device_context.cc 14.6 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Q
qijun 已提交
2 3 4 5
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
6

Q
qijun 已提交
7 8 9 10 11
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
#include "paddle/fluid/platform/device_context.h"
13
#include <set>
14
#include <string>
Y
Yu Yang 已提交
15
#include <unordered_set>
16 17
#include <vector>

Y
Yi Wang 已提交
18
#include "paddle/fluid/memory/memory.h"
19 20
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/rw_lock.h"
S
sneaxiy 已提交
21
#include "paddle/fluid/platform/cuda_device_guard.h"
22
#endif
23

24 25
#include "glog/logging.h"

Q
qijun 已提交
26 27 28
namespace paddle {
namespace platform {

D
dzhwinter 已提交
29 30
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
31
platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
D
dzhwinter 已提交
32 33 34
  auto it = device_contexts_.find(place);
  if (it == device_contexts_.end()) {
    PADDLE_THROW(
M
minqiyang 已提交
35 36 37
        "Place %s is not supported, Please re-compile with WITH_GPU "
        "option",
        place);
D
dzhwinter 已提交
38
  }
39
  return it->second.get().get();
D
dzhwinter 已提交
40 41
}

42 43 44 45 46 47 48 49 50 51 52
template <typename DevCtx, typename PlaceType>
inline void EmplaceDeviceContext(
    std::map<Place, std::shared_future<std::unique_ptr<DeviceContext>>>*
        map_ptr,
    platform::Place p) {
  using PtrType = std::unique_ptr<DeviceContext>;
  map_ptr->emplace(p, std::async(std::launch::deferred, [=] {
                     // lazy evaluation. i.e., only create device context at
                     // first `Get`
                     return PtrType(new DevCtx(boost::get<PlaceType>(p)));
                   }));
C
chengduozh 已提交
53 54
}

D
dzhwinter 已提交
55 56 57
DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
  PADDLE_ENFORCE_GT(places.size(), 0);
58
  std::set<Place> set;
Y
Yu Yang 已提交
59 60 61 62 63
  for (auto& p : places) {
    set.insert(p);
  }
  for (auto& p : set) {
    if (platform::is_cpu_place(p)) {
64
#ifdef PADDLE_WITH_MKLDNN
65
      EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
66
#else
67
      EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
68
#endif
Y
Yu Yang 已提交
69
    } else if (platform::is_gpu_place(p)) {
D
dzhwinter 已提交
70
#ifdef PADDLE_WITH_CUDA
71
      EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
D
dzhwinter 已提交
72 73
#else
      PADDLE_THROW(
D
dzhwinter 已提交
74
          "'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
D
dzhwinter 已提交
75
          "option");
C
chengduoZH 已提交
76 77 78
#endif
    } else if (platform::is_cuda_pinned_place(p)) {
#ifdef PADDLE_WITH_CUDA
79 80
      EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
          &device_contexts_, p);
C
chengduoZH 已提交
81 82 83 84
#else
      PADDLE_THROW(
          "'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
          "option");
D
dzhwinter 已提交
85 86 87 88 89
#endif
    }
  }
}

90 91 92 93 94 95 96
DeviceTemporaryAllocator* DeviceTemporaryAllocator::allocators = nullptr;

#ifdef PADDLE_WITH_CUDA
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
    const platform::Place& place, const cudaStream_t& stream) {
  PADDLE_ENFORCE(platform::is_gpu_place(place));
  auto place_stream = std::make_pair(place, stream);
97 98 99 100 101 102 103 104 105 106 107 108
  std::unique_lock<std::mutex> lock(mtx_);
  auto it = device_allocator_.find(place_stream);
  if (it == device_allocator_.end()) {
    auto tmp_allocator = new TemporaryAllocator(place);
    tmp_allocator->SetCallback([stream]() {
      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
      PADDLE_ENFORCE(cudaGetLastError());
    });
    device_allocator_[place_stream].reset(tmp_allocator);
    return *tmp_allocator;
  } else {
    return *it->second;
109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
  }
}

template <>
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
    const platform::CUDADeviceContext& dev_ctx) {
  return Get(dev_ctx.GetPlace(), dev_ctx.stream());
}
#endif

template <>
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
    const platform::CPUDeviceContext& dev_ctx) {
  return cpu_allocator_;
}

platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
    const platform::Place& place) {
  PADDLE_ENFORCE(platform::is_cpu_place(place), "You should pass CPUPlace");
  return cpu_allocator_;
}

