activation_grad_kernel.cu 20.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

15 16
#include "paddle/phi/kernels/activation_grad_kernel.h"

17
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
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
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/bfloat16.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/elementwise_base.h"
#include "paddle/phi/kernels/impl/activation_grad_impl.h"

namespace phi {

template <typename T, typename Context, typename Functor>
void ActivationGradGPUImpl(const Context& dev_ctx,
                           const DenseTensor* x,
                           const DenseTensor* out,
                           const DenseTensor* d_out,
                           DenseTensor* d_x,
                           const Functor& functor) {
  if (static_cast<int>(Functor::FwdDeps()) &
      static_cast<int>(funcs::ActBwdOpFwdDeps::kDepOut)) {
    PADDLE_ENFORCE_NOT_NULL(
        out, errors::NotFound("The input DenseTensor Out can not be nullptr"));
  }
  PADDLE_ENFORCE_NOT_NULL(
      d_out, errors::NotFound("The input DenseTensor dOut can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      d_x, errors::NotFound("The output DenseTensor dX can not be nullptr"));
Z
zhangbo9674 已提交
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
  if (!out) {
    out = d_out;  // fake out
  }
  if (static_cast<int>(Functor::FwdDeps()) &
      static_cast<int>(funcs::ActBwdOpFwdDeps::kDepX)) {
    PADDLE_ENFORCE_NOT_NULL(
        x, errors::NotFound("The input DenseTensor X can not be nullptr"));
  } else {
    VLOG(10) << "Inplace activation of Op Functor: " << typeid(Functor).name();
    x = d_x;
  }

  dev_ctx.template Alloc<T>(d_x);

  std::vector<const DenseTensor*> ins = {d_out};
  std::vector<DenseTensor*> outs = {d_x};

  if (static_cast<int>(Functor::FwdDeps()) ==
      static_cast<int>(funcs::ActBwdOpFwdDeps::kDepOut)) {
    // Only need forward output Out
    ins.push_back(out);
    funcs::ElementwiseKernel<T>(dev_ctx, ins, &outs, functor);
  } else if (static_cast<int>(Functor::FwdDeps()) ==
             static_cast<int>(funcs::ActBwdOpFwdDeps::kDepX)) {
    // Only need forward input X
    ins.push_back(x);
    funcs::ElementwiseKernel<T>(dev_ctx, ins, &outs, functor);
  } else {
    funcs::ElementwiseKernel<T>(dev_ctx, ins, &outs, functor);
  }
}

Y
YuanRisheng 已提交
76
#define DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(name, functor_class) \
77 78 79 80 81
  template <typename T, typename Context>                           \
  void name##GradKernel(const Context& dev_ctx,                     \
                        const DenseTensor& x,                       \
                        const DenseTensor& dout,                    \
                        DenseTensor* dx) {                          \
82 83
    funcs::functor_class<T> functor;                                \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>(     \
84 85 86
        dev_ctx, &x, nullptr, &dout, dx, functor);                  \
  }

Y
YuanRisheng 已提交
87
#define DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(         \
88 89 90 91 92 93 94 95 96 97 98 99 100 101
    name, functor_class, attr)                                  \
  template <typename T, typename Context>                       \
  void name##GradKernel(const Context& dev_ctx,                 \
                        const DenseTensor& x,                   \
                        const DenseTensor& dout,                \
                        float attr,                             \
                        DenseTensor* dx) {                      \
    funcs::functor_class<T> functor;                            \
    auto attrs = functor.GetAttrs();                            \
    *(attrs[0].second) = attr;                                  \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
        dev_ctx, &x, nullptr, &dout, dx, functor);              \
  }

Y
YuanRisheng 已提交
102
#define DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(         \
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118
    name, functor_class, attr1, attr2)                          \
  template <typename T, typename Context>                       \
  void name##GradKernel(const Context& dev_ctx,                 \
                        const DenseTensor& x,                   \
                        const DenseTensor& dout,                \
                        float attr1,                            \
                        float attr2,                            \
                        DenseTensor* dx) {                      \
    funcs::functor_class<T> functor;                            \
    auto attrs = functor.GetAttrs();                            \
    *(attrs[0].second) = attr1;                                 \
    *(attrs[1].second) = attr2;                                 \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
        dev_ctx, &x, nullptr, &dout, dx, functor);              \
  }

