elementwise_div_op.h 4.2 KB
Newer Older
G
gongweibao 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   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. */

F
fengjiayi 已提交
15 16
#pragma once

17
#include "paddle/operators/elementwise_op_function.h"
G
gongweibao 已提交
18 19 20 21

namespace paddle {
namespace operators {

Q
QI JUN 已提交
22
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
23
class ElementwiseDivKernel : public framework::OpKernel<T> {
G
gongweibao 已提交
24 25
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
Q
QI JUN 已提交
26
    ElementwiseCompute<EigenDivFunctor, DeviceContext, T>(ctx);
G
gongweibao 已提交
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
  }
};

template <typename T>
struct ElementwiseDivGradFunctor {
  template <typename Device, typename X, typename Y, typename Z, typename dX,
            typename dY, typename dZ>
  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
    auto y_e = framework::EigenVector<T>::Flatten(*y);
    auto z_e = framework::EigenVector<T>::Flatten(*z);
    auto dz_e = framework::EigenVector<T>::Flatten(*dz);

    if (dx) {
      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
      dx_e.device(d) = dz_e / y_e;
    }

    if (dy) {
      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
      dy_e.device(d) = -1.0 * dz_e * z_e / y_e;
    }
  }
};

template <typename T>
struct ElementwiseDivBroadCastGradFunctor {
  template <typename Device, typename X, typename Y, typename Z, typename dX,
            typename dY, typename dZ, typename Pre, typename N>
  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
    auto x_e = framework::EigenVector<T>::Flatten(*x);
    auto y_e = framework::EigenVector<T>::Flatten(*y);
    auto dz_e = framework::EigenVector<T>::Flatten(*dz);

    auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));

    if (dx) {
      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
      dx_e.device(d) = dz_e / y_e_bcast;
    }

    if (dy) {
      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
      dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
                           .reshape(Eigen::DSizes<int, 2>(pre, n))
                           .sum(Eigen::array<int, 1>{{0}});
    }
  }
};

template <typename T>
struct ElementwiseDivBroadCast2GradFunctor {
  template <typename Device, typename X, typename Y, typename Z, typename dX,
            typename dY, typename dZ, typename Pre, typename N, typename Post>
  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
                  Post post) {
    auto x_e = framework::EigenVector<T>::Flatten(*x);
    auto y_e = framework::EigenVector<T>::Flatten(*y);
    auto dz_e = framework::EigenVector<T>::Flatten(*dz);

    auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
    if (dx) {
      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
      dx_e.device(d) = dz_e / y_e_bcast;
    }

    if (dy) {
      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
      dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
                           .reshape(Eigen::DSizes<int, 3>(pre, n, post))
                           .sum(Eigen::array<int, 2>{{0, 2}});
    }
  }
};

Q
QI JUN 已提交
105
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
106
class ElementwiseDivGradKernel : public framework::OpKernel<T> {
G
gongweibao 已提交
107 108
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
Q
QI JUN 已提交
109
    ElementwiseGradCompute<DeviceContext, T, ElementwiseDivGradFunctor<T>,
G
gongweibao 已提交
110 111 112 113 114 115 116 117
                           ElementwiseDivGradFunctor<T>,
                           ElementwiseDivBroadCastGradFunctor<T>,
                           ElementwiseDivBroadCast2GradFunctor<T>>(ctx);
  }
};

}  // namespace operators
}  // namespace paddle