131 132 133 134
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
135
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
136 137 138 139 140 141 142
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

D
dzhwinter 已提交
143
Place CPUDeviceContext::GetPlace() const { return place_; }
144

145
#ifdef PADDLE_WITH_CUDA
146

Q
init  
qijun 已提交
147 148 149 150 151 152 153
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

D
dzhwinter 已提交
154
  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
155 156 157 158 159 160 161 162 163 164 165 166
    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 {
S
sneaxiy 已提交
167 168 169
    if (UNLIKELY(num_bytes == 0)) {
      return nullptr;
    }
170
    auto buf = paddle::memory::Alloc(place_, num_bytes);
171
    void* retv = buf->ptr();
S
sneaxiy 已提交
172 173 174 175
    {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.emplace(retv, std::move(buf));
    }
176
    return retv;
Q
init  
qijun 已提交
177 178
  }

S
sneaxiy 已提交
179 180 181 182 183 184
  void deallocate(void* buffer) const override {
    if (LIKELY(buffer)) {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.erase(buffer);
    }
  }
Q
init  
qijun 已提交
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204

  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
205
  CUDAPlace place_;
Q
init  
qijun 已提交
206 207
  const cudaStream_t* stream_;         // not owned;
  const cudaDeviceProp* device_prop_;  // not owned;
Q
qijun 已提交
208
  mutable void* scratch_;
Q
init  
qijun 已提交
209
  mutable unsigned int* semaphore_;
S
sneaxiy 已提交
210
  mutable std::mutex mtx_;  // to protect allocations_
Y
Yu Yang 已提交
211
  mutable std::unordered_map<void*, memory::AllocationPtr> allocations_;
Q
init  
qijun 已提交
212 213
};

S
sneaxiy 已提交
214
CudnnHolder::CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place)
Y
Yu Yang 已提交
215
    : workspace_(nullptr), stream_(stream), place_(place) {
N
nhzlx 已提交
216
  PADDLE_ENFORCE(cudaSetDevice(place_.device));
S
sneaxiy 已提交
217 218 219
  PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
  PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, *stream_));
}
220

S
sneaxiy 已提交
221 222
CudnnHolder::~CudnnHolder() {
  PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
S
sneaxiy 已提交
223
}
224

S
sneaxiy 已提交
225
void CudnnHolder::ReallocateWorkspace(size_t required_workspace_len) {
Y
Yu Yang 已提交
226
  if (required_workspace_len <= WorkspaceSize()) {
S
sneaxiy 已提交
227
    return;
Y
Yu Yang 已提交
228
  }
S
sneaxiy 已提交
229 230 231
  if (workspace_ != nullptr) {
    // Maybe someone is using the current workspace
    PADDLE_ENFORCE(cudaStreamSynchronize(*stream_));
Y
Yu Yang 已提交
232
    workspace_.reset();
233
  }
234
  workspace_ = paddle::memory::Alloc(place_, required_workspace_len);
S
sneaxiy 已提交
235
}
236 237 238

CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
    : place_(place), cudnn_holder_(nullptr) {
Y
Yu Yang 已提交
239
  CUDADeviceGuard guard(place_.device);
C
chengduo 已提交
240 241 242
  compute_capability_ = GetCUDAComputeCapability(place_.device);
  multi_process_ = GetCUDAMultiProcessors(place_.device);
  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
Q
init  
qijun 已提交
243 244 245
  PADDLE_ENFORCE(cudaStreamCreate(&stream_));
  eigen_stream_.reset(new EigenCudaStreamDevice());
  eigen_stream_->Reinitialize(&stream_, place);
246
  eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
247 248 249 250 251 252 253 254 255
  cublas_handle_.reset(new CublasHandleHolder(stream_, CUBLAS_DEFAULT_MATH));

  if (TensorCoreAvailable()) {
#if CUDA_VERSION >= 9000
    cublas_tensor_core_handle_.reset(
        new CublasHandleHolder(stream_, CUBLAS_TENSOR_OP_MATH));
#endif
  }

