[BUG]: BlockAdjacentDifference requests 2x the shared memory actually needed

[pauleonix]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Performance

Component

CUB

Describe the bug

Looking at

cccl/cub/cub/block/block_adjacent_difference.cuh

Lines 133 to 137 in 9b7333b

	struct _TempStorage
	{
	T first_items[BLOCK_THREADS];
	T last_items[BLOCK_THREADS];
	};

and the at all the member functions, it seems like none of them is using both first_items and last_items, i.e. a single array halo_items or similar would suffice. The unnecessary amount of requested shared memory can in practice result in reduced occupancy and therefore worse performance.

How to Reproduce

Not applicable.

Expected behavior

BlockAdjacentDifference should only request as much shared memory as it actually needs.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[pauleonix]

Arguably there should be a version of these algorithms that only use shared memory for inter-warp communication and warp shuffles otherwise for minimal shared memory usage at reduced performance like e.g. BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED, but that would be another issue.

[pauleonix]

Maybe at some point it was planned to also have algorithms that look both left and right, but it is a hard ask to pessimize these common algorithms just for those non-existent one to use the same API.