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

Add new rooflines #226

Closed JoseSantosAMD closed 9 months ago

maartenarnst commented 9 months ago

Just a question out of curiosity. Why are the rooflines provided as binaries? I would argue it could/would be of interest and of value to users/researchers to see the source code and see how these rooflines are obtained.

feizheng10 commented 9 months ago

Just a question out of curiosity. Why are the rooflines provided as binaries? I would argue it could/would be of interest and of value to users/researchers to see the source code and see how these rooflines are obtained.

The binary is only for measuring the peak BW on each hierarchical level. It is "debatable" open source or not:) The arithmetic part is open source.

coleramos425 commented 9 months ago

The binary is only for measuring the peak BW on each hierarchical level. It is "debatable" open source or not:) The arithmetic part is open source.

That's right. The source code for the roofline binary reveals microarchitectural details that our hardware team has not signed off on. @maartenarnst we'll be sure to update you if that changes in the future. As Fei mentioned, the derivation for AI (arithmetic intensity) calculation is open-source if that's of any value:

https://github.com/AMDResearch/omniperf/blob/b9e474ec5ff6afcb3ebafc653dcb475acd7096d0/src/omniperf_analyze/utils/roofline_calc.py#L188