size = 1000
for i in range(0, len(prev_layer_outputs), batch_size):
data = prev_layer_outputs[i:i + batch_size]
activations = relu(data - bias)
cumulative_ratios += (activations > 0.001).float().mean(dim
=(0, 2, 3)) * len(data)
fractions[j] = (cumulative_ratios / len(prev_layer_outputs)
).detach().cpu().numpy()
plt.plot(bias_values, fractions)
plt.show()
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'nb_channels': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import numpy as np
from torch import nn
from torch.nn.functional import relu
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_relu_sub_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(out_ptr0 + x3, tmp4, xmask)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4,), (1,))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_relu_sub_0[grid(256)](arg1_1, arg0_1, buf0, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del arg0_1
del arg1_1
return buf0,
def regula_falsi(func, a, b, iterations):
f_a = func(a, -1)
f_b = func(b, -1)
if torch.any(f_a * f_b >= 0):
None
raise Exception(
'You have not assumed right initial values in regula falsi')
c = a
break_indices = torch.zeros_like(a).bool()
for i in range(iterations):
c = (a * f_b - b * f_a) / (f_b - f_a)
f_c = func(c, i)
break_indices[f_c == 0] = True
b_eq_c_indices = (f_c * f_a < 0) & ~break_indices
b[b_eq_c_indices] = c[b_eq_c_indices]
a_eq_c_indices = ~(b_eq_c_indices | break_indices)
a[a_eq_c_indices] = c[a_eq_c_indices]
return c
class relu_constant_fractionNew(nn.Module):
def __init__(self, nb_channels):
super(relu_constant_fractionNew, self).__init__()
self.biases = nn.Parameter(torch.zeros(nb_channels))
self.biases.requires_grad = False
self.bias_buffer = None
def adjust_bias(self, desired_fraction, prev_layer_outputs):
if desired_fraction > 1 - 0.001:
self.biases.data = -10 * torch.ones_like(self.biases)
return
def get_fraction_deviation(biases, j):
activations = relu(prev_layer_outputs - biases.view(1, -1, 1, 1))
ratios = (activations > 0.001).float().mean(dim=(0, 2, 3))
return ratios - desired_fraction
with torch.no_grad():
solutions = regula_falsi(get_fraction_deviation, -3 * torch.
ones_like(self.biases), 3 * torch.ones_like(self.biases), 20)
momentum = 0.75
dampening = 0.0
lr = 0.5
delta = solutions - self.biases
buf = self.bias_buffer
if buf is None:
buf = torch.clone(delta).detach()
self.bias_buffer = buf
else:
buf.mul_(momentum).add_(delta, alpha=1 - dampening)
delta = buf
self.biases.add_(delta, alpha=lr)
def get_activation_fractions(self, prev_layer_outputs):
activations = relu(prev_layer_outputs - self.biases.view(1, -1, 1, 1))
ratios = (activations > 0.001).float().mean(dim=(0, 2, 3))
return ratios
def show_trajectory(self, prev_layer_outputs):
import matplotlib.pyplot as plt
bias_values = np.linspace(-10, 10, 1000)
fractions = np.zeros((1000, self.biases.shape[0]))
for j, bias in enumerate(bias_values):
cumulative_ratios = torch.zeros_like(self.biases)
batch_size = 1000
for i in range(0, len(prev_layer_outputs), batch_size):
data = prev_layer_outputs[i:i + batch_size]
activations = relu(data - bias)
cumulative_ratios += (activations > 0.001).float().mean(dim
=(0, 2, 3)) * len(data)
fractions[j] = (cumulative_ratios / len(prev_layer_outputs)
).detach().cpu().numpy()
plt.plot(bias_values, fractions)
plt.show()
def forward(self, input_0):
arg0_1 = self.biases
arg1_1 = input_0
output = call([arg0_1, arg1_1])
return output[0]
| Noppornying00/constant-fraction-activation | relu_constant_fraction | false | 915 | [
"Apache-2.0"
] | 0 | b25745e7339df13e3db34d8c8372d5cbaa3c13bb | https://github.com/Noppornying00/constant-fraction-activation/tree/b25745e7339df13e3db34d8c8372d5cbaa3c13bb | import torch
import numpy as np
from torch import nn
from torch.nn.functional import relu
def regula_falsi(func, a, b, iterations):
f_a = func(a, -1)
f_b = func(b, -1)
if torch.any(f_a * f_b >= 0):
None
raise Exception(
'You have not assumed right initial values in regula falsi')
c = a
break_indices = torch.zeros_like(a).bool()
for i in range(iterations):
c = (a * f_b - b * f_a) / (f_b - f_a)
f_c = func(c, i)
break_indices[f_c == 0] = True
b_eq_c_indices = (f_c * f_a < 0) & ~break_indices
b[b_eq_c_indices] = c[b_eq_c_indices]
a_eq_c_indices = ~(b_eq_c_indices | break_indices)
a[a_eq_c_indices] = c[a_eq_c_indices]
return c
class Model(nn.Module):
def __init__(self, nb_channels):
super().__init__()
self.biases = nn.Parameter(torch.zeros(nb_channels))
self.biases.requires_grad = False
self.bias_buffer = None
def forward(self, x):
return relu(x - self.biases.view(1, -1, 1, 1))
def adjust_bias(self, desired_fraction, prev_layer_outputs):
if desired_fraction > 1 - 0.001:
self.biases.data = -10 * torch.ones_like(self.biases)
return
def get_fraction_deviation(biases, j):
activations = relu(prev_layer_outputs - biases.view(1, -1, 1, 1))
ratios = (activations > 0.001).float().mean(dim=(0, 2, 3))
return ratios - desired_fraction
with torch.no_grad():
solutions = regula_falsi(get_fraction_deviation, -3 * torch.
ones_like(self.biases), 3 * torch.ones_like(self.biases), 20)
momentum = 0.75
dampening = 0.0
lr = 0.5
delta = solutions - self.biases
buf = self.bias_buffer
if buf is None:
buf = torch.clone(delta).detach()
self.bias_buffer = buf
else:
buf.mul_(momentum).add_(delta, alpha=1 - dampening)
delta = buf
self.biases.add_(delta, alpha=lr)
def get_activation_fractions(self, prev_layer_outputs):
activations = relu(prev_layer_outputs - self.biases.view(1, -1, 1, 1))
ratios = (activations > 0.001).float().mean(dim=(0, 2, 3))
return ratios
def show_trajectory(self, prev_layer_outputs):
import matplotlib.pyplot as plt
bias_values = np.linspace(-10, 10, 1000)
fractions = np.zeros((1000, self.biases.shape[0]))
for j, bias in enumerate(bias_values):
cumulative_ratios = torch.zeros_like(self.biases)
batch_size = 1000
for i in range(0, len(prev_layer_outputs), batch_size):
data = prev_layer_outputs[i:i + batch_size]
activations = relu(data - bias)
cumulative_ratios += (activations > 0.001).float().mean(dim
=(0, 2, 3)) * len(data)
fractions[j] = (cumulative_ratios / len(prev_layer_outputs)
).detach().cpu().numpy()
plt.plot(bias_values, fractions)
plt.show()
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
MarginDisparityDiscrepancy | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/td/ctdj5kazgiki6gdaadhqtp2x7tq2ee5ey5hqqdcoqmp54jyhf74f.py
# Topologically Sorted Source Nodes: [cross_entropy], Original ATen: [aten._log_softmax]
# Source node to ATen node mapping:
# cross_entropy => amax, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%arg2_1, [1], True), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg2_1, %amax), kwargs = {})
triton_poi_fused__log_softmax_0 = async_compile.triton('triton_poi_fused__log_softmax_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__log_softmax_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__log_softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tl.store(out_ptr0 + (x3), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/xk/cxkugsynlmnyrjhah42fewrhwovuvurnuv2qimo2qhxq27wjmq7q.py
# Topologically Sorted Source Nodes: [softmax], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# softmax => amax_1, exp_1, sub_2
# Graph fragment:
# %amax_1 : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%arg3_1, [1], True), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg3_1, %amax_1), kwargs = {})
# %exp_1 : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub_2,), kwargs = {})
triton_poi_fused__softmax_1 = async_compile.triton('triton_poi_fused__softmax_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x3), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/72/c723jrqk3swxzkkbncmap355zvmdp4omozimuaafxsjoczdy4tqn.py
# Topologically Sorted Source Nodes: [max_1, max_2, cross_entropy, mul, nll_loss, add_1], Original ATen: [aten.max, aten.nll_loss2d_forward, aten.mul, aten.add]
# Source node to ATen node mapping:
# add_1 => add_1
# cross_entropy => convert_element_type, div, full_default_1, ne_1, ne_2, neg, sum_2, sum_3, where_1
# max_1 => max_1
# max_2 => max_2
# mul => mul
# nll_loss => convert_element_type_1, div_2, full_default_3, ne_4, ne_5, neg_1, sum_5, sum_6, where_3
# Graph fragment:
# %max_1 : [num_users=1] = call_function[target=torch.ops.aten.max.dim](args = (%arg0_1, 1), kwargs = {})
# %max_2 : [num_users=1] = call_function[target=torch.ops.aten.max.dim](args = (%arg1_1, 1), kwargs = {})
# %ne_1 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%getitem_1, -100), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%squeeze,), kwargs = {})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_1 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%ne_1, %neg, %full_default_1), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%where_1,), kwargs = {})
# %ne_2 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%getitem_1, -100), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%ne_2,), kwargs = {})
# %convert_element_type : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%sum_2, torch.float32), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_3, %convert_element_type), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div, 4), kwargs = {})
# %ne_4 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%getitem_3, -100), kwargs = {})
# %neg_1 : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%squeeze_1,), kwargs = {})
# %full_default_3 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_3 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%ne_4, %neg_1, %full_default_3), kwargs = {})
# %sum_6 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%where_3,), kwargs = {})
# %ne_5 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%getitem_3, -100), kwargs = {})
# %sum_5 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%ne_5,), kwargs = {})
# %convert_element_type_1 : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%sum_5, torch.float32), kwargs = {})
# %div_2 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%sum_6, %convert_element_type_1), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %div_2), kwargs = {})
triton_per_fused_add_max_mul_nll_loss2d_forward_2 = async_compile.triton('triton_per_fused_add_max_mul_nll_loss2d_forward_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 64],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {5: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 6), equal_to_1=(5,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_max_mul_nll_loss2d_forward_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 16, 'num_reduction': 4, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_max_mul_nll_loss2d_forward_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, in_ptr3, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 64
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = (rindex // 16)
r2 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + (64*r1)), None)
tmp1 = tl.load(in_ptr0 + (16 + r0 + (64*r1)), None)
tmp17 = tl.load(in_ptr0 + (32 + r0 + (64*r1)), None)
tmp32 = tl.load(in_ptr0 + (48 + r0 + (64*r1)), None)
tmp56 = tl.load(in_ptr1 + (r0 + (64*r1)), None)
tmp58 = tl.load(in_ptr1 + (16 + r0 + (64*r1)), None)
tmp61 = tl.load(in_ptr1 + (32 + r0 + (64*r1)), None)
tmp64 = tl.load(in_ptr1 + (48 + r0 + (64*r1)), None)
tmp79 = tl.load(in_ptr2 + (r0 + (64*r1)), None)
tmp80 = tl.load(in_ptr2 + (16 + r0 + (64*r1)), None)
tmp93 = tl.load(in_ptr2 + (32 + r0 + (64*r1)), None)
tmp107 = tl.load(in_ptr2 + (48 + r0 + (64*r1)), None)
tmp128 = tl.load(in_ptr3 + (r0 + (64*r1)), None)
tmp129 = tl.load(in_ptr3 + (16 + r0 + (64*r1)), None)
tmp131 = tl.load(in_ptr3 + (32 + r0 + (64*r1)), None)
tmp133 = tl.load(in_ptr3 + (48 + r0 + (64*r1)), None)
tmp2 = tmp0 > tmp1
tmp3 = tmp0 == tmp1
tmp4 = tmp0 != tmp0
tmp5 = tmp1 != tmp1
tmp6 = tmp4 > tmp5
tmp7 = tmp2 | tmp6
tmp8 = tmp4 & tmp5
tmp9 = tmp3 | tmp8
tmp10 = tl.full([1, 1], 0, tl.int64)
tmp11 = tl.full([1, 1], 1, tl.int64)
tmp12 = tmp10 < tmp11
tmp13 = tmp9 & tmp12
tmp14 = tmp7 | tmp13
tmp15 = tl.where(tmp14, tmp0, tmp1)
tmp16 = tl.where(tmp14, tmp10, tmp11)
tmp18 = tmp15 > tmp17
tmp19 = tmp15 == tmp17
tmp20 = tmp15 != tmp15
tmp21 = tmp17 != tmp17
tmp22 = tmp20 > tmp21
tmp23 = tmp18 | tmp22
tmp24 = tmp20 & tmp21
tmp25 = tmp19 | tmp24
tmp26 = tl.full([1, 1], 2, tl.int64)
tmp27 = tmp16 < tmp26
tmp28 = tmp25 & tmp27
tmp29 = tmp23 | tmp28
tmp30 = tl.where(tmp29, tmp15, tmp17)
tmp31 = tl.where(tmp29, tmp16, tmp26)
tmp33 = tmp30 > tmp32
tmp34 = tmp30 == tmp32
tmp35 = tmp30 != tmp30
tmp36 = tmp32 != tmp32
tmp37 = tmp35 > tmp36
tmp38 = tmp33 | tmp37
tmp39 = tmp35 & tmp36
tmp40 = tmp34 | tmp39
tmp41 = tl.full([1, 1], 3, tl.int64)
tmp42 = tmp31 < tmp41
tmp43 = tmp40 & tmp42
tmp44 = tmp38 | tmp43
tmp45 = tl.where(tmp44, tmp30, tmp32)
tmp46 = tl.where(tmp44, tmp31, tmp41)
tmp47 = tl.full([1, 1], -100, tl.int64)
tmp48 = tmp46 != tmp47
tmp49 = tl.where(tmp48, tmp46, tmp10)
tmp50 = tl.full([XBLOCK, RBLOCK], 4, tl.int32)
tmp51 = tmp49 + tmp50
tmp52 = tmp49 < 0
tmp53 = tl.where(tmp52, tmp51, tmp49)
tl.device_assert((0 <= tmp53) & (tmp53 < 4), "index out of bounds: 0 <= tmp53 < 4")
tmp55 = tl.load(in_ptr1 + (r0 + (16*tmp53) + (64*r1)), None)
tmp57 = tl_math.exp(tmp56)
tmp59 = tl_math.exp(tmp58)
tmp60 = tmp57 + tmp59
tmp62 = tl_math.exp(tmp61)
tmp63 = tmp60 + tmp62
tmp65 = tl_math.exp(tmp64)
tmp66 = tmp63 + tmp65
tmp67 = tl_math.log(tmp66)
tmp68 = tmp55 - tmp67
tmp69 = -tmp68
tmp70 = 0.0
tmp71 = tl.where(tmp48, tmp69, tmp70)
tmp72 = tl.broadcast_to(tmp71, [XBLOCK, RBLOCK])
tmp74 = tl.sum(tmp72, 1)[:, None]
tmp75 = tmp48.to(tl.int64)
tmp76 = tl.broadcast_to(tmp75, [XBLOCK, RBLOCK])
tmp78 = tl.sum(tmp76, 1)[:, None]
tmp81 = tmp79 > tmp80
tmp82 = tmp79 == tmp80
tmp83 = tmp79 != tmp79
tmp84 = tmp80 != tmp80
tmp85 = tmp83 > tmp84
tmp86 = tmp81 | tmp85
tmp87 = tmp83 & tmp84
tmp88 = tmp82 | tmp87
tmp89 = tmp88 & tmp12
tmp90 = tmp86 | tmp89
tmp91 = tl.where(tmp90, tmp79, tmp80)
tmp92 = tl.where(tmp90, tmp10, tmp11)
tmp94 = tmp91 > tmp93
tmp95 = tmp91 == tmp93
tmp96 = tmp91 != tmp91
tmp97 = tmp93 != tmp93
tmp98 = tmp96 > tmp97
tmp99 = tmp94 | tmp98
tmp100 = tmp96 & tmp97
tmp101 = tmp95 | tmp100
tmp102 = tmp92 < tmp26
tmp103 = tmp101 & tmp102
tmp104 = tmp99 | tmp103
tmp105 = tl.where(tmp104, tmp91, tmp93)
tmp106 = tl.where(tmp104, tmp92, tmp26)
tmp108 = tmp105 > tmp107
tmp109 = tmp105 == tmp107
tmp110 = tmp105 != tmp105
tmp111 = tmp107 != tmp107
tmp112 = tmp110 > tmp111
tmp113 = tmp108 | tmp112
tmp114 = tmp110 & tmp111
tmp115 = tmp109 | tmp114
tmp116 = tmp106 < tmp41
tmp117 = tmp115 & tmp116
tmp118 = tmp113 | tmp117
tmp119 = tl.where(tmp118, tmp105, tmp107)
tmp120 = tl.where(tmp118, tmp106, tmp41)
tmp121 = tmp120 != tmp47
tmp122 = tl.where(tmp121, tmp120, tmp10)
tmp123 = tmp122 + tmp50
tmp124 = tmp122 < 0
tmp125 = tl.where(tmp124, tmp123, tmp122)
tl.device_assert((0 <= tmp125) & (tmp125 < 4), "index out of bounds: 0 <= tmp125 < 4")
tmp127 = tl.load(in_ptr3 + (r0 + (16*tmp125) + (64*r1)), None)
tmp130 = tmp128 + tmp129
tmp132 = tmp130 + tmp131
tmp134 = tmp132 + tmp133
tmp135 = tmp127 / tmp134
tmp136 = 1.0
tmp137 = tmp136 - tmp135
tmp138 = 1e-06
tmp139 = tmp137 + tmp138
tmp140 = triton_helpers.minimum(tmp139, tmp136)
tmp141 = tl_math.log(tmp140)
tmp142 = -tmp141
tmp143 = tl.where(tmp121, tmp142, tmp70)
tmp144 = tl.broadcast_to(tmp143, [XBLOCK, RBLOCK])
tmp146 = tl.sum(tmp144, 1)[:, None]
tmp147 = tmp121.to(tl.int64)
tmp148 = tl.broadcast_to(tmp147, [XBLOCK, RBLOCK])
tmp150 = tl.sum(tmp148, 1)[:, None]
tmp151 = tmp78.to(tl.float32)
tmp152 = tmp74 / tmp151
tmp153 = 4.0
tmp154 = tmp152 * tmp153
tmp155 = tmp150.to(tl.float32)
tmp156 = tmp146 / tmp155
tmp157 = tmp154 + tmp156
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp157, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg3_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [cross_entropy], Original ATen: [aten._log_softmax]
stream0 = get_raw_stream(0)
triton_poi_fused__log_softmax_0.run(arg2_1, buf2, 256, grid=grid(256), stream=stream0)
del arg2_1
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [softmax], Original ATen: [aten._softmax]
triton_poi_fused__softmax_1.run(arg3_1, buf5, 256, grid=grid(256), stream=stream0)
del arg3_1
buf3 = empty_strided_cuda((), (), torch.float32)
buf8 = buf3; del buf3 # reuse
# Topologically Sorted Source Nodes: [max_1, max_2, cross_entropy, mul, nll_loss, add_1], Original ATen: [aten.max, aten.nll_loss2d_forward, aten.mul, aten.add]
triton_per_fused_add_max_mul_nll_loss2d_forward_2.run(buf8, arg0_1, buf2, arg1_1, buf5, 1, 64, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
del buf2
del buf5
return (buf8, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg2_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg3_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1, arg2_1, arg3_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from typing import Optional
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
def shift_log(x: 'torch.Tensor', offset: 'Optional[float]'=1e-06
) ->torch.Tensor:
"""
First shift, then calculate log, which can be described as:
.. math::
y = \\max(\\log(x+\\text{offset}), 0)
Used to avoid the gradient explosion problem in log(x) function when x=0.
Parameters:
- **x**: input tensor
- **offset**: offset size. Default: 1e-6
.. note::
Input tensor falls in [0., 1.] and the output tensor falls in [-log(offset), 0]
"""
return torch.log(torch.clamp(x + offset, max=1.0))
class MarginDisparityDiscrepancy(nn.Module):
"""The margin disparity discrepancy (MDD) is proposed to measure the distribution discrepancy in domain adaptation.
The :math:`y^s` and :math:`y^t` are logits output by the main classifier on the source and target domain respectively.
The :math:`y_{adv}^s` and :math:`y_{adv}^t` are logits output by the adversarial classifier.
They are expected to contain raw, unnormalized scores for each class.
The definition can be described as:
.. math::
\\mathcal{D}_{\\gamma}(\\hat{\\mathcal{S}}, \\hat{\\mathcal{T}}) =
\\gamma \\mathbb{E}_{y^s, y_{adv}^s \\sim\\hat{\\mathcal{S}}} \\log\\left(\\frac{\\exp(y_{adv}^s[h_{y^s}])}{\\sum_j \\exp(y_{adv}^s[j])}\\right) +
\\mathbb{E}_{y^t, y_{adv}^t \\sim\\hat{\\mathcal{T}}} \\log\\left(1-\\frac{\\exp(y_{adv}^t[h_{y^t}])}{\\sum_j \\exp(y_{adv}^t[j])}\\right),
where :math:`\\gamma` is a margin hyper-parameter and :math:`h_y` refers to the predicted label when the logits output is :math:`y`.
You can see more details in `Bridging Theory and Algorithm for Domain Adaptation <https://arxiv.org/abs/1904.05801>`_.
Parameters:
- **margin** (float): margin :math:`\\gamma`. Default: 4
- **reduction** (string, optional): Specifies the reduction to apply to the output:
``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
``'mean'``: the sum of the output will be divided by the number of
elements in the output, ``'sum'``: the output will be summed. Default: ``'mean'``
Inputs: y_s, y_s_adv, y_t, y_t_adv
- **y_s**: logits output :math:`y^s` by the main classifier on the source domain
- **y_s_adv**: logits output :math:`y^s` by the adversarial classifier on the source domain
- **y_t**: logits output :math:`y^t` by the main classifier on the target domain
- **y_t_adv**: logits output :math:`y_{adv}^t` by the adversarial classifier on the target domain
Shape:
- Inputs: :math:`(minibatch, C)` where C = number of classes, or :math:`(minibatch, C, d_1, d_2, ..., d_K)`
with :math:`K \\geq 1` in the case of `K`-dimensional loss.
- Output: scalar. If :attr:`reduction` is ``'none'``, then the same size as the target: :math:`(minibatch)`, or
:math:`(minibatch, d_1, d_2, ..., d_K)` with :math:`K \\geq 1` in the case of K-dimensional loss.
Examples::
>>> num_classes = 2
>>> batch_size = 10
>>> loss = MarginDisparityDiscrepancy(margin=4.)
>>> # logits output from source domain and target domain
>>> y_s, y_t = torch.randn(batch_size, num_classes), torch.randn(batch_size, num_classes)
>>> # adversarial logits output from source domain and target domain
>>> y_s_adv, y_t_adv = torch.randn(batch_size, num_classes), torch.randn(batch_size, num_classes)
>>> output = loss(y_s, y_s_adv, y_t, y_t_adv)
"""
def __init__(self, margin: 'Optional[int]'=4, reduction:
'Optional[str]'='mean'):
super(MarginDisparityDiscrepancy, self).__init__()
self.margin = margin
self.reduction = reduction
def forward(self, y_s: 'torch.Tensor', y_s_adv: 'torch.Tensor', y_t:
'torch.Tensor', y_t_adv: 'torch.Tensor') ->torch.Tensor:
_, prediction_s = y_s.max(dim=1)
_, prediction_t = y_t.max(dim=1)
return self.margin * F.cross_entropy(y_s_adv, prediction_s,
reduction=self.reduction) + F.nll_loss(shift_log(1.0 - F.
softmax(y_t_adv, dim=1)), prediction_t, reduction=self.reduction)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from typing import Optional
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__log_softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tl.store(out_ptr0 + x3, tmp8, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_per_fused_add_max_mul_nll_loss2d_forward_2(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, in_ptr3, xnumel, rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = rindex // 16
tmp0 = tl.load(in_ptr0 + (r0 + 64 * r1), None)
tmp1 = tl.load(in_ptr0 + (16 + r0 + 64 * r1), None)
tmp17 = tl.load(in_ptr0 + (32 + r0 + 64 * r1), None)
tmp32 = tl.load(in_ptr0 + (48 + r0 + 64 * r1), None)
tmp56 = tl.load(in_ptr1 + (r0 + 64 * r1), None)
tmp58 = tl.load(in_ptr1 + (16 + r0 + 64 * r1), None)
tmp61 = tl.load(in_ptr1 + (32 + r0 + 64 * r1), None)
tmp64 = tl.load(in_ptr1 + (48 + r0 + 64 * r1), None)
tmp79 = tl.load(in_ptr2 + (r0 + 64 * r1), None)
tmp80 = tl.load(in_ptr2 + (16 + r0 + 64 * r1), None)
tmp93 = tl.load(in_ptr2 + (32 + r0 + 64 * r1), None)
tmp107 = tl.load(in_ptr2 + (48 + r0 + 64 * r1), None)
tmp128 = tl.load(in_ptr3 + (r0 + 64 * r1), None)
tmp129 = tl.load(in_ptr3 + (16 + r0 + 64 * r1), None)
tmp131 = tl.load(in_ptr3 + (32 + r0 + 64 * r1), None)
tmp133 = tl.load(in_ptr3 + (48 + r0 + 64 * r1), None)
tmp2 = tmp0 > tmp1
tmp3 = tmp0 == tmp1
tmp4 = tmp0 != tmp0
tmp5 = tmp1 != tmp1
tmp6 = tmp4 > tmp5
tmp7 = tmp2 | tmp6
tmp8 = tmp4 & tmp5
tmp9 = tmp3 | tmp8
tmp10 = tl.full([1, 1], 0, tl.int64)
tmp11 = tl.full([1, 1], 1, tl.int64)
tmp12 = tmp10 < tmp11
tmp13 = tmp9 & tmp12
tmp14 = tmp7 | tmp13
tmp15 = tl.where(tmp14, tmp0, tmp1)
tmp16 = tl.where(tmp14, tmp10, tmp11)
tmp18 = tmp15 > tmp17
tmp19 = tmp15 == tmp17
tmp20 = tmp15 != tmp15
tmp21 = tmp17 != tmp17
tmp22 = tmp20 > tmp21
tmp23 = tmp18 | tmp22
tmp24 = tmp20 & tmp21
tmp25 = tmp19 | tmp24
tmp26 = tl.full([1, 1], 2, tl.int64)
tmp27 = tmp16 < tmp26
tmp28 = tmp25 & tmp27
tmp29 = tmp23 | tmp28
tmp30 = tl.where(tmp29, tmp15, tmp17)
tmp31 = tl.where(tmp29, tmp16, tmp26)
tmp33 = tmp30 > tmp32
tmp34 = tmp30 == tmp32
tmp35 = tmp30 != tmp30
tmp36 = tmp32 != tmp32
tmp37 = tmp35 > tmp36
tmp38 = tmp33 | tmp37
tmp39 = tmp35 & tmp36
tmp40 = tmp34 | tmp39
tmp41 = tl.full([1, 1], 3, tl.int64)
tmp42 = tmp31 < tmp41
tmp43 = tmp40 & tmp42
tmp44 = tmp38 | tmp43
tl.where(tmp44, tmp30, tmp32)
tmp46 = tl.where(tmp44, tmp31, tmp41)
tmp47 = tl.full([1, 1], -100, tl.int64)
tmp48 = tmp46 != tmp47
tmp49 = tl.where(tmp48, tmp46, tmp10)
tmp50 = tl.full([XBLOCK, RBLOCK], 4, tl.int32)
tmp51 = tmp49 + tmp50
tmp52 = tmp49 < 0
tmp53 = tl.where(tmp52, tmp51, tmp49)
tl.device_assert((0 <= tmp53) & (tmp53 < 4),
'index out of bounds: 0 <= tmp53 < 4')
tmp55 = tl.load(in_ptr1 + (r0 + 16 * tmp53 + 64 * r1), None)
tmp57 = tl_math.exp(tmp56)
tmp59 = tl_math.exp(tmp58)
tmp60 = tmp57 + tmp59
tmp62 = tl_math.exp(tmp61)
tmp63 = tmp60 + tmp62
tmp65 = tl_math.exp(tmp64)
tmp66 = tmp63 + tmp65
tmp67 = tl_math.log(tmp66)
tmp68 = tmp55 - tmp67
tmp69 = -tmp68
tmp70 = 0.0
tmp71 = tl.where(tmp48, tmp69, tmp70)
tmp72 = tl.broadcast_to(tmp71, [XBLOCK, RBLOCK])
tmp74 = tl.sum(tmp72, 1)[:, None]
tmp75 = tmp48.to(tl.int64)
tmp76 = tl.broadcast_to(tmp75, [XBLOCK, RBLOCK])
tmp78 = tl.sum(tmp76, 1)[:, None]
tmp81 = tmp79 > tmp80
tmp82 = tmp79 == tmp80
tmp83 = tmp79 != tmp79
tmp84 = tmp80 != tmp80
tmp85 = tmp83 > tmp84
tmp86 = tmp81 | tmp85
tmp87 = tmp83 & tmp84
tmp88 = tmp82 | tmp87
tmp89 = tmp88 & tmp12
tmp90 = tmp86 | tmp89
tmp91 = tl.where(tmp90, tmp79, tmp80)
tmp92 = tl.where(tmp90, tmp10, tmp11)
tmp94 = tmp91 > tmp93
tmp95 = tmp91 == tmp93
tmp96 = tmp91 != tmp91
tmp97 = tmp93 != tmp93
tmp98 = tmp96 > tmp97
tmp99 = tmp94 | tmp98
tmp100 = tmp96 & tmp97
tmp101 = tmp95 | tmp100
tmp102 = tmp92 < tmp26
tmp103 = tmp101 & tmp102
tmp104 = tmp99 | tmp103
tmp105 = tl.where(tmp104, tmp91, tmp93)
tmp106 = tl.where(tmp104, tmp92, tmp26)
tmp108 = tmp105 > tmp107
tmp109 = tmp105 == tmp107
tmp110 = tmp105 != tmp105
tmp111 = tmp107 != tmp107
tmp112 = tmp110 > tmp111
tmp113 = tmp108 | tmp112
tmp114 = tmp110 & tmp111
tmp115 = tmp109 | tmp114
tmp116 = tmp106 < tmp41
tmp117 = tmp115 & tmp116
tmp118 = tmp113 | tmp117
tl.where(tmp118, tmp105, tmp107)
tmp120 = tl.where(tmp118, tmp106, tmp41)
tmp121 = tmp120 != tmp47
tmp122 = tl.where(tmp121, tmp120, tmp10)
tmp123 = tmp122 + tmp50
tmp124 = tmp122 < 0
tmp125 = tl.where(tmp124, tmp123, tmp122)
tl.device_assert((0 <= tmp125) & (tmp125 < 4),
'index out of bounds: 0 <= tmp125 < 4')
tmp127 = tl.load(in_ptr3 + (r0 + 16 * tmp125 + 64 * r1), None)
tmp130 = tmp128 + tmp129
tmp132 = tmp130 + tmp131
tmp134 = tmp132 + tmp133
tmp135 = tmp127 / tmp134
tmp136 = 1.0
tmp137 = tmp136 - tmp135
tmp138 = 1e-06
tmp139 = tmp137 + tmp138
tmp140 = triton_helpers.minimum(tmp139, tmp136)
tmp141 = tl_math.log(tmp140)
tmp142 = -tmp141
tmp143 = tl.where(tmp121, tmp142, tmp70)
tmp144 = tl.broadcast_to(tmp143, [XBLOCK, RBLOCK])
tmp146 = tl.sum(tmp144, 1)[:, None]
tmp147 = tmp121.to(tl.int64)
tmp148 = tl.broadcast_to(tmp147, [XBLOCK, RBLOCK])
tmp150 = tl.sum(tmp148, 1)[:, None]
tmp151 = tmp78.to(tl.float32)
tmp152 = tmp74 / tmp151
tmp153 = 4.0
tmp154 = tmp152 * tmp153
tmp155 = tmp150.to(tl.float32)
tmp156 = tmp146 / tmp155
tmp157 = tmp154 + tmp156
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp157, None)
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg3_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__log_softmax_0[grid(256)](arg2_1, buf2, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del arg2_1
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(256)](arg3_1, buf5, 256, XBLOCK=
256, num_warps=4, num_stages=1)
del arg3_1
buf3 = empty_strided_cuda((), (), torch.float32)
buf8 = buf3
del buf3
triton_per_fused_add_max_mul_nll_loss2d_forward_2[grid(1)](buf8,
arg0_1, buf2, arg1_1, buf5, 1, 64, XBLOCK=1, num_warps=2,
num_stages=1)
del arg0_1
del arg1_1
del buf2
del buf5
return buf8,
def shift_log(x: 'torch.Tensor', offset: 'Optional[float]'=1e-06
) ->torch.Tensor:
"""
First shift, then calculate log, which can be described as:
.. math::
y = \\max(\\log(x+\\text{offset}), 0)
Used to avoid the gradient explosion problem in log(x) function when x=0.
