mul_op.h 3.7 KB
Newer Older
1 2 3
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   Licensed under the Apache License, Version 2.0 (the "License");
F
fengjiayi 已提交
4 5
   You may not use this file except in compliance with the License.
   You may obtain a copy of the License at
6 7 8 9 10

   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,
F
fengjiayi 已提交
11
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 13 14 15
   See the License for the specific language governing permissions and
   limitations under the License. */

#pragma once
Y
Yu Yang 已提交
16

Q
qijun 已提交
17
#include "paddle/operators/math/math_function.h"
Q
qijun 已提交
18

D
dongzhihong 已提交
19 20
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
21 22 23 24

namespace paddle {
namespace operators {

D
dongzhihong 已提交
25 26 27 28 29
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

Q
qijun 已提交
30
template <typename Place, typename T>
D
dongzhihong 已提交
31
class MulKernel : public framework::OpKernel {
32
 public:
D
dongzhihong 已提交
33
  void Compute(const framework::ExecutionContext& context) const override {
34 35
    const Tensor* x = context.Input<Tensor>("X");
    const Tensor* y = context.Input<Tensor>("Y");
F
fengjiayi 已提交
36
    Tensor* z = context.Output<Tensor>("Out");
37 38
    const Tensor x_matrix =
        x->dims().size() > 2
F
fengjiayi 已提交
39
            ? framework::ReshapeToMatrix<T>(
F
fengjiayi 已提交
40
                  *x, context.template Attr<int>("x_num_col_dims"))
41 42 43
            : *x;
    const Tensor y_matrix =
        y->dims().size() > 2
F
fengjiayi 已提交
44
            ? framework::ReshapeToMatrix<T>(
F
fengjiayi 已提交
45
                  *y, context.template Attr<int>("y_num_col_dims"))
46
            : *y;
47

F
fengjiayi 已提交
48
    z->mutable_data<T>(context.GetPlace());
49 50
    math::matmul<Place, T>(context.device_context(), x_matrix, false, y_matrix,
                           false, 1, z, 0);
51 52
  }
};
D
dongzhihong 已提交
53

D
dongzhihong 已提交
54 55 56 57
template <typename Place, typename T>
class MulGradKernel : public framework::OpKernel {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
F
fengjiayi 已提交
58 59
    int x_num_col_dims = ctx.template Attr<int>("x_num_col_dims");
    int y_num_col_dims = ctx.template Attr<int>("y_num_col_dims");
60 61 62
    const Tensor* x = ctx.Input<Tensor>("X");
    const Tensor* y = ctx.Input<Tensor>("Y");
    const Tensor x_matrix =
F
fengjiayi 已提交
63
        x->dims().size() > 2 ? framework::ReshapeToMatrix<T>(*x, x_num_col_dims)
64 65
                             : *x;
    const Tensor y_matrix =
F
fengjiayi 已提交
66
        y->dims().size() > 2 ? framework::ReshapeToMatrix<T>(*y, y_num_col_dims)
67 68
                             : *y;
    const Tensor* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
D
dongzhihong 已提交
69

70 71
    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
    Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
72 73
    if (dx) {
      dx->mutable_data<T>(ctx.GetPlace());
F
fengjiayi 已提交
74 75
      Tensor dx_matrix = dx->dims().size() > 2 ? framework::ReshapeToMatrix<T>(
                                                     *dx, x_num_col_dims)
76
                                               : *dx;
77
      // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
78 79
      math::matmul<Place, T>(ctx.device_context(), *dout, false, y_matrix, true,
                             1, &dx_matrix, 0);
80 81 82
    }
    if (dy) {
      dy->mutable_data<T>(ctx.GetPlace());
F
fengjiayi 已提交
83 84
      Tensor dy_matrix = dy->dims().size() > 2 ? framework::ReshapeToMatrix<T>(
                                                     *dy, y_num_col_dims)
85
                                               : *dy;
86
      // dy = x' * dout. dy K x N, dout : M x N, x : M x K
87 88
      math::matmul<Place, T>(ctx.device_context(), x_matrix, true, *dout, false,
                             1, &dy_matrix, 0);
89
    }
D
dongzhihong 已提交
90 91 92
  }
};

93 94
}  // namespace operators
}  // namespace paddle