[XPU] Enable reduction optimization by default #2846
Conversation
Add reduction optimization pass to the pipeline by default.
victor-eds
requested review from
etiotto,
whitneywhtsang,
chengjunlu and
a team
|
Performance gains can be seen under @chengjunlu I'd say we go with this for now. You can disable it when doing your current work. This is giving performance now and having this would make testing the followups we have in mind easier. WDYT? |
|
Looks like the wide vector flag is giving some better performance, but this is noticeable enough. Just a minor bump from that flag. |
LGTM. |
| if os.getenv("TRITON_INTEL_OPTIMIZE_REDUCTION_LOCALITY", "0") == "1": | ||
| intel.passes.ttgpuir.add_optimize_reduction_locality(pm) | ||
| intel.passes.ttgpuir.add_optimize_elementwise_parallelism(pm) | ||
| intel.passes.ttgpuir.add_optimize_reduction_locality(pm) |
There was a problem hiding this comment.
Can we use if os.getenv("TRITON_INTEL_OPTIMIZE_REDUCTION_LOCALITY", "1") == "1": to allow the user to disable it?
There was a problem hiding this comment.
Of course. I'll add that when I finish the investigation
|
From https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12061551469, it appears that the reduction optimization has a negative geomean impact to FA (causal=false, dhead=64) with the new IGC driver (improved instruction scheduling). Can we investigate it further before enabling it by default? |
Yes! I can look into that. |
|
Waiting till regressions are fixed. #2907 is the best attempt we have currently. |
Add reduction optimization pass to the pipeline by default.