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

Timestamp inaccuracy using --roof-only #61

Closed coleramos425 closed 1 year ago

coleramos425 commented 1 year ago

When profiling with the --roof-only flag, timestamps are not corrected. You can see timestamps.csv isn't generated and replace_timestamps() isn't called.

https://github.com/AMDResearch/omniperf/blob/9bc41f3a85b4bea7fa7febdec104983da41b9e51/src/omniperf#L286-L300

This leads to inaccurate kernel duration readouts in analyze mode. This logic needs to be added to the standalone roofline.

CC: Georgios

coleramos425 commented 1 year ago

Updates have been made and are available in dev branch. Expect this to be included in our next release.