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

Enable rocprof (v1) support for MI300 profiling #388

Closed coleramos425 closed 2 months ago

coleramos425 commented 3 months ago

Is your feature request related to a problem? Please describe. Currently, Omniperf requires the user set export ROCPROF=rocprofv2 to profile on MI300 skews. Per internal conversation with the rocprofiler team, rocprof (v1) has been fully tested and verified for MI300 profiling. I suggest we relax the requirement in Omniperf such that users can profile with rocprof (v1) in addition to rocprofv2.

Describe the solution you'd like I see the following steps as required:

  1. Make appropriate changes to Omniperf backend. Start with modifying set_compatible_profilers() for each gfx94x child class.
  2. Verify changes on MI300 hardware, ensuring profiling goes smoothly in rocprofv1 environment.
  3. File PR into dev branch
  4. Update documentation PR with a comment instructing @peterjunpark to update core-install.rst accordingly
  5. Update internal ticket to notify all watchers of new change

Additional context Related GitHub convo to make this change: https://github.com/ROCm/omniperf/pull/362#discussion_r1690371801 Internal tracking: https://ontrack-internal.amd.com/browse/SWDEV-475098