On 14/06/16 18:47, Martin Peres wrote:
On 14/06/16 17:58, Juha-Pekka Heikkila wrote:
Here is fixed version of this ralloc set. Now I got to run this on many
different machines thanks to Mark Janes. There didn't show up any
regressions on different gen hw. On my IVB I've been running also many
different traces with Apitrace while having Valgrind running on
background
but Valgrind did seem to be happy with my changes.
As a performance test I did shader-db compile runs 10 times and compare
timing results against what Mesa master does on my IVB. To my surprise
this
does bring reasonable gain which also seem to be repeatable, on my IVB
shader compile time is around 5% faster with these changes.
On my SKL gt2, I only get a 0.35% improvement (10 runs also).
Ministat says that there is no difference proven at 95% confidence and
0.35% at 90%. Adding 100 more runs overnight, we'll see what we get in
the end.
With n=110: Difference at 95.0% confidence
-0.133665% +/- 0.0945651%
(Student's t, pooled s = 0.447829)
Martin
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
