elementwise_add_op.h 4.9 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 {

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
};

G
gongweibao 已提交
27
template <typename Place, 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
    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());
C
chengduoZH 已提交
37 38
    TransformFunctor<AddFunctor<T>, T, Place> functor(
        x, y, z, ctx.device_context(), AddFunctor<T>());
C
chengduoZH 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

    auto x_dims = x->dims();
    auto y_dims = y->dims();
    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
                      "Rank of first input must >= rank of second input.");

    if (x_dims == y_dims) {
      functor.Run();
      return;
    }

    int axis = ctx.Attr<int>("axis");
    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
    PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
                   "Axis should be in range [0, x_dims)");

    int pre, n, post;
    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
    if (post == 1) {
      functor.RunRowWise(n, pre);
      return;
    } else {
      functor.RunMidWise(n, pre, post);
      return;
    }
G
gongweibao 已提交
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
  }
};

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 ElementwiseAddOneGradFunctor {
  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.sum();
    }
  }
};

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}});
    }
  }
};

template <typename Place, typename T>
Y
Yu Yang 已提交
141
class ElementwiseAddGradKernel : public framework::OpKernel<T> {
G
gongweibao 已提交
142 143 144 145 146 147 148 149 150 151 152
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    ElementwiseGradCompute<Place, T, ElementwiseAddGradFunctor<T>,
                           ElementwiseAddOneGradFunctor<T>,
                           ElementwiseAddBroadCastGradFunctor<T>,
                           ElementwiseAddBroadCast2GradFunctor<T>>(ctx);
  }
};

}  // namespace operators
}  // namespace paddle