Parameters:
- **x**: input tensor
- **offset**: offset size. Default: 1e-6
.. note::
Input tensor falls in [0., 1.] and the output tensor falls in [-log(offset), 0]
"""
return torch.log(torch.clamp(x + offset, max=1.0))
class MarginDisparityDiscrepancyNew(nn.Module):
"""The margin disparity discrepancy (MDD) is proposed to measure the distribution discrepancy in domain adaptation.
The :math:`y^s` and :math:`y^t` are logits output by the main classifier on the source and target domain respectively.
The :math:`y_{adv}^s` and :math:`y_{adv}^t` are logits output by the adversarial classifier.
They are expected to contain raw, unnormalized scores for each class.
The definition can be described as:
.. math::
\\mathcal{D}_{\\gamma}(\\hat{\\mathcal{S}}, \\hat{\\mathcal{T}}) =
\\gamma \\mathbb{E}_{y^s, y_{adv}^s \\sim\\hat{\\mathcal{S}}} \\log\\left(\\frac{\\exp(y_{adv}^s[h_{y^s}])}{\\sum_j \\exp(y_{adv}^s[j])}\\right) +
\\mathbb{E}_{y^t, y_{adv}^t \\sim\\hat{\\mathcal{T}}} \\log\\left(1-\\frac{\\exp(y_{adv}^t[h_{y^t}])}{\\sum_j \\exp(y_{adv}^t[j])}\\right),
where :math:`\\gamma` is a margin hyper-parameter and :math:`h_y` refers to the predicted label when the logits output is :math:`y`.
You can see more details in `Bridging Theory and Algorithm for Domain Adaptation <https://arxiv.org/abs/1904.05801>`_.
Parameters:
- **margin** (float): margin :math:`\\gamma`. Default: 4
- **reduction** (string, optional): Specifies the reduction to apply to the output:
``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
``'mean'``: the sum of the output will be divided by the number of
elements in the output, ``'sum'``: the output will be summed. Default: ``'mean'``
Inputs: y_s, y_s_adv, y_t, y_t_adv
- **y_s**: logits output :math:`y^s` by the main classifier on the source domain
- **y_s_adv**: logits output :math:`y^s` by the adversarial classifier on the source domain
- **y_t**: logits output :math:`y^t` by the main classifier on the target domain
- **y_t_adv**: logits output :math:`y_{adv}^t` by the adversarial classifier on the target domain
Shape:
- Inputs: :math:`(minibatch, C)` where C = number of classes, or :math:`(minibatch, C, d_1, d_2, ..., d_K)`
with :math:`K \\geq 1` in the case of `K`-dimensional loss.
- Output: scalar. If :attr:`reduction` is ``'none'``, then the same size as the target: :math:`(minibatch)`, or
:math:`(minibatch, d_1, d_2, ..., d_K)` with :math:`K \\geq 1` in the case of K-dimensional loss.
Examples::
>>> num_classes = 2
>>> batch_size = 10
>>> loss = MarginDisparityDiscrepancy(margin=4.)
>>> # logits output from source domain and target domain
>>> y_s, y_t = torch.randn(batch_size, num_classes), torch.randn(batch_size, num_classes)
>>> # adversarial logits output from source domain and target domain
>>> y_s_adv, y_t_adv = torch.randn(batch_size, num_classes), torch.randn(batch_size, num_classes)
>>> output = loss(y_s, y_s_adv, y_t, y_t_adv)
"""
def __init__(self, margin: 'Optional[int]'=4, reduction:
'Optional[str]'='mean'):
super(MarginDisparityDiscrepancyNew, self).__init__()
self.margin = margin
self.reduction = reduction
def forward(self, input_0, input_1, input_2, input_3):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
arg3_1 = input_3
output = call([arg0_1, arg1_1, arg2_1, arg3_1])
return output[0]
| NiteshBharadwaj/ignoringhumanpose | MarginDisparityDiscrepancy | false | 916 | [
"MIT"
] | 0 | 1fb7a063fded9cff18f7de4e1d71845983077256 | https://github.com/NiteshBharadwaj/ignoringhumanpose/tree/1fb7a063fded9cff18f7de4e1d71845983077256 | import torch
from typing import Optional
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
def shift_log(x: 'torch.Tensor', offset: 'Optional[float]'=1e-06
) ->torch.Tensor:
"""
First shift, then calculate log, which can be described as:
.. math::
y = \\max(\\log(x+\\text{offset}), 0)
Used to avoid the gradient explosion problem in log(x) function when x=0.
Parameters:
- **x**: input tensor
- **offset**: offset size. Default: 1e-6
.. note::
Input tensor falls in [0., 1.] and the output tensor falls in [-log(offset), 0]
"""
return torch.log(torch.clamp(x + offset, max=1.0))
class Model(nn.Module):
"""The margin disparity discrepancy (MDD) is proposed to measure the distribution discrepancy in domain adaptation.
The :math:`y^s` and :math:`y^t` are logits output by the main classifier on the source and target domain respectively.
The :math:`y_{adv}^s` and :math:`y_{adv}^t` are logits output by the adversarial classifier.
They are expected to contain raw, unnormalized scores for each class.
The definition can be described as:
.. math::
\\mathcal{D}_{\\gamma}(\\hat{\\mathcal{S}}, \\hat{\\mathcal{T}}) =
\\gamma \\mathbb{E}_{y^s, y_{adv}^s \\sim\\hat{\\mathcal{S}}} \\log\\left(\\frac{\\exp(y_{adv}^s[h_{y^s}])}{\\sum_j \\exp(y_{adv}^s[j])}\\right) +
\\mathbb{E}_{y^t, y_{adv}^t \\sim\\hat{\\mathcal{T}}} \\log\\left(1-\\frac{\\exp(y_{adv}^t[h_{y^t}])}{\\sum_j \\exp(y_{adv}^t[j])}\\right),
where :math:`\\gamma` is a margin hyper-parameter and :math:`h_y` refers to the predicted label when the logits output is :math:`y`.
You can see more details in `Bridging Theory and Algorithm for Domain Adaptation <https://arxiv.org/abs/1904.05801>`_.
Parameters:
- **margin** (float): margin :math:`\\gamma`. Default: 4
- **reduction** (string, optional): Specifies the reduction to apply to the output:
``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
``'mean'``: the sum of the output will be divided by the number of
elements in the output, ``'sum'``: the output will be summed. Default: ``'mean'``
Inputs: y_s, y_s_adv, y_t, y_t_adv
- **y_s**: logits output :math:`y^s` by the main classifier on the source domain
- **y_s_adv**: logits output :math:`y^s` by the adversarial classifier on the source domain
- **y_t**: logits output :math:`y^t` by the main classifier on the target domain
- **y_t_adv**: logits output :math:`y_{adv}^t` by the adversarial classifier on the target domain
Shape:
- Inputs: :math:`(minibatch, C)` where C = number of classes, or :math:`(minibatch, C, d_1, d_2, ..., d_K)`
with :math:`K \\geq 1` in the case of `K`-dimensional loss.
- Output: scalar. If :attr:`reduction` is ``'none'``, then the same size as the target: :math:`(minibatch)`, or
:math:`(minibatch, d_1, d_2, ..., d_K)` with :math:`K \\geq 1` in the case of K-dimensional loss.
Examples::
>>> num_classes = 2
>>> batch_size = 10
>>> loss = MarginDisparityDiscrepancy(margin=4.)
>>> # logits output from source domain and target domain
>>> y_s, y_t = torch.randn(batch_size, num_classes), torch.randn(batch_size, num_classes)
>>> # adversarial logits output from source domain and target domain
>>> y_s_adv, y_t_adv = torch.randn(batch_size, num_classes), torch.randn(batch_size, num_classes)
>>> output = loss(y_s, y_s_adv, y_t, y_t_adv)
"""
def __init__(self, margin: 'Optional[int]'=4, reduction:
'Optional[str]'='mean'):
super().__init__()
self.margin = margin
self.reduction = reduction
def forward(self, y_s: 'torch.Tensor', y_s_adv: 'torch.Tensor', y_t:
'torch.Tensor',
# ... truncated (>4000 chars) for memory efficiency |
DivisiveNormalization2d | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/vg/cvgg2fijk3fjanwfoyz5tqsq5zl466ja33p5ct3cux3cbat6w24y.py
# Topologically Sorted Source Nodes: [abs_1, suppression_field, add, truediv], Original ATen: [aten.abs, aten.max_pool2d_with_indices, aten.add, aten.div]
# Source node to ATen node mapping:
# abs_1 => abs_1
# add => add
# suppression_field => _low_memory_max_pool2d_with_offsets
# truediv => div
# Graph fragment:
# %abs_1 : [num_users=1] = call_function[target=torch.ops.aten.abs.default](args = (%arg0_1,), kwargs = {})
# %_low_memory_max_pool2d_with_offsets : [num_users=1] = call_function[target=torch.ops.prims._low_memory_max_pool2d_with_offsets.default](args = (%abs_1, [5, 5], [1, 1], [2, 2], [1, 1], False), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem, 1.0), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%arg0_1, %add), kwargs = {})
triton_poi_fused_abs_add_div_max_pool2d_with_indices_0 = async_compile.triton('triton_poi_fused_abs_add_div_max_pool2d_with_indices_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_abs_add_div_max_pool2d_with_indices_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 26, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_abs_add_div_max_pool2d_with_indices_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4) % 4
x0 = xindex % 4
x3 = xindex
tmp191 = tl.load(in_ptr0 + (x3), xmask)
tmp0 = (-2) + x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tmp2 & tmp4
tmp6 = (-2) + x0
tmp7 = tmp6 >= tmp1
tmp8 = tmp6 < tmp3
tmp9 = tmp7 & tmp8
tmp10 = tmp5 & tmp9
tmp11 = tl.load(in_ptr0 + ((-10) + x3), tmp10 & xmask, other=0.0)
tmp12 = tl_math.abs(tmp11)
tmp13 = tl.full(tmp12.shape, float("-inf"), tmp12.dtype)
tmp14 = tl.where(tmp10, tmp12, tmp13)
tmp15 = (-1) + x0
tmp16 = tmp15 >= tmp1
tmp17 = tmp15 < tmp3
tmp18 = tmp16 & tmp17
tmp19 = tmp5 & tmp18
tmp20 = tl.load(in_ptr0 + ((-9) + x3), tmp19 & xmask, other=0.0)
tmp21 = tl_math.abs(tmp20)
tmp22 = tl.full(tmp21.shape, float("-inf"), tmp21.dtype)
tmp23 = tl.where(tmp19, tmp21, tmp22)
tmp24 = triton_helpers.maximum(tmp23, tmp14)
tmp25 = x0
tmp26 = tmp25 >= tmp1
tmp27 = tmp25 < tmp3
tmp28 = tmp26 & tmp27
tmp29 = tmp5 & tmp28
tmp30 = tl.load(in_ptr0 + ((-8) + x3), tmp29 & xmask, other=0.0)
tmp31 = tl_math.abs(tmp30)
tmp32 = tl.full(tmp31.shape, float("-inf"), tmp31.dtype)
tmp33 = tl.where(tmp29, tmp31, tmp32)
tmp34 = triton_helpers.maximum(tmp33, tmp24)
tmp35 = 1 + x0
tmp36 = tmp35 >= tmp1
tmp37 = tmp35 < tmp3
tmp38 = tmp36 & tmp37
tmp39 = tmp5 & tmp38
tmp40 = tl.load(in_ptr0 + ((-7) + x3), tmp39 & xmask, other=0.0)
tmp41 = tl_math.abs(tmp40)
tmp42 = tl.full(tmp41.shape, float("-inf"), tmp41.dtype)
tmp43 = tl.where(tmp39, tmp41, tmp42)
tmp44 = triton_helpers.maximum(tmp43, tmp34)
tmp45 = 2 + x0
tmp46 = tmp45 >= tmp1
tmp47 = tmp45 < tmp3
tmp48 = tmp46 & tmp47
tmp49 = tmp5 & tmp48
tmp50 = tl.load(in_ptr0 + ((-6) + x3), tmp49 & xmask, other=0.0)
tmp51 = tl_math.abs(tmp50)
tmp52 = tl.full(tmp51.shape, float("-inf"), tmp51.dtype)
tmp53 = tl.where(tmp49, tmp51, tmp52)
tmp54 = triton_helpers.maximum(tmp53, tmp44)
tmp55 = (-1) + x1
tmp56 = tmp55 >= tmp1
tmp57 = tmp55 < tmp3
tmp58 = tmp56 & tmp57
tmp59 = tmp58 & tmp9
tmp60 = tl.load(in_ptr0 + ((-6) + x3), tmp59 & xmask, other=0.0)
tmp61 = tl_math.abs(tmp60)
tmp62 = tl.full(tmp61.shape, float("-inf"), tmp61.dtype)
tmp63 = tl.where(tmp59, tmp61, tmp62)
tmp64 = triton_helpers.maximum(tmp63, tmp54)
tmp65 = tmp58 & tmp18
tmp66 = tl.load(in_ptr0 + ((-5) + x3), tmp65 & xmask, other=0.0)
tmp67 = tl_math.abs(tmp66)
tmp68 = tl.full(tmp67.shape, float("-inf"), tmp67.dtype)
tmp69 = tl.where(tmp65, tmp67, tmp68)
tmp70 = triton_helpers.maximum(tmp69, tmp64)
tmp71 = tmp58 & tmp28
tmp72 = tl.load(in_ptr0 + ((-4) + x3), tmp71 & xmask, other=0.0)
tmp73 = tl_math.abs(tmp72)
tmp74 = tl.full(tmp73.shape, float("-inf"), tmp73.dtype)
tmp75 = tl.where(tmp71, tmp73, tmp74)
tmp76 = triton_helpers.maximum(tmp75, tmp70)
tmp77 = tmp58 & tmp38
tmp78 = tl.load(in_ptr0 + ((-3) + x3), tmp77 & xmask, other=0.0)
tmp79 = tl_math.abs(tmp78)
tmp80 = tl.full(tmp79.shape, float("-inf"), tmp79.dtype)
tmp81 = tl.where(tmp77, tmp79, tmp80)
tmp82 = triton_helpers.maximum(tmp81, tmp76)
tmp83 = tmp58 & tmp48
tmp84 = tl.load(in_ptr0 + ((-2) + x3), tmp83 & xmask, other=0.0)
tmp85 = tl_math.abs(tmp84)
tmp86 = tl.full(tmp85.shape, float("-inf"), tmp85.dtype)
tmp87 = tl.where(tmp83, tmp85, tmp86)
tmp88 = triton_helpers.maximum(tmp87, tmp82)
tmp89 = x1
tmp90 = tmp89 >= tmp1
tmp91 = tmp89 < tmp3
tmp92 = tmp90 & tmp91
tmp93 = tmp92 & tmp9
tmp94 = tl.load(in_ptr0 + ((-2) + x3), tmp93 & xmask, other=0.0)
tmp95 = tl_math.abs(tmp94)
tmp96 = tl.full(tmp95.shape, float("-inf"), tmp95.dtype)
tmp97 = tl.where(tmp93, tmp95, tmp96)
tmp98 = triton_helpers.maximum(tmp97, tmp88)
tmp99 = tmp92 & tmp18
tmp100 = tl.load(in_ptr0 + ((-1) + x3), tmp99 & xmask, other=0.0)
tmp101 = tl_math.abs(tmp100)
tmp102 = tl.full(tmp101.shape, float("-inf"), tmp101.dtype)
tmp103 = tl.where(tmp99, tmp101, tmp102)
tmp104 = triton_helpers.maximum(tmp103, tmp98)
tmp105 = tmp92 & tmp28
tmp106 = tl.load(in_ptr0 + (x3), tmp105 & xmask, other=0.0)
tmp107 = tl_math.abs(tmp106)
tmp108 = tl.full(tmp107.shape, float("-inf"), tmp107.dtype)
tmp109 = tl.where(tmp105, tmp107, tmp108)
tmp110 = triton_helpers.maximum(tmp109, tmp104)
tmp111 = tmp92 & tmp38
tmp112 = tl.load(in_ptr0 + (1 + x3), tmp111 & xmask, other=0.0)
tmp113 = tl_math.abs(tmp112)
tmp114 = tl.full(tmp113.shape, float("-inf"), tmp113.dtype)
tmp115 = tl.where(tmp111, tmp113, tmp114)
tmp116 = triton_helpers.maximum(tmp115, tmp110)
tmp117 = tmp92 & tmp48
tmp118 = tl.load(in_ptr0 + (2 + x3), tmp117 & xmask, other=0.0)
tmp119 = tl_math.abs(tmp118)
tmp120 = tl.full(tmp119.shape, float("-inf"), tmp119.dtype)
tmp121 = tl.where(tmp117, tmp119, tmp120)
tmp122 = triton_helpers.maximum(tmp121, tmp116)
tmp123 = 1 + x1
tmp124 = tmp123 >= tmp1
tmp125 = tmp123 < tmp3
tmp126 = tmp124 & tmp125
tmp127 = tmp126 & tmp9
tmp128 = tl.load(in_ptr0 + (2 + x3), tmp127 & xmask, other=0.0)
tmp129 = tl_math.abs(tmp128)
tmp130 = tl.full(tmp129.shape, float("-inf"), tmp129.dtype)
tmp131 = tl.where(tmp127, tmp129, tmp130)
tmp132 = triton_helpers.maximum(tmp131, tmp122)
tmp133 = tmp126 & tmp18
tmp134 = tl.load(in_ptr0 + (3 + x3), tmp133 & xmask, other=0.0)
tmp135 = tl_math.abs(tmp134)
tmp136 = tl.full(tmp135.shape, float("-inf"), tmp135.dtype)
tmp137 = tl.where(tmp133, tmp135, tmp136)
tmp138 = triton_helpers.maximum(tmp137, tmp132)
tmp139 = tmp126 & tmp28
tmp140 = tl.load(in_ptr0 + (4 + x3), tmp139 & xmask, other=0.0)
tmp141 = tl_math.abs(tmp140)
tmp142 = tl.full(tmp141.shape, float("-inf"), tmp141.dtype)
tmp143 = tl.where(tmp139, tmp141, tmp142)
tmp144 = triton_helpers.maximum(tmp143, tmp138)
tmp145 = tmp126 & tmp38
tmp146 = tl.load(in_ptr0 + (5 + x3), tmp145 & xmask, other=0.0)
tmp147 = tl_math.abs(tmp146)
tmp148 = tl.full(tmp147.shape, float("-inf"), tmp147.dtype)
tmp149 = tl.where(tmp145, tmp147, tmp148)
tmp150 = triton_helpers.maximum(tmp149, tmp144)
tmp151 = tmp126 & tmp48
tmp152 = tl.load(in_ptr0 + (6 + x3), tmp151 & xmask, other=0.0)
tmp153 = tl_math.abs(tmp152)
tmp154 = tl.full(tmp153.shape, float("-inf"), tmp153.dtype)
tmp155 = tl.where(tmp151, tmp153, tmp154)
tmp156 = triton_helpers.maximum(tmp155, tmp150)
tmp157 = 2 + x1
tmp158 = tmp157 >= tmp1
tmp159 = tmp157 < tmp3
tmp160 = tmp158 & tmp159
tmp161 = tmp160 & tmp9
tmp162 = tl.load(in_ptr0 + (6 + x3), tmp161 & xmask, other=0.0)
tmp163 = tl_math.abs(tmp162)
tmp164 = tl.full(tmp163.shape, float("-inf"), tmp163.dtype)
tmp165 = tl.where(tmp161, tmp163, tmp164)
tmp166 = triton_helpers.maximum(tmp165, tmp156)
tmp167 = tmp160 & tmp18
tmp168 = tl.load(in_ptr0 + (7 + x3), tmp167 & xmask, other=0.0)
tmp169 = tl_math.abs(tmp168)
tmp170 = tl.full(tmp169.shape, float("-inf"), tmp169.dtype)
tmp171 = tl.where(tmp167, tmp169, tmp170)
tmp172 = triton_helpers.maximum(tmp171, tmp166)
tmp173 = tmp160 & tmp28
tmp174 = tl.load(in_ptr0 + (8 + x3), tmp173 & xmask, other=0.0)
tmp175 = tl_math.abs(tmp174)
tmp176 = tl.full(tmp175.shape, float("-inf"), tmp175.dtype)
tmp177 = tl.where(tmp173, tmp175, tmp176)
tmp178 = triton_helpers.maximum(tmp177, tmp172)
tmp179 = tmp160 & tmp38
tmp180 = tl.load(in_ptr0 + (9 + x3), tmp179 & xmask, other=0.0)
tmp181 = tl_math.abs(tmp180)
tmp182 = tl.full(tmp181.shape, float("-inf"), tmp181.dtype)
tmp183 = tl.where(tmp179, tmp181, tmp182)
tmp184 = triton_helpers.maximum(tmp183, tmp178)
tmp185 = tmp160 & tmp48
tmp186 = tl.load(in_ptr0 + (10 + x3), tmp185 & xmask, other=0.0)
tmp187 = tl_math.abs(tmp186)
tmp188 = tl.full(tmp187.shape, float("-inf"), tmp187.dtype)
tmp189 = tl.where(tmp185, tmp187, tmp188)
tmp190 = triton_helpers.maximum(tmp189, tmp184)
tmp192 = 1.0
tmp193 = tmp190 + tmp192
tmp194 = tmp191 / tmp193
tl.store(in_out_ptr0 + (x3), tmp194, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [abs_1, suppression_field, add, truediv], Original ATen: [aten.abs, aten.max_pool2d_with_indices, aten.add, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_abs_add_div_max_pool2d_with_indices_0.run(buf1, arg0_1, 256, grid=grid(256), stream=stream0)
del arg0_1
return (buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| from torch.nn import Module
import torch
from torch import Tensor
from typing import Union
from typing import Tuple
import torch.nn.functional as F
class DivisiveNormalization2d(Module):
"""Applies a 2D divisive normalization over an input signal composed of several input
planes.
In the simplest case, the output value of the layer with input size :math:`(N, C, H, W)`
and output :math:`(N, C, H, W)`.
Args:
b_type: Type of suppressin field, must be one of (`linf`, `l1`, `l2`).
b_size: The size of the suppression field, must be > 0.
sigma: Constant added to suppression field, must be > 0.
Shape:
- Input: :math:`(N, C, H, W)`
- Output: :math:`(N, C, H, W)`
Examples::
>>> # suppression of size=3, sigma=1
>>> d = DivisiveNormalization2d(b_size=3, sigma=1)
>>> input = torch.randn(20, 16, 50, 50)
>>> output = d(input)
"""
def __init__(self, b_type: 'str'='linf', b_size:
'Union[int, Tuple[int, int]]'=(5, 5), sigma: 'float'=1.0) ->None:
super(DivisiveNormalization2d, self).__init__()
self.sigma = sigma
if isinstance(b_size, int):
self.b_size = b_size, b_size
else:
self.b_size = b_size
self.padding = self.b_size[0] // 2, self.b_size[1] // 2
self.b_type = b_type
def forward(self, input: 'Tensor') ->Tensor:
if self.b_type == 'linf':
suppression_field = F.max_pool2d(torch.abs(input), self.b_size,
1, self.padding, 1)
elif self.b_type == 'l1':
weight = torch.ones((input.shape[1], 1, self.b_size[0], self.
b_size[1]))
suppression_field = F.conv2d(torch.abs(input), weight=weight,
padding=self.padding, groups=input.shape[1])
elif self.b_type == 'l2':
weight = torch.ones((input.shape[1], 1, self.b_size[0], self.
b_size[1]))
suppression_field = torch.sqrt(F.conv2d(input ** 2, weight=
weight, padding=self.padding, groups=input.shape[1]))
else:
raise NotImplementedError
return input / (self.sigma + suppression_field)
def __repr__(self) ->str:
s = 'DivisiveNormalization2d('
s += f'b_type={self.b_type}, b_size={self.b_size}, sigma={self.sigma}'
s += ')'
return s.format(**self.__dict__)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch.nn import Module
from typing import Union
from typing import Tuple
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_abs_add_div_max_pool2d_with_indices_0(in_out_ptr0,
in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 4
x0 = xindex % 4
x3 = xindex
tmp191 = tl.load(in_ptr0 + x3, xmask)
tmp0 = -2 + x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tmp2 & tmp4
tmp6 = -2 + x0
tmp7 = tmp6 >= tmp1
tmp8 = tmp6 < tmp3
tmp9 = tmp7 & tmp8
tmp10 = tmp5 & tmp9
tmp11 = tl.load(in_ptr0 + (-10 + x3), tmp10 & xmask, other=0.0)
tmp12 = tl_math.abs(tmp11)
tmp13 = tl.full(tmp12.shape, float('-inf'), tmp12.dtype)
tmp14 = tl.where(tmp10, tmp12, tmp13)
tmp15 = -1 + x0
tmp16 = tmp15 >= tmp1
tmp17 = tmp15 < tmp3
tmp18 = tmp16 & tmp17
tmp19 = tmp5 & tmp18
tmp20 = tl.load(in_ptr0 + (-9 + x3), tmp19 & xmask, other=0.0)
tmp21 = tl_math.abs(tmp20)
tmp22 = tl.full(tmp21.shape, float('-inf'), tmp21.dtype)
tmp23 = tl.where(tmp19, tmp21, tmp22)
tmp24 = triton_helpers.maximum(tmp23, tmp14)
tmp25 = x0
tmp26 = tmp25 >= tmp1
tmp27 = tmp25 < tmp3
tmp28 = tmp26 & tmp27
tmp29 = tmp5 & tmp28
tmp30 = tl.load(in_ptr0 + (-8 + x3), tmp29 & xmask, other=0.0)
tmp31 = tl_math.abs(tmp30)
tmp32 = tl.full(tmp31.shape, float('-inf'), tmp31.dtype)
tmp33 = tl.where(tmp29, tmp31, tmp32)
tmp34 = triton_helpers.maximum(tmp33, tmp24)
tmp35 = 1 + x0
tmp36 = tmp35 >= tmp1
tmp37 = tmp35 < tmp3
tmp38 = tmp36 & tmp37
tmp39 = tmp5 & tmp38
tmp40 = tl.load(in_ptr0 + (-7 + x3), tmp39 & xmask, other=0.0)
tmp41 = tl_math.abs(tmp40)
tmp42 = tl.full(tmp41.shape, float('-inf'), tmp41.dtype)
tmp43 = tl.where(tmp39, tmp41, tmp42)
tmp44 = triton_helpers.maximum(tmp43, tmp34)
tmp45 = 2 + x0
tmp46 = tmp45 >= tmp1
tmp47 = tmp45 < tmp3
tmp48 = tmp46 & tmp47
tmp49 = tmp5 & tmp48
tmp50 = tl.load(in_ptr0 + (-6 + x3), tmp49 & xmask, other=0.0)
tmp51 = tl_math.abs(tmp50)
tmp52 = tl.full(tmp51.shape, float('-inf'), tmp51.dtype)
tmp53 = tl.where(tmp49, tmp51, tmp52)
tmp54 = triton_helpers.maximum(tmp53, tmp44)
tmp55 = -1 + x1
tmp56 = tmp55 >= tmp1
tmp57 = tmp55 < tmp3
tmp58 = tmp56 & tmp57
tmp59 = tmp58 & tmp9
tmp60 = tl.load(in_ptr0 + (-6 + x3), tmp59 & xmask, other=0.0)
tmp61 = tl_math.abs(tmp60)
tmp62 = tl.full(tmp61.shape, float('-inf'), tmp61.dtype)
tmp63 = tl.where(tmp59, tmp61, tmp62)
tmp64 = triton_helpers.maximum(tmp63, tmp54)
tmp65 = tmp58 & tmp18
tmp66 = tl.load(in_ptr0 + (-5 + x3), tmp65 & xmask, other=0.0)
tmp67 = tl_math.abs(tmp66)
tmp68 = tl.full(tmp67.shape, float('-inf'), tmp67.dtype)
tmp69 = tl.where(tmp65, tmp67, tmp68)
tmp70 = triton_helpers.maximum(tmp69, tmp64)
tmp71 = tmp58 & tmp28
tmp72 = tl.load(in_ptr0 + (-4 + x3), tmp71 & xmask, other=0.0)
tmp73 = tl_math.abs(tmp72)
tmp74 = tl.full(tmp73.shape, float('-inf'), tmp73.dtype)
tmp75 = tl.where(tmp71, tmp73, tmp74)
tmp76 = triton_helpers.maximum(tmp75, tmp70)
tmp77 = tmp58 & tmp38
tmp78 = tl.load(in_ptr0 + (-3 + x3), tmp77 & xmask, other=0.0)
tmp79 = tl_math.abs(tmp78)
tmp80 = tl.full(tmp79.shape, float('-inf'), tmp79.dtype)
tmp81 = tl.where(tmp77, tmp79, tmp80)
tmp82 = triton_helpers.maximum(tmp81, tmp76)
tmp83 = tmp58 & tmp48
tmp84 = tl.load(in_ptr0 + (-2 + x3), tmp83 & xmask, other=0.0)
tmp85 = tl_math.abs(tmp84)
tmp86 = tl.full(tmp85.shape, float('-inf'), tmp85.dtype)
tmp87 = tl.where(tmp83, tmp85, tmp86)
tmp88 = triton_helpers.maximum(tmp87, tmp82)
tmp89 = x1
tmp90 = tmp89 >= tmp1
tmp91 = tmp89 < tmp3
tmp92 = tmp90 & tmp91
tmp93 = tmp92 & tmp9
tmp94 = tl.load(in_ptr0 + (-2 + x3), tmp93 & xmask, other=0.0)
tmp95 = tl_math.abs(tmp94)
tmp96 = tl.full(tmp95.shape, float('-inf'), tmp95.dtype)
tmp97 = tl.where(tmp93, tmp95, tmp96)
tmp98 = triton_helpers.maximum(tmp97, tmp88)
tmp99 = tmp92 & tmp18
tmp100 = tl.load(in_ptr0 + (-1 + x3), tmp99 & xmask, other=0.0)
tmp101 = tl_math.abs(tmp100)
tmp102 = tl.full(tmp101.shape, float('-inf'), tmp101.dtype)
tmp103 = tl.where(tmp99, tmp101, tmp102)
tmp104 = triton_helpers.maximum(tmp103, tmp98)
tmp105 = tmp92 & tmp28
tmp106 = tl.load(in_ptr0 + x3, tmp105 & xmask, other=0.0)
tmp107 = tl_math.abs(tmp106)
tmp108 = tl.full(tmp107.shape, float('-inf'), tmp107.dtype)
tmp109 = tl.where(tmp105, tmp107, tmp108)
tmp110 = triton_helpers.maximum(tmp109, tmp104)
tmp111 = tmp92 & tmp38
tmp112 = tl.load(in_ptr0 + (1 + x3), tmp111 & xmask, other=0.0)
tmp113 = tl_math.abs(tmp112)
tmp114 = tl.full(tmp113.shape, float('-inf'), tmp113.dtype)
tmp115 = tl.where(tmp111, tmp113, tmp114)
tmp116 = triton_helpers.maximum(tmp115, tmp110)
tmp117 = tmp92 & tmp48
tmp118 = tl.load(in_ptr0 + (2 + x3), tmp117 & xmask, other=0.0)
tmp119 = tl_math.abs(tmp118)
tmp120 = tl.full(tmp119.shape, float('-inf'), tmp119.dtype)
tmp121 = tl.where(tmp117, tmp119, tmp120)
tmp122 = triton_helpers.maximum(tmp121, tmp116)
tmp123 = 1 + x1
tmp124 = tmp123 >= tmp1
tmp125 = tmp123 < tmp3
tmp126 = tmp124 & tmp125
tmp127 = tmp126 & tmp9
tmp128 = tl.load(in_ptr0 + (2 + x3), tmp127 & xmask, other=0.0)
tmp129 = tl_math.abs(tmp128)
tmp130 = tl.full(tmp129.shape, float('-inf'), tmp129.dtype)
tmp131 = tl.where(tmp127, tmp129, tmp130)
tmp132 = triton_helpers.maximum(tmp131, tmp122)
tmp133 = tmp126 & tmp18
tmp134 = tl.load(in_ptr0 + (3 + x3), tmp133 & xmask, other=0.0)
tmp135 = tl_math.abs(tmp134)
tmp136 = tl.full(tmp135.shape, float('-inf'), tmp135.dtype)
tmp137 = tl.where(tmp133, tmp135, tmp136)
tmp138 = triton_helpers.maximum(tmp137, tmp132)
tmp139 = tmp126 & tmp28
tmp140 = tl.load(in_ptr0 + (4 + x3), tmp139 & xmask, other=0.0)
tmp141 = tl_math.abs(tmp140)
tmp142 = tl.full(tmp141.shape, float('-inf'), tmp141.dtype)
tmp143 = tl.where(tmp139, tmp141, tmp142)
tmp144 = triton_helpers.maximum(tmp143, tmp138)
tmp145 = tmp126 & tmp38
tmp146 = tl.load(in_ptr0 + (5 + x3), tmp145 & xmask, other=0.0)
tmp147 = tl_math.abs(tmp146)
tmp148 = tl.full(tmp147.shape, float('-inf'), tmp147.dtype)
tmp149 = tl.where(tmp145, tmp147, tmp148)
tmp150 = triton_helpers.maximum(tmp149, tmp144)
tmp151 = tmp126 & tmp48
tmp152 = tl.load(in_ptr0 + (6 + x3), tmp151 & xmask, other=0.0)
tmp153 = tl_math.abs(tmp152)
tmp154 = tl.full(tmp153.shape, float('-inf'), tmp153.dtype)
tmp155 = tl.where(tmp151, tmp153, tmp154)
tmp156 = triton_helpers.maximum(tmp155, tmp150)
tmp157 = 2 + x1
tmp158 = tmp157 >= tmp1
tmp159 = tmp157 < tmp3
tmp160 = tmp158 & tmp159
tmp161 = tmp160 & tmp9
tmp162 = tl.load(in_ptr0 + (6 + x3), tmp161 & xmask, other=0.0)
tmp163 = tl_math.abs(tmp162)
tmp164 = tl.full(tmp163.shape, float('-inf'), tmp163.dtype)
tmp165 = tl.where(tmp161, tmp163, tmp164)
tmp166 = triton_helpers.maximum(tmp165, tmp156)
tmp167 = tmp160 & tmp18
tmp168 = tl.load(in_ptr0 + (7 + x3), tmp167 & xmask, other=0.0)
tmp169 = tl_math.abs(tmp168)
tmp170 = tl.full(tmp169.shape, float('-inf'), tmp169.dtype)
tmp171 = tl.where(tmp167, tmp169, tmp170)
tmp172 = triton_helpers.maximum(tmp171, tmp166)
tmp173 = tmp160 & tmp28
tmp174 = tl.load(in_ptr0 + (8 + x3), tmp173 & xmask, other=0.0)
tmp175 = tl_math.abs(tmp174)
tmp176 = tl.full(tmp175.shape, float('-inf'), tmp175.dtype)
tmp177 = tl.where(tmp173, tmp175, tmp176)
tmp178 = triton_helpers.maximum(tmp177, tmp172)
tmp179 = tmp160 & tmp38
tmp180 = tl.load(in_ptr0 + (9 + x3), tmp179 & xmask, other=0.0)
tmp181 = tl_math.abs(tmp180)
tmp182 = tl.full(tmp181.shape, float('-inf'), tmp181.dtype)
tmp183 = tl.where(tmp179, tmp181, tmp182)
tmp184 = triton_helpers.maximum(tmp183, tmp178)
tmp185 = tmp160 & tmp48
tmp186 = tl.load(in_ptr0 + (10 + x3), tmp185 & xmask, other=0.0)
tmp187 = tl_math.abs(tmp186)
tmp188 = tl.full(tmp187.shape, float('-inf'), tmp187.dtype)
tmp189 = tl.where(tmp185, tmp187, tmp188)
tmp190 = triton_helpers.maximum(tmp189, tmp184)
tmp192 = 1.0
tmp193 = tmp190 + tmp192
tmp194 = tmp191 / tmp193
tl.store(in_out_ptr0 + x3, tmp194, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_abs_add_div_max_pool2d_with_indices_0[grid(256)](buf1,
arg0_1, 256, XBLOCK=128, num_warps=4, num_stages=1)
del arg0_1
return buf1,
class DivisiveNormalization2dNew(Module):
"""Applies a 2D divisive normalization over an input signal composed of several input
planes.
