On 31/05/2017 13:45, Chris Wilson wrote:
On Wed, May 31, 2017 at 11:28:07AM +0100, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Most of the subtest were failing on my SKL GT2 and on the
various CI systems as well. Try to fix that by different
tweaks per subtests:
performance:
We cannot say how big the performance drop will be once the
fences are contented, just that there will be one, so modify
the assert accordingly.
No. The goal is that the contention is not noticeable. Asserting that
contention is inevitable is baking in a kernel bug instead of fixing the
kernel.
Bug in the case of this specific subtest? I see now that it is different
that the others in a way that it goes through objects cyclically so
maybe really shouldn't be so hardly impacted by the oversubscribe of
fences. Need to look into what's happening there.
But what about the other subtests? Comments on those?
Threaded one certainly cannot be shielded from the oversubscription.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
