Re: [PATCH v1 1/2] perf vendor events arm64: Add N3 events/metrics

From: James Clark
Date: Mon Jan 20 2025 - 05:18:02 EST




On 18/01/2025 6:22 pm, Namhyung Kim wrote:
Hello,

On Mon, Jan 13, 2025 at 04:48:06PM +0000, James Clark wrote:


On 13/01/2025 3:51 pm, Ian Rogers wrote:
On Mon, Jan 13, 2025 at 4:35 AM James Clark <james.clark@xxxxxxxxxx> wrote:



On 10/01/2025 5:20 pm, Ian Rogers wrote:
Using the scripts at:
https://gitlab.arm.com/telemetry-solution/telemetry-solution/

Generate perf json for neoverse-n3 using the following command:
```
$ telemetry-solution/tools/perf_json_generator/generate.py \
tools/perf/ --telemetry-files \
telemetry-solution/data/pmu/cpu/neoverse/neoverse-n3.json:neoverse-n3
```

Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
---
.../arm64/arm/neoverse-n3/core-imp-def.json | 930 ++++++++++++++++++

I assume you got warnings about missing groups, if one is missing that
results in everything being dumped in core-imp-def.json. I have an
internal ticket open about it which is why I didn't post N3 and V3 jsons
yet, I'll check on it again.

I was testing with JEVENTS_ARCH=all. I'm not sure what the group
warning is, do you have a link?


From the perf_json_generator:

$ telemetry-solution/tools/perf_json_generator/generate.py \
tools/perf/ --telemetry-files \
telemetry-solution/data/pmu/cpu/neoverse/neoverse-n3.json:neoverse-n3

Warning: Not writing events which do not have a group (note that the
mnemonic printed may differ to the one in the source json if it differs
in the common Perf files):

IMP_CT_FLUSH: Counts flushes including architectural,
microarchitectural, and branch redirects.
...

Up to you whether you want to continue with it like this and I can send
a fix up later but there's also a conflict with the recent Fujitsu
Monaka events.

I'm happy to let you push the updates. An alternative, that may work
around the group issue, is just to push the event updates. Any chance
on an ETA?

Thanks,
Ian

Can't say on an eta. There was a fix already done but not made live yet, but
now it turns out maybe we don't want to put these particular events into
Perf because they're not in the TRM. Could be several weeks before anything
goes live.

It's a bit of a hack but I can regenerate them without these particular
events in a V2, and then whatever the outcome is from above will be an
incremental change later.
So what's the final decision?

James, do you want to publish the metrics later at once or are you ok
with this patch for now (and do the incremental change later)?

Thanks,
Namhyung


I'm just finishing off some changes to the generator script and the change will end up a bit different. Turns out it wasn't updating the common events properly either.

I think it's best to leave this one out and I'll send a v2 this week.