device_context.cc 20.3 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
namespace paddle {
namespace memory {

AllocationPtr Alloc(const platform::DeviceContext& dev_ctx, size_t size) {
  auto place = dev_ctx.GetPlace();
32
  if (size == 0) {
33 34
    return Alloc(place, size);
  }
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55

  if (platform::is_gpu_place(place)) {
#ifdef PADDLE_WITH_CUDA
    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
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Paddle can't use CUDA device since it's not compiled with CUDA,"
        "Please recompile or reinstall Paddle with GPU support."));
#endif
  } else if (platform::is_xpu_place(place)) {
#ifdef PADDLE_WITH_XPU
    // TODO(liuyuhui): Consider xpu stream later
56 57
    return Alloc(place, size);
#else
58 59 60
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Paddle can't use XPU device since it's not compiled with XPU,"
        "Please recompile or reinstall Paddle with XPU support."));
61
#endif
62 63 64
  } else {
    return Alloc(place, size);
  }
65 66 67 68 69
}

}  // namespace memory
}  // namespace paddle

Q
qijun 已提交
70 71 72
namespace paddle {
namespace platform {

73 74 75 76 77 78
#ifdef PADDLE_WITH_CUDA
bool allow_tf32_cublas = true;
void SetAllowTF32Cublas(bool active) { allow_tf32_cublas = active; }
bool AllowTF32Cublas() { return allow_tf32_cublas; }
#endif  // PADDLE_WITH_CUDA

D
dzhwinter 已提交
79 80
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
81
platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
D
dzhwinter 已提交
82 83
  auto it = device_contexts_.find(place);
  if (it == device_contexts_.end()) {
G
GaoWei8 已提交
84 85
    PADDLE_THROW(platform::errors::Unimplemented(
        "Place %s is not supported. Please check that your paddle compiles "
86 87
        "with WITH_GPU or WITH_XPU option or check that your train process "
        "hold the "
G
GaoWei8 已提交
88 89
        "correct gpu_id if you use Executor.",
        place));
D
dzhwinter 已提交
90
  }
91
  return it->second.get().get();
D
dzhwinter 已提交
92 93
}

94 95 96 97 98 99 100 101 102
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`
103
                     return PtrType(new DevCtx(BOOST_GET_CONST(PlaceType, p)));
104
                   }));
C
chengduozh 已提交
105 106
}

D
dzhwinter 已提交
107 108
DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
G
GaoWei8 已提交
109 110 111 112 113
  PADDLE_ENFORCE_GT(
      places.size(), 0,
      platform::errors::InvalidArgument("The number of platform places should "
                                        "be larger than 0. But received %d.",
                                        places.size()));
114
  std::set<Place> set;
Y
Yu Yang 已提交
115 116 117 118 119
  for (auto& p : places) {
    set.insert(p);
  }
  for (auto& p : set) {
    if (platform::is_cpu_place(p)) {
120
#ifdef PADDLE_WITH_MKLDNN
121
      EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
122
#else
123
      EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
124
#endif
Y
Yu Yang 已提交
125
    } else if (platform::is_gpu_place(p)) {
D
dzhwinter 已提交
126
#ifdef PADDLE_WITH_CUDA
127
      EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
D
dzhwinter 已提交
128
#else
G
GaoWei8 已提交
129 130 131
      PADDLE_THROW(
          platform::errors::Unimplemented("CUDAPlace is not supported. Please "
                                          "re-compile with WITH_GPU option."));
C
chengduoZH 已提交
132 133 134
#endif
    } else if (platform::is_cuda_pinned_place(p)) {
#ifdef PADDLE_WITH_CUDA
135 136
      EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
          &device_contexts_, p);
C
chengduoZH 已提交
137
#else
G
GaoWei8 已提交
138
      PADDLE_THROW(platform::errors::Unimplemented(
G
GaoWei8 已提交
139 140
          "CUDAPlace is not supported. Please re-compile with WITH_GPU "
          "option."));
141 142 143 144 145 146 147 148
#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
149 150 151 152 153
#endif
    }
  }
}

154 155 156 157
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
158
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
159 160 161 162 163 164 165
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

D
dzhwinter 已提交
166
Place CPUDeviceContext::GetPlace() const { return place_; }
167

168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210
#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

211
#ifdef PADDLE_WITH_CUDA
212

Q
init  
qijun 已提交
213 214 215 216 217 218 219
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

D
dzhwinter 已提交
220
  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
221 222 223 224 225 226 227 228 229 230 231 232
    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
233 234 235
    if (UNLIKELY(num_bytes == 0)) {
      return nullptr;
    }
236 237 238
    auto buf = memory::Alloc(place_, num_bytes);
    VLOG(4) << "Eigen allocated at " << buf->ptr() << ", size" << buf->size()
            << " requested " << num_bytes;
239
    void* retv = buf->ptr();
S
sneaxiy 已提交
240 241 242 243
    {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.emplace(retv, std::move(buf));
    }
244
    return retv;
Q
init  
qijun 已提交
245 246
  }

S
sneaxiy 已提交
247 248 249 250 251 252
  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
253 254 255

  void* scratchpad() const override {
    if (scratch_ == NULL) {
Z
Zhang Ting 已提交
256 257 258 259
// 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
260
      scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int));
Z
Zhang Ting 已提交
261 262 263
#else
      scratch_ = allocate(Eigen::kGpuScratchSize + sizeof(unsigned int));
#endif
Q
init  
qijun 已提交
264 265 266 267 268 269
    }
    return scratch_;
  }

  unsigned int* semaphore() const override {
    if (semaphore_ == NULL) {
Z
Zhang Ting 已提交
270
#ifdef _WIN32
Q
init  
qijun 已提交
271 272
      char* scratch =
          static_cast<char*>(scratchpad()) + Eigen::kCudaScratchSize;
Z
Zhang Ting 已提交
273 274 275
#else
      char* scratch = static_cast<char*>(scratchpad()) + Eigen::kGpuScratchSize;
#endif
Q
init  
qijun 已提交
276
      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
277
      PADDLE_ENFORCE_CUDA_SUCCESS(
Q
init  
qijun 已提交
278 279 280 281 282 283
          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
    }
    return semaphore_;
  }

 private:
D
dzhwinter 已提交
284
  CUDAPlace place_;
Q
init  
qijun 已提交
285 286
  const cudaStream_t* stream_;         // not owned;
  const cudaDeviceProp* device_prop_;  // not owned;
Q
qijun 已提交
287
  mutable void* scratch_;
Q
init  
qijun 已提交
288
  mutable unsigned int* semaphore_;
S
sneaxiy 已提交
289
  mutable std::mutex mtx_;  // to protect allocations_
Y
Yu Yang 已提交
290
  mutable std::unordered_map<void*, memory::AllocationPtr> allocations_;
Q
init  
qijun 已提交
291 292
};

293 294 295 296 297 298 299 300 301
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);
}

302 303 304 305 306 307 308 309 310 311 312 313
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,
314
                         const stream::Priority& priority) {
315 316 317 318 319 320
  place_ = place;
  CUDADeviceGuard guard(place_.device);
  stream_.reset(new stream::CUDAStream(place, priority));
  InitEigenContext();
  InitCuBlasContext();
  InitCuDNNContext();
G
Guo Sheng 已提交
321
  InitCuSolverContext();
322 323 324 325 326 327
}

CUDAContext::~CUDAContext() {
  CUDADeviceGuard guard(place_.device);
  DestoryCuDNNContext();
  DestoryCuBlasContext();
G
Guo Sheng 已提交
328
  DestoryCuSolverContext();
329 330
}

331
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
Y
Yu Yang 已提交
332
  CUDADeviceGuard guard(place_.device);
C
chengduo 已提交
333 334 335
  compute_capability_ = GetCUDAComputeCapability(place_.device);
  multi_process_ = GetCUDAMultiProcessors(place_.device);
  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
336
  max_grid_dim_size_ = GetGpuMaxGridDimSize(place_.device);
337
  max_threads_per_block_ = GetCUDAMaxThreadsPerBlock(place_.device);
338

C
chengduo 已提交
339 340 341
  driver_version_ = GetCUDADriverVersion(place_.device);
  runtime_version_ = GetCUDARuntimeVersion(place_.device);

