Comme=
nt # 14
on bug 10354=
4
from Roland Scheidegger
(In reply to Ilia Mirkin from comment #13)
> (In reply to Roland Scheidegger from comment #12)
> > Here's a apitrace for this (1GB, of course the corruption is only=
seen
> > towards the end...), should be available a week (?):
> > https://we.tl/EanuxRG7Yf
>=20
> Well, this all replays fine with nouveau, even if I remove the blending
> config which makes 0*nan=3D0. At least I couldn't see any artifacts.
>=20
> So this all points to ... sadness. Can you check if you still see arti=
facts
> with nosb?
Tried that, doesn't help.
I noticed some things looking questionable in the driver:
- the driver doesn't use the ieee variants consistently (e.g. muls from ler=
ps).
There's also some inconsistencies - for r600, the comments say for non-gl u=
sage
the ieee variant should be used for rcp, however eg/cayman will use that
anyway, and I don't think handling it differently between these drivers is =
done
on purpose.
- the driver uses min/max instead of min_dx10/max_dx10 (though at least the=
EG
ISA docs are wrong/inaccurate what they exactly do). These would kill off N=
aNs
(albeit I am not entirely sure if the non-dx10 version doesn't kill off NaNs
neither, given the docs aren't accurate in the first place...). I think this
probably should be changed (as far as I can tell, radeonsi also uses
effectively the dx10 versions, since llvm.min/maxnum is specified as select=
ing
non-nan operands, though I don't know if the backend honors it really). This
actually does help things, it removes the black holes in the fires, and the
"vertical blue stripes" (albeit the latter wasn't really in that =
trace). But
otherwise there's still lots of black boxes around.