Mila::Dnn::Compute::CudaPinnedMemoryResource Class Reference

CUDA pinned memory resource for fast host/device transfer memory. More...

Inheritance diagram for Mila::Dnn::Compute::CudaPinnedMemoryResource:

Public Member Functions
	CudaPinnedMemoryResource (int device_id)
	Constructs CUDA pinned memory resource with device ID.
Public Member Functions inherited from Mila::Dnn::Compute::MemoryResource
virtual	~MemoryResource ()=default
	Virtual destructor for proper cleanup of derived classes.

Static Public Attributes
static constexpr DeviceType	device_type = DeviceType::Cuda
static constexpr bool	is_device_accessible = DeviceAccessible::is_device_accessible
	Indicates pinned memory is accessible from device code.
static constexpr bool	is_host_accessible = HostAccessible::is_host_accessible
	Indicates pinned memory is accessible from host code.
Static Public Attributes inherited from Mila::Dnn::Compute::MemoryResource
static constexpr bool	is_device_accessible = false
	Checks if the memory is accessible from device code.
static constexpr bool	is_host_accessible = true
	Checks if the memory is accessible from host code.

Protected Member Functions
void *	do_allocate (std::size_t bytes, std::size_t=alignof(std::max_align_t)) override
	Allocates CUDA pinned memory.
void	do_deallocate (void *ptr, std::size_t, std::size_t=alignof(std::max_align_t)) override
	Deallocates CUDA pinned memory.
bool	do_is_equal (const std::pmr::memory_resource &other) const noexcept override
	Compares pinned memory resources for equality.

Detailed Description

CUDA pinned memory resource for fast host/device transfer memory.

Provides CUDA pinned (page-locked) memory allocation that is accessible from host code but provides faster transfers to/from GPU devices. Focuses purely on memory allocation responsibilities without tensor-specific operations.

Constructor & Destructor Documentation

CudaPinnedMemoryResource()

Mila::Dnn::Compute::CudaPinnedMemoryResource::CudaPinnedMemoryResource ( int device_id )

inlineexplicit

Constructs CUDA pinned memory resource with device ID.

Parameters

device_id

CUDA device ID (0, 1, 2, etc.)

Exceptions

std::invalid_argument

If device_id is invalid

Member Function Documentation

do_allocate()

void * Mila::Dnn::Compute::CudaPinnedMemoryResource::do_allocate ( std::size_t bytes, std::size_t = alignof(std::max_align_t) )

inlineoverrideprotectedvirtual

Allocates CUDA pinned memory.

Uses cudaMallocHost to allocate page-locked memory that provides faster transfers between host and device while remaining host-accessible.

Parameters

bytes

Number of bytes to allocate The alignment argument is ignored: CUDA allocations are already suitably aligned.

Returns

Pointer to allocated pinned memory

Exceptions

std::bad_alloc

If allocation fails

Implements Mila::Dnn::Compute::MemoryResource.

do_deallocate()

void Mila::Dnn::Compute::CudaPinnedMemoryResource::do_deallocate ( void * ptr, std::size_t , std::size_t = alignof(std::max_align_t) )

inlineoverrideprotectedvirtual

Deallocates CUDA pinned memory.

Uses cudaFreeHost to properly release page-locked memory. Ensures operation occurs on the correct device.

Parameters

ptr	Pointer to pinned memory to deallocate The alignment argument is unused, and kept for interface compatibility.

The size argument is unused and therefore intentionally unnamed.

Implements Mila::Dnn::Compute::MemoryResource.

do_is_equal()

bool Mila::Dnn::Compute::CudaPinnedMemoryResource::do_is_equal ( const std::pmr::memory_resource & other ) const

inlineoverrideprotectedvirtualnoexcept

Compares pinned memory resources for equality.

Pinned memory resources are equal if they are both CudaPinnedMemoryResource instances with the same device ID.

Parameters

other

The other memory resource to compare with

Returns

true if both are CudaPinnedMemoryResource with same device ID

Implements Mila::Dnn::Compute::MemoryResource.

---

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

• Mila/Src/Dnn/Compute/Devices/Cuda/CudaPinnedMemoryResource.ixx