Some benchmarks with Radeon and Xorg7.1

[Roland Scheidegger]

Paulo J. Matos wrote:
> I've tested X against three options: AGPFastWrite, EnablePageFlip and
>  AccelMethod. Both of these have 2 possible values, True/False for 
> the first two and XAA/EXA for the last one. I start fluxbox, 54 
> processes are running with an average of 0% CPU, 138112 used RAM and 
> 377896 free RAM. The benchmarks table is as follows (CPU is average 
> during glxgears runtime and FPS are given by glxgears): 
> Method     EnablePageFlip     AGPFastWrite     FPS     CPU
> XAA     False     False     702     10%
> XAA     False     True     702     10%
> XAA     True     False     978     14%
> XAA     True     True     978     14%
> EXA     False     False     661     10%
> EXA     False     True     661     10%
> EXA     True     False     661     10%
> EXA     True     True     661     10%
glxgears isn't really a benchmark.
That glxgears is slightly slower with exa is interesting though, I'm a bit
puzzled why the 2d accel method would make a difference, maybe there is
some additional overhead somewhere (dri client is more often interrupted
by 2d?).
Pageflip won't work with exa so naturally you don't get any difference
there.
> I should add that in all EXA configurations, the screen sometimes 
> would generated distorted images on the top, disallowing a 
> confortable use even if it happened to be better, which is not. So, 
> currently I'm using XAA with only EnablePageFlip enabled.
> 
> In fact, it is understandable that AGPFastWrite changes nothing since
>  by closely analysing lspci I notice that my motherboard as FW+ and 
> video card as not, so video card doesn't support FastWrite which is 
> unfortunate.
Even if it works, it simply won't do anything for performance. You're
fortunate if it doesn't just lockup (happens on a lot of card/chipset
combinations).
It should be noted that glxgears gains a lot more performance by
pageflip than normal apps (due to the very high framerate).
For another (big) boost you could enable hyperz (driconf or
hyperz=true), though again glxgears will benefit more than real apps.

> Hopefully, EXA gets better in later releases and performes better 
> than XAA and without the distortion.
Well, there really shouldn't be a difference with 3d performance if exa
or xaa is used. If something like, say, quake3, would show exa performing
slower by 5% I'd be worried.

Roland