In the simplest case, the output value of the layer with input size :math:`(N, C, H, W)`
and output :math:`(N, C, H, W)`.
Args:
b_type: Type of suppressin field, must be one of (`linf`, `l1`, `l2`).
b_size: The size of the suppression field, must be > 0.
sigma: Constant added to suppression field, must be > 0.
Shape:
- Input: :math:`(N, C, H, W)`
- Output: :math:`(N, C, H, W)`
Examples::
>>> # suppression of size=3, sigma=1
>>> d = DivisiveNormalization2d(b_size=3, sigma=1)
>>> input = torch.randn(20, 16, 50, 50)
>>> output = d(input)
"""
def __init__(self, b_type: 'str'='linf', b_size:
'Union[int, Tuple[int, int]]'=(5, 5), sigma: 'float'=1.0) ->None:
super(DivisiveNormalization2dNew, self).__init__()
self.sigma = sigma
if isinstance(b_size, int):
self.b_size = b_size, b_size
else:
self.b_size = b_size
self.padding = self.b_size[0] // 2, self.b_size[1] // 2
self.b_type = b_type
def __repr__(self) ->str:
s = 'DivisiveNormalization2d('
s += f'b_type={self.b_type}, b_size={self.b_size}, sigma={self.sigma}'
s += ')'
return s.format(**self.__dict__)
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| Noppornying00/constant-fraction-activation | DivisiveNormalization2d | false | 917 | [
"Apache-2.0"
] | 0 | b25745e7339df13e3db34d8c8372d5cbaa3c13bb | https://github.com/Noppornying00/constant-fraction-activation/tree/b25745e7339df13e3db34d8c8372d5cbaa3c13bb | from torch.nn import Module
import torch
from torch import Tensor
from typing import Union
from typing import Tuple
import torch.nn.functional as F
class Model(Module):
"""Applies a 2D divisive normalization over an input signal composed of several input
planes.
In the simplest case, the output value of the layer with input size :math:`(N, C, H, W)`
and output :math:`(N, C, H, W)`.
Args:
b_type: Type of suppressin field, must be one of (`linf`, `l1`, `l2`).
b_size: The size of the suppression field, must be > 0.
sigma: Constant added to suppression field, must be > 0.
Shape:
- Input: :math:`(N, C, H, W)`
- Output: :math:`(N, C, H, W)`
Examples::
>>> # suppression of size=3, sigma=1
>>> d = DivisiveNormalization2d(b_size=3, sigma=1)
>>> input = torch.randn(20, 16, 50, 50)
>>> output = d(input)
"""
def __init__(self, b_type: 'str'='linf', b_size:
'Union[int, Tuple[int, int]]'=(5, 5), sigma: 'float'=1.0) ->None:
super().__init__()
self.sigma = sigma
if isinstance(b_size, int):
self.b_size = b_size, b_size
else:
self.b_size = b_size
self.padding = self.b_size[0] // 2, self.b_size[1] // 2
self.b_type = b_type
def forward(self, input: 'Tensor') ->Tensor:
if self.b_type == 'linf':
suppression_field = F.max_pool2d(torch.abs(input), self.b_size,
1, self.padding, 1)
elif self.b_type == 'l1':
weight = torch.ones((input.shape[1], 1, self.b_size[0], self.
b_size[1]))
suppression_field = F.conv2d(torch.abs(input), weight=weight,
padding=self.padding, groups=input.shape[1])
elif self.b_type == 'l2':
weight = torch.ones((input.shape[1], 1, self.b_size[0], self.
b_size[1]))
suppression_field = torch.sqrt(F.conv2d(input ** 2, weight=
weight, padding=self.padding, groups=input.shape[1]))
else:
raise NotImplementedError
return input / (self.sigma + suppression_field)
def __repr__(self) ->str:
s = 'DivisiveNormalization2d('
s += f'b_type={self.b_type}, b_size={self.b_size}, sigma={self.sigma}'
s += ')'
return s.format(**self.__dict__)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
UPChannelRPN | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/h3/ch3uvw442xky75hq4sbmvpdmcbekbfhnxol6tssfy4yqvnujuz2e.py
# Topologically Sorted Source Nodes: [cls_kernel, pk], Original ATen: [aten.convolution, aten.view]
# Source node to ATen node mapping:
# cls_kernel => convolution
# pk => view
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %view : [num_users=2] = call_function[target=torch.ops.aten.reshape.default](args = (%convolution, [-1, 256, 62, 62]), kwargs = {})
triton_poi_fused_convolution_view_0 = async_compile.triton('triton_poi_fused_convolution_view_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[67108864],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_view_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_view_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 39362560
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x4 = xindex
x1 = (xindex // 3844) % 2560
tmp0 = tl.load(in_out_ptr0 + (x4), None)
tmp1 = tl.load(in_ptr0 + (x1), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x4), tmp2, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/y6/cy6qx57jpiu4vlxl2g25y7rel2psvzfz7ebvhvgjv3bc3t32qrmr.py
# Topologically Sorted Source Nodes: [cls_feature], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# cls_feature => convolution_2
# Graph fragment:
# %convolution_2 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_8, %primals_6, %primals_7, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_1 = async_compile.triton('triton_poi_fused_convolution_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4194304],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_1', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 3936256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = (xindex // 3844) % 256
tmp0 = tl.load(in_out_ptr0 + (x3), None)
tmp1 = tl.load(in_ptr0 + (x1), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x3), tmp2, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/f6/cf64zdbx27ms4xfo56np7zctuwwujlpnt54f3vrr32ypmoj5lfws.py
# Topologically Sorted Source Nodes: [loc_kernel, pk_1], Original ATen: [aten.convolution, aten.view]
# Source node to ATen node mapping:
# loc_kernel => convolution_1
# pk_1 => view_3
# Graph fragment:
# %convolution_1 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_4, %primals_5, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %view_3 : [num_users=2] = call_function[target=torch.ops.aten.reshape.default](args = (%convolution_1, [-1, 256, 62, 62]), kwargs = {})
triton_poi_fused_convolution_view_2 = async_compile.triton('triton_poi_fused_convolution_view_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[134217728],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_view_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_view_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 78725120
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x4 = xindex
x1 = (xindex // 3844) % 5120
tmp0 = tl.load(in_out_ptr0 + (x4), None)
tmp1 = tl.load(in_ptr0 + (x1), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x4), tmp2, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/sd/csdttsqq72udbm3kifpciu6nqccumiwd5q6fjnjfury3wkr56a4c.py
# Topologically Sorted Source Nodes: [loc], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# loc => convolution_6
# Graph fragment:
# %convolution_6 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%view_5, %primals_11, %primals_12, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_3 = async_compile.triton('triton_poi_fused_convolution_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 20
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12 = args
args.clear()
assert_size_stride(primals_1, (2560, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_2, (2560, ), (1, ))
assert_size_stride(primals_3, (4, 256, 64, 64), (1048576, 4096, 64, 1))
assert_size_stride(primals_4, (5120, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_5, (5120, ), (1, ))
assert_size_stride(primals_6, (256, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_7, (256, ), (1, ))
assert_size_stride(primals_8, (4, 256, 64, 64), (1048576, 4096, 64, 1))
assert_size_stride(primals_9, (256, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_10, (256, ), (1, ))
assert_size_stride(primals_11, (20, 20, 1, 1), (20, 1, 1, 1))
assert_size_stride(primals_12, (20, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [cls_kernel], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 2560, 62, 62), (9840640, 3844, 62, 1))
# Topologically Sorted Source Nodes: [loc_kernel], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(primals_3, primals_4, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 5120, 62, 62), (19681280, 3844, 62, 1))
# Topologically Sorted Source Nodes: [cls_feature], Original ATen: [aten.convolution]
buf2 = extern_kernels.convolution(primals_8, primals_6, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 256, 62, 62), (984064, 3844, 62, 1))
# Topologically Sorted Source Nodes: [loc_feature], Original ATen: [aten.convolution]
buf3 = extern_kernels.convolution(primals_8, primals_9, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 256, 62, 62), (984064, 3844, 62, 1))
buf4 = buf0; del buf0 # reuse
buf5 = reinterpret_tensor(buf4, (40, 256, 62, 62), (984064, 3844, 62, 1), 0); del buf4 # reuse
# Topologically Sorted Source Nodes: [cls_kernel, pk], Original ATen: [aten.convolution, aten.view]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_view_0.run(buf5, primals_2, 39362560, grid=grid(39362560), stream=stream0)
del primals_2
buf6 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [cls_feature], Original ATen: [aten.convolution]
triton_poi_fused_convolution_1.run(buf6, primals_7, 3936256, grid=grid(3936256), stream=stream0)
del primals_7
# Topologically Sorted Source Nodes: [po], Original ATen: [aten.convolution]
buf7 = extern_kernels.convolution(reinterpret_tensor(buf6, (1, 1024, 62, 62), (0, 3844, 62, 1), 0), buf5, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=4, bias=None)
assert_size_stride(buf7, (1, 40, 1, 1), (40, 1, 1, 1))
buf8 = buf1; del buf1 # reuse
buf9 = reinterpret_tensor(buf8, (80, 256, 62, 62), (984064, 3844, 62, 1), 0); del buf8 # reuse
# Topologically Sorted Source Nodes: [loc_kernel, pk_1], Original ATen: [aten.convolution, aten.view]
triton_poi_fused_convolution_view_2.run(buf9, primals_5, 78725120, grid=grid(78725120), stream=stream0)
del primals_5
buf10 = buf3; del buf3 # reuse
# Topologically Sorted Source Nodes: [loc_feature], Original ATen: [aten.convolution]
triton_poi_fused_convolution_1.run(buf10, primals_10, 3936256, grid=grid(3936256), stream=stream0)
del primals_10
# Topologically Sorted Source Nodes: [po_2], Original ATen: [aten.convolution]
buf11 = extern_kernels.convolution(reinterpret_tensor(buf10, (1, 1024, 62, 62), (0, 3844, 62, 1), 0), buf9, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=4, bias=None)
assert_size_stride(buf11, (1, 80, 1, 1), (80, 1, 1, 1))
# Topologically Sorted Source Nodes: [loc], Original ATen: [aten.convolution]
buf12 = extern_kernels.convolution(reinterpret_tensor(buf11, (4, 20, 1, 1), (20, 1, 1, 1), 0), primals_11, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf12, (4, 20, 1, 1), (20, 1, 1, 1))
buf13 = buf12; del buf12 # reuse
# Topologically Sorted Source Nodes: [loc], Original ATen: [aten.convolution]
triton_poi_fused_convolution_3.run(buf13, primals_12, 80, grid=grid(80), stream=stream0)
del primals_12
return (reinterpret_tensor(buf7, (4, 10, 1, 1), (10, 1, 1, 1), 0), buf13, primals_1, primals_3, primals_4, primals_6, primals_8, primals_9, primals_11, buf5, reinterpret_tensor(buf6, (1, 1024, 62, 62), (3936256, 3844, 62, 1), 0), buf9, reinterpret_tensor(buf10, (1, 1024, 62, 62), (3936256, 3844, 62, 1), 0), reinterpret_tensor(buf11, (4, 20, 1, 1), (20, 1, 1, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((2560, 256, 3, 3), (2304, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((2560, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 256, 64, 64), (1048576, 4096, 64, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((5120, 256, 3, 3), (2304, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((5120, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((256, 256, 3, 3), (2304, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 256, 64, 64), (1048576, 4096, 64, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((256, 256, 3, 3), (2304, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((20, 20, 1, 1), (20, 1, 1, 1), device='cuda:0', dtype=torch.float32)
primals_12 = rand_strided((20, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11, primals_12])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
def xcorr_fast(x, kernel):
"""group conv2d to calculate cross correlation, fast version
"""
batch = kernel.size()[0]
pk = kernel.view(-1, x.size()[1], kernel.size()[2], kernel.size()[3])
px = x.view(1, -1, x.size()[2], x.size()[3])
po = F.conv2d(px, pk, groups=batch)
po = po.view(batch, -1, po.size()[2], po.size()[3])
return po
class RPN(nn.Module):
def __init__(self):
super(RPN, self).__init__()
def forward(self, z_f, x_f):
raise NotImplementedError
class UPChannelRPN(RPN):
def __init__(self, anchor_num=5, feature_in=256):
super(UPChannelRPN, self).__init__()
cls_output = 2 * anchor_num
loc_output = 4 * anchor_num
self.template_cls_conv = nn.Conv2d(feature_in, feature_in *
cls_output, kernel_size=3)
self.template_loc_conv = nn.Conv2d(feature_in, feature_in *
loc_output, kernel_size=3)
self.search_cls_conv = nn.Conv2d(feature_in, feature_in, kernel_size=3)
self.search_loc_conv = nn.Conv2d(feature_in, feature_in, kernel_size=3)
self.loc_adjust = nn.Conv2d(loc_output, loc_output, kernel_size=1)
def forward(self, z_f, x_f):
cls_kernel = self.template_cls_conv(z_f)
loc_kernel = self.template_loc_conv(z_f)
cls_feature = self.search_cls_conv(x_f)
loc_feature = self.search_loc_conv(x_f)
cls = xcorr_fast(cls_feature, cls_kernel)
loc = self.loc_adjust(xcorr_fast(loc_feature, loc_kernel))
return cls, loc
def get_inputs():
return [torch.rand([4, 256, 64, 64]), torch.rand([4, 256, 64, 64])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_view_0(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x4 = xindex
x1 = xindex // 3844 % 2560
tmp0 = tl.load(in_out_ptr0 + x4, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x4, tmp2, None)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 3844 % 256
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, None)
@triton.jit
def triton_poi_fused_convolution_view_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x4 = xindex
x1 = xindex // 3844 % 5120
tmp0 = tl.load(in_out_ptr0 + x4, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x4, tmp2, None)
@triton.jit
def triton_poi_fused_convolution_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 20
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (2560, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_2, (2560,), (1,))
assert_size_stride(primals_3, (4, 256, 64, 64), (1048576, 4096, 64, 1))
assert_size_stride(primals_4, (5120, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_5, (5120,), (1,))
assert_size_stride(primals_6, (256, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_7, (256,), (1,))
assert_size_stride(primals_8, (4, 256, 64, 64), (1048576, 4096, 64, 1))
assert_size_stride(primals_9, (256, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_10, (256,), (1,))
assert_size_stride(primals_11, (20, 20, 1, 1), (20, 1, 1, 1))
assert_size_stride(primals_12, (20,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 2560, 62, 62), (9840640, 3844, 62, 1))
buf1 = extern_kernels.convolution(primals_3, primals_4, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 5120, 62, 62), (19681280, 3844, 62, 1))
buf2 = extern_kernels.convolution(primals_8, primals_6, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 256, 62, 62), (984064, 3844, 62, 1))
buf3 = extern_kernels.convolution(primals_8, primals_9, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 256, 62, 62), (984064, 3844, 62, 1))
buf4 = buf0
del buf0
buf5 = reinterpret_tensor(buf4, (40, 256, 62, 62), (984064, 3844,
62, 1), 0)
del buf4
get_raw_stream(0)
triton_poi_fused_convolution_view_0[grid(39362560)](buf5, primals_2,
39362560, XBLOCK=512, num_warps=8, num_stages=1)
del primals_2
buf6 = buf2
del buf2
triton_poi_fused_convolution_1[grid(3936256)](buf6, primals_7,
3936256, XBLOCK=512, num_warps=8, num_stages=1)
del primals_7
buf7 = extern_kernels.convolution(reinterpret_tensor(buf6, (1, 1024,
62, 62), (0, 3844, 62, 1), 0), buf5, stride=(1, 1), padding=(0,
0), dilation=(1, 1), transposed=False, output_padding=(0, 0),
groups=4, bias=None)
assert_size_stride(buf7, (1, 40, 1, 1), (40, 1, 1, 1))
buf8 = buf1
del buf1
buf9 = reinterpret_tensor(buf8, (80, 256, 62, 62), (984064, 3844,
62, 1), 0)
del buf8
triton_poi_fused_convolution_view_2[grid(78725120)](buf9, primals_5,
78725120, XBLOCK=512, num_warps=8, num_stages=1)
del primals_5
buf10 = buf3
del buf3
triton_poi_fused_convolution_1[grid(3936256)](buf10, primals_10,
3936256, XBLOCK=512, num_warps=8, num_stages=1)
del primals_10
buf11 = extern_kernels.convolution(reinterpret_tensor(buf10, (1,
1024, 62, 62), (0, 3844, 62, 1), 0), buf9, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=4, bias=None)
assert_size_stride(buf11, (1, 80, 1, 1), (80, 1, 1, 1))
buf12 = extern_kernels.convolution(reinterpret_tensor(buf11, (4, 20,
1, 1), (20, 1, 1, 1), 0), primals_11, stride=(1, 1), padding=(0,
0), dilation=(1, 1), transposed=False, output_padding=(0, 0),
groups=1, bias=None)
assert_size_stride(buf12, (4, 20, 1, 1), (20, 1, 1, 1))
buf13 = buf12
del buf12
triton_poi_fused_convolution_3[grid(80)](buf13, primals_12, 80,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_12
return (reinterpret_tensor(buf7, (4, 10, 1, 1), (10, 1, 1, 1), 0),
buf13, primals_1, primals_3, primals_4, primals_6, primals_8,
primals_9, primals_11, buf5, reinterpret_tensor(buf6, (1, 1024, 62,
62), (3936256, 3844, 62, 1), 0), buf9, reinterpret_tensor(buf10, (1,
1024, 62, 62), (3936256, 3844, 62, 1), 0), reinterpret_tensor(buf11,
(4, 20, 1, 1), (20, 1, 1, 1), 0))
def xcorr_fast(x, kernel):
"""group conv2d to calculate cross correlation, fast version
"""
batch = kernel.size()[0]
pk = kernel.view(-1, x.size()[1], kernel.size()[2], kernel.size()[3])
px = x.view(1, -1, x.size()[2], x.size()[3])
po = F.conv2d(px, pk, groups=batch)
po = po.view(batch, -1, po.size()[2], po.size()[3])
return po
class RPN(nn.Module):
def __init__(self):
super(RPN, self).__init__()
def forward(self, z_f, x_f):
raise NotImplementedError
class UPChannelRPNNew(RPN):
def __init__(self, anchor_num=5, feature_in=256):
super(UPChannelRPNNew, self).__init__()
cls_output = 2 * anchor_num
loc_output = 4 * anchor_num
self.template_cls_conv = nn.Conv2d(feature_in, feature_in *
cls_output, kernel_size=3)
self.template_loc_conv = nn.Conv2d(feature_in, feature_in *
loc_output, kernel_size=3)
self.search_cls_conv = nn.Conv2d(feature_in, feature_in, kernel_size=3)
self.search_loc_conv = nn.Conv2d(feature_in, feature_in, kernel_size=3)
self.loc_adjust = nn.Conv2d(loc_output, loc_output, kernel_size=1)
def forward(self, input_0, input_1):
primals_1 = self.template_cls_conv.weight
primals_2 = self.template_cls_conv.bias
primals_4 = self.template_loc_conv.weight
primals_5 = self.template_loc_conv.bias
primals_6 = self.search_cls_conv.weight
primals_7 = self.search_cls_conv.bias
primals_9 = self.search_loc_conv.weight
primals_10 = self.search_loc_conv.bias
primals_11 = self.loc_adjust.weight
primals_12 = self.loc_adjust.bias
primals_3 = input_0
primals_8 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0], output[1]
| LSH9832/MyPythonModules | UPChannelRPN | false | 918 | [
"MIT"
] | 0 | 442566a0fbd6ebe2bc20b6914686a1e2663d10c0 | https://github.com/LSH9832/MyPythonModules/tree/442566a0fbd6ebe2bc20b6914686a1e2663d10c0 | import torch
import torch.nn as nn
import torch.nn.functional as F
def xcorr_fast(x, kernel):
"""group conv2d to calculate cross correlation, fast version
"""
batch = kernel.size()[0]
pk = kernel.view(-1, x.size()[1], kernel.size()[2], kernel.size()[3])
px = x.view(1, -1, x.size()[2], x.size()[3])
po = F.conv2d(px, pk, groups=batch)
po = po.view(batch, -1, po.size()[2], po.size()[3])
return po
class RPN(nn.Module):
def __init__(self):
super().__init__()
def forward(self, z_f, x_f):
raise NotImplementedError
class Model(RPN):
def __init__(self, anchor_num=5, feature_in=256):
super().__init__()
cls_output = 2 * anchor_num
loc_output = 4 * anchor_num
self.template_cls_conv = nn.Conv2d(feature_in, feature_in *
cls_output, kernel_size=3)
self.template_loc_conv = nn.Conv2d(feature_in, feature_in *
loc_output, kernel_size=3)
self.search_cls_conv = nn.Conv2d(feature_in, feature_in, kernel_size=3)
self.search_loc_conv = nn.Conv2d(feature_in, feature_in, kernel_size=3)
self.loc_adjust = nn.Conv2d(loc_output, loc_output, kernel_size=1)
def forward(self, z_f, x_f):
cls_kernel = self.template_cls_conv(z_f)
loc_kernel = self.template_loc_conv(z_f)
cls_feature = self.search_cls_conv(x_f)
loc_feature = self.search_loc_conv(x_f)
cls = xcorr_fast(cls_feature, cls_kernel)
loc = self.loc_adjust(xcorr_fast(loc_feature, loc_kernel))
return cls, loc
def get_inputs():
return [torch.rand([4, 256, 64, 64]), torch.rand([4, 256, 64, 64])]
def get_init_inputs():
return []
|
MatrixTree | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/mw/cmwulpxxw2jt763yw6dihufmcwawkztohwf6my2rtf33kl2nbgds.py
# Topologically Sorted Source Nodes: [eye, ne, lap, sum_1], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
# Source node to ATen node mapping:
# eye => eq, full_default, full_default_1, iota_1, where
# lap => full_default_2, where_1
# ne => ne
# sum_1 => sum_1
# Graph fragment:
# %iota_1 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze, %iota_1), kwargs = {})
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq, %full_default, %full_default_1), kwargs = {})
# %ne : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where, 0), kwargs = {})
# %full_default_2 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_1 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne, %full_default_2, %select), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%where_1, [0]), kwargs = {})
triton_poi_fused_eye_masked_fill_ne_sum_0 = async_compile.triton('triton_poi_fused_eye_masked_fill_ne_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_eye_masked_fill_ne_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (x0), xmask)
tmp16 = tl.load(in_ptr0 + (4 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (8 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (12 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/5f/c5fvishccr56l7e7obdlxzywjkbh3davnf5cdvnuqxihtgywozfd.py
# Topologically Sorted Source Nodes: [eye, ne, lap, neg, diag, lap_1, diag_1, exp_1], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
# Source node to ATen node mapping:
# diag => eq_1, full_default_3, iota_2, where_2
# diag_1 => diagonal_copy
# exp_1 => exp_1
# eye => eq, full_default, full_default_1, iota_1, where
# lap => full_default_2, where_1
# lap_1 => add_1
# ne => ne
# neg => neg
# Graph fragment:
# %iota_1 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze, %iota_1), kwargs = {})
# %full_default : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_1 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq, %full_default, %full_default_1), kwargs = {})
# %ne : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where, 0), kwargs = {})
# %full_default_2 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_1 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne, %full_default_2, %select), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%where_1,), kwargs = {})
# %iota_2 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_1 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_2, %unsqueeze_2), kwargs = {})
# %full_default_3 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_2 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_1, %permute, %full_default_3), kwargs = {})
# %add_1 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%neg, %where_2), kwargs = {})
# %diagonal_copy : [num_users=1] = call_function[target=torch.ops.aten.diagonal_copy.default](args = (%select_1,), kwargs = {})
# %exp_1 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%diagonal_copy,), kwargs = {})
# %select_scatter_default : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%add_1, %exp_1, 0, 0), kwargs = {})
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1 = async_compile.triton('triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (5*x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (x2), xmask)
tmp18 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + (x2), tmp21, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/5k/c5kir7ouxg6vi6ka3fqtxb3atjlynhq3dcipimsepx55zvv5ls6k.py
# Topologically Sorted Source Nodes: [eye_1, ne_1, lap_2, sum_2], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
# Source node to ATen node mapping:
# eye_1 => eq_3, full_default_7, full_default_8, iota_7, where_4
# lap_2 => full_default_9, where_5
# ne_1 => ne_1
# sum_2 => sum_2
# Graph fragment:
# %iota_7 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_3 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze_6, %iota_7), kwargs = {})
# %full_default_7 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_8 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_4 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_3, %full_default_7, %full_default_8), kwargs = {})
# %ne_1 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where_4, 0), kwargs = {})
# %full_default_9 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_5 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne_1, %full_default_9, %select_20), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%where_5, [0]), kwargs = {})
triton_poi_fused_eye_masked_fill_ne_sum_2 = async_compile.triton('triton_poi_fused_eye_masked_fill_ne_sum_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_eye_masked_fill_ne_sum_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (16 + x0), xmask)
tmp16 = tl.load(in_ptr0 + (20 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (24 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (28 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ny/cnyxgqp3tmqmvliho4wkdr5wi6tnqd2oycsfgq3cjapc2phvva3v.py
# Topologically Sorted Source Nodes: [eye_1, ne_1, lap_2, neg_1, diag_5, lap_3, diag_6, exp_5], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
# Source node to ATen node mapping:
# diag_5 => eq_4, full_default_10, iota_8, where_6
# diag_6 => diagonal_copy_3
# exp_5 => exp_5
# eye_1 => eq_3, full_default_7, full_default_8, iota_7, where_4
# lap_2 => full_default_9, where_5
# lap_3 => add_3
# ne_1 => ne_1
# neg_1 => neg_1
# Graph fragment:
# %iota_7 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_3 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze_6, %iota_7), kwargs = {})
# %full_default_7 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_8 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_4 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_3, %full_default_7, %full_default_8), kwargs = {})
# %ne_1 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where_4, 0), kwargs = {})
# %full_default_9 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_5 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne_1, %full_default_9, %select_20), kwargs = {})
# %neg_1 : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%where_5,), kwargs = {})
# %iota_8 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_4 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_8, %unsqueeze_8), kwargs = {})
# %full_default_10 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_6 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_4, %permute_5, %full_default_10), kwargs = {})
# %add_3 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%neg_1, %where_6), kwargs = {})
# %diagonal_copy_3 : [num_users=1] = call_function[target=torch.ops.aten.diagonal_copy.default](args = (%select_21,), kwargs = {})
# %exp_5 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%diagonal_copy_3,), kwargs = {})
# %select_scatter_default_4 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%add_3, %exp_5, 0, 0), kwargs = {})
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3 = async_compile.triton('triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (16 + (5*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (16 + x2), xmask)
tmp18 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + (x2), tmp21, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/76/c764os7jr65oallcf3eupkgmn4rajwnkb7qa2a4qrpedfegb2xji.py
# Topologically Sorted Source Nodes: [diag_4, add_2], Original ATen: [aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_2 => add_2
# diag_4 => eq_2, full_default_6, iota_4, where_3
# Graph fragment:
# %iota_4 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_2 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_4, %unsqueeze_5), kwargs = {})
# %full_default_6 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_3 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_2, %permute_4, %full_default_6), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_16, %where_3), kwargs = {})
triton_poi_fused_add_diag_embed_4 = async_compile.triton('triton_poi_fused_add_diag_embed_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_4(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp4 = tl.load(in_ptr0 + (x2), xmask)
tmp6 = tl.load(in_ptr1 + (5*x0), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + (x2), xmask)
tmp18 = tl.load(in_ptr0 + (5*x0), xmask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 0, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tmp2 == tmp0
tmp5 = tl_math.exp(tmp4)
tmp7 = tmp5 * tmp6
tmp8 = 0.0
tmp9 = tl.where(tmp3, tmp8, tmp7)
tmp10 = x1
tmp11 = tmp10 == tmp0
tmp13 = tmp5 * tmp12
tmp14 = tl.where(tmp11, tmp8, tmp13)
tmp15 = tmp9 - tmp14
tmp16 = tl.where(tmp1, tmp15, tmp4)
tmp17 = tmp2 == tmp10
tmp19 = tl_math.exp(tmp18)
tmp21 = tmp19 * tmp20
tmp22 = tl.where(tmp17, tmp21, tmp8)
tmp23 = tmp16 + tmp22
tl.store(out_ptr0 + (x2), tmp23, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ry/crysshrrigakeht5r6neboffuaxbkjaegs3cqmvdij2ytziz5vck.py
# Topologically Sorted Source Nodes: [exp_2, mul, setitem_1, exp_3, mul_1, setitem_2, sub, diag_4, add_2], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_2 => add_2
# diag_4 => eq_2, full_default_6, iota_4, where_3
# exp_2 => exp_2
# exp_3 => exp_3
# mul => mul
# mul_1 => mul_1
# setitem_1 => copy_1, full_default_4
# setitem_2 => copy_2, full_default_5
# sub => sub
# Graph fragment:
# %exp_2 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_5,), kwargs = {})
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_2, %permute_1), kwargs = {})
# %full_default_4 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_1 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_7, %full_default_4), kwargs = {})
# %select_scatter_default_1 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul, %copy_1, 1, 0), kwargs = {})
# %exp_3 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_6,), kwargs = {})
# %mul_1 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_3, %permute_2), kwargs = {})
# %full_default_5 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_2 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_9, %full_default_5), kwargs = {})
# %select_scatter_default_2 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_1, %copy_2, 0, 0), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select_scatter_default_1, %select_scatter_default_2), kwargs = {})
# %select_scatter_default_3 : [num_users=3] = call_function[target=torch.ops.aten.select_scatter.default](args = (%arg0_1, %sub, 0, 0), kwargs = {})
# %iota_4 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_2 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_4, %unsqueeze_5), kwargs = {})
# %full_default_6 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_3 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_2, %permute_4, %full_default_6), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_16, %where_3), kwargs = {})
# %select_scatter_default_5 : [num_users=2] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_3, %add_2, 0, 0), kwargs = {})
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5 = async_compile.triton('triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = (xindex // 16)
