device_context.cc 19.2 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"
21
#include "paddle/fluid/memory/allocation/cuda_device_context_allocator.h"
S
sneaxiy 已提交
22
#include "paddle/fluid/platform/cuda_device_guard.h"
23
#endif
24

25 26
#include "glog/logging.h"

27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53
namespace paddle {
namespace memory {

AllocationPtr Alloc(const platform::DeviceContext& dev_ctx, size_t size) {
  auto place = dev_ctx.GetPlace();
#ifdef PADDLE_WITH_CUDA
  if (size == 0 || !platform::is_gpu_place(place)) {
    return Alloc(place, size);
  }
  auto* default_dev_ctx = static_cast<platform::CUDADeviceContext*>(
      platform::DeviceContextPool::Instance().Get(place));
  auto& desired_dev_ctx =
      static_cast<const platform::CUDADeviceContext&>(dev_ctx);
  if (default_dev_ctx->stream() == desired_dev_ctx.stream()) {
    return Alloc(place, size);
  } else {
    return allocation::CUDADeviceContextAllocatorPool::Instance().Alloc(
        desired_dev_ctx, size);
  }
#else
  return Alloc(place, size);
#endif
}

}  // namespace memory
}  // namespace paddle

Q
qijun 已提交
54 55 56
namespace paddle {
namespace platform {

D
dzhwinter 已提交
57 58
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
59
platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
D
dzhwinter 已提交
60 61
  auto it = device_contexts_.find(place);
  if (it == device_contexts_.end()) {
G
GaoWei8 已提交
62 63
    PADDLE_THROW(platform::errors::Unimplemented(
        "Place %s is not supported. Please check that your paddle compiles "
64 65
        "with WITH_GPU or WITH_XPU option or check that your train process "
        "hold the "
G
GaoWei8 已提交
66 67
        "correct gpu_id if you use Executor.",
        place));
D
dzhwinter 已提交
68
  }
69
  return it->second.get().get();
D
dzhwinter 已提交
70 71
}

72 73 74 75 76 77 78 79 80
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`
81
                     return PtrType(new DevCtx(BOOST_GET_CONST(PlaceType, p)));
82
                   }));
C
chengduozh 已提交
83 84
}

D
dzhwinter 已提交
85 86
DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
G
GaoWei8 已提交
87 88 89 90 91
  PADDLE_ENFORCE_GT(
      places.size(), 0,
      platform::errors::InvalidArgument("The number of platform places should "
                                        "be larger than 0. But received %d.",
                                        places.size()));
92
  std::set<Place> set;
Y
Yu Yang 已提交
93 94 95 96 97
  for (auto& p : places) {
    set.insert(p);
  }
  for (auto& p : set) {
    if (platform::is_cpu_place(p)) {
98
#ifdef PADDLE_WITH_MKLDNN
99
      EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
100
#else
101
      EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
102
#endif
Y
Yu Yang 已提交
103
    } else if (platform::is_gpu_place(p)) {
D
dzhwinter 已提交
104
#ifdef PADDLE_WITH_CUDA
105
      EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
D
dzhwinter 已提交
106
#else
G
GaoWei8 已提交
107 108 109
      PADDLE_THROW(
          platform::errors::Unimplemented("CUDAPlace is not supported. Please "
                                          "re-compile with WITH_GPU option."));
C
chengduoZH 已提交
110 111 112
#endif
    } else if (platform::is_cuda_pinned_place(p)) {
#ifdef PADDLE_WITH_CUDA
113 114
      EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
          &device_contexts_, p);
C
chengduoZH 已提交
115
#else
G
GaoWei8 已提交
116
      PADDLE_THROW(platform::errors::Unimplemented(
G
GaoWei8 已提交
117 118
          "CUDAPlace is not supported. Please re-compile with WITH_GPU "
          "option."));
119 120 121 122 123 124 125 126
#endif
    } else if (platform::is_xpu_place(p)) {
#ifdef PADDLE_WITH_XPU
      EmplaceDeviceContext<XPUDeviceContext, XPUPlace>(&device_contexts_, p);
#else
      PADDLE_THROW(
          platform::errors::Unimplemented("XPUPlace is not supported. Please "
                                          "re-compile with WITH_XPU option."));
D
dzhwinter 已提交
127 128 129 130 131
#endif
    }
  }
}

132 133 134 135
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
136
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
137 138 139 140 141 142 143
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

D
dzhwinter 已提交
144
Place CPUDeviceContext::GetPlace() const { return place_; }
145

146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188
#ifdef PADDLE_WITH_XPU
XPUDeviceContext::XPUDeviceContext() { context_ = xpu::create_context(); }

XPUDeviceContext::~XPUDeviceContext() { xpu::destroy_context(context_); }

XPUDeviceContext::XPUDeviceContext(XPUPlace place) : place_(place) {
  int dev_id = -1;
  int ret = xpu_current_device(&dev_id);
  PADDLE_ENFORCE_EQ(ret, XPU_SUCCESS,
                    platform::errors::External(
                        "XPU API return wrong value[%d], please check whether "
                        "Baidu Kunlun Card is properly installed.",
                        ret));
  ret = xpu_set_device(place.device);
  PADDLE_ENFORCE_EQ(ret, XPU_SUCCESS,
                    platform::errors::External(
                        "XPU API return wrong value[%d], please check whether "
                        "Baidu Kunlun Card is properly installed.",
                        ret));
  context_ = xpu::create_context();
  ret = xpu_set_device(dev_id);
  PADDLE_ENFORCE_EQ(ret, XPU_SUCCESS,
                    platform::errors::External(
                        "XPU API return wrong value[%d], please check whether "
                        "Baidu Kunlun Card is properly installed.",
                        ret));
}

void XPUDeviceContext::Wait() const {
  int ret = xpu_set_device(place_.device);
  PADDLE_ENFORCE_EQ(ret, XPU_SUCCESS,
                    platform::errors::External(
                        "XPU API return wrong value[%d], please check whether "
                        "Baidu Kunlun Card is properly installed.",
                        ret));
  xpu_wait();
}

Place XPUDeviceContext::GetPlace() const { return place_; }

xpu::Context* XPUDeviceContext::x_context() const { return context_; }
#endif

189
#ifdef PADDLE_WITH_CUDA
190

Q
init  
qijun 已提交
191 192 193 194 195 196 197
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

D
dzhwinter 已提交
198
  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
199 200 201 202 203 204 205 206 207 208 209 210
    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
211 212 213
    if (UNLIKELY(num_bytes == 0)) {
      return nullptr;
    }
214 215 216
    auto buf = memory::Alloc(place_, num_bytes);
    VLOG(4) << "Eigen allocated at " << buf->ptr() << ", size" << buf->size()
            << " requested " << num_bytes;
217
    void* retv = buf->ptr();
S
sneaxiy 已提交
218 219 220 221
    {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.emplace(retv, std::move(buf));
    }
222
    return retv;
Q
init  
qijun 已提交
223 224
  }

S
sneaxiy 已提交
225 226 227 228 229 230
  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
231 232 233

  void* scratchpad() const override {
    if (scratch_ == NULL) {
Z
Zhang Ting 已提交
234 235 236 237
// windows use an old version of eigen that uses kCudaScratchSize,
// once windows updates eigen to a recent version, the following code
// can use kGpuScratchSize uniformly
#ifdef _WIN32
Q
init  
qijun 已提交
238
      scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int));
Z
Zhang Ting 已提交
239 240 241
#else
      scratch_ = allocate(Eigen::kGpuScratchSize + sizeof(unsigned int));
#endif
Q
init  
qijun 已提交
242 243 244 245 246 247
    }
    return scratch_;
  }

  unsigned int* semaphore() const override {
    if (semaphore_ == NULL) {
Z
Zhang Ting 已提交
248
#ifdef _WIN32
Q
init  
qijun 已提交
249 250
      char* scratch =
          static_cast<char*>(scratchpad()) + Eigen::kCudaScratchSize;
Z
Zhang Ting 已提交
251 252 253
#else
      char* scratch = static_cast<char*>(scratchpad()) + Eigen::kGpuScratchSize;
#endif
Q
init  
qijun 已提交
254
      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
255
      PADDLE_ENFORCE_CUDA_SUCCESS(
Q
init  
qijun 已提交
256 257 258 259 260 261
          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
    }
    return semaphore_;
  }

