All Classes Files Functions Variables Typedefs Enumerations Enumerator Friends Modules Pages
aligned_resource_adaptor.hpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2021-2025, NVIDIA CORPORATION.
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 #pragma once
17 
18 #include <rmm/aligned.hpp>
19 #include <rmm/cuda_stream_view.hpp>
20 #include <rmm/detail/error.hpp>
21 #include <rmm/detail/export.hpp>
24 #include <rmm/resource_ref.hpp>
25 
26 #include <cstddef>
27 #include <mutex>
28 #include <unordered_map>
29 
30 namespace RMM_NAMESPACE {
31 namespace mr {
56 template <typename Upstream>
58  public:
70  std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT,
71  std::size_t alignment_threshold = default_alignment_threshold)
72  : upstream_{upstream}, alignment_{alignment}, alignment_threshold_{alignment_threshold}
73  {
74  RMM_EXPECTS(rmm::is_supported_alignment(alignment),
75  "Allocation alignment is not a power of 2.");
76  }
77 
89  explicit aligned_resource_adaptor(Upstream* upstream,
90  std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT,
91  std::size_t alignment_threshold = default_alignment_threshold)
92  : upstream_{to_device_async_resource_ref_checked(upstream)},
93  alignment_{alignment},
94  alignment_threshold_{alignment_threshold}
95  {
96  RMM_EXPECTS(rmm::is_supported_alignment(alignment),
97  "Allocation alignment is not a power of 2.");
98  }
99 
100  aligned_resource_adaptor() = delete;
101  ~aligned_resource_adaptor() override = default;
104  aligned_resource_adaptor& operator=(aligned_resource_adaptor const&) = delete;
105  aligned_resource_adaptor& operator=(aligned_resource_adaptor&&) = delete;
106 
111  {
112  return upstream_;
113  }
114 
118  static constexpr std::size_t default_alignment_threshold = 0;
119 
120  private:
121  using lock_guard = std::lock_guard<std::mutex>;
122 
134  void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
135  {
136  if (alignment_ == rmm::CUDA_ALLOCATION_ALIGNMENT || bytes < alignment_threshold_) {
137  return get_upstream_resource().allocate_async(bytes, 1, stream);
138  }
139  auto const size = upstream_allocation_size(bytes);
140  void* pointer = get_upstream_resource().allocate_async(size, 1, stream);
141  // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
142  auto const address = reinterpret_cast<std::size_t>(pointer);
143  auto const aligned_address = rmm::align_up(address, alignment_);
144  // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast,performance-no-int-to-ptr)
145  void* aligned_pointer = reinterpret_cast<void*>(aligned_address);
146  if (pointer != aligned_pointer) {
147  lock_guard lock(mtx_);
148  pointers_.emplace(aligned_pointer, pointer);
149  }
150  return aligned_pointer;
151  }
152 
160  void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) override
161  {
162  if (alignment_ == rmm::CUDA_ALLOCATION_ALIGNMENT || bytes < alignment_threshold_) {
163  get_upstream_resource().deallocate_async(ptr, bytes, 1, stream);
164  } else {
165  {
166  lock_guard lock(mtx_);
167  auto const iter = pointers_.find(ptr);
168  if (iter != pointers_.end()) {
169  ptr = iter->second;
170  pointers_.erase(iter);
171  }
172  }
173  get_upstream_resource().deallocate_async(ptr, upstream_allocation_size(bytes), 1, stream);
174  }
175  }
176 
184  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
185  {
186  if (this == &other) { return true; }
187  auto cast = dynamic_cast<aligned_resource_adaptor<Upstream> const*>(&other);
188  if (cast == nullptr) { return false; }
189  return get_upstream_resource() == cast->get_upstream_resource() &&
190  alignment_ == cast->alignment_ && alignment_threshold_ == cast->alignment_threshold_;
191  }
192 
200  std::size_t upstream_allocation_size(std::size_t bytes) const
201  {
202  auto const aligned_size = rmm::align_up(bytes, alignment_);
203  return aligned_size + alignment_ - rmm::CUDA_ALLOCATION_ALIGNMENT;
204  }
205 
207  device_async_resource_ref upstream_;
208  std::unordered_map<void*, void*> pointers_;
209  std::size_t alignment_;
210  std::size_t alignment_threshold_;
211  mutable std::mutex mtx_;
212 };
213  // end of group
215 } // namespace mr
216 } // namespace RMM_NAMESPACE
Strongly-typed non-owning wrapper for CUDA streams with default constructor.
Definition: cuda_stream_view.hpp:39
Resource that adapts Upstream memory resource to allocate memory in a specified alignment size.
Definition: aligned_resource_adaptor.hpp:57
aligned_resource_adaptor(device_async_resource_ref upstream, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT, std::size_t alignment_threshold=default_alignment_threshold)
Construct an aligned resource adaptor using upstream to satisfy allocation requests.
Definition: aligned_resource_adaptor.hpp:69
aligned_resource_adaptor(Upstream *upstream, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT, std::size_t alignment_threshold=default_alignment_threshold)
Construct an aligned resource adaptor using upstream to satisfy allocation requests.
Definition: aligned_resource_adaptor.hpp:89
rmm::device_async_resource_ref get_upstream_resource() const noexcept
rmm::device_async_resource_ref to the upstream resource
Definition: aligned_resource_adaptor.hpp:110
Base class for all librmm device memory allocation.
Definition: device_memory_resource.hpp:93
cuda::mr::async_resource_ref< cuda::mr::device_accessible > device_async_resource_ref
Alias for a cuda::mr::async_resource_ref with the property cuda::mr::device_accessible.
Definition: resource_ref.hpp:41
device_async_resource_ref to_device_async_resource_ref_checked(Resource *res)
Convert pointer to memory resource into device_async_resource_ref, checking for nullptr
Definition: resource_ref.hpp:79
static constexpr std::size_t CUDA_ALLOCATION_ALIGNMENT
Default alignment used for CUDA memory allocation.
Definition: aligned.hpp:43
constexpr bool is_supported_alignment(std::size_t alignment) noexcept
Returns whether or not alignment is a valid memory alignment.
Definition: aligned.hpp:64
constexpr std::size_t align_up(std::size_t value, std::size_t alignment) noexcept
Align up to nearest multiple of specified power of 2.
Definition: aligned.hpp:77
Management of per-device device_memory_resources.