x3 = xindex % 16
x0 = xindex % 4
x1 = (xindex // 4) % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + (x3), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr1 + (x3), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr2 + (5*x0), xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr2 + (x3), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr1 + (x5), xmask)
tmp0 = x2
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tmp4 == tmp1
tmp7 = tl_math.exp(tmp6)
tmp9 = tmp7 * tmp8
tmp10 = 0.0
tmp11 = tl.where(tmp5, tmp10, tmp9)
tmp12 = x1
tmp13 = tmp12 == tmp1
tmp15 = tmp7 * tmp14
tmp16 = tl.where(tmp13, tmp10, tmp15)
tmp17 = tmp11 - tmp16
tmp19 = tl.where(tmp2, tmp17, tmp18)
tmp20 = tl.where(tmp2, tmp3, tmp19)
tl.store(out_ptr0 + (x5), tmp20, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ds/cdsgimfrigpgjvo5w5eo5m4acelh6kweus6l5g6oa5euxfo5zryv.py
# Topologically Sorted Source Nodes: [eye_2, ne_2, lap_4, sum_3], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
# Source node to ATen node mapping:
# eye_2 => eq_6, full_default_14, full_default_15, iota_13, where_8
# lap_4 => full_default_16, where_9
# ne_2 => ne_2
# sum_3 => sum_3
# Graph fragment:
# %iota_13 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_6 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze_12, %iota_13), kwargs = {})
# %full_default_14 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_15 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_8 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_6, %full_default_14, %full_default_15), kwargs = {})
# %ne_2 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where_8, 0), kwargs = {})
# %full_default_16 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_9 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne_2, %full_default_16, %select_41), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%where_9, [0]), kwargs = {})
triton_poi_fused_eye_masked_fill_ne_sum_6 = async_compile.triton('triton_poi_fused_eye_masked_fill_ne_sum_6', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_eye_masked_fill_ne_sum_6', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_6(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (32 + x0), xmask)
tmp16 = tl.load(in_ptr0 + (36 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (40 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (44 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/bx/cbx52hvmunlgryyn5hbdsqyfni653yiknzgubzvil2q6yaseomvy.py
# Topologically Sorted Source Nodes: [eye_2, ne_2, lap_4, neg_2, diag_10, lap_5, diag_11, exp_9], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
# Source node to ATen node mapping:
# diag_10 => eq_7, full_default_17, iota_14, where_10
# diag_11 => diagonal_copy_6
# exp_9 => exp_9
# eye_2 => eq_6, full_default_14, full_default_15, iota_13, where_8
# lap_4 => full_default_16, where_9
# lap_5 => add_5
# ne_2 => ne_2
# neg_2 => neg_2
# Graph fragment:
# %iota_13 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_6 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze_12, %iota_13), kwargs = {})
# %full_default_14 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_15 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_8 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_6, %full_default_14, %full_default_15), kwargs = {})
# %ne_2 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where_8, 0), kwargs = {})
# %full_default_16 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_9 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne_2, %full_default_16, %select_41), kwargs = {})
# %neg_2 : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%where_9,), kwargs = {})
# %iota_14 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_7 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_14, %unsqueeze_14), kwargs = {})
# %full_default_17 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_10 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_7, %permute_10, %full_default_17), kwargs = {})
# %add_5 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%neg_2, %where_10), kwargs = {})
# %diagonal_copy_6 : [num_users=1] = call_function[target=torch.ops.aten.diagonal_copy.default](args = (%select_42,), kwargs = {})
# %exp_9 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%diagonal_copy_6,), kwargs = {})
# %select_scatter_default_9 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%add_5, %exp_9, 0, 0), kwargs = {})
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7 = async_compile.triton('triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (32 + (5*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (32 + x2), xmask)
tmp18 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + (x2), tmp21, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/mz/cmzl5zoke7j2kqgzcvjpkpdzkg2kwyb47nzibimhi2hklzb5s6ut.py
# Topologically Sorted Source Nodes: [diag_9, add_4], Original ATen: [aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_4 => add_4
# diag_9 => eq_5, full_default_13, iota_10, where_7
# Graph fragment:
# %iota_10 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_5 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_10, %unsqueeze_11), kwargs = {})
# %full_default_13 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_7 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_5, %permute_9, %full_default_13), kwargs = {})
# %add_4 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_37, %where_7), kwargs = {})
triton_poi_fused_add_diag_embed_8 = async_compile.triton('triton_poi_fused_add_diag_embed_8', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_8', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_8(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp5 = tl.load(in_ptr0 + (16 + x2), xmask)
tmp7 = tl.load(in_ptr1 + (5*x0), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr1 + (x2), xmask)
tmp17 = tl.load(in_ptr2 + (16 + x2), xmask)
tmp20 = tl.load(in_ptr0 + (16 + (5*x0)), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 1, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = tmp2 == tmp3
tmp6 = tl_math.exp(tmp5)
tmp8 = tmp6 * tmp7
tmp9 = 0.0
tmp10 = tl.where(tmp4, tmp9, tmp8)
tmp11 = x1
tmp12 = tmp11 == tmp3
tmp14 = tmp6 * tmp13
tmp15 = tl.where(tmp12, tmp9, tmp14)
tmp16 = tmp10 - tmp15
tmp18 = tl.where(tmp1, tmp16, tmp17)
tmp19 = tmp2 == tmp11
tmp21 = tl_math.exp(tmp20)
tmp23 = tmp21 * tmp22
tmp24 = tl.where(tmp19, tmp23, tmp9)
tmp25 = tmp18 + tmp24
tl.store(out_ptr0 + (x2), tmp25, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/4m/c4mfewqtg74k4hoz5gcjnz5z3hfa6tzmoe5hz6cil2jwbmx7cosw.py
# Topologically Sorted Source Nodes: [exp_6, mul_3, setitem_6, exp_7, mul_4, setitem_7, sub_1, diag_9, add_4], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_4 => add_4
# diag_9 => eq_5, full_default_13, iota_10, where_7
# exp_6 => exp_6
# exp_7 => exp_7
# mul_3 => mul_3
# mul_4 => mul_4
# setitem_6 => copy_6, full_default_11
# setitem_7 => copy_7, full_default_12
# sub_1 => sub_1
# Graph fragment:
# %exp_6 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_25,), kwargs = {})
# %mul_3 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_6, %permute_6), kwargs = {})
# %full_default_11 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_6 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_27, %full_default_11), kwargs = {})
# %select_scatter_default_6 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_3, %copy_6, 1, 0), kwargs = {})
# %exp_7 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_26,), kwargs = {})
# %mul_4 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_7, %permute_7), kwargs = {})
# %full_default_12 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_7 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_29, %full_default_12), kwargs = {})
# %select_scatter_default_7 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_4, %copy_7, 0, 0), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select_scatter_default_6, %select_scatter_default_7), kwargs = {})
# %select_scatter_default_8 : [num_users=3] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_5, %sub_1, 0, 1), kwargs = {})
# %iota_10 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_5 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_10, %unsqueeze_11), kwargs = {})
# %full_default_13 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_7 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_5, %permute_9, %full_default_13), kwargs = {})
# %add_4 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_37, %where_7), kwargs = {})
# %select_scatter_default_10 : [num_users=2] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_8, %add_4, 0, 1), kwargs = {})
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9 = async_compile.triton('triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = (xindex // 16)
x3 = xindex % 16
x0 = xindex % 4
x1 = (xindex // 4) % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + (x3), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (16 + x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + (5*x0), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr2 + (x3), xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_out_ptr0 + (x5), xmask)
tmp0 = x2
tmp1 = tl.full([1], 1, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = tmp4 == tmp5
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 * tmp9
tmp11 = 0.0
tmp12 = tl.where(tmp6, tmp11, tmp10)
tmp13 = x1
tmp14 = tmp13 == tmp5
tmp16 = tmp8 * tmp15
tmp17 = tl.where(tmp14, tmp11, tmp16)
tmp18 = tmp12 - tmp17
tmp20 = tl.where(tmp2, tmp18, tmp19)
tmp21 = tl.where(tmp2, tmp3, tmp20)
tl.store(in_out_ptr0 + (x5), tmp21, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/f4/cf43dymhxq3xler6ijzfxsjgm3oemfpzsjaoarspoimxqod3bkiv.py
# Topologically Sorted Source Nodes: [eye_3, ne_3, lap_6, sum_4], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
# Source node to ATen node mapping:
# eye_3 => eq_9, full_default_21, full_default_22, iota_19, where_12
# lap_6 => full_default_23, where_13
# ne_3 => ne_3
# sum_4 => sum_4
# Graph fragment:
# %iota_19 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_9 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze_18, %iota_19), kwargs = {})
# %full_default_21 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_22 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_12 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_9, %full_default_21, %full_default_22), kwargs = {})
# %ne_3 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where_12, 0), kwargs = {})
# %full_default_23 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_13 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne_3, %full_default_23, %select_62), kwargs = {})
# %sum_4 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%where_13, [0]), kwargs = {})
triton_poi_fused_eye_masked_fill_ne_sum_10 = async_compile.triton('triton_poi_fused_eye_masked_fill_ne_sum_10', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_eye_masked_fill_ne_sum_10', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_10(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (48 + x0), xmask)
tmp16 = tl.load(in_ptr0 + (52 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (56 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (60 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/sg/csgmzo5h24yeta5phehu6lzwmojkxl6w5o6pxa7fu4tywve6h72o.py
# Topologically Sorted Source Nodes: [eye_3, ne_3, lap_6, neg_3, diag_15, lap_7, diag_16, exp_13], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
# Source node to ATen node mapping:
# diag_15 => eq_10, full_default_24, iota_20, where_14
# diag_16 => diagonal_copy_9
# exp_13 => exp_13
# eye_3 => eq_9, full_default_21, full_default_22, iota_19, where_12
# lap_6 => full_default_23, where_13
# lap_7 => add_7
# ne_3 => ne_3
# neg_3 => neg_3
# Graph fragment:
# %iota_19 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_9 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%unsqueeze_18, %iota_19), kwargs = {})
# %full_default_21 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([1], 1), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %full_default_22 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_12 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_9, %full_default_21, %full_default_22), kwargs = {})
# %ne_3 : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%where_12, 0), kwargs = {})
# %full_default_23 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_13 : [num_users=2] = call_function[target=torch.ops.aten.where.self](args = (%ne_3, %full_default_23, %select_62), kwargs = {})
# %neg_3 : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%where_13,), kwargs = {})
# %iota_20 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_10 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_20, %unsqueeze_20), kwargs = {})
# %full_default_24 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_14 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_10, %permute_15, %full_default_24), kwargs = {})
# %add_7 : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%neg_3, %where_14), kwargs = {})
# %diagonal_copy_9 : [num_users=1] = call_function[target=torch.ops.aten.diagonal_copy.default](args = (%select_63,), kwargs = {})
# %exp_13 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%diagonal_copy_9,), kwargs = {})
# %select_scatter_default_14 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%add_7, %exp_13, 0, 0), kwargs = {})
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11 = async_compile.triton('triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = (xindex // 4)
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (48 + (5*x0)), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (48 + x2), xmask)
tmp18 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + (x2), tmp21, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/s3/cs3gvd4vxrcwwln23ccoasr7omajc7scigkbutbxw7gq332qtsp2.py
# Topologically Sorted Source Nodes: [diag_14, add_6], Original ATen: [aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_6 => add_6
# diag_14 => eq_8, full_default_20, iota_16, where_11
# Graph fragment:
# %iota_16 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_8 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_16, %unsqueeze_17), kwargs = {})
# %full_default_20 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_11 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_8, %permute_14, %full_default_20), kwargs = {})
# %add_6 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_58, %where_11), kwargs = {})
triton_poi_fused_add_diag_embed_12 = async_compile.triton('triton_poi_fused_add_diag_embed_12', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_12', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_12(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp5 = tl.load(in_ptr0 + (32 + x2), xmask)
tmp7 = tl.load(in_ptr1 + (5*x0), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr1 + (x2), xmask)
tmp17 = tl.load(in_ptr2 + (32 + x2), xmask)
tmp20 = tl.load(in_ptr0 + (32 + (5*x0)), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 2, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = tmp2 == tmp3
tmp6 = tl_math.exp(tmp5)
tmp8 = tmp6 * tmp7
tmp9 = 0.0
tmp10 = tl.where(tmp4, tmp9, tmp8)
tmp11 = x1
tmp12 = tmp11 == tmp3
tmp14 = tmp6 * tmp13
tmp15 = tl.where(tmp12, tmp9, tmp14)
tmp16 = tmp10 - tmp15
tmp18 = tl.where(tmp1, tmp16, tmp17)
tmp19 = tmp2 == tmp11
tmp21 = tl_math.exp(tmp20)
tmp23 = tmp21 * tmp22
tmp24 = tl.where(tmp19, tmp23, tmp9)
tmp25 = tmp18 + tmp24
tl.store(out_ptr0 + (x2), tmp25, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/me/cmej2qnos2ywrczshe25nqkszxvcm53ics5lgh4v3txzow4adwk6.py
# Topologically Sorted Source Nodes: [exp_10, mul_6, setitem_11, exp_11, mul_7, setitem_12, sub_2, diag_14, add_6], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_6 => add_6
# diag_14 => eq_8, full_default_20, iota_16, where_11
# exp_10 => exp_10
# exp_11 => exp_11
# mul_6 => mul_6
# mul_7 => mul_7
# setitem_11 => copy_11, full_default_18
# setitem_12 => copy_12, full_default_19
# sub_2 => sub_2
# Graph fragment:
# %exp_10 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_46,), kwargs = {})
# %mul_6 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_10, %permute_11), kwargs = {})
# %full_default_18 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_11 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_48, %full_default_18), kwargs = {})
# %select_scatter_default_11 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_6, %copy_11, 1, 0), kwargs = {})
# %exp_11 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_47,), kwargs = {})
# %mul_7 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_11, %permute_12), kwargs = {})
# %full_default_19 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_12 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_50, %full_default_19), kwargs = {})
# %select_scatter_default_12 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_7, %copy_12, 0, 0), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select_scatter_default_11, %select_scatter_default_12), kwargs = {})
# %select_scatter_default_13 : [num_users=3] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_10, %sub_2, 0, 2), kwargs = {})
# %iota_16 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_8 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_16, %unsqueeze_17), kwargs = {})
# %full_default_20 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_11 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_8, %permute_14, %full_default_20), kwargs = {})
# %add_6 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_58, %where_11), kwargs = {})
# %select_scatter_default_15 : [num_users=2] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_13, %add_6, 0, 2), kwargs = {})
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13 = async_compile.triton('triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = (xindex // 16)
x3 = xindex % 16
x0 = xindex % 4
x1 = (xindex // 4) % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + (x3), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (32 + x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + (5*x0), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr2 + (x3), xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_out_ptr0 + (x5), xmask)
tmp0 = x2
tmp1 = tl.full([1], 2, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = tmp4 == tmp5
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 * tmp9
tmp11 = 0.0
tmp12 = tl.where(tmp6, tmp11, tmp10)
tmp13 = x1
tmp14 = tmp13 == tmp5
tmp16 = tmp8 * tmp15
tmp17 = tl.where(tmp14, tmp11, tmp16)
tmp18 = tmp12 - tmp17
tmp20 = tl.where(tmp2, tmp18, tmp19)
tmp21 = tl.where(tmp2, tmp3, tmp20)
tl.store(in_out_ptr0 + (x5), tmp21, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/7m/c7mcscnhipdy5wvhd42yefmq35fu4oqgmcy7p53evfyoamhb6fsp.py
# Topologically Sorted Source Nodes: [diag_19, add_8], Original ATen: [aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_8 => add_8
# diag_19 => eq_11, full_default_27, iota_22, where_15
# Graph fragment:
# %iota_22 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_11 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_22, %unsqueeze_23), kwargs = {})
# %full_default_27 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_15 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_11, %permute_19, %full_default_27), kwargs = {})
# %add_8 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_79, %where_15), kwargs = {})
triton_poi_fused_add_diag_embed_14 = async_compile.triton('triton_poi_fused_add_diag_embed_14', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_14', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_14(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp5 = tl.load(in_ptr0 + (48 + x2), xmask)
tmp7 = tl.load(in_ptr1 + (5*x0), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr1 + (x2), xmask)
tmp17 = tl.load(in_ptr2 + (48 + x2), xmask)
tmp20 = tl.load(in_ptr0 + (48 + (5*x0)), xmask, eviction_policy='evict_last')
tmp22 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 3, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = tmp2 == tmp3
tmp6 = tl_math.exp(tmp5)
tmp8 = tmp6 * tmp7
tmp9 = 0.0
tmp10 = tl.where(tmp4, tmp9, tmp8)
tmp11 = x1
tmp12 = tmp11 == tmp3
tmp14 = tmp6 * tmp13
tmp15 = tl.where(tmp12, tmp9, tmp14)
tmp16 = tmp10 - tmp15
tmp18 = tl.where(tmp1, tmp16, tmp17)
tmp19 = tmp2 == tmp11
tmp21 = tl_math.exp(tmp20)
tmp23 = tmp21 * tmp22
tmp24 = tl.where(tmp19, tmp23, tmp9)
tmp25 = tmp18 + tmp24
tl.store(out_ptr0 + (x2), tmp25, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/xd/cxdxskm7n54mdwp6h2mbtokxdjclp2po3usor7fw3tqm473pzik6.py
# Topologically Sorted Source Nodes: [exp_14, mul_9, setitem_16, exp_15, mul_10, setitem_17, sub_3, diag_19, add_8], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
# Source node to ATen node mapping:
# add_8 => add_8
# diag_19 => eq_11, full_default_27, iota_22, where_15
# exp_14 => exp_14
# exp_15 => exp_15
# mul_10 => mul_10
# mul_9 => mul_9
# setitem_16 => copy_16, full_default_25
# setitem_17 => copy_17, full_default_26
# sub_3 => sub_3
# Graph fragment:
# %exp_14 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_67,), kwargs = {})
# %mul_9 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_14, %permute_16), kwargs = {})
# %full_default_25 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_16 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_69, %full_default_25), kwargs = {})
# %select_scatter_default_16 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_9, %copy_16, 1, 0), kwargs = {})
# %exp_15 : [num_users=1] = call_function[target=torch.ops.aten.exp.default](args = (%select_68,), kwargs = {})
# %mul_10 : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%exp_15, %permute_17), kwargs = {})
# %full_default_26 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %copy_17 : [num_users=1] = call_function[target=torch.ops.aten.copy.default](args = (%select_71, %full_default_26), kwargs = {})
# %select_scatter_default_17 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%mul_10, %copy_17, 0, 0), kwargs = {})
# %sub_3 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%select_scatter_default_16, %select_scatter_default_17), kwargs = {})
# %select_scatter_default_18 : [num_users=3] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_15, %sub_3, 0, 3), kwargs = {})
# %iota_22 : [num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (4,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
# %eq_11 : [num_users=1] = call_function[target=torch.ops.aten.eq.Tensor](args = (%iota_22, %unsqueeze_23), kwargs = {})
# %full_default_27 : [num_users=1] = call_function[target=torch.ops.aten.full.default](args = ([], 0.0), kwargs = {dtype: torch.float32, layout: torch.strided, device: cuda:0, pin_memory: False})
# %where_15 : [num_users=1] = call_function[target=torch.ops.aten.where.self](args = (%eq_11, %permute_19, %full_default_27), kwargs = {})
# %add_8 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%select_79, %where_15), kwargs = {})
# %select_scatter_default_19 : [num_users=1] = call_function[target=torch.ops.aten.select_scatter.default](args = (%select_scatter_default_18, %add_8, 0, 3), kwargs = {})
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15 = async_compile.triton('triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = (xindex // 16)
x3 = xindex % 16
x0 = xindex % 4
x1 = (xindex // 4) % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + (x3), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (48 + x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + (5*x0), xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr2 + (x3), xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_out_ptr0 + (x5), xmask)
tmp0 = x2
tmp1 = tl.full([1], 3, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = tmp4 == tmp5
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 * tmp9
tmp11 = 0.0
tmp12 = tl.where(tmp6, tmp11, tmp10)
tmp13 = x1
tmp14 = tmp13 == tmp5
tmp16 = tmp8 * tmp15
tmp17 = tl.where(tmp14, tmp11, tmp16)
tmp18 = tmp12 - tmp17
tmp20 = tl.where(tmp2, tmp18, tmp19)
tmp21 = tl.where(tmp2, tmp3, tmp20)
tl.store(in_out_ptr0 + (x5), tmp21, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [eye, ne, lap, sum_1], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
stream0 = get_raw_stream(0)
triton_poi_fused_eye_masked_fill_ne_sum_0.run(arg0_1, buf0, 4, grid=grid(4), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [eye, ne, lap, neg, diag, lap_1, diag_1, exp_1], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1.run(arg0_1, buf0, buf1, 16, grid=grid(16), stream=stream0)
# Topologically Sorted Source Nodes: [eye, ne, lap, neg, diag, lap_1, diag_1, exp_1, inv_laplacian], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp, aten.linalg_inv_ex]
buf2 = torch.ops.aten.linalg_inv_ex.default(buf1)
buf3 = buf2[0]
del buf2
buf5 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [eye_1, ne_1, lap_2, sum_2], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
triton_poi_fused_eye_masked_fill_ne_sum_2.run(arg0_1, buf5, 4, grid=grid(4), stream=stream0)
buf6 = buf1; del buf1 # reuse
# Topologically Sorted Source Nodes: [eye_1, ne_1, lap_2, neg_1, diag_5, lap_3, diag_6, exp_5], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3.run(arg0_1, buf5, buf6, 16, grid=grid(16), stream=stream0)
# Topologically Sorted Source Nodes: [eye_1, ne_1, lap_2, neg_1, diag_5, lap_3, diag_6, exp_5, inv_laplacian_1], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp, aten.linalg_inv_ex]
buf7 = torch.ops.aten.linalg_inv_ex.default(buf6)
buf8 = buf7[0]
del buf7
buf10 = buf6; del buf6 # reuse
# Topologically Sorted Source Nodes: [diag_4, add_2], Original ATen: [aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_4.run(arg0_1, buf3, buf10, 16, grid=grid(16), stream=stream0)
buf11 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [exp_2, mul, setitem_1, exp_3, mul_1, setitem_2, sub, diag_4, add_2], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5.run(buf10, arg0_1, buf3, buf11, 64, grid=grid(64), stream=stream0)
del buf10
buf12 = buf5; del buf5 # reuse
# Topologically Sorted Source Nodes: [eye_2, ne_2, lap_4, sum_3], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
triton_poi_fused_eye_masked_fill_ne_sum_6.run(arg0_1, buf12, 4, grid=grid(4), stream=stream0)
buf13 = reinterpret_tensor(buf3, (4, 4), (4, 1), 0); del buf3 # reuse
# Topologically Sorted Source Nodes: [eye_2, ne_2, lap_4, neg_2, diag_10, lap_5, diag_11, exp_9], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7.run(arg0_1, buf12, buf13, 16, grid=grid(16), stream=stream0)
# Topologically Sorted Source Nodes: [eye_2, ne_2, lap_4, neg_2, diag_10, lap_5, diag_11, exp_9, inv_laplacian_2], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp, aten.linalg_inv_ex]
buf14 = torch.ops.aten.linalg_inv_ex.default(buf13)
buf15 = buf14[0]
del buf14
buf17 = buf13; del buf13 # reuse
# Topologically Sorted Source Nodes: [diag_9, add_4], Original ATen: [aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_8.run(arg0_1, buf8, buf11, buf17, 16, grid=grid(16), stream=stream0)
buf18 = buf11; del buf11 # reuse
# Topologically Sorted Source Nodes: [exp_6, mul_3, setitem_6, exp_7, mul_4, setitem_7, sub_1, diag_9, add_4], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9.run(buf18, buf17, arg0_1, buf8, 64, grid=grid(64), stream=stream0)
del buf17
buf19 = buf12; del buf12 # reuse
# Topologically Sorted Source Nodes: [eye_3, ne_3, lap_6, sum_4], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.sum]
triton_poi_fused_eye_masked_fill_ne_sum_10.run(arg0_1, buf19, 4, grid=grid(4), stream=stream0)
buf20 = reinterpret_tensor(buf8, (4, 4), (4, 1), 0); del buf8 # reuse
# Topologically Sorted Source Nodes: [eye_3, ne_3, lap_6, neg_3, diag_15, lap_7, diag_16, exp_13], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp]
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11.run(arg0_1, buf19, buf20, 16, grid=grid(16), stream=stream0)
del buf19
# Topologically Sorted Source Nodes: [eye_3, ne_3, lap_6, neg_3, diag_15, lap_7, diag_16, exp_13, inv_laplacian_3], Original ATen: [aten.eye, aten.ne, aten.masked_fill, aten.neg, aten.diag_embed, aten.add, aten.diagonal_copy, aten.exp, aten.linalg_inv_ex]
buf21 = torch.ops.aten.linalg_inv_ex.default(buf20)
buf22 = buf21[0]
del buf21
buf24 = buf20; del buf20 # reuse
# Topologically Sorted Source Nodes: [diag_14, add_6], Original ATen: [aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_12.run(arg0_1, buf15, buf18, buf24, 16, grid=grid(16), stream=stream0)
buf25 = buf18; del buf18 # reuse
# Topologically Sorted Source Nodes: [exp_10, mul_6, setitem_11, exp_11, mul_7, setitem_12, sub_2, diag_14, add_6], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13.run(buf25, buf24, arg0_1, buf15, 64, grid=grid(64), stream=stream0)
del buf15
buf26 = buf24; del buf24 # reuse
# Topologically Sorted Source Nodes: [diag_19, add_8], Original ATen: [aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_14.run(arg0_1, buf22, buf25, buf26, 16, grid=grid(16), stream=stream0)
buf27 = buf25; del buf25 # reuse
# Topologically Sorted Source Nodes: [exp_14, mul_9, setitem_16, exp_15, mul_10, setitem_17, sub_3, diag_19, add_8], Original ATen: [aten.exp, aten.mul, aten.lift_fresh, aten.fill, aten.sub, aten.diag_embed, aten.add]
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15.run(buf27, buf26, arg0_1, buf22, 64, grid=grid(64), stream=stream0)
del arg0_1
del buf22
del buf26
return (buf27, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class MatrixTree(nn.Module):
"""Implementation of the matrix-tree theorem for computing marginals
of non-projective dependency parsing. This attention layer is used
in the paper "Learning Structured Text Representations"
:cite:`DBLP:journals/corr/LiuL17d`.
"""
def __init__(self, eps=1e-05):
self.eps = eps
super(MatrixTree, self).__init__()
def forward(self, input):
laplacian = input.exp() + self.eps
output = input.clone()
for b in range(input.size(0)):
lap = laplacian[b].masked_fill(torch.eye(input.size(1), device=
input.device).ne(0), 0)
lap = -lap + torch.diag(lap.sum(0))
lap[0] = input[b].diag().exp()
inv_laplacian = lap.inverse()
factor = inv_laplacian.diag().unsqueeze(1).expand_as(input[b]
).transpose(0, 1)
term1 = input[b].exp().mul(factor).clone()
term2 = input[b].exp().mul(inv_laplacian.transpose(0, 1)).clone()
term1[:, 0] = 0
term2[0] = 0
output[b] = term1 - term2
roots_output = input[b].diag().exp().mul(inv_laplacian.
