pool_memory_resource.hpp

Go to the documentation of this file.

/*
 * SPDX-FileCopyrightText: Copyright (c) 2020-2026, 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/detail/format.hpp>
#include <rmm/detail/logging_assert.hpp>
#include <rmm/logger.hpp>
#include <rmm/mr/detail/coalescing_free_list.hpp>
#include <rmm/mr/detail/stream_ordered_memory_resource.hpp>
#include <rmm/mr/device_memory_resource.hpp>
#include <rmm/mr/per_device_resource.hpp>
#include <rmm/resource_ref.hpp>
 
#include <cuda/std/type_traits>
#include <cuda_runtime_api.h>
 
#include <algorithm>
#include <cstddef>
#include <memory>
#include <mutex>
#include <optional>
#include <set>
 
namespace RMM_NAMESPACE {
namespace mr {
namespace detail {
template <class PoolResource, class Upstream, class Property, class = void>
struct maybe_remove_property {};
 
template <class PoolResource, class Upstream, class Property>
struct maybe_remove_property<PoolResource,
                             Upstream,
                             Property,
                             cuda::std::enable_if_t<!cuda::has_property<Upstream, Property>>> {
#if defined(__GNUC__) && !defined(__clang__)  // GCC warns about compatibility
                                              // issues with pre ISO C++ code
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wnon-template-friend"
#endif  // __GNUC__ and not __clang__
  friend void get_property(const PoolResource&, Property) = delete;
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic pop
#endif  // __GNUC__ and not __clang__
};
}  // namespace detail
 
template <typename Upstream>
class pool_memory_resource final
  : public detail::
      maybe_remove_property<pool_memory_resource<Upstream>, Upstream, cuda::mr::device_accessible>,
    public detail::stream_ordered_memory_resource<pool_memory_resource<Upstream>,
                                                  detail::coalescing_free_list>,
    public cuda::forward_property<pool_memory_resource<Upstream>, Upstream> {
 public:
  friend class detail::stream_ordered_memory_resource<pool_memory_resource<Upstream>,
                                                      detail::coalescing_free_list>;
 
  explicit pool_memory_resource(device_async_resource_ref upstream_mr,
                                std::size_t initial_pool_size,
                                std::optional<std::size_t> maximum_pool_size = std::nullopt)
    : upstream_mr_{upstream_mr}
  {
    RMM_EXPECTS(rmm::is_aligned(initial_pool_size, rmm::CUDA_ALLOCATION_ALIGNMENT),
                "Error, Initial pool size required to be a multiple of 256 bytes");
    RMM_EXPECTS(rmm::is_aligned(maximum_pool_size.value_or(0), rmm::CUDA_ALLOCATION_ALIGNMENT),
                "Error, Maximum pool size required to be a multiple of 256 bytes");
 
    initialize_pool(initial_pool_size, maximum_pool_size);
  }
 
  explicit pool_memory_resource(Upstream* upstream_mr,
                                std::size_t initial_pool_size,
                                std::optional<std::size_t> maximum_pool_size = std::nullopt)
    : upstream_mr_{to_device_async_resource_ref_checked(upstream_mr)}
  {
    RMM_EXPECTS(rmm::is_aligned(initial_pool_size, rmm::CUDA_ALLOCATION_ALIGNMENT),
                "Error, Initial pool size required to be a multiple of 256 bytes");
    RMM_EXPECTS(rmm::is_aligned(maximum_pool_size.value_or(0), rmm::CUDA_ALLOCATION_ALIGNMENT),
                "Error, Maximum pool size required to be a multiple of 256 bytes");
 
    initialize_pool(initial_pool_size, maximum_pool_size);
  }
 
  template <typename Upstream2 = Upstream>
  explicit pool_memory_resource(Upstream2& upstream_mr,
                                std::size_t initial_pool_size,
                                std::optional<std::size_t> maximum_pool_size = std::nullopt)
    : pool_memory_resource(cuda::std::addressof(upstream_mr), initial_pool_size, maximum_pool_size)
  {
  }
 
  ~pool_memory_resource() override { release(); }
 
  pool_memory_resource()                                       = delete;
  pool_memory_resource(pool_memory_resource const&)            = delete;
  pool_memory_resource(pool_memory_resource&&)                 = delete;
  pool_memory_resource& operator=(pool_memory_resource const&) = delete;
  pool_memory_resource& operator=(pool_memory_resource&&)      = delete;
 
  [[nodiscard]] device_async_resource_ref get_upstream_resource() const noexcept
  {
    return upstream_mr_;
  }
 
  [[nodiscard]] std::size_t pool_size() const noexcept { return current_pool_size_; }
 
 protected:
  using free_list  = detail::coalescing_free_list;  
  using block_type = free_list::block_type;         
  using typename detail::stream_ordered_memory_resource<pool_memory_resource<Upstream>,
                                                        detail::coalescing_free_list>::split_block;
  using lock_guard = std::lock_guard<std::mutex>;  
 
  [[nodiscard]] std::size_t get_maximum_allocation_size() const
  {
    return std::numeric_limits<std::size_t>::max();
  }
 
  block_type try_to_expand(std::size_t try_size, std::size_t min_size, cuda_stream_view stream)
  {
    auto report_error = [&](const char* reason) {
      RMM_LOG_ERROR("[A][Stream %s][Upstream %zuB][FAILURE maximum pool size exceeded: %s]",
                    rmm::detail::format_stream(stream),
                    min_size,
                    reason);
      auto const msg = std::string("Maximum pool size exceeded (failed to allocate ") +
                       rmm::detail::format_bytes(min_size) + std::string("): ") + reason;
      RMM_FAIL(msg.c_str(), rmm::out_of_memory);
    };
 
    while (try_size >= min_size) {
      try {
        auto block = block_from_upstream(try_size, stream);
        current_pool_size_ += block.size();
        return block;
      } catch (std::exception const& e) {
        if (try_size == min_size) { report_error(e.what()); }
      }
      try_size = std::max(min_size, try_size / 2);
    }
 
    auto const max_size = maximum_pool_size_.value_or(std::numeric_limits<std::size_t>::max());
    auto const msg      = std::string("Not enough room to grow, current/max/try size = ") +
                     rmm::detail::format_bytes(pool_size()) + ", " +
                     rmm::detail::format_bytes(max_size) + ", " +
                     rmm::detail::format_bytes(min_size);
    report_error(msg.c_str());
    return {};
  }
 
  void initialize_pool(std::size_t initial_size, std::optional<std::size_t> maximum_size)
  {
    current_pool_size_ = 0;  // try_to_expand will set this if it succeeds
    maximum_pool_size_ = maximum_size;
 
    RMM_EXPECTS(
      initial_size <= maximum_pool_size_.value_or(std::numeric_limits<std::size_t>::max()),
      "Initial pool size exceeds the maximum pool size!");
 
    if (initial_size > 0) {
      auto const block = try_to_expand(initial_size, initial_size, cuda_stream_legacy);
      this->insert_block(block, cuda_stream_legacy);
    }
  }
 
  block_type expand_pool(std::size_t size,
                         [[maybe_unused]] free_list& blocks,
                         cuda_stream_view stream)
  {
    // Strategy: If maximum_pool_size_ is set, then grow geometrically, e.g. by halfway to the
    // limit each time. If it is not set, grow exponentially, e.g. by doubling the pool size each
    // time. Upon failure, attempt to back off exponentially, e.g. by half the attempted size,
    // until either success or the attempt is less than the requested size.
 