342
  LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device
343 344 345
                          << ", GPU Compute Capability: "
                          << compute_capability_ / 10 << "."
                          << compute_capability_ % 10
C
chengduo 已提交
346
                          << ", Driver API Version: " << driver_version_ / 1000
347
                          << "." << (driver_version_ % 100) / 10
C
chengduo 已提交
348 349 350
                          << ", Runtime API Version: "
                          << runtime_version_ / 1000 << "."
                          << (runtime_version_ % 100) / 10;
351 352 353
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
354
                          << (cudnn_dso_ver % 1000) / 100 << ".";
S
sneaxiy 已提交
355 356 357

  {
    // Check CUDA/CUDNN version compatiblity
358 359 360 361
    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
362 363 364 365 366 367 368 369 370 371 372 373
    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.";
    }
  }
374
  default_ctx_.reset(new CUDAContext(place_));
375 376 377 378
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
379 380 381 382 383
#if defined(PADDLE_WITH_NCCL)
  if (nccl_comm_) {
    PADDLE_ENFORCE_CUDA_SUCCESS(dynload::ncclCommDestroy(nccl_comm_));
  }
#endif
384 385
}

L
liaogang 已提交
386
Place CUDADeviceContext::GetPlace() const { return place_; }
387

388
void CUDADeviceContext::Wait() const { context()->Stream()->Wait(); }
389

K
Kexin Zhao 已提交
390
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
391
  return compute_capability_;
K
Kexin Zhao 已提交
392 393
}

394
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
395
  return multi_process_ * max_threads_per_mp_;
396 397
}

398 399 400 401 402 403
int CUDADeviceContext::GetSMCount() const { return multi_process_; }

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

404
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
405
  return context()->EigenDevice().get();
406 407
}

408
bool CUDADeviceContext::tensor_core_available() const {
409
  return context()->CublasTensorCoreHandle() != nullptr;
S
sneaxiy 已提交
410 411
}

412 413 414 415
dim3 CUDADeviceContext::GetCUDAMaxGridDimSize() const {
  return max_grid_dim_size_;
}

416 417 418
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
  return context()->CudnnHandle();
}
419

S
sneaxiy 已提交
420
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
421
  return CudnnWorkspaceHandle(*this, &cudnn_handle_mtx_);
422
}
423

G
Guo Sheng 已提交
424 425 426 427
cusolverDnHandle_t CUDADeviceContext::cusolver_dn_handle() const {
  return context()->CusolverDnHandle();
}

428 429 430
cudaStream_t CUDADeviceContext::stream() const {
  return context()->RawStream();
}
Q
qijun 已提交
431

C
chengduoZH 已提交
432 433 434 435 436 437 438 439 440 441 442 443 444 445
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
446
#endif
Q
qijun 已提交
447

T
tensor-tang 已提交
448 449
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
A
Adam 已提交
450 451 452
    : CPUDeviceContext(place),
      engine_(mkldnn::engine::kind::cpu, 0),
      p_blobmap_() {
453 454
  p_blobmap_.reset(new BlobMap());
  p_mutex_.reset(new std::mutex());
T
tensor-tang 已提交
455 456
}

457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
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) {
474 475
  cur_input_shape_str = input_shape_str;
}
476 477
void MKLDNNDeviceContextThreadLocals::Body::set_cur_input_shape_cache_capacity(
    int input_shape_cache_capacity) {
478 479
  cur_input_shape_cache_capacity = input_shape_cache_capacity;
}
S
Sylwester Fraczek 已提交
480

481 482
void MKLDNNDeviceContextThreadLocals::Body::set_cur_paddle_data_layout(
    framework::DataLayout dl) {
483 484 485
  cur_paddle_data_layout = dl;
}

486 487
framework::DataLayout
MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
488 489 490
  return cur_paddle_data_layout;
}

491 492 493 494 495 496 497 498 499
void MKLDNNDeviceContextThreadLocals::Body::log_lib_version(void) {
  if (!said_once) {
    said_once = true;
    auto dv = dnnl::version();
    LOG(INFO) << "oneDNN v" << dv->major << "." << dv->minor << "."
              << dv->patch;
  }
}

