提交 555b0a72 编写于 作者: Y Yu Yang 提交者: GitHub

Merge pull request #2814 from QiJune/feature/refine_device_context

Add GetEigenDevice To DeviceContext.
add_subdirectory(dynload)
nv_test(cuda_test SRCS cuda_test.cu DEPS dyload_cuda)
nv_test(cuda_test SRCS cuda_test.cu)
cc_library(place SRCS place.cc)
cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
IF(WITH_GPU)
set(GPU_CTX_DEPS dyload_cuda dynamic_loader )
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
ELSE()
set(GPU_CTX_DEPS)
ENDIF()
......
#include <paddle/platform/device_context.h>
/* 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
template <>
Eigen::DefaultDevice* DeviceContext::get_eigen_device<Eigen::DefaultDevice>() {
return reinterpret_cast<CPUDeviceContext*>(this)->eigen_device();
}
#ifndef PADDLE_ONLY_CPU
template <>
Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() {
return reinterpret_cast<CUDADeviceContext*>(this)->eigen_device();
}
#endif
} // namespace platform
} // namespace paddle
\ No newline at end of file
} // 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.
......@@ -23,6 +20,7 @@ limitations under the License. */
#define EIGEN_USE_GPU
#endif
#include <paddle/platform/place.h>
#include <memory>
#include <unsupported/Eigen/CXX11/Tensor>
namespace paddle {
......@@ -32,14 +30,27 @@ class DeviceContext {
public:
virtual ~DeviceContext() {}
virtual Place GetPlace() const = 0;
template <typename DeviceType>
DeviceType* get_eigen_device();
};
class CPUDeviceContext : public DeviceContext {
public:
Eigen::DefaultDevice* eigen_device() {
if (!eigen_device_) {
eigen_device_.reset(new Eigen::DefaultDevice());
}
return eigen_device_.get();
}
Place GetPlace() const override {
Place retv = CPUPlace();
return retv;
}
private:
std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
};
#ifndef PADDLE_ONLY_CPU
......@@ -64,8 +75,8 @@ class CUDADeviceContext : public DeviceContext {
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_);
eigen_stream_.reset(new Eigen::CudaStreamDevice(&stream_));
eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
}
Place GetPlace() const override {
......@@ -80,7 +91,7 @@ class CUDADeviceContext : public DeviceContext {
cudaStream_t stream() { return stream_; }
Eigen::GpuDevice eigen_device() { return *eigen_device_; }
Eigen::GpuDevice* eigen_device() { return eigen_device_.get(); }
cublasHandle_t cublas_handle() {
if (!blas_handle_) {
......@@ -145,10 +156,8 @@ class CUDADeviceContext : public DeviceContext {
rand_generator_) == CURAND_STATUS_SUCCESS,
"curandDestroyGenerator failed");
}
delete eigen_stream_;
delete eigen_device_;
eigen_stream_.reset();
eigen_device_.reset();
paddle::platform::throw_on_error(cudaStreamDestroy(stream_),
"cudaStreamDestroy failed");
}
......@@ -157,8 +166,8 @@ class CUDADeviceContext : public DeviceContext {
GPUPlace gpu_place_;
cudaStream_t stream_;
Eigen::CudaStreamDevice* eigen_stream_;
Eigen::GpuDevice* eigen_device_;
std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
std::unique_ptr<Eigen::GpuDevice> eigen_device_;
cublasHandle_t blas_handle_{nullptr};
......@@ -167,6 +176,8 @@ class CUDADeviceContext : public DeviceContext {
int random_seed_;
curandGenerator_t rand_generator_{nullptr};
};
#endif
} // namespace platform
} // namespace paddle
......@@ -15,13 +15,26 @@ limitations under the License. */
#include "paddle/platform/device_context.h"
#include "gtest/gtest.h"
TEST(CUDADeviceContext, Init) {
using DEVICE_GPU = Eigen::GpuDevice;
TEST(Device, Init) {
int count = paddle::platform::GetDeviceCount();
for (int i = 0; i < count; i++) {
paddle::platform::DeviceContext* device_context =
new paddle::platform::CUDADeviceContext(i);
Eigen::GpuDevice* gpu_device =
device_context->template get_eigen_device<DEVICE_GPU>();
ASSERT_NE(nullptr, gpu_device);
delete device_context;
}
}
TEST(Device, CUDADeviceContext) {
int count = paddle::platform::GetDeviceCount();
for (int i = 0; i < count; i++) {
paddle::platform::CUDADeviceContext* device_context =
new paddle::platform::CUDADeviceContext(i);
Eigen::GpuDevice gpu_device = device_context->eigen_device();
ASSERT_NE(nullptr, gpu_device.stream());
Eigen::GpuDevice* gpu_device = device_context->eigen_device();
ASSERT_NE(nullptr, gpu_device);
cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
ASSERT_NE(nullptr, cudnn_handle);
cublasHandle_t cublas_handle = device_context->cublas_handle();
......
cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags)
nv_library(dyload_cuda SRCS cublas.cc cudnn.cc curand.cc)
nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc)
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