20 #include <rmm/detail/error.hpp>
21 #include <rmm/detail/export.hpp>
25 #include <cuda/memory_resource>
26 #include <cuda_runtime_api.h>
32 namespace RMM_NAMESPACE {
111 : _stream{stream}, _mr{mr}
114 allocate_async(size);
140 : _stream{stream}, _mr{mr}
143 allocate_async(size);
144 copy_async(source_data, size);
187 : _data{other._data},
189 _capacity{other._capacity},
192 _device{other._device}
194 other._data =
nullptr;
198 other._device = cuda_device_id{-1};
217 if (&other !=
this) {
223 _capacity = other._capacity;
224 set_stream(other.
stream());
226 _device = other._device;
228 other._data =
nullptr;
272 if (new_capacity > capacity()) {
275 auto const old_size = size();
276 RMM_CUDA_TRY(cudaMemcpyAsync(tmp.data(), data(), size(), cudaMemcpyDefault, stream.
value()));
277 *
this = std::move(tmp);
312 if (new_size <= capacity()) {
317 RMM_CUDA_TRY(cudaMemcpyAsync(tmp.data(), data(), size(), cudaMemcpyDefault, stream.
value()));
318 *
this = std::move(tmp);
338 if (size() != capacity()) {
344 std::swap(tmp, *
this);
351 [[nodiscard]]
void const*
data() const noexcept {
return _data; }
356 void*
data() noexcept {
return _data; }
361 [[nodiscard]] std::size_t
size() const noexcept {
return _size; }
366 [[nodiscard]] std::int64_t
ssize() const noexcept
368 assert(size() <
static_cast<std::size_t
>(std::numeric_limits<int64_t>::max()) &&
369 "Size overflows signed integer");
370 return static_cast<int64_t
>(size());
379 [[nodiscard]]
bool is_empty() const noexcept {
return 0 == size(); }
388 [[nodiscard]] std::size_t
capacity() const noexcept {
return _capacity; }
414 void* _data{
nullptr};
416 std::size_t _capacity{};
417 cuda_stream_view _stream{};
433 void allocate_async(std::size_t bytes)
437 _data = (bytes > 0) ? _mr.allocate_async(bytes, stream()) :
nullptr;
449 void deallocate_async() noexcept
451 if (capacity() > 0) { _mr.deallocate_async(data(), capacity(), stream()); }
469 void copy_async(
void const* source, std::size_t bytes)
472 RMM_EXPECTS(
nullptr != source,
"Invalid copy from nullptr.");
473 RMM_EXPECTS(
nullptr != _data,
"Invalid copy to nullptr.");
475 RMM_CUDA_TRY(cudaMemcpyAsync(_data, source, bytes, cudaMemcpyDefault, stream().value()));
Strongly-typed non-owning wrapper for CUDA streams with default constructor.
Definition: cuda_stream_view.hpp:39
constexpr cudaStream_t value() const noexcept
Get the wrapped stream.
Definition: cuda_stream_view.hpp:73
RAII construct for device memory allocation.
Definition: device_buffer.hpp:83
cuda_stream_view stream() const noexcept
The stream most recently specified for allocation/deallocation.
Definition: device_buffer.hpp:393
void resize(std::size_t new_size, cuda_stream_view stream)
Resize the device memory allocation.
Definition: device_buffer.hpp:307
void * data() noexcept
Pointer to the device memory allocation.
Definition: device_buffer.hpp:356
~device_buffer() noexcept
Destroy the device buffer object.
Definition: device_buffer.hpp:244
device_buffer & operator=(device_buffer &&other) noexcept
Move assignment operator moves the contents from other.
Definition: device_buffer.hpp:215
device_buffer()
Default constructor creates an empty device_buffer
Definition: device_buffer.hpp:96
std::size_t capacity() const noexcept
Returns actual size in bytes of device memory allocation.
Definition: device_buffer.hpp:388
void const * data() const noexcept
Const pointer to the device memory allocation.
Definition: device_buffer.hpp:351
void reserve(std::size_t new_capacity, cuda_stream_view stream)
Increase the capacity of the device memory allocation.
Definition: device_buffer.hpp:269
device_buffer(std::size_t size, cuda_stream_view stream, device_async_resource_ref mr=mr::get_current_device_resource_ref())
Constructs a new device buffer of size uninitialized bytes.
Definition: device_buffer.hpp:108
void set_stream(cuda_stream_view stream) noexcept
Sets the stream to be used for deallocation.
Definition: device_buffer.hpp:406
std::size_t size() const noexcept
The number of bytes.
Definition: device_buffer.hpp:361
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())
Construct a new device buffer by copying from a raw pointer to an existing host or device memory allo...
Definition: device_buffer.hpp:136
device_buffer(device_buffer &&other) noexcept
Constructs a new device_buffer by moving the contents of another device_buffer into the newly constru...
Definition: device_buffer.hpp:186
device_buffer(device_buffer const &other, cuda_stream_view stream, device_async_resource_ref mr=mr::get_current_device_resource_ref())
Construct a new device_buffer by deep copying the contents of another device_buffer,...
Definition: device_buffer.hpp:168
void shrink_to_fit(cuda_stream_view stream)
Forces the deallocation of unused memory.
Definition: device_buffer.hpp:335
std::int64_t ssize() const noexcept
The signed number of bytes.
Definition: device_buffer.hpp:366
bool is_empty() const noexcept
Whether or not the buffer currently holds any data.
Definition: device_buffer.hpp:379
rmm::device_async_resource_ref memory_resource() const noexcept
The resource used to allocate and deallocate.
Definition: device_buffer.hpp:411
cuda_device_id get_current_cuda_device()
Returns a cuda_device_id for the current device.
Definition: cuda_device.hpp:99
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 get_current_device_resource_ref()
Get the device_async_resource_ref for the current device.
Definition: per_device_resource.hpp:411
Management of per-device device_memory_resources.
Strong type for a CUDA device identifier.
Definition: cuda_device.hpp:41
RAII class that sets the current CUDA device to the specified device on construction and restores the...
Definition: cuda_device.hpp:151