math_function.cu 11.3 KB
Newer Older
Q
qijun 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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. */

#include "paddle/operators/math/math_function.h"
Q
qijun 已提交
16

Q
qijun 已提交
17 18 19 20 21
namespace paddle {
namespace operators {
namespace math {

template <>
22 23
void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
                                     const CBLAS_TRANSPOSE transA,
Q
qijun 已提交
24 25 26
                                     const CBLAS_TRANSPOSE transB, const int M,
                                     const int N, const int K,
                                     const float alpha, const float* A,
27 28
                                     const float* B, const float beta,
                                     float* C) {
Q
qijun 已提交
29 30
  // Note that cublas follows fortran order, so the order is different from
  // the cblas convention.
Q
qijun 已提交
31 32
  int lda = (transA == CblasNoTrans) ? K : M;
  int ldb = (transB == CblasNoTrans) ? N : K;
Q
qijun 已提交
33
  cublasOperation_t cuTransA =
Q
qijun 已提交
34
      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
Q
qijun 已提交
35
  cublasOperation_t cuTransB =
Q
qijun 已提交
36
      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
Q
qijun 已提交
37

Q
qijun 已提交
38
  PADDLE_ENFORCE(platform::dynload::cublasSgemm(
39 40
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
Q
qijun 已提交
41
      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
Q
qijun 已提交
42 43 44
}

template <>
45 46
void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
                                      const CBLAS_TRANSPOSE transA,
Q
qijun 已提交
47 48 49 50
                                      const CBLAS_TRANSPOSE transB, const int M,
                                      const int N, const int K,
                                      const double alpha, const double* A,
                                      const double* B, const double beta,
51
                                      double* C) {
Q
qijun 已提交
52 53
  // Note that cublas follows fortran order, so the order is different from
  // the cblas convention.
Q
qijun 已提交
54 55
  int lda = (transA == CblasNoTrans) ? K : M;
  int ldb = (transB == CblasNoTrans) ? N : K;
Q
qijun 已提交
56
  cublasOperation_t cuTransA =
Q
qijun 已提交
57
      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
Q
qijun 已提交
58
  cublasOperation_t cuTransB =
Q
qijun 已提交
59
      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
Q
qijun 已提交
60
  PADDLE_ENFORCE(platform::dynload::cublasDgemm(
61 62
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
Q
qijun 已提交
63
      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
Q
qijun 已提交
64 65
}

G
guosheng 已提交
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101
template <>
void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
                                     const bool transA, const bool transB,
                                     const int M, const int N, const int K,
                                     const float alpha, const float* A,
                                     const int lda, const float* B,
                                     const int ldb, const float beta, float* C,
                                     const int ldc) {
  // Note that cublas follows fortran order, so the order is different from
  // the cblas convention.
  cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T;
  cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T;
  PADDLE_ENFORCE(platform::dynload::cublasSgemm(
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc));
}

template <>
void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
                                      const bool transA, const bool transB,
                                      const int M, const int N, const int K,
                                      const double alpha, const double* A,
                                      const int lda, const double* B,
                                      const int ldb, const double beta,
                                      double* C, const int ldc) {
  // Note that cublas follows fortran order, so the order is different from
  // the cblas convention.
  cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T;
  cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T;
  PADDLE_ENFORCE(platform::dynload::cublasDgemm(
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc));
}

Q
qijun 已提交
102
template <>
103 104 105 106
void matmul<platform::GPUPlace, float>(
    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
    framework::Tensor* matrix_out, float beta) {
Q
qijun 已提交
107 108 109 110 111 112 113 114 115
  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_gpu_place(matrix_a.place()) &&
                     platform::is_gpu_place(matrix_b.place()) &&
                     platform::is_gpu_place(matrix_out->place()),
Q
qijun 已提交
116
                 "Matrix must all be in GPUPlace");
Q
qijun 已提交
117

Q
qijun 已提交
118 119 120
  int M = dim_out[0];
  int N = dim_out[1];
  int K = (trans_a == false) ? dim_a[1] : dim_a[0];
Q
qijun 已提交
121

Q
qijun 已提交
122 123
  CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
  CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
Q
qijun 已提交
124

