device_context.cc 27.1 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

15
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
16
#include "paddle/fluid/memory/allocation/cuda_device_context_allocator.h"
S
sneaxiy 已提交
17
#include "paddle/fluid/platform/cuda_device_guard.h"
18
#endif
19
#include "glog/logging.h"
20
#include "paddle/fluid/platform/profiler.h"
21

22 23 24 25 26
namespace paddle {
namespace memory {

AllocationPtr Alloc(const platform::DeviceContext& dev_ctx, size_t size) {
  auto place = dev_ctx.GetPlace();
27
  if (size == 0) {
28 29
    return Alloc(place, size);
  }
30 31

  if (platform::is_gpu_place(place)) {
32
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
    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
51 52
    return Alloc(place, size);
#else
53 54 55
    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."));
56
#endif
57 58 59
  } else {
    return Alloc(place, size);
  }
60 61 62 63 64
}

}  // namespace memory
}  // namespace paddle

Q
qijun 已提交
65 66 67
namespace paddle {
namespace platform {

68
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
69 70 71
bool allow_tf32_cublas = true;
void SetAllowTF32Cublas(bool active) { allow_tf32_cublas = active; }
bool AllowTF32Cublas() { return allow_tf32_cublas; }
A
AshburnLee 已提交
72 73 74 75

bool allow_tf32_cudnn = true;
void SetAllowTF32Cudnn(bool active) { allow_tf32_cudnn = active; }
bool AllowTF32Cudnn() { return allow_tf32_cudnn; }
76 77
#endif  // PADDLE_WITH_CUDA

D
dzhwinter 已提交
78 79
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
80
platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
81
  VLOG(4) << "DeviceContextPool Get: " << 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, WITH_XPU or WITH_ASCEND_CL option or check that "
        "your train process set the correct device id if you use Executor.",
G
GaoWei8 已提交
88
        place));
D
dzhwinter 已提交
89
  }
90
  return it->second.get().get();
D
dzhwinter 已提交
91 92
}

93 94 95 96 97 98 99 100 101
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`
102
                     return PtrType(new DevCtx(BOOST_GET_CONST(PlaceType, p)));
103
                   }));
C
chengduozh 已提交
104 105
}

D
dzhwinter 已提交
106 107
DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
G
GaoWei8 已提交
108 109 110 111 112
  PADDLE_ENFORCE_GT(
      places.size(), 0,
      platform::errors::InvalidArgument("The number of platform places should "
                                        "be larger than 0. But received %d.",
                                        places.size()));
113
  std::set<Place> set;
Y
Yu Yang 已提交
114 115 116 117 118
  for (auto& p : places) {
    set.insert(p);
  }
  for (auto& p : set) {
    if (platform::is_cpu_place(p)) {
119
#ifdef PADDLE_WITH_MKLDNN
120
      EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
121
#else
122
      EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
123
#endif
Y
Yu Yang 已提交
124
    } else if (platform::is_gpu_place(p)) {
125
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
126
      EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
D
dzhwinter 已提交
127
#else
G
GaoWei8 已提交
128 129 130
      PADDLE_THROW(
          platform::errors::Unimplemented("CUDAPlace is not supported. Please "
                                          "re-compile with WITH_GPU option."));
C
chengduoZH 已提交
131 132
#endif
    } else if (platform::is_cuda_pinned_place(p)) {
133
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
134 135
      EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
          &device_contexts_, p);
C
chengduoZH 已提交
136
#else
G
GaoWei8 已提交
137
      PADDLE_THROW(platform::errors::Unimplemented(
G
GaoWei8 已提交
138 139
          "CUDAPlace is not supported. Please re-compile with WITH_GPU "
          "option."));
140 141 142 143 144 145 146 147
#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."));
148 149 150 151 152 153 154 155
#endif
    } else if (platform::is_npu_place(p)) {
#ifdef PADDLE_WITH_ASCEND_CL
      EmplaceDeviceContext<NPUDeviceContext, NPUPlace>(&device_contexts_, p);
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "NPUPlace is not supported. Please "
          "re-compile with WITH_ASCEND_CL option."));
156 157 158 159 160 161 162 163 164 165
#endif
    } else if (platform::is_npu_pinned_place(p)) {
#ifdef PADDLE_WITH_ASCEND_CL
      EmplaceDeviceContext<NPUPinnedDeviceContext, NPUPinnedPlace>(
          &device_contexts_, p);
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "NPUPinnedPlace is not supported. Please re-compile with "
          "WITH_ASCEND_CL "
          "option."));
D
dzhwinter 已提交
166 167 168 169 170
#endif
    }
  }
}

