radeon and EXA

[Michel Dänzer]

On Tue, 2005-09-20 at 14:35 -0400, Michel Dänzer wrote:
> On Tue, 2005-09-20 at 10:12 +0200, Michael Olbrich wrote:
> > 
> > The basics are pretty much what was reported so far.
> > Without xcompmgr scrolling is rather unusable.
> > With xcompmgr scrolling is better (but not as good as XAA) but switching
> > between desktops (and windows to a lesser extend) is very slow.
> 
> Do you have Eric Anholt's latest exaGlyphs improvements? They made a big
> difference here, at least for scrolling.
> 
> Unfortunately, they cause problems of their own here: Only the first
> word (sometimes the first two, if they're 'small enough') of most text
> gets rendered, e.g.
> http://penguinppc.org/~daenzer/EXA/Evolution-about-broken.png is what
> the Evolution about dialog looks like. I've tried playing with the
> radeon EXA cache flush code, building EXA with gcc-3.3 instead of
> gcc-4.0, building it with -fsigned-char, but the only way to work around
> it seems to be to disable exaGlyphs.
> 
> This is on a Titanium PowerBook IV with an M9 with basically the same
> software setup except for kernel 2.6.12, but the weirdest thing is that
> Benjamin Herrenschmidt doesn't see this on a same machine, and we have
> yet to discover any relevant difference between our setups...

Found something: I downgraded GTK+ from 2.8 to 2.6 on a hunch, and this
particular problem is gone. Now I'm only seeing the occasional minor
corruption Ben is seeing as well. What I'm not sure is whether GTK+ 2.8
(Cairo?) triggers a bug in the exaGlyphs code or maybe just causes the
apparently still remaining cache issues in the radeon driver to manifest
themselves differently.


> > What worries me most is the influence of CPU load. I've been running a
> > compiler in background and the system was more or less unusable.
> > Switching between windows takes seconds. The clock on my panel (which is
> > usually the last thing that still works when the load gets to high)
> > sometimes skips 2-3 second. Even the cursor gets jerky. While that can
> > be improved with a higher priority for the xserver it is far from the
> > responsiveness I get with XAA.
> 
> Note that EXA vs. XAA doesn't have any direct impact on how cursor
> movement is handled, unless you're using SW cursor maybe.

Actually, I can now confirm this problem. :\

> > (This is all with xcompmgr which is what we want EXA for anyway, right?)
> 
> If you don't run the compile with nice, it might help to renice the X
> server and/or xcompmgr to -1 or something.

Also, xcompmgr -a seems much snappier than client-side compositing here.


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