Skip to content

port dyload to platform #2722

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Jul 4, 2017
Merged

port dyload to platform #2722

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
3ba7a73
add dynamic_load
QiJune Jul 3, 2017
a30754b
test device_context
QiJune Jul 3, 2017
a77fcef
fix cuda compile error
QiJune Jul 3, 2017
ed18647
finish test
QiJune Jul 4, 2017
76b7be4
add deps for dyload cc_library
QiJune Jul 4, 2017
9eeabe9
follow comments
QiJune Jul 4, 2017
3567ea6
move to dynload directory
QiJune Jul 4, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions paddle/platform/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,5 @@
add_subdirectory(dynload)

nv_test(cuda_test SRCS cuda_test.cu)

cc_library(place SRCS place.cc)
Expand Down
10 changes: 10 additions & 0 deletions paddle/platform/cuda.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,16 @@ int GetDeviceCount(void) {
return count;
}

int GetCurrentDeviceId(void) {
int device_id;
throw_on_error(cudaGetDevice(&device_id), "cudaGetDevice failed");
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we only use PADDLE_ENFORCE and not throw thrust::system_error?

But it seems that we need to move enforce to platform if we want to use PADDLE_ENFORCE in platform

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more. 

#define PADDLE_ENFORCE(condition, ...) \
  do {                                 \
    if (UNLIKELY(!(condition))) {      \
      PADDLE_THROW(__VA_ARGS__);       \
    }                                  \
  } while (0)

UNLIKELY can not handle the following situation since different libs have a different return struct.

#pragma once

#include <sstream>
#include <stdexcept>
#include <string>

#ifndef PADDLE_ONLY_CPU

#include <cublas_v2.h>
#include <cudnn.h>
#include <curand.h>
#include <thrust/system/cuda/error.h>
#include <thrust/system_error.h>

#endif  // PADDLE_ONLY_CPU

namespace paddle {
namespace platform {

#ifndef PADDLE_ONLY_CPU

inline void throw_on_error(cudaError_t e, const char* message) {
  if (e) {
    throw thrust::system_error(e, thrust::cuda_category(), message);
  }
}

inline void throw_on_error(curandStatus_t stat, const char* message) {
  if (stat != CURAND_STATUS_SUCCESS) {
    throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(),
                               message);
  }
}

inline void throw_on_error(cudnnStatus_t stat, const char* message) {
  std::stringstream ss;
  if (stat == CUDNN_STATUS_SUCCESS) {
    return;
  } else {
    ss << cudnnGetErrorString(stat);
    ss << ", " << message;
    throw std::runtime_error(ss.str());
  }
}

inline void throw_on_error(cublasStatus_t stat, const char* message) {
  std::stringstream ss;
  if (stat == CUBLAS_STATUS_SUCCESS) {
    return;
  } else if (stat == CUBLAS_STATUS_NOT_INITIALIZED) {
    ss << "CUBLAS: not initialized";
  } else if (stat == CUBLAS_STATUS_ALLOC_FAILED) {
    ss << "CUBLAS: alloc failed";
  } else if (stat == CUBLAS_STATUS_INVALID_VALUE) {
    ss << "CUBLAS: invalid value";
  } else if (stat == CUBLAS_STATUS_ARCH_MISMATCH) {
    ss << "CUBLAS: arch mismatch";
  } else if (stat == CUBLAS_STATUS_MAPPING_ERROR) {
    ss << "CUBLAS: mapping error";
  } else if (stat == CUBLAS_STATUS_EXECUTION_FAILED) {
    ss << "CUBLAS: execution failed";
  } else if (stat == CUBLAS_STATUS_INTERNAL_ERROR) {
    ss << "CUBLAS: internal error";
  } else if (stat == CUBLAS_STATUS_NOT_SUPPORTED) {
    ss << "CUBLAS: not supported";
  } else if (stat == CUBLAS_STATUS_LICENSE_ERROR) {
    ss << "CUBLAS: license error";
  }
  ss << ", " << message;
  throw std::runtime_error(ss.str());
}

inline void throw_on_error(cublasStatus_t stat) {
  const char* message = "";
  throw_on_error(stat, message);
}

#endif  // PADDLE_ONLY_CPU

inline void throw_on_error(int stat, const char* message) {
  if (stat) {
    throw std::runtime_error(message + (", stat = " + std::to_string(stat)));
  }
}

}  // namespace platform
}  // namespace paddle

return device_id;
}

void SetDeviceId(int device_id) {
throw_on_error(cudaSetDevice(device_id), "cudaSetDevice failed");
}

} // namespace platform
} // namespace paddle

Expand Down
1 change: 1 addition & 0 deletions paddle/platform/dynload/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags)
104 changes: 104 additions & 0 deletions paddle/platform/dynload/cublas.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,104 @@
/* 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

#include <cublas_v2.h>
#include "paddle/platform/dynamic_loader.h"

namespace paddle {
namespace platform {
namespace dynload {

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
#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...); \
} \
} __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
#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
#else
#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
#endif
} // namespace dynload
} // namespace platform
} // namespace paddle
134 changes: 134 additions & 0 deletions paddle/platform/dynload/cudnn.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,134 @@
/* 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

#include <cudnn.h>
#include "paddle/platform/dynamic_loader.h"

namespace paddle {
namespace platform {
namespace dynload {

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

#ifdef PADDLE_USE_DSO

#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...); \
} \
} __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
} // namespace dynload
} // namespace platform
} // namespace paddle
65 changes: 65 additions & 0 deletions paddle/platform/dynload/curand.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,65 @@
/* 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

#include <curand.h>
#include "paddle/platform/dynamic_loader.h"

namespace paddle {
namespace platform {
namespace dynload {
std::once_flag curand_dso_flag;
void *curand_dso_handle = nullptr;
#ifdef PADDLE_USE_DSO
#define DYNAMIC_LOAD_CURAND_WRAP(__name) \
struct DynLoad__##__name { \
template <typename... Args> \
curandStatus_t operator()(Args... args) { \
typedef curandStatus_t (*curandFunc)(Args...); \
std::call_once(curand_dso_flag, \
paddle::platform::dynload::GetCurandDsoHandle, \
&curand_dso_handle); \
void *p_##__name = dlsym(curand_dso_handle, #__name); \
return reinterpret_cast<curandFunc>(p_##__name)(args...); \
} \
} __name; /* struct DynLoad__##__name */
#else
#define DYNAMIC_LOAD_CURAND_WRAP(__name) \
struct DynLoad__##__name { \
template <typename... Args> \
curandStatus_t operator()(Args... args) { \
return __name(args...); \
} \
} __name; /* struct DynLoad__##__name */
#endif

/* include all needed curand functions in HPPL */
// clang-format off
#define CURAND_RAND_ROUTINE_EACH(__macro) \
__macro(curandCreateGenerator) \
__macro(curandSetStream) \
__macro(curandSetPseudoRandomGeneratorSeed)\
__macro(curandGenerateUniform) \
__macro(curandGenerateUniformDouble) \
__macro(curandDestroyGenerator)
// clang-format on

CURAND_RAND_ROUTINE_EACH(DYNAMIC_LOAD_CURAND_WRAP)

#undef CURAND_RAND_ROUTINE_EACH
#undef DYNAMIC_LOAD_CURAND_WRAP
} // namespace dynload
} // namespace platform
} // namespace paddle
Loading