math_function.h 5.0 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Q
qijun 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15

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. */

#pragma once
T
tensor-tang 已提交
16
#ifdef PADDLE_WITH_MKLML
Q
qijun 已提交
17 18 19 20 21 22 23 24 25 26 27 28
#include <mkl_cblas.h>
#include <mkl_lapacke.h>
#include <mkl_vml_functions.h>
#endif

#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#include <lapacke.h>
#endif

#ifndef LAPACK_FOUND
extern "C" {
Y
Yu Yang 已提交
29
#include <cblas.h>  // NOLINT
Q
qijun 已提交
30 31 32 33 34 35 36 37 38 39 40 41
int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
                   int* ipiv);
int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
                   int* ipiv);
int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
                   const int* ipiv);
int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
                   const int* ipiv);
}
#endif

#include <cmath>
Y
Yu Yang 已提交
42
#include <vector>
Q
qijun 已提交
43

Y
Yi Wang 已提交
44 45 46 47 48
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
Q
qijun 已提交
49 50 51 52 53

namespace paddle {
namespace operators {
namespace math {

Q
qijun 已提交
54 55
// Support continuous memory now
// If transA = N, and transB = N
M
Markus Kliegl 已提交
56
// Then matrixA: M * K, matrixB: K * N, matrixC : M * N
Q
qijun 已提交
57 58
// For more detailed info, please refer to
// http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html
Q
QI JUN 已提交
59 60
template <typename DeviceContext, typename T>
void gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA,
61 62
          const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
          const T alpha, const T* A, const T* B, const T beta, T* C);
Q
qijun 已提交
63

G
guosheng 已提交
64
// gemm wrapper with stride args for matrix uncontinuous in memory
Q
QI JUN 已提交
65 66 67 68 69
template <typename DeviceContext, typename T>
void gemm(const DeviceContext& context, const bool transA, const bool transB,
          const int M, const int N, const int K, const T alpha, const T* A,
          const int lda, const T* B, const int ldb, const T beta, T* C,
          const int ldc);
G
guosheng 已提交
70

Q
qijun 已提交
71
// matrix multiply with continuous memory
Q
QI JUN 已提交
72 73 74 75
template <typename DeviceContext, typename T>
void matmul(const DeviceContext& context, const framework::Tensor& matrix_a,
            bool trans_a, const framework::Tensor& matrix_b, bool trans_b,
            T alpha, framework::Tensor* matrix_out, T beta);
Q
qijun 已提交
76

M
Markus Kliegl 已提交
77
// Batched gemm
Q
QI JUN 已提交
78 79 80 81
template <typename DeviceContext, typename T>
void batched_gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA,
                  const CBLAS_TRANSPOSE transB, const int M, const int N,
                  const int K, const T alpha, const T* A, const T* B,
Y
Yu Yang 已提交
82 83
                  const T beta, T* C, const int batchCount,
                  const int64_t strideA, const int64_t strideB);
Q
QI JUN 已提交
84 85 86 87 88 89 90 91 92 93 94

template <typename DeviceContext, typename T>
void gemv(const DeviceContext& context, const bool trans_a, const int M,
          const int N, const T alpha, const T* A, const T* B, const T beta,
          T* C);

template <typename DeviceContext, typename T>
void axpy(const DeviceContext& context, const int n, const T alpha, const T* x,
          T* y);

template <typename DeviceContext, typename T, int Rank>
95
struct Transpose {
Q
QI JUN 已提交
96 97
  void operator()(const DeviceContext& context, const framework::Tensor& in,
                  framework::Tensor* out, const std::vector<int>& axis);
98 99
};

Q
QI JUN 已提交
100
template <typename DeviceContext, typename T>
Q
qijun 已提交
101
struct SetConstant {
Q
QI JUN 已提交
102 103
  void operator()(const DeviceContext& context, framework::Tensor* tensor,
                  T num);
Q
qijun 已提交
104 105
};

106 107 108 109 110 111 112
template <typename Place>
void set_constant_with_place(const platform::DeviceContext& context,
                             framework::Tensor* tensor, float value);

void set_constant(const platform::DeviceContext& context,
                  framework::Tensor* tensor, float value);

Q
QI JUN 已提交
113
template <typename DeviceContext, typename T>
114
struct RowwiseAdd {
Q
QI JUN 已提交
115 116
  void operator()(const DeviceContext& context, const framework::Tensor& input,
                  const framework::Tensor& vec, framework::Tensor* output);
117 118
};

Q
QI JUN 已提交
119
template <typename DeviceContext, typename T>
120
struct ColwiseSum {
Q
QI JUN 已提交
121 122
  void operator()(const DeviceContext& context, const framework::Tensor& input,
                  framework::Tensor* vec);
123 124
};

C
chengduoZH 已提交
125 126 127 128 129 130 131 132 133 134 135 136
template <typename DeviceContext, typename T>
struct RowwiseSum {
  void operator()(const DeviceContext& context, const framework::Tensor& input,
                  framework::Tensor* vec);
};

template <typename DeviceContext, typename T>
struct RowwiseMean {
  void operator()(const DeviceContext& context, const framework::Tensor& input,
                  framework::Tensor* vec);
};

Q
qijun 已提交
137 138 139
}  // namespace math
}  // namespace operators
}  // namespace paddle