transpose(0, 1)[0])
output[b] = output[b] + torch.diag(roots_output)
return output
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
import torch.cuda
import torch.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_0(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + x0, xmask)
tmp16 = tl.load(in_ptr0 + (4 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (8 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (12 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1(
in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + 5 * x0, xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + x2, xmask)
tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + x2, tmp21, xmask)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_2(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (16 + x0), xmask)
tmp16 = tl.load(in_ptr0 + (20 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (24 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (28 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3(
in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (16 + 5 * x0), xmask, eviction_policy='evict_last'
)
tmp11 = tl.load(in_ptr0 + (16 + x2), xmask)
tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + x2, tmp21, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_4(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp4 = tl.load(in_ptr0 + x2, xmask)
tmp6 = tl.load(in_ptr1 + 5 * x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr1 + x2, xmask)
tmp18 = tl.load(in_ptr0 + 5 * x0, xmask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 0, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tmp2 == tmp0
tmp5 = tl_math.exp(tmp4)
tmp7 = tmp5 * tmp6
tmp8 = 0.0
tmp9 = tl.where(tmp3, tmp8, tmp7)
tmp10 = x1
tmp11 = tmp10 == tmp0
tmp13 = tmp5 * tmp12
tmp14 = tl.where(tmp11, tmp8, tmp13)
tmp15 = tmp9 - tmp14
tmp16 = tl.where(tmp1, tmp15, tmp4)
tmp17 = tmp2 == tmp10
tmp19 = tl_math.exp(tmp18)
tmp21 = tmp19 * tmp20
tmp22 = tl.where(tmp17, tmp21, tmp8)
tmp23 = tmp16 + tmp22
tl.store(out_ptr0 + x2, tmp23, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5(in_ptr0,
in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16
x3 = xindex % 16
x0 = xindex % 4
x1 = xindex // 4 % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr1 + x3, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr2 + 5 * x0, xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr1 + x5, xmask)
tmp0 = x2
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tmp4 == tmp1
tmp7 = tl_math.exp(tmp6)
tmp9 = tmp7 * tmp8
tmp10 = 0.0
tmp11 = tl.where(tmp5, tmp10, tmp9)
tmp12 = x1
tmp13 = tmp12 == tmp1
tmp15 = tmp7 * tmp14
tmp16 = tl.where(tmp13, tmp10, tmp15)
tmp17 = tmp11 - tmp16
tmp19 = tl.where(tmp2, tmp17, tmp18)
tmp20 = tl.where(tmp2, tmp3, tmp19)
tl.store(out_ptr0 + x5, tmp20, xmask)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_6(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (32 + x0), xmask)
tmp16 = tl.load(in_ptr0 + (36 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (40 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (44 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7(
in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (32 + 5 * x0), xmask, eviction_policy='evict_last'
)
tmp11 = tl.load(in_ptr0 + (32 + x2), xmask)
tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + x2, tmp21, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_8(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp5 = tl.load(in_ptr0 + (16 + x2), xmask)
tmp7 = tl.load(in_ptr1 + 5 * x0, xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr1 + x2, xmask)
tmp17 = tl.load(in_ptr2 + (16 + x2), xmask)
tmp20 = tl.load(in_ptr0 + (16 + 5 * x0), xmask, eviction_policy=
'evict_last')
tmp22 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 1, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = tmp2 == tmp3
tmp6 = tl_math.exp(tmp5)
tmp8 = tmp6 * tmp7
tmp9 = 0.0
tmp10 = tl.where(tmp4, tmp9, tmp8)
tmp11 = x1
tmp12 = tmp11 == tmp3
tmp14 = tmp6 * tmp13
tmp15 = tl.where(tmp12, tmp9, tmp14)
tmp16 = tmp10 - tmp15
tmp18 = tl.where(tmp1, tmp16, tmp17)
tmp19 = tmp2 == tmp11
tmp21 = tl_math.exp(tmp20)
tmp23 = tmp21 * tmp22
tmp24 = tl.where(tmp19, tmp23, tmp9)
tmp25 = tmp18 + tmp24
tl.store(out_ptr0 + x2, tmp25, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9(in_out_ptr0,
in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16
x3 = xindex % 16
x0 = xindex % 4
x1 = xindex // 4 % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (16 + x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + 5 * x0, xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_out_ptr0 + x5, xmask)
tmp0 = x2
tmp1 = tl.full([1], 1, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = tmp4 == tmp5
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 * tmp9
tmp11 = 0.0
tmp12 = tl.where(tmp6, tmp11, tmp10)
tmp13 = x1
tmp14 = tmp13 == tmp5
tmp16 = tmp8 * tmp15
tmp17 = tl.where(tmp14, tmp11, tmp16)
tmp18 = tmp12 - tmp17
tmp20 = tl.where(tmp2, tmp18, tmp19)
tmp21 = tl.where(tmp2, tmp3, tmp20)
tl.store(in_out_ptr0 + x5, tmp21, xmask)
@triton.jit
def triton_poi_fused_eye_masked_fill_ne_sum_10(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp7 = tl.load(in_ptr0 + (48 + x0), xmask)
tmp16 = tl.load(in_ptr0 + (52 + x0), xmask)
tmp25 = tl.load(in_ptr0 + (56 + x0), xmask)
tmp34 = tl.load(in_ptr0 + (60 + x0), xmask)
tmp0 = tl.full([1], 0, tl.int64)
tmp1 = x0
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != tmp4
tmp8 = tl_math.exp(tmp7)
tmp9 = 1e-05
tmp10 = tmp8 + tmp9
tmp11 = tl.where(tmp6, tmp4, tmp10)
tmp12 = tl.full([1], 1, tl.int64)
tmp13 = tmp12 == tmp1
tmp14 = tl.where(tmp13, tmp3, tmp4)
tmp15 = tmp14 != tmp4
tmp17 = tl_math.exp(tmp16)
tmp18 = tmp17 + tmp9
tmp19 = tl.where(tmp15, tmp4, tmp18)
tmp20 = tmp11 + tmp19
tmp21 = tl.full([1], 2, tl.int64)
tmp22 = tmp21 == tmp1
tmp23 = tl.where(tmp22, tmp3, tmp4)
tmp24 = tmp23 != tmp4
tmp26 = tl_math.exp(tmp25)
tmp27 = tmp26 + tmp9
tmp28 = tl.where(tmp24, tmp4, tmp27)
tmp29 = tmp20 + tmp28
tmp30 = tl.full([1], 3, tl.int64)
tmp31 = tmp30 == tmp1
tmp32 = tl.where(tmp31, tmp3, tmp4)
tmp33 = tmp32 != tmp4
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp35 + tmp9
tmp37 = tl.where(tmp33, tmp4, tmp36)
tmp38 = tmp29 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11(
in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x0 = xindex % 4
x2 = xindex
tmp3 = tl.load(in_ptr0 + (48 + 5 * x0), xmask, eviction_policy='evict_last'
)
tmp11 = tl.load(in_ptr0 + (48 + x2), xmask)
tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = tl_math.exp(tmp3)
tmp5 = x0
tmp6 = tmp0 == tmp5
tmp7 = 1.0
tmp8 = 0.0
tmp9 = tl.where(tmp6, tmp7, tmp8)
tmp10 = tmp9 != tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = 1e-05
tmp14 = tmp12 + tmp13
tmp15 = tl.where(tmp10, tmp8, tmp14)
tmp16 = -tmp15
tmp17 = tmp5 == tmp0
tmp19 = tl.where(tmp17, tmp18, tmp8)
tmp20 = tmp16 + tmp19
tmp21 = tl.where(tmp2, tmp4, tmp20)
tl.store(out_ptr0 + x2, tmp21, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_12(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp5 = tl.load(in_ptr0 + (32 + x2), xmask)
tmp7 = tl.load(in_ptr1 + 5 * x0, xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr1 + x2, xmask)
tmp17 = tl.load(in_ptr2 + (32 + x2), xmask)
tmp20 = tl.load(in_ptr0 + (32 + 5 * x0), xmask, eviction_policy=
'evict_last')
tmp22 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 2, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = tmp2 == tmp3
tmp6 = tl_math.exp(tmp5)
tmp8 = tmp6 * tmp7
tmp9 = 0.0
tmp10 = tl.where(tmp4, tmp9, tmp8)
tmp11 = x1
tmp12 = tmp11 == tmp3
tmp14 = tmp6 * tmp13
tmp15 = tl.where(tmp12, tmp9, tmp14)
tmp16 = tmp10 - tmp15
tmp18 = tl.where(tmp1, tmp16, tmp17)
tmp19 = tmp2 == tmp11
tmp21 = tl_math.exp(tmp20)
tmp23 = tmp21 * tmp22
tmp24 = tl.where(tmp19, tmp23, tmp9)
tmp25 = tmp18 + tmp24
tl.store(out_ptr0 + x2, tmp25, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13(in_out_ptr0,
in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16
x3 = xindex % 16
x0 = xindex % 4
x1 = xindex // 4 % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (32 + x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + 5 * x0, xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_out_ptr0 + x5, xmask)
tmp0 = x2
tmp1 = tl.full([1], 2, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = tmp4 == tmp5
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 * tmp9
tmp11 = 0.0
tmp12 = tl.where(tmp6, tmp11, tmp10)
tmp13 = x1
tmp14 = tmp13 == tmp5
tmp16 = tmp8 * tmp15
tmp17 = tl.where(tmp14, tmp11, tmp16)
tmp18 = tmp12 - tmp17
tmp20 = tl.where(tmp2, tmp18, tmp19)
tmp21 = tl.where(tmp2, tmp3, tmp20)
tl.store(in_out_ptr0 + x5, tmp21, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_14(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp5 = tl.load(in_ptr0 + (48 + x2), xmask)
tmp7 = tl.load(in_ptr1 + 5 * x0, xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr1 + x2, xmask)
tmp17 = tl.load(in_ptr2 + (48 + x2), xmask)
tmp20 = tl.load(in_ptr0 + (48 + 5 * x0), xmask, eviction_policy=
'evict_last')
tmp22 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp0 = tl.full([1], 3, tl.int32)
tmp1 = tmp0 == tmp0
tmp2 = x0
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = tmp2 == tmp3
tmp6 = tl_math.exp(tmp5)
tmp8 = tmp6 * tmp7
tmp9 = 0.0
tmp10 = tl.where(tmp4, tmp9, tmp8)
tmp11 = x1
tmp12 = tmp11 == tmp3
tmp14 = tmp6 * tmp13
tmp15 = tl.where(tmp12, tmp9, tmp14)
tmp16 = tmp10 - tmp15
tmp18 = tl.where(tmp1, tmp16, tmp17)
tmp19 = tmp2 == tmp11
tmp21 = tl_math.exp(tmp20)
tmp23 = tmp21 * tmp22
tmp24 = tl.where(tmp19, tmp23, tmp9)
tmp25 = tmp18 + tmp24
tl.store(out_ptr0 + x2, tmp25, xmask)
@triton.jit
def triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15(in_out_ptr0,
in_ptr0, in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16
x3 = xindex % 16
x0 = xindex % 4
x1 = xindex // 4 % 4
x5 = xindex
tmp3 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (48 + x3), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + 5 * x0, xmask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp19 = tl.load(in_out_ptr0 + x5, xmask)
tmp0 = x2
tmp1 = tl.full([1], 3, tl.int32)
tmp2 = tmp0 == tmp1
tmp4 = x0
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = tmp4 == tmp5
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 * tmp9
tmp11 = 0.0
tmp12 = tl.where(tmp6, tmp11, tmp10)
tmp13 = x1
tmp14 = tmp13 == tmp5
tmp16 = tmp8 * tmp15
tmp17 = tl.where(tmp14, tmp11, tmp16)
tmp18 = tmp12 - tmp17
tmp20 = tl.where(tmp2, tmp18, tmp19)
tmp21 = tl.where(tmp2, tmp3, tmp20)
tl.store(in_out_ptr0 + x5, tmp21, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4,), (1,), torch.float32)
get_raw_stream(0)
triton_poi_fused_eye_masked_fill_ne_sum_0[grid(4)](arg0_1, buf0, 4,
XBLOCK=4, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_1[
grid(16)](arg0_1, buf0, buf1, 16, XBLOCK=16, num_warps=1,
num_stages=1)
buf2 = torch.ops.aten.linalg_inv_ex.default(buf1)
buf3 = buf2[0]
del buf2
buf5 = buf0
del buf0
triton_poi_fused_eye_masked_fill_ne_sum_2[grid(4)](arg0_1, buf5, 4,
XBLOCK=4, num_warps=1, num_stages=1)
buf6 = buf1
del buf1
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_3[
grid(16)](arg0_1, buf5, buf6, 16, XBLOCK=16, num_warps=1,
num_stages=1)
buf7 = torch.ops.aten.linalg_inv_ex.default(buf6)
buf8 = buf7[0]
del buf7
buf10 = buf6
del buf6
triton_poi_fused_add_diag_embed_4[grid(16)](arg0_1, buf3, buf10, 16,
XBLOCK=16, num_warps=1, num_stages=1)
buf11 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_5[grid(64)
](buf10, arg0_1, buf3, buf11, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del buf10
buf12 = buf5
del buf5
triton_poi_fused_eye_masked_fill_ne_sum_6[grid(4)](arg0_1, buf12, 4,
XBLOCK=4, num_warps=1, num_stages=1)
buf13 = reinterpret_tensor(buf3, (4, 4), (4, 1), 0)
del buf3
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_7[
grid(16)](arg0_1, buf12, buf13, 16, XBLOCK=16, num_warps=1,
num_stages=1)
buf14 = torch.ops.aten.linalg_inv_ex.default(buf13)
buf15 = buf14[0]
del buf14
buf17 = buf13
del buf13
triton_poi_fused_add_diag_embed_8[grid(16)](arg0_1, buf8, buf11,
buf17, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf18 = buf11
del buf11
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_9[grid(64)
](buf18, buf17, arg0_1, buf8, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del buf17
buf19 = buf12
del buf12
triton_poi_fused_eye_masked_fill_ne_sum_10[grid(4)](arg0_1, buf19,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf20 = reinterpret_tensor(buf8, (4, 4), (4, 1), 0)
del buf8
triton_poi_fused_add_diag_embed_diagonal_copy_exp_eye_masked_fill_ne_neg_11[
grid(16)](arg0_1, buf19, buf20, 16, XBLOCK=16, num_warps=1,
num_stages=1)
del buf19
buf21 = torch.ops.aten.linalg_inv_ex.default(buf20)
buf22 = buf21[0]
del buf21
buf24 = buf20
del buf20
triton_poi_fused_add_diag_embed_12[grid(16)](arg0_1, buf15, buf18,
buf24, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf25 = buf18
del buf18
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_13[grid(64)
](buf25, buf24, arg0_1, buf15, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del buf15
buf26 = buf24
del buf24
triton_poi_fused_add_diag_embed_14[grid(16)](arg0_1, buf22, buf25,
buf26, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf27 = buf25
del buf25
triton_poi_fused_add_diag_embed_exp_fill_lift_fresh_mul_sub_15[grid(64)
](buf27, buf26, arg0_1, buf22, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del arg0_1
del buf22
del buf26
return buf27,
class MatrixTreeNew(nn.Module):
"""Implementation of the matrix-tree theorem for computing marginals
of non-projective dependency parsing. This attention layer is used
in the paper "Learning Structured Text Representations"
:cite:`DBLP:journals/corr/LiuL17d`.
"""
def __init__(self, eps=1e-05):
self.eps = eps
super(MatrixTreeNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| NaomiatLibrary/OpenNMT-kpg-release | MatrixTree | false | 919 | [
"MIT"
] | 0 | 1da3468d7dad22529a77f3526abf9b373bd3dc4c | https://github.com/NaomiatLibrary/OpenNMT-kpg-release/tree/1da3468d7dad22529a77f3526abf9b373bd3dc4c | import torch
import torch.nn as nn
import torch.cuda
import torch.distributed
class Model(nn.Module):
"""Implementation of the matrix-tree theorem for computing marginals
of non-projective dependency parsing. This attention layer is used
in the paper "Learning Structured Text Representations"
:cite:`DBLP:journals/corr/LiuL17d`.
"""
def __init__(self, eps=1e-05):
self.eps = eps
super().__init__()
def forward(self, input):
laplacian = input.exp() + self.eps
output = input.clone()
for b in range(input.size(0)):
lap = laplacian[b].masked_fill(torch.eye(input.size(1), device=
input.device).ne(0), 0)
lap = -lap + torch.diag(lap.sum(0))
lap[0] = input[b].diag().exp()
inv_laplacian = lap.inverse()
factor = inv_laplacian.diag().unsqueeze(1).expand_as(input[b]
).transpose(0, 1)
term1 = input[b].exp().mul(factor).clone()
term2 = input[b].exp().mul(inv_laplacian.transpose(0, 1)).clone()
term1[:, 0] = 0
term2[0] = 0
output[b] = term1 - term2
roots_output = input[b].diag().exp().mul(inv_laplacian.
transpose(0, 1)[0])
output[b] = output[b] + torch.diag(roots_output)
return output
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return []
|
JointsMSELossNoReduction | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/w4/cw47hqyegjtpbj76g3m6wdm5y3basxa4forj3ogmivmltvv2n6l5.py
# Topologically Sorted Source Nodes: [heatmap_pred_1, mul_1, sub, pow_1, sum_1], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
# Source node to ATen node mapping:
# heatmap_pred_1 => mul
# mul_1 => mul_1
# pow_1 => pow_1
# sub => sub
# sum_1 => sum_1
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze, %select), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_1, %select_1), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul, %mul_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [1]), kwargs = {})
triton_poi_fused_mul_pow_sub_sum_0 = async_compile.triton('triton_poi_fused_mul_pow_sub_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_pow_sub_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 12, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (0))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (16*x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + (0))
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + (4))
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (1 + (16*x0)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr2 + (4))
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + (8))
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (2 + (16*x0)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr2 + (8))
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + (12))
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (3 + (16*x0)), xmask, eviction_policy='evict_last')
tmp33 = tl.load(in_ptr2 + (12))
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/gf/cgfdg57m2kq25gjnusrmcqfjbxfmsvhnozc46s4wultemkiwbfv3.py
# Topologically Sorted Source Nodes: [heatmap_pred_3, mul_4, sub_1, pow_2, sum_2], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
# Source node to ATen node mapping:
# heatmap_pred_3 => mul_3
# mul_4 => mul_4
# pow_2 => pow_2
# sub_1 => sub_1
# sum_2 => sum_2
# Graph fragment:
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_2, %select_2), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_3, %select_3), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_3, %mul_4), kwargs = {})
# %pow_2 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub_1, 2), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_2, [1]), kwargs = {})
triton_poi_fused_mul_pow_sub_sum_1 = async_compile.triton('triton_poi_fused_mul_pow_sub_sum_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_pow_sub_sum_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 12, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (1))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (4 + (16*x0)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + (1))
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + (5))
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (5 + (16*x0)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr2 + (5))
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + (9))
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (6 + (16*x0)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr2 + (9))
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + (13))
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (7 + (16*x0)), xmask, eviction_policy='evict_last')
tmp33 = tl.load(in_ptr2 + (13))
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ud/cudmstgzhe4d5hgeycx5jjjsr7p3mlugt2thu4sph43qkmo2yhuw.py
# Topologically Sorted Source Nodes: [heatmap_pred_5, mul_7, sub_2, pow_3, sum_3], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
# Source node to ATen node mapping:
# heatmap_pred_5 => mul_6
# mul_7 => mul_7
# pow_3 => pow_3
# sub_2 => sub_2
# sum_3 => sum_3
# Graph fragment:
# %mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_4, %select_4), kwargs = {})
# %mul_7 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_5, %select_5), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_6, %mul_7), kwargs = {})
# %pow_3 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub_2, 2), kwargs = {})
# %sum_3 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_3, [1]), kwargs = {})
triton_poi_fused_mul_pow_sub_sum_2 = async_compile.triton('triton_poi_fused_mul_pow_sub_sum_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_pow_sub_sum_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 12, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_2(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (2))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (8 + (16*x0)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + (2))
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + (6))
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (9 + (16*x0)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr2 + (6))
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + (10))
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (10 + (16*x0)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr2 + (10))
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + (14))
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (11 + (16*x0)), xmask, eviction_policy='evict_last')
tmp33 = tl.load(in_ptr2 + (14))
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fe/cfe55sltrebf2lbsv5frianiofe6asqmd35tky7ipprmweivgp3c.py
# Topologically Sorted Source Nodes: [heatmap_pred_7, mul_10, sub_3, pow_4, sum_4], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
# Source node to ATen node mapping:
# heatmap_pred_7 => mul_9
# mul_10 => mul_10
# pow_4 => pow_4
# sub_3 => sub_3
# sum_4 => sum_4
# Graph fragment:
# %mul_9 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_6, %select_6), kwargs = {})
# %mul_10 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%squeeze_7, %select_7), kwargs = {})
# %sub_3 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%mul_9, %mul_10), kwargs = {})
# %pow_4 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub_3, 2), kwargs = {})
# %sum_4 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_4, [1]), kwargs = {})
triton_poi_fused_mul_pow_sub_sum_3 = async_compile.triton('triton_poi_fused_mul_pow_sub_sum_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_pow_sub_sum_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 12, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_3(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (3))
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (12 + (16*x0)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + (3))
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + (7))
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (13 + (16*x0)), xmask, eviction_policy='evict_last')
tmp13 = tl.load(in_ptr2 + (7))
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + (11))
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (14 + (16*x0)), xmask, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr2 + (11))
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + (15))
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (15 + (16*x0)), xmask, eviction_policy='evict_last')
tmp33 = tl.load(in_ptr2 + (15))
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + (x0), tmp38, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/gd/cgdrg63pr72f73rp5nnrjrvij7iiwg77mw4d5cjipod5c5henjos.py
# Topologically Sorted Source Nodes: [loss], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# loss => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%mul_2, %mul_5, %mul_8, %mul_11], 1), kwargs = {})
triton_poi_fused_cat_4 = async_compile.triton('triton_poi_fused_cat_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_4(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = (xindex // 4)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x1), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = 0.5
tmp7 = tmp5 * tmp6
tmp8 = tl.full(tmp7.shape, 0.0, tmp7.dtype)
tmp9 = tl.where(tmp4, tmp7, tmp8)
tmp10 = tmp0 >= tmp3
tmp11 = tl.full([1], 2, tl.int64)
tmp12 = tmp0 < tmp11
tmp13 = tmp10 & tmp12
tmp14 = tl.load(in_ptr1 + (x1), tmp13 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tmp14 * tmp6
tmp16 = tl.full(tmp15.shape, 0.0, tmp15.dtype)
tmp17 = tl.where(tmp13, tmp15, tmp16)
tmp18 = tmp0 >= tmp11
tmp19 = tl.full([1], 3, tl.int64)
tmp20 = tmp0 < tmp19
tmp21 = tmp18 & tmp20
tmp22 = tl.load(in_ptr2 + (x1), tmp21 & xmask, eviction_policy='evict_last', other=0.0)
tmp23 = tmp22 * tmp6
tmp24 = tl.full(tmp23.shape, 0.0, tmp23.dtype)
tmp25 = tl.where(tmp21, tmp23, tmp24)
tmp26 = tmp0 >= tmp19
tmp27 = tl.full([1], 4, tl.int64)
tmp28 = tmp0 < tmp27
tmp29 = tl.load(in_ptr3 + (x1), tmp26 & xmask, eviction_policy='evict_last', other=0.0)
tmp30 = tmp29 * tmp6
tmp31 = tl.full(tmp30.shape, 0.0, tmp30.dtype)
tmp32 = tl.where(tmp26, tmp30, tmp31)
tmp33 = tl.where(tmp21, tmp25, tmp32)
tmp34 = tl.where(tmp13, tmp17, tmp33)
tmp35 = tl.where(tmp4, tmp9, tmp34)
tl.store(out_ptr0 + (x2), tmp35, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
assert_size_stride(arg2_1, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [heatmap_pred_1, mul_1, sub, pow_1, sum_1], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_pow_sub_sum_0.run(arg0_1, arg2_1, arg1_1, buf0, 4, grid=grid(4), stream=stream0)
buf1 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [heatmap_pred_3, mul_4, sub_1, pow_2, sum_2], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
triton_poi_fused_mul_pow_sub_sum_1.run(arg0_1, arg2_1, arg1_1, buf1, 4, grid=grid(4), stream=stream0)
buf2 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [heatmap_pred_5, mul_7, sub_2, pow_3, sum_3], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
triton_poi_fused_mul_pow_sub_sum_2.run(arg0_1, arg2_1, arg1_1, buf2, 4, grid=grid(4), stream=stream0)
buf3 = empty_strided_cuda((4, ), (1, ), torch.float32)
# Topologically Sorted Source Nodes: [heatmap_pred_7, mul_10, sub_3, pow_4, sum_4], Original ATen: [aten.mul, aten.sub, aten.pow, aten.sum]
triton_poi_fused_mul_pow_sub_sum_3.run(arg0_1, arg2_1, arg1_1, buf3, 4, grid=grid(4), stream=stream0)
del arg0_1
del arg1_1
del arg2_1
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [loss], Original ATen: [aten.cat]
triton_poi_fused_cat_4.run(buf0, buf1, buf2, buf3, buf4, 16, grid=grid(16), stream=stream0)
del buf0
del buf1
del buf2
del buf3
return (buf4, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
arg2_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1, arg2_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class JointsMSELossNoReduction(nn.Module):
def __init__(self, use_target_weight, logger):
super(JointsMSELossNoReduction, self).__init__()
self.criterion = lambda x, y: ((x - y) ** 2).sum(1).unsqueeze(1)
self.use_target_weight = use_target_weight
self.logger = logger
def forward(self, output, target, target_weight):
batch_size = output.size(0)
num_joints = output.size(1)
heatmaps_pred = output.reshape((batch_size, num_joints, -1)).split(1, 1
)
heatmaps_gt = target.reshape((batch_size, num_joints, -1)).split(1, 1)
loss = []
for idx in range(num_joints):
heatmap_pred = heatmaps_pred[idx].squeeze()
heatmap_gt = heatmaps_gt[idx].squeeze()
if self.use_target_weight:
heatmap_pred = heatmap_pred.mul(target_weight[:, idx])
loss_val = self.criterion(heatmap_pred, heatmap_gt.mul(
target_weight[:, idx]))
loss.append(0.5 * loss_val)
else:
loss.append(0.5 * self.criterion(heatmap_pred, heatmap_gt))
loss = torch.cat(loss, dim=1)
return loss
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'use_target_weight': 4, 'logger': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + 16 * x0, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + 0)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + 4)
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (1 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp13 = tl.load(in_ptr2 + 4)
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + 8)
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (2 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr2 + 8)
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + 12)
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (3 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp33 = tl.load(in_ptr2 + 12)
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 1)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (4 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp4 = tl.load(in_ptr2 + 1)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + 5)
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (5 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp13 = tl.load(in_ptr2 + 5)
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + 9)
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (6 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr2 + 9)
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + 13)
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (7 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp33 = tl.load(in_ptr2 + 13)
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_2(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 2)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (8 + 16 * x0), xmask, eviction_policy='evict_last'
)
tmp4 = tl.load(in_ptr2 + 2)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + 6)
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (9 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp13 = tl.load(in_ptr2 + 6)
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + 10)
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (10 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr2 + 10)
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + 14)
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (11 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp33 = tl.load(in_ptr2 + 14)
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_mul_pow_sub_sum_3(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 3)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = tl.load(in_ptr1 + (12 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr2 + 3)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK])
tmp9 = tl.load(in_ptr0 + 7)
tmp10 = tl.broadcast_to(tmp9, [XBLOCK])
tmp11 = tl.load(in_ptr1 + (13 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp13 = tl.load(in_ptr2 + 7)
tmp14 = tl.broadcast_to(tmp13, [XBLOCK])
tmp19 = tl.load(in_ptr0 + 11)
tmp20 = tl.broadcast_to(tmp19, [XBLOCK])
tmp21 = tl.load(in_ptr1 + (14 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp23 = tl.load(in_ptr2 + 11)
tmp24 = tl.broadcast_to(tmp23, [XBLOCK])
tmp29 = tl.load(in_ptr0 + 15)
tmp30 = tl.broadcast_to(tmp29, [XBLOCK])
tmp31 = tl.load(in_ptr1 + (15 + 16 * x0), xmask, eviction_policy=
'evict_last')
tmp33 = tl.load(in_ptr2 + 15)
tmp34 = tl.broadcast_to(tmp33, [XBLOCK])
tmp3 = tmp1 * tmp2
tmp6 = tmp5 * tmp2
tmp7 = tmp3 - tmp6
tmp8 = tmp7 * tmp7
tmp12 = tmp10 * tmp11
tmp15 = tmp14 * tmp11
tmp16 = tmp12 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp8 + tmp17
tmp22 = tmp20 * tmp21
tmp25 = tmp24 * tmp21
tmp26 = tmp22 - tmp25
tmp27 = tmp26 * tmp26
tmp28 = tmp18 + tmp27
tmp32 = tmp30 * tmp31
tmp35 = tmp34 * tmp31
tmp36 = tmp32 - tmp35
tmp37 = tmp36 * tmp36
tmp38 = tmp28 + tmp37
tl.store(out_ptr0 + x0, tmp38, xmask)
@triton.jit
def triton_poi_fused_cat_4(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + x1, tmp4 & xmask, eviction_policy='evict_last',
other=0.0)
tmp6 = 0.5
tmp7 = tmp5 * tmp6
tmp8 = tl.full(tmp7.shape, 0.0, tmp7.dtype)
tmp9 = tl.where(tmp4, tmp7, tmp8)
tmp10 = tmp0 >= tmp3
tmp11 = tl.full([1], 2, tl.int64)
tmp12 = tmp0 < tmp11
tmp13 = tmp10 & tmp12
tmp14 = tl.load(in_ptr1 + x1, tmp13 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp15 = tmp14 * tmp6
tmp16 = tl.full(tmp15.shape, 0.0, tmp15.dtype)
tmp17 = tl.where(tmp13, tmp15, tmp16)
tmp18 = tmp0 >= tmp11
tmp19 = tl.full([1], 3, tl.int64)
tmp20 = tmp0 < tmp19
tmp21 = tmp18 & tmp20
tmp22 = tl.load(in_ptr2 + x1, tmp21 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp23 = tmp22 * tmp6
tmp24 = tl.full(tmp23.shape, 0.0, tmp23.dtype)
tmp25 = tl.where(tmp21, tmp23, tmp24)
tmp26 = tmp0 >= tmp19
tl.full([1], 4, tl.int64)
tmp29 = tl.load(in_ptr3 + x1, tmp26 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp30 = tmp29 * tmp6
tmp31 = tl.full(tmp30.shape, 0.0, tmp30.dtype)
tmp32 = tl.where(tmp26, tmp30, tmp31)
tmp33 = tl.where(tmp21, tmp25, tmp32)
tmp34 = tl.where(tmp13, tmp17, tmp33)
tmp35 = tl.where(tmp4, tmp9, tmp34)
tl.store(out_ptr0 + x2, tmp35, xmask)
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4, 4), (4, 1))
assert_size_stride(arg2_1, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4,), (1,), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_pow_sub_sum_0[grid(4)](arg0_1, arg2_1, arg1_1,
buf0, 4, XBLOCK=4, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4,), (1,), torch.float32)
triton_poi_fused_mul_pow_sub_sum_1[grid(4)](arg0_1, arg2_1, arg1_1,
buf1, 4, XBLOCK=4, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4,), (1,), torch.float32)
triton_poi_fused_mul_pow_sub_sum_2[grid(4)](arg0_1, arg2_1, arg1_1,
buf2, 4, XBLOCK=4, num_warps=1, num_stages=1)
buf3 = empty_strided_cuda((4,), (1,), torch.float32)
triton_poi_fused_mul_pow_sub_sum_3[grid(4)](arg0_1, arg2_1, arg1_1,
buf3, 4, XBLOCK=4, num_warps=1, num_stages=1)
del arg0_1
del arg1_1
del arg2_1
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_cat_4[grid(16)](buf0, buf1, buf2, buf3, buf4, 16,
XBLOCK=16, num_warps=1, num_stages=1)
del buf0
del buf1
del buf2
del buf3
return buf4,
class JointsMSELossNoReductionNew(nn.Module):
def __init__(self, use_target_weight, logger):
super(JointsMSELossNoReductionNew, self).__init__()
self.criterion = lambda x, y: ((x - y) ** 2).sum(1).unsqueeze(1)
self.use_target_weight = use_target_weight
self.logger = logger
def forward(self, input_0, input_1, input_2):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
output = call([arg0_1, arg1_1, arg2_1])
return output[0]
| NiteshBharadwaj/ignoringhumanpose | JointsMSELossNoReduction | false | 920 | [
"MIT"
] | 0 | 1fb7a063fded9cff18f7de4e1d71845983077256 | https://github.com/NiteshBharadwaj/ignoringhumanpose/tree/1fb7a063fded9cff18f7de4e1d71845983077256 | import torch
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
class Model(nn.Module):
def __init__(self, use_target_weight, logger):
super().__init__()
self.criterion = lambda x, y: ((x - y) ** 2).sum(1).unsqueeze(1)
self.use_target_weight = use_target_weight
self.logger = logger
def forward(self, output, target, target_weight):
batch_size = output.size(0)
num_joints = output.size(1)
heatmaps_pred = output.reshape((batch_size, num_joints, -1)).split(1, 1
)
heatmaps_gt = target.reshape((batch_size, num_joints, -1)).split(1, 1)
loss = []
for idx in range(num_joints):
heatmap_pred = heatmaps_pred[idx].squeeze()
heatmap_gt = heatmaps_gt[idx].squeeze()
if self.use_target_weight:
heatmap_pred = heatmap_pred.mul(target_weight[:, idx])
loss_val = self.criterion(heatmap_pred, heatmap_gt.mul(
target_weight[:, idx]))
loss.append(0.5 * loss_val)
else:
loss.append(0.5 * self.criterion(heatmap_pred, heatmap_gt))
loss = torch.cat(loss, dim=1)
return loss
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
Attention | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/nj/cnjqzm7hm3u6ggjfvpspnl6pii56jckgwol3ydau4qesjyoteutl.py
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%permute, %permute_1], 2), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[128],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x2 = (xindex // 32)
x1 = (xindex // 8) % 4
x3 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + ((4*x2) + x0), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tl.load(in_ptr1 + ((4*x2) + (16*x1) + ((-4) + x0)), tmp6 & xmask, eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + (x3), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/hz/chzi3aam26mikdhljz5x7jlqazm7kpktzeptsf36thgfhsg7ub6a.py
# Topologically Sorted Source Nodes: [energy], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# energy => amax, exp, sub
# Graph fragment:
# %amax : [num_users=1] = call_function[target=torch.ops.aten.amax.default](args = (%view_1, [2], True), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%view_1, %amax), kwargs = {})
# %exp : [num_users=2] = call_function[target=torch.ops.aten.exp.default](args = (%sub,), kwargs = {})
triton_poi_fused__softmax_1 = async_compile.triton('triton_poi_fused__softmax_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + (x2), tmp9, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/em/cem6qbxwbiqnjqybzk5arf2obt5uggy4qs7otwwpovvnrhvdc6h4.py
# Topologically Sorted Source Nodes: [energy], Original ATen: [aten._softmax]
# Source node to ATen node mapping:
# energy => div, sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%exp, [2], True), kwargs = {})
# %div : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%exp, %sum_1), kwargs = {})
triton_poi_fused__softmax_2 = async_compile.triton('triton_poi_fused__softmax_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + (x2), tmp8, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/2d/c2dr3b76oqwlimt46tz4bftt63mdbj2eclhl6o6li35qwyf5jdzo.py
# Topologically Sorted Source Nodes: [repeat_1], Original ATen: [aten.repeat]
# Source node to ATen node mapping:
# repeat_1 => repeat_1
# Graph fragment:
# %repeat_1 : [num_users=1] = call_function[target=torch.ops.aten.repeat.default](args = (%primals_5, [4, 1]), kwargs = {})
triton_poi_fused_repeat_3 = async_compile.triton('triton_poi_fused_repeat_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_repeat_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_repeat_3(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + (x2), tmp0, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/wi/cwihcoqftbkppgwobea5wbaobihbvfvl6jipij3fbwihs3o366tl.py
# Topologically Sorted Source Nodes: [relu], Original ATen: [aten.relu, aten.threshold_backward]
# Source node to ATen node mapping:
# relu => relu
# Graph fragment:
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%squeeze,), kwargs = {})
# %le : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu, 0), kwargs = {})
triton_poi_fused_relu_threshold_backward_4 = async_compile.triton('triton_poi_fused_relu_threshold_backward_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*i1', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_threshold_backward_4', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_threshold_backward_4(in_out_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp3 = 0.0
tmp4 = tmp2 <= tmp3
tl.store(in_out_ptr0 + (x0), tmp2, xmask)
tl.store(out_ptr0 + (x0), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 8), (8, 1))
assert_size_stride(primals_4, (4, ), (1, ))
assert_size_stride(primals_5, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(primals_2, primals_1, buf0, 128, grid=grid(128), stream=stream0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_4, reinterpret_tensor(buf0, (16, 8), (8, 1), 0), reinterpret_tensor(primals_3, (8, 4), (1, 8), 0), alpha=1, beta=1, out=buf1)
del primals_3
del primals_4
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [energy], Original ATen: [aten._softmax]
triton_poi_fused__softmax_1.run(buf1, buf2, 64, grid=grid(64), stream=stream0)
buf3 = reinterpret_tensor(buf1, (4, 4, 4), (16, 4, 1), 0); del buf1 # reuse
# Topologically Sorted Source Nodes: [energy], Original ATen: [aten._softmax]
triton_poi_fused__softmax_2.run(buf2, buf3, 64, grid=grid(64), stream=stream0)
del buf2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [repeat_1], Original ATen: [aten.repeat]
triton_poi_fused_repeat_3.run(primals_5, buf4, 16, grid=grid(16), stream=stream0)
del primals_5
