search

ROCm / omniperf

Advanced Profiling and Analytics for AMD Hardware
https://rocm.docs.amd.com/projects/omniperf/en/latest/
MIT License
132 stars 39 forks source link

gfx908 doesn't support roofline plot #385

Closed aymane-eljerari closed 2 months ago

aymane-eljerari commented 2 months ago

Is your feature request related to a problem? Please describe. I am unable to generate roofline plots using omniperf on MI100 (gfx908)

Describe the solution you'd like Being able to generate the roofline plots automatically when profiling using omniperf.

Describe alternatives you've considered So far, I only have access to speed of light values.

Additional context

root@febc8da47e3f:~/git/rocm-llm-profile# ./omniperf_profile.sh 
  DEBUG ROC Profiler: /opt/rocm-6.0.0/bin/rocprof
  DEBUG Execution mode = profile

  ___                  _                  __ 
 / _ \ _ __ ___  _ __ (_)_ __   ___ _ __ / _|
| | | | '_ ` _ \| '_ \| | '_ \ / _ \ '__| |_ 
| |_| | | | | | | | | | | |_) |  __/ |  |  _|
 \___/|_| |_| |_|_| |_|_| .__/ \___|_|  |_|  
                        |_|                  

  DEBUG [profiling] perform SoC profiling setup for gfx908
  ERROR gfx908 does not support roofline analysis
coleramos425 commented 2 months ago

@aymane-eljerari MI100 (gfx908) hardware doesn't support matrix fused multiply-add (MFMA) instructions. MFMA is a critical part of Omniperf's roofline methodology, thus to maintain this consistency, we do not plan to add support for this generation of hardware.