Commen=
t # 6
on bug 10354=
4
from Ilia Mirkin
(In reply to Roland Scheidegger from comment #5)
> I've actually got that game myself here.
> So I did the bisect and the winner is:
> ce7a045feeef8cad155f1c9aa07f166e146e3d00 is the first bad commit
> commit ce7a045feeef8cad155f1c9aa07f166e146e3d00
> Author: Ilia Mirkin <im=
irkin@alum.mit.edu>
> Date: Mon Jan 23 20:53:50 2017 -0500
>=20
> r600g: use ieee variants of multiplication instructions
>=20=20=20=20=20
> This matches the behavior of most other drivers, including nouveau,
> radeonsi, and i965.
>=20=20=20=20=20
> Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
> Reviewed-by: Nicolai H=C3=A4hnle <nicolai.haehnle@amd.com>
>=20
> Looks like some numerical issue then, albeit I don't know if the game =
is at
> fault here.
Apologies for the trouble.
The main difference between IEEE and non-IEEE is whether 0 * infinity =3D 0=
or
NaN. IEEE makes it mean NaN. DX9 behavior is 0. I added a flag to be used by
st/nine to enable the DX9 behavior optionally, but leave the IEEE behavior =
for
GLSL. (There was some additional desire to expose that in a GL ext for WINE=
to
use, but it got shot down pretty quickly.)
Perhaps there are other changes from using the IEEE instruction variants, e=
.g.
denorms, which would be undesirable. I was never too familiar with the R600
ISA.