cudnn.h 5.6 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
#include <cudnn.h>
18 19 20
#include <dlfcn.h>
#include <mutex>
#include "paddle/platform/dynload/dynamic_loader.h"
Q
qijun 已提交
21 22

namespace paddle {
Q
qijun 已提交
23
namespace platform {
Q
qijun 已提交
24
namespace dynload {
Q
qijun 已提交
25 26 27 28 29 30

std::once_flag cudnn_dso_flag;
void* cudnn_dso_handle = nullptr;

#ifdef PADDLE_USE_DSO

Q
qijun 已提交
31 32 33 34 35 36 37 38 39 40 41
#define DYNAMIC_LOAD_CUDNN_WRAP(__name)                            \
  struct DynLoad__##__name {                                       \
    template <typename... Args>                                    \
    auto operator()(Args... args) -> decltype(__name(args...)) {   \
      using cudnn_func = decltype(__name(args...)) (*)(Args...);   \
      std::call_once(cudnn_dso_flag,                               \
                     paddle::platform::dynload::GetCudnnDsoHandle, \
                     &cudnn_dso_handle);                           \
      void* p_##__name = dlsym(cudnn_dso_handle, #__name);         \
      return reinterpret_cast<cudnn_func>(p_##__name)(args...);    \
    }                                                              \
Q
qijun 已提交
42 43 44 45 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 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
  } __name; /* struct DynLoad__##__name */

#else

#define DYNAMIC_LOAD_CUDNN_WRAP(__name)                          \
  struct DynLoad__##__name {                                     \
    template <typename... Args>                                  \
    auto operator()(Args... args) -> decltype(__name(args...)) { \
      return __name(args...);                                    \
    }                                                            \
  } __name; /* struct DynLoad__##__name */

#endif

/**
 * include all needed cudnn functions in HPPL
 * different cudnn version has different interfaces
 **/
// clang-format off
#define CUDNN_DNN_ROUTINE_EACH(__macro)                   \
  __macro(cudnnSetTensor4dDescriptor)                     \
  __macro(cudnnSetTensor4dDescriptorEx)                   \
  __macro(cudnnGetConvolutionNdForwardOutputDim)          \
  __macro(cudnnGetConvolutionForwardAlgorithm)            \
  __macro(cudnnCreateTensorDescriptor)                    \
  __macro(cudnnDestroyTensorDescriptor)                   \
  __macro(cudnnCreateFilterDescriptor)                    \
  __macro(cudnnSetFilter4dDescriptor)                     \
  __macro(cudnnSetPooling2dDescriptor)                    \
  __macro(cudnnDestroyFilterDescriptor)                   \
  __macro(cudnnCreateConvolutionDescriptor)               \
  __macro(cudnnCreatePoolingDescriptor)                   \
  __macro(cudnnDestroyPoolingDescriptor)                  \
  __macro(cudnnSetConvolution2dDescriptor)                \
  __macro(cudnnDestroyConvolutionDescriptor)              \
  __macro(cudnnCreate)                                    \
  __macro(cudnnDestroy)                                   \
  __macro(cudnnSetStream)                                 \
  __macro(cudnnActivationForward)                         \
  __macro(cudnnConvolutionForward)                        \
  __macro(cudnnConvolutionBackwardBias)                   \
  __macro(cudnnGetConvolutionForwardWorkspaceSize)        \
  __macro(cudnnTransformTensor)                           \
  __macro(cudnnPoolingForward)                            \
  __macro(cudnnPoolingBackward)                           \
  __macro(cudnnSoftmaxBackward)                           \
  __macro(cudnnSoftmaxForward)                            \
  __macro(cudnnGetVersion)                                \
  __macro(cudnnGetErrorString)
CUDNN_DNN_ROUTINE_EACH(DYNAMIC_LOAD_CUDNN_WRAP)

#define CUDNN_DNN_ROUTINE_EACH_R2(__macro)                \
  __macro(cudnnAddTensor)                                 \
  __macro(cudnnConvolutionBackwardData)                   \
  __macro(cudnnConvolutionBackwardFilter)
CUDNN_DNN_ROUTINE_EACH_R2(DYNAMIC_LOAD_CUDNN_WRAP)

// APIs available after R3:
#if CUDNN_VERSION >= 3000
#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro)              \
  __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize)     \
  __macro(cudnnGetConvolutionBackwardDataAlgorithm)           \
  __macro(cudnnGetConvolutionBackwardFilterAlgorithm)         \
  __macro(cudnnGetConvolutionBackwardDataWorkspaceSize)
CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DYNAMIC_LOAD_CUDNN_WRAP)
#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R3
#endif


// APIs available after R4:
#if CUDNN_VERSION >= 4007
#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro)             \
  __macro(cudnnBatchNormalizationForwardTraining)            \
  __macro(cudnnBatchNormalizationForwardInference)           \
  __macro(cudnnBatchNormalizationBackward)
CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DYNAMIC_LOAD_CUDNN_WRAP)
#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R4
#endif

// APIs in R5
#if CUDNN_VERSION >= 5000
#define CUDNN_DNN_ROUTINE_EACH_R5(__macro)                    \
  __macro(cudnnCreateActivationDescriptor)                    \
  __macro(cudnnSetActivationDescriptor)                       \
  __macro(cudnnGetActivationDescriptor)                       \
  __macro(cudnnDestroyActivationDescriptor)
CUDNN_DNN_ROUTINE_EACH_R5(DYNAMIC_LOAD_CUDNN_WRAP)
#undef CUDNN_DNN_ROUTINE_EACH_R5
#endif

#undef CUDNN_DNN_ROUTINE_EACH
// clang-format on
Q
qijun 已提交
134
}  // namespace dynload
Q
qijun 已提交
135
}  // namespace platform
Q
qijun 已提交
136
}  // namespace paddle