buf5 = empty_strided_cuda((4, 1, 4), (4, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [energy_2], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf4, (4, 1, 4), (4, 0, 1), 0), reinterpret_tensor(buf3, (4, 4, 4), (16, 1, 4), 0), out=buf5)
buf6 = reinterpret_tensor(buf5, (4, 4), (4, 1), 0); del buf5 # reuse
buf7 = empty_strided_cuda((4, 4), (4, 1), torch.bool)
# Topologically Sorted Source Nodes: [relu], Original ATen: [aten.relu, aten.threshold_backward]
triton_poi_fused_relu_threshold_backward_4.run(buf6, buf7, 16, grid=grid(16), stream=stream0)
return (reinterpret_tensor(buf6, (4, 1, 4), (4, 4, 1), 0), reinterpret_tensor(buf0, (16, 8), (8, 1), 0), buf3, buf7, reinterpret_tensor(buf4, (4, 4, 1), (4, 1, 4), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4), (16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 8), (8, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.nn
import torch.optim
from torch.nn import functional as F
from torch import nn
class Attention(nn.Module):
def __init__(self, hidden_size):
super(Attention, self).__init__()
self.hidden_size = hidden_size
self.attn = nn.Linear(self.hidden_size * 2, hidden_size)
self.v = nn.Parameter(torch.rand(hidden_size))
stdv = 1.0 / math.sqrt(self.v.size(0))
self.v.data.uniform_(-stdv, stdv)
def forward(self, hidden, encoder_outputs):
timestep = encoder_outputs.size(0)
h = hidden.repeat(timestep, 1, 1).transpose(0, 1)
encoder_outputs = encoder_outputs.transpose(0, 1)
attn_energies = self.score(h, encoder_outputs)
return F.relu(attn_energies).unsqueeze(1)
def score(self, hidden, encoder_outputs):
energy = F.softmax(self.attn(torch.cat([hidden, encoder_outputs], 2
)), dim=2)
energy = energy.transpose(1, 2)
v = self.v.repeat(encoder_outputs.size(0), 1).unsqueeze(1)
energy = torch.bmm(v, energy)
return energy.squeeze(1)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'hidden_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import math
import torch.nn
import torch.optim
from torch.nn import functional as F
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x2 = xindex // 32
x1 = xindex // 8 % 4
x3 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x2 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x2 + 16 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x3, tmp10, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_repeat_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + x2, tmp0, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_4(in_out_ptr0, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp3 = 0.0
tmp4 = tmp2 <= tmp3
tl.store(in_out_ptr0 + x0, tmp2, xmask)
tl.store(out_ptr0 + x0, tmp4, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 8), (8, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(128)](primals_2, primals_1, buf0, 128,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_1
del primals_2
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_4, reinterpret_tensor(buf0, (16, 8), (
8, 1), 0), reinterpret_tensor(primals_3, (8, 4), (1, 8), 0),
alpha=1, beta=1, out=buf1)
del primals_3
del primals_4
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(64)](buf1, buf2, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf3 = reinterpret_tensor(buf1, (4, 4, 4), (16, 4, 1), 0)
del buf1
triton_poi_fused__softmax_2[grid(64)](buf2, buf3, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf2
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_repeat_3[grid(16)](primals_5, buf4, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((4, 1, 4), (4, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf4, (4, 1, 4), (4, 0, 1), 0
), reinterpret_tensor(buf3, (4, 4, 4), (16, 1, 4), 0), out=buf5)
buf6 = reinterpret_tensor(buf5, (4, 4), (4, 1), 0)
del buf5
buf7 = empty_strided_cuda((4, 4), (4, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_4[grid(16)](buf6, buf7, 16,
XBLOCK=16, num_warps=1, num_stages=1)
return reinterpret_tensor(buf6, (4, 1, 4), (4, 4, 1), 0
), reinterpret_tensor(buf0, (16, 8), (8, 1), 0
), buf3, buf7, reinterpret_tensor(buf4, (4, 4, 1), (4, 1, 4), 0)
class AttentionNew(nn.Module):
def __init__(self, hidden_size):
super(AttentionNew, self).__init__()
self.hidden_size = hidden_size
self.attn = nn.Linear(self.hidden_size * 2, hidden_size)
self.v = nn.Parameter(torch.rand(hidden_size))
stdv = 1.0 / math.sqrt(self.v.size(0))
self.v.data.uniform_(-stdv, stdv)
def score(self, hidden, encoder_outputs):
energy = F.softmax(self.attn(torch.cat([hidden, encoder_outputs], 2
)), dim=2)
energy = energy.transpose(1, 2)
v = self.v.repeat(encoder_outputs.size(0), 1).unsqueeze(1)
energy = torch.bmm(v, energy)
return energy.squeeze(1)
def forward(self, input_0, input_1):
primals_4 = self.v
primals_3 = self.attn.weight
primals_5 = self.attn.bias
primals_2 = input_0
primals_1 = input_1
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| OneAdder/hseling-repo-chukchi-type | Attention | false | 921 | [
"MIT"
] | 0 | 5f5e651510bca7cfb89dc2e98b07bcc63b6330a4 | https://github.com/OneAdder/hseling-repo-chukchi-type/tree/5f5e651510bca7cfb89dc2e98b07bcc63b6330a4 | import math
import torch
import torch.nn
import torch.optim
from torch.nn import functional as F
from torch import nn
class Model(nn.Module):
def __init__(self, hidden_size):
super().__init__()
self.hidden_size = hidden_size
self.attn = nn.Linear(self.hidden_size * 2, hidden_size)
self.v = nn.Parameter(torch.rand(hidden_size))
stdv = 1.0 / math.sqrt(self.v.size(0))
self.v.data.uniform_(-stdv, stdv)
def forward(self, hidden, encoder_outputs):
timestep = encoder_outputs.size(0)
h = hidden.repeat(timestep, 1, 1).transpose(0, 1)
encoder_outputs = encoder_outputs.transpose(0, 1)
attn_energies = self.score(h, encoder_outputs)
return F.relu(attn_energies).unsqueeze(1)
def score(self, hidden, encoder_outputs):
energy = F.softmax(self.attn(torch.cat([hidden, encoder_outputs], 2
)), dim=2)
energy = energy.transpose(1, 2)
v = self.v.repeat(encoder_outputs.size(0), 1).unsqueeze(1)
energy = torch.bmm(v, energy)
return energy.squeeze(1)
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [4]
|
myLoss3 | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/iy/ciyeaaq4drcnk4ofbs3khggbmxbh2l7t7bquvanj6fq7ttjsksnz.py
# Topologically Sorted Source Nodes: [loss_1, loss_2, mul, add, loss_3, mul_1, add_1, loss], Original ATen: [aten.mse_loss, aten.mul, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# loss => div
# loss_1 => mean, pow_1, sub
# loss_2 => mean_1, pow_2, sub_1
# loss_3 => mean_2, pow_3, sub_2
# mul => mul
# mul_1 => mul_1
# Graph fragment:
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg1_1, %arg0_1), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%pow_1,), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg3_1, %arg2_1), kwargs = {})
# %pow_2 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub_1, 2), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%pow_2,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mean_1, 1.0), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean, %mul), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg5_1, %arg4_1), kwargs = {})
# %pow_3 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub_2, 2), kwargs = {})
# %mean_2 : [num_users=1] = call_function[target=torch.ops.aten.mean.default](args = (%pow_3,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mean_2, 1.0), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%add, %mul_1), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%add_1, 3.0), kwargs = {})
triton_per_fused_add_div_mse_loss_mul_0 = async_compile.triton('triton_per_fused_add_div_mse_loss_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp32', 7: 'i32', 8: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {7: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 8), equal_to_1=(7,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_div_mse_loss_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 6, 'num_reduction': 3, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_div_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = tl.load(in_ptr1 + (r0), None)
tmp7 = tl.load(in_ptr2 + (r0), None)
tmp8 = tl.load(in_ptr3 + (r0), None)
tmp14 = tl.load(in_ptr4 + (r0), None)
tmp15 = tl.load(in_ptr5 + (r0), None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp9 = tmp7 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tl.broadcast_to(tmp10, [RBLOCK])
tmp13 = triton_helpers.promote_to_tensor(tl.sum(tmp11, 0))
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tl.broadcast_to(tmp17, [RBLOCK])
tmp20 = triton_helpers.promote_to_tensor(tl.sum(tmp18, 0))
tmp21 = 256.0
tmp22 = tmp6 / tmp21
tmp23 = tmp13 / tmp21
tmp24 = 1.0
tmp25 = tmp23 * tmp24
tmp26 = tmp22 + tmp25
tmp27 = tmp20 / tmp21
tmp28 = tmp27 * tmp24
tmp29 = tmp26 + tmp28
tmp30 = 0.3333333333333333
tmp31 = tmp29 * tmp30
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp31, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg3_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg4_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg5_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf3 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [loss_1, loss_2, mul, add, loss_3, mul_1, add_1, loss], Original ATen: [aten.mse_loss, aten.mul, aten.add, aten.div]
stream0 = get_raw_stream(0)
triton_per_fused_add_div_mse_loss_mul_0.run(buf3, arg1_1, arg0_1, arg3_1, arg2_1, arg5_1, arg4_1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
del arg2_1
del arg3_1
del arg4_1
del arg5_1
return (buf3, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg2_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg3_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg4_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg5_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
import torch.nn as nn
class myLoss3(nn.Module):
def __init__(self, alpha=1.0, beta=1.0):
super(myLoss3, self).__init__()
self.alpha = alpha
self.beta = beta
def forward(self, sent_probs, doc_probs, event_probs, sent_targets,
doc_targets, event_targets):
loss_1 = F.mse_loss(sent_probs, sent_targets)
loss_2 = F.mse_loss(doc_probs, doc_targets)
loss_3 = F.mse_loss(event_probs, event_targets)
norm = 1.0 + self.alpha + self.beta
loss = (loss_1 + self.alpha * loss_2 + self.beta * loss_3) / norm
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]),
torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_div_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, in_ptr3, in_ptr4, in_ptr5, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp7 = tl.load(in_ptr2 + r0, None)
tmp8 = tl.load(in_ptr3 + r0, None)
tmp14 = tl.load(in_ptr4 + r0, None)
tmp15 = tl.load(in_ptr5 + r0, None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp9 = tmp7 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tl.broadcast_to(tmp10, [RBLOCK])
tmp13 = triton_helpers.promote_to_tensor(tl.sum(tmp11, 0))
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tl.broadcast_to(tmp17, [RBLOCK])
tmp20 = triton_helpers.promote_to_tensor(tl.sum(tmp18, 0))
tmp21 = 256.0
tmp22 = tmp6 / tmp21
tmp23 = tmp13 / tmp21
tmp24 = 1.0
tmp25 = tmp23 * tmp24
tmp26 = tmp22 + tmp25
tmp27 = tmp20 / tmp21
tmp28 = tmp27 * tmp24
tmp29 = tmp26 + tmp28
tmp30 = 0.3333333333333333
tmp31 = tmp29 * tmp30
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp31, None)
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg3_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg4_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg5_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf3 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_div_mse_loss_mul_0[grid(1)](buf3, arg1_1,
arg0_1, arg3_1, arg2_1, arg5_1, arg4_1, 1, 256, num_warps=2,
num_stages=1)
del arg0_1
del arg1_1
del arg2_1
del arg3_1
del arg4_1
del arg5_1
return buf3,
class myLoss3New(nn.Module):
def __init__(self, alpha=1.0, beta=1.0):
super(myLoss3New, self).__init__()
self.alpha = alpha
self.beta = beta
def forward(self, input_0, input_1, input_2, input_3, input_4, input_5):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
arg3_1 = input_3
arg4_1 = input_4
arg5_1 = input_5
output = call([arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1])
return output[0]
| PKULiuHui/LiveBlogSum | myLoss3 | false | 924 | [
"MIT"
] | 0 | b6a22521ee454e649981d70ddca6c89a1bac5a4c | https://github.com/PKULiuHui/LiveBlogSum/tree/b6a22521ee454e649981d70ddca6c89a1bac5a4c | import torch
import torch.nn.functional as F
import torch.nn as nn
class Model(nn.Module):
def __init__(self, alpha=1.0, beta=1.0):
super().__init__()
self.alpha = alpha
self.beta = beta
def forward(self, sent_probs, doc_probs, event_probs, sent_targets,
doc_targets, event_targets):
loss_1 = F.mse_loss(sent_probs, sent_targets)
loss_2 = F.mse_loss(doc_probs, doc_targets)
loss_3 = F.mse_loss(event_probs, event_targets)
norm = 1.0 + self.alpha + self.beta
loss = (loss_1 + self.alpha * loss_2 + self.beta * loss_3) / norm
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]),
torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
GraphLearner | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/m6/cm645lheesrjji6wgkstt4nu675ugbbjruised3fke4juyuyosol.py
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
# Source node to ATen node mapping:
# _weight_norm => pow_1, pow_2, sum_1
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_2, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [1], True), kwargs = {})
# %pow_2 : [num_users=2] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_1, 0.5), kwargs = {})
triton_poi_fused__weight_norm_interface_0 = async_compile.triton('triton_poi_fused__weight_norm_interface_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__weight_norm_interface_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__weight_norm_interface_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (4*x0), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x0)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (2 + (4*x0)), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (3 + (4*x0)), xmask, eviction_policy='evict_last')
tmp1 = tmp0 * tmp0
tmp3 = tmp2 * tmp2
tmp4 = tmp1 + tmp3
tmp6 = tmp5 * tmp5
tmp7 = tmp4 + tmp6
tmp9 = tmp8 * tmp8
tmp10 = tmp7 + tmp9
tmp11 = libdevice.sqrt(tmp10)
tl.store(out_ptr0 + (x0), tmp11, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/dp/cdpmihjazxc2dpfye4tlkemiovtq5jgmt3cquzgrtbm3gn32us7u.py
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
# Source node to ATen node mapping:
# _weight_norm => div, mul
# Graph fragment:
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, %pow_2), kwargs = {})
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %div), kwargs = {})
triton_poi_fused__weight_norm_interface_1 = async_compile.triton('triton_poi_fused__weight_norm_interface_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__weight_norm_interface_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused__weight_norm_interface_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr1 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + (x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 / tmp2
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/gm/cgmflgdlpeeb52xctoa47uvw47ycyf7ahlj5wdscxdatpbwcboco.py
# Topologically Sorted Source Nodes: [h_1], Original ATen: [aten.relu, aten.threshold_backward]
# Source node to ATen node mapping:
# h_1 => relu
# Graph fragment:
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%view_1,), kwargs = {})
# %le_1 : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu, 0), kwargs = {})
triton_poi_fused_relu_threshold_backward_2 = async_compile.triton('triton_poi_fused_relu_threshold_backward_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*i1', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_threshold_backward_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_threshold_backward_2(in_out_ptr0, in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
tl.store(out_ptr0 + (x2), tmp6, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/mb/cmbwfe33bqdlolenpgjirxfhtdzznq3zzu4sgjelogzwsbh4242z.py
# Topologically Sorted Source Nodes: [h_3], Original ATen: [aten.relu]
# Source node to ATen node mapping:
# h_3 => relu_1
# Graph fragment:
# %relu_1 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%view_3,), kwargs = {})
triton_poi_fused_relu_3 = async_compile.triton('triton_poi_fused_relu_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7 = args
args.clear()
assert_size_stride(primals_1, (4, 1), (1, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4, 1), (1, 1))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
stream0 = get_raw_stream(0)
triton_poi_fused__weight_norm_interface_0.run(primals_2, buf0, 4, grid=grid(4), stream=stream0)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [_weight_norm], Original ATen: [aten._weight_norm_interface]
triton_poi_fused__weight_norm_interface_1.run(primals_2, primals_1, buf0, buf1, 16, grid=grid(16), stream=stream0)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_4, (64, 4), (4, 1), 0), reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf2)
buf3 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
# Topologically Sorted Source Nodes: [_weight_norm_1], Original ATen: [aten._weight_norm_interface]
triton_poi_fused__weight_norm_interface_0.run(primals_6, buf3, 4, grid=grid(4), stream=stream0)
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [_weight_norm_1], Original ATen: [aten._weight_norm_interface]
triton_poi_fused__weight_norm_interface_1.run(primals_6, primals_5, buf3, buf4, 16, grid=grid(16), stream=stream0)
buf5 = reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf2 # reuse
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
# Topologically Sorted Source Nodes: [h_1], Original ATen: [aten.relu, aten.threshold_backward]
triton_poi_fused_relu_threshold_backward_2.run(buf5, primals_3, buf9, 256, grid=grid(256), stream=stream0)
del primals_3
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf5, (64, 4), (4, 1), 0), reinterpret_tensor(buf4, (4, 4), (1, 4), 0), out=buf6)
buf7 = reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf6 # reuse
# Topologically Sorted Source Nodes: [h_3], Original ATen: [aten.relu]
triton_poi_fused_relu_3.run(buf7, primals_7, 256, grid=grid(256), stream=stream0)
del primals_7
buf8 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [adjacency_matrix], Original ATen: [aten.bmm]
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf7, (16, 4, 4), (16, 1, 4), 0), out=buf8)
return (buf8, buf1, buf4, primals_1, primals_2, primals_5, primals_6, buf0, reinterpret_tensor(primals_4, (64, 4), (4, 1), 0), buf3, reinterpret_tensor(buf5, (64, 4), (4, 1), 0), buf7, buf4, buf9, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 1), (1, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((4, 1), (1, 1), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| from torch.nn import Module
import torch
from torch.nn.modules.module import Module
import torch.nn as nn
import torch.nn.functional as F
class GraphLearner(Module):
def __init__(self, in_feature_dim, combined_feature_dim, n_obj, dropout=0.0
):
"""
eq(1): A=EE^T, build adj matrix
## Variables:
- in_feature_dim: dimensionality of input features
- combined_feature_dim: dimensionality of the joint hidden embedding
- k: number of graph nodes/objects on the image
"""
super(GraphLearner, self).__init__()
self.in_dim = in_feature_dim
self.combined_dim = combined_feature_dim
self.n_obj = n_obj
self.edge_layer_1 = nn.Linear(in_feature_dim, combined_feature_dim)
self.edge_layer_2 = nn.Linear(combined_feature_dim,
combined_feature_dim)
self.dropout = nn.Dropout(p=dropout)
self.edge_layer_1 = nn.utils.weight_norm(self.edge_layer_1)
self.edge_layer_2 = nn.utils.weight_norm(self.edge_layer_2)
def forward(self, graph_nodes):
"""
## Inputs:
- graph_nodes (batch_size, K, in_feat_dim): input features
## Returns:
- adjacency matrix (batch_size, K, K)
"""
h = self.edge_layer_1(graph_nodes)
h = F.relu(h)
h = self.edge_layer_2(h)
h = F.relu(h)
h = h.view(-1, self.n_obj, self.combined_dim)
adjacency_matrix = torch.matmul(h, h.transpose(1, 2))
return adjacency_matrix
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_feature_dim': 4, 'combined_feature_dim': 4, 'n_obj': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
from torch.nn import Module
from torch.nn.modules.module import Module
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__weight_norm_interface_0(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp1 = tmp0 * tmp0
tmp3 = tmp2 * tmp2
tmp4 = tmp1 + tmp3
tmp6 = tmp5 * tmp5
tmp7 = tmp4 + tmp6
tmp9 = tmp8 * tmp8
tmp10 = tmp7 + tmp9
tmp11 = libdevice.sqrt(tmp10)
tl.store(out_ptr0 + x0, tmp11, xmask)
@triton.jit
def triton_poi_fused__weight_norm_interface_1(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp3 = tmp1 / tmp2
tmp4 = tmp0 * tmp3
tl.store(out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_2(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr0 + x2, tmp6, xmask)
@triton.jit
def triton_poi_fused_relu_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 1), (1, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4, 1), (1, 1))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__weight_norm_interface_0[grid(4)](primals_2, buf0,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused__weight_norm_interface_1[grid(16)](primals_2,
primals_1, buf0, buf1, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_4, (64, 4), (4, 1), 0),
reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf2)
buf3 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
triton_poi_fused__weight_norm_interface_0[grid(4)](primals_6, buf3,
4, XBLOCK=4, num_warps=1, num_stages=1)
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused__weight_norm_interface_1[grid(16)](primals_6,
primals_5, buf3, buf4, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf5 = reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf2
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_2[grid(256)](buf5,
primals_3, buf9, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_3
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf5, (64, 4), (4, 1), 0),
reinterpret_tensor(buf4, (4, 4), (1, 4), 0), out=buf6)
buf7 = reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf6
triton_poi_fused_relu_3[grid(256)](buf7, primals_7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_7
buf8 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf7, (16, 4, 4), (16, 1, 4), 0), out=buf8)
return (buf8, buf1, buf4, primals_1, primals_2, primals_5, primals_6,
buf0, reinterpret_tensor(primals_4, (64, 4), (4, 1), 0), buf3,
reinterpret_tensor(buf5, (64, 4), (4, 1), 0), buf7, buf4, buf9)
class GraphLearnerNew(Module):
def __init__(self, in_feature_dim, combined_feature_dim, n_obj, dropout=0.0
):
"""
eq(1): A=EE^T, build adj matrix
## Variables:
- in_feature_dim: dimensionality of input features
- combined_feature_dim: dimensionality of the joint hidden embedding
- k: number of graph nodes/objects on the image
"""
super(GraphLearnerNew, self).__init__()
self.in_dim = in_feature_dim
self.combined_dim = combined_feature_dim
self.n_obj = n_obj
self.edge_layer_1 = nn.Linear(in_feature_dim, combined_feature_dim)
self.edge_layer_2 = nn.Linear(combined_feature_dim,
combined_feature_dim)
self.dropout = nn.Dropout(p=dropout)
self.edge_layer_1 = nn.utils.weight_norm(self.edge_layer_1)
self.edge_layer_2 = nn.utils.weight_norm(self.edge_layer_2)
def forward(self, input_0):
primals_3 = self.edge_layer_1.bias
primals_1 = self.edge_layer_1.weight_g
primals_2 = self.edge_layer_1.weight_v
primals_7 = self.edge_layer_2.bias
primals_5 = self.edge_layer_2.weight_g
primals_6 = self.edge_layer_2.weight_v
primals_4 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
| Originofamonia/vqa-project | GraphLearner | false | 925 | [
"Apache-2.0"
] | 0 | cf67b62ddf5732881dfb4278478accfd15c4df6d | https://github.com/Originofamonia/vqa-project/tree/cf67b62ddf5732881dfb4278478accfd15c4df6d | from torch.nn import Module
import torch
from torch.nn.modules.module import Module
import torch.nn as nn
import torch.nn.functional as F
class Model(Module):
def __init__(self, in_feature_dim, combined_feature_dim, n_obj, dropout=0.0
):
"""
eq(1): A=EE^T, build adj matrix
## Variables:
- in_feature_dim: dimensionality of input features
- combined_feature_dim: dimensionality of the joint hidden embedding
- k: number of graph nodes/objects on the image
"""
super().__init__()
self.in_dim = in_feature_dim
self.combined_dim = combined_feature_dim
self.n_obj = n_obj
self.edge_layer_1 = nn.Linear(in_feature_dim, combined_feature_dim)
self.edge_layer_2 = nn.Linear(combined_feature_dim,
combined_feature_dim)
self.dropout = nn.Dropout(p=dropout)
self.edge_layer_1 = nn.utils.weight_norm(self.edge_layer_1)
self.edge_layer_2 = nn.utils.weight_norm(self.edge_layer_2)
def forward(self, graph_nodes):
"""
## Inputs:
- graph_nodes (batch_size, K, in_feat_dim): input features
## Returns:
- adjacency matrix (batch_size, K, K)
"""
h = self.edge_layer_1(graph_nodes)
h = F.relu(h)
h = self.edge_layer_2(h)
h = F.relu(h)
h = h.view(-1, self.n_obj, self.combined_dim)
adjacency_matrix = torch.matmul(h, h.transpose(1, 2))
return adjacency_matrix
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
AE_Net | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/md/cmd3ewacyhu5w5hausgbjbmtnt5rr66cgczh4ibdypq7dz6p4v7g.py
# Topologically Sorted Source Nodes: [activation_1], Original ATen: [aten.relu, aten.threshold_backward]
# Source node to ATen node mapping:
# activation_1 => relu
# Graph fragment:
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%view_1,), kwargs = {})
# %le_2 : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu, 0), kwargs = {})
triton_poi_fused_relu_threshold_backward_0 = async_compile.triton('triton_poi_fused_relu_threshold_backward_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[8192],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*i1', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_threshold_backward_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 8192
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + (x2), None)
tmp1 = tl.load(in_ptr0 + (x0), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + (x2), tmp4, None)
tl.store(out_ptr0 + (x2), tmp6, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/pn/cpnlnh3b3wlvct5iqro5dsv2ynht445jjp4wh5go67vobwu2lnni.py
# Topologically Sorted Source Nodes: [code], Original ATen: [aten.relu, aten.threshold_backward]
# Source node to ATen node mapping:
# code => relu_1
# Graph fragment:
# %relu_1 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%view_3,), kwargs = {})
# %le_1 : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu_1, 0), kwargs = {})
triton_poi_fused_relu_threshold_backward_1 = async_compile.triton('triton_poi_fused_relu_threshold_backward_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[4096],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*i1', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_threshold_backward_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_threshold_backward_1(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), None)
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp3 = 0.0
tmp4 = tmp2 <= tmp3
tl.store(out_ptr0 + (x0), tmp2, None)
tl.store(out_ptr1 + (x0), tmp4, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9 = args
args.clear()
assert_size_stride(primals_1, (128, 4), (4, 1))
assert_size_stride(primals_2, (128, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (64, 128), (128, 1))
assert_size_stride(primals_5, (64, ), (1, ))
assert_size_stride(primals_6, (128, 64), (64, 1))
assert_size_stride(primals_7, (128, ), (1, ))
assert_size_stride(primals_8, (4, 128), (128, 1))
assert_size_stride(primals_9, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 128), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 128), (2048, 512, 128, 1), 0); del buf0 # reuse
buf9 = empty_strided_cuda((4, 4, 4, 128), (2048, 512, 128, 1), torch.bool)
# Topologically Sorted Source Nodes: [activation_1], Original ATen: [aten.relu, aten.threshold_backward]
stream0 = get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0.run(buf1, primals_2, buf9, 8192, grid=grid(8192), stream=stream0)
del primals_2
buf2 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
# Topologically Sorted Source Nodes: [state_logit], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 128), (128, 1), 0), reinterpret_tensor(primals_4, (128, 64), (1, 128), 0), alpha=1, beta=1, out=buf2)
del primals_5
buf3 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.float32)
buf8 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.bool)
# Topologically Sorted Source Nodes: [code], Original ATen: [aten.relu, aten.threshold_backward]
triton_poi_fused_relu_threshold_backward_1.run(buf2, buf3, buf8, 4096, grid=grid(4096), stream=stream0)
buf4 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf3, (64, 64), (64, 1), 0), reinterpret_tensor(primals_6, (64, 128), (1, 64), 0), out=buf4)
buf5 = reinterpret_tensor(buf4, (4, 4, 4, 128), (2048, 512, 128, 1), 0); del buf4 # reuse
buf7 = empty_strided_cuda((4, 4, 4, 128), (2048, 512, 128, 1), torch.bool)
# Topologically Sorted Source Nodes: [activation_3], Original ATen: [aten.relu, aten.threshold_backward]
triton_poi_fused_relu_threshold_backward_0.run(buf5, primals_7, buf7, 8192, grid=grid(8192), stream=stream0)
del primals_7
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [activation_4], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_9, reinterpret_tensor(buf5, (64, 128), (128, 1), 0), reinterpret_tensor(primals_8, (128, 4), (1, 128), 0), alpha=1, beta=1, out=buf6)
del primals_9
return (reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0), reinterpret_tensor(buf2, (4, 4, 4, 64), (1024, 256, 64, 1), 0), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(buf1, (64, 128), (128, 1), 0), reinterpret_tensor(buf3, (64, 64), (64, 1), 0), reinterpret_tensor(buf5, (64, 128), (128, 1), 0), primals_8, buf7, primals_6, buf8, primals_4, buf9, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((128, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((128, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((64, 128), (128, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((128, 64), (64, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((128, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((4, 128), (128, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch.optim import *
from torch import nn
class AE_Net(nn.Module):
"""docstring for AE_Net."""
def __init__(self, input_shape):
super(AE_Net, self).__init__()
self.encoder_hidden_layer = nn.Linear(in_features=input_shape,
out_features=128)
self.encoder_output_layer = nn.Linear(in_features=128, out_features=64)
self.decoder_hidden_layer = nn.Linear(in_features=64, out_features=128)
self.decoder_output_layer = nn.Linear(in_features=128, out_features
=input_shape)
def forward(self, features):
activation = self.encoder_hidden_layer(features)
activation = torch.relu(activation)
state_logit = self.encoder_output_layer(activation)
code = torch.relu(state_logit)
activation = self.decoder_hidden_layer(code)
activation = torch.relu(activation)
activation = self.decoder_output_layer(activation)
reconstructed = activation
return reconstructed, state_logit
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_shape': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch.optim import *
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
@triton.jit
def triton_poi_fused_relu_threshold_backward_1(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, None)
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp3 = 0.0
tmp4 = tmp2 <= tmp3
tl.store(out_ptr0 + x0, tmp2, None)
tl.store(out_ptr1 + x0, tmp4, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (128, 4), (4, 1))
assert_size_stride(primals_2, (128,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (64, 128), (128, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (128, 64), (64, 1))
assert_size_stride(primals_7, (128,), (1,))
assert_size_stride(primals_8, (4, 128), (128, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 128), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 128), (2048, 512, 128, 1), 0)
del buf0
buf9 = empty_strided_cuda((4, 4, 4, 128), (2048, 512, 128, 1),
torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(8192)](buf1,
primals_2, buf9, 8192, XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 128),
(128, 1), 0), reinterpret_tensor(primals_4, (128, 64), (1, 128),
0), alpha=1, beta=1, out=buf2)
del primals_5
buf3 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.
float32)
buf8 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.bool
)
triton_poi_fused_relu_threshold_backward_1[grid(4096)](buf2, buf3,
buf8, 4096, XBLOCK=128, num_warps=4, num_stages=1)
buf4 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf3, (64, 64), (64, 1), 0),
reinterpret_tensor(primals_6, (64, 128), (1, 64), 0), out=buf4)
buf5 = reinterpret_tensor(buf4, (4, 4, 4, 128), (2048, 512, 128, 1), 0)
del buf4
buf7 = empty_strided_cuda((4, 4, 4, 128), (2048, 512, 128, 1),
torch.bool)
triton_poi_fused_relu_threshold_backward_0[grid(8192)](buf5,
primals_7, buf7, 8192, XBLOCK=128, num_warps=4, num_stages=1)
del primals_7
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, reinterpret_tensor(buf5, (64, 128),
(128, 1), 0), reinterpret_tensor(primals_8, (128, 4), (1, 128),
0), alpha=1, beta=1, out=buf6)
del primals_9
return reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(buf2, (4, 4, 4, 64), (1024, 256, 64, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), reinterpret_tensor(buf1, (64, 128), (128, 1), 0
), reinterpret_tensor(buf3, (64, 64), (64, 1), 0), reinterpret_tensor(
buf5, (64, 128), (128, 1), 0
), primals_8, buf7, primals_6, buf8, primals_4, buf9
class AE_NetNew(nn.Module):
"""docstring for AE_Net."""
def __init__(self, input_shape):
super(AE_NetNew, self).__init__()
self.encoder_hidden_layer = nn.Linear(in_features=input_shape,
out_features=128)
self.encoder_output_layer = nn.Linear(in_features=128, out_features=64)
self.decoder_hidden_layer = nn.Linear(in_features=64, out_features=128)
self.decoder_output_layer = nn.Linear(in_features=128, out_features
=input_shape)
def forward(self, input_0):
primals_1 = self.encoder_hidden_layer.weight
primals_2 = self.encoder_hidden_layer.bias
primals_4 = self.encoder_output_layer.weight
primals_5 = self.encoder_output_layer.bias
primals_6 = self.decoder_hidden_layer.weight
primals_7 = self.decoder_hidden_layer.bias
primals_8 = self.decoder_output_layer.weight
primals_9 = self.decoder_output_layer.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0], output[1]
| Paratra/IoTAnalytics_pub | AE_Net | false | 926 | [
"MIT"
] | 0 | 8c1d02b60ef609c3cba654ce4a5568c39fc63edf | https://github.com/Paratra/IoTAnalytics_pub/tree/8c1d02b60ef609c3cba654ce4a5568c39fc63edf | import torch
from torch.optim import *
from torch import nn
class Model(nn.Module):
"""docstring for AE_Net."""
def __init__(self, input_shape):
super().__init__()
self.encoder_hidden_layer = nn.Linear(in_features=input_shape,
out_features=128)
self.encoder_output_layer = nn.Linear(in_features=128, out_features=64)
self.decoder_hidden_layer = nn.Linear(in_features=64, out_features=128)
self.decoder_output_layer = nn.Linear(in_features=128, out_features
=input_shape)
def forward(self, features):
activation = self.encoder_hidden_layer(features)
activation = torch.relu(activation)
state_logit = self.encoder_output_layer(activation)
code = torch.relu(state_logit)
activation = self.decoder_hidden_layer(code)
activation = torch.relu(activation)
activation = self.decoder_output_layer(activation)
reconstructed = activation
return reconstructed, state_logit
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4]
|
NormedLinear | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/c4/cc4mo2hxcdeykamlqh3edsugtiwa5qgnvsckmhygtvamwp2crsqf.py
# Topologically Sorted Source Nodes: [norm_1, pow_2, add_1, x_, x__1], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div, aten.mul]
# Source node to ATen node mapping:
# add_1 => add_1
# norm_1 => pow_4, pow_5, sum_2
# pow_2 => pow_6
# x_ => div_1
# x__1 => mul
# Graph fragment:
# %pow_4 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_2, 2), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_4, [1], True), kwargs = {})
# %pow_5 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_2, 0.5), kwargs = {})
# %pow_6 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%pow_5, 1.0), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%pow_6, 1e-06), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_2, %add_1), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div_1, 20), kwargs = {})
triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0 = async_compile.triton('triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tmp16 = 20.0
tmp17 = tmp15 * tmp16
tl.store(out_ptr0 + (x3), tmp17, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/gr/cgrujw4uhohajox4hmoi742gbs4qqn5s46hnwwj2orhjgu6tkgrn.py
# Topologically Sorted Source Nodes: [norm, pow_1, add, weight_], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# norm => pow_1, pow_2, sum_1
# pow_1 => pow_3
# weight_ => div
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_1, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [1], True), kwargs = {})
# %pow_2 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_1, 0.5), kwargs = {})
# %pow_3 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%pow_2, 1.0), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%pow_3, 1e-06), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, %add), kwargs = {})
triton_poi_fused_add_div_linalg_vector_norm_pow_1 = async_compile.triton('triton_poi_fused_add_div_linalg_vector_norm_pow_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_linalg_vector_norm_pow_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_pow_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tl.store(out_ptr0 + (x2), tmp15, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [norm_1, pow_2, add_1, x_, x__1], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div, aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0.run(primals_2, buf0, 256, grid=grid(256), stream=stream0)
del primals_2
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [norm, pow_1, add, weight_], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div]
triton_poi_fused_add_div_linalg_vector_norm_pow_1.run(primals_1, buf1, 16, grid=grid(16), stream=stream0)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_3, reinterpret_tensor(buf0, (64, 4), (4, 1), 0), reinterpret_tensor(buf1, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf2)
del buf1
del primals_3
return (reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0), primals_1, reinterpret_tensor(buf0, (64, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn.functional as F
from torch import nn
class NormedLinear(nn.Linear):
"""Normalized Linear Layer.
Args:
tempeature (float, optional): Tempeature term. Default to 20.
power (int, optional): Power term. Default to 1.0.
eps (float, optional): The minimal value of divisor to
keep numerical stability. Default to 1e-6.
"""
def __init__(self, *args, tempearture=20, power=1.0, eps=1e-06, **kwargs):
super(NormedLinear, self).__init__(*args, **kwargs)
self.tempearture = tempearture
self.power = power
self.eps = eps
self.init_weights()
def init_weights(self):
nn.init.normal_(self.weight, mean=0, std=0.01)
if self.bias is not None:
nn.init.constant_(self.bias, 0)
def forward(self, x):
weight_ = self.weight / (self.weight.norm(dim=1, keepdim=True).pow(
self.power) + self.eps)
x_ = x / (x.norm(dim=1, keepdim=True).pow(self.power) + self.eps)
x_ = x_ * self.tempearture
return F.linear(x_, weight_, self.bias)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_features': 4, 'out_features': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0(in_ptr0, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tmp16 = 20.0
tmp17 = tmp15 * tmp16
tl.store(out_ptr0 + x3, tmp17, xmask)
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_pow_1(in_ptr0, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tl.store(out_ptr0 + x2, tmp15, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_linalg_vector_norm_mul_pow_0[grid(256)](
primals_2, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_add_div_linalg_vector_norm_pow_1[grid(16)](primals_1,
buf1, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_3, reinterpret_tensor(buf0, (64, 4), (
4, 1), 0), reinterpret_tensor(buf1, (4, 4), (1, 4), 0), alpha=1,
beta=1, out=buf2)
del buf1
del primals_3
return reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0
), primals_1, reinterpret_tensor(buf0, (64, 4), (4, 1), 0)
class NormedLinearNew(nn.Linear):
"""Normalized Linear Layer.
Args:
tempeature (float, optional): Tempeature term. Default to 20.
power (int, optional): Power term. Default to 1.0.
eps (float, optional): The minimal value of divisor to
keep numerical stability. Default to 1e-6.
"""
def __init__(self, *args, tempearture=20, power=1.0, eps=1e-06, **kwargs):
super(NormedLinearNew, self).__init__(*args, **kwargs)
self.tempearture = tempearture
self.power = power
self.eps = eps
self.init_weights()
def init_weights(self):
nn.init.normal_(self.weight, mean=0, std=0.01)
if self.bias is not None:
nn.init.constant_(self.bias, 0)
def forward(self, input_0):
primals_1 = self.weight
primals_3 = self.bias
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Parskatt/mmdetection | NormedLinear | false | 927 | [
"Apache-2.0"
] | 0 | ee4cfa29e7f479b2454b1f1355f8c05be62d8466 | https://github.com/Parskatt/mmdetection/tree/ee4cfa29e7f479b2454b1f1355f8c05be62d8466 | import torch
import torch.nn.functional as F
from torch import nn
class Model(nn.Linear):
"""Normalized Linear Layer.
Args:
tempeature (float, optional): Tempeature term. Default to 20.
power (int, optional): Power term. Default to 1.0.
eps (float, optional): The minimal value of divisor to
keep numerical stability. Default to 1e-6.
"""
def __init__(self, *args, tempearture=20, power=1.0, eps=1e-06, **kwargs):
super().__init__(*args, **kwargs)
self.tempearture = tempearture
self.power = power
self.eps = eps
self.init_weights()
def init_weights(self):
nn.init.normal_(self.weight, mean=0, std=0.01)
if self.bias is not None:
nn.init.constant_(self.bias, 0)
def forward(self, x):
weight_ = self.weight / (self.weight.norm(dim=1, keepdim=True).pow(
self.power) + self.eps)
x_ = x / (x.norm(dim=1, keepdim=True).pow(self.power) + self.eps)
x_ = x_ * self.tempearture
return F.linear(x_, weight_, self.bias)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4]
|
NormedConv2d | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/qs/cqsb5uykgs6dxtfkaqn2dtogrggqadkzu2etk6n3k2ycbuu3j5xj.py
# Topologically Sorted Source Nodes: [norm, pow_1, add, weight_], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# norm => pow_1, pow_2, sum_1
# pow_1 => pow_3
# weight_ => div
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_1, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [1], True), kwargs = {})
# %pow_2 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_1, 0.5), kwargs = {})
# %pow_3 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%pow_2, 1.0), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%pow_3, 1e-06), kwargs = {})
# %div : [num_users=2] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, %add), kwargs = {})
triton_poi_fused_add_div_linalg_vector_norm_pow_0 = async_compile.triton('triton_poi_fused_add_div_linalg_vector_norm_pow_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_linalg_vector_norm_pow_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_pow_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tl.store(out_ptr0 + (x3), tmp15, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/fb/cfb2mhtqkqxdf4at5lwvqvxh4rnjehaoepwqh6tscuxefrmmf3in.py
# Topologically Sorted Source Nodes: [norm_1, pow_2, add_1, x_, x__1], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div, aten.mul]
# Source node to ATen node mapping:
# add_1 => add_1
# norm_1 => pow_4, pow_5, sum_2
# pow_2 => pow_6
# x_ => div_1
# x__1 => mul
# Graph fragment:
# %pow_4 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_2, 2), kwargs = {})
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_4, [1], True), kwargs = {})
# %pow_5 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sum_2, 0.5), kwargs = {})
# %pow_6 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%pow_5, 1.0), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%pow_6, 1e-06), kwargs = {})
# %div_1 : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_2, %add_1), kwargs = {})
# %mul : [num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div_1, 20), kwargs = {})
triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1 = async_compile.triton('triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = (xindex // 64)
tmp0 = tl.load(in_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + (64*x2)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tmp16 = 20.0
tmp17 = tmp15 * tmp16
tl.store(out_ptr0 + (x3), tmp17, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/k2/ck2mamkqpmuzem4n3p4ij6fmfpy2bcbblg6sx6wwslgqwuqq5ifh.py
# Topologically Sorted Source Nodes: [x__2], Original ATen: [aten.convolution]
# Source node to ATen node mapping:
# x__2 => convolution
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%mul, %div, %primals_3, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
triton_poi_fused_convolution_2 = async_compile.triton('triton_poi_fused_convolution_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[16],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [norm, pow_1, add, weight_], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_linalg_vector_norm_pow_0.run(primals_1, buf0, 256, grid=grid(256), stream=stream0)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [norm_1, pow_2, add_1, x_, x__1], Original ATen: [aten.linalg_vector_norm, aten.pow, aten.add, aten.div, aten.mul]
triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1.run(primals_2, buf1, 256, grid=grid(256), stream=stream0)
del primals_2
# Topologically Sorted Source Nodes: [x__2], Original ATen: [aten.convolution]
buf2 = extern_kernels.convolution(buf1, buf0, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 4, 1, 1), (4, 1, 1, 1))
buf3 = buf2; del buf2 # reuse
# Topologically Sorted Source Nodes: [x__2], Original ATen: [aten.convolution]
triton_poi_fused_convolution_2.run(buf3, primals_3, 16, grid=grid(16), stream=stream0)
del primals_3
return (buf3, primals_1, buf0, buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn
class NormedConv2d(nn.Conv2d):
"""Normalized Conv2d Layer.
Args:
tempeature (float, optional): Tempeature term. Default to 20.
power (int, optional): Power term. Default to 1.0.
eps (float, optional): The minimal value of divisor to
keep numerical stability. Default to 1e-6.
norm_over_kernel (bool, optional): Normalize over kernel.
Default to False.
"""
def __init__(self, *args, tempearture=20, power=1.0, eps=1e-06,
norm_over_kernel=False, **kwargs):
super(NormedConv2d, self).__init__(*args, **kwargs)
self.tempearture = tempearture
self.power = power
self.norm_over_kernel = norm_over_kernel
self.eps = eps
def forward(self, x):
if not self.norm_over_kernel:
weight_ = self.weight / (self.weight.norm(dim=1, keepdim=True).
pow(self.power) + self.eps)
else:
weight_ = self.weight / (self.weight.view(self.weight.size(0),
-1).norm(dim=1, keepdim=True).pow(self.power)[..., None,
None] + self.eps)
x_ = x / (x.norm(dim=1, keepdim=True).pow(self.power) + self.eps)
x_ = x_ * self.tempearture
if hasattr(self, 'conv2d_forward'):
x_ = self.conv2d_forward(x_, weight_)
elif torch.__version__ >= '1.8':
x_ = self._conv_forward(x_, weight_, self.bias)
else:
x_ = self._conv_forward(x_, weight_)
return x_
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'kernel_size': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_pow_0(in_ptr0, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tl.store(out_ptr0 + x3, tmp15, xmask)
@triton.jit
def triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1(in_ptr0, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1e-06
tmp14 = tmp12 + tmp13
tmp15 = tmp0 / tmp14
tmp16 = 20.0
tmp17 = tmp15 * tmp16
tl.store(out_ptr0 + x3, tmp17, xmask)
@triton.jit
def triton_poi_fused_convolution_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_linalg_vector_norm_pow_0[grid(256)](primals_1,
buf0, 256, XBLOCK=256, num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_div_linalg_vector_norm_mul_pow_1[grid(256)](
primals_2, buf1, 256, XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf2 = extern_kernels.convolution(buf1, buf0, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 4, 1, 1), (4, 1, 1, 1))
buf3 = buf2
del buf2
triton_poi_fused_convolution_2[grid(16)](buf3, primals_3, 16,
XBLOCK=16, num_warps=1, num_stages=1)
del primals_3
return buf3, primals_1, buf0, buf1
class NormedConv2dNew(nn.Conv2d):
"""Normalized Conv2d Layer.
Args:
tempeature (float, optional): Tempeature term. Default to 20.
power (int, optional): Power term. Default to 1.0.
eps (float, optional): The minimal value of divisor to
keep numerical stability. Default to 1e-6.
norm_over_kernel (bool, optional): Normalize over kernel.
Default to False.
"""
def __init__(self, *args, tempearture=20, power=1.0, eps=1e-06,
norm_over_kernel=False, **kwargs):
super(NormedConv2dNew, self).__init__(*args, **kwargs)
self.tempearture = tempearture
self.power = power
self.norm_over_kernel = norm_over_kernel
self.eps = eps
def forward(self, input_0):
primals_1 = self.weight
primals_3 = self.bias
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Parskatt/mmdetection | NormedConv2d | false | 928 | [
"Apache-2.0"
] | 0 | ee4cfa29e7f479b2454b1f1355f8c05be62d8466 | https://github.com/Parskatt/mmdetection/tree/ee4cfa29e7f479b2454b1f1355f8c05be62d8466 | import torch
from torch import nn
class Model(nn.Conv2d):
"""Normalized Conv2d Layer.
Args:
tempeature (float, optional): Tempeature term. Default to 20.
power (int, optional): Power term. Default to 1.0.
eps (float, optional): The minimal value of divisor to
keep numerical stability. Default to 1e-6.
norm_over_kernel (bool, optional): Normalize over kernel.
Default to False.
"""
def __init__(self, *args, tempearture=20, power=1.0, eps=1e-06,
norm_over_kernel=False, **kwargs):
super().__init__(*args, **kwargs)
self.tempearture = tempearture
self.power = power
self.norm_over_kernel = norm_over_kernel
self.eps = eps
def forward(self, x):
if not self.norm_over_kernel:
weight_ = self.weight / (self.weight.norm(dim=1, keepdim=True).
pow(self.power) + self.eps)
else:
weight_ = self.weight / (self.weight.view(self.weight.size(0),
-1).norm(dim=1, keepdim=True).pow(self.power)[..., None,
None] + self.eps)
x_ = x / (x.norm(dim=1, keepdim=True).pow(self.power) + self.eps)
x_ = x_ * self.tempearture
if hasattr(self, 'conv2d_forward'):
x_ = self.conv2d_forward(x_, weight_)
elif torch.__version__ >= '1.8':
x_ = self._conv_forward(x_, weight_, self.bias)
else:
x_ = self._conv_forward(x_, weight_)
return x_
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [4, 4, 4]
|
Net | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/zv/czvfpj3ah2lefbwpcuw4esv23bxs5a3ab63ply3ntgbsdktepd5v.py
# Topologically Sorted Source Nodes: [conv2d, relu], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# conv2d => convolution
# relu => relu
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution,), kwargs = {})
triton_poi_fused_convolution_relu_0 = async_compile.triton('triton_poi_fused_convolution_relu_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[32768],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 18816
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 784) % 6
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x3), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/v7/cv7qi7gg3bpfwb3hj7zgy5jlgh7x7wdgqsfsodkjsoverxdjlf6z.py
# Topologically Sorted Source Nodes: [h], Original ATen: [aten.max_pool2d_with_indices]
# Source node to ATen node mapping:
# h => getitem, getitem_1
# Graph fragment:
# %getitem : [num_users=2] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets, 0), kwargs = {})
# %getitem_1 : [num_users=1] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets, 1), kwargs = {})
triton_poi_fused_max_pool2d_with_indices_1 = async_compile.triton('triton_poi_fused_max_pool2d_with_indices_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[8192],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*i8', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_pool2d_with_indices_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_1(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 4704
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 14
x3 = (xindex // 14)
x2 = (xindex // 1176)
x4 = xindex % 1176
tmp0 = tl.load(in_ptr0 + ((2*x0) + (56*x3)), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (2*x0) + (56*x3)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (28 + (2*x0) + (56*x3)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (29 + (2*x0) + (56*x3)), xmask, eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + (x4 + (1184*x2)), tmp6, xmask)
tl.store(out_ptr1 + (x4 + (1280*x2)), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/xe/cxelxvpw3asckozc53rh36773aohp5hqpbp2nos5ymcdqhxvo4bl.py
# Topologically Sorted Source Nodes: [conv2d_1, relu_1], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# conv2d_1 => convolution_1
# relu_1 => relu_1
# Graph fragment:
# %convolution_1 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%getitem, %primals_4, %primals_5, [1, 1], [0, 0], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu_1 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution_1,), kwargs = {})
triton_poi_fused_convolution_relu_2 = async_compile.triton('triton_poi_fused_convolution_relu_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[8192],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_2', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 6400
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = (xindex // 100) % 16
tmp0 = tl.load(in_out_ptr0 + (x3), xmask)
tmp1 = tl.load(in_ptr0 + (x1), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x3), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/tn/ctnw4tbgfy47ppke77vu7rtiz7dl5o3ahickx4p64n7c5rmrrix6.py
# Topologically Sorted Source Nodes: [h_1], Original ATen: [aten.max_pool2d_with_indices]
# Source node to ATen node mapping:
# h_1 => _low_memory_max_pool2d_with_offsets_1, getitem_3
# Graph fragment:
# %_low_memory_max_pool2d_with_offsets_1 : [num_users=2] = call_function[target=torch.ops.prims._low_memory_max_pool2d_with_offsets.default](args = (%relu_1, [2, 2], [2, 2], [0, 0], [1, 1], False), kwargs = {})
# %getitem_3 : [num_users=1] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets_1, 1), kwargs = {})
triton_poi_fused_max_pool2d_with_indices_3 = async_compile.triton('triton_poi_fused_max_pool2d_with_indices_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[2048],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*i8', 2: '*fp32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_pool2d_with_indices_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_3(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 1600
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 5
x1 = (xindex // 5)
x2 = xindex
tmp0 = tl.load(in_ptr0 + ((2*x0) + (20*x1)), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + (2*x0) + (20*x1)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (10 + (2*x0) + (20*x1)), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (11 + (2*x0) + (20*x1)), xmask, eviction_policy='evict_last')
tmp2 = tmp1 > tmp0
tmp3 = tl.full([1], 1, tl.int8)
tmp4 = tl.full([1], 0, tl.int8)
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = triton_helpers.maximum(tmp1, tmp0)
tmp8 = tmp7 > tmp6
tmp9 = tl.full([1], 2, tl.int8)
tmp10 = tl.where(tmp8, tmp9, tmp5)
tmp11 = triton_helpers.maximum(tmp7, tmp6)
tmp13 = tmp12 > tmp11
tmp14 = tl.full([1], 3, tl.int8)
tmp15 = tl.where(tmp13, tmp14, tmp10)
tmp16 = triton_helpers.maximum(tmp12, tmp11)
tl.store(out_ptr0 + (x2), tmp15, xmask)
tl.store(out_ptr1 + (x2), tmp16, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/jn/cjnqv3sgcv5x2iz7ij5zdad6ofabcnonrlksgsxu2ob7n274gz6b.py
# Topologically Sorted Source Nodes: [h_3], Original ATen: [aten.relu]
# Source node to ATen node mapping:
# h_3 => relu_2
# Graph fragment:
# %add_tensor_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default_1, %primals_7), kwargs = {})
# %relu_2 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%add_tensor_1,), kwargs = {})
triton_poi_fused_relu_4 = async_compile.triton('triton_poi_fused_relu_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_4', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_4(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 480
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 120
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/6m/c6m6u2ctjb4r4ra3sizrwezzkzegfp2ombflmfg3dwjfci2pen7h.py
# Topologically Sorted Source Nodes: [h_4], Original ATen: [aten.relu]
# Source node to ATen node mapping:
# h_4 => relu_3
# Graph fragment:
# %add_tensor : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm_default, %primals_9), kwargs = {})
# %relu_3 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%add_tensor,), kwargs = {})
triton_poi_fused_relu_5 = async_compile.triton('triton_poi_fused_relu_5', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_relu_5', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_relu_5(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 336
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 84
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11 = args
args.clear()
assert_size_stride(primals_1, (6, 3, 5, 5), (75, 25, 5, 1))
assert_size_stride(primals_2, (6, ), (1, ))
assert_size_stride(primals_3, (4, 3, 32, 32), (3072, 1024, 32, 1))
assert_size_stride(primals_4, (16, 6, 5, 5), (150, 25, 5, 1))
assert_size_stride(primals_5, (16, ), (1, ))
assert_size_stride(primals_6, (120, 400), (400, 1))
assert_size_stride(primals_7, (120, ), (1, ))
assert_size_stride(primals_8, (84, 120), (120, 1))
assert_size_stride(primals_9, (84, ), (1, ))
assert_size_stride(primals_10, (11, 84), (84, 1))
assert_size_stride(primals_11, (11, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
# Topologically Sorted Source Nodes: [conv2d], Original ATen: [aten.convolution]
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 6, 28, 28), (4704, 784, 28, 1))
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [conv2d, relu], Original ATen: [aten.convolution, aten.relu]
stream0 = get_raw_stream(0)
triton_poi_fused_convolution_relu_0.run(buf1, primals_2, 18816, grid=grid(18816), stream=stream0)
del primals_2
buf2 = empty_strided_cuda((4, 6, 14, 14), (1184, 196, 14, 1), torch.float32)
buf3 = empty_strided_cuda((4, 6, 14, 14), (1280, 196, 14, 1), torch.int8)
# Topologically Sorted Source Nodes: [h], Original ATen: [aten.max_pool2d_with_indices]
triton_poi_fused_max_pool2d_with_indices_1.run(buf1, buf2, buf3, 4704, grid=grid(4704), stream=stream0)
# Topologically Sorted Source Nodes: [conv2d_1], Original ATen: [aten.convolution]
buf4 = extern_kernels.convolution(buf2, primals_4, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (4, 16, 10, 10), (1600, 100, 10, 1))
buf5 = buf4; del buf4 # reuse
# Topologically Sorted Source Nodes: [conv2d_1, relu_1], Original ATen: [aten.convolution, aten.relu]
triton_poi_fused_convolution_relu_2.run(buf5, primals_5, 6400, grid=grid(6400), stream=stream0)
del primals_5
buf6 = empty_strided_cuda((4, 16, 5, 5), (400, 25, 5, 1), torch.int8)
buf7 = empty_strided_cuda((4, 16, 5, 5), (400, 25, 5, 1), torch.float32)
# Topologically Sorted Source Nodes: [h_1], Original ATen: [aten.max_pool2d_with_indices]
triton_poi_fused_max_pool2d_with_indices_3.run(buf5, buf6, buf7, 1600, grid=grid(1600), stream=stream0)
buf8 = empty_strided_cuda((4, 120), (120, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(buf7, (4, 400), (400, 1), 0), reinterpret_tensor(primals_6, (400, 120), (1, 400), 0), out=buf8)
buf9 = buf8; del buf8 # reuse
# Topologically Sorted Source Nodes: [h_3], Original ATen: [aten.relu]
triton_poi_fused_relu_4.run(buf9, primals_7, 480, grid=grid(480), stream=stream0)
del primals_7
buf10 = empty_strided_cuda((4, 84), (84, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(buf9, reinterpret_tensor(primals_8, (120, 84), (1, 120), 0), out=buf10)
buf11 = buf10; del buf10 # reuse
# Topologically Sorted Source Nodes: [h_4], Original ATen: [aten.relu]
triton_poi_fused_relu_5.run(buf11, primals_9, 336, grid=grid(336), stream=stream0)
del primals_9
buf12 = empty_strided_cuda((4, 11), (11, 1), torch.float32)
# Topologically Sorted Source Nodes: [h_5], Original ATen: [aten.addmm]
extern_kernels.addmm(primals_11, buf11, reinterpret_tensor(primals_10, (84, 11), (1, 84), 0), alpha=1, beta=1, out=buf12)
del primals_11
return (buf12, reinterpret_tensor(buf12, (4, 10), (11, 1), 0), reinterpret_tensor(buf12, (4, 1), (11, 1), 10), primals_1, primals_3, primals_4, buf1, buf2, buf3, buf5, buf6, reinterpret_tensor(buf7, (4, 400), (400, 1), 0), buf9, buf11, primals_10, primals_8, primals_6, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((6, 3, 5, 5), (75, 25, 5, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((6, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 3, 32, 32), (3072, 1024, 32, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((16, 6, 5, 5), (150, 25, 5, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((16, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_6 = rand_strided((120, 400), (400, 1), device='cuda:0', dtype=torch.float32)
primals_7 = rand_strided((120, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_8 = rand_strided((84, 120), (120, 1), device='cuda:0', dtype=torch.float32)
primals_9 = rand_strided((84, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_10 = rand_strided((11, 84), (84, 1), device='cuda:0', dtype=torch.float32)
primals_11 = rand_strided((11, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8, primals_9, primals_10, primals_11])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data
class Net(nn.Module):
def __init__(self, num_rej=0):
super(Net, self).__init__()
self.num_rej = num_rej + 1
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10 + self.num_rej)
self.loss = lambda g: -F.logsigmoid(-g)
def forward(self, x):
h = self.pool(F.relu(self.conv1(x)))
h = self.pool(F.relu(self.conv2(h)))
h = h.view(-1, 16 * 5 * 5)
h = F.relu(self.fc1(h))
h = F.relu(self.fc2(h))
h = self.fc3(h)
h1 = h[:, :-self.num_rej]
h2 = h[:, -self.num_rej:]
return h, h1, h2
def pconf(self, hidden, t, loss, rate_all=1.0, rate_inv=0.8):
for i in range(self.num_rej):
if i == 0:
h = hidden[:, i]
loss += torch.mean(self.loss(h) + (1 - rate_all) / rate_all *
self.loss(-h))
else:
h = hidden[:, i]
temp = (t == i).type(torch.FloatTensor)
if type(rate_inv) == float:
temp *= rate_inv
else:
temp *= rate_inv[i]
loss += torch.mean((self.loss(h) + (1 - rate_inv) /
rate_inv * self.loss(-h)) * temp)
return loss
def get_inputs():
return [torch.rand([4, 3, 32, 32])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 18816
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 784 % 6
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_1(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 4704
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 14
x3 = xindex // 14
x2 = xindex // 1176
x4 = xindex % 1176
tmp0 = tl.load(in_ptr0 + (2 * x0 + 56 * x3), xmask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 56 * x3), xmask, eviction_policy
='evict_last')
tmp3 = tl.load(in_ptr0 + (28 + 2 * x0 + 56 * x3), xmask,
eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (29 + 2 * x0 + 56 * x3), xmask,
eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + (x4 + 1184 * x2), tmp6, xmask)
tl.store(out_ptr1 + (x4 + 1280 * x2), tmp16, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 6400
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 100 % 16
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_3(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 1600
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 5
x1 = xindex // 5
x2 = xindex
tmp0 = tl.load(in_ptr0 + (2 * x0 + 20 * x1), xmask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 20 * x1), xmask, eviction_policy
='evict_last')
tmp7 = tl.load(in_ptr0 + (10 + 2 * x0 + 20 * x1), xmask,
eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (11 + 2 * x0 + 20 * x1), xmask,
eviction_policy='evict_last')
tmp2 = tmp1 > tmp0
tmp3 = tl.full([1], 1, tl.int8)
tmp4 = tl.full([1], 0, tl.int8)
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = triton_helpers.maximum(tmp1, tmp0)
tmp8 = tmp7 > tmp6
tmp9 = tl.full([1], 2, tl.int8)
tmp10 = tl.where(tmp8, tmp9, tmp5)
tmp11 = triton_helpers.maximum(tmp7, tmp6)
tmp13 = tmp12 > tmp11
tmp14 = tl.full([1], 3, tl.int8)
tmp15 = tl.where(tmp13, tmp14, tmp10)
tmp16 = triton_helpers.maximum(tmp12, tmp11)
tl.store(out_ptr0 + x2, tmp15, xmask)
tl.store(out_ptr1 + x2, tmp16, xmask)
@triton.jit
def triton_poi_fused_relu_4(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 480
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 120
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_relu_5(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 336
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 84
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11) = args
args.clear()
assert_size_stride(primals_1, (6, 3, 5, 5), (75, 25, 5, 1))
assert_size_stride(primals_2, (6,), (1,))
assert_size_stride(primals_3, (4, 3, 32, 32), (3072, 1024, 32, 1))
assert_size_stride(primals_4, (16, 6, 5, 5), (150, 25, 5, 1))
assert_size_stride(primals_5, (16,), (1,))
assert_size_stride(primals_6, (120, 400), (400, 1))
assert_size_stride(primals_7, (120,), (1,))
assert_size_stride(primals_8, (84, 120), (120, 1))
assert_size_stride(primals_9, (84,), (1,))
assert_size_stride(primals_10, (11, 84), (84, 1))
assert_size_stride(primals_11, (11,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 6, 28, 28), (4704, 784, 28, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_relu_0[grid(18816)](buf1, primals_2,
18816, XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((4, 6, 14, 14), (1184, 196, 14, 1), torch
.float32)
buf3 = empty_strided_cuda((4, 6, 14, 14), (1280, 196, 14, 1), torch
.int8)
triton_poi_fused_max_pool2d_with_indices_1[grid(4704)](buf1, buf2,
buf3, 4704, XBLOCK=128, num_warps=4, num_stages=1)
buf4 = extern_kernels.convolution(buf2, primals_4, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (4, 16, 10, 10), (1600, 100, 10, 1))
buf5 = buf4
del buf4
triton_poi_fused_convolution_relu_2[grid(6400)](buf5, primals_5,
6400, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf6 = empty_strided_cuda((4, 16, 5, 5), (400, 25, 5, 1), torch.int8)
buf7 = empty_strided_cuda((4, 16, 5, 5), (400, 25, 5, 1), torch.float32
)
triton_poi_fused_max_pool2d_with_indices_3[grid(1600)](buf5, buf6,
buf7, 1600, XBLOCK=256, num_warps=4, num_stages=1)
buf8 = empty_strided_cuda((4, 120), (120, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf7, (4, 400), (400, 1), 0),
reinterpret_tensor(primals_6, (400, 120), (1, 400), 0), out=buf8)
buf9 = buf8
del buf8
triton_poi_fused_relu_4[grid(480)](buf9, primals_7, 480, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_7
buf10 = empty_strided_cuda((4, 84), (84, 1), torch.float32)
extern_kernels.mm(buf9, reinterpret_tensor(primals_8, (120, 84), (1,
120), 0), out=buf10)
buf11 = buf10
del buf10
triton_poi_fused_relu_5[grid(336)](buf11, primals_9, 336, XBLOCK=
128, num_warps=4, num_stages=1)
del primals_9
buf12 = empty_strided_cuda((4, 11), (11, 1), torch.float32)
extern_kernels.addmm(primals_11, buf11, reinterpret_tensor(
primals_10, (84, 11), (1, 84), 0), alpha=1, beta=1, out=buf12)
del primals_11
return (buf12, reinterpret_tensor(buf12, (4, 10), (11, 1), 0),
reinterpret_tensor(buf12, (4, 1), (11, 1), 10), primals_1,
primals_3, primals_4, buf1, buf2, buf3, buf5, buf6,
reinterpret_tensor(buf7, (4, 400), (400, 1), 0), buf9, buf11,
primals_10, primals_8, primals_6)
class NetNew(nn.Module):
def __init__(self, num_rej=0):
super(NetNew, self).__init__()
self.num_rej = num_rej + 1
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10 + self.num_rej)
self.loss = lambda g: -F.logsigmoid(-g)
def pconf(self, hidden, t, loss, rate_all=1.0, rate_inv=0.8):
for i in range(self.num_rej):
if i == 0:
h = hidden[:, i]
loss += torch.mean(self.loss(h) + (1 - rate_all) / rate_all *
self.loss(-h))
else:
h = hidden[:, i]
temp = (t == i).type(torch.FloatTensor)
if type(rate_inv) == float:
temp *= rate_inv
else:
temp *= rate_inv[i]
loss += torch.mean((self.loss(h) + (1 - rate_inv) /
rate_inv * self.loss(-h)) * temp)
return loss
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_6 = self.fc1.weight
primals_7 = self.fc1.bias
primals_8 = self.fc2.weight
primals_9 = self.fc2.bias
primals_10 = self.fc3.weight
primals_11 = self.fc3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11])
return output[0], output[1], output[2]
| NlGG/Rejection | Net | false | 929 | [
"MIT"
] | 0 | 5f7cc64b71dacc2eb794b3f7c48390457e363cc5 | https://github.com/NlGG/Rejection/tree/5f7cc64b71dacc2eb794b3f7c48390457e363cc5 | import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data
class Model(nn.Module):
def __init__(self, num_rej=0):
super().__init__()
self.num_rej = num_rej + 1
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10 + self.num_rej)
self.loss = lambda g: -F.logsigmoid(-g)
def forward(self, x):
h = self.pool(F.relu(self.conv1(x)))
h = self.pool(F.relu(self.conv2(h)))
h = h.view(-1, 16 * 5 * 5)
h = F.relu(self.fc1(h))
h = F.relu(self.fc2(h))
h = self.fc3(h)
h1 = h[:, :-self.num_rej]
h2 = h[:, -self.num_rej:]
return h, h1, h2
def pconf(self, hidden, t, loss, rate_all=1.0, rate_inv=0.8):
for i in range(self.num_rej):
if i == 0:
h = hidden[:, i]
loss += torch.mean(self.loss(h) + (1 - rate_all) / rate_all *
self.loss(-h))
else:
h = hidden[:, i]
temp = (t == i).type(torch.FloatTensor)
if type(rate_inv) == float:
temp *= rate_inv
else:
temp *= rate_inv[i]
loss += torch.mean((self.loss(h) + (1 - rate_inv) /
rate_inv * self.loss(-h)) * temp)
return loss
def get_inputs():
return [torch.rand([4, 3, 32, 32])]
def get_init_inputs():
return []
|
BPMLLLoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/im/cimboyc56ame5x37npkmkfw2xve7rvkkasb2ozskd2qtmm4jkmoq.py
# Topologically Sorted Source Nodes: [neg, y_bar], Original ATen: [aten.neg, aten.add]
# Source node to ATen node mapping:
# neg => neg
# y_bar => add
# Graph fragment:
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%arg1_1,), kwargs = {})
# %add : [num_users=2] = call_function[target=torch.ops.aten.add.Tensor](args = (%neg, 1), kwargs = {})
triton_poi_fused_add_neg_0 = async_compile.triton('triton_poi_fused_add_neg_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_neg_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_neg_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = -tmp0
tmp2 = 1.0
tmp3 = tmp1 + tmp2
tl.store(out_ptr0 + (x0), tmp3, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/26/c26z3t2s2o7fbb3srtzysccunz5ychms4ztfx6xensu3es5gjxgt.py
# Topologically Sorted Source Nodes: [sum_2], Original ATen: [aten.sum]
# Source node to ATen node mapping:
# sum_2 => sum_2
# Graph fragment:
# %sum_2 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%add, [1]), kwargs = {})
triton_poi_fused_sum_1 = async_compile.triton('triton_poi_fused_sum_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sum_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_sum_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = (xindex // 16)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (64*x1)), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + (64*x1)), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + (64*x1)), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + (64*x1)), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tl.store(out_ptr0 + (x2), tmp6, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/n7/cn7k7pwrs5ncxfptoddu6p5xxefbnneyomtuq6vuyowcr6krxgln.py
# Topologically Sorted Source Nodes: [c], Original ATen: [aten.sigmoid]
# Source node to ATen node mapping:
# c => sigmoid
# Graph fragment:
# %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%arg0_1,), kwargs = {})
triton_poi_fused_sigmoid_2 = async_compile.triton('triton_poi_fused_sigmoid_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sigmoid_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_sigmoid_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), xmask)
tmp1 = tl.sigmoid(tmp0)
tl.store(out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/k4/ck4um4cmkwovrlgbdy2tsszyj7xlrduydpwxmf4nqecx35nrgqjy.py
# Topologically Sorted Source Nodes: [sum_1, ne, all_1], Original ATen: [aten.sum, aten.ne, aten.all]
# Source node to ATen node mapping:
# all_1 => any_1, logical_not, logical_not_1
# ne => ne
# sum_1 => sum_1
# Graph fragment:
# %sum_1 : [num_users=2] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%arg1_1, [1]), kwargs = {})
# %ne : [num_users=1] = call_function[target=torch.ops.aten.ne.Scalar](args = (%sum_1, 0), kwargs = {})
# %logical_not : [num_users=1] = call_function[target=torch.ops.aten.logical_not.default](args = (%ne,), kwargs = {})
# %any_1 : [num_users=1] = call_function[target=torch.ops.aten.any.dims](args = (%logical_not,), kwargs = {})
# %logical_not_1 : [num_users=1] = call_function[target=torch.ops.aten.logical_not.default](args = (%any_1,), kwargs = {})
triton_per_fused_all_ne_sum_3 = async_compile.triton('triton_per_fused_all_ne_sum_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 64],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*i1', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_all_ne_sum_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_all_ne_sum_3(in_out_ptr0, in_ptr0, out_ptr0, xnumel, rnumel, XBLOCK : tl.constexpr):
xnumel = 1
rnumel = 64
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
roffset = 0
rmask = tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = (rindex // 16)
r2 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + (64*r1)), None)
tmp1 = tl.load(in_ptr0 + (16 + r0 + (64*r1)), None)
tmp3 = tl.load(in_ptr0 + (32 + r0 + (64*r1)), None)
tmp5 = tl.load(in_ptr0 + (48 + r0 + (64*r1)), None)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 0.0
tmp8 = tmp6 != tmp7
tmp9 = tmp8 == 0
tmp10 = tl.broadcast_to(tmp9, [XBLOCK, RBLOCK])
tmp12 = triton_helpers.any(tmp10, 1)[:, None]
tmp13 = tmp12 == 0
tl.store(out_ptr0 + (tl.broadcast_to(r2, [XBLOCK, RBLOCK])), tmp6, None)
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([XBLOCK, 1], 0, tl.int32)), tmp13, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [neg, y_bar], Original ATen: [aten.neg, aten.add]
stream0 = get_raw_stream(0)
triton_poi_fused_add_neg_0.run(arg1_1, buf0, 256, grid=grid(256), stream=stream0)
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sum_2], Original ATen: [aten.sum]
triton_poi_fused_sum_1.run(buf0, buf1, 64, grid=grid(64), stream=stream0)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [c], Original ATen: [aten.sigmoid]
triton_poi_fused_sigmoid_2.run(arg0_1, buf2, 256, grid=grid(256), stream=stream0)
del arg0_1
buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf4 = empty_strided_cuda((), (), torch.bool)
buf5 = buf4; del buf4 # reuse
# Topologically Sorted Source Nodes: [sum_1, ne, all_1], Original ATen: [aten.sum, aten.ne, aten.all]
triton_per_fused_all_ne_sum_3.run(buf5, arg1_1, buf3, 1, 64, grid=grid(1), stream=stream0)
return (buf1, buf3, buf0, arg1_1, buf2, buf5, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import Tensor
class BPMLLLoss(torch.nn.Module):
def __init__(self, bias=(1, 1)):
super(BPMLLLoss, self).__init__()
self.bias = bias
assert len(self.bias) == 2 and all(map(lambda x: isinstance(x, int) and
x > 0, bias)), 'bias must be positive integers'
def forward(self, c: 'Tensor', y: 'Tensor') ->Tensor:
"""
compute the loss, which has the form:
L = \\sum_{i=1}^{m} \\frac{1}{|Y_i| \\cdot |\\bar{Y}_i|} \\sum_{(k, l) \\in Y_i \\times \\bar{Y}_i} \\exp{-c^i_k+c^i_l}
:param c: prediction tensor, size: batch_size * n_labels
:param y: target tensor, size: batch_size * n_labels
:return: size: scalar tensor
"""
c = torch.sigmoid(c)
y = y.float()
y_bar = -y + 1
y_norm = torch.pow(y.sum(dim=(1,)), self.bias[0])
y_bar_norm = torch.pow(y_bar.sum(dim=(1,)), self.bias[1])
assert torch.all(y_norm != 0) or torch.all(y_bar_norm != 0
), 'an instance cannot have none or all the labels'
return torch.mean(1 / torch.mul(y_norm, y_bar_norm) * self.
