ROCmSoftwarePlatform
Add Compilation Flag for enabling/disabling clipping, and tune number of blocks for mscclpp allreduce8
Details
Do not mention proprietary info or link to internal work items in this PR.
Work item: "Internal", or link to GitHub issue (if applicable).
What were the changes?
Added performance optmization in mscclpp by tuning the number of blocks and disabling clipping by default.
Why were the changes made? observed significant regression between current mscclpp commit in rccl, and prior mscclpp commit.
How was the outcome achieved?
Compilation flag that is set by default to false, prevents clipping routine from wrapping each addition. Additionally, the number of blocks in allreduce8 were increased from 512 to 1024.
Additional Documentation:
Disabling clipping has impact on correctness but may be necessary for particular workloads.
Approval Checklist
Do not approve until these items are satisfied.
- [ ] Verify the CHANGELOG has been updated, if
- there are any NCCL API version changes,
- any changes impact library users, and/or
- any changes impact any other ROCm library.
@nusislam I haven't merged yet since I have added a few commits since your approval. Please let me know if there are any more reviews
@nusislam I haven't merged yet since I have added a few commits since your approval. Please let me know if there are any more reviews
I would suggest merging this one after https://github.com/ROCm/rccl/pull/1615
