math_function.cpp 5.9 KB
Newer Older
Z
zhaojiaying01 已提交
1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
朔-望's avatar
朔-望 已提交
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. */

Z
zhaojiaying01 已提交
15
#include "operators/math/math_function.h"
xiebaiyuan's avatar
xiebaiyuan 已提交
16
#include <cstring>
Z
zhaojiaying01 已提交
17
#include "operators/math/gemm.h"
朔-望's avatar
朔-望 已提交
18 19

namespace paddle_mobile {
朔-望's avatar
朔-望 已提交
20 21 22
namespace operators {
namespace math {

朔-望's avatar
朔-望 已提交
23
template <>
朔-望's avatar
朔-望 已提交
24
void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
朔-望's avatar
朔-望 已提交
25
                   const framework::Tensor &matrix_b, bool trans_b, float alpha,
26 27
                   framework::Tensor *matrix_out, float beta, bool relu,
                   float *bias) {
28 29 30
  auto dim_a = matrix_a.dims();
  auto dim_b = matrix_b.dims();
  auto dim_out = matrix_out->dims();
Z
zhaojiaying01 已提交
31 32 33 34 35 36 37 38 39 40
  //  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
  //  dim_out.size() ==
  //  2,
  //                 "The input and output of matmul be matrix");
  //
  //  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
  //                     platform::is_cpu_place(matrix_b.place())
  //                     &&
  //                     platform::is_cpu_place(matrix_out->place()),
  //                 "Matrix must all be in CPUPlace");
41 42 43

  int M = dim_out[0];
  int N = dim_out[1];
44
  int K = (!trans_a) ? dim_a[1] : dim_a[0];
45

46 47 48 49
#ifdef _OPENMP
  Sgemm_omp(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
            N, beta, matrix_out->data<float>(), N, relu, bias);
#else
Z
zhaojiaying01 已提交
50
  Sgemm(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(), N,
51
        beta, matrix_out->data<float>(), N, relu, bias);
52
#endif
53
}
朔-望's avatar
朔-望 已提交
54

朔-望's avatar
朔-望 已提交
55
template <>
56 57 58 59
void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
                         const framework::Tensor &matrix_b, bool trans_b,
                         float alpha, framework::Tensor *matrix_out, float beta,
                         bool relu, framework::Tensor *new_scale,
Y
yangfei 已提交
60
                         framework::Tensor *new_bias, int group, float *bias) {
61 62 63
  auto dim_a = matrix_a.dims();
  auto dim_b = matrix_b.dims();
  auto dim_out = matrix_out->dims();
Z
zhaojiaying01 已提交
64 65 66 67 68 69 70 71 72 73
  //  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
  //  dim_out.size() ==
  //  2,
  //                 "The input and output of matmul be matrix");
  //
  //  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
  //                     platform::is_cpu_place(matrix_b.place())
  //                     &&
  //                     platform::is_cpu_place(matrix_out->place()),
  //                 "Matrix must all be in CPUPlace");
74 75 76

  int M = dim_out[0];
  int N = dim_out[1];
77 78
  int K = (!trans_a) ? dim_a[1] : dim_a[0];

79 80 81 82
#ifdef _OPENMP
  SgemmWithBn_omp(M, N, K, alpha, matrix_a.data<float>(), K,
                  matrix_b.data<float>(), N, beta, matrix_out->data<float>(), N,
                  relu, new_scale->data<float>() + group,
Y
yangfei 已提交
83
                  new_bias->data<float>() + group, bias);
84
#else
Z
zhaojiaying01 已提交
85 86
  SgemmWithBn(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
              N, beta, matrix_out->data<float>(), N, relu,
Y
yangfei 已提交
87 88
              new_scale->data<float>() + group, new_bias->data<float>() + group,
              bias);
89 90
#endif
}
91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
void matmulWithPRelu(const framework::Tensor &matrix_a, bool trans_a,
                     const framework::Tensor &matrix_b, bool trans_b,
                     framework::Tensor *matrix_out, float *p, std::string mode,
                     float *bias, float *bias1) {
  auto dim_a = matrix_a.dims();
  auto dim_b = matrix_b.dims();
  auto dim_out = matrix_out->dims();
  //  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
  //  dim_out.size() ==
  //  2,
  //                 "The input and output of matmul be matrix");
  //
  //  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
  //                     platform::is_cpu_place(matrix_b.place())
  //                     &&
  //                     platform::is_cpu_place(matrix_out->place()),
  //                 "Matrix must all be in CPUPlace");

  int M = dim_out[0];
  int N = dim_out[1];
  int K = (!trans_a) ? dim_a[1] : dim_a[0];

#ifdef _OPENMP
114 115
  SgemmWithPRelu_omp(M, N, K, matrix_a.data<float>(), K, matrix_b.data<float>(),
                     N, matrix_out->data<float>(), N, p, mode, bias, bias1);
116 117 118 119 120 121
#else
  SgemmWithPRelu(M, N, K, matrix_a.data<float>(), K, matrix_b.data<float>(), N,
                 matrix_out->data<float>(), N, p, mode, bias, bias1);

#endif
}
朔-望's avatar
朔-望 已提交
122

xiebaiyuan's avatar
xiebaiyuan 已提交
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
template <typename T>
struct ClearTensor<CPU, T> {
  void operator()(framework::Tensor *tensor) {
    auto size = tensor->numel();
    auto *tensor_data = tensor->data<float>();
    memset((void *)tensor_data, 0, sizeof(T) * size);
  }
};

template <typename T>
struct RowwiseAdd<CPU, T> {
  void operator()(const framework::Tensor &input,
                  const framework::Tensor &vector, framework::Tensor *output) {
    auto in_dims = input.dims();
    auto size = input.numel() / in_dims[0];
    PADDLE_MOBILE_ENFORCE((vector.numel() == size),
                          "vector.numel() must be equal to size.");
    PADDLE_MOBILE_ENFORCE((output->dims() == in_dims),
                          "output->dims() must be equal to in_dims.");

    auto *input_data = input.data<float>();
    auto *out_data = output->data<float>();
    auto *vec_data = vector.data<float>();
    for (int64_t i = 0; i < in_dims[0]; ++i) {
      for (int64_t j = 0; j < size; ++j) {
        out_data[i * size + j] = input_data[i * size + j] + vec_data[j];
      }
    }
  }
};

template struct RowwiseAdd<CPU, float>;
template struct ClearTensor<CPU, float>;

朔-望's avatar
朔-望 已提交
157 158 159
}  // namespace math
}  // namespace operators
}  // namespace paddle_mobile