Improving rotated ATI performance

Jerome Glisse glisse at freedesktop.org
Fri Dec 14 06:25:06 PST 2007


On Thu, 13 Dec 2007 13:24:22 -0500
"Alex Deucher" <alexdeucher at gmail.com> wrote:

> On Dec 13, 2007 1:19 PM, Adr3nal D0S <adr3nald0s at gmail.com> wrote:
> > On Dec 13, 2007 11:37 AM, Alex Deucher <alexdeucher at gmail.com> wrote:
> >
> > > rotation can use EXA composite to accelerate transforms, so you'll
> > > have to use EXA.  For r1xx-r2xx the composite code is written, but
> > > there are some bug when doing transforms, so it doesn't work properly
> > > for rotation.  for r3xx and newer, the code needs to be written.
> > > there should be enough info available for r3xx-r4xx in the 3D driver
> > > to add composite support, but no one has done the work yet.
> >
> > I've done a fair amount of OpenGL over the years, but haven't dealt
> > with 3D much at the hardware level.  Am I correct that 3D acceleration
> > only works up to a maximum virtual size of 2048x2048? If so, does this
> > mean it would be impossible to have accelerated compositing under EXA
> > for larger virtuals?
> 
> It depends on the hardware.  r1xx and r2xx had 2048 3d coordinate
> limits and tiling surface limits.  r3xx/r4xx are 4096.  textures are
> limited to 2048 on r1xx-r4xx.

Alex would be good if you can double check but r300 hw i got can't
properly render 3d above 2650x2650 (offset things you can see in the
drm). As side note fglrx doesn't render properly above this limit.
So i think we should change advertized viewport size to 2650, i
will do that soon if no one object with valid arguments.

We still can work around this but doing it with current mesa
infrastructure is likely too much pain to be worth it. Likely
a lot easier to work around this in exa.

Cheers,
Jerome Glisse



More information about the xorg mailing list