device_context.cc 27.6 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

78 79 80 81 82 83 84 85 86 87 88 89 90
DeviceType Place2DeviceType(const platform::Place& place) {
  if (platform::is_cpu_place(place)) {
    return platform::DeviceType::CPU;
  } else if (platform::is_gpu_place(place)) {
    return platform::DeviceType::CUDA;
  } else if (platform::is_xpu_place(place)) {
    return platform::DeviceType::XPU;
  } else {
    PADDLE_THROW(platform::errors::Unavailable(
        "Unsupported place %s to convert into platform::DeviceType.", place));
  }
}

D
dzhwinter 已提交
91 92
DeviceContextPool* DeviceContextPool::pool = nullptr;

Y
Yu Yang 已提交
93
platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
94
  VLOG(4) << "DeviceContextPool Get: " << place;
D
dzhwinter 已提交
95 96
  auto it = device_contexts_.find(place);
  if (it == device_contexts_.end()) {
G
GaoWei8 已提交
97 98
    PADDLE_THROW(platform::errors::Unimplemented(
        "Place %s is not supported. Please check that your paddle compiles "
99 100
        "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
101
        place));
D
dzhwinter 已提交
102
  }
103
  return it->second.get().get();
D
dzhwinter 已提交
104 105
}

106 107 108 109 110 111 112 113 114
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`
115
                     return PtrType(new DevCtx(BOOST_GET_CONST(PlaceType, p)));
116
                   }));
C
chengduozh 已提交
117 118
}

D
dzhwinter 已提交
119 120
DeviceContextPool::DeviceContextPool(
    const std::vector<platform::Place>& places) {
G
GaoWei8 已提交
121 122 123 124 125
  PADDLE_ENFORCE_GT(
      places.size(), 0,
      platform::errors::InvalidArgument("The number of platform places should "
                                        "be larger than 0. But received %d.",
                                        places.size()));
126
  std::set<Place> set;
Y
Yu Yang 已提交
127 128 129 130 131
  for (auto& p : places) {
    set.insert(p);
  }
  for (auto& p : set) {
    if (platform::is_cpu_place(p)) {
132
#ifdef PADDLE_WITH_MKLDNN
133
      EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
134
#else
135
      EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
136
#endif
Y
Yu Yang 已提交
137
    } else if (platform::is_gpu_place(p)) {
138
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
139
      EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
D
dzhwinter 已提交
140
#else
G
GaoWei8 已提交
141 142 143
      PADDLE_THROW(
          platform::errors::Unimplemented("CUDAPlace is not supported. Please "
                                          "re-compile with WITH_GPU option."));
C
chengduoZH 已提交
144 145
#endif
    } else if (platform::is_cuda_pinned_place(p)) {
146
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
147 148
      EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
          &device_contexts_, p);
C
chengduoZH 已提交
149
#else
G
GaoWei8 已提交
150
      PADDLE_THROW(platform::errors::Unimplemented(
G
GaoWei8 已提交
151 152
          "CUDAPlace is not supported. Please re-compile with WITH_GPU "
          "option."));
153 154 155 156 157 158 159 160
#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."));
161 162 163 164 165 166 167 168
#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."));
169 170 171 172 173 174 175 176 177 178
#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
179 180 181 182 183
#endif
    }
  }
}

184 185 186 187
CPUDeviceContext::CPUDeviceContext() {
  eigen_device_.reset(new Eigen::DefaultDevice());
}

D
dzhwinter 已提交
188
CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) {
189 190 191 192 193 194 195
  eigen_device_.reset(new Eigen::DefaultDevice());
}

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

D
dzhwinter 已提交
196
Place CPUDeviceContext::GetPlace() const { return place_; }
197

198
#ifdef PADDLE_WITH_XPU
Q
QingshuChen 已提交
199 200 201 202
XPUDeviceContext::XPUDeviceContext() {
  context_ = xpu::create_context();
  xpu_version_ = get_xpu_version(place_.device);
}
203

204
XPUDeviceContext::~XPUDeviceContext() {}
205 206 207 208 209 210 211 212 213 214 215 216 217 218 219

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));
220 221 222

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

223
  context_ = xpu::create_context();
224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240
  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;
    }
241
  }
242

243 244 245 246 247 248 249 250 251 252 253 254 255 256 257
  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));
258
  xpu_wait(context_->xpu_stream);
259 260 261 262 263 264 265
}

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

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

266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
#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_));
}
281

