device_context.cc 17.0 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)
319 320 321
  if (nccl_comm_) {
    PADDLE_ENFORCE(dynload::ncclCommDestroy(nccl_comm_));
  }
Q
qingqing01 已提交
322
#endif
323 324
}

L
liaogang 已提交
325
Place CUDADeviceContext::GetPlace() const { return place_; }
326

L
liaogang 已提交
327
void CUDADeviceContext::Wait() const {
328 329
  auto& allocator =
      DeviceTemporaryAllocator::Instance().Get<CUDADeviceContext>(*this);
330
  allocator.Release([this]() {
331 332 333 334 335 336 337 338 339 340 341
    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;
    }
342
  });
343 344
}

K
Kexin Zhao 已提交
345
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
346
  return compute_capability_;
K
Kexin Zhao 已提交
347 348
}

349
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
350
  return multi_process_ * max_threads_per_mp_;
351 352
}

353 354 355 356
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
  return eigen_device_.get();
}

357 358
bool CUDADeviceContext::tensor_core_available() const {
  return cublas_tensor_core_handle_ != nullptr;
S
sneaxiy 已提交
359 360
}

361 362 363 364 365 366 367 368 369
CudnnHolder* CUDADeviceContext::cudnn_holder() const {
  std::call_once(init_cudnn_, [&]() {
    if (dynload::HasCUDNN()) {
      cudnn_holder_.reset(new CudnnHolder(&stream_, place_));
    }
  });
  return cudnn_holder_.get();
}

370
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
371
  return cudnn_holder()->cudnn_handle();
372 373
}

S
sneaxiy 已提交
374
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
375
  return CudnnWorkspaceHandle(cudnn_holder());
376
}
377

378
cudaStream_t CUDADeviceContext::stream() const { return stream_; }
Q
qijun 已提交
379

C
chengduoZH 已提交
380 381 382 383 384 385 386 387 388 389 390 391 392 393
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
394
#endif
Q
qijun 已提交
395

T
tensor-tang 已提交
396 397
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
398 399 400
    : 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
401 402
}

S
Sylwester Fraczek 已提交
403
namespace {
404 405
// Current mkldnn session id.
thread_local size_t cur_mkldnn_session_id = kMKLDNNSessionID_Default;
406 407 408 409
// Current data input shape string.
// - For fixed-shape, it's a null string in default.
// - For dynamic-shape, it's user specific.
thread_local std::string cur_input_shape_str = "";
410 411 412
// the cache capacity of different input shapes for MKLDNN.
// Default 1 means fixed input shape, not dynamic shape.
thread_local int cur_input_shape_cache_capacity = 1;
413
}  // namespace
S
Sylwester Fraczek 已提交
414

415 416
void set_cur_mkldnn_session_id(size_t sid) { cur_mkldnn_session_id = sid; }
size_t get_cur_mkldnn_session_id(void) { return cur_mkldnn_session_id; }
417 418 419
void set_cur_input_shape_str(std::string input_shape_str) {
  cur_input_shape_str = input_shape_str;
}
420 421 422
void set_cur_input_shape_cache_capacity(int input_shape_cache_capacity) {
  cur_input_shape_cache_capacity = input_shape_cache_capacity;
}
S
Sylwester Fraczek 已提交
423

424 425
void MKLDNNDeviceContext::ResetBlobMap() const { p_blobmap_->clear(); }

426 427 428 429 430 431 432 433 434 435 436
size_t MKLDNNDeviceContext::GetShapeBlobSize() const {
  std::lock_guard<std::mutex> lock(*p_mutex_);
  BlobMap* pMap = p_blobmap_.get();
  auto map_it = pMap->find(cur_mkldnn_session_id);
  if (map_it == pMap->end()) {
    LOG(FATAL) << "MKLDNNDeviceContext don't find cur_mkldnn_session_id : "
               << cur_mkldnn_session_id;
  }
  return map_it->second->size();
}

