Mapped (zero-copy) pinned host memory resource: page-locked host memory also mapped into the device address space (GPU::memAllocHostMapped / GPU::memFreeHost). More...

#include <EBGeometry_MemoryResource.hpp>

Inheritance diagram for EBGeometry::MappedMemoryResource:

[legend]

Collaboration diagram for EBGeometry::MappedMemoryResource:

[legend]

Public Member Functions
	MappedMemoryResource ()=default
	Default constructor.

	~MappedMemoryResource () override=default
	Destructor.

void *	allocate (size_t a_bytes, size_t a_alignment) override
	Allocate a block of mapped (zero-copy) pinned host memory.

void	deallocate (void *a_ptr, size_t a_bytes, size_t a_alignment) noexcept override
	Free a block previously returned by allocate.

bool	isHostAccessible () const noexcept override
	Mapped blocks are host-accessible.

bool	isDeviceAccessible () const noexcept override
	Mapped blocks are device-accessible.

void	copy (void a_dst, const MemoryResource &a_dstResource, const void a_src, const MemoryResource &a_srcResource, size_t a_bytes) const noexcept override
	Copy between host and device memory with the GPU runtime; see MemoryResource::copy.

Public Member Functions inherited from EBGeometry::MemoryResource
	MemoryResource ()=default
	Default constructor.

virtual	~MemoryResource ()=default
	Destructor.

	MemoryResource (const MemoryResource &)=delete
	Copy constructor. Deleted: a resource's identity is meaningful and non-copyable.

MemoryResource &	operator= (const MemoryResource &)=delete
	Copy assignment. Deleted: a resource's identity is meaningful and non-copyable.

Detailed Description

Mapped (zero-copy) pinned host memory resource: page-locked host memory also mapped into the device address space (GPU::memAllocHostMapped / GPU::memFreeHost).

Compiled only under an offload backend (CUDA/HIP) or for documentation. Blocks are both host- and device-accessible: the same base pointer is dereferenced by the CPU and by a kernel, which reaches it over the host/device interconnect on each access. Unlike ManagedMemoryResource there is no page migration – the block stays in host RAM and the device reads it in place, which suits small or write-once/streamed data rather than hot, reused data. Requires unified virtual addressing (all 64-bit CUDA; HIP under HSA) so the host pointer is directly valid on the device.

Member Function Documentation

◆ allocate()

void * EBGeometry::MappedMemoryResource::allocate	(	size_t	a_bytes,
		size_t	a_alignment
	)

overridevirtual

Allocate a block of mapped (zero-copy) pinned host memory.

Parameters

[in]	a_bytes	Number of bytes to allocate.
[in]	a_alignment	Required alignment (validated to be <= PoolBaseAlign).

Returns: Host pointer to the mapped block (never null on success; aborts on failure).

Implements EBGeometry::MemoryResource.

◆ copy()

void EBGeometry::MappedMemoryResource::copy	(	void *	a_dst,
		const MemoryResource &	a_dstResource,
		const void *	a_src,
		const MemoryResource &	a_srcResource,
		size_t	a_bytes
	)		const

overridevirtualnoexcept

Copy between host and device memory with the GPU runtime; see MemoryResource::copy.

Host-to-host is a plain copy; device-to-device aborts, in every build.

Parameters

[out]	a_dst	Destination block.
[in]	a_dstResource	Resource that allocated a_dst.
[in]	a_src	Source block.
[in]	a_srcResource	Resource that allocated a_src.
[in]	a_bytes	Number of bytes to copy.