device_context.h

/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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
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. */
#pragma once

#include <future>  // NOLINT
#include <memory>
#include <mutex>  // NOLINT
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>

#include "paddle/pten/backends/cpu/cpu_context.h"
#include "paddle/pten/core/device_context.h"

#include "paddle/fluid/memory/malloc.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/device/gpu/gpu_helper.h"
#include "paddle/fluid/platform/dynload/cublas.h"
#include "paddle/fluid/platform/dynload/cudnn.h"
#include "paddle/fluid/platform/dynload/cusolver.h"
#include "paddle/fluid/platform/dynload/cusparse.h"
#if !defined(__APPLE__) && defined(PADDLE_WITH_NCCL)
#include "paddle/fluid/platform/dynload/nccl.h"
#endif
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
#endif

#ifdef PADDLE_WITH_HIP
#include "paddle/fluid/platform/device/gpu/gpu_helper.h"  // NOLINT
#include "paddle/fluid/platform/dynload/miopen.h"
#include "paddle/fluid/platform/dynload/rocblas.h"
#if !defined(__APPLE__) && defined(PADDLE_WITH_RCCL)
#include "paddle/fluid/platform/dynload/rccl.h"
#endif
#include "paddle/fluid/platform/device/gpu/gpu_info.h"  // NOLINT
#endif

#if defined(PADDLE_WITH_XPU_BKCL)
#include "xpu/bkcl.h"
#endif

#ifdef PADDLE_WITH_MKLDNN
#include "dnnl.hpp"
#include "paddle/fluid/framework/data_layout.h"
#endif

#include <map>

#include "glog/logging.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/place.h"
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#include "paddle/fluid/platform/stream/cuda_stream.h"
#endif
#ifdef PADDLE_WITH_ASCEND_CL
#include "paddle/fluid/platform/device/npu/enforce_npu.h"
#include "paddle/fluid/platform/device/npu/npu_stream.h"
#endif
#ifdef PADDLE_WITH_IPU
#include "paddle/fluid/platform/device/ipu/device.h"
#endif
#include "unsupported/Eigen/CXX11/Tensor"

namespace Eigen {
struct DefaultDevice;
struct GpuDevice;
}  // namespace Eigen

#ifdef PADDLE_WITH_XPU
#include "paddle/fluid/platform/device/xpu/xpu_header.h"
#include "paddle/fluid/platform/device/xpu/xpu_info.h"
#endif

#ifdef PADDLE_WITH_ASCEND_CL
#include "acl/acl.h"
#include "paddle/fluid/platform/device/npu/npu_info.h"
#endif

namespace paddle {
namespace platform {

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
/*Set the value of the global variable allow_tf32_cublas*/
void SetAllowTF32Cublas(bool active);
/*Get the global variable allow_tf32_cublas value*/
bool AllowTF32Cublas();
extern bool allow_tf32_cudnn;
/*Set the value of the global variable allow_tf32_cudnn*/
void SetAllowTF32Cudnn(bool active);
/*Get the global variable allow_tf32_cudnn value*/
bool AllowTF32Cudnn();
#endif  // PADDLE_WITH_CUDA

enum DeviceType {
  CPU = 0,
  CUDA = 1,
  XPU = 2,
  NPU = 3,
  IPU = 4,
  MLU = 5,

  MAX_DEVICE_TYPES = 6,
};

DeviceType Place2DeviceType(const platform::Place& place);

constexpr DeviceType kCPU = DeviceType::CPU;
constexpr DeviceType kCUDA = DeviceType::CUDA;
constexpr DeviceType kXPU = DeviceType::XPU;
constexpr DeviceType kNPU = DeviceType::NPU;
constexpr DeviceType kIPU = DeviceType::IPU;
constexpr DeviceType kMLU = DeviceType::MLU;

using DeviceContext = pten::DeviceContext;

// using CPUDeviceContext = pten::CPUContext;
// TODO(wilber): The place constructor is used in many places, it is more
// difficult to use CPUDeviceContext = pten::CPUContext directly.
class CPUDeviceContext : public pten::CPUContext {
 public:
  CPUDeviceContext();
  explicit CPUDeviceContext(CPUPlace place);
};

template <typename Place>
struct DefaultDeviceContextType;

template <>
struct DefaultDeviceContextType<platform::CPUPlace> {
  using TYPE = CPUDeviceContext;
};

// Graphcore IPU
#ifdef PADDLE_WITH_IPU
class IPUDeviceContext : public DeviceContext {
 public:
  IPUDeviceContext() = delete;
  explicit IPUDeviceContext(IPUPlace place);
  virtual ~IPUDeviceContext();
  Eigen::DefaultDevice* eigen_device() const { return nullptr; }
  Place GetPlace() const override;
  /*! \brief  Wait for all operations completion in the stream. */
  void Wait() const override;
  int DeviceId() const { return device_.getId(); }

 private:
  IPUPlace place_;
  platform::ipu::Device device_;
};
template <>
struct DefaultDeviceContextType<platform::IPUPlace> {
  using TYPE = IPUDeviceContext;
};
#endif

#ifdef PADDLE_WITH_MLU
class MLUDeviceContext;

template <>
struct DefaultDeviceContextType<platform::MLUPlace>;
#endif

#ifdef PADDLE_WITH_XPU
namespace xpu = baidu::xpu::api;
class XPUDeviceContext : public DeviceContext {
 public:
  XPUDeviceContext();
  explicit XPUDeviceContext(XPUPlace place);
  virtual ~XPUDeviceContext();
  Eigen::DefaultDevice* eigen_device() const { return nullptr; }
  XPUVersion xpu_version() const { return xpu_version_; }
  Place GetPlace() const override;
  xpu::Context* x_context() const;

  /*! \brief  Wait for all operations completion in the stream. */
  void Wait() const override;

#ifdef PADDLE_WITH_XPU_BKCL
  /*! \brief  Return bkcl context. */
  BKCLContext_t bkcl_context() const { return bkcl_context_; }

  /*! \brief  Set bkcl context. */
  void set_bkcl_context(BKCLContext_t context) { bkcl_context_ = context; }
#endif

 private:
  XPUPlace place_;
  XPUVersion xpu_version_;
  xpu::Context* context_;
#ifdef PADDLE_WITH_XPU_BKCL
  BKCLContext_t bkcl_context_;
#endif

  // Need to be the same with other DeviceContext,
  // Eventhough eigen_device_ is not used in XPU
  std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
  DISABLE_COPY_AND_ASSIGN(XPUDeviceContext);
};

template <>
struct DefaultDeviceContextType<platform::XPUPlace> {
  using TYPE = XPUDeviceContext;
};
#endif

#ifdef PADDLE_WITH_ASCEND_CL
class NPUDeviceContext : public DeviceContext {
 public:
  explicit NPUDeviceContext(NPUPlace place);
  virtual ~NPUDeviceContext();
  Eigen::DefaultDevice* eigen_device() const { return nullptr; }
  Place GetPlace() const override;
  aclrtContext context() const;

  /*! \brief  Wait for all operations completion in the stream. */
  void Wait() const override;

  /*! \brief  Return npu stream in the device context. */
  aclrtStream stream() const;

  template <typename Callback>
  void AddStreamCallback(Callback&& callback) const {
    return stream_->AddCallback(callback);
  }

  void WaitStreamCallback() const { return stream_->WaitCallback(); }

#if defined(PADDLE_WITH_ASCEND_CL)
  /*! \brief  Return hccl communicators. */
  HcclComm hccl_comm() const { return hccl_comm_; }

  /*! \brief  Set hccl communicators. */
  void set_hccl_comm(HcclComm comm) { hccl_comm_ = comm; }
#endif

  // template <typename Callback>
  // void AddStreamCallback(Callback&& callback) const {
  //   return stream_->AddCallback(callback);
  // }

  // void WaitStreamCallback() const { return stream_->WaitCallback(); }

 private:
  NPUPlace place_;
  aclrtContext context_;

#ifdef PADDLE_WITH_ASCEND_CL
  // HCCLContext_t hccl_context_;
  HcclComm hccl_comm_{nullptr};
#endif

