device_context.cc 20.7 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
#ifdef PADDLE_WITH_CUDA
bool allow_tf32_cublas = true;
void SetAllowTF32Cublas(bool active) { allow_tf32_cublas = active; }
bool AllowTF32Cublas() { return allow_tf32_cublas; }
A
AshburnLee 已提交
77 78 79 80

bool allow_tf32_cudnn = true;
void SetAllowTF32Cudnn(bool active) { allow_tf32_cudnn = active; }
bool AllowTF32Cudnn() { return allow_tf32_cudnn; }
81 82
#endif  // PADDLE_WITH_CUDA

D
dzhwinter 已提交
83 84
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
85
platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
D
dzhwinter 已提交
86 87
  auto it = device_contexts_.find(place);
  if (it == device_contexts_.end()) {
G
GaoWei8 已提交
88 89
    PADDLE_THROW(platform::errors::Unimplemented(
        "Place %s is not supported. Please check that your paddle compiles "
90 91
        "with WITH_GPU or WITH_XPU option or check that your train process "
        "hold the "
G
GaoWei8 已提交
92 93
        "correct gpu_id if you use Executor.",
        place));
D
dzhwinter 已提交
94
  }
95
  return it->second.get().get();
D
dzhwinter 已提交
96 97
}

98 99 100 101 102 103 104 105 106
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`
107
                     return PtrType(new DevCtx(BOOST_GET_CONST(PlaceType, p)));
108
                   }));
C
chengduozh 已提交
109 110
}

D
dzhwinter 已提交
111 112
DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
G
GaoWei8 已提交
113 114 115 116 117
  PADDLE_ENFORCE_GT(
      places.size(), 0,
      platform::errors::InvalidArgument("The number of platform places should "
                                        "be larger than 0. But received %d.",
                                        places.size()));
118
  std::set<Place> set;
Y
Yu Yang 已提交
119 120 121 122 123
  for (auto& p : places) {
    set.insert(p);
  }
  for (auto& p : set) {
    if (platform::is_cpu_place(p)) {
124
#ifdef PADDLE_WITH_MKLDNN
125
      EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
126
#else
127
      EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
128
#endif
Y
Yu Yang 已提交
129
    } else if (platform::is_gpu_place(p)) {
D
dzhwinter 已提交
130
#ifdef PADDLE_WITH_CUDA
131
      EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
D
dzhwinter 已提交
132
#else
G
GaoWei8 已提交
133 134 135
      PADDLE_THROW(
          platform::errors::Unimplemented("CUDAPlace is not supported. Please "
                                          "re-compile with WITH_GPU option."));
C
chengduoZH 已提交
136 137 138
#endif
    } else if (platform::is_cuda_pinned_place(p)) {
#ifdef PADDLE_WITH_CUDA
139 140
      EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
          &device_contexts_, p);
C
chengduoZH 已提交
141
#else
G
GaoWei8 已提交
142
      PADDLE_THROW(platform::errors::Unimplemented(
G
GaoWei8 已提交
143 144
          "CUDAPlace is not supported. Please re-compile with WITH_GPU "
          "option."));
145 146 147 148 149 150 151 152
#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
153 154 155 156 157
#endif
    }
  }
}

158 159 160 161
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
162
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
163 164 165 166 167 168 169
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

D
dzhwinter 已提交
170
Place CPUDeviceContext::GetPlace() const { return place_; }
171

172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
#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

215
#ifdef PADDLE_WITH_CUDA
216

Q
init  
qijun 已提交
217 218 219 220 221 222 223
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

D
dzhwinter 已提交
224
  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
225 226 227 228 229 230 231 232 233 234 235 236
    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
237 238 239
    if (UNLIKELY(num_bytes == 0)) {
      return nullptr;
    }
240 241 242
    auto buf = memory::Alloc(place_, num_bytes);
    VLOG(4) << "Eigen allocated at " << buf->ptr() << ", size" << buf->size()
            << " requested " << num_bytes;
243
    void* retv = buf->ptr();
S
sneaxiy 已提交
244 245 246 247
    {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.emplace(retv, std::move(buf));
    }
248
    return retv;
Q
init  
qijun 已提交
249 250
  }