Y
YuanRisheng 已提交
119
#define DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(name, functor_class) \
120 121 122 123 124
  template <typename T, typename Context>                             \
  void name##GradKernel(const Context& dev_ctx,                       \
                        const DenseTensor& out,                       \
                        const DenseTensor& dout,                      \
                        DenseTensor* dx) {                            \
125 126
    funcs::functor_class<T> functor;                                  \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>(       \
127 128 129
        dev_ctx, nullptr, &out, &dout, dx, functor);                  \
  }

Y
YuanRisheng 已提交
130
#define DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPOUT(       \
131 132 133 134 135 136 137 138 139 140 141 142 143 144
    name, functor_class, attr)                                  \
  template <typename T, typename Context>                       \
  void name##GradKernel(const Context& dev_ctx,                 \
                        const DenseTensor& out,                 \
                        const DenseTensor& dout,                \
                        float attr,                             \
                        DenseTensor* dx) {                      \
    funcs::functor_class<T> functor;                            \
    auto attrs = functor.GetAttrs();                            \
    *(attrs[0].second) = attr;                                  \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
        dev_ctx, nullptr, &out, &dout, dx, functor);            \
  }

Y
YuanRisheng 已提交
145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
#define DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT(       \
    name, functor_class, attr1, attr2)                          \
  template <typename T, typename Context>                       \
  void name##GradKernel(const Context& dev_ctx,                 \
                        const DenseTensor& out,                 \
                        const DenseTensor& dout,                \
                        float attr1,                            \
                        float attr2,                            \
                        DenseTensor* dx) {                      \
    funcs::functor_class<T> functor;                            \
    auto attrs = functor.GetAttrs();                            \
    *(attrs[0].second) = attr1;                                 \
    *(attrs[1].second) = attr2;                                 \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
        dev_ctx, nullptr, &out, &dout, dx, functor);            \
  }

Y
YuanRisheng 已提交
162 163 164 165 166 167 168 169 170
#define DEFINE_GPU_ACTIVATION_GRAD_KERNEL_NODEP(name, functor_class)      \
  template <typename T, typename Context>                                 \
  void name##GradKernel(                                                  \
      const Context& dev_ctx, const DenseTensor& dout, DenseTensor* dx) { \
    funcs::functor_class<T> functor;                                      \
    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>(           \
        dev_ctx, nullptr, nullptr, &dout, dx, functor);                   \
  }

Y
YuanRisheng 已提交
171 172
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Relu, CudaReluGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Tanh, CudaTanhGradFunctor);
Y
YuanRisheng 已提交
173 174
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Sigmoid, CudaSigmoidGradFunctor);

Y
YuanRisheng 已提交
175 176 177 178
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_NODEP(Round, CudaZeroGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_NODEP(Floor, CudaZeroGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_NODEP(Ceil, CudaZeroGradFunctor);

Y
YuanRisheng 已提交
179 180 181 182 183 184 185 186 187 188 189 190 191
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Cos, CudaCosGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Tan, CudaTanGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Acos, CudaAcosGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Sin, CudaSinGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Asin, CudaAsinGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Atan, CudaAtanGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Sinh, CudaSinhGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Cosh, CudaCoshGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Asinh, CudaAsinhGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Acosh, CudaAcoshGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Atanh, CudaAtanhGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(TanhShrink, CudaTanhShrinkGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Silu, CudaSiluGradFunctor);
192 193 194 195 196 197 198
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Square, CudaSquareGradFunctor);

DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Exp, CudaExpGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Expm1, CudaExpm1GradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Reciprocal, CudaReciprocalGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Sqrt, CudaSqrtGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Rsqrt, CudaRsqrtGradFunctor);
199
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Softsign, CudaSoftsignGradFunctor);
Y
YuanRisheng 已提交
200
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(LogSigmoid, CudaLogSigmoidGradFunctor);
201 202 203 204
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Log, CudaLogGradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Log2, CudaLog2GradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Log10, CudaLog10GradFunctor);
DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Log1p, CudaLog1pGradFunctor);
Y
YuanRisheng 已提交
205 206

DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(LeakyRelu,
207 208
                                               CudaLeakyReluGradFunctor,
                                               alpha);
Y
YuanRisheng 已提交
209
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(ThresholdedRelu,
210 211
                                               CudaThresholdedReluGradFunctor,
                                               threshold);
Y
YuanRisheng 已提交
212 213 214 215 216 217
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(SoftShrink,
                                               CudaSoftShrinkGradFunctor,
                                               lambda);
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(HardShrink,
                                               CudaHardShrinkGradFunctor,
                                               threshold);
Y
YuanRisheng 已提交
218 219 220
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(Swish,
                                               CudaSwishGradFunctor,
                                               beta);
221

222 223 224
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(Mish,
                                               CudaMishGradFunctor,
                                               threshold);
Y
YuanRisheng 已提交
225 226 227
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPX(Celu,
                                               CudaCELUGradFunctor,
                                               alpha);
228 229 230
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_ONE_ATTRS_DEPOUT(Relu6,
                                                 CudaRelu6GradFunctor,
                                                 threshold);
231

232 233
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(HardTanh,
                                               CudaHardTanhGradFunctor,
234 235
                                               t_min,
                                               t_max);
236

237 238 239 240 241 242 243 244 245
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(STanh,
                                               CudaSTanhGradFunctor,
                                               scale_a,
                                               scale_b);

DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(Softplus,
                                               CudaSoftplusGradFunctor,
                                               beta,
                                               threshold);
Y
YuanRisheng 已提交
246 247 248 249 250
DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT(HardSigmoid,
                                                 CudaHardSigmoidGradFunctor,
                                                 slope,
                                                 offset);

Y
YuanRisheng 已提交
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272
template <typename T, typename Context>
void EluGradKernel(const Context& dev_ctx,
                   const DenseTensor& x,
                   const DenseTensor& out,
                   const DenseTensor& dout,
                   float alpha,
                   DenseTensor* dx) {
  dev_ctx.template Alloc<T>(dx);
  std::vector<const DenseTensor*> ins = {&dout, &out};
  std::vector<DenseTensor*> outs = {dx};
  if (alpha > 0) {
    funcs::CudaELUGradFunctor<T> functor;
    functor.alpha = alpha;
    funcs::ElementwiseKernel<T>(dev_ctx, ins, &outs, functor);
  } else {
    funcs::CudaELUGradNegativeAlphaFunctor<T> functor;
    functor.alpha = alpha;
    ins.push_back(&x);
    funcs::ElementwiseKernel<T>(dev_ctx, ins, &outs, functor);
  }
}

Y
YuanRisheng 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
template <typename T, typename Context>
void HardSwishGradKernel(const Context& dev_ctx,
                         const DenseTensor& x,
                         const DenseTensor& dout,
                         float threshold,
                         float scale,
                         float offset,
                         DenseTensor* dx) {
  funcs::CudaHardSwishGradFunctor<T> functor;
  auto attrs = functor.GetAttrs();
  *(attrs[0].second) = threshold;
  *(attrs[1].second) = scale;
  *(attrs[2].second) = offset;
  ActivationGradGPUImpl<T, Context, funcs::CudaHardSwishGradFunctor<T>>(
      dev_ctx, &x, nullptr, &dout, dx, functor);
}

290
}  // namespace phi
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
#ifdef PADDLE_WITH_HIP
PD_REGISTER_KERNEL(relu_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::ReluGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}
PD_REGISTER_KERNEL(relu_double_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::ReluDoubleGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}
#else
PD_REGISTER_KERNEL(relu_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::ReluGradKernel,
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
PD_REGISTER_KERNEL(relu_double_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::ReluDoubleGradKernel,
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
#endif
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

#define PD_REGISTER_ACTIVATION_GRAD_KERNEL(name, func) \
  PD_REGISTER_KERNEL(name,                             \
                     GPU,                              \
                     ALL_LAYOUT,                       \
                     phi::func,                        \
                     float,                            \
                     double,                           \
                     phi::dtype::float16,              \
                     phi::dtype::bfloat16) {}

PD_REGISTER_ACTIVATION_GRAD_KERNEL(sin_grad, SinGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(cos_grad, CosGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(tan_grad, TanGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(acos_grad, AcosGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(asin_grad, AsinGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(atan_grad, AtanGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sinh_grad, SinhGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(cosh_grad, CoshGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(asinh_grad, AsinhGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(acosh_grad, AcoshGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(atanh_grad, AtanhGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(tanh_grad, TanhGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(tanh_double_grad, TanhDoubleGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(tanh_triple_grad, TanhTripleGradKernel)
350
PD_REGISTER_ACTIVATION_GRAD_KERNEL(hard_tanh_grad, HardTanhGradKernel)
351 352 353 354 355
PD_REGISTER_ACTIVATION_GRAD_KERNEL(leaky_relu_grad, LeakyReluGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(leaky_relu_double_grad,
                                   LeakyReluDoubleGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(thresholded_relu_grad,
                                   ThresholdedReluGradKernel)
356
PD_REGISTER_ACTIVATION_GRAD_KERNEL(relu6_grad, Relu6GradKernel)
357 358 359 360 361
PD_REGISTER_ACTIVATION_GRAD_KERNEL(mish_grad, MishGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(stanh_grad, STanhGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(reciprocal_grad, ReciprocalGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(softplus_grad, SoftplusGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_grad, SqrtGradKernel)
Y
YuanRisheng 已提交
362
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_double_grad, SqrtDoubleGradKernel)
363
PD_REGISTER_ACTIVATION_GRAD_KERNEL(rsqrt_grad, RsqrtGradKernel)
Y
YuanRisheng 已提交
364
PD_REGISTER_ACTIVATION_GRAD_KERNEL(rsqrt_double_grad, RsqrtDoubleGradKernel)
365 366 367 368 369 370 371 372 373 374 375

PD_REGISTER_KERNEL(exp_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::ExpGradKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16) {}

Y
YuanRisheng 已提交
376 377 378 379 380 381
PD_REGISTER_ACTIVATION_GRAD_KERNEL(soft_shrink_grad, SoftShrinkGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(hard_shrink_grad, HardShrinkGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(tanh_shrink_grad, TanhShrinkGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(silu_grad, SiluGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(elu_grad, EluGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(elu_double_grad, EluDoubleGradKernel)
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408

PD_REGISTER_KERNEL(expm1_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::Expm1GradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}

PD_REGISTER_KERNEL(logit_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::LogitGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}

PD_REGISTER_KERNEL(square_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::SquareGradKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
Y
YuanRisheng 已提交
409 410 411 412 413 414 415 416 417 418
PD_REGISTER_KERNEL(square_double_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::SquareDoubleGradKernel,
                   float,
                   double,
                   int,
                   int64_t,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}
419

420
PD_REGISTER_ACTIVATION_GRAD_KERNEL(softsign_grad, SoftsignGradKernel)
Y
YuanRisheng 已提交
421 422 423 424 425
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sigmoid_grad, SigmoidGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sigmoid_double_grad, SigmoidDoubleGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(sigmoid_triple_grad, SigmoidTripleGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(hard_sigmoid_grad, HardSigmoidGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(logsigmoid_grad, LogSigmoidGradKernel)
426 427 428 429 430 431 432 433 434 435 436
PD_REGISTER_ACTIVATION_GRAD_KERNEL(log_grad, LogGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(log2_grad, Log2GradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(log10_grad, Log10GradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(log1p_grad, Log1pGradKernel)
PD_REGISTER_KERNEL(log_double_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::LogDoubleGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}
Y
YuanRisheng 已提交
437 438 439 440 441
PD_REGISTER_ACTIVATION_GRAD_KERNEL(hard_swish_grad, HardSwishGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(swish_grad, SwishGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(round_grad, RoundGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(floor_grad, FloorGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(ceil_grad, CeilGradKernel)
Y
YuanRisheng 已提交
442 443
PD_REGISTER_ACTIVATION_GRAD_KERNEL(celu_grad, CeluGradKernel)
PD_REGISTER_ACTIVATION_GRAD_KERNEL(celu_double_grad, CeluDoubleGradKernel)
Y
YuanRisheng 已提交
444 445 446 447 448 449 450 451 452

PD_REGISTER_KERNEL(pow_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::PowGradKernel,
                   float,
                   double,
                   int,
                   int64_t,
453 454
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}