math_function_impl.h 6.2 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

15
#pragma once
16
#include <vector>
Y
Yi Wang 已提交
17 18
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/operators/math/math_function.h"
19 20 21 22 23

namespace paddle {
namespace operators {
namespace math {

Q
QI JUN 已提交
24 25 26 27
template <typename DeviceContext, typename T>
void SetConstant<DeviceContext, T>::operator()(const DeviceContext& context,
                                               framework::Tensor* tensor,
                                               T num) {
28
  auto t = framework::EigenVector<T>::Flatten(*tensor);
Q
QI JUN 已提交
29
  t.device(*context.eigen_device()) = t.constant(static_cast<T>(num));
30 31
}

Q
QI JUN 已提交
32 33 34
template <typename DeviceContext, typename T, int Rank>
void Transpose<DeviceContext, T, Rank>::operator()(
    const DeviceContext& context, const framework::Tensor& in,
35 36 37 38 39 40 41 42 43 44
    framework::Tensor* out, const std::vector<int>& axis) {
  Eigen::array<int, Rank> permute;
  for (int i = 0; i < Rank; i++) {
    permute[i] = axis[i];
  }
  auto in_dim = in.dims();
  auto out_dim = out->dims();

  auto eigen_in = framework::EigenTensor<T, Rank>::From(in);
  auto eigen_out = framework::EigenTensor<T, Rank>::From(*out);
Q
QI JUN 已提交
45
  auto* dev = context.eigen_device();
46 47
  eigen_out.device(*dev) = eigen_in.shuffle(permute);
}
48

Q
QI JUN 已提交
49 50 51
template <typename DeviceContext, typename T>
void ColwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
                                              const framework::Tensor& input,
Y
Yu Yang 已提交
52
                                              framework::Tensor* out) {
53 54
  auto in_dims = input.dims();
  auto size = input.numel() / in_dims[0];
Y
Yu Yang 已提交
55
  PADDLE_ENFORCE_EQ(out->numel(), size);
56 57

  auto in = framework::EigenMatrix<T>::From(input);
Y
Yu Yang 已提交
58 59 60
  auto vec = framework::EigenVector<T>::Flatten(*out);

  vec.device(*context.eigen_device()) = in.sum(Eigen::array<int, 1>({{0}}));
61
}
62

Y
Yu Yang 已提交
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
// Specialize for CPU, since Eigen implement a general reduce. However,
// colwise-sum can be easily implemented. General reduce has a huge overhead in
// CPU
template <typename T>
class ColwiseSum<platform::CPUDeviceContext, T> {
 public:
  void operator()(const platform::CPUDeviceContext& context,
                  const framework::Tensor& input, framework::Tensor* out) {
    auto& in_dims = input.dims();
    auto height = in_dims[0];
    auto size = in_dims[1];
    PADDLE_ENFORCE_EQ(out->numel(), size);

    T* out_buf = out->mutable_data<T>(out->place());
    const T* in_buf = input.data<T>();

Q
qiaolongfei 已提交
79 80
    for (size_t i = 0; i < static_cast<size_t>(height); ++i) {
      for (size_t j = 0; j < static_cast<size_t>(size); ++j) {
Y
Yu Yang 已提交
81 82 83 84 85 86 87 88 89 90
        if (i == 0) {
          out_buf[j] = in_buf[i * size + j];
        } else {
          out_buf[j] += in_buf[i * size + j];
        }
      }
    }
  }
};

C
chengduoZH 已提交
91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
template <typename DeviceContext, typename T>
void RowwiseMean<DeviceContext, T>::operator()(const DeviceContext& context,
                                               const framework::Tensor& input,
                                               framework::Tensor* out) {
  auto in_dims = input.dims();
  PADDLE_ENFORCE_EQ(in_dims.size(), 2U);
  PADDLE_ENFORCE_EQ(out->numel(), in_dims[0]);

  auto in = framework::EigenMatrix<T>::From(input);
  auto vec = framework::EigenVector<T>::Flatten(*out);

  vec.device(*context.eigen_device()) = in.mean(Eigen::array<int, 1>({{1}}));
}
// TODO(zcd): Following ColwiseSum format, need to confirm.
// Specialize for CPU, since Eigen implement a general reduce. However,
// rowwise-sum can be easily implemented. General reduce has a huge overhead in
// CPU
template <typename T>
class RowwiseMean<platform::CPUDeviceContext, T> {
 public:
  void operator()(const platform::CPUDeviceContext& context,
                  const framework::Tensor& input, framework::Tensor* out) {
    auto& in_dims = input.dims();
    PADDLE_ENFORCE_EQ(in_dims.size(), 2U);
    auto height = in_dims[0];
    auto size = in_dims[1];
    PADDLE_ENFORCE_EQ(out->numel(), height);
    auto inv_size = 1.0 / size;
    T* out_buf = out->mutable_data<T>(out->place());
    const T* in_buf = input.data<T>();

    for (size_t i = 0; i < static_cast<size_t>(height); ++i) {
      T sum = 0;
      for (size_t j = 0; j < static_cast<size_t>(size); ++j) {
        sum += in_buf[i * size + j];
      }
      out_buf[i] = sum * inv_size;
    }
  }
};

template <typename DeviceContext, typename T>
void RowwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
                                              const framework::Tensor& input,
                                              framework::Tensor* out) {
  auto in_dims = input.dims();
  PADDLE_ENFORCE_EQ(in_dims.size(), 2U);
  PADDLE_ENFORCE_EQ(out->numel(), in_dims[0]);

  auto in = framework::EigenMatrix<T>::From(input);
  auto vec = framework::EigenVector<T>::Flatten(*out);

  vec.device(*context.eigen_device()) = in.sum(Eigen::array<int, 1>({{1}}));
}
// TODO(zcd): Following ColwiseSum format, need to confirm.
// Specialize for CPU, since Eigen implement a general reduce. However,
// rowwise-sum can be easily implemented. General reduce has a huge overhead in
// CPU
template <typename T>
class RowwiseSum<platform::CPUDeviceContext, T> {
 public:
  void operator()(const platform::CPUDeviceContext& context,
                  const framework::Tensor& input, framework::Tensor* out) {
    auto& in_dims = input.dims();
    PADDLE_ENFORCE_EQ(in_dims.size(), 2U);
    auto height = in_dims[0];
    auto size = in_dims[1];
W
weixing02 已提交
158
    PADDLE_ENFORCE_EQ(out->numel(), height);
C
chengduoZH 已提交
159 160 161 162 163 164 165 166 167 168 169 170 171 172

    T* out_buf = out->mutable_data<T>(out->place());
    const T* in_buf = input.data<T>();

    for (size_t i = 0; i < static_cast<size_t>(height); ++i) {
      T sum = 0;
      for (size_t j = 0; j < static_cast<size_t>(size); ++j) {
        sum += in_buf[i * size + j];
      }
      out_buf[i] = sum;
    }
  }
};

173 174 175
}  // namespace math
}  // namespace operators
}  // namespace paddle