mul_op.h 3.8 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

19
#include "paddle/framework/op_registry.h"
20 21 22 23

namespace paddle {
namespace operators {

24 25
using Tensor = framework::Tensor;

Q
qijun 已提交
26
template <typename Place, typename T>
27
class MulKernel : public framework::OpKernel<T> {
28
 public:
29
  void Compute(const framework::ExecutionContext& context) const override {
30 31
    const Tensor* x = context.Input<Tensor>("X");
    const Tensor* y = context.Input<Tensor>("Y");
F
fengjiayi 已提交
32
    Tensor* z = context.Output<Tensor>("Out");
33 34
    const Tensor x_matrix =
        x->dims().size() > 2
35
            ? framework::ReshapeToMatrix(
F
fengjiayi 已提交
36
                  *x, context.template Attr<int>("x_num_col_dims"))
37 38 39
            : *x;
    const Tensor y_matrix =
        y->dims().size() > 2
40
            ? framework::ReshapeToMatrix(
F
fengjiayi 已提交
41
                  *y, context.template Attr<int>("y_num_col_dims"))
42
            : *y;
43

F
fengjiayi 已提交
44
    z->mutable_data<T>(context.GetPlace());
45 46 47 48
    auto z_dim = z->dims();
    if (z_dim.size() != 2) {
      z->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
    }
49 50
    math::matmul<Place, T>(context.device_context(), x_matrix, false, y_matrix,
                           false, 1, z, 0);
51 52 53
    if (z_dim.size() != 2) {
      z->Resize(z_dim);
    }
54 55
  }
};
D
dongzhihong 已提交
56

D
dongzhihong 已提交
57
template <typename Place, typename T>
58
class MulGradKernel : public framework::OpKernel<T> {
D
dongzhihong 已提交
59 60
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
F
fengjiayi 已提交
61 62
    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");
63 64
    const Tensor* x = ctx.Input<Tensor>("X");
    const Tensor* y = ctx.Input<Tensor>("Y");
65 66 67 68 69 70
    const Tensor x_matrix = x->dims().size() > 2
                                ? framework::ReshapeToMatrix(*x, x_num_col_dims)
                                : *x;
    const Tensor y_matrix = y->dims().size() > 2
                                ? framework::ReshapeToMatrix(*y, y_num_col_dims)
                                : *y;
71
    const Tensor* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
D
dongzhihong 已提交
72

73 74 75 76 77
    Tensor dout_mat;
    dout_mat.ShareDataWith(*dout);
    dout_mat.Resize({framework::flatten_to_2d(x->dims(), x_num_col_dims)[0],
                     framework::flatten_to_2d(y->dims(), y_num_col_dims)[1]});

78 79
    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
    Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
80 81
    if (dx) {
      dx->mutable_data<T>(ctx.GetPlace());
82 83 84
      Tensor dx_matrix = dx->dims().size() > 2
                             ? framework::ReshapeToMatrix(*dx, x_num_col_dims)
                             : *dx;
85

86
      // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
87 88
      math::matmul<Place, T>(ctx.device_context(), dout_mat, false, y_matrix,
                             true, 1, &dx_matrix, 0);
89 90 91
    }
    if (dy) {
      dy->mutable_data<T>(ctx.GetPlace());
92 93 94
      Tensor dy_matrix = dy->dims().size() > 2
                             ? framework::ReshapeToMatrix(*dy, y_num_col_dims)
                             : *dy;
95
      // dy = x' * dout. dy K x N, dout : M x N, x : M x K
96 97
      math::matmul<Place, T>(ctx.device_context(), x_matrix, true, dout_mat,
                             false, 1, &dy_matrix, 0);
98
    }
D
dongzhihong 已提交
99 100 101
  }
};

102 103
}  // namespace operators
}  // namespace paddle
反馈
建议
客服 返回
顶部