提交 85806e75 编写于 作者: Q qijun

follow comments

上级 8ee50a35
...@@ -5,4 +5,11 @@ nv_test(cuda_test SRCS cuda_test.cu) ...@@ -5,4 +5,11 @@ nv_test(cuda_test SRCS cuda_test.cu)
cc_library(place SRCS place.cc) cc_library(place SRCS place.cc)
cc_test(place_test SRCS place_test.cc DEPS place glog gflags) cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
nv_test(device_context_test SRCS device_context_test.cc DEPS dynload_cuda dynamic_loader eigen3 place) IF(WITH_GPU)
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
ELSE()
set(GPU_CTX_DEPS)
ENDIF()
cc_library(device_context SRCS device_context.cc DEPS place eigen3 ${GPU_CTX_DEPS})
nv_test(device_context_test SRCS device_context_test.cc DEPS device_context glog gflags)
/* 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 "paddle/framework/enforce.h"
#include "paddle/platform/cuda.h"
#include "paddle/platform/dynload/cublas.h"
#include "paddle/platform/dynload/cudnn.h"
#include "paddle/platform/dynload/curand.h"
#define EIGEN_USE_GPU
#include "paddle/platform/device_context.h"
#include "paddle/platform/place.h"
#include "unsupported/Eigen/CXX11/Tensor"
namespace paddle {
namespace platform {
class GPUPlaceGuard {
public:
explicit GPUPlaceGuard(GPUPlace new_place) : previous_(GetCurrentDeviceId()) {
if (previous_ != new_place) {
paddle::platform::SetDeviceId(new_place.device);
}
}
~GPUPlaceGuard() { paddle::platform::SetDeviceId(previous_.device); }
private:
GPUPlace previous_;
};
class CUDADeviceContext : public DeviceContext {
public:
explicit CUDADeviceContext(const GPUPlace gpu_place) : gpu_place_(gpu_place) {
GPUPlaceGuard guard(gpu_place_);
paddle::platform::throw_on_error(cudaStreamCreate(&stream_),
"cudaStreamCreate failed");
eigen_stream_ = new Eigen::CudaStreamDevice(&stream_);
eigen_device_ = new Eigen::GpuDevice(eigen_stream_);
}
Place GetPlace() const override {
Place retv = GPUPlace();
return retv;
}
void Wait() {
paddle::platform::throw_on_error(cudaStreamSynchronize(stream_),
"cudaStreamSynchronize failed");
}
cudaStream_t stream() { return stream_; }
Eigen::GpuDevice eigen_device() { return *eigen_device_; }
cublasHandle_t cublas_handle() {
if (!blas_handle_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_) ==
CUBLAS_STATUS_SUCCESS,
"cublasCreate failed");
PADDLE_ENFORCE(paddle::platform::dynload::cublasSetStream(
blas_handle_, stream_) == CUBLAS_STATUS_SUCCESS,
"cublasSetStream failed");
}
return blas_handle_;
}
cudnnHandle_t cudnn_handle() {
if (!dnn_handle_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_) ==
CUDNN_STATUS_SUCCESS,
"cudnnCreate failed");
PADDLE_ENFORCE(paddle::platform::dynload::cudnnSetStream(
dnn_handle_, stream_) == CUDNN_STATUS_SUCCESS,
"cudnnSetStream failed");
}
return dnn_handle_;
}
curandGenerator_t curand_generator() {
if (!rand_generator_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::curandCreateGenerator(
&rand_generator_, CURAND_RNG_PSEUDO_DEFAULT) ==
CURAND_STATUS_SUCCESS,
"curandCreateGenerator failed");
PADDLE_ENFORCE(
paddle::platform::dynload::curandSetPseudoRandomGeneratorSeed(
rand_generator_, random_seed_) == CURAND_STATUS_SUCCESS,
"curandSetPseudoRandomGeneratorSeed failed");
PADDLE_ENFORCE(paddle::platform::dynload::curandSetStream(
rand_generator_, stream_) == CURAND_STATUS_SUCCESS,
"curandSetStream failed");
}
return rand_generator_;
}
~CUDADeviceContext() {
Wait();
if (blas_handle_) {
PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_) ==
CUBLAS_STATUS_SUCCESS,
"cublasDestroy failed");
