limiting_resource_adaptor.hpp

Go to the documentation of this file.

/*
 * Copyright (c) 2021-2024, 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/aligned.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 <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_async(bytes, stream);
      } catch (...) {
        allocated_bytes_ -= proposed_size;
        throw;
      }
    }
 
    allocated_bytes_ -= proposed_size;
    RMM_FAIL("Exceeded memory limit", rmm::out_of_memory);
  }
 
  void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) override
  {
    std::size_t allocated_size = align_up(bytes, alignment_);
    get_upstream_resource().deallocate_async(ptr, bytes, stream);
    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_;
};
 
template <typename Upstream>
[[deprecated(
  "make_limiting_adaptor is deprecated in RMM 24.10. Use the limiting_resource_adaptor constructor "
  "instead.")]]
limiting_resource_adaptor<Upstream> make_limiting_adaptor(Upstream* upstream,
                                                          std::size_t allocation_limit)
{
  return limiting_resource_adaptor<Upstream>{upstream, allocation_limit};
}
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE