[CUB] Replace Shuffle(Up|Down|Index) with cuda::device::warp_shuffle

[fbusato]

Description

Address the aliasing issue in:

• Avoid alias violations in SHUFFLE_{UP, DOWN, INDEX} #8117

Additionally, deprecate legacy shuffle functions in favor of new CUDA device APIs.
Notes: SHFL instruction counts are preserved, reduce benchmark kernels show significantly fewer instructions (up to 20%)

[github-actions]

😬 CI Workflow Results

🟥 Finished in 1h 52m: Pass: 97%/298 | Total: 10d 08h | Max: 1h 50m | Hits: 81%/373048

See results here.

[bernhardmgruber]

That's a very impactful change. I see it affects CUB's radix sort, reduce and scan. I am afraid we need to see benchmarks for all those algorithms on at least one architecture.

Q: Are the new warp_shuffle_idx functions implement the same on all CUDA architectures? If not, we may want to get an additional benchmark of one of the algorithms on an architecture where the shuffle implementation is different.

[bernhardmgruber]

SHFL instruction counts are preserved, reduce benchmark kernels show significantly fewer instructions (up to 20%)

Unfortunately, instruction count does not always translate to runtime improvement. For a change to our algorithms, there either needs to be no changes in SASS or a benchmark that shows the impact. I know it's annoying, but letting a regression slip in would be even worse.

We could try the new benchmark CI to do this ;)

[github-actions]

⚠️ Benchmark Results

CUB benchmark comparison cancelled.

Results
Artifacts