  // Need to be the same with other DeviceContext,
  // Eventhough eigen_device_ is not used in NPU
  // NOTE(zhiqiu): why need?
  std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
  std::shared_ptr<stream::NPUStream> stream_;

  DISABLE_COPY_AND_ASSIGN(NPUDeviceContext);
};

template <>
struct DefaultDeviceContextType<platform::NPUPlace> {
  using TYPE = NPUDeviceContext;
};

// Currently, NPUPinnedDeviceContext is only used to data copying.
class NPUPinnedDeviceContext : public DeviceContext {
 public:
  NPUPinnedDeviceContext();
  explicit NPUPinnedDeviceContext(NPUPinnedPlace place);

  Place GetPlace() const override;

  Eigen::DefaultDevice* eigen_device() const;

 private:
  NPUPinnedPlace place_;
  std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
};

template <>
struct DefaultDeviceContextType<platform::NPUPinnedPlace> {
  using TYPE = NPUPinnedDeviceContext;
};

#endif

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
class CudnnWorkspaceHandle;
class EigenCudaStreamDevice;

class CUDAContext {
 public:
  CUDAContext() = default;
  explicit CUDAContext(
      const CUDAPlace& place,
      const stream::Priority& priority = stream::Priority::kNormal,
      const stream::StreamFlag& flag = stream::StreamFlag::kDefaultFlag);

  ~CUDAContext();

  const CUDAPlace& Place() const { return place_; }

  const std::unique_ptr<Eigen::GpuDevice>& EigenDevice() const {
    return eigen_device_;
  }

  const std::unique_ptr<EigenCudaStreamDevice>& EigenStream() const {
    return eigen_stream_;
  }

  const std::unique_ptr<stream::CUDAStream>& Stream() const { return stream_; }

  stream::CUDAStream* SetStream(stream::CUDAStream* new_stream_ptr) {
    auto* old_stream_ptr = stream_.release();
    stream_.reset(new_stream_ptr);
    return old_stream_ptr;
  }

  void SetStream(gpuStream_t stream);

  const gpuStream_t& RawStream() { return stream_->raw_stream(); }

#ifdef PADDLE_WITH_HIP
  const miopenHandle_t& CudnnHandle() const { return cudnn_handle_; }
#else
  const cudnnHandle_t& CudnnHandle() const { return cudnn_handle_; }
#endif

#ifndef PADDLE_WITH_HIP
  const cusolverDnHandle_t& CusolverDnHandle() const {
    return cusolver_dn_handle_;
  }
#endif

  const std::unique_ptr<CublasHandleHolder>& CublasHandle() const {
    return cublas_handle_;
  }

  const std::unique_ptr<CublasHandleHolder>& CublasTensorCoreHandle() const {
    return cublas_tensor_core_handle_;
  }

#ifndef PADDLE_WITH_HIP
  const std::unique_ptr<CusparseHandleHolder>& CusparseHandle() const {
    return cusparse_handle_;
  }
#endif

  /*! \brief  Call cublas function safely. */
  template <typename Callback>
  inline void CublasCall(Callback&& callback) const {
    if (cublas_tf32_tensor_core_handle_) {
      cublas_tf32_tensor_core_handle_->Call(std::forward<Callback>(callback));
    } else {
      cublas_handle_->Call(std::forward<Callback>(callback));
    }
  }

#ifndef PADDLE_WITH_HIP
  /*! \brief  Call cusparse function safely. */
  template <typename Callback>
  inline void CusparseCall(Callback&& callback) const {
    cusparse_handle_->Call(std::forward<Callback>(callback));
  }
#endif

  /*! \brief  Check whether tensor core is supported */
  bool tensor_core_available() const;

  /*! \brief  Call cublas function with Tensor Core safely. If
      Tensor Core is not available, use DEFAULT_MATH instead. */
  template <typename Callback>
  inline void TensorCoreCublasCallIfAvailable(Callback&& callback) const {
    if (cublas_tensor_core_handle_) {
      cublas_tensor_core_handle_->Call(std::forward<Callback>(callback));
    } else {
      cublas_handle_->Call(std::forward<Callback>(callback));
    }
  }

 private:
  void InitEigenContext();

