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
|
#include <c10/util/accumulate.h>
#include "caffe2/operators/elementwise_mul_op.h"
#include "caffe2/utils/math/broadcast.h"
#include <algorithm>
#include <functional>
#include <string>
#include <vector>
namespace caffe2 {
namespace {
template <typename TGrad, typename TIn>
void ComputeMulGradientFastpath(
const int A_size,
const int B_size,
const int C_size,
const TGrad* dC,
const TIn* A,
const TIn* B,
TGrad* dA,
TGrad* dB) {
int A_index = 0;
int B_index = 0;
for (int C_index = 0; C_index < C_size; ++C_index) {
dA[A_index] += dC[C_index] * B[B_index];
dB[B_index] += dC[C_index] * A[A_index];
A_index++;
B_index++;
if (A_index >= A_size) {
A_index = 0;
}
if (B_index >= B_size) {
B_index = 0;
}
}
}
template <typename TGrad, typename TIn>
void ComputeMulGradient(
const int ndim,
const int* A_dims,
const int* B_dims,
const int* C_dims,
const TGrad* dC,
const TIn* A,
const TIn* B,
TGrad* dA,
TGrad* dB,
CPUContext* context) {
const auto A_size = c10::multiply_integers(A_dims, A_dims + ndim);
const auto B_size = c10::multiply_integers(B_dims, B_dims + ndim);
const auto C_size = c10::multiply_integers(C_dims, C_dims + ndim);
math::Set<TGrad, CPUContext>(A_size, TGrad(0), dA, context);
math::Set<TGrad, CPUContext>(B_size, TGrad(0), dB, context);
if (
math::can_use_broadcast_fastpath(ndim, A_dims)
&& math::can_use_broadcast_fastpath(ndim, B_dims)) {
ComputeMulGradientFastpath(A_size, B_size, C_size, dC, A, B, dA, dB);
return;
}
std::vector<int> index(ndim, 0);
for (int C_index = 0; C_index < C_size; ++C_index) {
const int A_index =
math::utils::GetIndexFromDims(ndim, A_dims, index.data());
const int B_index =
math::utils::GetIndexFromDims(ndim, B_dims, index.data());
dA[A_index] += dC[C_index] * B[B_index];
dB[B_index] += dC[C_index] * A[A_index];
math::utils::IncreaseIndexInDims(ndim, C_dims, index.data());
}
}
// A : input not to broadcast whose size is common_size x broadcast_size
// B : input to broadcast whose size is common_size
void ComputeMulGradient(
const int common_size,
const int broadcast_size,
const float* dC,
const float* A,
const float* B,
float* dA,
float* dB,
CPUContext* context) {
for (int i = 0; i < common_size; ++i) {
caffe2::math::Scale(
broadcast_size,
B[i],
dC + i * broadcast_size,
dA + i * broadcast_size,
context);
caffe2::math::Dot(
broadcast_size,
dC + i * broadcast_size,
A + i * broadcast_size,
dB + i,
context);
}
}
void ComputeMulGradient(
const int size,
const float* dC,
const float* A,
const float* B,
float* dA,
float* dB) {
if (dA != nullptr) {
CAFFE_ENFORCE_NE(dA, dB, "Outputs dA and dB should point to distinct blobs");
}
if (dC == dA) {
// Ensure operation can be performed in-place.
// See below comment in `MulFunctor::Backward`.
std::swap(A, B);
std::swap(dA, dB);
}
for (int i = 0; i < size; ++i) {
dA[i] = dC[i] * B[i];
dB[i] = dC[i] * A[i];
}
}
} // namespace
template <>
template <typename TGrad, typename TIn, typename TOut>
bool MulFunctor<CPUContext>::Backward(
const std::vector<int>& A_dims,
const std::vector<int>& B_dims,
const TGrad* dC,
const TIn* A,
const TIn* B,
const TOut* /* C */,
TGrad* dA,
TGrad* dB,
CPUContext* context) const {
if (dA != nullptr) {
CAFFE_ENFORCE_NE(dA, dB, "Outputs dA and dB should point to distinct blobs");
}
if (A_dims == B_dims) {
const auto size = c10::multiply_integers(A_dims);
if (dC == dA) {
// A, B, and dC are inputs (dC is the output of the previous gradient op
// in the dag), and dA and dB are outputs. If the op is performed
// in-place, either dA or dB could alias dC. In the dC == dA case, we need
// to make sure we don't overwrite dC when we write to dA, so swap the
// inputs to avoid clobbering dC. Semantically this is equivalent with
// writing to dB first. The other case (dC == dB) is already safe because
// we are writing to dA first.
std::swap(A, B);
std::swap(dA, dB);
}
math::Mul(size, dC, B, dA, context);
math::Mul(size, dC, A, dB, context);
return true;
}
const int ndim = std::max(A_dims.size(), B_dims.size());
if (ndim == 0) {
return true;
}
std::vector<int> A_broadcast_dims(ndim);
std::vector<int> B_broadcast_dims(ndim);
std::vector<int> C_broadcast_dims(ndim);
math::utils::ComputeBroadcastBinaryOpDims(
A_dims.size(),
A_dims.data(),
B_dims.size(),
B_dims.data(),
A_broadcast_dims.data(),
B_broadcast_dims.data(),
C_broadcast_dims.data());
const int C_size = std::accumulate(
C_broadcast_dims.cbegin(),
C_broadcast_dims.cbegin() + ndim,
1,
// NOLINTNEXTLINE(modernize-use-transparent-functors)
std::multiplies<int>());
if (C_size == 0) {
const auto A_size = c10::multiply_integers(A_dims);
const auto B_size = c10::multiply_integers(B_dims);
math::Set<TGrad, CPUContext>(A_size, TGrad(0), dA, context);
math::Set<TGrad, CPUContext>(B_size, TGrad(0), dB, context);
return true;
}
// Flatten dims as much as possible
// We call A is broadcasted at dim d if A_broadcast_dims[d] <= 1
// Two consecutive dims d and d+1 can be flattened if
// A and B are broadcasted at dim d, or
// A and B are broadcasted at dim d + 1, or
// A is broadcasted at dim d and d + 1, or
// B is broadcasted at dim d and d + 1, or
// A and B are not broadcasted at dim d and d + 1
std::vector<int> A_broadcast_dims_flattened, B_broadcast_dims_flattened,
C_broadcast_dims_flattened;
A_broadcast_dims_flattened.reserve(ndim);
B_broadcast_dims_flattened.reserve(ndim);
A_broadcast_dims_flattened.push_back(A_broadcast_dims[0]);
B_broadcast_dims_flattened.push_back(B_broadcast_dims[0]);
for (int i = 1; i < ndim; ++i) {
int A_old = A_broadcast_dims_flattened.back();
int B_old = B_broadcast_dims_flattened.back();
int A_new = A_broadcast_dims[i];
int B_new = B_broadcast_dims[i];
if ((A_old == 1 && B_old == 1) || (A_new == 1 && B_new == 1) ||
(A_old == 1 && A_new == 1) || (B_old == 1 && B_new == 1) ||
(A_old > 1 && B_old > 1 && A_new > 1 && B_new > 1)) {
A_broadcast_dims_flattened.back() *= A_new;
B_broadcast_dims_flattened.back() *= B_new;
} else {
A_broadcast_dims_flattened.push_back(A_new);
B_broadcast_dims_flattened.push_back(B_new);
}
}
int ndim_flattened = A_broadcast_dims_flattened.size();
C_broadcast_dims_flattened.resize(ndim_flattened);
for (int i = 0; i < ndim_flattened; ++i) {
C_broadcast_dims_flattened[i] =
std::max(A_broadcast_dims_flattened[i], B_broadcast_dims_flattened[i]);
}
if (std::is_same<TGrad, float>::value && std::is_same<TIn, float>::value &&
ndim_flattened <= 2 &&
A_broadcast_dims_flattened[0] == B_broadcast_dims_flattened[0] &&
(ndim_flattened == 1 || A_broadcast_dims_flattened[1] <= 1 ||
B_broadcast_dims_flattened[1] <= 1)) {
if (ndim_flattened == 2) {
// fast path when we have 2 flattened dimensions and the second dimension
// is broadcasted.
bool broadcast_B = B_broadcast_dims_flattened[1] <= 1;
ComputeMulGradient(
C_broadcast_dims_flattened[0],
C_broadcast_dims_flattened[1],
reinterpret_cast<const float*>(dC),
reinterpret_cast<const float*>(broadcast_B ? A : B),
reinterpret_cast<const float*>(broadcast_B ? B : A),
reinterpret_cast<float*>(broadcast_B ? dA : dB),
reinterpret_cast<float*>(broadcast_B ? dB : dA),
context);
} else {
// fast path when we have 1 flattened dimension
assert(ndim_flattened == 1);
ComputeMulGradient(
C_broadcast_dims_flattened[0],
reinterpret_cast<const float*>(dC),
reinterpret_cast<const float*>(A),
reinterpret_cast<const float*>(B),
reinterpret_cast<float*>(dA),
reinterpret_cast<float*>(dB));
}
} else {
ComputeMulGradient<TGrad, TIn>(
ndim_flattened,
A_broadcast_dims_flattened.data(),
B_broadcast_dims_flattened.data(),
C_broadcast_dims_flattened.data(),
dC,
A,
B,
dA,
dB,
context);
}
return true;
}
// Used in fallback ops
template bool MulFunctor<CPUContext>::Backward<float, float, float>(
const std::vector<int>& A_dims,
const std::vector<int>& B_dims,
const float* dC,
const float* A,
const float* B,
const float* /* C */,
float* dA,
float* dB,
CPUContext* context) const;
template bool MulFunctor<CPUContext>::Backward<int32_t, int32_t, int32_t>(
const std::vector<int>& A_dims,
const std::vector<int>& B_dims,
const int* dC,
const int* A,
const int* B,
const int* /* C */,
int* dA,
int* dB,
CPUContext* context) const;
template bool MulFunctor<CPUContext>::Backward<double, double, double>(
const std::vector<int>& A_dims,
const std::vector<int>& B_dims,
const double* dC,
const double* A,
const double* B,
const double* /* C */,
double* dA,
double* dB,
CPUContext* context) const;
template bool MulFunctor<CPUContext>::Backward<int64_t, int64_t, int64_t>(
const std::vector<int>& A_dims,
const std::vector<int>& B_dims,
const int64_t* dC,
const int64_t* A,
const int64_t* B,
const int64_t* /* C */,
int64_t* dA,
int64_t* dB,
CPUContext* context) const;
REGISTER_CPU_OPERATOR(
MulGradient,
BinaryElementwiseGradientOp<
NumericTypes,
CPUContext,
MulFunctor<CPUContext>>);
namespace {
class GetMulGradient final : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
std::vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"MulGradient",
"",
std::vector<std::string>{GO(0), I(0), I(1)},
std::vector<std::string>{GI(0), GI(1)});
}
};
} // namespace
REGISTER_GRADIENT(Mul, GetMulGradient);
} // namespace caffe2
|