i built Mesa 6.4 with -O3 -mcpu=iwmmxt -mtune=iwmmxt to get it going faster and it runs faster from my tests:
from 16.5FPS to 18FPS @416MHZ in glxgears
from 26FPS to 28FPS @624MHZ in glxgears

there some demos on the build also.
Sash:about the pcsx build that is on the feed,does it build with my patches? coz it is slow by 2-3 fps that my build!

one thing i realy pissed is that we have such poor 3d preformence with ur Zaurus.
i searched for a software-render OpenGL such as mesa with integer instead of floats but i didn`t find one.
if you look at PPC they get such a good speed.
i compiled foobillard with OpenGL and it runs for 3-5FPS(estimate) realy slow.

where is your patches?
i added but not tested it

so how did you build it? my patches aren`t for boost but are for building it but they also make it faster.
patches here.
it runs damn slow just 6-8FPS(your build) and 8-11FPS(my build).
something else that could build is Visual Boy that run as slow as 50-60%@624MHZ.

i added pcsx to builder, so it will building every new release automatic

i mean ,how did it build it coz you need to hack the makefiles and some files to get it build.
and check my patches

hacking around in the floating point emulation should speed things up, not sure if gcc supports autovactorisateion for the iwmmx but i am trying to find out at the moment

only with gcc 4.0.(because of that Apple use 4.0 for VMX(aka Altivec) optimz).
the better thing i planing to do is write a arm dynamic recompiler so we will have alot of speed boost i think about 25-30FPS which will be damn good.
think about this: Fifa 2003(or sooner) on the Z just amazing.
tekken 3,Gta2 and all the good thigs

Sounds cool - is this the kind of thing which profiles itself then recompiles to make the parts which are run the most run the fastest?


Si

i thoght that gcc had that built in, you just had to do multiple passes, the thing i wonder about fir autovectorisation is does gcc have support for it or only mmx/sse/altivec

i thing the speed boost from vectorising the floating point rouitnes would be better than dynamic recompalation scince thats the area thats cauising the most performance degerdation

Yes, but one of those passes has to be performed on the machine.



Although vfp has vector support, I thought that this was only for hardware floating point units? I'd assumed that as it's emulated there would be no difference (or can the emulation perform vector calculations somehow?).


Si

this thread has quite some info about vfp: http://maemo.org/pipermail/maemo-developer...ber/001602.html

a interpter core runs a lot slower than dynamic recompiler so building one will be the best option

thanks for the link to the mailing list, i would be intrested to see how much of an improvment we could get, i might even do some iwmmx asm if i feel there is a good performance boost, i am planning to conver the mmx AES routines to iwmmx so mabey after that

just a note, iwmmx includes all integer SSE stuff, so its actually a superset of your standard mmx, i dont know how this effects the code but i assume this would allow for some instructions that would take longer to do in pure mmx