171 172 173 174
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
175
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
176 177 178 179 180 181 182
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

D
dzhwinter 已提交
183
Place CPUDeviceContext::GetPlace() const { return place_; }
184

185 186 187
#ifdef PADDLE_WITH_XPU
XPUDeviceContext::XPUDeviceContext() { context_ = xpu::create_context(); }

188
XPUDeviceContext::~XPUDeviceContext() {}
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203

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));
204 205 206

  LOG_FIRST_N(WARNING, 1) << "Please NOTE: xpu device: " << place_.device;

207
  context_ = xpu::create_context();
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
  const int MAX_XPU_NUM = 16;
  const int l3_size = 13.5 * 1024 * 1024;
  static void* l3ptrs[MAX_XPU_NUM] = {nullptr};

  auto selected_xpus = GetXPUSelectedDevices();
  for (unsigned int i = 0; i < selected_xpus.size(); i++) {
    if (place.device == selected_xpus[i]) {
      if (l3ptrs[place.device] == nullptr) {
        xpu_malloc(static_cast<void**>(&l3ptrs[place.device]), l3_size,
                   XPU_MEM_L3);
      }
      if (l3ptrs[place.device] != nullptr) {
        context_->_l3_mgr.set(l3ptrs[place.device], l3_size);
        VLOG(3) << "xpu place " << place.device << " set l3 size " << l3_size;
      }
      break;
    }
225
  }
226

227 228 229 230 231 232 233 234 235 236 237 238 239 240 241
  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));
242
  xpu_wait(context_->xpu_stream);
243 244 245 246 247 248 249
}

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

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

250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
#ifdef PADDLE_WITH_ASCEND_CL
NPUDeviceContext::NPUDeviceContext(NPUPlace place) : place_(place) {
  NPUDeviceGuard guard(place_.device);
  // PADDLE_ENFORCE_NPU_SUCCESS(aclrtCreateContext(&context_, place_.device));
  // NOTE(zhiqiu): Usually, no need to create context explicitly,
  // ACL creates a default context which contains 1 default stream
  // and 1 sync strean after aclrtSetDevice.
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtGetCurrentContext(&context_));
  stream_.reset(new stream::NPUStream(place));
}

NPUDeviceContext::~NPUDeviceContext() {
  // NPUDeviceGuard guard(place_.device);
  // PADDLE_ENFORCE_NPU_SUCCESS(aclrtDestroyContext(context_));
}
265

266
void NPUDeviceContext::Wait() const {
267 268 269
  platform::RecordEvent record_event("NPUDeviceContext/wait");
  VLOG(4) << "NPU context(" << this << ")  Wait";
  stream_->Wait();
270 271 272 273 274 275 276
}

aclrtStream NPUDeviceContext::stream() const { return stream_->raw_stream(); }

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

aclrtContext NPUDeviceContext::context() const { return context_; }
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292

NPUPinnedDeviceContext::NPUPinnedDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

NPUPinnedDeviceContext::NPUPinnedDeviceContext(NPUPinnedPlace place)
    : place_(place) {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

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

293 294 295
#endif

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
Q
init  
qijun 已提交
296 297 298 299 300 301 302
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

303
  void Reinitialize(const gpuStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
304 305 306 307 308
    stream_ = cuda_stream;
    place_ = place;
    device_prop_ = &Eigen::m_deviceProperties[place.device];
  }

309
  const gpuStream_t& stream() const override { return *stream_; }
Q
init  
qijun 已提交
310

311 312 313
#ifdef PADDLE_WITH_HIP
  const hipDeviceProp_t& deviceProperties() const override {
#else
Q
init  
qijun 已提交
314
  const cudaDeviceProp& deviceProperties() const override {
315
#endif
Q
init  
qijun 已提交
316 317 318 319
    return *device_prop_;
  }

  void* allocate(size_t num_bytes) const override {
S
sneaxiy 已提交
320 321 322
    if (UNLIKELY(num_bytes == 0)) {
      return nullptr;
    }
323 324 325
    auto buf = memory::Alloc(place_, num_bytes);
    VLOG(4) << "Eigen allocated at " << buf->ptr() << ", size" << buf->size()
            << " requested " << num_bytes;
326
    void* retv = buf->ptr();
S
sneaxiy 已提交
327 328 329 330
    {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.emplace(retv, std::move(buf));
    }
331
    return retv;
Q
init  
qijun 已提交
332 333
  }

S
sneaxiy 已提交
334 335 336 337 338 339
  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
340 341 342

  void* scratchpad() const override {
    if (scratch_ == NULL) {
Z
Zhang Ting 已提交
343
      scratch_ = allocate(Eigen::kGpuScratchSize + sizeof(unsigned int));
Q
init  
qijun 已提交
344 345 346 347 348 349
    }
    return scratch_;
  }

  unsigned int* semaphore() const override {
    if (semaphore_ == NULL) {
Z
Zhang Ting 已提交
350
      char* scratch = static_cast<char*>(scratchpad()) + Eigen::kGpuScratchSize;
Q
init  
qijun 已提交
351
      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
352 353 354 355
#ifdef PADDLE_WITH_HIP
      PADDLE_ENFORCE_CUDA_SUCCESS(
          hipMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
#else
356
      PADDLE_ENFORCE_CUDA_SUCCESS(
Q
init  
qijun 已提交
357
          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
358
#endif
Q
init  
qijun 已提交
359 360 361 362 363
    }
    return semaphore_;
  }

 private:
D
dzhwinter 已提交
364
  CUDAPlace place_;
365 366 367 368
  const gpuStream_t* stream_;  // not owned;
#ifdef PADDLE_WITH_HIP
  const hipDeviceProp_t* device_prop_;
#else
Q
init  
qijun 已提交
369
  const cudaDeviceProp* device_prop_;  // not owned;
370
#endif
Q
qijun 已提交
371
  mutable void* scratch_;
Q
init  
qijun 已提交
372
  mutable unsigned int* semaphore_;
S
sneaxiy 已提交
373
  mutable std::mutex mtx_;  // to protect allocations_
Y
Yu Yang 已提交
374
  mutable std::unordered_map<void*, memory::AllocationPtr> allocations_;
Q
init  
qijun 已提交
375 376
};

377 378 379 380 381 382 383 384 385
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);
}

386 387 388 389 390 391 392 393 394 395 396 397
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,
398
                         const stream::Priority& priority) {
399 400 401 402 403 404
  place_ = place;
  CUDADeviceGuard guard(place_.device);
  stream_.reset(new stream::CUDAStream(place, priority));
  InitEigenContext();
  InitCuBlasContext();
  InitCuDNNContext();
405
#ifndef PADDLE_WITH_HIP
G
Guo Sheng 已提交
406
  InitCuSolverContext();
407
#endif
408 409 410 411 412 413
}

CUDAContext::~CUDAContext() {
  CUDADeviceGuard guard(place_.device);
  DestoryCuDNNContext();
  DestoryCuBlasContext();
414
#ifndef PADDLE_WITH_HIP
G
Guo Sheng 已提交
415
  DestoryCuSolverContext();
416
#endif
417 418
}

419
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
Y
Yu Yang 已提交
420
  CUDADeviceGuard guard(place_.device);
C
chengduo 已提交
421 422 423
  compute_capability_ = GetCUDAComputeCapability(place_.device);
  multi_process_ = GetCUDAMultiProcessors(place_.device);
  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
424
  max_grid_dim_size_ = GetGpuMaxGridDimSize(place_.device);
425
  max_threads_per_block_ = GetCUDAMaxThreadsPerBlock(place_.device);
426

C
chengduo 已提交
427 428 429
  driver_version_ = GetCUDADriverVersion(place_.device);
  runtime_version_ = GetCUDARuntimeVersion(place_.device);

