[Intel-gfx] [PATCH igt 3/3] igt/perf_pmu: Idle the GPU before starting to measure busyness

[Chris Wilson]

Quoting Tvrtko Ursulin (2017-11-24 07:32:30)
> 
> On 23/11/17 16:07, Chris Wilson wrote:
> > Quoting Chris Wilson (2017-11-23 16:02:20)
> >> Quoting Tvrtko Ursulin (2017-11-23 13:37:30)
> >>>
> >>> On 23/11/2017 08:22, Chris Wilson wrote:
> >>>> Make sure the HW is idle before we start sampling the GPU for busyness.
> >>>> If we do not rest for long enough between tests, we may carry the
> >>>> sampling over.
> >>>
> >>> I'd rather not have this since as I said yesterday each opened PMU event
> >>> is supposed to record the initial counter value as reference. If that is
> >>> failing or not good enough on some tests/platforms I would rather first
> >>> understand why and how.
> >>
> >> All legacy sampling fails... :| I'm putting it back!
> > 
> > So presumably it is coupling with the parking.
> 
> Or more precisely averaging with the previous sample. Every time one PMU
> client signals the timer to self-disarm, but the new PMU client enables
> the timer again in this window, the previous sample value does not get
> cleared.

[snip]
 
> But it is too evil to wait for the sampling timer in case of MMIO in the
> irq disabled section. Even though it would happen only on disable-enable
> transitions quicker than sampling period. It is not an use case, but still..
> 
> Or we could drop sample averaging.. ?

Yeah, we are accumulating too much cruft for a small improvement (that
only affects the endpoints).
-Chris