}
if (dnn_handle_) {
PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_) ==
CUDNN_STATUS_SUCCESS,
"cudnnDestroy failed");
}
if (rand_generator_) {
PADDLE_ENFORCE(paddle::platform::dynload::curandDestroyGenerator(
rand_generator_) == CURAND_STATUS_SUCCESS,
"curandDestroyGenerator failed");
}
delete eigen_stream_;
delete eigen_device_;
paddle::platform::throw_on_error(cudaStreamDestroy(stream_),
"cudaStreamDestroy failed");
}
private:
GPUPlace gpu_place_;
cudaStream_t stream_;
Eigen::CudaStreamDevice* eigen_stream_;
Eigen::GpuDevice* eigen_device_;
cublasHandle_t blas_handle_{nullptr};
cudnnHandle_t dnn_handle_{nullptr};
int random_seed_;
curandGenerator_t rand_generator_{nullptr};
};
template <>
Eigen::GpuDevice DeviceContext::get_eigen_device<Eigen::GpuDevice>() {
return dynamic_cast<CUDADeviceContext*>(this)->eigen_device();
}
} // namespace platform
} // namespace paddle
/* 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. */
#include "paddle/platform/device_context.h"
namespace paddle {
namespace platform {
namespace dynload {
namespace dummy {
// Make DeviceContext A library.
int DUMMY_VAR_FOR_DEV_CTX = 0;
} // namespace dummy
} // namespace dynload
} // namespace platform
} // namespace paddle
\ No newline at end of file
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License"); Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License. you may not use this file except in compliance with the License.
You may obtain a copy of the License at You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0 http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS, distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
...@@ -13,9 +10,17 @@ See the License for the specific language governing permissions and ...@@ -13,9 +10,17 @@ See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#pragma once #pragma once
#include "paddle/framework/enforce.h" #include "paddle/framework/enforce.h"
#include "paddle/platform/place.h" #ifndef PADDLE_ONLY_CPU
#include "unsupported/Eigen/CXX11/Tensor" #include "paddle/platform/cuda.h"
#include "paddle/platform/dynload/cublas.h"
#include "paddle/platform/dynload/cudnn.h"
#include "paddle/platform/dynload/curand.h"
#define EIGEN_USE_GPU
#endif
#include <paddle/platform/place.h>
#include <unsupported/Eigen/CXX11/Tensor>
namespace paddle { namespace paddle {
namespace platform { namespace platform {
...@@ -23,11 +28,10 @@ namespace platform { ...@@ -23,11 +28,10 @@ namespace platform {
class DeviceContext { class DeviceContext {
public: public:
virtual ~DeviceContext() {} virtual ~DeviceContext() {}
virtual Place GetPlace() const = 0;
template <typename DeviceType> template <typename DeviceType>
inline DeviceType get_eigen_device(); inline DeviceType get_eigen_device();
virtual Place GetPlace() const = 0;
}; };
class CPUDeviceContext : public DeviceContext { class CPUDeviceContext : public DeviceContext {
...@@ -52,5 +56,138 @@ template <> ...@@ -52,5 +56,138 @@ template <>
Eigen::DefaultDevice DeviceContext::get_eigen_device<Eigen::DefaultDevice>() { Eigen::DefaultDevice DeviceContext::get_eigen_device<Eigen::DefaultDevice>() {
return dynamic_cast<CPUDeviceContext*>(this)->eigen_device(); return dynamic_cast<CPUDeviceContext*>(this)->eigen_device();
} }
#ifndef PADDLE_ONLY_CPU
class GPUPlaceGuard {
public:
explicit GPUPlaceGuard(GPUPlace new_place) : previous_(GetCurrentDeviceId()) {
if (previous_ != new_place) {
paddle::platform::SetDeviceId(new_place.device);
}
}