 private:
D
dzhwinter 已提交
262
  CUDAPlace place_;
Q
init  
qijun 已提交
263 264
  const cudaStream_t* stream_;         // not owned;
  const cudaDeviceProp* device_prop_;  // not owned;
Q
qijun 已提交
265
  mutable void* scratch_;
Q
init  
qijun 已提交
266
  mutable unsigned int* semaphore_;
S
sneaxiy 已提交
267
  mutable std::mutex mtx_;  // to protect allocations_
Y
Yu Yang 已提交
268
  mutable std::unordered_map<void*, memory::AllocationPtr> allocations_;
Q
init  
qijun 已提交
269 270
};

271 272 273 274 275 276 277 278 279
void CudnnWorkspaceHandle::ReallocWorkspace(size_t required_workspace_bytes) {
  if (required_workspace_bytes <= WorkspaceSize()) {
    return;
  }
  // reset allocation first before re-allocate to save memory
  allocation_.reset();
  allocation_ = memory::Alloc(device_context_, required_workspace_bytes);
}

280 281 282 283 284 285 286 287 288 289 290 291
thread_local std::unordered_map<const CUDADeviceContext*,
                                std::shared_ptr<CUDAContext>>
    CUDADeviceContext::thread_ctx_;
thread_local std::mutex CUDADeviceContext::ctx_mtx_;

void CUDAContext::InitEigenContext() {
  eigen_stream_.reset(new EigenCudaStreamDevice());
  eigen_stream_->Reinitialize(&RawStream(), place_);
  eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
}

CUDAContext::CUDAContext(const CUDAPlace& place,
292
                         const stream::Priority& priority) {
293 294 295 296 297 298
  place_ = place;
  CUDADeviceGuard guard(place_.device);
  stream_.reset(new stream::CUDAStream(place, priority));
  InitEigenContext();
  InitCuBlasContext();
  InitCuDNNContext();
G
Guo Sheng 已提交
299
  InitCuSolverContext();
300 301 302 303 304 305
}

CUDAContext::~CUDAContext() {
  CUDADeviceGuard guard(place_.device);
  DestoryCuDNNContext();
  DestoryCuBlasContext();
G
Guo Sheng 已提交
306
  DestoryCuSolverContext();
307 308
}

309
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
Y
Yu Yang 已提交
310
  CUDADeviceGuard guard(place_.device);
C
chengduo 已提交
311 312 313
  compute_capability_ = GetCUDAComputeCapability(place_.device);
  multi_process_ = GetCUDAMultiProcessors(place_.device);
  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
314
  max_grid_dim_size_ = GetGpuMaxGridDimSize(place_.device);
315
  max_threads_per_block_ = GetCUDAMaxThreadsPerBlock(place_.device);
316

C
chengduo 已提交
317 318 319
  driver_version_ = GetCUDADriverVersion(place_.device);
  runtime_version_ = GetCUDARuntimeVersion(place_.device);

320 321
  LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device
                          << ", CUDA Capability: " << compute_capability_
C
chengduo 已提交
322
                          << ", Driver API Version: " << driver_version_ / 1000
323
                          << "." << (driver_version_ % 100) / 10
C
chengduo 已提交
324 325 326
                          << ", Runtime API Version: "
                          << runtime_version_ / 1000 << "."
                          << (runtime_version_ % 100) / 10;
327 328 329
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
330
                          << (cudnn_dso_ver % 1000) / 100 << ".";
S
sneaxiy 已提交
331 332 333

  {
    // Check CUDA/CUDNN version compatiblity
334 335 336 337
    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
338 339 340 341 342 343 344 345 346 347 348 349
    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.";
    }
  }
350
  default_ctx_.reset(new CUDAContext(place_));
351 352 353 354
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
355 356 357 358 359
#if defined(PADDLE_WITH_NCCL)
  if (nccl_comm_) {
    PADDLE_ENFORCE_CUDA_SUCCESS(dynload::ncclCommDestroy(nccl_comm_));
  }
#endif
360 361
}

