DeviceResource Class Reference

Default stream-ordered device memory resource. Allocations are made with cudaMallocAsync and freed with cudaFreeAsync via the util::cuda wrappers. More...

#include <nanovdb/cuda/DeviceResource.h>

Public Member Functions
void *	allocate_async (size_t bytes, size_t, cudaStream_t stream)
	Stream-ordered allocation.
void	deallocate_async (void *p, size_t, size_t, cudaStream_t stream)
	Stream-ordered deallocation.
void *	allocate (size_t bytes, size_t alignment)
	Synchronous allocation; the returned memory is immediately valid on every stream.
void	deallocate (void *p, size_t bytes, size_t alignment)
	Synchronous deallocation; the caller guarantees that device work touching the memory has completed.

Static Public Attributes
static constexpr size_t	DEFAULT_ALIGNMENT = 256

Friends
bool	operator== (const DeviceResource &, const DeviceResource &)
	Stateless: memory allocated through one instance may be deallocated through any other, so every two instances compare equal (the equality cuda::mr's resource concepts require).
bool	operator!= (const DeviceResource &, const DeviceResource &)

Detailed Description

Default stream-ordered device memory resource. Allocations are made with cudaMallocAsync and freed with cudaFreeAsync via the util::cuda wrappers.

Note

Models the AsyncResource concept, which refines the synchronous Resource concept (as in CCCL's cuda::mr): the instance methods provide both the stream-ordered allocate_async / deallocate_async and the synchronous allocate / deallocate. The type is stateless, so a default-constructed instance adds no overhead.

Member Function Documentation

allocate()

void * allocate ( size_t bytes, size_t alignment )

inline

Synchronous allocation; the returned memory is immediately valid on every stream.

Parameters

bytes	number of bytes to allocate
alignment	requested alignment

allocate_async()

void * allocate_async ( size_t bytes, size_t , cudaStream_t stream )

inline

Stream-ordered allocation.

Parameters

bytes	number of bytes to allocate
stream	cuda stream the allocation is ordered on

Note

the alignment parameter is unnamed: cudaMallocAsync always 256B-aligns

deallocate()

void deallocate ( void * p, size_t bytes, size_t alignment )

inline

Synchronous deallocation; the caller guarantees that device work touching the memory has completed.

Parameters

p	pointer previously returned by allocate or allocate_async
bytes	size of the allocation in bytes
alignment	alignment of the allocation in bytes

deallocate_async()

void deallocate_async ( void * p, size_t , size_t , cudaStream_t stream )

inline

Stream-ordered deallocation.

Parameters

p	pointer previously returned by allocate_async
stream	cuda stream the deallocation is ordered on

Friends And Related Symbol Documentation

operator!=

bool operator!= ( const DeviceResource & , const DeviceResource &  )

friend

operator==

bool operator== ( const DeviceResource & , const DeviceResource &  )

friend

Stateless: memory allocated through one instance may be deallocated through any other, so every two instances compare equal (the equality cuda::mr's resource concepts require).

Member Data Documentation

DEFAULT_ALIGNMENT

size_t DEFAULT_ALIGNMENT = 256

staticconstexpr