Hi Ian,

On 8/5/2021 1:09 PM, Ian Rogers wrote:

Hi,

I'm not sure what's happening but when I generate this using:
https://github.com/intel/event-converter-for-linux-perf

I see here:
         "MetricExpr": "1 / IPC",

There are various other differences and so I've included the generated file.

Thanks,
Ian

Please see following commit in event-converter-for-linux-perf repo.

commit b2265b953d57130a2ad8255d126bb69cf8552898
Author: Jin Yao [off-list ref]
Date:   Thu Jun 3 09:57:01 2021 +0800

     Don't expand inner metric if possible

     Before:

         {
             "MetricExpr": "1 / (INST_RETIRED.ANY / CPU_CLK_UNHALTED.THREAD)",
             "BriefDescription": "Cycles Per Instruction (per Logical Processor)",
             "MetricGroup": "Pipeline",
             "MetricName": "CPI"
         },

     After:

         {
             "MetricExpr": "1 / IPC",
             "BriefDescription": "Cycles Per Instruction (per Logical Processor)",
             "MetricGroup": "Pipeline",
             "MetricName": "CPI"
         },

     We also provide a new option '--expr-events' for checking if it's a complex
     metric. For a simple metric, we probablly don't need to reuse the inner
     metric. For a complex metric, it's useful to reuse the inner metric.

I enable the option '--expr-events' when generating the metrics JSON. Since CPI is a simple metric, 
so the inner metric is still expanded.

For cstate metrics, only Core C1/C6 and Package C2/C6 are supported for ICX. The cstate metrics are 
appended manually right now.

Thanks
Jin Yao