cudnn.h 6.1 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

Y
Yu Yang 已提交
26 27
extern std::once_flag cudnn_dso_flag;
extern void* cudnn_dso_handle;
28
extern bool HasCUDNN();
Q
qijun 已提交
29 30 31

#ifdef PADDLE_USE_DSO

32
extern void EnforceCUDNNLoaded(const char* fn_name);
Y
Yu Yang 已提交
33
#define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)                    \
Q
qijun 已提交
34 35 36 37 38
  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,                               \
39
                     paddle::platform::dynload::GetCUDNNDsoHandle, \
Q
qijun 已提交
40
                     &cudnn_dso_handle);                           \
41
      EnforceCUDNNLoaded(#__name);                                 \
Q
qijun 已提交
42 43 44
      void* p_##__name = dlsym(cudnn_dso_handle, #__name);         \
      return reinterpret_cast<cudnn_func>(p_##__name)(args...);    \
    }                                                              \
Y
Yu Yang 已提交
45 46
  };                                                               \
  extern struct DynLoad__##__name __name
Q
qijun 已提交
47 48 49

#else

Y
Yu Yang 已提交
50
#define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)                  \
Q
qijun 已提交
51 52 53 54 55
  struct DynLoad__##__name {                                     \
    template <typename... Args>                                  \
    auto operator()(Args... args) -> decltype(__name(args...)) { \
      return __name(args...);                                    \
    }                                                            \
Y
Yu Yang 已提交
56 57
  };                                                             \
  extern DynLoad__##__name __name
Q
qijun 已提交
58 59 60 61 62 63 64

#endif

/**
 * include all needed cudnn functions in HPPL
 * different cudnn version has different interfaces
 **/
Y
Yu Yang 已提交
65 66 67
#define CUDNN_DNN_ROUTINE_EACH(__macro)             \
  __macro(cudnnSetTensor4dDescriptor);              \
  __macro(cudnnSetTensor4dDescriptorEx);            \
68 69
  __macro(cudnnSetTensorNdDescriptor);              \
  __macro(cudnnGetTensorNdDescriptor);              \
Y
Yu Yang 已提交
70 71 72 73 74 75
  __macro(cudnnGetConvolutionNdForwardOutputDim);   \
  __macro(cudnnGetConvolutionForwardAlgorithm);     \
  __macro(cudnnCreateTensorDescriptor);             \
  __macro(cudnnDestroyTensorDescriptor);            \
  __macro(cudnnCreateFilterDescriptor);             \
  __macro(cudnnSetFilter4dDescriptor);              \
76 77
  __macro(cudnnSetFilterNdDescriptor);              \
  __macro(cudnnGetFilterNdDescriptor);              \
Y
Yu Yang 已提交
78
  __macro(cudnnSetPooling2dDescriptor);             \
79 80
  __macro(cudnnSetPoolingNdDescriptor);             \
  __macro(cudnnGetPoolingNdDescriptor);             \
Y
Yu Yang 已提交
81 82 83 84 85 86
  __macro(cudnnDestroyFilterDescriptor);            \
  __macro(cudnnCreateConvolutionDescriptor);        \
  __macro(cudnnCreatePoolingDescriptor);            \
  __macro(cudnnDestroyPoolingDescriptor);           \
  __macro(cudnnSetConvolution2dDescriptor);         \
  __macro(cudnnDestroyConvolutionDescriptor);       \
87 88
  __macro(cudnnSetConvolutionNdDescriptor);         \
  __macro(cudnnGetConvolutionNdDescriptor);         \
Q
Qiao Longfei 已提交
89
  __macro(cudnnDeriveBNTensorDescriptor);           \
Y
Yu Yang 已提交
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
  __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(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)

#define CUDNN_DNN_ROUTINE_EACH_R2(__macro) \
  __macro(cudnnAddTensor);                 \
  __macro(cudnnConvolutionBackwardData);   \
  __macro(cudnnConvolutionBackwardFilter);
CUDNN_DNN_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
Q
qijun 已提交
111 112 113

// APIs available after R3:
#if CUDNN_VERSION >= 3000
Y
Yu Yang 已提交
114 115 116 117 118 119
#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro)           \
  __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize); \
  __macro(cudnnGetConvolutionBackwardDataAlgorithm);       \
  __macro(cudnnGetConvolutionBackwardFilterAlgorithm);     \
  __macro(cudnnGetConvolutionBackwardDataWorkspaceSize);
CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
Q
qijun 已提交
120 121 122 123
#endif

// APIs available after R4:
#if CUDNN_VERSION >= 4007
Y
Yu Yang 已提交
124 125 126 127 128
#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro)    \
  __macro(cudnnBatchNormalizationForwardTraining);  \
  __macro(cudnnBatchNormalizationForwardInference); \
  __macro(cudnnBatchNormalizationBackward);
CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
Q
qijun 已提交
129 130 131 132
#endif

// APIs in R5
#if CUDNN_VERSION >= 5000
Y
Yu Yang 已提交
133 134 135 136 137 138
#define CUDNN_DNN_ROUTINE_EACH_R5(__macro)  \
  __macro(cudnnCreateActivationDescriptor); \
  __macro(cudnnSetActivationDescriptor);    \
  __macro(cudnnGetActivationDescriptor);    \
  __macro(cudnnDestroyActivationDescriptor);
CUDNN_DNN_ROUTINE_EACH_R5(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
Q
qijun 已提交
139 140
#endif

武毅 已提交
141 142 143 144 145 146
#if CUDNN_VERSION >= 7001
#define CUDNN_DNN_ROUTINE_EACH_R7(__macro) \
  __macro(cudnnSetConvolutionGroupCount);
CUDNN_DNN_ROUTINE_EACH_R7(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
#endif

Q
qijun 已提交
147
}  // namespace dynload
Q
qijun 已提交
148
}  // namespace platform
Q
qijun 已提交
149
}  // namespace paddle