sam_headroom_memory_resource.hpp

Go to the documentation of this file.

/*
 * SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION.
 * SPDX-License-Identifier: Apache-2.0
 */
#pragma once
 
#include <rmm/cuda_device.hpp>
#include <rmm/cuda_stream_view.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/mr/device/device_memory_resource.hpp>
#include <rmm/mr/device/system_memory_resource.hpp>
#include <rmm/resource_ref.hpp>
 
#include <algorithm>
#include <cstddef>
 
namespace RMM_NAMESPACE {
namespace mr {
class sam_headroom_memory_resource final : public device_memory_resource {
 public:
  explicit sam_headroom_memory_resource(std::size_t headroom) : system_mr_{}, headroom_{headroom} {}
 
  sam_headroom_memory_resource()                                               = delete;
  ~sam_headroom_memory_resource() override                                     = default;
  sam_headroom_memory_resource(sam_headroom_memory_resource const&)            = delete;
  sam_headroom_memory_resource(sam_headroom_memory_resource&&)                 = delete;
  sam_headroom_memory_resource& operator=(sam_headroom_memory_resource const&) = delete;
  sam_headroom_memory_resource& operator=(sam_headroom_memory_resource&&)      = delete;
 
 private:
  void* do_allocate(std::size_t bytes, [[maybe_unused]] cuda_stream_view stream) override
  {
    void* pointer = system_mr_.allocate(stream, bytes, rmm::CUDA_ALLOCATION_ALIGNMENT);
 
    auto const free        = rmm::available_device_memory().first;
    auto const allocatable = free > headroom_ ? free - headroom_ : 0UL;
    auto const gpu_portion =
      rmm::align_down(std::min(allocatable, bytes), rmm::CUDA_ALLOCATION_ALIGNMENT);
    auto const cpu_portion = bytes - gpu_portion;
 
    if (gpu_portion != 0) {
#if defined(CUDART_VERSION) && CUDART_VERSION >= 13000
      cudaMemLocation location{cudaMemLocationTypeDevice, rmm::get_current_cuda_device().value()};
      RMM_CUDA_TRY(
        cudaMemAdvise(pointer, gpu_portion, cudaMemAdviseSetPreferredLocation, location));
#else
      RMM_CUDA_TRY(cudaMemAdvise(pointer,
                                 gpu_portion,
                                 cudaMemAdviseSetPreferredLocation,
                                 rmm::get_current_cuda_device().value()));
#endif
    }
    if (cpu_portion != 0) {
#if defined(CUDART_VERSION) && CUDART_VERSION >= 13000
      cudaMemLocation location{cudaMemLocationTypeHost, 0};
      RMM_CUDA_TRY(cudaMemAdvise(static_cast<char*>(pointer) + gpu_portion,
                                 cpu_portion,
                                 cudaMemAdviseSetPreferredLocation,
                                 location));
#else
      RMM_CUDA_TRY(cudaMemAdvise(static_cast<char*>(pointer) + gpu_portion,
                                 cpu_portion,
                                 cudaMemAdviseSetPreferredLocation,
                                 cudaCpuDeviceId));
#endif
    }
 
    return pointer;
  }
 
  void do_deallocate(void* ptr,
                     [[maybe_unused]] std::size_t bytes,
                     [[maybe_unused]] cuda_stream_view stream) noexcept override
  {
    system_mr_.deallocate(stream, ptr, bytes, rmm::CUDA_ALLOCATION_ALIGNMENT);
  }
 
  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
  {
    if (this == &other) { return true; }
    auto cast = dynamic_cast<sam_headroom_memory_resource const*>(&other);
    if (cast == nullptr) { return false; }
    return headroom_ == cast->headroom_;
  }
 
  system_memory_resource system_mr_;
  std::size_t headroom_;
};  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE