All Classes Files Functions Variables Typedefs Enumerations Enumerator Friends Modules Pages
limiting_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/detail/error.hpp>
20 #include <rmm/detail/export.hpp>
21 #include <rmm/detail/format.hpp>
24 #include <rmm/resource_ref.hpp>
25 
26 #include <atomic>
27 #include <cstddef>
28 
29 namespace RMM_NAMESPACE {
30 namespace mr {
48 template <typename Upstream>
50  public:
60  std::size_t allocation_limit,
61  std::size_t alignment = CUDA_ALLOCATION_ALIGNMENT)
62  : upstream_{upstream},
63  allocation_limit_{allocation_limit},
64  allocated_bytes_(0),
65  alignment_(alignment)
66  {
67  }
68 
79  limiting_resource_adaptor(Upstream* upstream,
80  std::size_t allocation_limit,
81  std::size_t alignment = CUDA_ALLOCATION_ALIGNMENT)
82  : upstream_{to_device_async_resource_ref_checked(upstream)},
83  allocation_limit_{allocation_limit},
84  allocated_bytes_(0),
85  alignment_(alignment)
86  {
87  }
88 
89  limiting_resource_adaptor() = delete;
90  ~limiting_resource_adaptor() override = default;
93  default;
94  limiting_resource_adaptor& operator=(limiting_resource_adaptor const&) = delete;
96  default;
97 
101  [[nodiscard]] device_async_resource_ref get_upstream_resource() const noexcept
102  {
103  return upstream_;
104  }
105 
115  [[nodiscard]] std::size_t get_allocated_bytes() const { return allocated_bytes_; }
116 
124  [[nodiscard]] std::size_t get_allocation_limit() const { return allocation_limit_; }
125 
126  private:
140  void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
141  {
142  auto const proposed_size = align_up(bytes, alignment_);
143  auto const old = allocated_bytes_.fetch_add(proposed_size);
144  if (old + proposed_size <= allocation_limit_) {
145  try {
146  return get_upstream_resource().allocate_async(bytes, stream);
147  } catch (...) {
148  allocated_bytes_ -= proposed_size;
149  throw;
150  }
151  }
152 
153  allocated_bytes_ -= proposed_size;
154  auto const msg = std::string("Exceeded memory limit (failed to allocate ") +
155  rmm::detail::format_bytes(bytes) + ")";
156  RMM_FAIL(msg.c_str(), rmm::out_of_memory);
157  }
158 
166  void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) override
167  {
168  std::size_t allocated_size = align_up(bytes, alignment_);
169  get_upstream_resource().deallocate_async(ptr, bytes, stream);
170  allocated_bytes_ -= allocated_size;
171  }
172 
180  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
181  {
182  if (this == &other) { return true; }
183  auto const* cast = dynamic_cast<limiting_resource_adaptor<Upstream> const*>(&other);
184  if (cast == nullptr) { return false; }
185  return get_upstream_resource() == cast->get_upstream_resource();
186  }
187 
188  // The upstream resource used for satisfying allocation requests
189  device_async_resource_ref upstream_;
190 
191  // maximum bytes this allocator is allowed to allocate.
192  std::size_t allocation_limit_;
193 
194  // number of currently-allocated bytes
195  std::atomic<std::size_t> allocated_bytes_;
196 
197  // todo: should be some way to ask the upstream...
198  std::size_t alignment_;
199 };
200  // end of group
202 } // namespace mr
203 } // namespace RMM_NAMESPACE
Strongly-typed non-owning wrapper for CUDA streams with default constructor.
Definition: cuda_stream_view.hpp:39
Base class for all librmm device memory allocation.
Definition: device_memory_resource.hpp:93
Resource that uses Upstream to allocate memory and limits the total allocations possible.
Definition: limiting_resource_adaptor.hpp:49
limiting_resource_adaptor(device_async_resource_ref upstream, std::size_t allocation_limit, std::size_t alignment=CUDA_ALLOCATION_ALIGNMENT)
Construct a new limiting resource adaptor using upstream to satisfy allocation requests and limiting ...
Definition: limiting_resource_adaptor.hpp:59
limiting_resource_adaptor(limiting_resource_adaptor &&) noexcept=default
Default move constructor.
std::size_t get_allocated_bytes() const
Query the number of bytes that have been allocated. Note that this can not be used to know how large ...
Definition: limiting_resource_adaptor.hpp:115
std::size_t get_allocation_limit() const
Query the maximum number of bytes that this allocator is allowed to allocate. This is the limit on th...
Definition: limiting_resource_adaptor.hpp:124
limiting_resource_adaptor(Upstream *upstream, std::size_t allocation_limit, std::size_t alignment=CUDA_ALLOCATION_ALIGNMENT)
Construct a new limiting resource adaptor using upstream to satisfy allocation requests and limiting ...
Definition: limiting_resource_adaptor.hpp:79
Exception thrown when RMM runs out of memory.
Definition: error.hpp:87
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 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.