limiting_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/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/detail/format.hpp>
#include <rmm/mr/device/device_memory_resource.hpp>
#include <rmm/mr/device/per_device_resource.hpp>
#include <rmm/resource_ref.hpp>
 
#include <atomic>
#include <cstddef>
 
namespace RMM_NAMESPACE {
namespace mr {
template <typename Upstream>
class limiting_resource_adaptor final : public device_memory_resource {
 public:
  limiting_resource_adaptor(device_async_resource_ref upstream,
                            std::size_t allocation_limit,
                            std::size_t alignment = CUDA_ALLOCATION_ALIGNMENT)
    : upstream_{upstream},
      allocation_limit_{allocation_limit},
      allocated_bytes_(0),
      alignment_(alignment)
  {
  }
 
  limiting_resource_adaptor(Upstream* upstream,
                            std::size_t allocation_limit,
                            std::size_t alignment = CUDA_ALLOCATION_ALIGNMENT)
    : upstream_{to_device_async_resource_ref_checked(upstream)},
      allocation_limit_{allocation_limit},
      allocated_bytes_(0),
      alignment_(alignment)
  {
  }
 
  limiting_resource_adaptor()                                 = delete;
  ~limiting_resource_adaptor() override                       = default;
  limiting_resource_adaptor(limiting_resource_adaptor const&) = delete;
  limiting_resource_adaptor(limiting_resource_adaptor&&) noexcept =
    default;  
  limiting_resource_adaptor& operator=(limiting_resource_adaptor const&) = delete;
  limiting_resource_adaptor& operator=(limiting_resource_adaptor&&) noexcept =
    default;  
 
  [[nodiscard]] device_async_resource_ref get_upstream_resource() const noexcept
  {
    return upstream_;
  }
 
  [[nodiscard]] std::size_t get_allocated_bytes() const { return allocated_bytes_; }
 
  [[nodiscard]] std::size_t get_allocation_limit() const { return allocation_limit_; }
 
 private:
  void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
  {
    auto const proposed_size = align_up(bytes, alignment_);
    auto const old           = allocated_bytes_.fetch_add(proposed_size);
    if (old + proposed_size <= allocation_limit_) {
      try {
        return get_upstream_resource().allocate(stream, bytes);
      } catch (...) {
        allocated_bytes_ -= proposed_size;
        throw;
      }
    }
 
    allocated_bytes_ -= proposed_size;
    auto const msg = std::string("Exceeded memory limit (failed to allocate ") +
                     rmm::detail::format_bytes(bytes) + ")";
    RMM_FAIL(msg.c_str(), rmm::out_of_memory);
  }
 
  void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) noexcept override
  {
    std::size_t allocated_size = align_up(bytes, alignment_);
    get_upstream_resource().deallocate(stream, ptr, bytes);
    allocated_bytes_ -= allocated_size;
  }
 
  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
  {
    if (this == &other) { return true; }
    auto const* cast = dynamic_cast<limiting_resource_adaptor<Upstream> const*>(&other);
    if (cast == nullptr) { return false; }
    return get_upstream_resource() == cast->get_upstream_resource();
  }
 
  // The upstream resource used for satisfying allocation requests
  device_async_resource_ref upstream_;
 
  // maximum bytes this allocator is allowed to allocate.
  std::size_t allocation_limit_;
 
  // number of currently-allocated bytes
  std::atomic<std::size_t> allocated_bytes_;
 
  // todo: should be some way to ask the upstream...
  std::size_t alignment_;
};
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE