TMA 4.7 Update

[calebbiggers]

Updated HSX, BDX, CLX, SKX, ICX, and SPR platforms with TMA 4.7

[captain5050]

Hi, has the substitution of #smt_on in the perf metrics been broken? It looks like "1" is always substituted, e.g.

        "MetricExpr": "( ( IDQ.MITE_CYCLES_ANY - IDQ.MITE_CYCLES_OK ) / ( CPU_CLK_UNHALTED.DISTRIBUTED if ( 1 ) else ( CPU_CLK_UNHALTED.THREAD ) ) / 2 )",
        "MetricGroup": "DSBmiss;FetchBW;TopdownL3;tma_L3_group;tma_fetch_bandwidth_group",
        "MetricName": "tma_mite",

[captain5050]

Hi Caleb, would it be possible to squash thing change into the TMA 4.7 update?

[calebbiggers]

Hi Caleb, would it be possible to squash thing change into the TMA 4.7 update?
@captain5050 As I merged some recent changes in-between the two commits, I cannot squash the changes into the original commit. I can close this PR and make a new one that only includes the fixed metrics. Is that preferred?

[edwarddavidbaker]

@captain5050 As I merged some recent changes in-between the two commits, I cannot squash the changes into the original commit. I can close this PR and make a new one that only includes the fixed metrics. Is that preferred?

Your development branch will cleanly rebase on top of main. Try git rebase -i origin/main and select s for the second commit 8dfd9e7 Fix for removal of SMT_on constant. Fixup the combined commit message as necessary.

[edwarddavidbaker]

@calebbiggers I spent a few minutes looking at the current failures.

Creating event json for BDX in /home/runner/work/perfmon/perfmon/scripts/perf/broadwellx
    Mapfile(basepath).to_perf_json(outdir)
  File "/home/runner/work/perfmon/perfmon/scripts/create_perf_json.py", line 1842, in to_perf_json
    model.to_perf_json(modeldir)
  File "/home/runner/work/perfmon/perfmon/scripts/create_perf_json.py", line 1676, in to_perf_json
    csv_metrics = self.extract_tma_metrics(metric_csv, pmu_prefix, events)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/runner/work/perfmon/perfmon/scripts/create_perf_json.py", line 1514, in extract_tma_metrics
    save_form(em['MetricName'], em['MetricGroup'], em['MetricExpr'],
  File "/home/runner/work/perfmon/perfmon/scripts/create_perf_json.py", line [120](https://github.com/intel/perfmon/actions/runs/7660092214/job/20876636275?pr=133#step:5:121)5, in save_form
    assert v in events or v.upper() in events or v in infoname or v in aux, \
AssertionError: Expected power to be an event in "tma_info_system_power": "( power@energy\-pkg@ * ( 61 ) + 15.6 * power@energy\-ram@ ) / ( ( ( duration_time * 1000 ) / 1000 ) * ( 1000000 ) )" on BDX
Error: Process completed with exit code 1.

I believe we want to skip tma_info_system_power using the existing ignore list.

perfmon/scripts/create_perf_json.py

Lines 637 to 642 in 09c155f

	# metrics redundant with perf or unusable
	ignore = {
	'tma_info_system_mux': 'MUX',
	'tma_info_system_power': 'Power',
	'tma_info_system_time': 'Time',
	}

A patch similar to the following should work.

@@ -1171,6 +1180,9 @@ class Model:
 
             def save_form(name, group, form, desc, locate, scale_unit, threshold,
                           issues):
+                if name in ignore:
+                    _verboseprint2(f'Dropping {self.shortname} metric {name}.')
+                    return
                 if self.shortname == 'BDW-DE':
                     if name in ['tma_false_sharing']:
                         # Uncore events missing for BDW-DE, so drop.

[captain5050]

I think we want to keep the metrics but need to update the skip here:

perfmon/scripts/create_perf_json.py

Lines 1193 to 1201 in 09c155f

	if v in ['if', 'then', 'else', 'min', 'max', 'core_wide',
	'SMT_on', 'duration_time', 'cmask', 'umask',
	'u', 'k', 'cpu', 'cpu_atom', 'cpu_core', 'edge',
	'inv', 'TSC', 'filter_opc', 'cha_0', 'event',
	'imc_0', 'uncore_cha_0', 'cbox_0', 'arb', 'cbox',
	'num_packages', 'num_cores', 'SYSTEM_TSC_FREQ',
	'filter_tid', 'TSC', 'cha', 'config1',
	'source_count', 'slots', 'thresh', 'has_pmem',
	'num_dies', 'num_cpus_online']:

to handle the power metrics case, like it is doing for cpu. The assert assumes the events for the metric are either in that list or are defined in the architecture's event json files.

resolved.