    return try_to_expand(size_to_grow(size), size, stream);
  }
 
  [[nodiscard]] std::size_t size_to_grow(std::size_t size) const
  {
    if (maximum_pool_size_.has_value()) {
      auto const unaligned_remaining = maximum_pool_size_.value() - pool_size();
      auto const remaining    = rmm::align_up(unaligned_remaining, rmm::CUDA_ALLOCATION_ALIGNMENT);
      auto const aligned_size = rmm::align_up(size, rmm::CUDA_ALLOCATION_ALIGNMENT);
      return (aligned_size <= remaining) ? std::max(aligned_size, remaining / 2) : 0;
    }
    return std::max(size, pool_size());
  };
 
  block_type block_from_upstream(std::size_t size, cuda_stream_view stream)
  {
    RMM_LOG_DEBUG("[A][Stream %s][Upstream %zuB]", rmm::detail::format_stream(stream), size);
 
    if (size == 0) { return {}; }
 
    void* ptr = get_upstream_resource().allocate(stream, size);
    return *upstream_blocks_.emplace(static_cast<char*>(ptr), size, true).first;
  }
 
  split_block allocate_from_block(block_type const& block, std::size_t size)
  {
    block_type const alloc{block.pointer(), size, block.is_head()};
#ifdef RMM_POOL_TRACK_ALLOCATIONS
    allocated_blocks_.insert(alloc);
#endif
 
    auto rest = (block.size() > size)
                  // NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
                  ? block_type{block.pointer() + size, block.size() - size, false}
                  : block_type{};
    return {alloc, rest};
  }
 
  block_type free_block(void* ptr, std::size_t size) noexcept
  {
#ifdef RMM_POOL_TRACK_ALLOCATIONS
    if (ptr == nullptr) return block_type{};
    auto const iter = allocated_blocks_.find(static_cast<char*>(ptr));
    RMM_LOGGING_ASSERT(iter != allocated_blocks_.end());
 
    auto block = *iter;
    RMM_LOGGING_ASSERT(block.size() == rmm::align_up(size, allocation_alignment));
    allocated_blocks_.erase(iter);
 
    return block;
#else
    auto const iter = upstream_blocks_.find(static_cast<char*>(ptr));
    return block_type{static_cast<char*>(ptr), size, (iter != upstream_blocks_.end())};
#endif
  }
 
  void release()
  {
    lock_guard lock(this->get_mutex());
 
    for (auto block : upstream_blocks_) {
      get_upstream_resource().deallocate_sync(block.pointer(), block.size());
    }
    upstream_blocks_.clear();
#ifdef RMM_POOL_TRACK_ALLOCATIONS
    allocated_blocks_.clear();
#endif
 
    current_pool_size_ = 0;
  }
 
#ifdef RMM_DEBUG_PRINT
  void print()
  {
    lock_guard lock(this->get_mutex());
 
    auto const [free, total] = rmm::available_device_memory();
    std::cout << "GPU free memory: " << free << " total: " << total << "\n";
 
    std::cout << "upstream_blocks: " << upstream_blocks_.size() << "\n";
    std::size_t upstream_total{0};
 
    for (auto blocks : upstream_blocks_) {
      blocks.print();
      upstream_total += blocks.size();
    }
    std::cout << "total upstream: " << upstream_total << " B\n";
 
#ifdef RMM_POOL_TRACK_ALLOCATIONS
    std::cout << "allocated_blocks: " << allocated_blocks_.size() << "\n";
    for (auto block : allocated_blocks_)
      block.print();
#endif
 
    this->print_free_blocks();
  }
#endif
 
  std::pair<std::size_t, std::size_t> free_list_summary(free_list const& blocks)
  {
    std::size_t largest{};
    std::size_t total{};
    std::for_each(blocks.cbegin(), blocks.cend(), [&largest, &total](auto const& block) {
      total += block.size();
      largest = std::max(largest, block.size());
    });
    return {largest, total};
  }
 
 private:
  // The "heap" to allocate the pool from
  device_async_resource_ref upstream_mr_;
  std::size_t current_pool_size_{};
  std::optional<std::size_t> maximum_pool_size_{};
 
#ifdef RMM_POOL_TRACK_ALLOCATIONS
  std::set<block_type, rmm::mr::detail::compare_blocks<block_type>> allocated_blocks_;
#endif
 
  // blocks allocated from upstream
  std::set<block_type, rmm::mr::detail::compare_blocks<block_type>> upstream_blocks_;
};  // namespace mr
  // end of group
}  // namespace mr
}  // namespace RMM_NAMESPACE