cuda_async_memory_resource.hpp

Go to the documentation of this file.

/*
 * Copyright (c) 2021-2025, NVIDIA CORPORATION.
 *
 * 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 <rmm/cuda_device.hpp>
#include <rmm/cuda_stream_view.hpp>
#include <rmm/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/detail/runtime_async_alloc.hpp>
#include <rmm/detail/thrust_namespace.h>
#include <rmm/mr/device/cuda_async_view_memory_resource.hpp>
#include <rmm/mr/device/device_memory_resource.hpp>
 
#include <cuda/std/type_traits>
#include <cuda_runtime_api.h>
 
#include <cstddef>
#include <cstdint>
#include <optional>
 
namespace RMM_NAMESPACE {
namespace mr {
class cuda_async_memory_resource final : public device_memory_resource {
 public:
  enum class allocation_handle_type : std::int32_t {
    none                  = 0x0,  
    posix_file_descriptor = 0x1,  
    win32     = 0x2,              
    win32_kmt = 0x4,  
    fabric    = 0x8   
  };
 
  enum class mempool_usage : unsigned short {
    hw_decompress = 0x2,  
  };
 
  // NOLINTNEXTLINE(bugprone-easily-swappable-parameters)
  cuda_async_memory_resource(std::optional<std::size_t> initial_pool_size             = {},
                             std::optional<std::size_t> release_threshold             = {},
                             std::optional<allocation_handle_type> export_handle_type = {})
  {
    // Check if cudaMallocAsync Memory pool supported
    RMM_EXPECTS(rmm::detail::runtime_async_alloc::is_supported(),
                "cudaMallocAsync not supported with this CUDA driver/runtime version");
 
    // Construct explicit pool
    cudaMemPoolProps pool_props{};
    pool_props.allocType   = cudaMemAllocationTypePinned;
    pool_props.handleTypes = static_cast<cudaMemAllocationHandleType>(
      export_handle_type.value_or(allocation_handle_type::none));
 
#if defined(CUDA_VERSION) && CUDA_VERSION >= RMM_MIN_HWDECOMPRESS_CUDA_DRIVER_VERSION
    // Enable hardware decompression if supported (requires CUDA 12.8 driver or higher)
    if (rmm::detail::runtime_async_alloc::is_hwdecompress_supported()) {
      pool_props.usage = static_cast<unsigned short>(mempool_usage::hw_decompress);
    }
#endif
 
    RMM_EXPECTS(
      rmm::detail::runtime_async_alloc::is_export_handle_type_supported(pool_props.handleTypes),
      "Requested IPC memory handle type not supported");
    pool_props.location.type = cudaMemLocationTypeDevice;
    pool_props.location.id   = rmm::get_current_cuda_device().value();
    cudaMemPool_t cuda_pool_handle{};
    RMM_CUDA_TRY(cudaMemPoolCreate(&cuda_pool_handle, &pool_props));
    pool_ = cuda_async_view_memory_resource{cuda_pool_handle};
 
    // CUDA drivers before 11.5 have known incompatibilities with the async allocator.
    // We'll disable `cudaMemPoolReuseAllowOpportunistic` if cuda driver < 11.5.
    // See https://github.com/NVIDIA/spark-rapids/issues/4710.
    int driver_version{};
    RMM_CUDA_TRY(cudaDriverGetVersion(&driver_version));
    constexpr auto min_async_driver_version{11050};
    if (driver_version < min_async_driver_version) {
      int disabled{0};
      RMM_CUDA_TRY(
        cudaMemPoolSetAttribute(pool_handle(), cudaMemPoolReuseAllowOpportunistic, &disabled));
    }
 
    auto const [free, total] = rmm::available_device_memory();
 
    // Need an l-value to take address to pass to cudaMemPoolSetAttribute
    uint64_t threshold = release_threshold.value_or(total);
    RMM_CUDA_TRY(
      cudaMemPoolSetAttribute(pool_handle(), cudaMemPoolAttrReleaseThreshold, &threshold));
 
    // Allocate and immediately deallocate the initial_pool_size to prime the pool with the
    // specified size
    auto const pool_size = initial_pool_size.value_or(free / 2);
    auto* ptr            = do_allocate(pool_size, cuda_stream_default);
    do_deallocate(ptr, pool_size, cuda_stream_default);
  }
 
  [[nodiscard]] cudaMemPool_t pool_handle() const noexcept { return pool_.pool_handle(); }
 
  ~cuda_async_memory_resource() override
  {
    RMM_ASSERT_CUDA_SUCCESS(cudaMemPoolDestroy(pool_handle()));
  }
  cuda_async_memory_resource(cuda_async_memory_resource const&)            = delete;
  cuda_async_memory_resource(cuda_async_memory_resource&&)                 = delete;
  cuda_async_memory_resource& operator=(cuda_async_memory_resource const&) = delete;
  cuda_async_memory_resource& operator=(cuda_async_memory_resource&&)      = delete;
 
 private:
  cuda_async_view_memory_resource pool_{};
 
  void* do_allocate(std::size_t bytes, rmm::cuda_stream_view stream) override
  {
    void* ptr{nullptr};
    ptr = pool_.allocate(bytes, stream);
    return ptr;
  }
 
  void do_deallocate(void* ptr, std::size_t bytes, rmm::cuda_stream_view stream) override
  {
    pool_.deallocate(ptr, bytes, stream);
  }
 
  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
  {
    auto const* async_mr = dynamic_cast<cuda_async_memory_resource const*>(&other);
    return (async_mr != nullptr) && (this->pool_handle() == async_mr->pool_handle());
  }
};
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE