File size: 16,291 Bytes
587665f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 |
import torch
import re
import cupy
from modules.cupy_module.cupy_utils import cupy_launch
# Code from https://github.com/sniklaus/softmax-splatting/blob/master/softsplat.py
kernel_Softsplat_updateOutput = '''
extern "C" __global__ void kernel_Softsplat_updateOutput(
const int n,
const float* input,
const float* flow,
float* output
) { for (int intIndex = (blockIdx.x * blockDim.x) + threadIdx.x; intIndex < n; intIndex += blockDim.x * gridDim.x) {
const int intN = ( intIndex / SIZE_3(output) / SIZE_2(output) / SIZE_1(output) ) % SIZE_0(output);
const int intC = ( intIndex / SIZE_3(output) / SIZE_2(output) ) % SIZE_1(output);
const int intY = ( intIndex / SIZE_3(output) ) % SIZE_2(output);
const int intX = ( intIndex ) % SIZE_3(output);
float fltOutputX = (float) (intX) + VALUE_4(flow, intN, 0, intY, intX);
float fltOutputY = (float) (intY) + VALUE_4(flow, intN, 1, intY, intX);
int intNorthwestX = (int) (floor(fltOutputX));
int intNorthwestY = (int) (floor(fltOutputY));
int intNortheastX = intNorthwestX + 1;
int intNortheastY = intNorthwestY;
int intSouthwestX = intNorthwestX;
int intSouthwestY = intNorthwestY + 1;
int intSoutheastX = intNorthwestX + 1;
int intSoutheastY = intNorthwestY + 1;
float fltNorthwest = ((float) (intSoutheastX) - fltOutputX) * ((float) (intSoutheastY) - fltOutputY);
float fltNortheast = (fltOutputX - (float) (intSouthwestX)) * ((float) (intSouthwestY) - fltOutputY);
float fltSouthwest = ((float) (intNortheastX) - fltOutputX) * (fltOutputY - (float) (intNortheastY));
float fltSoutheast = (fltOutputX - (float) (intNorthwestX)) * (fltOutputY - (float) (intNorthwestY));
if ((intNorthwestX >= 0) & (intNorthwestX < SIZE_3(output)) & (intNorthwestY >= 0) & (intNorthwestY < SIZE_2(output))) {
atomicAdd(&output[OFFSET_4(output, intN, intC, intNorthwestY, intNorthwestX)], VALUE_4(input, intN, intC, intY, intX) * fltNorthwest);
}
if ((intNortheastX >= 0) & (intNortheastX < SIZE_3(output)) & (intNortheastY >= 0) & (intNortheastY < SIZE_2(output))) {
atomicAdd(&output[OFFSET_4(output, intN, intC, intNortheastY, intNortheastX)], VALUE_4(input, intN, intC, intY, intX) * fltNortheast);
}
if ((intSouthwestX >= 0) & (intSouthwestX < SIZE_3(output)) & (intSouthwestY >= 0) & (intSouthwestY < SIZE_2(output))) {
atomicAdd(&output[OFFSET_4(output, intN, intC, intSouthwestY, intSouthwestX)], VALUE_4(input, intN, intC, intY, intX) * fltSouthwest);
}
if ((intSoutheastX >= 0) & (intSoutheastX < SIZE_3(output)) & (intSoutheastY >= 0) & (intSoutheastY < SIZE_2(output))) {
atomicAdd(&output[OFFSET_4(output, intN, intC, intSoutheastY, intSoutheastX)], VALUE_4(input, intN, intC, intY, intX) * fltSoutheast);
}
} }
'''
kernel_Softsplat_updateGradInput = '''
extern "C" __global__ void kernel_Softsplat_updateGradInput(
const int n,
const float* input,
const float* flow,
const float* gradOutput,
float* gradInput,
float* gradFlow
) { for (int intIndex = (blockIdx.x * blockDim.x) + threadIdx.x; intIndex < n; intIndex += blockDim.x * gridDim.x) {
const int intN = ( intIndex / SIZE_3(gradInput) / SIZE_2(gradInput) / SIZE_1(gradInput) ) % SIZE_0(gradInput);
const int intC = ( intIndex / SIZE_3(gradInput) / SIZE_2(gradInput) ) % SIZE_1(gradInput);
const int intY = ( intIndex / SIZE_3(gradInput) ) % SIZE_2(gradInput);
const int intX = ( intIndex ) % SIZE_3(gradInput);
float fltGradInput = 0.0;
float fltOutputX = (float) (intX) + VALUE_4(flow, intN, 0, intY, intX);
float fltOutputY = (float) (intY) + VALUE_4(flow, intN, 1, intY, intX);
int intNorthwestX = (int) (floor(fltOutputX));
int intNorthwestY = (int) (floor(fltOutputY));
int intNortheastX = intNorthwestX + 1;
int intNortheastY = intNorthwestY;
int intSouthwestX = intNorthwestX;
int intSouthwestY = intNorthwestY + 1;
int intSoutheastX = intNorthwestX + 1;
int intSoutheastY = intNorthwestY + 1;
float fltNorthwest = ((float) (intSoutheastX) - fltOutputX) * ((float) (intSoutheastY) - fltOutputY);
float fltNortheast = (fltOutputX - (float) (intSouthwestX)) * ((float) (intSouthwestY) - fltOutputY);
float fltSouthwest = ((float) (intNortheastX) - fltOutputX) * (fltOutputY - (float) (intNortheastY));
float fltSoutheast = (fltOutputX - (float) (intNorthwestX)) * (fltOutputY - (float) (intNorthwestY));
if ((intNorthwestX >= 0) & (intNorthwestX < SIZE_3(gradOutput)) & (intNorthwestY >= 0) & (intNorthwestY < SIZE_2(gradOutput))) {
fltGradInput += VALUE_4(gradOutput, intN, intC, intNorthwestY, intNorthwestX) * fltNorthwest;
}
if ((intNortheastX >= 0) & (intNortheastX < SIZE_3(gradOutput)) & (intNortheastY >= 0) & (intNortheastY < SIZE_2(gradOutput))) {
fltGradInput += VALUE_4(gradOutput, intN, intC, intNortheastY, intNortheastX) * fltNortheast;
}
if ((intSouthwestX >= 0) & (intSouthwestX < SIZE_3(gradOutput)) & (intSouthwestY >= 0) & (intSouthwestY < SIZE_2(gradOutput))) {
fltGradInput += VALUE_4(gradOutput, intN, intC, intSouthwestY, intSouthwestX) * fltSouthwest;
}
if ((intSoutheastX >= 0) & (intSoutheastX < SIZE_3(gradOutput)) & (intSoutheastY >= 0) & (intSoutheastY < SIZE_2(gradOutput))) {
fltGradInput += VALUE_4(gradOutput, intN, intC, intSoutheastY, intSoutheastX) * fltSoutheast;
}
gradInput[intIndex] = fltGradInput;
} }
'''
kernel_Softsplat_updateGradFlow = '''
extern "C" __global__ void kernel_Softsplat_updateGradFlow(
const int n,
const float* input,
const float* flow,
const float* gradOutput,
float* gradInput,
float* gradFlow
) { for (int intIndex = (blockIdx.x * blockDim.x) + threadIdx.x; intIndex < n; intIndex += blockDim.x * gridDim.x) {
float fltGradFlow = 0.0;
const int intN = ( intIndex / SIZE_3(gradFlow) / SIZE_2(gradFlow) / SIZE_1(gradFlow) ) % SIZE_0(gradFlow);
const int intC = ( intIndex / SIZE_3(gradFlow) / SIZE_2(gradFlow) ) % SIZE_1(gradFlow);
const int intY = ( intIndex / SIZE_3(gradFlow) ) % SIZE_2(gradFlow);
const int intX = ( intIndex ) % SIZE_3(gradFlow);
float fltOutputX = (float) (intX) + VALUE_4(flow, intN, 0, intY, intX);
float fltOutputY = (float) (intY) + VALUE_4(flow, intN, 1, intY, intX);
int intNorthwestX = (int) (floor(fltOutputX));
int intNorthwestY = (int) (floor(fltOutputY));
int intNortheastX = intNorthwestX + 1;
int intNortheastY = intNorthwestY;
int intSouthwestX = intNorthwestX;
int intSouthwestY = intNorthwestY + 1;
int intSoutheastX = intNorthwestX + 1;
int intSoutheastY = intNorthwestY + 1;
float fltNorthwest = 0.0;
float fltNortheast = 0.0;
float fltSouthwest = 0.0;
float fltSoutheast = 0.0;
if (intC == 0) {
fltNorthwest = ((float) (-1.0)) * ((float) (intSoutheastY) - fltOutputY);
fltNortheast = ((float) (+1.0)) * ((float) (intSouthwestY) - fltOutputY);
fltSouthwest = ((float) (-1.0)) * (fltOutputY - (float) (intNortheastY));
fltSoutheast = ((float) (+1.0)) * (fltOutputY - (float) (intNorthwestY));
