blas_impl.h 9.8 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// 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
Y
Yu Yang 已提交
15
#include <vector>
Y
Yu Yang 已提交
16 17 18 19 20 21 22 23 24
#include "paddle/fluid/operators/math/math_function.h"

namespace paddle {
namespace operators {
namespace math {

template <typename T>
struct CBlas;

25
#ifdef PADDLE_WITH_MKLML
Y
Yu Yang 已提交
26 27
template <>
struct CBlas<float> {
Y
Yu Yang 已提交
28 29
  template <typename... ARGS>
  static void GEMM(ARGS... args) {
30
    platform::dynload::cblas_sgemm(args...);
Y
Yu Yang 已提交
31
  }
Y
Yu Yang 已提交
32

T
tensor-tang 已提交
33 34 35 36 37 38
#ifdef PADDLE_WITH_LIBXSMM
  template <typename... ARGS>
  static void SMM_GEMM(ARGS... args) {
    libxsmm_sgemm(args...);
  }
#endif
Y
Yu Yang 已提交
39 40
  template <typename... ARGS>
  static void AXPY(ARGS... args) {
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
    platform::dynload::cblas_saxpy(args...);
  }

  template <typename... ARGS>
  static void VCOPY(ARGS... args) {
    platform::dynload::cblas_scopy(args...);
  }

  template <typename... ARGS>
  static void GEMV(ARGS... args) {
    platform::dynload::cblas_sgemv(args...);
  }

  template <typename... ARGS>
  static void GEMM_BATCH(ARGS... args) {
    platform::dynload::cblas_sgemm_batch(args...);
Y
Yu Yang 已提交
57 58
  }

59 60
  template <typename... ARGS>
  static void VADD(ARGS... args) {
61 62 63 64 65 66 67 68 69 70 71
    platform::dynload::vsAdd(args...);
  }
};

template <>
struct CBlas<double> {
  template <typename... ARGS>
  static void GEMM(ARGS... args) {
    platform::dynload::cblas_dgemm(args...);
  }

T
tensor-tang 已提交
72 73 74 75 76 77
#ifdef PADDLE_WITH_LIBXSMM
  template <typename... ARGS>
  static void SMM_GEMM(ARGS... args) {
    libxsmm_dgemm(args...);
  }
#endif
78 79 80
  template <typename... ARGS>
  static void AXPY(ARGS... args) {
    platform::dynload::cblas_daxpy(args...);
81 82 83 84
  }

  template <typename... ARGS>
  static void VCOPY(ARGS... args) {
85
    platform::dynload::cblas_dcopy(args...);
86 87
  }

Y
Yu Yang 已提交
88 89
  template <typename... ARGS>
  static void GEMV(ARGS... args) {
90
    platform::dynload::cblas_dgemv(args...);
Y
Yu Yang 已提交
91 92 93 94
  }

  template <typename... ARGS>
  static void GEMM_BATCH(ARGS... args) {
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
    platform::dynload::cblas_dgemm_batch(args...);
  }

  template <typename... ARGS>
  static void VADD(ARGS... args) {
    platform::dynload::vdAdd(args...);
  }
};

#else

template <>
struct CBlas<float> {
  template <typename... ARGS>
  static void GEMM(ARGS... args) {
    cblas_sgemm(args...);
  }

  template <typename... ARGS>
  static void AXPY(ARGS... args) {
    cblas_saxpy(args...);
  }

  template <typename... ARGS>
  static void VCOPY(ARGS... args) {
    cblas_scopy(args...);
  }

  template <typename... ARGS>
  static void GEMV(ARGS... args) {
    cblas_sgemv(args...);
Y
Yu Yang 已提交
126
  }
Y
Yu Yang 已提交
127 128 129 130
};

template <>
struct CBlas<double> {
Y
Yu Yang 已提交
131 132 133 134
  template <typename... ARGS>
  static void GEMM(ARGS... args) {
    cblas_dgemm(args...);
  }
Y
Yu Yang 已提交
135 136 137 138 139 140

  template <typename... ARGS>
  static void AXPY(ARGS... args) {
    cblas_daxpy(args...);
  }

141 142 143 144 145
  template <typename... ARGS>
  static void VCOPY(ARGS... args) {
    cblas_dcopy(args...);
  }

Y
Yu Yang 已提交
146 147 148 149
  template <typename... ARGS>
  static void GEMV(ARGS... args) {
    cblas_dgemv(args...);
  }
Y
Yu Yang 已提交
150
};
151
#endif
Y
Yu Yang 已提交
152 153
template <>
struct CBlas<platform::float16> {
Y
Yu Yang 已提交
154
  static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); }
T
tensor-tang 已提交
155 156 157
  static void SMM_GEMM(...) {
    PADDLE_THROW("float16 SMM_GEMM not supported on CPU");
  }
Y
Yu Yang 已提交
158 159 160 161 162
#ifdef PADDLE_WITH_MKLML
  static void GEMM_BATCH(...) {
    PADDLE_THROW("float16 GEMM_BATCH not supported on CPU");
  }
#endif
Y
Yu Yang 已提交
163 164 165 166
};

template <>
template <typename T>
Y
Yu Yang 已提交
167 168 169 170
void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
                                            CBLAS_TRANSPOSE transB, int M,
                                            int N, int K, T alpha, const T *A,
                                            const T *B, T beta, T *C) const {
T
tensor-tang 已提交
171 172 173 174 175 176 177 178 179 180
#ifdef PADDLE_WITH_LIBXSMM
  if (M * N * K < 128 * 128 * 128 && transA == CblasNoTrans &&
      transB == CblasNoTrans) {
    // refer to https://github.com/hfp/libxsmm/blob/master/README.md
    // Note: SMM use ColMajor
    const char transa = 'N';
    const char transb = 'N';
    const int lda = M;
    const int ldb = K;
    const int ldc = M;
181
    CBlas<T>::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &ldb, A, &lda,
T
tensor-tang 已提交
182 183 184 185 186 187 188 189 190 191 192
                       &beta, C, &ldc);
  } else {
#endif
    int lda = (transA == CblasNoTrans) ? K : M;
    int ldb = (transB == CblasNoTrans) ? N : K;
    int ldc = N;
    CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B,
                   ldb, beta, C, ldc);
#ifdef PADDLE_WITH_LIBXSMM
  }
#endif
Y
Yu Yang 已提交
193 194 195 196
}

template <>
template <typename T>
Y
Yu Yang 已提交
197 198 199 200
void Blas<platform::CPUDeviceContext>::GEMM(bool transA, bool transB, int M,
                                            int N, int K, T alpha, const T *A,
                                            int lda, const T *B, int ldb,
                                            T beta, T *C, int ldc) const {
Y
Yu Yang 已提交
201 202 203 204 205
  CBlas<T>::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
                 transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
                 lda, B, ldb, beta, C, ldc);
}

Y
Yu Yang 已提交
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
template <typename DeviceContext>
template <typename T>
void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a, bool trans_a,
                                 const framework::Tensor &mat_b, bool trans_b,
                                 T alpha, framework::Tensor *mat_out,
                                 T beta) const {
  auto dim_a = mat_a.dims();
  auto dim_b = mat_b.dims();
  auto dim_out = mat_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(
      mat_a.place() == mat_b.place() && mat_a.place() == mat_out->place(),
      "The places of matrices must be same");

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

  CBLAS_TRANSPOSE transA = !trans_a ? CblasNoTrans : CblasTrans;
  CBLAS_TRANSPOSE transB = !trans_b ? CblasNoTrans : CblasTrans;

  this->GEMM(transA, transB, M, N, K, alpha, mat_a.data<T>(), mat_b.data<T>(),
             beta, mat_out->data<T>());
}

Y
Yu Yang 已提交
232 233 234 235 236 237 238
template <>
template <typename T>
void Blas<platform::CPUDeviceContext>::AXPY(int n, T alpha, const T *x,
                                            T *y) const {
  CBlas<T>::AXPY(n, alpha, x, 1, y, 1);
}

239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
template <>
template <typename T>
void Blas<platform::CPUDeviceContext>::VCOPY(int n, const T *x, T *y) const {
  CBlas<T>::VCOPY(n, x, 1, y, 1);
}

template <>
template <typename T>
void Blas<platform::CPUDeviceContext>::VADD(int n, const T *x, const T *y,
                                            T *z) const {
#ifdef PADDLE_WITH_MKLML
  CBlas<T>::VADD(n, x, y, z);
#else
  this->template VCOPY<T>(n, y, z);
  this->template AXPY<T>(n, 1., x, z);
#endif
}

Y
Yu Yang 已提交
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
template <>
template <typename T>
void Blas<platform::CPUDeviceContext>::GEMV(bool trans_a, int M, int N, T alpha,
                                            const T *A, const T *B, T beta,
                                            T *C) const {
  CBLAS_TRANSPOSE transA = !trans_a ? CblasNoTrans : CblasTrans;
  CBlas<T>::GEMV(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1);
}

template <>
template <typename T>
void Blas<platform::CPUDeviceContext>::BatchedGEMM(
    CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
    T alpha, const T *A, const T *B, T beta, T *C, int batchCount,
    int64_t strideA, int64_t strideB) const {
#ifdef PADDLE_WITH_MKLML
  int lda = (transA == CblasNoTrans) ? K : M;
  int ldb = (transB == CblasNoTrans) ? N : K;
  int ldc = N;
  auto a_array = std::vector<const T *>(batchCount);
  auto b_array = std::vector<const T *>(batchCount);
  auto c_array = std::vector<T *>(batchCount);
  for (int k = 0; k < batchCount; ++k) {
    a_array[k] = &A[k * strideA];
    b_array[k] = &B[k * strideB];
    c_array[k] = &C[k * M * N];
  }

  CBlas<T>::GEMM_BATCH(CblasRowMajor, &transA, &transB, &M, &N, &K, &alpha,
                       a_array.data(), &lda, b_array.data(), &ldb, &beta,
                       c_array.data(), &ldc, 1 /* group_count */, &batchCount);
#else
  for (int k = 0; k < batchCount; ++k) {
Y
yuyang18 已提交
290 291 292
    auto *Ak = &A[k * strideA];
    auto *Bk = &B[k * strideB];
    auto *Ck = &C[k * M * N];
Y
Yu Yang 已提交
293 294 295 296 297
    this->template GEMM<T>(transA, transB, M, N, K, alpha, Ak, Bk, beta, Ck);
  }
#endif
}

Y
Yu Yang 已提交
298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322
template <typename DeviceContext>
template <typename T>
void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a,
                                 const MatDescriptor &dim_a,
                                 const framework::Tensor &mat_b,
                                 const MatDescriptor &dim_b, T alpha,
                                 framework::Tensor *mat_out, T beta) const {
  PADDLE_ENFORCE_EQ(dim_a.width_, dim_b.height_);
  CBLAS_TRANSPOSE transA = !dim_a.trans_ ? CblasNoTrans : CblasTrans;
  CBLAS_TRANSPOSE transB = !dim_b.trans_ ? CblasNoTrans : CblasTrans;
  if (dim_a.batch_size_ == 0 && dim_b.batch_size_ == 0) {
    this->template GEMM<T>(transA, transB, dim_a.height_, dim_b.width_,
                           dim_a.width_, alpha, mat_a.data<T>(),
                           mat_b.data<T>(), beta, mat_out->data<T>());
  } else {
    PADDLE_ENFORCE(dim_a.batch_size_ == dim_b.batch_size_ ||
                   dim_a.batch_size_ == 0 || dim_b.batch_size_ == 0);
    this->template BatchedGEMM<T>(
        transA, transB, dim_a.height_, dim_b.width_, dim_a.width_, alpha,
        mat_a.data<T>(), mat_b.data<T>(), beta, mat_out->data<T>(),
        dim_a.batch_size_ == 0 ? dim_b.batch_size_ : dim_a.batch_size_,
        dim_a.stride_, dim_b.stride_);
  }
}

Y
Yu Yang 已提交
323 324 325
}  // namespace math
}  // namespace operators
}  // namespace paddle