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

PMC metric not found #75

Closed aosewski closed 1 year ago

aosewski commented 1 year ago

Hi,

I'm trying to profile a program on MI100 GPU and got following error:

ERROR: rocprofiler_open(), Initialize(), input metric 'TCP_GATE_EN1_sum' is not found

Similarly the metric SQ_IFETCH is also not found. I'm using ROCm 5.3.0 and I can't find those metrics for MI100 in rocprofile https://github.com/ROCm-Developer-Tools/rocprofiler/blob/rocm-5.3.0/test/tool/gfx_metrics.xml#L80

I'm using version 1.0.6.

[EDIT] I see that such expressions are defined here: https://github.com/ROCm-Developer-Tools/rocprofiler/blob/amd-master/test/tool/metrics.xml#L127 However this one and a lot of others are defined only for MI200. And you are using them for MI100: https://github.com/AMDResearch/omniperf/blob/main/src/perfmon_pub/mi100/pmc_tcp_perf.txt

coleramos425 commented 1 year ago

Thank you for reporting this discrepancy @aosewski. It's possible there are some counters that may have slipped into MI100 configs we overlooked.

Let me cross-check our listings and I'll report my findings in this thread.

coleramos425 commented 1 year ago

Update: I've confirmed this issue on ROCm 5.3.x and 5.4.x releases. Looking into possible solutions to include in our v1.0.7 release

coleramos425 commented 1 year ago

Fix has been merged into dev branch. Closing issue.