} else if (intC == 1) {
fltNorthwest = ((float) (intSoutheastX) - fltOutputX) * ((float) (-1.0));
fltNortheast = (fltOutputX - (float) (intSouthwestX)) * ((float) (-1.0));
fltSouthwest = ((float) (intNortheastX) - fltOutputX) * ((float) (+1.0));
fltSoutheast = (fltOutputX - (float) (intNorthwestX)) * ((float) (+1.0));
}
for (int intChannel = 0; intChannel < SIZE_1(gradOutput); intChannel += 1) {
float fltInput = VALUE_4(input, intN, intChannel, intY, intX);
if ((intNorthwestX >= 0) & (intNorthwestX < SIZE_3(gradOutput)) & (intNorthwestY >= 0) & (intNorthwestY < SIZE_2(gradOutput))) {
fltGradFlow += fltInput * VALUE_4(gradOutput, intN, intChannel, intNorthwestY, intNorthwestX) * fltNorthwest;
}
if ((intNortheastX >= 0) & (intNortheastX < SIZE_3(gradOutput)) & (intNortheastY >= 0) & (intNortheastY < SIZE_2(gradOutput))) {
fltGradFlow += fltInput * VALUE_4(gradOutput, intN, intChannel, intNortheastY, intNortheastX) * fltNortheast;
}
if ((intSouthwestX >= 0) & (intSouthwestX < SIZE_3(gradOutput)) & (intSouthwestY >= 0) & (intSouthwestY < SIZE_2(gradOutput))) {
fltGradFlow += fltInput * VALUE_4(gradOutput, intN, intChannel, intSouthwestY, intSouthwestX) * fltSouthwest;
}
if ((intSoutheastX >= 0) & (intSoutheastX < SIZE_3(gradOutput)) & (intSoutheastY >= 0) & (intSoutheastY < SIZE_2(gradOutput))) {
fltGradFlow += fltInput * VALUE_4(gradOutput, intN, intChannel, intSoutheastY, intSoutheastX) * fltSoutheast;
}
}
gradFlow[intIndex] = fltGradFlow;
} }
'''
def cupy_kernel(strFunction, objVariables):
strKernel = globals()[strFunction]
while True:
objMatch = re.search('(SIZE_)([0-4])(\()([^\)]*)(\))', strKernel)
if objMatch is None:
break
# end
intArg = int(objMatch.group(2))
strTensor = objMatch.group(4)
intSizes = objVariables[strTensor].size()
strKernel = strKernel.replace(objMatch.group(), str(intSizes[intArg]))
# end
while True:
objMatch = re.search('(OFFSET_)([0-4])(\()([^\)]+)(\))', strKernel)
if objMatch is None:
break
# end
intArgs = int(objMatch.group(2))
strArgs = objMatch.group(4).split(',')
strTensor = strArgs[0]
intStrides = objVariables[strTensor].stride()
strIndex = [ '((' + strArgs[intArg + 1].replace('{', '(').replace('}', ')').strip() + ')*' + str(intStrides[intArg]) + ')' for intArg in range(intArgs) ]
strKernel = strKernel.replace(objMatch.group(0), '(' + str.join('+', strIndex) + ')')
# end
while True:
objMatch = re.search('(VALUE_)([0-4])(\()([^\)]+)(\))', strKernel)
if objMatch is None:
break
# end
intArgs = int(objMatch.group(2))
strArgs = objMatch.group(4).split(',')
strTensor = strArgs[0]
intStrides = objVariables[strTensor].stride()
strIndex = [ '((' + strArgs[intArg + 1].replace('{', '(').replace('}', ')').strip() + ')*' + str(intStrides[intArg]) + ')' for intArg in range(intArgs) ]
strKernel = strKernel.replace(objMatch.group(0), strTensor + '[' + str.join('+', strIndex) + ']')
# end
return strKernel
# end
class _FunctionSoftsplat(torch.autograd.Function):
@staticmethod
def forward(self, input, flow):
intSamples = input.shape[0]
intInputDepth, intInputHeight, intInputWidth = input.shape[1], input.shape[2], input.shape[3]
intFlowDepth, intFlowHeight, intFlowWidth = flow.shape[1], flow.shape[2], flow.shape[3]
assert(intFlowDepth == 2)
assert(intInputHeight == intFlowHeight)
assert(intInputWidth == intFlowWidth)
input = input.contiguous(); assert(input.is_cuda == True)