282
void NPUDeviceContext::Wait() const {
283 284 285
  platform::RecordEvent record_event("NPUDeviceContext/wait");
  VLOG(4) << "NPU context(" << this << ")  Wait";
  stream_->Wait();
286 287 288 289 290 291 292
}

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

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

aclrtContext NPUDeviceContext::context() const { return context_; }
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308

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_; }

309 310 311
#endif

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
Q
init  
qijun 已提交
312 313 314 315 316 317 318
class EigenCudaStreamDevice : public Eigen::StreamInterface {
 public:
  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
    Eigen::initializeDeviceProp();
  }
  ~EigenCudaStreamDevice() override {}

319
  void Reinitialize(const gpuStream_t* cuda_stream, CUDAPlace place) {
Q
init  
qijun 已提交
320 321 322 323 324
    stream_ = cuda_stream;
    place_ = place;
    device_prop_ = &Eigen::m_deviceProperties[place.device];
  }

325
  const gpuStream_t& stream() const override { return *stream_; }
Q
init  
qijun 已提交
326

327 328 329
#ifdef PADDLE_WITH_HIP
  const hipDeviceProp_t& deviceProperties() const override {
#else
Q
init  
qijun 已提交
330
  const cudaDeviceProp& deviceProperties() const override {
331
#endif
Q
init  
qijun 已提交
332 333 334 335
    return *device_prop_;
  }

  void* allocate(size_t num_bytes) const override {
S
sneaxiy 已提交
336 337 338
    if (UNLIKELY(num_bytes == 0)) {
      return nullptr;
    }
339 340 341
    auto buf = memory::Alloc(place_, num_bytes);
    VLOG(4) << "Eigen allocated at " << buf->ptr() << ", size" << buf->size()
            << " requested " << num_bytes;
342
    void* retv = buf->ptr();
S
sneaxiy 已提交
343 344 345 346
    {
      std::lock_guard<std::mutex> lock(mtx_);
      allocations_.emplace(retv, std::move(buf));
    }
347
    return retv;
Q
init  
qijun 已提交
348 349
  }

S
sneaxiy 已提交
350 351 352 353 354 355
  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
356 357 358

  void* scratchpad() const override {
    if (scratch_ == NULL) {
Z
Zhang Ting 已提交
359
      scratch_ = allocate(Eigen::kGpuScratchSize + sizeof(unsigned int));
Q
init  
qijun 已提交
360 361 362 363 364 365
    }
    return scratch_;
  }

  unsigned int* semaphore() const override {
    if (semaphore_ == NULL) {
Z
Zhang Ting 已提交
366
      char* scratch = static_cast<char*>(scratchpad()) + Eigen::kGpuScratchSize;
Q
init  
qijun 已提交
367
      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
368 369 370 371
#ifdef PADDLE_WITH_HIP
      PADDLE_ENFORCE_CUDA_SUCCESS(
          hipMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
#else
372
      PADDLE_ENFORCE_CUDA_SUCCESS(
Q
init  
qijun 已提交
373
          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
374
#endif
Q
init  
qijun 已提交
375 376 377 378 379
    }
    return semaphore_;
  }

 private:
D
dzhwinter 已提交
380
  CUDAPlace place_;
381 382 383 384
  const gpuStream_t* stream_;  // not owned;
#ifdef PADDLE_WITH_HIP
  const hipDeviceProp_t* device_prop_;
#else
Q
init  
qijun 已提交
385
  const cudaDeviceProp* device_prop_;  // not owned;
386
#endif
Q
qijun 已提交
387
  mutable void* scratch_;
Q
init  
qijun 已提交
388
  mutable unsigned int* semaphore_;
S
sneaxiy 已提交
389
  mutable std::mutex mtx_;  // to protect allocations_
Y
Yu Yang 已提交
390
  mutable std::unordered_map<void*, memory::AllocationPtr> allocations_;
Q
init  
qijun 已提交
391 392
};

393 394 395 396 397 398 399 400 401
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);
}

402 403 404 405 406 407 408 409 410 411 412 413
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,
414 415
                         const stream::Priority& priority,
                         const stream::StreamFlag& flag) {
416 417
  place_ = place;
  CUDADeviceGuard guard(place_.device);
418
  stream_.reset(new stream::CUDAStream(place, priority, flag));
419 420 421
  InitEigenContext();
  InitCuBlasContext();
  InitCuDNNContext();
422
#ifndef PADDLE_WITH_HIP
G
Guo Sheng 已提交
423
  InitCuSolverContext();
424
#endif
425 426 427 428 429 430
}

