ROCm / rocprofiler-compute

Advanced Profiling and Analytics for AMD Hardware
https://rocm.docs.amd.com/projects/omniperf/en/latest/
MIT License
138 stars 50 forks source link

MI300 Roofline Enablement in Omniperf #454

Closed cfallows-amd closed 2 weeks ago

cfallows-amd commented 1 month ago

requirements.txt

Note- other distros will get roofline bins to enable mi300, but need to be further tested before putting into branch.

coleramos425 commented 1 month ago

Overall, this looks good. My only comments are:

  1. The workaround for MCLK fixed to 1300 MHz. Is this related to broken --showmclkrange ticket? If so, we should reconnect with rocm-smi team because this issue is likely affecting many other projects. Any reason the old conditional we had in place wasn't sufficient? https://github.com/ROCm/omniperf/blob/75870cd3c7e1614af7d8a822030417fb0623ec92/src/omniperf_soc/soc_gfx942.py#L74-L76
  2. Still need Python formatting via Black formatter (i.e. black --diff --check .)