Re: [PATCH] perf stat: Fix the hard-coded metrics calculation on the hybrid

[Liang, Kan]

On 2024-06-06 3:34 a.m., Namhyung Kim wrote:
> On Wed, Jun 5, 2024 at 10:21 AM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>>
>> On Wed, Jun 5, 2024 at 9:10 AM <kan.liang@xxxxxxxxxxxxxxx> wrote:
>>>
>>> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>>>
>>> The hard-coded metrics is wrongly calculated on the hybrid machine.
>>>
>>> $ perf stat -e cycles,instructions -a sleep 1
>>>
>>>  Performance counter stats for 'system wide':
>>>
>>>         18,205,487      cpu_atom/cycles/
>>>          9,733,603      cpu_core/cycles/
>>>          9,423,111      cpu_atom/instructions/     #  0.52  insn per cycle
>>>          4,268,965      cpu_core/instructions/     #  0.23  insn per cycle
>>>
>>> The insn per cycle for cpu_core should be 4,268,965 / 9,733,603 = 0.44.
>>>
>>> When finding the metric events, the find_stat() doesn't take the PMU
>>> type into account. The cpu_atom/cycles/ is wrongly used to calculate
>>> the IPC of the cpu_core.
>>>
>>> Fixes: 0a57b910807a ("perf stat: Use counts rather than saved_value")
>>> Reported-by: "Khalil, Amiri" <amiri.khalil@xxxxxxxxx>
>>> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>>
>> Reviewed-by: Ian Rogers <irogers@xxxxxxxxxx>
>>
>> Thanks,
>> Ian
>>
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> ---
>>>  tools/perf/util/stat-shadow.c | 4 ++++
>>>  1 file changed, 4 insertions(+)
>>>
>>> diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
>>> index 3466aa952442..4d0edc061f1a 100644
>>> --- a/tools/perf/util/stat-shadow.c
>>> +++ b/tools/perf/util/stat-shadow.c
>>> @@ -176,6 +176,10 @@ static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type
>>>                 if (type != evsel__stat_type(cur))
>>>                         continue;
>>>
>>> +               /* Ignore if not the PMU we're looking for. */
>>> +               if (evsel->pmu != cur->pmu)
>>> +                       continue;
> 
> Hmm.. Don't some metrics need events from different PMU?
> Like cycles per sec or branch instructions per sec..
>

Right.

In the hard-coded metrics, the events from a different PMU are
SW_CPU_CLOCK and SW_TASK_CLOCK. They both have the stat type,
STAT_NSECS. Perf should ignore the PMU checking for the type as below.
I will send a V2 to fix it.

diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
index 3466aa952442..d01335f18808 100644
--- a/tools/perf/util/stat-shadow.c
+++ b/tools/perf/util/stat-shadow.c
@@ -176,6 +176,9 @@ static double find_stat(const struct evsel *evsel,
int aggr_idx, enum stat_type
		if (type != evsel__stat_type(cur))
			continue;

+		if ((type != STAT_NSECS) && (evsel->pmu != cur->pmu))
+			continue;
+
		aggr = &cur->stats->aggr[aggr_idx];
		if (type == STAT_NSECS)
			return aggr->counts.val;


Thanks,
Kan

> Thanks,
> Namhyung
> 
> 
>>> +
>>>                 aggr = &cur->stats->aggr[aggr_idx];
>>>                 if (type == STAT_NSECS)
>>>                         return aggr->counts.val;
>>> --
>>> 2.35.1
>>>
>