#ifdef PADDLE_WITH_HIP
  void InitCuBlasContext() {
    cublas_handle_.reset(new CublasHandleHolder(RawStream()));
  }
#else
  void InitCuBlasContext() {
    cublas_handle_.reset(
        new CublasHandleHolder(RawStream(), CUBLAS_DEFAULT_MATH));
    if (TensorCoreAvailable()) {
#if CUDA_VERSION >= 9000
      cublas_tensor_core_handle_.reset(
          new CublasHandleHolder(RawStream(), CUBLAS_TENSOR_OP_MATH));
#if CUDA_VERSION >= 11000
      cublas_tf32_tensor_core_handle_.reset(
          new CublasHandleHolder(RawStream(), CUBLAS_TF32_TENSOR_OP_MATH));
#endif  // CUDA_VERSION >= 11000
#endif  // CUDA_VERSION >= 9000
    }
  }
#endif

#ifndef PADDLE_WITH_HIP
  void InitCuSparseContext() {
    cusparse_handle_.reset(new CusparseHandleHolder(RawStream()));
  }
#endif

  void InitCuDNNContext() {
    if (dynload::HasCUDNN()) {
#ifdef PADDLE_WITH_HIP
      size_t miopen_major, miopen_minor, miopen_patch;
      PADDLE_ENFORCE_GPU_SUCCESS(dynload::miopenGetVersion(
          &miopen_major, &miopen_minor, &miopen_patch));
      auto local_miopen_version =
          (miopen_major * 1000 + miopen_minor * 10 + miopen_patch) / 10;
      auto compile_miopen_version = MIOPEN_VERSION / 10;
      if (local_miopen_version < static_cast<size_t>(compile_miopen_version)) {
        LOG_FIRST_N(WARNING, 1)
            << "WARNING: device: " << place_.device
            << ". The installed Paddle is compiled with MIOPEN "
            << compile_miopen_version / 100 << "."
            << compile_miopen_version % 100
            << ", but MIOPEN version in your machine is "
            << local_miopen_version / 100 << "." << local_miopen_version % 100
            << ", which may cause serious incompatible bug. "
            << "Please recompile or reinstall Paddle with compatible MIOPEN "
               "version.";
      }
      PADDLE_ENFORCE_GPU_SUCCESS(dynload::miopenCreate(&cudnn_handle_));
      PADDLE_ENFORCE_GPU_SUCCESS(
          dynload::miopenSetStream(cudnn_handle_, RawStream()));
#else
      auto local_cudnn_version = dynload::cudnnGetVersion() / 100;
      auto compile_cudnn_version = CUDNN_VERSION / 100;
      if (local_cudnn_version < static_cast<size_t>(compile_cudnn_version)) {
        LOG_FIRST_N(WARNING, 1)
            << "WARNING: device: " << place_.device
            << ". The installed Paddle is compiled with CUDNN "
            << compile_cudnn_version / 10 << "." << compile_cudnn_version % 10
            << ", but CUDNN version in your machine is "
            << local_cudnn_version / 10 << "." << local_cudnn_version % 10
            << ", which may cause serious incompatible bug. "
            << "Please recompile or reinstall Paddle with compatible CUDNN "
               "version.";
      }
      PADDLE_RETRY_CUDA_SUCCESS(dynload::cudnnCreate(&cudnn_handle_));
      PADDLE_RETRY_CUDA_SUCCESS(
          dynload::cudnnSetStream(cudnn_handle_, RawStream()));
#endif
    } else {
      cudnn_handle_ = nullptr;
    }
  }

#ifndef PADDLE_WITH_HIP
  void InitCuSolverContext() {
    PADDLE_RETRY_CUDA_SUCCESS(dynload::cusolverDnCreate(&cusolver_dn_handle_));
    PADDLE_RETRY_CUDA_SUCCESS(
        dynload::cusolverDnSetStream(cusolver_dn_handle_, RawStream()));
  }
#endif

  void DestoryCuDNNContext() {
    if (cudnn_handle_) {
#ifdef PADDLE_WITH_HIP
      PADDLE_ENFORCE_GPU_SUCCESS(dynload::miopenDestroy(cudnn_handle_));
#else
      PADDLE_ENFORCE_GPU_SUCCESS(dynload::cudnnDestroy(cudnn_handle_));
#endif
    }
    cudnn_handle_ = nullptr;
  }

  void DestoryCuBlasContext() {
    cublas_handle_.reset();
    cublas_tensor_core_handle_.reset();
    cublas_tf32_tensor_core_handle_.reset();
  }

