Reproducing uarch pipeline metrics: Mem Bound, Core Bound, Retiring, Front-End Bound, Bad Speculatio

[kaplannp]

Hi,

I've been using this built in graphic from the uarch exploration profiling set.

I find it quite helpful, but I'd like to be able to generate it with the minimum number of performance counters (not the whole uarch exploration) because as I understand it, less counters improves accuracy, and reduces storage space. My question:

What are formulas to compute these metrics? Which hardware events do I need to measure (on Ice Lake)? The results should add up to 100%

Thanks for your help!

[yuzhang3_intel]

You can get TMA metrics formulas based on PMU events from the below repo:

https://github.com/intel/perfmon

[kaplannp]

Thanks! This is a great resource.

One question, when I look at the Icelake formulas, I see they depend on `PERF_METRICS.***` e.g. `PERF_METRICS.FRONTEND_BOUND`.

What are these, and how do I collect them? They don't seem to be PMU events because they're not listed in the events json, and they can't be collected with `-knob event-config=...`

[yuzhang3_intel]

After ICX , we add some new PMU events about TMAM, the prefix is PERF_METRICS.

[kaplannp]

Strange. If they're PMU events, why can't they be specified with `-knob event-config=...` and why aren't they listed here https://github.com/intel/perfmon/tree/main/CLX/events ?

[yuzhang3_intel]

If you want to know which hardware events the current platform supports, you can open VTune GUI, select uarch analysis type, and then click 'plus' below. Scroll down to see all the supported events.

All the supported events, including default and others.