L
liaogang 已提交
362
Place CUDADeviceContext::GetPlace() const { return place_; }
363

364
void CUDADeviceContext::Wait() const { context()->Stream()->Wait(); }
365

K
Kexin Zhao 已提交
366
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
367
  return compute_capability_;
K
Kexin Zhao 已提交
368 369
}

370
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
371
  return multi_process_ * max_threads_per_mp_;
372 373
}

374 375 376 377 378 379
int CUDADeviceContext::GetSMCount() const { return multi_process_; }

int CUDADeviceContext::GetMaxThreadsPerBlock() const {
  return max_threads_per_block_;
}

380
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
381
  return context()->EigenDevice().get();
382 383
}

384
bool CUDADeviceContext::tensor_core_available() const {
385
  return context()->CublasTensorCoreHandle() != nullptr;
S
sneaxiy 已提交
386 387
}

388 389 390 391
dim3 CUDADeviceContext::GetCUDAMaxGridDimSize() const {
  return max_grid_dim_size_;
}

392 393 394
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
  return context()->CudnnHandle();
}
395

S
sneaxiy 已提交
396
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
397
  return CudnnWorkspaceHandle(*this, &cudnn_handle_mtx_);
398
}
399

G
Guo Sheng 已提交
400 401 402 403
cusolverDnHandle_t CUDADeviceContext::cusolver_dn_handle() const {
  return context()->CusolverDnHandle();
}

404 405 406
cudaStream_t CUDADeviceContext::stream() const {
  return context()->RawStream();
}
Q
qijun 已提交
407

C
chengduoZH 已提交
408 409 410 411 412 413 414 415 416 417 418 419 420 421
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
422
#endif
Q
qijun 已提交
423

T
tensor-tang 已提交
424 425
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
A
Adam 已提交
426 427 428
    : CPUDeviceContext(place),
      engine_(mkldnn::engine::kind::cpu, 0),
      p_blobmap_() {
429 430
  p_blobmap_.reset(new BlobMap());
  p_mutex_.reset(new std::mutex());
T
tensor-tang 已提交
431 432
}

433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
MKLDNNDeviceContextThreadLocals::Body::Body() {
  cur_mkldnn_session_id = kMKLDNNSessionID_Default;
  cur_input_shape_str = "";
  cur_input_shape_cache_capacity = 1;
  cur_paddle_data_layout = paddle::framework::DataLayout::kNCHW;
}

void MKLDNNDeviceContextThreadLocals::Body::set_cur_mkldnn_session_id(
    size_t sid) {
  cur_mkldnn_session_id = sid;
}
size_t MKLDNNDeviceContextThreadLocals::Body::get_cur_mkldnn_session_id(void) {
  return cur_mkldnn_session_id;
}

void MKLDNNDeviceContextThreadLocals::Body::set_cur_input_shape_str(
    std::string input_shape_str) {
450 451
  cur_input_shape_str = input_shape_str;
}
452 453
void MKLDNNDeviceContextThreadLocals::Body::set_cur_input_shape_cache_capacity(
    int input_shape_cache_capacity) {
454 455
  cur_input_shape_cache_capacity = input_shape_cache_capacity;
}
S
Sylwester Fraczek 已提交
456

457 458
void MKLDNNDeviceContextThreadLocals::Body::set_cur_paddle_data_layout(
    framework::DataLayout dl) {
459 460 461
  cur_paddle_data_layout = dl;
}

462 463
framework::DataLayout
MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
464 465 466
  return cur_paddle_data_layout;
}

467 468 469 470 471 472 473 474 475 476 477 478 479 480 481
void MKLDNNDeviceContext::ResetBlobMap() {
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
  if (!block_next_cache_clearing_) {
    VLOG(3) << "Clearing DNNL cache.";
    p_blobmap_->clear();
  } else {
    VLOG(3) << "Prevented Clearing DNNL cache.";
    block_next_cache_clearing_ = false;
  }
}

void MKLDNNDeviceContext::BlockNextCacheClearing() {
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
  VLOG(3) << "Next DNNL cache clearing has been blocked.";
  block_next_cache_clearing_ = true;
482
}
483

