device_context.cc 7.0 KB
Newer Older
Q
qijun 已提交
1 2 3 4 5 6 7 8 9 10 11
/* 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. */

Y
Yi Wang 已提交
12 13
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/memory/memory.h"
Q
qijun 已提交
14 15 16 17

namespace paddle {
namespace platform {

D
dzhwinter 已提交
18 19
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
20
const platform::DeviceContext* DeviceContextPool::Get(
D
dzhwinter 已提交
21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
    const platform::Place& place) {
  auto it = device_contexts_.find(place);
  if (it == device_contexts_.end()) {
    PADDLE_THROW(
        "'Place' is not supported, Please re-compile with WITH_GPU "
        "option");
  }
  return it->second;
}

DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
  PADDLE_ENFORCE_GT(places.size(), 0);
  for (size_t i = 0; i < places.size(); i++) {
    if (platform::is_cpu_place(places[i])) {
      device_contexts_.emplace(places[i],
                               new platform::CPUDeviceContext(
                                   boost::get<platform::CPUPlace>(places[i])));
    } else if (platform::is_gpu_place(places[i])) {
#ifdef PADDLE_WITH_CUDA
      device_contexts_.emplace(places[i],
                               new platform::CUDADeviceContext(
D
dzhwinter 已提交
43
                                   boost::get<platform::CUDAPlace>(places[i])));
D
dzhwinter 已提交
44 45
#else
      PADDLE_THROW(
D
dzhwinter 已提交
46
          "'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
D
dzhwinter 已提交
47 48 49 50 51 52
          "option");
#endif
    }
  }
}

53 54 55 56
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
57
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
58 59 60 61 62 63 64
  eigen_device_.reset(new Eigen::DefaultDevice());
}

Eigen::DefaultDevice* CPUDeviceContext::eigen_device() const {
  return eigen_device_.get();
}

D
dzhwinter 已提交
65
Place CPUDeviceContext::GetPlace() const { return place_; }
66

67
#ifdef PADDLE_WITH_CUDA
68

Q
init  
qijun 已提交
69 70 71 72 73 74 75
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

D
dzhwinter 已提交
76
  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
77 78 79 80 81 82 83 84 85 86 87 88
    stream_ = cuda_stream;
    place_ = place;
    device_prop_ = &Eigen::m_deviceProperties[place.device];
  }

  const cudaStream_t& stream() const override { return *stream_; }

  const cudaDeviceProp& deviceProperties() const override {
    return *device_prop_;
  }

  void* allocate(size_t num_bytes) const override {
Q
qijun 已提交
89
    return paddle::memory::Alloc(place_, num_bytes);
Q
init  
qijun 已提交
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
  }

  void deallocate(void* buffer) const override {
    paddle::memory::Free(place_, buffer);
  }

  void* scratchpad() const override {
    if (scratch_ == NULL) {
      scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int));
    }
    return scratch_;
  }

  unsigned int* semaphore() const override {
    if (semaphore_ == NULL) {
      char* scratch =
          static_cast<char*>(scratchpad()) + Eigen::kCudaScratchSize;
      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
      PADDLE_ENFORCE(
          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
    }
    return semaphore_;
  }

 private:
D
dzhwinter 已提交
115
  CUDAPlace place_;
Q
init  
qijun 已提交
116 117
  const cudaStream_t* stream_;         // not owned;
  const cudaDeviceProp* device_prop_;  // not owned;
Q
qijun 已提交
118
  mutable void* scratch_;
Q
init  
qijun 已提交
119 120 121
  mutable unsigned int* semaphore_;
};

D
dzhwinter 已提交
122
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
123
  SetDeviceId(place_.device);
Q
init  
qijun 已提交
124 125 126
  PADDLE_ENFORCE(cudaStreamCreate(&stream_));
  eigen_stream_.reset(new EigenCudaStreamDevice());
  eigen_stream_->Reinitialize(&stream_, place);
127
  eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
128 129
  PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
  PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
D
dzhwinter 已提交
130 131 132 133 134 135
  if (dynload::HasCUDNN()) {
    PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
    PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
  } else {
    cudnn_handle_ = nullptr;
  }
136 137 138 139
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
L
liaogang 已提交
140
  Wait();
141
  PADDLE_ENFORCE(dynload::cublasDestroy(cublas_handle_));
D
dzhwinter 已提交
142 143 144
  if (cudnn_handle_ != nullptr) {
    PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
  }
145 146
  eigen_stream_.reset();
  eigen_device_.reset();
Q
init  
qijun 已提交
147
  PADDLE_ENFORCE(cudaStreamDestroy(stream_));
148 149
}

L
liaogang 已提交
150
Place CUDADeviceContext::GetPlace() const { return place_; }
151

L
liaogang 已提交
152
void CUDADeviceContext::Wait() const {
Q
init  
qijun 已提交
153
  PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
154 155 156
  PADDLE_ENFORCE(cudaGetLastError());
}

157 158 159 160
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
  return eigen_device_.get();
}

161
cublasHandle_t CUDADeviceContext::cublas_handle() const {
162 163 164
  return cublas_handle_;
}

165
cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
166

167
cudaStream_t CUDADeviceContext::stream() const { return stream_; }
Q
qijun 已提交
168

L
Luo Tao 已提交
169
#endif
Q
qijun 已提交
170

T
tensor-tang 已提交
171 172 173 174 175 176 177 178 179 180 181 182 183
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
    : CPUDeviceContext(place), ready_(false) {
  stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
  engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0));
}

template <typename T>
void MKLDNNDeviceContext::AddElement(const std::string& op_key,
                                     const T& value) {
  if (GetElement<T>(op_key)) {
    return;
  }
184
  GetElementPool<T>().emplace(op_key, std::move(value));
T
tensor-tang 已提交
185 186 187
}

template <typename T>
188
const T& MKLDNNDeviceContext::GetElement(const std::string& op_key) const {
T
tensor-tang 已提交
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234
  auto it = GetElementPool<T>().find(op_key);
  return it == GetElementPool<T>().end() ? nullptr : it->second;
}

template <>
const std::unordered_map<const std::string, const MKLDNNMemoryPtr,
                         std::hash<std::string>>&
MKLDNNDeviceContext::GetElementPool<MKLDNNMemoryPtr>() const {
  return memory_pool_;
}

template <>
const std::unordered_map<const std::string, const MKLDNNPrimitivePtr,
                         std::hash<std::string>>&
MKLDNNDeviceContext::GetElementPool<MKLDNNPrimitivePtr>() const {
  return primitive_pool_;
}

template <>
const std::unordered_map<const std::string, const MKLDNNPrimitiveDescPtr,
                         std::hash<std::string>>&
MKLDNNDeviceContext::GetElementPool<MKLDNNPrimitiveDescPtr>() const {
  return primitive_desc_pool_;
}

void MKLDNNDeviceContext::Execute(bool block) {
  if (pipeline_.empty()) {
    return;
  }
  ResetStream();
  stream_->submit(pipeline_).wait(block);
  ready_ = false;
  pipeline_.clear();
}

void MKLDNNDeviceContext::ResetStream() {
  if (ready_) {
    return;
  }
  // TODO(TJ): change me when mkldnn have specific method to reset this state
  stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
  ready_ = true;
}

#endif

Q
qijun 已提交
235
}  // namespace platform
Q
qijun 已提交
236
}  // namespace paddle