pairwise_sub_exp(y, y_bar, c))
def pairwise_sub_exp(self, y: 'Tensor', y_bar: 'Tensor', c: 'Tensor'
) ->Tensor:
"""
compute \\sum_{(k, l) \\in Y_i \\times \\bar{Y}_i} \\exp{-c^i_k+c^i_l}
"""
truth_matrix = y.unsqueeze(2).float() @ y_bar.unsqueeze(1).float()
exp_matrix = torch.exp(c.unsqueeze(1) - c.unsqueeze(2))
return torch.mul(truth_matrix, exp_matrix).sum(dim=(1, 2))
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch import Tensor
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_neg_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = -tmp0
tmp2 = 1.0
tmp3 = tmp1 + tmp2
tl.store(out_ptr0 + x0, tmp3, xmask)
@triton.jit
def triton_poi_fused_sum_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tl.store(out_ptr0 + x2, tmp6, xmask)
@triton.jit
def triton_poi_fused_sigmoid_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.sigmoid(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_per_fused_all_ne_sum_3(in_out_ptr0, in_ptr0, out_ptr0, xnumel,
rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex % 16
r1 = rindex // 16
r2 = rindex
tmp0 = tl.load(in_ptr0 + (r0 + 64 * r1), None)
tmp1 = tl.load(in_ptr0 + (16 + r0 + 64 * r1), None)
tmp3 = tl.load(in_ptr0 + (32 + r0 + 64 * r1), None)
tmp5 = tl.load(in_ptr0 + (48 + r0 + 64 * r1), None)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 0.0
tmp8 = tmp6 != tmp7
tmp9 = tmp8 == 0
tmp10 = tl.broadcast_to(tmp9, [XBLOCK, RBLOCK])
tmp12 = triton_helpers.any(tmp10, 1)[:, None]
tmp13 = tmp12 == 0
tl.store(out_ptr0 + tl.broadcast_to(r2, [XBLOCK, RBLOCK]), tmp6, None)
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp13, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_neg_0[grid(256)](arg1_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_sum_1[grid(64)](buf0, buf1, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_sigmoid_2[grid(256)](arg0_1, buf2, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf4 = empty_strided_cuda((), (), torch.bool)
buf5 = buf4
del buf4
triton_per_fused_all_ne_sum_3[grid(1)](buf5, arg1_1, buf3, 1, 64,
XBLOCK=1, num_warps=2, num_stages=1)
return buf1, buf3, buf0, arg1_1, buf2, buf5
class BPMLLLossNew(torch.nn.Module):
def __init__(self, bias=(1, 1)):
super(BPMLLLossNew, self).__init__()
self.bias = bias
assert len(self.bias) == 2 and all(map(lambda x: isinstance(x, int) and
x > 0, bias)), 'bias must be positive integers'
def pairwise_sub_exp(self, y: 'Tensor', y_bar: 'Tensor', c: 'Tensor'
) ->Tensor:
"""
compute \\sum_{(k, l) \\in Y_i \\times \\bar{Y}_i} \\exp{-c^i_k+c^i_l}
"""
truth_matrix = y.unsqueeze(2).float() @ y_bar.unsqueeze(1).float()
exp_matrix = torch.exp(c.unsqueeze(1) - c.unsqueeze(2))
return torch.mul(truth_matrix, exp_matrix).sum(dim=(1, 2))
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| PaulStryck/nih-chest-x-ray | BPMLLLoss | false | 931 | [
"MIT"
] | 0 | 92bfef80f49e7e38ad8a0156be43f10879b3f737 | https://github.com/PaulStryck/nih-chest-x-ray/tree/92bfef80f49e7e38ad8a0156be43f10879b3f737 | import torch
from torch import Tensor
class Model(torch.nn.Module):
def __init__(self, bias=(1, 1)):
super().__init__()
self.bias = bias
assert len(self.bias) == 2 and all(map(lambda x: isinstance(x, int) and
x > 0, bias)), 'bias must be positive integers'
def forward(self, c: 'Tensor', y: 'Tensor') ->Tensor:
"""
compute the loss, which has the form:
L = \\sum_{i=1}^{m} \\frac{1}{|Y_i| \\cdot |\\bar{Y}_i|} \\sum_{(k, l) \\in Y_i \\times \\bar{Y}_i} \\exp{-c^i_k+c^i_l}
:param c: prediction tensor, size: batch_size * n_labels
:param y: target tensor, size: batch_size * n_labels
:return: size: scalar tensor
"""
c = torch.sigmoid(c)
y = y.float()
y_bar = -y + 1
y_norm = torch.pow(y.sum(dim=(1,)), self.bias[0])
y_bar_norm = torch.pow(y_bar.sum(dim=(1,)), self.bias[1])
assert torch.all(y_norm != 0) or torch.all(y_bar_norm != 0
), 'an instance cannot have none or all the labels'
return torch.mean(1 / torch.mul(y_norm, y_bar_norm) * self.
pairwise_sub_exp(y, y_bar, c))
def pairwise_sub_exp(self, y: 'Tensor', y_bar: 'Tensor', c: 'Tensor'
) ->Tensor:
"""
compute \\sum_{(k, l) \\in Y_i \\times \\bar{Y}_i} \\exp{-c^i_k+c^i_l}
"""
truth_matrix = y.unsqueeze(2).float() @ y_bar.unsqueeze(1).float()
exp_matrix = torch.exp(c.unsqueeze(1) - c.unsqueeze(2))
return torch.mul(truth_matrix, exp_matrix).sum(dim=(1, 2))
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
FirstNet | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/rn/crng7m5mguccwv3xvtgv4yl47k24ov5e26h7ejsq2geg3uuvz5og.py
# Unsorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
triton_poi_fused_0 = async_compile.triton('triton_poi_fused_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[8192, 16], tile_hint=TileHint.SQUARE,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 8192
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = (yindex // 64)
tmp0 = tl.load(in_ptr0 + (x2 + (9*y3)), xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + (64*x2) + (576*y1)), tmp0, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/yv/cyvnlfdfywtr376ic26vmyul36c3pelngrtpsxdvhbavnd7igj7j.py
# Topologically Sorted Source Nodes: [x, x_1], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# x => convolution
# x_1 => relu
# Graph fragment:
# %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%primals_3, %primals_1, %primals_2, [1, 1], [1, 1], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution,), kwargs = {})
triton_poi_fused_convolution_relu_1 = async_compile.triton('triton_poi_fused_convolution_relu_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256, 4096], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 256
xnumel = 4096
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = (yindex // 64)
tmp0 = tl.load(in_ptr0 + (x2 + (4096*y3)), ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (y0), ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(out_ptr0 + (y0 + (64*x2) + (262144*y1)), tmp4, ymask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ew/cewxxvrvhc6zilpxrflpmda4gd33px7lnuquxldeeixilzlxc7vb.py
# Topologically Sorted Source Nodes: [x_2], Original ATen: [aten.max_pool2d_with_indices]
# Source node to ATen node mapping:
# x_2 => getitem, getitem_1
# Graph fragment:
# %getitem : [num_users=2] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets, 0), kwargs = {})
# %getitem_1 : [num_users=1] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets, 1), kwargs = {})
triton_poi_fused_max_pool2d_with_indices_2 = async_compile.triton('triton_poi_fused_max_pool2d_with_indices_2', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[262144],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*i8', 3: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_pool2d_with_indices_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_2(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
xnumel = 262144
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 64
x1 = (xindex // 64) % 32
x2 = (xindex // 2048)
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (128*x1) + (8192*x2)), None)
tmp1 = tl.load(in_ptr0 + (64 + x0 + (128*x1) + (8192*x2)), None)
tmp3 = tl.load(in_ptr0 + (4096 + x0 + (128*x1) + (8192*x2)), None)
tmp5 = tl.load(in_ptr0 + (4160 + x0 + (128*x1) + (8192*x2)), None)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + (x3), tmp6, None)
tl.store(out_ptr1 + (x3), tmp16, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/yi/cyidhxg2d6xrmtctl52zmpussqimohbzbcjj4eqnhj4gvukrl5pp.py
# Topologically Sorted Source Nodes: [x_3, x_4], Original ATen: [aten.convolution, aten.relu]
# Source node to ATen node mapping:
# x_3 => convolution_1
# x_4 => relu_1
# Graph fragment:
# %convolution_1 : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%getitem, %primals_4, %primals_5, [1, 1], [1, 1], [1, 1], False, [0, 0], 1), kwargs = {})
# %relu_1 : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%convolution_1,), kwargs = {})
triton_poi_fused_convolution_relu_3 = async_compile.triton('triton_poi_fused_convolution_relu_3', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[524288],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_convolution_relu_3', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_convolution_relu_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 524288
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + (x2), None)
tmp1 = tl.load(in_ptr0 + (x0), None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + (x2), tmp4, None)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/ig/cigpj5h5jh2tj3hhtkru242ebvodusam2qljytm5vrcpm336toz6.py
# Topologically Sorted Source Nodes: [x_5], Original ATen: [aten.max_pool2d_with_indices]
# Source node to ATen node mapping:
# x_5 => getitem_2, getitem_3
# Graph fragment:
# %getitem_2 : [num_users=1] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets_1, 0), kwargs = {})
# %getitem_3 : [num_users=1] = call_function[target=operator.getitem](args = (%_low_memory_max_pool2d_with_offsets_1, 1), kwargs = {})
triton_poi_fused_max_pool2d_with_indices_4 = async_compile.triton('triton_poi_fused_max_pool2d_with_indices_4', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[1024, 128], tile_hint=TileHint.DEFAULT,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*i8', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_pool2d_with_indices_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_4(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
ynumel = 1024
xnumel = 128
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x3 = xindex
y0 = yindex % 16
y4 = (yindex // 16)
y2 = (yindex // 256)
y5 = yindex % 256
y6 = yindex
tmp0 = tl.load(in_ptr0 + (x3 + (256*y0) + (8192*y4)), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (128 + x3 + (256*y0) + (8192*y4)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (4096 + x3 + (256*y0) + (8192*y4)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (4224 + x3 + (256*y0) + (8192*y4)), xmask, eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1, 1], 1, tl.int8)
tmp9 = tl.full([1, 1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1, 1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1, 1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + (y5 + (256*x3) + (32768*y2)), tmp6, xmask)
tl.store(out_ptr1 + (x3 + (128*y6)), tmp16, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (64, 1, 3, 3), (9, 9, 3, 1))
assert_size_stride(primals_2, (64, ), (1, ))
assert_size_stride(primals_3, (4, 1, 64, 64), (4096, 4096, 64, 1))
assert_size_stride(primals_4, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_5, (128, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((128, 64, 3, 3), (576, 1, 192, 64), torch.float32)
# Unsorted Source Nodes: [], Original ATen: []
stream0 = get_raw_stream(0)
triton_poi_fused_0.run(primals_4, buf0, 8192, 9, grid=grid(8192, 9), stream=stream0)
del primals_4
# Topologically Sorted Source Nodes: [x], Original ATen: [aten.convolution]
buf1 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 64, 64, 64), (262144, 4096, 64, 1))
buf2 = empty_strided_cuda((4, 64, 64, 64), (262144, 1, 4096, 64), torch.float32)
# Topologically Sorted Source Nodes: [x, x_1], Original ATen: [aten.convolution, aten.relu]
triton_poi_fused_convolution_relu_1.run(buf1, primals_2, buf2, 256, 4096, grid=grid(256, 4096), stream=stream0)
del buf1
del primals_2
buf3 = empty_strided_cuda((4, 64, 32, 32), (65536, 1, 2048, 64), torch.float32)
buf4 = empty_strided_cuda((4, 64, 32, 32), (65536, 1, 2048, 64), torch.int8)
# Topologically Sorted Source Nodes: [x_2], Original ATen: [aten.max_pool2d_with_indices]
triton_poi_fused_max_pool2d_with_indices_2.run(buf2, buf3, buf4, 262144, grid=grid(262144), stream=stream0)
# Topologically Sorted Source Nodes: [x_3], Original ATen: [aten.convolution]
buf5 = extern_kernels.convolution(buf3, buf0, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf5, (4, 128, 32, 32), (131072, 1, 4096, 128))
buf6 = buf5; del buf5 # reuse
# Topologically Sorted Source Nodes: [x_3, x_4], Original ATen: [aten.convolution, aten.relu]
triton_poi_fused_convolution_relu_3.run(buf6, primals_5, 524288, grid=grid(524288), stream=stream0)
del primals_5
buf7 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1), torch.float32)
buf8 = empty_strided_cuda((4, 128, 16, 16), (32768, 1, 2048, 128), torch.int8)
# Topologically Sorted Source Nodes: [x_5], Original ATen: [aten.max_pool2d_with_indices]
triton_poi_fused_max_pool2d_with_indices_4.run(buf6, buf7, buf8, 1024, 128, grid=grid(1024, 128), stream=stream0)
return (buf7, primals_1, primals_3, buf0, buf2, buf3, buf4, buf6, buf8, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((64, 1, 3, 3), (9, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((64, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 1, 64, 64), (4096, 4096, 64, 1), device='cuda:0', dtype=torch.float32)
primals_4 = rand_strided((128, 64, 3, 3), (576, 9, 3, 1), device='cuda:0', dtype=torch.float32)
primals_5 = rand_strided((128, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3, primals_4, primals_5])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
import torch.nn.functional as F
class FirstNet(nn.Module):
def __init__(self):
super(FirstNet, self).__init__()
self.conv1 = nn.Conv2d(in_channels=1, out_channels=64, kernel_size=
3, padding=1, stride=1)
self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
def forward(self, x):
x = self.conv1(x)
x = F.relu(x)
x = F.max_pool2d(x, 2, 2)
x = self.conv2(x)
x = F.relu(x)
x = F.max_pool2d(x, 2, 2)
return x
def get_inputs():
return [torch.rand([4, 1, 64, 64])]
def get_init_inputs():
return [[], {}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 64 * x2 + 576 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_1(in_ptr0, in_ptr1, out_ptr0, ynumel,
xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 256
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 4096 * y3), ymask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(out_ptr0 + (y0 + 64 * x2 + 262144 * y1), tmp4, ymask)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_2(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 64
x1 = xindex // 64 % 32
x2 = xindex // 2048
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 128 * x1 + 8192 * x2), None)
tmp1 = tl.load(in_ptr0 + (64 + x0 + 128 * x1 + 8192 * x2), None)
tmp3 = tl.load(in_ptr0 + (4096 + x0 + 128 * x1 + 8192 * x2), None)
tmp5 = tl.load(in_ptr0 + (4160 + x0 + 128 * x1 + 8192 * x2), None)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + x3, tmp6, None)
tl.store(out_ptr1 + x3, tmp16, None)
@triton.jit
def triton_poi_fused_convolution_relu_3(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_4(in_ptr0, out_ptr0, out_ptr1,
ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
xnumel = 128
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x3 = xindex
y0 = yindex % 16
y4 = yindex // 16
y2 = yindex // 256
y5 = yindex % 256
y6 = yindex
tmp0 = tl.load(in_ptr0 + (x3 + 256 * y0 + 8192 * y4), xmask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (128 + x3 + 256 * y0 + 8192 * y4), xmask,
eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (4096 + x3 + 256 * y0 + 8192 * y4), xmask,
eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (4224 + x3 + 256 * y0 + 8192 * y4), xmask,
eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1, 1], 1, tl.int8)
tmp9 = tl.full([1, 1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1, 1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1, 1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + (y5 + 256 * x3 + 32768 * y2), tmp6, xmask)
tl.store(out_ptr1 + (x3 + 128 * y6), tmp16, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (64, 1, 3, 3), (9, 9, 3, 1))
assert_size_stride(primals_2, (64,), (1,))
assert_size_stride(primals_3, (4, 1, 64, 64), (4096, 4096, 64, 1))
assert_size_stride(primals_4, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_5, (128,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((128, 64, 3, 3), (576, 1, 192, 64), torch
.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(8192, 9)](primals_4, buf0, 8192, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_4
buf1 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 64, 64, 64), (262144, 4096, 64, 1))
buf2 = empty_strided_cuda((4, 64, 64, 64), (262144, 1, 4096, 64),
torch.float32)
triton_poi_fused_convolution_relu_1[grid(256, 4096)](buf1,
primals_2, buf2, 256, 4096, XBLOCK=32, YBLOCK=32, num_warps=4,
num_stages=1)
del buf1
del primals_2
buf3 = empty_strided_cuda((4, 64, 32, 32), (65536, 1, 2048, 64),
torch.float32)
buf4 = empty_strided_cuda((4, 64, 32, 32), (65536, 1, 2048, 64),
torch.int8)
triton_poi_fused_max_pool2d_with_indices_2[grid(262144)](buf2, buf3,
buf4, 262144, XBLOCK=512, num_warps=8, num_stages=1)
buf5 = extern_kernels.convolution(buf3, buf0, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf5, (4, 128, 32, 32), (131072, 1, 4096, 128))
buf6 = buf5
del buf5
triton_poi_fused_convolution_relu_3[grid(524288)](buf6, primals_5,
524288, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_5
buf7 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1),
torch.float32)
buf8 = empty_strided_cuda((4, 128, 16, 16), (32768, 1, 2048, 128),
torch.int8)
triton_poi_fused_max_pool2d_with_indices_4[grid(1024, 128)](buf6,
buf7, buf8, 1024, 128, XBLOCK=32, YBLOCK=32, num_warps=4,
num_stages=1)
return buf7, primals_1, primals_3, buf0, buf2, buf3, buf4, buf6, buf8
class FirstNetNew(nn.Module):
def __init__(self):
super(FirstNetNew, self).__init__()
self.conv1 = nn.Conv2d(in_channels=1, out_channels=64, kernel_size=
3, padding=1, stride=1)
self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
| PacktPublishing/Designing-Models-for-Responsible-AI | FirstNet | false | 932 | [
"MIT"
] | 0 | 36b60f1e3e9db8b3d2db3ace873dbdee1b076b74 | https://github.com/PacktPublishing/Designing-Models-for-Responsible-AI/tree/36b60f1e3e9db8b3d2db3ace873dbdee1b076b74 | import torch
import torch.nn as nn
import torch.nn.functional as F
class Model(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(in_channels=1, out_channels=64, kernel_size=
3, padding=1, stride=1)
self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
def forward(self, x):
x = self.conv1(x)
x = F.relu(x)
x = F.max_pool2d(x, 2, 2)
x = self.conv2(x)
x = F.relu(x)
x = F.max_pool2d(x, 2, 2)
return x
def get_inputs():
return [torch.rand([4, 1, 64, 64])]
def get_init_inputs():
return []
|
AsymmetricLoss | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/rs/crsoklhylfs7zp45n7ulxgjy4g42ckalikfyb6hqsywsj7mfpf6h.py
# Topologically Sorted Source Nodes: [x_sigmoid, clamp_1, log, los_pos, sub_1, xs_neg, add, xs_neg_1, clamp_2, log_1, los_neg, loss, pt0, sub_2, pt1, pt, sub_4, mul_4, sub_3, mul_5, one_sided_gamma, one_sided_w, loss_1, sum_1, neg], Original ATen: [aten.sigmoid, aten.clamp, aten.log, aten.mul, aten.rsub, aten.add, aten.pow, aten.sum, aten.neg]
# Source node to ATen node mapping:
# add => add
# clamp_1 => clamp_min
# clamp_2 => clamp_min_1
# log => log
# log_1 => log_1
# los_neg => mul_1
# los_pos => mul
# loss => add_1
# loss_1 => mul_6
# mul_4 => mul_4
# mul_5 => mul_5
# neg => neg
# one_sided_gamma => add_3
# one_sided_w => pow_1
# pt => add_2
# pt0 => mul_2
# pt1 => mul_3
# sub_1 => sub_1
# sub_2 => sub_2
# sub_3 => sub_3
# sub_4 => sub_4
# sum_1 => sum_1
# x_sigmoid => sigmoid
# xs_neg => sub
# xs_neg_1 => clamp_max
# Graph fragment:
# %sigmoid : [num_users=3] = call_function[target=torch.ops.aten.sigmoid.default](args = (%arg0_1,), kwargs = {})
# %clamp_min : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%sigmoid, 1e-08), kwargs = {})
# %log : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_min,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg1_1, %log), kwargs = {})
# %sub_1 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %arg1_1), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %sigmoid), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sub, 0.05), kwargs = {})
# %clamp_max : [num_users=2] = call_function[target=torch.ops.aten.clamp_max.default](args = (%add, 1), kwargs = {})
# %clamp_min_1 : [num_users=1] = call_function[target=torch.ops.aten.clamp_min.default](args = (%clamp_max, 1e-08), kwargs = {})
# %log_1 : [num_users=1] = call_function[target=torch.ops.aten.log.default](args = (%clamp_min_1,), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_1, %log_1), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %mul_1), kwargs = {})
# %mul_2 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sigmoid, %arg1_1), kwargs = {})
# %sub_2 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %arg1_1), kwargs = {})
# %mul_3 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%clamp_max, %sub_2), kwargs = {})
# %add_2 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_2, %mul_3), kwargs = {})
# %sub_4 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %add_2), kwargs = {})
# %mul_4 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%arg1_1, 1), kwargs = {})
# %sub_3 : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (1, %arg1_1), kwargs = {})
# %mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub_3, 4), kwargs = {})
# %add_3 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_4, %mul_5), kwargs = {})
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Tensor](args = (%sub_4, %add_3), kwargs = {})
# %mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%add_1, %pow_1), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.default](args = (%mul_6,), kwargs = {})
# %neg : [num_users=1] = call_function[target=torch.ops.aten.neg.default](args = (%sum_1,), kwargs = {})
triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0 = async_compile.triton('triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.persistent_reduction(
size_hints=[1, 256],
reduction_hint=ReductionHint.INNER,
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {3: 1}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=(3,))]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': True, 'num_load': 2, 'num_reduction': 1, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel):
xnumel = 1
XBLOCK: tl.constexpr = 1
rnumel = 256
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
xmask = tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
roffset = 0
rmask = tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + (r0), None)
tmp1 = tl.load(in_ptr1 + (r0), None)
tmp2 = tl.sigmoid(tmp1)
tmp3 = 1e-08
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp5 = tl_math.log(tmp4)
tmp6 = tmp0 * tmp5
tmp7 = 1.0
tmp8 = tmp7 - tmp0
tmp9 = tmp7 - tmp2
tmp10 = 0.05
tmp11 = tmp9 + tmp10
tmp12 = triton_helpers.minimum(tmp11, tmp7)
tmp13 = triton_helpers.maximum(tmp12, tmp3)
tmp14 = tl_math.log(tmp13)
tmp15 = tmp8 * tmp14
tmp16 = tmp6 + tmp15
tmp17 = tmp2 * tmp0
tmp18 = tmp12 * tmp8
tmp19 = tmp17 + tmp18
tmp20 = tmp7 - tmp19
tmp21 = tmp0 * tmp7
tmp22 = 4.0
tmp23 = tmp8 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = libdevice.pow(tmp20, tmp24)
tmp26 = tmp16 * tmp25
tmp27 = tl.broadcast_to(tmp26, [RBLOCK])
tmp29 = triton_helpers.promote_to_tensor(tl.sum(tmp27, 0))
tmp30 = -tmp29
tl.debug_barrier()
tl.store(in_out_ptr0 + (tl.full([1], 0, tl.int32)), tmp30, None)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0; del buf0 # reuse
# Topologically Sorted Source Nodes: [x_sigmoid, clamp_1, log, los_pos, sub_1, xs_neg, add, xs_neg_1, clamp_2, log_1, los_neg, loss, pt0, sub_2, pt1, pt, sub_4, mul_4, sub_3, mul_5, one_sided_gamma, one_sided_w, loss_1, sum_1, neg], Original ATen: [aten.sigmoid, aten.clamp, aten.log, aten.mul, aten.rsub, aten.add, aten.pow, aten.sum, aten.neg]
stream0 = get_raw_stream(0)
triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0.run(buf1, arg1_1, arg0_1, 1, 256, grid=grid(1), stream=stream0)
del arg0_1
del arg1_1
return (buf1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
arg1_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1, arg1_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class AsymmetricLoss(nn.Module):
def __init__(self, gamma_neg=4, gamma_pos=1, clip=0.05, eps=1e-08,
disable_torch_grad_focal_loss=False):
super(AsymmetricLoss, self).__init__()
self.gamma_neg = gamma_neg
self.gamma_pos = gamma_pos
self.clip = clip
self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
self.eps = eps
def forward(self, x, y):
""""
Parameters
----------
x: input logits
y: targets (multi-label binarized vector)
"""
x_sigmoid = torch.sigmoid(x)
xs_pos = x_sigmoid
xs_neg = 1 - x_sigmoid
if self.clip is not None and self.clip > 0:
xs_neg = (xs_neg + self.clip).clamp(max=1)
los_pos = y * torch.log(xs_pos.clamp(min=self.eps))
los_neg = (1 - y) * torch.log(xs_neg.clamp(min=self.eps))
loss = los_pos + los_neg
if self.gamma_neg > 0 or self.gamma_pos > 0:
pt0 = xs_pos * y
pt1 = xs_neg * (1 - y)
pt = pt0 + pt1
one_sided_gamma = self.gamma_pos * y + self.gamma_neg * (1 - y)
one_sided_w = torch.pow(1 - pt, one_sided_gamma)
loss *= one_sided_w
return -loss.sum()
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0(in_out_ptr0,
in_ptr0, in_ptr1, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp2 = tl.sigmoid(tmp1)
tmp3 = 1e-08
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp5 = tl_math.log(tmp4)
tmp6 = tmp0 * tmp5
tmp7 = 1.0
tmp8 = tmp7 - tmp0
tmp9 = tmp7 - tmp2
tmp10 = 0.05
tmp11 = tmp9 + tmp10
tmp12 = triton_helpers.minimum(tmp11, tmp7)
tmp13 = triton_helpers.maximum(tmp12, tmp3)
tmp14 = tl_math.log(tmp13)
tmp15 = tmp8 * tmp14
tmp16 = tmp6 + tmp15
tmp17 = tmp2 * tmp0
tmp18 = tmp12 * tmp8
tmp19 = tmp17 + tmp18
tmp20 = tmp7 - tmp19
tmp21 = tmp0 * tmp7
tmp22 = 4.0
tmp23 = tmp8 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = libdevice.pow(tmp20, tmp24)
tmp26 = tmp16 * tmp25
tmp27 = tl.broadcast_to(tmp26, [RBLOCK])
tmp29 = triton_helpers.promote_to_tensor(tl.sum(tmp27, 0))
tmp30 = -tmp29
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp30, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_clamp_log_mul_neg_pow_rsub_sigmoid_sum_0[grid(1)](
buf1, arg1_1, arg0_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf1,
class AsymmetricLossNew(nn.Module):
def __init__(self, gamma_neg=4, gamma_pos=1, clip=0.05, eps=1e-08,
disable_torch_grad_focal_loss=False):
super(AsymmetricLossNew, self).__init__()
self.gamma_neg = gamma_neg
self.gamma_pos = gamma_pos
self.clip = clip
self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
self.eps = eps
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
| Pepijnnn/MasterThesis | AsymmetricLoss | false | 933 | [
"MIT"
] | 0 | 7ec831f5e55f5f181e0196fa78284e2846ce2e26 | https://github.com/Pepijnnn/MasterThesis/tree/7ec831f5e55f5f181e0196fa78284e2846ce2e26 | import torch
import torch.nn as nn
class Model(nn.Module):
def __init__(self, gamma_neg=4, gamma_pos=1, clip=0.05, eps=1e-08,
disable_torch_grad_focal_loss=False):
super().__init__()
self.gamma_neg = gamma_neg
self.gamma_pos = gamma_pos
self.clip = clip
self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
self.eps = eps
def forward(self, x, y):
""""
Parameters
----------
x: input logits
y: targets (multi-label binarized vector)
"""
x_sigmoid = torch.sigmoid(x)
xs_pos = x_sigmoid
xs_neg = 1 - x_sigmoid
if self.clip is not None and self.clip > 0:
xs_neg = (xs_neg + self.clip).clamp(max=1)
los_pos = y * torch.log(xs_pos.clamp(min=self.eps))
los_neg = (1 - y) * torch.log(xs_neg.clamp(min=self.eps))
loss = los_pos + los_neg
if self.gamma_neg > 0 or self.gamma_pos > 0:
pt0 = xs_pos * y
pt1 = xs_neg * (1 - y)
pt = pt0 + pt1
one_sided_gamma = self.gamma_pos * y + self.gamma_neg * (1 - y)
one_sided_w = torch.pow(1 - pt, one_sided_gamma)
loss *= one_sided_w
return -loss.sum()
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return []
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.