CUDAContext::~CUDAContext() {
  CUDADeviceGuard guard(place_.device);
  DestoryCuDNNContext();
  DestoryCuBlasContext();
431
#ifndef PADDLE_WITH_HIP
G
Guo Sheng 已提交
432
  DestoryCuSolverContext();
433
#endif
434 435
}

436
CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
Y
Yu Yang 已提交
437
  CUDADeviceGuard guard(place_.device);
C
chengduo 已提交
438 439 440
  compute_capability_ = GetCUDAComputeCapability(place_.device);
  multi_process_ = GetCUDAMultiProcessors(place_.device);
  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
441
  max_grid_dim_size_ = GetGpuMaxGridDimSize(place_.device);
442
  max_threads_per_block_ = GetCUDAMaxThreadsPerBlock(place_.device);
443

C
chengduo 已提交
444 445 446
  driver_version_ = GetCUDADriverVersion(place_.device);
  runtime_version_ = GetCUDARuntimeVersion(place_.device);

447
  LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device
448 449 450
                          << ", GPU Compute Capability: "
                          << compute_capability_ / 10 << "."
                          << compute_capability_ % 10
C
chengduo 已提交
451
                          << ", Driver API Version: " << driver_version_ / 1000
452
                          << "." << (driver_version_ % 100) / 10
C
chengduo 已提交
453 454 455
                          << ", Runtime API Version: "
                          << runtime_version_ / 1000 << "."
                          << (runtime_version_ % 100) / 10;
456 457 458 459 460 461 462 463
#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
464 465 466
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                          << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
467
                          << (cudnn_dso_ver % 1000) / 100 << ".";
468
#endif
S
sneaxiy 已提交
469 470
  {
    // Check CUDA/CUDNN version compatiblity
471 472
    auto local_cuda_version =
        (driver_version_ / 1000) * 10 + (driver_version_ % 100) / 10;
473 474 475
#ifdef PADDLE_WITH_HIP
    auto compile_cuda_version = (HIP_VERSION / 100) * 10 + (HIP_VERSION % 10);
#else
476 477
    auto compile_cuda_version =
        (CUDA_VERSION / 1000) * 10 + (CUDA_VERSION % 100) / 10;
478
#endif
S
sneaxiy 已提交
479 480 481 482 483 484 485 486 487 488 489 490
    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.";
    }
  }
491
  default_ctx_.reset(new CUDAContext(place_));
492 493 494 495
}

CUDADeviceContext::~CUDADeviceContext() {
  SetDeviceId(place_.device);
496
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
497 498 499 500
  if (nccl_comm_) {
    PADDLE_ENFORCE_CUDA_SUCCESS(dynload::ncclCommDestroy(nccl_comm_));
  }
#endif
501 502
}

L
liaogang 已提交
503
Place CUDADeviceContext::GetPlace() const { return place_; }
504

505
void CUDADeviceContext::Wait() const { context()->Stream()->Wait(); }
506

K
Kexin Zhao 已提交
507
int CUDADeviceContext::GetComputeCapability() const {
C
chengduo 已提交
508
  return compute_capability_;
K
Kexin Zhao 已提交
509 510
}

511
int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
C
chengduo 已提交
512
  return multi_process_ * max_threads_per_mp_;
513 514
}

515 516 517 518 519 520
int CUDADeviceContext::GetSMCount() const { return multi_process_; }

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

521
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
522
  return context()->EigenDevice().get();
523 524
}

525
bool CUDADeviceContext::tensor_core_available() const {
526
  return context()->CublasTensorCoreHandle() != nullptr;
S
sneaxiy 已提交
527 528
}

529 530 531 532
dim3 CUDADeviceContext::GetCUDAMaxGridDimSize() const {
  return max_grid_dim_size_;
}

533 534 535
#ifdef PADDLE_WITH_HIP
miopenHandle_t CUDADeviceContext::cudnn_handle() const {
#else
536
cudnnHandle_t CUDADeviceContext::cudnn_handle() const {
537
#endif
538 539
  return context()->CudnnHandle();
}
540

541 542 543 544 545
#ifdef PADDLE_WITH_HIP
rocblas_handle CUDADeviceContext::cublas_handle() const {
  return context()->CublasHandle()->GetCublasHandle();
}
#else
546 547 548
cublasHandle_t CUDADeviceContext::cublas_handle() const {
  return context()->CublasHandle()->GetCublasHandle();
}
549
#endif
550

S
sneaxiy 已提交
551
CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
552
  return CudnnWorkspaceHandle(*this, &cudnn_handle_mtx_);
553
}
554

