cuda_async_view_memory_resource.hpp

Go to the documentation of this file.

/*
 * SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.
 * SPDX-License-Identifier: Apache-2.0
 */
#pragma once
 
#include <rmm/aligned.hpp>
#include <rmm/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/detail/runtime_capabilities.hpp>
 
#include <cuda/memory_resource>
#include <cuda/stream_ref>
#include <cuda_runtime_api.h>
 
#include <cstddef>
 
namespace RMM_NAMESPACE {
namespace mr {
class cuda_async_view_memory_resource final {
 public:
  cuda_async_view_memory_resource(cudaMemPool_t pool_handle)
    : cuda_pool_handle_{[pool_handle]() {
        RMM_EXPECTS(nullptr != pool_handle, "Unexpected null pool handle.");
        return pool_handle;
      }()}
  {
    // Check if cudaMallocAsync Memory pool supported
    RMM_EXPECTS(rmm::detail::runtime_async_alloc::is_supported(),
                "cudaMallocAsync not supported with this CUDA driver/runtime version");
  }
 
  [[nodiscard]] cudaMemPool_t pool_handle() const noexcept { return cuda_pool_handle_; }
 
  cuda_async_view_memory_resource()  = default;
  ~cuda_async_view_memory_resource() = default;
  cuda_async_view_memory_resource(cuda_async_view_memory_resource const&) =
    default;  
  cuda_async_view_memory_resource(cuda_async_view_memory_resource&&) =
    default;  
  cuda_async_view_memory_resource& operator=(cuda_async_view_memory_resource const&) =
    default;  
  cuda_async_view_memory_resource& operator=(cuda_async_view_memory_resource&&) =
    default;  
 
  void* allocate(cuda::stream_ref stream,
                 std::size_t bytes,
                 [[maybe_unused]] std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT)
  {
    void* ptr{nullptr};
    if (bytes > 0) {
      RMM_CUDA_TRY_ALLOC(cudaMallocFromPoolAsync(&ptr, bytes, pool_handle(), stream.get()), bytes);
    }
    return ptr;
  }
 
  void deallocate(cuda::stream_ref stream,
                  void* ptr,
                  [[maybe_unused]] std::size_t bytes,
                  [[maybe_unused]] std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT) noexcept
  {
    if (ptr != nullptr) { RMM_ASSERT_CUDA_SUCCESS_SAFE_SHUTDOWN(cudaFreeAsync(ptr, stream.get())); }
  }
 
  void* allocate_sync(std::size_t bytes, std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT)
  {
    auto* ptr = allocate(cuda::stream_ref{cudaStream_t{nullptr}}, bytes, alignment);
    RMM_CUDA_TRY(cudaStreamSynchronize(cudaStream_t{nullptr}));
    return ptr;
  }
 
  void deallocate_sync(void* ptr,
                       std::size_t bytes,
                       std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT) noexcept
  {
    deallocate(cuda::stream_ref{cudaStream_t{nullptr}}, ptr, bytes, alignment);
  }
 
  [[nodiscard]] bool operator==(cuda_async_view_memory_resource const& other) const noexcept
  {
    return pool_handle() == other.pool_handle();
  }
 
  [[nodiscard]] bool operator!=(cuda_async_view_memory_resource const& other) const noexcept
  {
    return !operator==(other);
  }
 
  RMM_CONSTEXPR_FRIEND void get_property(cuda_async_view_memory_resource const&,
                                         cuda::mr::device_accessible) noexcept
  {
  }
 
 private:
  cudaMemPool_t cuda_pool_handle_{};
};
 
// static property checks
static_assert(cuda::mr::synchronous_resource<cuda_async_view_memory_resource>);
static_assert(cuda::mr::resource<cuda_async_view_memory_resource>);
static_assert(cuda::mr::synchronous_resource_with<cuda_async_view_memory_resource,
                                                  cuda::mr::device_accessible>);
static_assert(
  cuda::mr::resource_with<cuda_async_view_memory_resource, cuda::mr::device_accessible>);
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE