[Intel-gfx] [PATCH igt] igt/perf_pmu: Speed up frequency measurement

[Chris Wilson]

Quoting Tvrtko Ursulin (2017-12-19 12:57:54)
> 
> On 15/12/2017 21:05, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2017-12-15 18:24:42)
> >>
> >> Ah I see.. only for the spin batch. Why not then gem_sync or maybe we
> >> should add igt_spin_batch_free_sync?
> > 
> > gem_quiescent_gpu goes one step further than gem_sync and says the system is
> > idle / parked afterwards. Which is often quite important
> > 
> > Yes, seems like I'm repeating this pattern often enough that throwing it
> > into igt_spin_batch is worthwhile. Also I want to include a spin_batch
> > variant that guarantees it has started executing before returning.
> > Sadly will require MI_STORE_DWORD so limit it's availability. I think
> > I'll wait for the spin_batch options to land before adding more
> > parameters.
> 
> Not sure what is the status. I'd be OK with either gem_quiescent_gpu at 
> the start of subtests, or making igt_spin_batch_free ensure batch 
> finished. That should also stop all leaks between tests AFAICT.

We could add gem_sync() to batch_free (and have a __variant to avoid it
for certain tests like pm_rps, hmm, the more I think about it the more
places I can think of where we want to avoid the sync....). That still
leaves us with a quandary if we want to be sure the system is idle/parked
(i.e. !gt.awake). I think we still end up with both; a few judiciously
placed gem_quiescent_gpu().
-Chris