S
sneaxiy 已提交
251 252 253 254 255 256
  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
257 258 259

  void* scratchpad() const override {
    if (scratch_ == NULL) {
Z
Zhang Ting 已提交
260 261 262 263
// 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
264
      scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int));
Z
Zhang Ting 已提交
265 266 267
#else
      scratch_ = allocate(Eigen::kGpuScratchSize + sizeof(unsigned int));
#endif
Q
init  
qijun 已提交
268 269 270 271 272 273
    }
    return scratch_;
  }

  unsigned int* semaphore() const override {
    if (semaphore_ == NULL) {
Z
Zhang Ting 已提交
274
#ifdef _WIN32
Q
init  
qijun 已提交
275 276
      char* scratch =
          static_cast<char*>(scratchpad()) + Eigen::kCudaScratchSize;
Z
Zhang Ting 已提交
277 278 279
#else
      char* scratch = static_cast<char*>(scratchpad()) + Eigen::kGpuScratchSize;
#endif
Q
init  
qijun 已提交
280
      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
281
      PADDLE_ENFORCE_CUDA_SUCCESS(
Q
init  
qijun 已提交
282 283 284 285 286 287
          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
    }
    return semaphore_;
  }

 private:
D
dzhwinter 已提交
288
  CUDAPlace place_;
Q
init  
qijun 已提交
289 290
  const cudaStream_t* stream_;         // not owned;
  const cudaDeviceProp* device_prop_;  // not owned;
Q
qijun 已提交
291
  mutable void* scratch_;
Q
init  
qijun 已提交
292
  mutable unsigned int* semaphore_;
S
sneaxiy 已提交
293
  mutable std::mutex mtx_;  // to protect allocations_
Y
Yu Yang 已提交
294
  mutable std::unordered_map<void*, memory::AllocationPtr> allocations_;
Q
init  
qijun 已提交
295 296
};

297 298 299 300 301 302 303 304 305
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);
}

306 307 308 309 310 311 312 313 314 315 316 317
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,
318
                         const stream::Priority& priority) {
319 320 321 322 323 324
  place_ = place;
  CUDADeviceGuard guard(place_.device);
  stream_.reset(new stream::CUDAStream(place, priority));
  InitEigenContext();
  InitCuBlasContext();
  InitCuDNNContext();
G
Guo Sheng 已提交
325
  InitCuSolverContext();
326 327 328 329 330 331
}

CUDAContext::~CUDAContext() {
  CUDADeviceGuard guard(place_.device);
  DestoryCuDNNContext();
  DestoryCuBlasContext();
G
Guo Sheng 已提交
332
  DestoryCuSolverContext();
333 334
}

335
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
Y
Yu Yang 已提交
336
  CUDADeviceGuard guard(place_.device);
C
chengduo 已提交
337 338 339
  compute_capability_ = GetCUDAComputeCapability(place_.device);
  multi_process_ = GetCUDAMultiProcessors(place_.device);
  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
340
  max_grid_dim_size_ = GetGpuMaxGridDimSize(place_.device);
341
  max_threads_per_block_ = GetCUDAMaxThreadsPerBlock(place_.device);
342

C
chengduo 已提交
343 344 345
  driver_version_ = GetCUDADriverVersion(place_.device);
  runtime_version_ = GetCUDARuntimeVersion(place_.device);

346
  LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device
347 348 349
                          << ", GPU Compute Capability: "
                          << compute_capability_ / 10 << "."
                          << compute_capability_ % 10
C
chengduo 已提交
350
                          << ", Driver API Version: " << driver_version_ / 1000
351
                          << "." << (driver_version_ % 100) / 10
C
chengduo 已提交
352 353 354
                          << ", Runtime API Version: "
                          << runtime_version_ / 1000 << "."
                          << (runtime_version_ % 100) / 10;
355 356 357
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
358
                          << (cudnn_dso_ver % 1000) / 100 << ".";
S
sneaxiy 已提交
359 360 361

