Module aligned 

Aligned GPU memory allocation for optimal access patterns.

This module provides AlignedBuffer<T>, a device memory buffer that guarantees a specific alignment for the starting address. Proper alignment is critical for coalesced memory accesses on GPUs — misaligned loads and stores can incur extra memory transactions, significantly hurting throughput.

Alignment options

Variant	Bytes	Use case
Default	256	CUDA’s natural allocation alignment
Align256	256	Explicit 256-byte alignment
Align512	512	Optimal for many GPU texture/surface ops
Align1024	1024	Large-stride access patterns
Align4096	4096	Page-aligned for unified/mapped memory
Custom(n)	n	User-specified (must be a power of two)

Example

let buf = AlignedBuffer::<f32>::alloc(1024, Alignment::Align512)?;
assert!(buf.is_aligned());
assert_eq!(buf.as_device_ptr() % 512, 0);

Structs

AlignedBuffer

A device memory buffer whose starting address is guaranteed to meet the requested Alignment.

AlignmentInfo

Information about the alignment of an existing device pointer.

Enums

Alignment

Specifies the byte alignment for a device memory allocation.

Functions

check_alignment

Inspects a device pointer and reports its alignment characteristics.

coalesce_alignment

Computes the smallest alignment that ensures coalesced memory access for a given access_width (in bytes) across a warp of warp_size threads.

optimal_alignment_for_type

Recommends an optimal Alignment for a type based on its size.

round_up_to_alignment

Rounds bytes up to the next multiple of alignment.

validate_alignment

Validates that an Alignment is a power of two and within a reasonable range.