[igt-dev] [PATCH i-g-t 1/3] tests/perf_pmu: Tighten busy measurement

[Chris Wilson]

Quoting Chris Wilson (2018-02-01 17:16:28)
> Quoting Chris Wilson (2018-02-01 17:08:09)
> > Quoting Tvrtko Ursulin (2018-02-01 16:58:21)
> > > 
> > > I am just moving the two sample read-outs closer together in time. Two 
> > > samples are measure of how long we slept, and measure of PMU busyness. 
> > > There should be as little time as possible between the two readouts. 
> > > Ending the spinner I don't see how it belongs in this sandwich, or 
> > > anything else really.
> > 
> > Then we disagree on what we think we are measuring. The usleep should
> > equal the busy value defined by the spin-batch.
> 
> How about if it was phrased:
> 
>         create spin-batch per-engine queued on a unsignal vgem
>         open_group
> 
>         signal vgem // submit all queued spinners simultaneously (well, that's our goal)
>         duration = measured_usleep
>         end spinners
> 
>         busy = read_group
> 
> That's what I want to measure. The passage of time outside of the
> spinner phase is then irrelevant, and the error inside is something that
> we are trying to minimise as part of the driver overhead.

Ok, so we are looking at slightly different angles here. The patch
applies your idea of excluding the vagaries in controlling the spinning
batch from measuring the accuracy of the PMU event itself,

Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>

But I think there's some merit in my pov still :)
-Chris