C
chengduo 已提交
256 257 258
  driver_version_ = GetCUDADriverVersion(place_.device);
  runtime_version_ = GetCUDARuntimeVersion(place_.device);

259 260
  LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device
                          << ", CUDA Capability: " << compute_capability_
C
chengduo 已提交
261
                          << ", Driver API Version: " << driver_version_ / 1000
262
                          << "." << (driver_version_ % 100) / 10
C
chengduo 已提交
263 264 265
                          << ", Runtime API Version: "
                          << runtime_version_ / 1000 << "."
                          << (runtime_version_ % 100) / 10;
266 267 268
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
269
                          << (cudnn_dso_ver % 1000) / 100 << ".";
S
sneaxiy 已提交
270 271 272

  {
    // Check CUDA/CUDNN version compatiblity
273 274 275 276
    auto local_cuda_version =
        (driver_version_ / 1000) * 10 + (driver_version_ % 100) / 10;
    auto compile_cuda_version =
        (CUDA_VERSION / 1000) * 10 + (CUDA_VERSION % 100) / 10;
S
sneaxiy 已提交
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291
    if (local_cuda_version < compile_cuda_version) {
      LOG_FIRST_N(WARNING, 1)
          << "WARNING: device: " << place_.device
          << ". The installed Paddle is compiled with CUDA "
          << compile_cuda_version / 10 << "." << compile_cuda_version % 10
          << ", but CUDA runtime version in your machine is "
          << local_cuda_version / 10 << "." << local_cuda_version % 10
          << ", which may cause serious incompatible bug. "
          << "Please recompile or reinstall Paddle with compatible CUDA "
             "version.";
    }

    if (dynload::HasCUDNN()) {
      auto local_cudnn_version = cudnn_dso_ver / 100;
      auto compile_cudnn_version = CUDNN_VERSION / 100;
S
sneaxiy 已提交
292
      if (local_cudnn_version < static_cast<size_t>(compile_cudnn_version)) {
S
sneaxiy 已提交
293 294 295 296 297 298 299 300 301 302 303 304 305
        LOG_FIRST_N(WARNING, 1)
            << "WARNING: device: " << place_.device
            << ". The installed Paddle is compiled with CUDNN "
            << compile_cudnn_version / 10 << "." << compile_cudnn_version % 10
            << ", but CUDNN version in your machine is "
            << local_cudnn_version / 10 << "." << local_cudnn_version % 10
            << ", which may cause serious incompatible bug. "
            << "Please recompile or reinstall Paddle with compatible CUDNN "
               "version.";
      }
    }
  }

S
sneaxiy 已提交
306
  callback_manager_.reset(new StreamCallbackManager(stream_));
307 308 309 310
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
L
liaogang 已提交
311
  Wait();
S
sneaxiy 已提交
312
  WaitStreamCallback();
313 314
  cublas_handle_.reset();
  cublas_tensor_core_handle_.reset();
315 316
  eigen_stream_.reset();
  eigen_device_.reset();
Q
init  
qijun 已提交
317
  PADDLE_ENFORCE(cudaStreamDestroy(stream_));
Q
qingqing01 已提交
318
#if !defined(_WIN32)
Q
qingqing01 已提交
319
  PADDLE_ENFORCE(dynload::ncclCommDestroy(nccl_comm_));
Q
qingqing01 已提交
320
#endif
321 322
}

L
liaogang 已提交
323
Place CUDADeviceContext::GetPlace() const { return place_; }
324

L
liaogang 已提交
325
void CUDADeviceContext::Wait() const {
326 327
  auto& allocator =
      DeviceTemporaryAllocator::Instance().Get<CUDADeviceContext>(*this);
328
  allocator.Release([this]() {
329 330 331 332 333 334 335 336 337 338 339
    cudaError_t e_sync = cudaStreamSynchronize(stream_);
    if (e_sync != 0) {
      LOG(FATAL) << "cudaStreamSynchronize " << cudaGetErrorString(e_sync)
                 << " errno:" << e_sync;
    }

    cudaError_t e_get = cudaGetLastError();
    if (e_get != 0) {
      LOG(FATAL) << "cudaGetLastError  " << cudaGetErrorString(e_get)
                 << " errno:" << e_get;
    }
340
  });
341 342
}

K
Kexin Zhao 已提交
343
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
344
  return compute_capability_;
K
Kexin Zhao 已提交
345 346
}

347
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
348
  return multi_process_ * max_threads_per_mp_;
349 350
}

351 352 353 354
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
  return eigen_device_.get();
}

355 356
bool CUDADeviceContext::tensor_core_available() const {
  return cublas_tensor_core_handle_ != nullptr;
S
sneaxiy 已提交
357 358
}

359 360 361 362 363 364 365 366 367
CudnnHolder* CUDADeviceContext::cudnn_holder() const {
  std::call_once(init_cudnn_, [&]() {
    if (dynload::HasCUDNN()) {
      cudnn_holder_.reset(new CudnnHolder(&stream_, place_));
    }
  });
  return cudnn_holder_.get();
}

368
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
369
  return cudnn_holder()->cudnn_handle();
370 371
}

S
sneaxiy 已提交
372
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
373
  return CudnnWorkspaceHandle(cudnn_holder());
374
}
375

376
cudaStream_t CUDADeviceContext::stream() const { return stream_; }
Q
qijun 已提交
377

C
chengduoZH 已提交
378 379 380 381 382 383 384 385 386 387 388 389 390 391
CUDAPinnedDeviceContext::CUDAPinnedDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

CUDAPinnedDeviceContext::CUDAPinnedDeviceContext(CUDAPinnedPlace place)
    : place_(place) {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

Place CUDAPinnedDeviceContext::GetPlace() const { return place_; }
L
Luo Tao 已提交
392
#endif
Q
qijun 已提交
393

T
tensor-tang 已提交
394 395
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
396 397 398
    : CPUDeviceContext(place), engine_(mkldnn::engine::cpu, 0), p_blobmap_() {
  p_blobmap_.reset(new BlobMap());
  p_mutex_.reset(new std::mutex());
T
tensor-tang 已提交
399 400
}

S
Sylwester Fraczek 已提交
401 402 403 404 405 406 407 408
namespace {
// Current thread's id.
thread_local int cur_thread_id = 0;
}

void set_cur_thread_id(int tid) { cur_thread_id = tid; }
int get_cur_thread_id(void) { return cur_thread_id; }

409 410
void MKLDNNDeviceContext::SetBlob(const std::string& name,
                                  std::shared_ptr<void> data) const {
411 412 413 414
  BlobMap* pMap = p_blobmap_.get();
  std::shared_ptr<KeyBlob> pBlob = nullptr;

  int tid = platform::get_cur_thread_id();
T
tensor-tang 已提交
415

416
  std::lock_guard<std::mutex> lock(*p_mutex_);
T
tensor-tang 已提交
417

418 419 420 421 422 423 424
  // Find KeyBlob for current thread
  auto map_it = pMap->find(tid);

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
    pBlob = std::shared_ptr<KeyBlob>(new KeyBlob());
    (*pMap)[tid] = pBlob;
425
  } else {
426
    pBlob = map_it->second;
427
  }
T
tensor-tang 已提交
428

429 430 431 432 433 434 435 436 437 438
  // Find Key in found (or newly created) KeyBlob
  auto key_it = pBlob->find(name);

  if (key_it == pBlob->end()) {
    (*pBlob)[name] = data;  // create new blob
  } else {
    key_it->second = data;  // set data to existing blob
  }

  // lock will be automatically released when out of scope
439
  return;
T
tensor-tang 已提交
440 441
}

442 443
std::shared_ptr<void> MKLDNNDeviceContext::GetBlob(
    const std::string& name) const {
444 445
  BlobMap* pMap = p_blobmap_.get();
  std::shared_ptr<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
446

447
  int tid = platform::get_cur_thread_id();
T
tensor-tang 已提交
448

449
  std::lock_guard<std::mutex> lock(*p_mutex_);
450 451 452 453 454 455 456 457 458 459

  // Find KeyBlob for current thread firstly
  auto map_it = pMap->find(tid);
  if (map_it == pMap->end()) return nullptr;
  pBlob = map_it->second;

  // Find Blob via name
  auto key_it = pBlob->find(name);

  if (key_it == pBlob->end()) return nullptr;
460

461 462
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
463 464 465 466
}

#endif

Q
qijun 已提交
467
}  // namespace platform
Q
qijun 已提交
468
}  // namespace paddle