#ifndef PADDLE_WITH_HIP
  void DestoryCuSparseContext() { cusparse_handle_.reset(); }
#endif

#ifndef PADDLE_WITH_HIP
  void DestoryCuSolverContext() {
    if (cusolver_dn_handle_) {
      PADDLE_ENFORCE_GPU_SUCCESS(
          dynload::cusolverDnDestroy(cusolver_dn_handle_));
    }
  }
#endif

  CUDAPlace place_;
  std::unique_ptr<Eigen::GpuDevice> eigen_device_;
  std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
  std::unique_ptr<stream::CUDAStream> stream_;
#ifdef PADDLE_WITH_HIP
  miopenHandle_t cudnn_handle_;
#else
  cudnnHandle_t cudnn_handle_;
#endif
  std::unique_ptr<CublasHandleHolder> cublas_handle_;
  std::unique_ptr<CublasHandleHolder> cublas_tensor_core_handle_;
  std::unique_ptr<CublasHandleHolder> cublas_tf32_tensor_core_handle_;
#ifndef PADDLE_WITH_HIP
  cusolverDnHandle_t cusolver_dn_handle_;
  std::unique_ptr<CusparseHandleHolder> cusparse_handle_;
#endif
  DISABLE_COPY_AND_ASSIGN(CUDAContext);
};

class CUDADeviceContext : public DeviceContext {
 public:
  explicit CUDADeviceContext(CUDAPlace place);
  virtual ~CUDADeviceContext();

  /*! \brief  Wait for all operations completion in the stream. */
  void Wait() const override;

  /*! \brief  Return place in the device context. */
  Place GetPlace() const override;

  /*! \brief  Return compute capability in the device context. */
  int GetComputeCapability() const;

  /*! \brief  Return the max physical thread count in the device context */
  int GetMaxPhysicalThreadCount() const;

  /*! \brief  Return the SM count in the device context */
  int GetSMCount() const;

  /*! \brief  Return the Max thread num of block in the device context */
  int GetMaxThreadsPerBlock() const;

  /*! \brief  Return the max grid dim size in the device context */
  dim3 GetCUDAMaxGridDimSize() const;

  /*! \brief  Return eigen device in the device context. */
  Eigen::GpuDevice* eigen_device() const;

  /*! \brief  Call cublas function safely. */
  template <typename Callback>
  inline void CublasCall(Callback&& callback) const {
    return context()->CublasCall(callback);
  }

#ifndef PADDLE_WITH_HIP
  /*! \brief  Call cusparse function safely. */
  template <typename Callback>
  inline void CusparseCall(Callback&& callback) const {
    return context()->CusparseCall(callback);
  }
#endif

  /*! \brief  Check whether tensor core is supported */
  bool tensor_core_available() const;

  /*! \brief  Call cublas function with Tensor Core safely. If
      Tensor Core is not available, use DEFAULT_MATH instead. */
  template <typename Callback>
  inline void TensorCoreCublasCallIfAvailable(Callback&& callback) const {
    return context()->TensorCoreCublasCallIfAvailable(callback);
  }

/*! \brief  Return cudnn  handle in the device context. */
#ifdef PADDLE_WITH_HIP
  miopenHandle_t cudnn_handle() const;
#else
  cudnnHandle_t cudnn_handle() const;
#endif

/*! \brief  Return cublas handle in the device context. */
#ifdef PADDLE_WITH_HIP
  rocblas_handle cublas_handle() const;
#else
  cublasHandle_t cublas_handle() const;
  cusparseHandle_t cusparse_handle() const;
#endif

  /*! \brief  Return a cudnn workspace handle to call multiple cudnn
   *  functions without interrupting by other threads.
   *  Once the first cudnn function is called by the handle, a lock
   *  would be acquired to prevent other threads from accessing the
   *  workspace. Once the handle is destructed, the lock would be released.
   *  CudnnWorkspaceHandle is an RAII object to implement thread-safe
   *  sequential cudnn function calls. */
  CudnnWorkspaceHandle cudnn_workspace_handle() const;

#ifndef PADDLE_WITH_HIP
  cusolverDnHandle_t cusolver_dn_handle() const;
#endif

  /*! \brief  Return cuda stream in the device context. */
  gpuStream_t stream() const;

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
  /*! \brief  Return nccl communicators. */
  ncclComm_t nccl_comm() const { return nccl_comm_; }

  /*! \brief  Set nccl communicators. */
  void set_nccl_comm(ncclComm_t comm) { nccl_comm_ = comm; }
#endif

  template <typename Callback>
  void RecordEvent(gpuEvent_t ev, Callback callback) const {
    return context()->Stream()->RecordEvent(ev, callback);
  }

  template <typename Callback>
  void AddStreamCallback(Callback&& callback) const {
    return context()->Stream()->AddCallback(callback);
  }

  void WaitStreamCallback() const {
    return context()->Stream()->WaitCallback();
  }

  void ResetDefaultContext(const stream::Priority& priority) {
    default_ctx_.reset(new CUDAContext(place_, priority));
  }

  void ResetThreadContext(const stream::Priority& priority) {
    std::lock_guard<std::mutex> guard(ctx_mtx_);
    thread_ctx_[this].reset(new CUDAContext(place_, priority));
  }

  std::shared_ptr<CUDAContext> context() const {
    if (!thread_ctx_.count(this)) {
      return default_ctx_;
    }
    return thread_ctx_.at(this);
  }

  // Note: Can only be used under thread_local semantics.
  void SetThreadLocalStream(const gpuStream_t stream) {
    thread_ctx_.at(this)->SetStream(stream);
  }

 private:
  CUDAPlace place_;
  std::shared_ptr<CUDAContext> default_ctx_;

  // The thread_local static variable will be released before the
  // global static variable, so avoid using it in dtor.
  static thread_local std::unordered_map<const CUDADeviceContext*,
                                         std::shared_ptr<CUDAContext>>
      thread_ctx_;
  static thread_local std::mutex ctx_mtx_;

  mutable std::mutex cudnn_handle_mtx_;

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
  // NCCL communicator (single process version) for NCCL collective operations.
  // NCCL collective operations provides fast collectives over multiple GPUs
  // both within and across nodes.
  // But, this collectives is used for collectives over multiple GPUs within
  // nodes.
  ncclComm_t nccl_comm_{nullptr};
#endif

  int compute_capability_;
  int runtime_version_;
  int driver_version_;
  int multi_process_;
  int max_threads_per_mp_;
  int max_threads_per_block_;
  dim3 max_grid_dim_size_;

  DISABLE_COPY_AND_ASSIGN(CUDADeviceContext);
};

class CudnnWorkspaceHandle {
 public:
  inline CudnnWorkspaceHandle(const CUDADeviceContext& dev_ctx, std::mutex* mtx)
      : device_context_(dev_ctx), mtx_(mtx) {}

  template <typename Callback>
  inline void RunFunc(Callback&& cudnn_func, size_t required_workspace_bytes) {
    if (required_workspace_bytes > WorkspaceSize()) {
      ReallocWorkspace(required_workspace_bytes);
    }
    VLOG(2) << "Cudnn workspace size at RunFunc: "
            << static_cast<double>(WorkspaceSize()) / (1 << 20) << " MB";
    {
      std::lock_guard<std::mutex> guard(*mtx_);
      cudnn_func(allocation_ ? allocation_->ptr() : nullptr);
    }
  }

