Xorg 7.0-rc1 and EXA (radeon 9200)

Michel Dänzer michel at daenzer.net
Wed Oct 26 09:26:50 PDT 2005


On Wed, 2005-10-26 at 17:36 +0200, Stephane Marchesin wrote:
> On Wed, Oct 26, 2005 at 04:10:11PM +0200, Michel D?nzer wrote:
> > On Wed, 2005-10-26 at 16:00 +0200, Stephane Marchesin wrote:
> > > 
> > > Now, for your surface, I suggest you just do tiled pixel acess, the
> > > tiling functions are in the radeon_span* files in the radeon dri driver.
> > 
> > It's not that simple because it's not the driver itself that touches the
> > framebuffer, so this approach would mean that fb would have to deal with
> > all kinds of tiling, endianness, ... Doesn't seem feasible.
> 
> Surfaces are limited resources. You'll run out of them sooner or later,
> so you have to have fallbacks anyway.

Sure, the fallback is to download the pixmaps to system memory for
software rendering. Actually, some people argue this might be faster
than doing direct CPU access to the framebuffer anyway, it might be a
good idea to benchmark that before spending too much time on tuning the
latter...


> > I agree with Ben that a way for the X server to reserve a couple of 
> > physical surface registers (as opposed to virtual surfaces, which can 
> > be merged by the DRM) would be a good solution for this problem.
> 
> Well, look at the memory allocation policy between ddx and dri : it's
> ilt upon the same "reservation" kind of thinking, and the dri
> performance suffers a lot from that.

There's no question that's bad, people are working on the unified memory
manager to fix that. I don't agree that the situation with surfaces is
all that similar though. What kind of surface usage in clients do you
anticipate that would make it critical that clients can get 8 as opposed
to 5 physical surfaces?

> But doing the same with surfaces, you'fre just reproducing the same
> error...

But it's better to hurt X server performance by requiring up to 6 ioctls
per software fallback to allocate and free virtual surfaces? I wonder
how much time a typical user spends using X vs. GL.


-- 
Earthling Michel Dänzer      |     Debian (powerpc), X and DRI developer
Libre software enthusiast    |   http://svcs.affero.net/rm.php?r=daenzer



More information about the xorg mailing list