~GPUPlaceGuard() { paddle::platform::SetDeviceId(previous_.device); }
private:
GPUPlace previous_;
};
class CUDADeviceContext : public DeviceContext {
public:
explicit CUDADeviceContext(const GPUPlace gpu_place) : gpu_place_(gpu_place) {
GPUPlaceGuard guard(gpu_place_);
paddle::platform::throw_on_error(cudaStreamCreate(&stream_),
"cudaStreamCreate failed");
eigen_stream_ = new Eigen::CudaStreamDevice(&stream_);
eigen_device_ = new Eigen::GpuDevice(eigen_stream_);
}
Place GetPlace() const override {
Place retv = GPUPlace();
return retv;
}
void Wait() {
paddle::platform::throw_on_error(cudaStreamSynchronize(stream_),
"cudaStreamSynchronize failed");
}
cudaStream_t stream() { return stream_; }
Eigen::GpuDevice eigen_device() { return *eigen_device_; }
cublasHandle_t cublas_handle() {
if (!blas_handle_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_) ==
CUBLAS_STATUS_SUCCESS,
"cublasCreate failed");
PADDLE_ENFORCE(paddle::platform::dynload::cublasSetStream(
blas_handle_, stream_) == CUBLAS_STATUS_SUCCESS,
"cublasSetStream failed");
}
return blas_handle_;
}
cudnnHandle_t cudnn_handle() {
if (!dnn_handle_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_) ==
CUDNN_STATUS_SUCCESS,
"cudnnCreate failed");
PADDLE_ENFORCE(paddle::platform::dynload::cudnnSetStream(
dnn_handle_, stream_) == CUDNN_STATUS_SUCCESS,
"cudnnSetStream failed");
}
return dnn_handle_;
}
curandGenerator_t curand_generator() {
if (!rand_generator_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::curandCreateGenerator(
&rand_generator_, CURAND_RNG_PSEUDO_DEFAULT) ==
CURAND_STATUS_SUCCESS,
"curandCreateGenerator failed");
PADDLE_ENFORCE(
paddle::platform::dynload::curandSetPseudoRandomGeneratorSeed(
rand_generator_, random_seed_) == CURAND_STATUS_SUCCESS,
"curandSetPseudoRandomGeneratorSeed failed");
PADDLE_ENFORCE(paddle::platform::dynload::curandSetStream(
rand_generator_, stream_) == CURAND_STATUS_SUCCESS,
"curandSetStream failed");
}
return rand_generator_;
}
~CUDADeviceContext() {
Wait();
if (blas_handle_) {
PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_) ==
CUBLAS_STATUS_SUCCESS,
"cublasDestroy failed");
}
if (dnn_handle_) {
PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_) ==
CUDNN_STATUS_SUCCESS,
"cudnnDestroy failed");
}
if (rand_generator_) {
PADDLE_ENFORCE(paddle::platform::dynload::curandDestroyGenerator(
rand_generator_) == CURAND_STATUS_SUCCESS,
"curandDestroyGenerator failed");
}
delete eigen_stream_;
delete eigen_device_;
paddle::platform::throw_on_error(cudaStreamDestroy(stream_),
"cudaStreamDestroy failed");
}
private:
GPUPlace gpu_place_;
cudaStream_t stream_;
Eigen::CudaStreamDevice* eigen_stream_;
Eigen::GpuDevice* eigen_device_;
cublasHandle_t blas_handle_{nullptr};
cudnnHandle_t dnn_handle_{nullptr};
int random_seed_;
curandGenerator_t rand_generator_{nullptr};
};
template <>
Eigen::GpuDevice DeviceContext::get_eigen_device<Eigen::GpuDevice>() {
return dynamic_cast<CUDADeviceContext*>(this)->eigen_device();
}
#endif
} // namespace platform } // namespace platform
} // namespace paddle } // namespace paddle
...@@ -12,8 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,8 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "paddle/platform/device_context.h"
#include "gtest/gtest.h" #include "gtest/gtest.h"
#include "paddle/platform/cuda_device_context.h"
using DEVICE_GPU = Eigen::GpuDevice; using DEVICE_GPU = Eigen::GpuDevice;
TEST(Device, Init) { TEST(Device, Init) {
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册