430
  LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device
431 432 433
                          << ", GPU Compute Capability: "
                          << compute_capability_ / 10 << "."
                          << compute_capability_ % 10
C
chengduo 已提交
434
                          << ", Driver API Version: " << driver_version_ / 1000
435
                          << "." << (driver_version_ % 100) / 10
C
chengduo 已提交
436 437 438
                          << ", Runtime API Version: "
                          << runtime_version_ / 1000 << "."
                          << (runtime_version_ % 100) / 10;
439 440 441 442 443 444 445 446
#ifdef PADDLE_WITH_HIP
  size_t version_major, version_minor, version_patch;
  PADDLE_ENFORCE_CUDA_SUCCESS(dynload::miopenGetVersion(
      &version_major, &version_minor, &version_patch));
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", MIOpen Version: " << version_major << "."
                          << version_minor << "." << version_patch;
#else
447 448 449
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
450
                          << (cudnn_dso_ver % 1000) / 100 << ".";
451
#endif
S
sneaxiy 已提交
452 453
  {
    // Check CUDA/CUDNN version compatiblity
454 455
    auto local_cuda_version =
        (driver_version_ / 1000) * 10 + (driver_version_ % 100) / 10;
456 457 458
#ifdef PADDLE_WITH_HIP
    auto compile_cuda_version = (HIP_VERSION / 100) * 10 + (HIP_VERSION % 10);
#else
459 460
    auto compile_cuda_version =
        (CUDA_VERSION / 1000) * 10 + (CUDA_VERSION % 100) / 10;
461
#endif
S
sneaxiy 已提交
462 463 464 465 466 467 468 469 470 471 472 473
    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.";
    }
  }
474
  default_ctx_.reset(new CUDAContext(place_));
475 476 477 478
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
479
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
480 481 482 483
  if (nccl_comm_) {
    PADDLE_ENFORCE_CUDA_SUCCESS(dynload::ncclCommDestroy(nccl_comm_));
  }
#endif
484 485
}

L
liaogang 已提交
486
Place CUDADeviceContext::GetPlace() const { return place_; }
487

488
void CUDADeviceContext::Wait() const { context()->Stream()->Wait(); }
489

K
Kexin Zhao 已提交
490
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
491
  return compute_capability_;
K
Kexin Zhao 已提交
492 493
}

494
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
495
  return multi_process_ * max_threads_per_mp_;
496 497
}

498 499 500 501 502 503
int CUDADeviceContext::GetSMCount() const { return multi_process_; }

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

504
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
505
  return context()->EigenDevice().get();
506 507
}

508
bool CUDADeviceContext::tensor_core_available() const {
509
  return context()->CublasTensorCoreHandle() != nullptr;
S
sneaxiy 已提交
510 511
}

512 513 514 515
dim3 CUDADeviceContext::GetCUDAMaxGridDimSize() const {
  return max_grid_dim_size_;
}

516 517 518
#ifdef PADDLE_WITH_HIP
miopenHandle_t CUDADeviceContext::cudnn_handle() const {
#else
519
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
520
#endif
521 522
  return context()->CudnnHandle();
}
523

524 525 526 527 528
#ifdef PADDLE_WITH_HIP
rocblas_handle CUDADeviceContext::cublas_handle() const {
  return context()->CublasHandle()->GetCublasHandle();
}
#else
529 530 531
cublasHandle_t CUDADeviceContext::cublas_handle() const {
  return context()->CublasHandle()->GetCublasHandle();
}
532
#endif
533

S
sneaxiy 已提交
534
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
535
  return CudnnWorkspaceHandle(*this, &cudnn_handle_mtx_);
536
}
537

538
#ifndef PADDLE_WITH_HIP
G
Guo Sheng 已提交
539 540 541
cusolverDnHandle_t CUDADeviceContext::cusolver_dn_handle() const {
  return context()->CusolverDnHandle();
}
542
#endif
G
Guo Sheng 已提交
543

544
gpuStream_t CUDADeviceContext::stream() const { return context()->RawStream(); }
Q
qijun 已提交
545

C
chengduoZH 已提交
546 547 548 549 550 551 552 553 554 555 556 557 558 559
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
560
#endif
Q
qijun 已提交
561

T
tensor-tang 已提交
562 563
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
564
    : CPUDeviceContext(place), p_blobmap_() {
565
  p_blobmap_.reset(new BlobMap());
566
  p_exec_items_.reset(new ExecShape());
567
  p_mutex_.reset(new std::mutex());
T
tensor-tang 已提交
568 569
}

570 571
MKLDNNDeviceContextThreadLocals::Body::Body()
    : cur_engine(mkldnn::engine::kind::cpu, 0), cur_stream(cur_engine) {
572 573 574 575 576 577
  cur_mkldnn_session_id = kMKLDNNSessionID_Default;
  cur_input_shape_str = "";
  cur_input_shape_cache_capacity = 1;
  cur_paddle_data_layout = paddle::framework::DataLayout::kNCHW;
}