  {
    // Check CUDA/CUDNN version compatiblity
362 363 364 365
    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
366 367 368 369 370 371 372 373 374 375 376 377
    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.";
    }
  }
378
  default_ctx_.reset(new CUDAContext(place_));
379 380 381 382
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
383 384 385 386 387
#if defined(PADDLE_WITH_NCCL)
  if (nccl_comm_) {
    PADDLE_ENFORCE_CUDA_SUCCESS(dynload::ncclCommDestroy(nccl_comm_));
  }
#endif
388 389
}

L
liaogang 已提交
390
Place CUDADeviceContext::GetPlace() const { return place_; }
391

392
void CUDADeviceContext::Wait() const { context()->Stream()->Wait(); }
393

K
Kexin Zhao 已提交
394
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
395
  return compute_capability_;
K
Kexin Zhao 已提交
396 397
}

398
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
399
  return multi_process_ * max_threads_per_mp_;
400 401
}

402 403 404 405 406 407
int CUDADeviceContext::GetSMCount() const { return multi_process_; }

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

408
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
409
  return context()->EigenDevice().get();
410 411
}

412
bool CUDADeviceContext::tensor_core_available() const {
413
  return context()->CublasTensorCoreHandle() != nullptr;
S
sneaxiy 已提交
414 415
}

416 417 418 419
dim3 CUDADeviceContext::GetCUDAMaxGridDimSize() const {
  return max_grid_dim_size_;
}

420 421 422
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
  return context()->CudnnHandle();
}
423

S
sneaxiy 已提交
424
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
425
  return CudnnWorkspaceHandle(*this, &cudnn_handle_mtx_);
426
}
427

G
Guo Sheng 已提交
428 429 430 431
cusolverDnHandle_t CUDADeviceContext::cusolver_dn_handle() const {
  return context()->CusolverDnHandle();
}

432 433 434
cudaStream_t CUDADeviceContext::stream() const {
  return context()->RawStream();
}
Q
qijun 已提交
435

C
chengduoZH 已提交
436 437 438 439 440 441 442 443 444 445 446 447 448 449
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
450
#endif
Q
qijun 已提交
451

T
tensor-tang 已提交
452 453
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
A
Adam 已提交
454 455 456
    : CPUDeviceContext(place),
      engine_(mkldnn::engine::kind::cpu, 0),
      p_blobmap_() {
457 458
  p_blobmap_.reset(new BlobMap());
  p_mutex_.reset(new std::mutex());
T
tensor-tang 已提交
459 460
}

461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477
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) {
478 479
  cur_input_shape_str = input_shape_str;
}
480 481
void MKLDNNDeviceContextThreadLocals::Body::set_cur_input_shape_cache_capacity(
    int input_shape_cache_capacity) {
482 483
  cur_input_shape_cache_capacity = input_shape_cache_capacity;
}
S
Sylwester Fraczek 已提交
484

485 486
void MKLDNNDeviceContextThreadLocals::Body::set_cur_paddle_data_layout(
    framework::DataLayout dl) {
487 488 489
  cur_paddle_data_layout = dl;
}

490 491
framework::DataLayout
MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
492 493 494
  return cur_paddle_data_layout;
}

495 496 497 498 499 500 501 502 503
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;
  }
}

504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
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;
519
}
520

521
size_t MKLDNNDeviceContext::GetShapeBlobSize() const {
522
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
523
  BlobMap* pMap = p_blobmap_.get();
524
  auto map_it = pMap->find(tls().cur_mkldnn_session_id);
525
  if (map_it == pMap->end()) {
526 527 528
    PADDLE_THROW(platform::errors::NotFound(
        "MKLDNNDeviceContext don't find cur_mkldnn_session_id: %d.",
        tls().cur_mkldnn_session_id));
529 530 531 532
  }
  return map_it->second->size();
}