437 438
void MKLDNNDeviceContext::SetBlob(const std::string& name,
                                  std::shared_ptr<void> data) const {
439
  BlobMap* pMap = p_blobmap_.get();
440
  std::shared_ptr<ShapeBlob> sBlob = nullptr;
441 442
  std::shared_ptr<KeyBlob> pBlob = nullptr;

443
  int sid = platform::get_cur_mkldnn_session_id();
T
tensor-tang 已提交
444

445
  std::lock_guard<std::mutex> lock(*p_mutex_);
T
tensor-tang 已提交
446

447 448
  // Find ShapeBlob for current mkldnn session id.
  auto map_it = pMap->find(sid);
449 450 451

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
452
    sBlob = std::shared_ptr<ShapeBlob>(new ShapeBlob());
453 454
    (*pMap)[sid] = sBlob;
    VLOG(2) << "SetBlob: sid=" << sid << ", add new sid\n";
455
  } else {
456
    sBlob = map_it->second;
457
  }
T
tensor-tang 已提交
458

459 460
  // Find KeyBlob for current input shape
  auto key_it = sBlob->find(cur_input_shape_str);
461

462
  if (key_it == sBlob->end()) {
463 464 465 466 467 468 469 470 471
    // In cache clearing mode, cur_input_shape_cache_capacity defines
    // max pblob capacity
    if ((sid == kMKLDNNSessionID_CacheClearing) &&
        (sBlob->size() >=
         static_cast<size_t>(cur_input_shape_cache_capacity))) {
      VLOG(2) << "sid=" << sid
              << ", remove all blobs of shape: " << sBlob->begin()->first;
      sBlob->erase(sBlob->begin()->first);
    }
472 473
    pBlob = std::shared_ptr<KeyBlob>(new KeyBlob());
    (*sBlob)[cur_input_shape_str] = pBlob;
474
  } else {
475
    pBlob = key_it->second;
476 477
  }

478 479 480 481 482 483 484
  // Find Blob via name
  auto blob_it = pBlob->find(name);
  if (blob_it == pBlob->end()) {
    (*pBlob)[name] = data;
  } else {
    blob_it->second = data;  // set data to existing blob
  }
485
  VLOG(2) << "SetBlob: sid=" << sid << ", add blob=" << name << "\n";
486
  // lock will be automatically released when out of scope
487
  return;
T
tensor-tang 已提交
488 489
}

490 491
std::shared_ptr<void> MKLDNNDeviceContext::GetBlob(
    const std::string& name) const {
492
  BlobMap* pMap = p_blobmap_.get();
493
  std::shared_ptr<ShapeBlob> sBlob = nullptr;
494
  std::shared_ptr<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
495

496
  int sid = platform::get_cur_mkldnn_session_id();
T
tensor-tang 已提交
497

498
  std::lock_guard<std::mutex> lock(*p_mutex_);
499

500 501
  // Find ShapeBlob for current mkldnn session id firstly
  auto map_it = pMap->find(sid);
502
  if (map_it == pMap->end()) {
503
    VLOG(2) << "GetBlob: sid=" << sid << ", miss sid\n";
504 505 506 507 508 509 510
    return nullptr;
  }
  sBlob = map_it->second;

  // Find KeyBlob for current input shape secondly
  auto sBlob_it = sBlob->find(cur_input_shape_str);
  if (sBlob_it == sBlob->end()) {
511
    VLOG(2) << "GetBlob: sid=" << cur_input_shape_str
512 513 514 515
            << ", miss input_shape_str\n";
    return nullptr;
  }
  pBlob = sBlob_it->second;
516 517 518 519

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

520
  if (key_it == pBlob->end()) {
521
    VLOG(2) << "GetBlob sid=" << sid << ", miss blob=" << name << "\n";
522 523
    return nullptr;
  }
524

525
  VLOG(2) << "GetBlob sid=" << sid << ", get blob=" << name << "\n";
526 527
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
528 529 530 531
}

#endif

Q
qijun 已提交
532
}  // namespace platform
Q
qijun 已提交
533
}  // namespace paddle