math_function_impl.h 9.6 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 <memory>
17
#include <vector>
Y
Yi Wang 已提交
18
#include "paddle/fluid/framework/data_type.h"
19
#include "paddle/pten/kernels/funcs/math_function.h"
20

21 22
namespace pten {
namespace funcs {
23

24
using paddle::framework::To32BitIndex;
25

Q
QI JUN 已提交
26
template <typename DeviceContext, typename T>
27 28
void SetConstant<DeviceContext, T>::operator()(
    const DeviceContext& context, paddle::framework::Tensor* tensor, T num) {
29 30
  bool xpu_place = false;
#ifdef PADDLE_WITH_XPU
31
  if (paddle::platform::is_xpu_place(context.GetPlace())) {
32
    xpu_place = true;
33 34
    pten::VisitDataType(
        tensor->dtype(),
35
        TensorSetConstantXPU<T>(tensor, num, context.GetPlace()));
36 37 38
  }
#endif
  if (!xpu_place) {
39
    auto t = paddle::framework::EigenVector<T>::Flatten(*tensor);
40 41
    t.device(*context.eigen_device()) = t.constant(static_cast<T>(num));
  }
42 43
}

Q
QI JUN 已提交
44 45
template <typename DeviceContext, typename T, int Rank>
void Transpose<DeviceContext, T, Rank>::operator()(
46 47 48 49
    const DeviceContext& context,
    const paddle::framework::Tensor& in,
    paddle::framework::Tensor* out,
    const std::vector<int>& axis) {
50 51 52 53
  Eigen::array<int, Rank> permute;
  for (int i = 0; i < Rank; i++) {
    permute[i] = axis[i];
  }
54 55
  auto eigen_in = paddle::framework::EigenTensor<T, Rank>::From(in);
  auto eigen_out = paddle::framework::EigenTensor<T, Rank>::From(*out);
Q
QI JUN 已提交
56
  auto* dev = context.eigen_device();
57 58
  // use 32bit index to speed up computation
  bool use_32bit_index = eigen_out.size() < Eigen::NumTraits<int>::highest();
59
  bool is_gpu_place = paddle::platform::is_gpu_place(context.GetPlace());
60 61 62 63 64 65
  if (use_32bit_index && is_gpu_place) {
    To32BitIndex(eigen_out).device(*dev) =
        To32BitIndex(eigen_in).shuffle(permute);
  } else {
    eigen_out.device(*dev) = eigen_in.shuffle(permute);
  }
66
}
67

Q
QI JUN 已提交
68
template <typename DeviceContext, typename T>
69 70 71 72
void ColwiseSum<DeviceContext, T>::operator()(
    const DeviceContext& context,
    const paddle::framework::Tensor& input,
    paddle::framework::Tensor* out) {
73 74
  auto in_dims = input.dims();
  auto size = input.numel() / in_dims[0];
75 76 77
  PADDLE_ENFORCE_EQ(out->numel(),
                    size,
                    paddle::platform::errors::InvalidArgument(
78 79 80
                        "The size of output tensor "
                        "should be equal to the size of input tensor column"
                        " dimension. Expected output size=%d, but received %d",
81 82
                        size,
                        out->numel()));
83

84 85
  auto in = paddle::framework::EigenMatrix<T>::From(input);
  auto vec = paddle::framework::EigenVector<T>::Flatten(*out);
Y
Yu Yang 已提交
86 87

  vec.device(*context.eigen_device()) = in.sum(Eigen::array<int, 1>({{0}}));
88
}
89

Y
Yu Yang 已提交
90 91 92 93
// 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>
94
class ColwiseSum<paddle::platform::CPUDeviceContext, T> {
Y
Yu Yang 已提交
95
 public:
96 97 98
  void operator()(const paddle::platform::CPUDeviceContext& context,
                  const paddle::framework::Tensor& input,
                  paddle::framework::Tensor* out) {
Y
Yu Yang 已提交
99 100 101
    auto& in_dims = input.dims();
    auto height = in_dims[0];
    auto size = in_dims[1];
102
    PADDLE_ENFORCE_EQ(
103 104 105
        out->numel(),
        size,
        paddle::platform::errors::InvalidArgument(
106 107 108
            "The size of output tensor "
            "should be equal to the size of input tensor column"
            " dimension. Expected output size=%d, but received %d",
109 110
            size,
            out->numel()));
Y
Yu Yang 已提交
111 112 113 114

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

Q
qiaolongfei 已提交
115 116
    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
117 118 119 120 121 122 123 124 125 126
        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
127
template <typename DeviceContext, typename T>
128 129 130 131
void RowwiseMean<DeviceContext, T>::operator()(
    const DeviceContext& context,
    const paddle::framework::Tensor& input,
    paddle::framework::Tensor* out) {
C
chengduoZH 已提交
132
  auto in_dims = input.dims();
133 134 135 136 137 138 139 140 141
  PADDLE_ENFORCE_EQ(
      in_dims.size(),
      2U,
      paddle::platform::errors::InvalidArgument("The rank of input tensor "
                                                "should be 2, but received %d",
                                                in_dims.size()));
  PADDLE_ENFORCE_EQ(out->numel(),
                    in_dims[0],
                    paddle::platform::errors::InvalidArgument(
142 143 144
                        "The size of output tensor "
                        "should be equal to the size of input tensor row"
                        " dimension. Expected output size=%d, but received %d",
145 146
                        in_dims[0],
                        out->numel()));
C
chengduoZH 已提交
147

148 149
  auto in = paddle::framework::EigenMatrix<T>::From(input);
  auto vec = paddle::framework::EigenVector<T>::Flatten(*out);
C
chengduoZH 已提交
150 151 152 153 154 155 156 157

  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>
158
class RowwiseMean<paddle::platform::CPUDeviceContext, T> {
C
chengduoZH 已提交
159
 public:
160 161 162
  void operator()(const paddle::platform::CPUDeviceContext& context,
                  const paddle::framework::Tensor& input,
                  paddle::framework::Tensor* out) {
C
chengduoZH 已提交
163
    auto& in_dims = input.dims();
164 165 166 167 168 169
    PADDLE_ENFORCE_EQ(in_dims.size(),
                      2U,
                      paddle::platform::errors::InvalidArgument(
                          "The rank of input tensor "
                          "should be 2, but received %d",
                          in_dims.size()));
C
chengduoZH 已提交
170 171
    auto height = in_dims[0];
    auto size = in_dims[1];
172
    PADDLE_ENFORCE_EQ(
173 174 175
        out->numel(),
        height,
        paddle::platform::errors::InvalidArgument(
176 177 178
            "The size of output tensor "
            "should be equal to the size of input tensor row"
            " dimension. Expected output size=%d, but received %d",
179 180
            height,
            out->numel()));
C
chengduoZH 已提交
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
    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>
196 197 198 199
void RowwiseSum<DeviceContext, T>::operator()(
    const DeviceContext& context,
    const paddle::framework::Tensor& input,
    paddle::framework::Tensor* out) {
C
chengduoZH 已提交
200
  auto in_dims = input.dims();
201 202 203 204 205 206 207 208 209
  PADDLE_ENFORCE_EQ(
      in_dims.size(),
      2U,
      paddle::platform::errors::InvalidArgument("The rank of input tensor "
                                                "should be 2, but received %d",
                                                in_dims.size()));
  PADDLE_ENFORCE_EQ(out->numel(),
                    in_dims[0],
                    paddle::platform::errors::InvalidArgument(
210 211 212
                        "The size of output tensor "
                        "should be equal to the size of input tensor row"
                        " dimension. Expected output size=%d, but received %d",
213 214
                        in_dims[0],
                        out->numel()));
C
chengduoZH 已提交
215

216 217
  auto in = paddle::framework::EigenMatrix<T>::From(input);
  auto vec = paddle::framework::EigenVector<T>::Flatten(*out);
C
chengduoZH 已提交
218 219 220 221 222 223 224 225

  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>
226
class RowwiseSum<paddle::platform::CPUDeviceContext, T> {
C
chengduoZH 已提交
227
 public:
228 229 230
  void operator()(const paddle::platform::CPUDeviceContext& context,
                  const paddle::framework::Tensor& input,
                  paddle::framework::Tensor* out) {
C
chengduoZH 已提交
231
    auto& in_dims = input.dims();
232 233 234 235 236 237
    PADDLE_ENFORCE_EQ(in_dims.size(),
                      2U,
                      paddle::platform::errors::InvalidArgument(
                          "The rank of input tensor "
                          "should be 2, but received %d",
                          in_dims.size()));
C
chengduoZH 已提交
238 239
    auto height = in_dims[0];
    auto size = in_dims[1];
240
    PADDLE_ENFORCE_EQ(
241 242 243
        out->numel(),
        height,
        paddle::platform::errors::InvalidArgument(
244 245 246
            "The size of output tensor "
            "should be equal to the size of input tensor row"
            " dimension. Expected output size=%d, but received %d",
247 248
            height,
            out->numel()));
C
chengduoZH 已提交
249 250 251 252 253 254 255 256 257 258 259 260 261 262

    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;
    }
  }
};

263 264
}  // namespace funcs
}  // namespace pten