Mila::Dnn::Compute::MemoryResource Class Reference

Clean memory resource abstraction for device-specific memory allocation. More...

Inheritance diagram for Mila::Dnn::Compute::MemoryResource:

Collaboration diagram for Mila::Dnn::Compute::MemoryResource:

Public Member Functions
virtual	~MemoryResource ()=default
	Virtual destructor for proper cleanup of derived classes.

Static Public Attributes
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
virtual void *	do_allocate (std::size_t bytes, std::size_t alignment) override=0
	Allocates memory with specified size and alignment.
virtual void	do_deallocate (void *ptr, std::size_t bytes, std::size_t alignment) override=0
	Deallocates previously allocated memory.
virtual bool	do_is_equal (const std::pmr::memory_resource &other) const noexcept override=0
	Compares this memory resource with another for equality.

Detailed Description

Clean memory resource abstraction for device-specific memory allocation.

This class extends std::pmr::memory_resource with device-awareness while maintaining clear separation of concerns. Memory resources are responsible only for allocation, deallocation, and basic memory operations. Tensor-specific operations like type conversion and fill operations are handled by separate tensor operation classes.

Device-specific implementations (CPU, CUDA) override these methods with optimized implementations appropriate for their memory types and access patterns.

Constructor & Destructor Documentation

~MemoryResource()

virtual Mila::Dnn::Compute::MemoryResource::~MemoryResource ( )

virtualdefault

Virtual destructor for proper cleanup of derived classes.

Member Function Documentation

do_allocate()

virtual void * Mila::Dnn::Compute::MemoryResource::do_allocate ( std::size_t bytes, std::size_t alignment )

overrideprotectedpure virtual

Allocates memory with specified size and alignment.

Pure virtual function that must be implemented by derived classes to provide device-specific memory allocation. Implementation should handle device-specific allocation strategies and error conditions.

Parameters

bytes	Number of bytes to allocate
alignment	Memory alignment requirement

Returns

Pointer to allocated memory

Exceptions

std::bad_alloc

If allocation fails

Implemented in Mila::Dnn::Compute::CpuMemoryResource, Mila::Dnn::Compute::CudaDeviceMemoryResource, Mila::Dnn::Compute::CudaManagedMemoryResource, Mila::Dnn::Compute::CudaPinnedMemoryResource, Mila::Dnn::Compute::MetalMemoryResource, Mila::Dnn::Compute::TrackedMemoryResource, and Mila::Dnn::Compute::VulkanMemoryResource.

do_deallocate()

virtual void Mila::Dnn::Compute::MemoryResource::do_deallocate ( void * ptr, std::size_t bytes, std::size_t alignment )

overrideprotectedpure virtual

Deallocates previously allocated memory.

Pure virtual function that must be implemented by derived classes to provide device-specific memory deallocation. Implementation should handle proper cleanup and device synchronization if necessary.

Parameters

ptr	Pointer to memory to deallocate
bytes	Size of memory block (may be used for debugging/validation)
alignment	Alignment that was used during allocation

Implemented in Mila::Dnn::Compute::CpuMemoryResource, Mila::Dnn::Compute::CudaDeviceMemoryResource, Mila::Dnn::Compute::CudaManagedMemoryResource, Mila::Dnn::Compute::CudaPinnedMemoryResource, Mila::Dnn::Compute::MetalMemoryResource, Mila::Dnn::Compute::TrackedMemoryResource, and Mila::Dnn::Compute::VulkanMemoryResource.

do_is_equal()

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

overrideprotectedpure virtualnoexcept

Compares this memory resource with another for equality.

Two memory resources are equal if they can be used interchangeably for allocation and deallocation. Typically this means they are the same type and manage the same underlying memory pool or device.

Parameters

other

The other memory resource to compare with

Returns

true if the memory resources are equivalent

Implemented in Mila::Dnn::Compute::CpuMemoryResource, Mila::Dnn::Compute::CudaDeviceMemoryResource, Mila::Dnn::Compute::CudaManagedMemoryResource, Mila::Dnn::Compute::CudaPinnedMemoryResource, Mila::Dnn::Compute::MetalMemoryResource, Mila::Dnn::Compute::TrackedMemoryResource, and Mila::Dnn::Compute::VulkanMemoryResource.

Member Data Documentation

is_device_accessible

bool Mila::Dnn::Compute::MemoryResource::is_device_accessible = false

staticconstexpr

Checks if the memory is accessible from device code.

Derived classes should override this to reflect their accessibility characteristics.

is_host_accessible

bool Mila::Dnn::Compute::MemoryResource::is_host_accessible = true

staticconstexpr

Checks if the memory is accessible from host code.

Derived classes should override this to reflect their accessibility characteristics.

---

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

• /__w/Mila/Mila/Mila/Src/Dnn/Compute/MemoryResource.ixx