device_buffer.hpp

Go to the documentation of this file.

/*
 * Copyright (c) 2019-2024, 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/cuda_device.hpp>
#include <rmm/cuda_stream_view.hpp>
#include <rmm/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/mr/device/per_device_resource.hpp>
#include <rmm/resource_ref.hpp>
 
#include <cuda/memory_resource>
#include <cuda_runtime_api.h>
 
#include <cassert>
#include <cstddef>
#include <stdexcept>
#include <utility>
 
namespace RMM_NAMESPACE {
class device_buffer {
 public:
  // The copy constructor and copy assignment operator without a stream are deleted because they
  // provide no way to specify an explicit stream
  device_buffer(device_buffer const& other)            = delete;
  device_buffer& operator=(device_buffer const& other) = delete;
 
  // Note: we cannot use `device_buffer() = default;` because nvcc implicitly adds
  // `__host__ __device__` specifiers to the defaulted constructor when it is called within the
  // context of both host and device functions. Specifically, the `cudf::type_dispatcher` is a host-
  // device function. This causes warnings/errors because this ctor invokes host-only functions.
  device_buffer() : _mr{rmm::mr::get_current_device_resource_ref()} {}
 
  explicit device_buffer(std::size_t size,
                         cuda_stream_view stream,
                         device_async_resource_ref mr = mr::get_current_device_resource_ref())
    : _stream{stream}, _mr{mr}
  {
    cuda_set_device_raii dev{_device};
    allocate_async(size);
  }
 
  device_buffer(void const* source_data,
                std::size_t size,
                cuda_stream_view stream,
                device_async_resource_ref mr = mr::get_current_device_resource_ref())
    : _stream{stream}, _mr{mr}
  {
    cuda_set_device_raii dev{_device};
    allocate_async(size);
    copy_async(source_data, size);
  }
 
  device_buffer(device_buffer const& other,
                cuda_stream_view stream,
                device_async_resource_ref mr = mr::get_current_device_resource_ref())
    : device_buffer{other.data(), other.size(), stream, mr}
  {
  }
 
  device_buffer(device_buffer&& other) noexcept
    : _data{other._data},
      _size{other._size},
      _capacity{other._capacity},
      _stream{other.stream()},
      _mr{other._mr},
      _device{other._device}
  {
    other._data     = nullptr;
    other._size     = 0;
    other._capacity = 0;
    other.set_stream(cuda_stream_view{});
    other._device = cuda_device_id{-1};
  }
 
  device_buffer& operator=(device_buffer&& other) noexcept
  {
    if (&other != this) {
      cuda_set_device_raii dev{_device};
      deallocate_async();
 
      _data     = other._data;
      _size     = other._size;
      _capacity = other._capacity;
      set_stream(other.stream());
      _mr     = other._mr;
      _device = other._device;
 
      other._data     = nullptr;
      other._size     = 0;
      other._capacity = 0;
      other.set_stream(cuda_stream_view{});
      other._device = cuda_device_id{-1};
    }
    return *this;
  }
 
  ~device_buffer() noexcept
  {
    cuda_set_device_raii dev{_device};
    deallocate_async();
    _stream = cuda_stream_view{};
  }
 
  void reserve(std::size_t new_capacity, cuda_stream_view stream)
  {
    set_stream(stream);
    if (new_capacity > capacity()) {
      cuda_set_device_raii dev{_device};
      auto tmp            = device_buffer{new_capacity, stream, _mr};
      auto const old_size = size();
      RMM_CUDA_TRY(cudaMemcpyAsync(tmp.data(), data(), size(), cudaMemcpyDefault, stream.value()));
      *this = std::move(tmp);
      _size = old_size;
    }
  }
 
  void resize(std::size_t new_size, cuda_stream_view stream)
  {
    set_stream(stream);
    // If the requested size is smaller than the current capacity, just update
    // the size without any allocations
    if (new_size <= capacity()) {
      _size = new_size;
    } else {
      cuda_set_device_raii dev{_device};
      auto tmp = device_buffer{new_size, stream, _mr};
      RMM_CUDA_TRY(cudaMemcpyAsync(tmp.data(), data(), size(), cudaMemcpyDefault, stream.value()));
      *this = std::move(tmp);
    }
  }
 
  void shrink_to_fit(cuda_stream_view stream)
  {
    set_stream(stream);
    if (size() != capacity()) {
      cuda_set_device_raii dev{_device};
      // Invoke copy ctor on self which only copies `[0, size())` and swap it
      // with self. The temporary `device_buffer` will hold the old contents
      // which will then be destroyed
      auto tmp = device_buffer{*this, stream, _mr};
      std::swap(tmp, *this);
    }
  }
 
  [[nodiscard]] void const* data() const noexcept { return _data; }
 
  void* data() noexcept { return _data; }
 
  [[nodiscard]] std::size_t size() const noexcept { return _size; }
 
  [[nodiscard]] std::int64_t ssize() const noexcept
  {
    assert(size() < static_cast<std::size_t>(std::numeric_limits<int64_t>::max()) &&
           "Size overflows signed integer");
    return static_cast<int64_t>(size());
  }
 
  [[nodiscard]] bool is_empty() const noexcept { return 0 == size(); }
 
  [[nodiscard]] std::size_t capacity() const noexcept { return _capacity; }
 
  [[nodiscard]] cuda_stream_view stream() const noexcept { return _stream; }
 
  void set_stream(cuda_stream_view stream) noexcept { _stream = stream; }
 
  [[nodiscard]] rmm::device_async_resource_ref memory_resource() const noexcept { return _mr; }
 
 private:
  void* _data{nullptr};        
  std::size_t _size{};         
  std::size_t _capacity{};     
  cuda_stream_view _stream{};  
 
  rmm::device_async_resource_ref _mr{
    rmm::mr::get_current_device_resource_ref()};  
  cuda_device_id _device{get_current_cuda_device()};
 
  void allocate_async(std::size_t bytes)
  {
    _size     = bytes;
    _capacity = bytes;
    _data     = (bytes > 0) ? _mr.allocate_async(bytes, stream()) : nullptr;
  }
 
  void deallocate_async() noexcept
  {
    if (capacity() > 0) { _mr.deallocate_async(data(), capacity(), stream()); }
    _size     = 0;
    _capacity = 0;
    _data     = nullptr;
  }
 
  void copy_async(void const* source, std::size_t bytes)
  {
    if (bytes > 0) {
      RMM_EXPECTS(nullptr != source, "Invalid copy from nullptr.");
      RMM_EXPECTS(nullptr != _data, "Invalid copy to nullptr.");
 
      RMM_CUDA_TRY(cudaMemcpyAsync(_data, source, bytes, cudaMemcpyDefault, stream().value()));
    }
  }
};
  // end of group
}  // namespace RMM_NAMESPACE