  /*! \brief Thread which call RunFuncSync() would release gpu memory after
   *  running the function. Currently this function is only used when cudnn
   *  exhaustive searching and callers have to guarantee that the input function
   *  is host blocking */
  template <typename Callback>
  inline void RunFuncSync(Callback&& cudnn_func,
                          size_t required_workspace_bytes) {
    RunFunc(cudnn_func, required_workspace_bytes);
    ResetWorkspace();
  }

  void ReallocWorkspace(size_t required_workspace_bytes);

  inline void ResetWorkspace() { allocation_ = nullptr; }

  inline size_t WorkspaceSize() {
    if (allocation_ == nullptr) {
      return 0;
    }
    return allocation_->size();
  }

  CudnnWorkspaceHandle(CudnnWorkspaceHandle&&) = default;
  CudnnWorkspaceHandle& operator=(CudnnWorkspaceHandle&&) = delete;

 private:
  memory::allocation::AllocationPtr allocation_;
  const CUDADeviceContext& device_context_;
  std::mutex* mtx_;
};

template <>
struct DefaultDeviceContextType<platform::CUDAPlace> {
  using TYPE = CUDADeviceContext;
};

// Currently, CUDAPinnedDeviceContext is only used to data copying.
class CUDAPinnedDeviceContext : public DeviceContext {
 public:
  CUDAPinnedDeviceContext();
  explicit CUDAPinnedDeviceContext(CUDAPinnedPlace place);

  Place GetPlace() const override;

  Eigen::DefaultDevice* eigen_device() const;

 private:
  CUDAPinnedPlace place_;
  std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
};

template <>
struct DefaultDeviceContextType<platform::CUDAPinnedPlace> {
  using TYPE = CUDAPinnedDeviceContext;
};
#endif

#ifdef PADDLE_WITH_MKLDNN

class MKLDNNDeviceContextThreadLocals {
  // default mkldnn session id

  typedef MKLDNNDeviceContextThreadLocals self;
  struct Body {
    bool said_once = false;
    size_t cur_mkldnn_session_id;
    // Current data input shape string.
    // - For fixed-shape, it's a null string in default.
    // - For dynamic-shape, it's user specific.
    std::string cur_input_shape_str;
    // the cache capacity of different input shapes for MKLDNN.
    // Default 1 means fixed input shape, not dynamic shape.
    int cur_input_shape_cache_capacity;
    // Recently registered data_format. This is needed to
    // know for converting MKL-DNN Tensor to non MKL-DNN
    paddle::framework::DataLayout cur_paddle_data_layout;
    // MKL-DNN stream used for execution of primitives (per-thread)
    dnnl::engine cur_engine;
    dnnl::stream cur_stream;
    std::string key_suffix;  // Key identifying current Executor
    bool key_attach_thread_id = true;
    void* exec_ptr_ = nullptr;

    Body();
    ~Body();
    void set_cur_mkldnn_session_id(size_t sid);
    size_t get_cur_mkldnn_session_id(void);
    void set_cur_input_shape_str(std::string input_shape_str);
    void set_cur_input_shape_cache_capacity(int input_shape_cache_capacity);
    void set_cur_paddle_data_layout(framework::DataLayout dl);
    framework::DataLayout get_cur_paddle_data_layout(void);
    void log_lib_version(void);
    const dnnl::engine& get_engine(void);
    dnnl::stream& get_stream(void);
    void set_key_suffix(const std::string& suffix) { key_suffix = suffix; }
    const std::string& get_key_suffix(void) const { return key_suffix; }
    void disable_tid_in_key(void) { key_attach_thread_id = false; }
    bool is_tid_used_in_key(void) const { return key_attach_thread_id; }
    void set_curr_exec(void* exec_ptr) { exec_ptr_ = exec_ptr; }
    void* get_curr_exec(void) const { return exec_ptr_; }
  };
  MKLDNNDeviceContextThreadLocals() = default;
  MKLDNNDeviceContextThreadLocals(const MKLDNNDeviceContextThreadLocals& c) =
      delete;

 public:
  // default mkldnn session id
  static constexpr size_t kMKLDNNSessionID_Default = 0;
  // mkldnn session id for cache clearing mode
  static constexpr size_t kMKLDNNSessionID_CacheClearing = -1;
  static Body& fetch() {
    thread_local Body b;
    return b;
  }
};

class MKLDNNDeviceContext : public CPUDeviceContext {
 public:
  template <class T>
  using BlobPtr_t = std::shared_ptr<T>;
  template <class P1, class P2>
  using umap_value_smart_t = std::unordered_map<P1, BlobPtr_t<P2>>;
  template <class T>
  using umap_key_string_t = umap_value_smart_t<std::string, T>;