500 501 502 503 504 505 506 507 508 509 510 511 512 513 514
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;
515
}
516

517
size_t MKLDNNDeviceContext::GetShapeBlobSize() const {
518
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
519
  BlobMap* pMap = p_blobmap_.get();
520
  auto map_it = pMap->find(tls().cur_mkldnn_session_id);
521
  if (map_it == pMap->end()) {
522 523 524
    PADDLE_THROW(platform::errors::NotFound(
        "MKLDNNDeviceContext don't find cur_mkldnn_session_id: %d.",
        tls().cur_mkldnn_session_id));
525 526 527 528
  }
  return map_it->second->size();
}

529
void MKLDNNDeviceContext::SetBlob(const std::string& name,
530
                                  BlobPtr_t<void> data) const {
531
  BlobMap* pMap = p_blobmap_.get();
532 533
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
534

535
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
536

537
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
T
tensor-tang 已提交
538

539 540
  // Find ShapeBlob for current mkldnn session id.
  auto map_it = pMap->find(sid);
541 542 543

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
544
    sBlob = std::make_shared<ShapeBlob>();
545 546
    (*pMap)[sid] = sBlob;
    VLOG(2) << "SetBlob: sid=" << sid << ", add new sid\n";
547
  } else {
548
    sBlob = map_it->second;
549
  }
T
tensor-tang 已提交
550

551
  // Find KeyBlob for current input shape
552
  auto key_it = sBlob->find(tls().cur_input_shape_str);
553

554
  if (key_it == sBlob->end()) {
555 556
    // In cache clearing mode, cur_input_shape_cache_capacity defines
    // max pblob capacity
557 558
    if ((static_cast<size_t>(sid) ==
         MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_CacheClearing) &&
559
        sBlob->size() &&
560
        (sBlob->size() >=
561
         static_cast<size_t>(tls().cur_input_shape_cache_capacity))) {
562 563 564 565
      VLOG(2) << "sid=" << sid
              << ", remove all blobs of shape: " << sBlob->begin()->first;
      sBlob->erase(sBlob->begin()->first);
    }
566 567
    pBlob = std::make_shared<KeyBlob>();
    (*sBlob)[tls().cur_input_shape_str] = pBlob;
568
  } else {
569
    pBlob = key_it->second;
570 571
  }

572 573 574 575 576 577 578
  // 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
  }
579
  VLOG(2) << "SetBlob: sid=" << sid << ", add blob=" << name << "\n";
580
  // lock will be automatically released when out of scope
581
  return;
T
tensor-tang 已提交
582 583
}

584
MKLDNNDeviceContext::BlobPtr_t<void> MKLDNNDeviceContext::GetBlob(
585
    const std::string& name) const {
586
  BlobMap* pMap = p_blobmap_.get();
587 588
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
589

590
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
591

592
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
593

594 595
  // Find ShapeBlob for current mkldnn session id firstly
  auto map_it = pMap->find(sid);
596
  if (map_it == pMap->end()) {
597
    VLOG(2) << "GetBlob: sid=" << sid << ", miss sid\n";
598 599 600 601 602
    return nullptr;
  }
  sBlob = map_it->second;

  // Find KeyBlob for current input shape secondly
603
  auto sBlob_it = sBlob->find(tls().cur_input_shape_str);
604
  if (sBlob_it == sBlob->end()) {
605
    VLOG(2) << "GetBlob: sid=" << tls().cur_input_shape_str
606 607 608 609
            << ", miss input_shape_str\n";
    return nullptr;
  }
  pBlob = sBlob_it->second;
610 611 612 613

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

614
  if (key_it == pBlob->end()) {
615
    VLOG(2) << "GetBlob sid=" << sid << ", miss blob=" << name << "\n";
616 617
    return nullptr;
  }
618

619
  VLOG(2) << "GetBlob sid=" << sid << ", get blob=" << name << "\n";
620 621
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
622 623 624 625
}

#endif

Q
qijun 已提交
626
}  // namespace platform
Q
qijun 已提交
627
}  // namespace paddle