484
size_t MKLDNNDeviceContext::GetShapeBlobSize() const {
485
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
486
  BlobMap* pMap = p_blobmap_.get();
487
  auto map_it = pMap->find(tls().cur_mkldnn_session_id);
488
  if (map_it == pMap->end()) {
489 490 491
    PADDLE_THROW(platform::errors::NotFound(
        "MKLDNNDeviceContext don't find cur_mkldnn_session_id: %d.",
        tls().cur_mkldnn_session_id));
492 493 494 495
  }
  return map_it->second->size();
}

496
void MKLDNNDeviceContext::SetBlob(const std::string& name,
497
                                  BlobPtr_t<void> data) const {
498
  BlobMap* pMap = p_blobmap_.get();
499 500
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
501

502
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
503

504
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
T
tensor-tang 已提交
505

506 507
  // Find ShapeBlob for current mkldnn session id.
  auto map_it = pMap->find(sid);
508 509 510

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
511
    sBlob = std::make_shared<ShapeBlob>();
512 513
    (*pMap)[sid] = sBlob;
    VLOG(2) << "SetBlob: sid=" << sid << ", add new sid\n";
514
  } else {
515
    sBlob = map_it->second;
516
  }
T
tensor-tang 已提交
517

518
  // Find KeyBlob for current input shape
519
  auto key_it = sBlob->find(tls().cur_input_shape_str);
520

521
  if (key_it == sBlob->end()) {
522 523
    // In cache clearing mode, cur_input_shape_cache_capacity defines
    // max pblob capacity
524 525
    if ((static_cast<size_t>(sid) ==
         MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_CacheClearing) &&
526
        sBlob->size() &&
527
        (sBlob->size() >=
528
         static_cast<size_t>(tls().cur_input_shape_cache_capacity))) {
529 530 531 532
      VLOG(2) << "sid=" << sid
              << ", remove all blobs of shape: " << sBlob->begin()->first;
      sBlob->erase(sBlob->begin()->first);
    }
533 534
    pBlob = std::make_shared<KeyBlob>();
    (*sBlob)[tls().cur_input_shape_str] = pBlob;
535
  } else {
536
    pBlob = key_it->second;
537 538
  }

539 540 541 542 543 544 545
  // 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
  }
546
  VLOG(2) << "SetBlob: sid=" << sid << ", add blob=" << name << "\n";
547
  // lock will be automatically released when out of scope
548
  return;
T
tensor-tang 已提交
549 550
}

551
MKLDNNDeviceContext::BlobPtr_t<void> MKLDNNDeviceContext::GetBlob(
552
    const std::string& name) const {
553
  BlobMap* pMap = p_blobmap_.get();
554 555
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
556

557
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
558

559
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
560

561 562
  // Find ShapeBlob for current mkldnn session id firstly
  auto map_it = pMap->find(sid);
563
  if (map_it == pMap->end()) {
564
    VLOG(2) << "GetBlob: sid=" << sid << ", miss sid\n";
565 566 567 568 569
    return nullptr;
  }
  sBlob = map_it->second;

  // Find KeyBlob for current input shape secondly
570
  auto sBlob_it = sBlob->find(tls().cur_input_shape_str);
571
  if (sBlob_it == sBlob->end()) {
572
    VLOG(2) << "GetBlob: sid=" << tls().cur_input_shape_str
573 574 575 576
            << ", miss input_shape_str\n";
    return nullptr;
  }
  pBlob = sBlob_it->second;
577 578 579 580

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

581
  if (key_it == pBlob->end()) {
582
    VLOG(2) << "GetBlob sid=" << sid << ", miss blob=" << name << "\n";
583 584
    return nullptr;
  }
585

586
  VLOG(2) << "GetBlob sid=" << sid << ", get blob=" << name << "\n";
587 588
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
589 590 591 592
}

#endif

Q
qijun 已提交
593
}  // namespace platform
Q
qijun 已提交
594
}  // namespace paddle