Q
qijun 已提交
125
  gemm<platform::GPUPlace, float>(
126 127
      context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
      matrix_b.data<float>(), beta, matrix_out->data<float>());
Q
qijun 已提交
128 129 130
}

template <>
131 132 133 134
void matmul<platform::GPUPlace, double>(
    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
    framework::Tensor* matrix_out, double beta) {
Q
qijun 已提交
135 136 137 138 139 140 141 142 143
  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_gpu_place(matrix_a.place()) &&
                     platform::is_gpu_place(matrix_b.place()) &&
                     platform::is_gpu_place(matrix_out->place()),
Q
qijun 已提交
144
                 "Matrix must all be in GPUPlace");
Q
qijun 已提交
145

Q
qijun 已提交
146 147 148 149 150 151
  int M = dim_out[0];
  int N = dim_out[1];
  int K = (trans_a == false) ? dim_a[1] : dim_a[0];

  CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
  CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
Q
qijun 已提交
152

Q
qijun 已提交
153
  gemm<platform::GPUPlace, double>(
154 155
      context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
      matrix_b.data<double>(), beta, matrix_out->data<double>());
Q
qijun 已提交
156
}
Q
qijun 已提交
157

M
Markus Kliegl 已提交
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
template <>
void batched_gemm<platform::GPUPlace, float>(
    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
    const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
    const float alpha, const float* A, const float* B, const float beta,
    float* C, const int batchCount, const int strideA, const int strideB) {
  // Note that cublas follows fortran order, so the order is different from
  // the cblas convention.
  int lda = (transA == CblasNoTrans) ? K : M;
  int ldb = (transB == CblasNoTrans) ? N : K;
  int ldc = N;
  cublasOperation_t cuTransA =
      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
  cublasOperation_t cuTransB =
      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
  const int strideC = M * N;

  PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched(
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, strideB, A, lda, strideA,
      &beta, C, ldc, strideC, batchCount));
}

template <>
void batched_gemm<platform::GPUPlace, double>(
    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
    const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
    const double alpha, const double* A, const double* B, const double beta,
    double* C, const int batchCount, const int strideA, const int strideB) {
  // Note that cublas follows fortran order, so the order is different from
  // the cblas convention.
  int lda = (transA == CblasNoTrans) ? K : M;
  int ldb = (transB == CblasNoTrans) ? N : K;
  int ldc = N;
  cublasOperation_t cuTransA =
      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
  cublasOperation_t cuTransB =
      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
  const int strideC = M * N;

  PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched(
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, strideB, A, lda, strideA,
      &beta, C, ldc, strideC, batchCount));
}

206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
template <>
void gemv<platform::GPUPlace, float>(const platform::DeviceContext& context,
                                     const bool trans_a, const int M,
                                     const int N, const float alpha,
                                     const float* A, const float* B,
                                     const float beta, float* C) {
  cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N;

  PADDLE_ENFORCE(platform::dynload::cublasSgemv(
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
      cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1));
}

template <>
void gemv<platform::GPUPlace, double>(const platform::DeviceContext& context,
                                      const bool trans_a, const int M,
                                      const int N, const double alpha,
                                      const double* A, const double* B,
                                      const double beta, double* C) {
  cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N;
  PADDLE_ENFORCE(platform::dynload::cublasDgemv(
      reinterpret_cast<const platform::CUDADeviceContext&>(context)
          .cublas_handle(),
      cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1));
}

233
template struct SetConstant<platform::GPUPlace, float>;
Q
qijun 已提交
234

235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258
struct TensorSetConstant {
  TensorSetConstant(const platform::DeviceContext& context,
                    framework::Tensor* tensor, float value)
      : context_(context), tensor_(tensor), value_(value) {}

  template <typename T>
  void operator()() const {
    SetConstant<platform::GPUPlace, T> functor;
    functor(context_, tensor_, static_cast<T>(value_));
  }

  const platform::DeviceContext& context_;
  framework::Tensor* tensor_;
  float value_;
};

template <>
void set_constant_with_place<platform::GPUPlace>(
    const platform::DeviceContext& context, framework::Tensor* tensor,
    float value) {
  framework::VisitDataType(framework::ToDataType(tensor->type()),
                           TensorSetConstant(context, tensor, value));
}

Q
qijun 已提交
259 260 261
}  // namespace math
}  // namespace operators
}  // namespace paddle