fixed_size_memory_resource.hpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2020-2021, 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/cuda_stream_view.hpp>
19 #include <rmm/detail/aligned.hpp>
20 #include <rmm/detail/error.hpp>
21 #include <rmm/detail/logging_assert.hpp>
22 #include <rmm/mr/device/detail/fixed_size_free_list.hpp>
23 #include <rmm/mr/device/detail/stream_ordered_memory_resource.hpp>
24 
25 #include <rmm/detail/thrust_namespace.h>
26 #include <thrust/iterator/counting_iterator.h>
27 #include <thrust/iterator/transform_iterator.h>
28 
29 #include <cuda_runtime_api.h>
30 
31 #include <algorithm>
32 #include <cstddef>
33 #include <list>
34 #include <map>
35 #include <utility>
36 #include <vector>
37 
38 namespace rmm::mr {
50 template <typename Upstream>
52  : public detail::stream_ordered_memory_resource<fixed_size_memory_resource<Upstream>,
53  detail::fixed_size_free_list> {
54  public:
55  friend class detail::stream_ordered_memory_resource<fixed_size_memory_resource<Upstream>,
56  detail::fixed_size_free_list>;
57 
58  static constexpr std::size_t default_block_size = 1 << 20;
59 
62  static constexpr std::size_t default_blocks_to_preallocate = 128;
63 
75  explicit fixed_size_memory_resource(
76  Upstream* upstream_mr,
77  std::size_t block_size = default_block_size,
78  std::size_t blocks_to_preallocate = default_blocks_to_preallocate)
79  : upstream_mr_{upstream_mr},
80  block_size_{rmm::detail::align_up(block_size, rmm::detail::CUDA_ALLOCATION_ALIGNMENT)},
81  upstream_chunk_size_{block_size * blocks_to_preallocate}
82  {
83  // allocate initial blocks and insert into free list
84  this->insert_blocks(std::move(blocks_from_upstream(cuda_stream_legacy)), cuda_stream_legacy);
85  }
86 
91  ~fixed_size_memory_resource() override { release(); }
92 
93  fixed_size_memory_resource() = delete;
94  fixed_size_memory_resource(fixed_size_memory_resource const&) = delete;
95  fixed_size_memory_resource(fixed_size_memory_resource&&) = delete;
96  fixed_size_memory_resource& operator=(fixed_size_memory_resource const&) = delete;
97  fixed_size_memory_resource& operator=(fixed_size_memory_resource&&) = delete;
98 
105  [[nodiscard]] bool supports_streams() const noexcept override { return true; }
106 
112  [[nodiscard]] bool supports_get_mem_info() const noexcept override { return false; }
113 
119  Upstream* get_upstream() const noexcept { return upstream_mr_; }
120 
126  [[nodiscard]] std::size_t get_block_size() const noexcept { return block_size_; }
127 
128  protected:
129  using free_list = detail::fixed_size_free_list;
130  using block_type = free_list::block_type;
131  using typename detail::stream_ordered_memory_resource<fixed_size_memory_resource<Upstream>,
132  detail::fixed_size_free_list>::split_block;
133  using lock_guard = std::lock_guard<std::mutex>;
134 
141  [[nodiscard]] std::size_t get_maximum_allocation_size() const { return get_block_size(); }
142 
154  block_type expand_pool(std::size_t size, free_list& blocks, cuda_stream_view stream)
155  {
156  blocks.insert(std::move(blocks_from_upstream(stream)));
157  return blocks.get_block(size);
158  }
159 
166  free_list blocks_from_upstream(cuda_stream_view stream)
167  {
168  void* ptr = get_upstream()->allocate(upstream_chunk_size_, stream);
169  block_type block{ptr};
170  upstream_blocks_.push_back(block);
171 
172  auto num_blocks = upstream_chunk_size_ / block_size_;
173 
174  auto block_gen = [ptr, this](int index) {
175  // NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
176  return block_type{static_cast<char*>(ptr) + index * block_size_};
177  };
178  auto first =
179  thrust::make_transform_iterator(thrust::make_counting_iterator(std::size_t{0}), block_gen);
180  return free_list(first, first + num_blocks);
181  }
182 