555
#ifndef PADDLE_WITH_HIP
G
Guo Sheng 已提交
556 557 558
cusolverDnHandle_t CUDADeviceContext::cusolver_dn_handle() const {
  return context()->CusolverDnHandle();
}
559
#endif
G
Guo Sheng 已提交
560

561
gpuStream_t CUDADeviceContext::stream() const { return context()->RawStream(); }
Q
qijun 已提交
562

C
chengduoZH 已提交
563 564 565 566 567 568 569 570 571 572 573 574 575 576
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
577
#endif
Q
qijun 已提交
578

T
tensor-tang 已提交
579 580
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
581
    : CPUDeviceContext(place), p_blobmap_() {
582
  p_blobmap_.reset(new BlobMap());
583
  p_exec_items_.reset(new ExecShape());
584
  p_mutex_.reset(new std::mutex());
T
tensor-tang 已提交
585 586
}

587 588
MKLDNNDeviceContextThreadLocals::Body::Body()
    : cur_engine(mkldnn::engine::kind::cpu, 0), cur_stream(cur_engine) {
589 590 591 592 593 594
  cur_mkldnn_session_id = kMKLDNNSessionID_Default;
  cur_input_shape_str = "";
  cur_input_shape_cache_capacity = 1;
  cur_paddle_data_layout = paddle::framework::DataLayout::kNCHW;
}

595 596 597 598 599 600 601 602 603 604 605 606
// 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);
607
  dev_ctx->ResetBlobMap(exec_ptr_);
608 609
}

610 611 612 613 614 615 616 617 618 619
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) {
620 621
  cur_input_shape_str = input_shape_str;
}
622 623
void MKLDNNDeviceContextThreadLocals::Body::set_cur_input_shape_cache_capacity(
    int input_shape_cache_capacity) {
624 625
  cur_input_shape_cache_capacity = input_shape_cache_capacity;
}
S
Sylwester Fraczek 已提交
626

627 628
void MKLDNNDeviceContextThreadLocals::Body::set_cur_paddle_data_layout(
    framework::DataLayout dl) {
629 630 631
  cur_paddle_data_layout = dl;
}

632 633
framework::DataLayout
MKLDNNDeviceContextThreadLocals::Body::get_cur_paddle_data_layout(void) {
634 635 636
  return cur_paddle_data_layout;
}

637 638 639 640 641 642 643 644 645
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;
  }
}

646 647 648 649 650 651 652 653
const mkldnn::engine& MKLDNNDeviceContextThreadLocals::Body::get_engine(void) {
  return cur_engine;
}

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

654
void MKLDNNDeviceContext::ResetBlobMap(void* ptr) {
655 656 657
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
  if (!block_next_cache_clearing_) {
    VLOG(3) << "Clearing DNNL cache.";
658 659 660 661 662 663
    // 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 {
664 665 666 667 668
      // 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]) {
669
          (v.first)->erase(v.second);
670 671
        }
        s.second->erase(ptr);
672 673
      }
    }
674 675 676 677 678 679
  } else {
    VLOG(3) << "Prevented Clearing DNNL cache.";
    block_next_cache_clearing_ = false;
  }
}

680 681
void MKLDNNDeviceContext::RemoveShapeEntriesWithExecutor(void) const {
  p_exec_items_->erase(p_exec_items_->begin());
682 683
}

684 685
void MKLDNNDeviceContext::LinkEntryWithExecutor(BlobPtr_t<KeyBlob> pblob,
                                                KeyBlob::iterator it) const {
686
  // Take current input shape from TLS
687 688
  // Take current executor addess from TLS
  // and for this executor's items add the one defined with arguments
689 690 691 692 693 694 695 696 697
  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";
698 699
}

700 701 702 703
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;
704
}
705

706
size_t MKLDNNDeviceContext::GetShapeBlobSize() const {
707
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
708
  BlobMap* pMap = p_blobmap_.get();
709
  auto map_it = pMap->find(tls().cur_mkldnn_session_id);
710
  if (map_it == pMap->end()) {
711 712 713
    PADDLE_THROW(platform::errors::NotFound(
        "MKLDNNDeviceContext don't find cur_mkldnn_session_id: %d.",
        tls().cur_mkldnn_session_id));
714 715 716 717
  }
  return map_it->second->size();
}

718
void MKLDNNDeviceContext::SetBlob(const std::string& name,
719
                                  BlobPtr_t<void> data) const {
720
  BlobMap* pMap = p_blobmap_.get();
721
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
722
  BlobPtr_t<KeyBlob> pBlob = nullptr;
723

