Implementation class for RmmResourceAdaptor. More...

#include <rmm_resource_adaptor_impl.hpp>

Public Member Functions
	RmmResourceAdaptorImpl (PrimaryMR primary_mr)
	Construct with a primary memory resource. More...

template<typename... Args>
	RmmResourceAdaptorImpl (std::in_place_t, Args &&... args)
	Construct the primary resource in-place from forwarded arguments. More...

	RmmResourceAdaptorImpl (RmmResourceAdaptorImpl const &)=delete

	RmmResourceAdaptorImpl (RmmResourceAdaptorImpl &&)=delete

RmmResourceAdaptorImpl &	operator= (RmmResourceAdaptorImpl const &)=delete

RmmResourceAdaptorImpl &	operator= (RmmResourceAdaptorImpl &&)=delete

bool	operator== (RmmResourceAdaptorImpl const &other) const noexcept
	Equality comparison. More...

PrimaryMR const &	get_upstream_resource () const noexcept
	Returns a reference to the primary upstream resource. More...

ScopedMemoryRecord	get_main_record () const
	Returns a copy of the main memory record. More...

std::int64_t	current_allocated () const noexcept
	Get the total current allocated memory through this resource. More...

void	begin_scoped_memory_record ()
	Begin recording a new scoped memory usage record for the current thread. More...

ScopedMemoryRecord	end_scoped_memory_record ()
	End the current scoped memory record and return it. More...

void *	allocate (cuda::stream_ref stream, std::size_t bytes, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT)
	Allocate memory asynchronously on the given stream. More...

void	deallocate (cuda::stream_ref stream, void *ptr, std::size_t bytes, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT) noexcept
	Deallocate memory asynchronously on the given stream. More...

void *	allocate_sync (std::size_t bytes, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT)
	Allocate memory synchronously. More...

void	deallocate_sync (void *ptr, std::size_t bytes, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT) noexcept
	Deallocate memory synchronously. More...

Friends
void	get_property (RmmResourceAdaptorImpl const &, cuda::mr::device_accessible) noexcept
	Tag this resource as device-accessible for the CCCL concept.

Detailed Description

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>
class rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >

Implementation class for RmmResourceAdaptor.

Holds all mutable state for memory tracking. This class satisfies the CCCL cuda::mr::resource concept and is held by RmmResourceAdaptor via cuda::mr::shared_resource for reference-counted ownership.

Template Parameters

PrimaryMR The type of the primary memory resource. Use a concrete resource type (e.g. cuda::pinned_memory_pool) to store the resource directly inside the shared control block, avoiding an extra heap allocation.

Definition at line 43 of file rmm_resource_adaptor_impl.hpp.

Constructor & Destructor Documentation

◆ RmmResourceAdaptorImpl() [1/2]

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::RmmResourceAdaptorImpl ( PrimaryMR primary_mr )

inlineexplicit

Construct with a primary memory resource.

Parameters

primary_mr The primary memory resource (moved in).

Definition at line 54 of file rmm_resource_adaptor_impl.hpp.

◆ RmmResourceAdaptorImpl() [2/2]

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

template<typename... Args>

rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::RmmResourceAdaptorImpl	(	std::in_place_t	,
		Args &&...	args
	)

inlineexplicit

Construct the primary resource in-place from forwarded arguments.

Use this overload when the primary resource type is non-movable (e.g. cuda::pinned_memory_pool). The resource is constructed directly inside the shared-ownership control block, avoiding an extra heap allocation.

Parameters

args	Arguments forwarded to the `PrimaryMR` constructor.

Definition at line 71 of file rmm_resource_adaptor_impl.hpp.

Member Function Documentation

◆ allocate()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

void* rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::allocate	(	cuda::stream_ref	stream,
		std::size_t	bytes,
		std::size_t	alignment = `rmm::CUDA_ALLOCATION_ALIGNMENT`
	)

inline

Allocate memory asynchronously on the given stream.

Parameters

stream	The CUDA stream for the allocation.
bytes	Number of bytes to allocate.
alignment	Alignment requirement.

Returns: Pointer to the allocated memory.

Definition at line 142 of file rmm_resource_adaptor_impl.hpp.

◆ allocate_sync()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

void* rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::allocate_sync	(	std::size_t	bytes,
		std::size_t	alignment = `rmm::CUDA_ALLOCATION_ALIGNMENT`
	)

inline

Allocate memory synchronously.

Parameters

bytes	Number of bytes to allocate.
alignment	Alignment requirement.

Returns: Pointer to the allocated memory.

Definition at line 202 of file rmm_resource_adaptor_impl.hpp.

◆ begin_scoped_memory_record()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

void rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::begin_scoped_memory_record ( )

inline

Begin recording a new scoped memory usage record for the current thread.

This method pushes a new empty ScopedMemoryRecord onto the thread-local record stack, allowing for nested memory tracking scopes.

Must be paired with a matching call to end_scoped_memory_record().

See also: end_scoped_memory_record()

Definition at line 112 of file rmm_resource_adaptor_impl.hpp.

◆ current_allocated()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

std::int64_t rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::current_allocated ( ) const

inlinenoexcept

Get the total current allocated memory through this resource.

Returns: Total number of currently allocated bytes.

Definition at line 106 of file rmm_resource_adaptor_impl.hpp.

◆ deallocate()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

void rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::deallocate	(	cuda::stream_ref	stream,
		void *	ptr,
		std::size_t	bytes,
		std::size_t	alignment = `rmm::CUDA_ALLOCATION_ALIGNMENT`
	)

inlinenoexcept

Deallocate memory asynchronously on the given stream.

Parameters

stream	The CUDA stream for the deallocation.
ptr	Pointer to the memory to deallocate.
bytes	Number of bytes to deallocate.
alignment	Alignment of the original allocation.

Definition at line 172 of file rmm_resource_adaptor_impl.hpp.

◆ deallocate_sync()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

void rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::deallocate_sync	(	void *	ptr,
		std::size_t	bytes,
		std::size_t	alignment = `rmm::CUDA_ALLOCATION_ALIGNMENT`
	)

inlinenoexcept

Deallocate memory synchronously.

Parameters

ptr	Pointer to the memory to deallocate.
bytes	Number of bytes to deallocate.
alignment	Alignment of the original allocation.

Definition at line 217 of file rmm_resource_adaptor_impl.hpp.

◆ end_scoped_memory_record()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

ScopedMemoryRecord rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::end_scoped_memory_record ( )

inline

End the current scoped memory record and return it.

Pops the top ScopedMemoryRecord from the thread-local stack and returns it. If this scope was nested within another (i.e. if begin_scoped_memory_record() was called multiple times in a row), the returned scope is automatically added as a subscope to the next scope remaining on the stack.

This allows nesting of scoped memory tracking, where each scope can contain one or more subscopes. When analyzing or reporting memory statistics, the memory usage of each scope can be calculated inclusive of its subscopes. This behavior mimics standard hierarchical memory profilers, where the total memory attributed to a scope includes all allocations made within it, plus those made in its nested regions.

Returns: The scope that was just ended.

Exceptions

std::out_of_range if called without a matching begin_scoped_memory_record().

See also: begin_scoped_memory_record()

Definition at line 118 of file rmm_resource_adaptor_impl.hpp.

◆ get_main_record()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

ScopedMemoryRecord rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::get_main_record ( ) const

inline

Returns a copy of the main memory record.

The main record tracks memory statistics for the lifetime of the resource.

Returns: A copy of the current main memory record.

Definition at line 100 of file rmm_resource_adaptor_impl.hpp.

◆ get_upstream_resource()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

PrimaryMR const& rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::get_upstream_resource ( ) const

inlinenoexcept

Returns a reference to the primary upstream resource.

Returns: Reference to the primary resource.

Definition at line 95 of file rmm_resource_adaptor_impl.hpp.

◆ operator==()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>

bool rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >::operator== ( RmmResourceAdaptorImpl< PrimaryMR > const & other ) const

inlinenoexcept

Equality comparison.

Parameters

other The other impl to compare.

Returns: True if the two instances are the same.

Definition at line 87 of file rmm_resource_adaptor_impl.hpp.

The documentation for this class was generated from the following file:

/__w/rapidsmpf/rapidsmpf/cpp/include/rapidsmpf/detail/rmm_resource_adaptor_impl.hpp

Public Member Functions

Friends

Detailed Description

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR> class rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >

Constructor & Destructor Documentation

◆ RmmResourceAdaptorImpl() [1/2]

◆ RmmResourceAdaptorImpl() [2/2]

Member Function Documentation

◆ allocate()

◆ allocate_sync()

◆ begin_scoped_memory_record()

◆ current_allocated()

◆ deallocate()

◆ deallocate_sync()

◆ end_scoped_memory_record()

◆ get_main_record()

◆ get_upstream_resource()

◆ operator==()

template<cuda::mr::resource_with< cuda::mr::device_accessible > PrimaryMR>
class rapidsmpf::detail::RmmResourceAdaptorImpl< PrimaryMR >