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

All latency metrics are missing in CLI #244

Closed skyreflectedinmirrors closed 8 months ago

skyreflectedinmirrors commented 8 months ago

Describe the bug

image

The latency metrics are missing, even with all blocks collected

Development Environment:

To Reproduce Steps to reproduce the behavior:

using e.g., the vcopy sample:

$ omniperf profile --no-roof -n whatever -- ./a.out $ omniperf analyze -p workloads/whateverMI300A_A1/ -b 1 2

Expected behavior

The latency metrics should be present

Screenshots

See above

Additional context

I'm fairly sure this has something to do with the renaming of "raw_pmc_df" to "raw_pmc", e.g., in the Workload schema(https://github.com/AMDResearch/omniperf/blob/b9e474ec5ff6afcb3ebafc653dcb475acd7096d0/src/omniperf_analyze/utils/schema.py#L63) and the analyze (https://github.com/AMDResearch/omniperf/blob/b9e474ec5ff6afcb3ebafc653dcb475acd7096d0/src/omniperf_analyze/omniperf_analyze.py#L182). I've verified the latency expressions seem to get built sanely, but they just show up as none when eval'd. I don't super understand how this is supposed to work.

skyreflectedinmirrors commented 8 months ago

Ah, I realized I'm an idiot :). Happens.

These are explicitly left empty because there's no sane "peak"