193  split_block allocate_from_block(block_type const& block, std::size_t size)
194  {
195  return {block, block_type{nullptr}};
196  }
197 
206  block_type free_block(void* ptr, std::size_t size) noexcept
207  {
208  // Deallocating a fixed-size block just inserts it in the free list, which is
209  // handled by the parent class
210  RMM_LOGGING_ASSERT(rmm::detail::align_up(size, rmm::detail::CUDA_ALLOCATION_ALIGNMENT) <=
211  block_size_);
212  return block_type{ptr};
213  }
214 
223  [[nodiscard]] std::pair<std::size_t, std::size_t> do_get_mem_info(
224  [[maybe_unused]] cuda_stream_view stream) const override
225  {
226  return std::make_pair(0, 0);
227  }
228 
233  void release()
234  {
235  lock_guard lock(this->get_mutex());
236 
237  for (auto block : upstream_blocks_) {
238  get_upstream()->deallocate(block.pointer(), upstream_chunk_size_);
239  }
240  upstream_blocks_.clear();
241  }
242 
243 #ifdef RMM_DEBUG_PRINT
244  void print()
245  {
246  lock_guard lock(this->get_mutex());
247 
248  auto const [free, total] = get_upstream()->get_mem_info(rmm::cuda_stream_default);
249  std::cout << "GPU free memory: " << free << " total: " << total << "\n";
250 
251  std::cout << "upstream_blocks: " << upstream_blocks_.size() << "\n";
252  std::size_t upstream_total{0};
253 
254  for (auto blocks : upstream_blocks_) {
255  blocks.print();
256  upstream_total += upstream_chunk_size_;
257  }
258  std::cout << "total upstream: " << upstream_total << " B\n";
259 
260  this->print_free_blocks();
261  }
262 #endif
263 
272  std::pair<std::size_t, std::size_t> free_list_summary(free_list const& blocks)
273  {
274  return blocks.is_empty() ? std::make_pair(std::size_t{0}, std::size_t{0})
275  : std::make_pair(block_size_, blocks.size() * block_size_);
276  }
277 
278  private:
279  Upstream* upstream_mr_; // The resource from which to allocate new blocks
280 
281  std::size_t const block_size_; // size of blocks this MR allocates
282  std::size_t const upstream_chunk_size_; // size of chunks allocated from heap MR
283 
284  // blocks allocated from heap: so they can be easily freed
285  std::vector<block_type> upstream_blocks_;
286 };
287  // end of group
289 } // namespace rmm::mr
Strongly-typed non-owning wrapper for CUDA streams with default constructor.
Definition: cuda_stream_view.hpp:41
A device_memory_resource which allocates memory blocks of a single fixed size.
Definition: fixed_size_memory_resource.hpp:53
std::pair< std::size_t, std::size_t > free_list_summary(free_list const &blocks)
Get the largest available block size and total free size in the specified free list.
Definition: fixed_size_memory_resource.hpp:272
static constexpr std::size_t default_block_size
Default allocation block size.
Definition: fixed_size_memory_resource.hpp:58
std::pair< std::size_t, std::size_t > do_get_mem_info([[maybe_unused]] cuda_stream_view stream) const override
Get free and available memory for memory resource.
Definition: fixed_size_memory_resource.hpp:223
detail::fixed_size_free_list free_list
The free list type.
Definition: fixed_size_memory_resource.hpp:129
block_type free_block(void *ptr, std::size_t size) noexcept
Finds, frees and returns the block associated with pointer.
Definition: fixed_size_memory_resource.hpp:206
free_list::block_type block_type
The type of block managed by the free list.
Definition: fixed_size_memory_resource.hpp:130
std::lock_guard< std::mutex > lock_guard
Type of lock used to synchronize access.
Definition: fixed_size_memory_resource.hpp:133
void release()
free all memory allocated using the upstream resource.
Definition: fixed_size_memory_resource.hpp:233
split_block allocate_from_block(block_type const &block, std::size_t size)
Splits block if necessary to return a pointer to memory of size bytes.
Definition: fixed_size_memory_resource.hpp:193