activation_grad_impl.h 4.6 KB
Newer Older
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
// 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.

#pragma once

#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/kernels/funcs/activation_functor.h"

#include "paddle/fluid/platform/device_context.h"

namespace phi {

template <typename T, typename Context, typename Functor>
void ActivationGradImpl(const Context& dev_ctx,
                        const DenseTensor* X,
                        const DenseTensor* Out,
                        const DenseTensor* dOut,
                        DenseTensor* dX,
                        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(
      dOut, errors::NotFound("The input DenseTensor dOut can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      dX, errors::NotFound("The output DenseTensor dX can not be nullptr"));
  if (!Out) {
    Out = dOut;  // 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 = dX;
  }

  dev_ctx.template Alloc<T>(dX);
  auto dout = phi::EigenVector<T>::Flatten(
      GET_DATA_SAFELY(dOut, "Input", "Out@GRAD", "ActivationGrad"));
  auto out = phi::EigenVector<T>::Flatten(
      GET_DATA_SAFELY(Out, "Input", "Out", "ActivationGrad"));
  auto dx = phi::EigenVector<T>::Flatten(
      GET_DATA_SAFELY(dX, "Input", "X@GRAD", "ActivationGrad"));
  auto x = phi::EigenVector<T>::Flatten(
      GET_DATA_SAFELY(X, "Input", "X", "ActivationGrad"));
  auto* place = dev_ctx.eigen_device();
  // use 32bit index to speed up computation
  bool use_32bit_index = out.size() < Eigen::NumTraits<int>::highest();
  bool is_gpu_place = paddle::platform::is_gpu_place(dev_ctx.GetPlace());
  if (use_32bit_index && is_gpu_place) {
    functor(*place,
            To32BitIndex(x),
            To32BitIndex(out),
            To32BitIndex(dout),
            To32BitIndex(dx));
  } else {
    functor(*place, x, out, dout, dx);
  }
}

template <typename T, typename Context, typename Functor>
void ActivationDoubleGradImpl(const Context& dev_ctx,
                              const DenseTensor* X,
                              const DenseTensor* Out,
                              const DenseTensor* ddX,
                              DenseTensor* dX,
                              DenseTensor* dOut,
                              DenseTensor* ddOut,
                              const Functor& functor) {
  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 = ddX;
  }
  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"));
  } else {
    VLOG(10) << "Inplace activation of Op Functor: " << typeid(Functor).name();
    Out = ddX;
  }

  if (ddOut) {
    dev_ctx.template Alloc<T>(ddOut);
  }
  if (dOut) {
    dev_ctx.template Alloc<T>(dOut);
  }
  if (dX) {
    dX->Resize(Out->dims());
    dev_ctx.template Alloc<T>(dX);
  }

  functor(dev_ctx, X, Out, ddX, ddOut, dOut, dX);
}

template <typename T, typename Context>
void ReluDoubleGradKernel(const Context& dev_ctx,
                          const DenseTensor& out,
                          const DenseTensor& ddx,
                          DenseTensor* ddout) {
  funcs::ReluGradGradFunctor<T> relu_double_grad_functor;
  ActivationDoubleGradImpl<T, Context, funcs::ReluGradGradFunctor<T>>(
      dev_ctx,
      nullptr,
      &out,
      &ddx,
      nullptr,
      nullptr,
      ddout,
      relu_double_grad_functor);
}

}  // namespace phi