flow = flow.contiguous(); assert(flow.is_cuda == True)
output = input.new_zeros([ intSamples, intInputDepth, intInputHeight, intInputWidth ])
if input.is_cuda == True:
n = output.nelement()
cupy_launch('kernel_Softsplat_updateOutput', cupy_kernel('kernel_Softsplat_updateOutput', {
'input': input,
'flow': flow,
'output': output
}))(
grid=tuple([ int((n + 512 - 1) / 512), 1, 1 ]),
block=tuple([ 512, 1, 1 ]),
args=[ cupy.int32(n), input.data_ptr(), flow.data_ptr(), output.data_ptr() ]
)
elif input.is_cuda == False:
raise NotImplementedError()
# end
self.save_for_backward(input, flow)
return output
# end
@staticmethod
def backward(self, gradOutput):
input, flow = self.saved_tensors
intSamples = input.shape[0]
intInputDepth, intInputHeight, intInputWidth = input.shape[1], input.shape[2], input.shape[3]
intFlowDepth, intFlowHeight, intFlowWidth = flow.shape[1], flow.shape[2], flow.shape[3]
assert(intFlowDepth == 2)
assert(intInputHeight == intFlowHeight)
assert(intInputWidth == intFlowWidth)
gradOutput = gradOutput.contiguous(); assert(gradOutput.is_cuda == True)
gradInput = input.new_zeros([ intSamples, intInputDepth, intInputHeight, intInputWidth ]) if self.needs_input_grad[0] == True else None
gradFlow = input.new_zeros([ intSamples, intFlowDepth, intFlowHeight, intFlowWidth ]) if self.needs_input_grad[1] == True else None
if input.is_cuda == True:
if gradInput is not None:
n = gradInput.nelement()
cupy_launch('kernel_Softsplat_updateGradInput', cupy_kernel('kernel_Softsplat_updateGradInput', {
'input': input,
'flow': flow,
'gradOutput': gradOutput,
'gradInput': gradInput,
'gradFlow': gradFlow
}))(
grid=tuple([ int((n + 512 - 1) / 512), 1, 1 ]),
block=tuple([ 512, 1, 1 ]),
args=[ cupy.int32(n), input.data_ptr(), flow.data_ptr(), gradOutput.data_ptr(), gradInput.data_ptr(), None ]
)
# end
if gradFlow is not None:
n = gradFlow.nelement()
cupy_launch('kernel_Softsplat_updateGradFlow', cupy_kernel('kernel_Softsplat_updateGradFlow', {
'input': input,
'flow': flow,
'gradOutput': gradOutput,
'gradInput': gradInput,
'gradFlow': gradFlow
}))(
grid=tuple([ int((n + 512 - 1) / 512), 1, 1 ]),
block=tuple([ 512, 1, 1 ]),
args=[ cupy.int32(n), input.data_ptr(), flow.data_ptr(), gradOutput.data_ptr(), None, gradFlow.data_ptr() ]
)
# end
elif input.is_cuda == False:
raise NotImplementedError()
# end
return gradInput, gradFlow
# end
# end
def FunctionSoftsplat(tenInput, tenFlow, tenMetric, strType):
assert(tenMetric is None or tenMetric.shape[1] == 1)
assert(strType in ['summation', 'average', 'linear', 'softmax'])
if strType == 'average':
tenInput = torch.cat([ tenInput, tenInput.new_ones(tenInput.shape[0], 1, tenInput.shape[2], tenInput.shape[3]) ], 1)
elif strType == 'linear':
tenInput = torch.cat([ tenInput * tenMetric, tenMetric ], 1)
elif strType == 'softmax':
tenInput = torch.cat([ tenInput * tenMetric.exp(), tenMetric.exp() ], 1)
# end
tenOutput = _FunctionSoftsplat.apply(tenInput, tenFlow)
if strType != 'summation':
tenNormalize = tenOutput[:, -1:, :, :]
tenNormalize[tenNormalize == 0.0] = 1.0
tenOutput = tenOutput[:, :-1, :, :] / tenNormalize
# end
return tenOutput
# end
class ModuleSoftsplat(torch.nn.Module):
def __init__(self, strType):
super().__init__()
self.strType = strType
# end
def forward(self, tenInput, tenFlow, tenMetric):
return FunctionSoftsplat(tenInput, tenFlow, tenMetric, self.strType)
# end
# end |