search

ROCm / rocprofiler-compute

Advanced Profiling and Analytics for AMD Hardware
https://rocm.docs.amd.com/projects/omniperf/en/latest/
MIT License
135 stars 49 forks source link

Fixes #274

Closed skyreflectedinmirrors closed 8 months ago

coleramos425 commented 8 months ago

LGTM

feizheng10 commented 8 months ago

@skyreflectedinmirrors Do you get a chance to verify other metrics using GRBM_GUI_ACTIVE on MI300? Should some of them be replaced with $GRBM_GUI_ACTIVE_PER_XCD?

skyreflectedinmirrors commented 8 months ago

I believe all of the YAML files that use GRBM_GUI_ACTIVE have been updated in the refactor-specs branch now. Essentially: yes, all GRBM counts need to use the "per-xcd" version (IIRC, the only other one we use is SPI_BUSY).

skyreflectedinmirrors commented 8 months ago

and yes, I spot-checked them using some of my occupancy tests :)

feizheng10 commented 8 months ago

and yes, I spot-checked them using some of my occupancy tests :)

How about https://github.com/AMDResearch/omniperf/blob/fe3b8e06d4db2e75ea612c747eb48ccfaf31f8eb/src/omniperf_soc/analysis_configs/gfx942/0700_wavefront-launch.yaml#L137

coleramos425 commented 8 months ago

@feizheng10 I think you're just confused because we're working out of feature branch refactor-specs. See here https://github.com/AMDResearch/omniperf/blob/2bdd037a9ad68c93c28368da12f06fcca185b4a9/src/omniperf_soc/analysis_configs/gfx942/0700_wavefront-launch.yaml#L136-L142

feizheng10 commented 8 months ago

Thanks! My bad.