533
void MKLDNNDeviceContext::SetBlob(const std::string& name,
534
                                  BlobPtr_t<void> data) const {
535
  BlobMap* pMap = p_blobmap_.get();
536 537
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
538

539
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
540

541
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
T
tensor-tang 已提交
542

543 544
  // Find ShapeBlob for current mkldnn session id.
  auto map_it = pMap->find(sid);
545 546 547

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
548
    sBlob = std::make_shared<ShapeBlob>();
549 550
    (*pMap)[sid] = sBlob;
    VLOG(2) << "SetBlob: sid=" << sid << ", add new sid\n";
551
  } else {
552
    sBlob = map_it->second;
553
  }
T
tensor-tang 已提交
554

555
  // Find KeyBlob for current input shape
556
  auto key_it = sBlob->find(tls().cur_input_shape_str);
557

558
  if (key_it == sBlob->end()) {
559 560
    // In cache clearing mode, cur_input_shape_cache_capacity defines
    // max pblob capacity
561 562
    if ((static_cast<size_t>(sid) ==
         MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_CacheClearing) &&
563
        sBlob->size() &&
564
        (sBlob->size() >=
565
         static_cast<size_t>(tls().cur_input_shape_cache_capacity))) {
566 567 568 569
      VLOG(2) << "sid=" << sid
              << ", remove all blobs of shape: " << sBlob->begin()->first;
      sBlob->erase(sBlob->begin()->first);
    }
570 571
    pBlob = std::make_shared<KeyBlob>();
    (*sBlob)[tls().cur_input_shape_str] = pBlob;
572
  } else {
573
    pBlob = key_it->second;
574 575
  }

576 577 578 579 580 581 582
  // 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
  }
583
  VLOG(2) << "SetBlob: sid=" << sid << ", add blob=" << name << "\n";
584
  // lock will be automatically released when out of scope
585
  return;
T
tensor-tang 已提交
586 587
}

588 589 590 591 592 593 594 595 596 597
unsigned int MKLDNNDeviceContext::GetCachedObjectsNumber(void) {
  unsigned int num_entries = 0;
  for (auto const& l3 : *p_blobmap_) {
    for (auto const& l2 : *(l3.second)) {
      num_entries += (l2.second)->size();
    }
  }
  return num_entries;
}

598
MKLDNNDeviceContext::BlobPtr_t<void> MKLDNNDeviceContext::GetBlob(
599
    const std::string& name) const {
600
  BlobMap* pMap = p_blobmap_.get();
601 602
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
603

604
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
605

606
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
607

608 609
  // Find ShapeBlob for current mkldnn session id firstly
  auto map_it = pMap->find(sid);
610
  if (map_it == pMap->end()) {
611
    VLOG(2) << "GetBlob: sid=" << sid << ", miss sid\n";
612 613 614 615 616
    return nullptr;
  }
  sBlob = map_it->second;

  // Find KeyBlob for current input shape secondly
617
  auto sBlob_it = sBlob->find(tls().cur_input_shape_str);
618
  if (sBlob_it == sBlob->end()) {
619
    VLOG(2) << "GetBlob: sid=" << tls().cur_input_shape_str
620 621 622 623
            << ", miss input_shape_str\n";
    return nullptr;
  }
  pBlob = sBlob_it->second;
624 625 626 627

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

628
  if (key_it == pBlob->end()) {
629
    VLOG(2) << "GetBlob sid=" << sid << ", miss blob=" << name << "\n";
630 631
    return nullptr;
  }
632

633
  VLOG(2) << "GetBlob sid=" << sid << ", get blob=" << name << "\n";
634 635
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
636 637 638 639
}

#endif

Q
qijun 已提交
640
}  // namespace platform
Q
qijun 已提交
641
}  // namespace paddle