cuda_async_view_memory_resource.hpp

Go to the documentation of this file.

/*
 * SPDX-FileCopyrightText: Copyright (c) 2021-2025, NVIDIA CORPORATION.
 * SPDX-License-Identifier: Apache-2.0
 */
#pragma once
 
#include <rmm/cuda_device.hpp>
#include <rmm/cuda_stream_view.hpp>
#include <rmm/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/detail/thrust_namespace.h>
#include <rmm/mr/device/device_memory_resource.hpp>
 
#include <cuda_runtime_api.h>
 
#include <cstddef>
 
namespace RMM_NAMESPACE {
namespace mr {
class cuda_async_view_memory_resource final : public device_memory_resource {
 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;  
 
 private:
  cudaMemPool_t cuda_pool_handle_{};
 
  void* do_allocate(std::size_t bytes, rmm::cuda_stream_view stream) override
  {
    void* ptr{nullptr};
    if (bytes > 0) {
      RMM_CUDA_TRY_ALLOC(cudaMallocFromPoolAsync(&ptr, bytes, pool_handle(), stream.value()),
                         bytes);
    }
    return ptr;
  }
 
  void do_deallocate(void* ptr,
                     [[maybe_unused]] std::size_t bytes,
                     rmm::cuda_stream_view stream) noexcept override
  {
    if (ptr != nullptr) { RMM_ASSERT_CUDA_SUCCESS(cudaFreeAsync(ptr, stream.value())); }
  }
 
  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
  {
    return dynamic_cast<cuda_async_view_memory_resource const*>(&other) != nullptr;
  }
};
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE