cublas.h 4.0 KB
Newer Older
Q
qijun 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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. */

#pragma once

Q
qijun 已提交
17 18 19 20
#include <cublas_v2.h>
#include "paddle/platform/dynamic_loader.h"

namespace paddle {
Q
qijun 已提交
21
namespace platform {
Q
qijun 已提交
22
namespace dynload {
Q
qijun 已提交
23 24 25 26 27 28 29 30 31 32 33 34

std::once_flag cublas_dso_flag;
void *cublas_dso_handle = nullptr;

/**
 * The following macro definition can generate structs
 * (for each function) to dynamic load cublas routine
 * via operator overloading.
 *
 * note: default dynamic linked libs
 */
#ifdef PADDLE_USE_DSO
Q
qijun 已提交
35 36 37 38 39 40 41 42 43 44 45
#define DYNAMIC_LOAD_CUBLAS_WRAP(__name)                            \
  struct DynLoad__##__name {                                        \
    template <typename... Args>                                     \
    cublasStatus_t operator()(Args... args) {                       \
      typedef cublasStatus_t (*cublasFunc)(Args...);                \
      std::call_once(cublas_dso_flag,                               \
                     paddle::platform::dynload::GetCublasDsoHandle, \
                     &cublas_dso_handle);                           \
      void *p_##__name = dlsym(cublas_dso_handle, #__name);         \
      return reinterpret_cast<cublasFunc>(p_##__name)(args...);     \
    }                                                               \
Q
qijun 已提交
46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
  } __name;  // struct DynLoad__##__name
#else
#define DYNAMIC_LOAD_CUBLAS_WRAP(__name)      \
  struct DynLoad__##__name {                  \
    template <typename... Args>               \
    cublasStatus_t operator()(Args... args) { \
      return __name(args...);                 \
    }                                         \
  } __name;  // struct DynLoad__##__name
#endif

#define DYNAMIC_LOAD_CUBLAS_V2_WRAP(__name) DYNAMIC_LOAD_CUBLAS_WRAP(__name)

// include all needed cublas functions in HPPL
// clang-format off
#define CUBLAS_BLAS_ROUTINE_EACH(__macro) \
  __macro(cublasSgemv)                    \
  __macro(cublasDgemv)                    \
  __macro(cublasSgemm)                    \
  __macro(cublasDgemm)                    \
  __macro(cublasSgeam)                    \
  __macro(cublasDgeam)                    \

DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasCreate)
DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasDestroy)
DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetStream)
DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetPointerMode)
DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasGetPointerMode)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgemmBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgemmBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasCgemmBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasZgemmBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetrfBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetriBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetrfBatched)
DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetriBatched)
CUBLAS_BLAS_ROUTINE_EACH(DYNAMIC_LOAD_CUBLAS_V2_WRAP)

#undef DYNAMIC_LOAD_CUBLAS_WRAP
#undef DYNAMIC_LOAD_CUBLAS_V2_WRAP
#undef CUBLAS_BLAS_ROUTINE_EACH

// clang-format on
#ifndef PADDLE_TYPE_DOUBLE
Q
qijun 已提交
90 91 92 93 94
#define CUBLAS_GEAM paddle::platform::dynload::cublasSgeam
#define CUBLAS_GEMV paddle::platform::dynload::cublasSgemv
#define CUBLAS_GEMM paddle::platform::dynload::cublasSgemm
#define CUBLAS_GETRF paddle::platform::dynload::cublasSgetrfBatched
#define CUBLAS_GETRI paddle::platform::dynload::cublasSgetriBatched
Q
qijun 已提交
95
#else
Q
qijun 已提交
96 97 98 99 100
#define CUBLAS_GEAM paddle::platform::dynload::cublasDgeam
#define CUBLAS_GEMV paddle::platform::dynload::cublasDgemv
#define CUBLAS_GEMM paddle::platform::dynload::cublasDgemm
#define CUBLAS_GETRF paddle::platform::dynload::cublasDgetrfBatched
#define CUBLAS_GETRI paddle::platform::dynload::cublasDgetriBatched
Q
qijun 已提交
101
#endif
Q
qijun 已提交
102
}  // namespace dynload
Q
qijun 已提交
103
}  // namespace platform
Q
qijun 已提交
104
}  // namespace paddle