single node AG/RS enable warp speed

[isaki001]

Motivation

Enable warpSpeed for AllGather and AllReduce on gfx950 to take advantage of reducing CUs by half.

Technical Details

• Switch to using single-slice, single chunk-step, and unroll=2, for allReduce and AllGather on gfx950
• For single-node gfx950 AllGather:

• warpSpeed is enabled for message sizes >= 128MB
• CU reduction applies even when warpSpeed is not enabled
• 1-3% regression observed at 32MB-64MB when compared against unroll=1, CU=112, warpSpeed Disabled AllGather

• For single-node gfx950 ReduceScatter:

• warpSpeed is enabled for message sizes >= 2GB
• CU reduction applies even when warpSpeed is not enabled
• 18-24% regression observed at 1MB-4MB when compared against unroll=1, CU=112, warpSpeed Disabled AllGather
• Protocol adjustment from LL to Simple limits this regression to 13-18% for 2MB-4MB

JIRA ID

AICOMRCCL-415
AICOMRCCL-416

Test Plan

gfx950 correctness and performance tests

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.