578 579 580 581 582 583 584 585 586 587 588 589
// When Thread finish we clear oneDNN cache
// This is needed when we have one executor used by many threads
// e.g. test_analyzer_detect. Thread ID is not part of caching key
// (for naive executor) so we need to clear cache when one thread finish
// and other is to start inference
// TODO(jczaja): Ideally it would be good to clear only part of cache
// related to thread that is to be terminated
MKLDNNDeviceContextThreadLocals::Body::~Body() {
  auto cpu_place = paddle::platform::CPUPlace();
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
  platform::MKLDNNDeviceContext* dev_ctx =
      (platform::MKLDNNDeviceContext*)pool.Get(cpu_place);
590
  dev_ctx->ResetBlobMap(exec_ptr_);
591 592
}

593 594 595 596 597 598 599 600 601 602
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) {
603 604
  cur_input_shape_str = input_shape_str;
}
605 606
void MKLDNNDeviceContextThreadLocals::Body::set_cur_input_shape_cache_capacity(
    int input_shape_cache_capacity) {
607 608
  cur_input_shape_cache_capacity = input_shape_cache_capacity;
}
S
Sylwester Fraczek 已提交
609

610 611
void MKLDNNDeviceContextThreadLocals::Body::set_cur_paddle_data_layout(
    framework::DataLayout dl) {
612 613 614
  cur_paddle_data_layout = dl;
}

615 616
framework::DataLayout
MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
617 618 619
  return cur_paddle_data_layout;
}

620 621 622 623 624 625 626 627 628
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;
  }
}

629 630 631 632 633 634 635 636
const mkldnn::engine& MKLDNNDeviceContextThreadLocals::Body::get_engine(void) {
  return cur_engine;
}

mkldnn::stream& MKLDNNDeviceContextThreadLocals::Body::get_stream(void) {
  return cur_stream;
}

637
void MKLDNNDeviceContext::ResetBlobMap(void* ptr) {
638 639 640
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
  if (!block_next_cache_clearing_) {
    VLOG(3) << "Clearing DNNL cache.";
641 642 643 644 645 646
    // If no specific executor pointer then clear
    // everything. For executor pointer then clear only
    // objects allocated when using given executor
    if (ptr == nullptr) {
      p_blobmap_->clear();
    } else {
647 648 649 650 651 652 653 654
      // Iterate through all shapes and release
      // for each shape and active executor all entries
      // of this executor
      for (auto& s : *p_exec_items_) {
        for (auto& v : (*s.second)[ptr]) {
          (v.first)->erase(v.second);
        }
        s.second->erase(ptr);
655 656
      }
    }
657 658 659 660 661 662
  } else {
    VLOG(3) << "Prevented Clearing DNNL cache.";
    block_next_cache_clearing_ = false;
  }
}

663 664 665 666
void MKLDNNDeviceContext::RemoveShapeEntriesWithExecutor(void) const {
  p_exec_items_->erase(p_exec_items_->begin());
}

667 668
void MKLDNNDeviceContext::LinkEntryWithExecutor(BlobPtr_t<KeyBlob> pblob,
                                                KeyBlob::iterator it) const {
669
  // Take current input shape from TLS
670 671
  // Take current executor addess from TLS
  // and for this executor's items add the one defined with arguments
672 673 674 675 676 677 678 679 680
  auto key_it = p_exec_items_
                    ->insert(std::make_pair(tls().cur_input_shape_str,
                                            std::make_shared<ExecMap>()))
                    .first;
  (*key_it->second)[tls().get_curr_exec()].push_back(std::make_pair(pblob, it));

  VLOG(3) << "LinkEntryWithExecutor, shapes: " << p_exec_items_->size()
          << " curr exec size: "
          << (*key_it->second)[tls().get_curr_exec()].size() << "\n";
681 682
}

683 684 685 686
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;
687
}
688

689
size_t MKLDNNDeviceContext::GetShapeBlobSize() const {
690
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
691
  BlobMap* pMap = p_blobmap_.get();
692
  auto map_it = pMap->find(tls().cur_mkldnn_session_id);
693
  if (map_it == pMap->end()) {
694 695 696
    PADDLE_THROW(platform::errors::NotFound(
        "MKLDNNDeviceContext don't find cur_mkldnn_session_id: %d.",
        tls().cur_mkldnn_session_id));
697 698 699 700
  }
  return map_it->second->size();
}

701
void MKLDNNDeviceContext::SetBlob(const std::string& name,
702
                                  BlobPtr_t<void> data) const {
703
  BlobMap* pMap = p_blobmap_.get();
704 705
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
706

707
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
708

709
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
T
tensor-tang 已提交
710

711 712
  // Find ShapeBlob for current mkldnn session id.
  auto map_it = pMap->find(sid);
713 714 715

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
716
    sBlob = std::make_shared<ShapeBlob>();
717 718
    (*pMap)[sid] = sBlob;
    VLOG(2) << "SetBlob: sid=" << sid << ", add new sid\n";
719
  } else {
720
    sBlob = map_it->second;
721
  }
T
tensor-tang 已提交
722

723
  // Find KeyBlob for current input shape
724
  auto key_it = sBlob->find(tls().cur_input_shape_str);
725

726
  if (key_it == sBlob->end()) {
727 728
    // In cache clearing mode, cur_input_shape_cache_capacity defines
    // max pblob capacity
729 730
    if ((static_cast<size_t>(sid) ==
         MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_CacheClearing) &&
731
        sBlob->size() &&
732
        (sBlob->size() >=
733
         static_cast<size_t>(tls().cur_input_shape_cache_capacity))) {
734 735 736
      VLOG(2) << "sid=" << sid
              << ", remove all blobs of shape: " << sBlob->begin()->first;
      sBlob->erase(sBlob->begin()->first);
737
      RemoveShapeEntriesWithExecutor();
738
    }
739 740
    pBlob = std::make_shared<KeyBlob>();
    (*sBlob)[tls().cur_input_shape_str] = pBlob;
741
  } else {
742
    pBlob = key_it->second;
743 744
  }

745 746 747
  // Find Blob via name
  auto blob_it = pBlob->find(name);
  if (blob_it == pBlob->end()) {
748 749 750 751 752
    auto el =
        pBlob->insert(std::make_pair(name, data));  //  (*pBlob)[name] = data;
    // Register new element in per executor map
    // to have easily erased when executor terminated
    LinkEntryWithExecutor(pBlob, el.first);
753 754 755
  } else {
    blob_it->second = data;  // set data to existing blob
  }
756
  VLOG(2) << "SetBlob: sid=" << sid << ", add blob=" << name << "\n";
757
  // lock will be automatically released when out of scope
758
  return;
T
tensor-tang 已提交
759 760
}

761
unsigned int MKLDNNDeviceContext::GetCachedObjectsNumber(void) const {
762 763 764 765 766 767 768 769 770
  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;
}

771
MKLDNNDeviceContext::BlobPtr_t<void> MKLDNNDeviceContext::GetBlob(
772
    const std::string& name) const {
773
  BlobMap* pMap = p_blobmap_.get();
774 775
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
  BlobPtr_t<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
776

777
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
778

779
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
780

781 782
  // Find ShapeBlob for current mkldnn session id firstly
  auto map_it = pMap->find(sid);
783
  if (map_it == pMap->end()) {
784
    VLOG(2) << "GetBlob: sid=" << sid << ", miss sid\n";
785 786 787 788 789
    return nullptr;
  }
  sBlob = map_it->second;

  // Find KeyBlob for current input shape secondly
790
  auto sBlob_it = sBlob->find(tls().cur_input_shape_str);
791
  if (sBlob_it == sBlob->end()) {
792
    VLOG(2) << "GetBlob: sid=" << tls().cur_input_shape_str
793 794 795 796
            << ", miss input_shape_str\n";
    return nullptr;
  }
  pBlob = sBlob_it->second;
797 798 799 800

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

801
  if (key_it == pBlob->end()) {
802
    VLOG(2) << "GetBlob sid=" << sid << ", miss blob=" << name << "\n";
803 804
    return nullptr;
  }
805

806
  VLOG(2) << "GetBlob sid=" << sid << ", get blob=" << name << "\n";
807 808
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
809 810 811
}

#endif
Q
qijun 已提交
812
}  // namespace platform
Q
qijun 已提交
813
}  // namespace paddle