elementwise_add_op.h 4.0 KB
Newer Older
G
gongweibao 已提交
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

L
Luo Tao 已提交
3 4 5
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
G
gongweibao 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
G
gongweibao 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
G
gongweibao 已提交
14

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 {

C
chengduoZH 已提交
22 23
template <typename T>
struct AddFunctor {
C
chengduoZH 已提交
24
  inline HOSTDEVICE T operator()(T a, T b) const { return a + b; }
C
chengduoZH 已提交
25 26
};

Q
QI JUN 已提交
27
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
28
class ElementwiseAddKernel : public framework::OpKernel<T> {
G
gongweibao 已提交
29 30
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
C
chengduoZH 已提交
31 32 33 34 35 36 37 38
    using Tensor = framework::Tensor;

    auto* x = ctx.Input<Tensor>("X");
    auto* y = ctx.Input<Tensor>("Y");
    auto* z = ctx.Output<Tensor>("Out");
    z->mutable_data<T>(ctx.GetPlace());
    int axis = ctx.Attr<int>("axis");
    ElementwiseComputeEx<AddFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
G
gongweibao 已提交
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
  }
};

template <typename T>
struct ElementwiseAddGradFunctor {
  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 dz_e = framework::EigenVector<T>::Flatten(*dz);
    if (dx) {
      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
      dx_e.device(d) = dz_e;
    }
    if (dy) {
      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
      dy_e.device(d) = dz_e;
    }
  }
};

template <typename T>
struct ElementwiseAddBroadCastGradFunctor {
  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 dz_e = framework::EigenVector<T>::Flatten(*dz);
    if (dx) {
      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
      dx_e.device(d) = dz_e;
    }

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

template <typename T>
struct ElementwiseAddBroadCast2GradFunctor {
  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 dz_e = framework::EigenVector<T>::Flatten(*dz);
    if (dx) {
      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
      dx_e.device(d) = dz_e;
    }

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

Q
QI JUN 已提交
98
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
99
class ElementwiseAddGradKernel : public framework::OpKernel<T> {
G
gongweibao 已提交
100 101
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
C
chengduoZH 已提交
102 103 104 105 106 107 108 109 110
    using Tensor = framework::Tensor;

    auto* x = ctx.Input<Tensor>("X");
    auto* y = ctx.Input<Tensor>("Y");
    auto* out = ctx.Input<Tensor>("Out");
    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
    int axis = ctx.Attr<int>("axis");
Q
QI JUN 已提交
111
    ElementwiseGradCompute<DeviceContext, T, ElementwiseAddGradFunctor<T>,
G
gongweibao 已提交
112
                           ElementwiseAddBroadCastGradFunctor<T>,
C
chengduoZH 已提交
113 114
                           ElementwiseAddBroadCast2GradFunctor<T>>(
        ctx, x, y, out, dout, axis, dx, dy);
G
gongweibao 已提交
115 116 117 118 119
  }
};

}  // namespace operators
}  // namespace paddle