  // Following three maps are used to cache MKLDNN primitives.
  // There relations are:
  // - BlobMap = Map<cur_thread_id, ShapeBlob>
  // - ShapeBlob = Map<cur_input_shape_str, KeyBlob>
  // - KeyBlob  = Map<blob_name, blob>

  using KeyBlob = umap_key_string_t<void>;
  using ShapeBlob = umap_key_string_t<KeyBlob>;
  using BlobMap = umap_value_smart_t<int, ShapeBlob>;

  // Auxillary two-level structure (shape, executor) to easier control
  // clearing cache objects related to specific executor

  using ExecKey = void*;
  using ExecMapCacheIterPair = std::pair<BlobPtr_t<KeyBlob>, KeyBlob::iterator>;
  using ExecMap =
      std::unordered_map<ExecKey, std::vector<ExecMapCacheIterPair>>;
  using ExecShape = std::unordered_map<std::string, std::shared_ptr<ExecMap>>;

  explicit MKLDNNDeviceContext(CPUPlace place);

  /* \brief  Get the active engine */
  const dnnl::engine& GetEngine() const { return tls().get_engine(); }

  // Register object to currently used executor's map
  void LinkEntryWithExecutor(BlobPtr_t<KeyBlob>, KeyBlob::iterator) const;
  void RemoveShapeEntriesWithExecutor(void) const;

  // Remove all entries from the blob map
  void ResetBlobMap(void* ptr);

  // Prevent next ResetBlobMap()
  void BlockNextCacheClearing();

  // Get the ShapeBlob size in cur_mkldnn_session_id.
  size_t GetShapeBlobSize() const;

  // Set data to blob (i.e. name/data pair). Create blob if not existing
  void SetBlob(const std::string& name, std::shared_ptr<void> data) const;

  // Calculate number of oneDNN objects cached
  unsigned int GetCachedObjectsNumber(void) const;

  // Find a saved blob. Return nullptr if not found
  std::shared_ptr<void> GetBlob(const std::string& name) const;

  static auto tls() -> decltype(MKLDNNDeviceContextThreadLocals::fetch()) {
    return MKLDNNDeviceContextThreadLocals::fetch();
  }

 private:
  std::shared_ptr<BlobMap> p_blobmap_;
  // Map key is pointer of executor and value is a data(iterator in map) needed
  // to erase
  std::shared_ptr<ExecShape> p_exec_items_;
  std::shared_ptr<std::mutex> p_mutex_;
  bool block_next_cache_clearing_ = false;
};
#endif

/*! \brief device context pool singleton */
class DeviceContextPool {
 public:
  explicit DeviceContextPool(const std::vector<platform::Place>& places);

  static DeviceContextPool& Instance() {
    PADDLE_ENFORCE_NOT_NULL(pool,
                            platform::errors::PreconditionNotMet(
                                "Need to Create DeviceContextPool firstly!"));
    return *pool;
  }

  /*! \brief  Create should only called by Init function */
  static DeviceContextPool& Init(const std::vector<platform::Place>& places) {
    if (pool == nullptr) {
      pool = new DeviceContextPool(places);
    }
    return *pool;
  }

  static void SetPool(DeviceContextPool* dev_pool) { pool = dev_pool; }

  /*! \brief  Return handle of single device context. */
  platform::DeviceContext* Get(const platform::Place& place);

  template <typename Place>
  const typename DefaultDeviceContextType<Place>::TYPE* GetByPlace(
      const Place& place) {
    return reinterpret_cast<
        const typename DefaultDeviceContextType<Place>::TYPE*>(Get(place));
  }

  size_t size() const { return device_contexts_.size(); }

 private:
  static DeviceContextPool* pool;
  std::map<Place, std::shared_future<std::unique_ptr<DeviceContext>>>
      device_contexts_;
  DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
};

}  // namespace platform
}  // namespace paddle