724
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
725

726
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
T
tensor-tang 已提交
727

728 729
  // Find ShapeBlob for current mkldnn session id.
  auto map_it = pMap->find(sid);
730 731 732

  if (map_it == pMap->end()) {
    // 1st time to set blob in current thread
733
    sBlob = std::make_shared<ShapeBlob>();
734 735
    (*pMap)[sid] = sBlob;
    VLOG(2) << "SetBlob: sid=" << sid << ", add new sid\n";
736
  } else {
737
    sBlob = map_it->second;
738
  }
T
tensor-tang 已提交
739

740
  // Find KeyBlob for current input shape
741
  auto key_it = sBlob->find(tls().cur_input_shape_str);
742

743
  if (key_it == sBlob->end()) {
744 745
    // In cache clearing mode, cur_input_shape_cache_capacity defines
    // max pblob capacity
746 747
    if ((static_cast<size_t>(sid) ==
         MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_CacheClearing) &&
748
        sBlob->size() &&
749
        (sBlob->size() >=
750
         static_cast<size_t>(tls().cur_input_shape_cache_capacity))) {
751 752 753 754
      VLOG(2) << "sid=" << sid
              << ", remove all blobs of shape: " << sBlob->begin()->first;
      sBlob->erase(sBlob->begin()->first);
      RemoveShapeEntriesWithExecutor();
755
    }
756
    pBlob = std::make_shared<KeyBlob>();
757
    (*sBlob)[tls().cur_input_shape_str] = pBlob;
758
  } else {
759
    pBlob = key_it->second;
760 761
  }

762
  // Find Blob via name
763 764 765 766
  auto blob_it = pBlob->find(name);
  if (blob_it == pBlob->end()) {
    auto el =
        pBlob->insert(std::make_pair(name, data));  //  (*pBlob)[name] = data;
767 768 769
    // Register new element in per executor map
    // to have easily erased when executor terminated
    LinkEntryWithExecutor(pBlob, el.first);
770 771 772
  } else {
    blob_it->second = data;  // set data to existing blob
  }
773
  VLOG(2) << "SetBlob: sid=" << sid << ", add blob=" << name << "\n";
774
  // lock will be automatically released when out of scope
775
  return;
T
tensor-tang 已提交
776 777
}

778
unsigned int MKLDNNDeviceContext::GetCachedObjectsNumber(void) const {
779 780 781
  unsigned int num_entries = 0;
  for (auto const& l3 : *p_blobmap_) {
    for (auto const& l2 : *(l3.second)) {
782
      num_entries += (l2.second)->size();
783 784 785 786 787
    }
  }
  return num_entries;
}

788
MKLDNNDeviceContext::BlobPtr_t<void> MKLDNNDeviceContext::GetBlob(
789
    const std::string& name) const {
790
  BlobMap* pMap = p_blobmap_.get();
791
  BlobPtr_t<ShapeBlob> sBlob = nullptr;
792
  BlobPtr_t<KeyBlob> pBlob = nullptr;
T
tensor-tang 已提交
793

794
  int sid = tls().get_cur_mkldnn_session_id();
T
tensor-tang 已提交
795

796
  std::lock_guard<decltype(*p_mutex_)> lock(*p_mutex_);
797

798 799
  // Find ShapeBlob for current mkldnn session id firstly
  auto map_it = pMap->find(sid);
800
  if (map_it == pMap->end()) {
801
    VLOG(2) << "GetBlob: sid=" << sid << ", miss sid\n";
802 803 804 805 806
    return nullptr;
  }
  sBlob = map_it->second;

  // Find KeyBlob for current input shape secondly
807
  auto sBlob_it = sBlob->find(tls().cur_input_shape_str);
808
  if (sBlob_it == sBlob->end()) {
809
    VLOG(2) << "GetBlob: sid=" << tls().cur_input_shape_str
810 811 812 813
            << ", miss input_shape_str\n";
    return nullptr;
  }
  pBlob = sBlob_it->second;
814 815

  // Find Blob via name
816
  auto key_it = pBlob->find(name);
817

818
  if (key_it == pBlob->end()) {
819
    VLOG(2) << "GetBlob sid=" << sid << ", miss blob=" << name << "\n";
820 821
    return nullptr;
  }
822

823
  VLOG(2) << "GetBlob sid=" << sid << ", get blob=" << name << "\n";
824 825
  // lock will be automatically released when out of scope
  return key_it->second;
T
tensor-tang 已提交
826 827 828
}

#endif
Q
qijun 已提交
829
}  // namespace platform
Q
qijun 已提交
830
}  // namespace paddle