aligned_resource_adaptor.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/aligned.hpp>
#include <rmm/cuda_stream_view.hpp>
#include <rmm/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/mr/device/device_memory_resource.hpp>
#include <rmm/mr/device/per_device_resource.hpp>
#include <rmm/resource_ref.hpp>
 
#include <algorithm>
#include <cstddef>
#include <mutex>
#include <unordered_map>
 
namespace RMM_NAMESPACE {
namespace mr {
template <typename Upstream>
class aligned_resource_adaptor final : public device_memory_resource {
 public:
  explicit aligned_resource_adaptor(device_async_resource_ref upstream,
                                    std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT,
                                    std::size_t alignment_threshold = default_alignment_threshold)
    : upstream_{upstream},
      alignment_{std::max(alignment, rmm::CUDA_ALLOCATION_ALIGNMENT)},
      alignment_threshold_{alignment_threshold}
  {
    RMM_EXPECTS(rmm::is_supported_alignment(alignment),
                "Allocation alignment is not a power of 2.");
  }
 
  explicit aligned_resource_adaptor(Upstream* upstream,
                                    std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT,
                                    std::size_t alignment_threshold = default_alignment_threshold)
    : upstream_{to_device_async_resource_ref_checked(upstream)},
      alignment_{std::max(alignment, rmm::CUDA_ALLOCATION_ALIGNMENT)},
      alignment_threshold_{alignment_threshold}
  {
    RMM_EXPECTS(rmm::is_supported_alignment(alignment),
                "Allocation alignment is not a power of 2.");
  }
 
  aligned_resource_adaptor()                                           = delete;
  ~aligned_resource_adaptor() override                                 = default;
  aligned_resource_adaptor(aligned_resource_adaptor const&)            = delete;
  aligned_resource_adaptor(aligned_resource_adaptor&&)                 = delete;
  aligned_resource_adaptor& operator=(aligned_resource_adaptor const&) = delete;
  aligned_resource_adaptor& operator=(aligned_resource_adaptor&&)      = delete;
 
  [[nodiscard]] rmm::device_async_resource_ref get_upstream_resource() const noexcept
  {
    return upstream_;
  }
 
  static constexpr std::size_t default_alignment_threshold = 0;
 
 private:
  using lock_guard = std::lock_guard<std::mutex>;
 
  void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
  {
    if (alignment_ == rmm::CUDA_ALLOCATION_ALIGNMENT || bytes < alignment_threshold_) {
      return get_upstream_resource().allocate(stream, bytes, 1);
    }
    auto const size = upstream_allocation_size(bytes);
    void* pointer   = get_upstream_resource().allocate(stream, size, 1);
    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
    auto const address         = reinterpret_cast<std::size_t>(pointer);
    auto const aligned_address = rmm::align_up(address, alignment_);
    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast,performance-no-int-to-ptr)
    void* aligned_pointer = reinterpret_cast<void*>(aligned_address);
    if (pointer != aligned_pointer) {
      lock_guard lock(mtx_);
      pointers_.emplace(aligned_pointer, pointer);
    }
    return aligned_pointer;
  }
 
  void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) noexcept override
  {
    if (alignment_ == rmm::CUDA_ALLOCATION_ALIGNMENT || bytes < alignment_threshold_) {
      get_upstream_resource().deallocate(stream, ptr, bytes, 1);
    } else {
      {
        lock_guard lock(mtx_);
        auto const iter = pointers_.find(ptr);
        if (iter != pointers_.end()) {
          ptr = iter->second;
          pointers_.erase(iter);
        }
      }
      get_upstream_resource().deallocate(stream, ptr, upstream_allocation_size(bytes), 1);
    }
  }
 
  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
  {
    if (this == &other) { return true; }
    auto cast = dynamic_cast<aligned_resource_adaptor<Upstream> const*>(&other);
    if (cast == nullptr) { return false; }
    return get_upstream_resource() == cast->get_upstream_resource() &&
           alignment_ == cast->alignment_ && alignment_threshold_ == cast->alignment_threshold_;
  }
 
  std::size_t upstream_allocation_size(std::size_t bytes) const
  {
    auto const aligned_size = rmm::align_up(bytes, alignment_);
    // aligned_size: bytes of properly aligned space needed
    // (alignment_ - rmm::CUDA_ALLOCATION_ALIGNMENT): maximum "waste" due to pointer misalignment
    return aligned_size + (alignment_ - rmm::CUDA_ALLOCATION_ALIGNMENT);
  }
 
  device_async_resource_ref upstream_;
  std::unordered_map<void*, void*> pointers_;  
  std::size_t alignment_;                      
  std::size_t alignment_threshold_;  
  mutable std::mutex mtx_;           
};
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE