[Mesa-users] How to get the most out of llvmpipe?

[Brian Paul]

On 04/19/2012 09:40 AM, Henn, Thomas wrote:
> Hi all,
>
> I investigated the performance of Mesa's llvmpipe with our application. Unfortunately, I don't seem to get the most out of llvmpipe, as my 8 cores never exceed 1% fractional CPU load by the llvmpipe threads while my framerates are rather poor. Using a hardware accelerated driver, I get reasonable high framerates. Thus, there should be no problem with my application logic throttling everything down...
>
> As openarena is often used for benchmarking, I did another test with llvmpipe and openarena. And wow, all my cores are loaded with 6 to 8% by llvmpipe threads making up a beautiful 50 to 60fps.
>
> Now my question is: What am I doing wrong?
> Is there something I have to take care of to get a good rendering performance with llvmpipe?
> Are there some special guidelines I have to follow?
> Some Dos and Don'ts?

Does the application use any unusual GL features?  Are the shaders 
unusually long or complex?  Can you disable various GL features one by 
one and see if there's any effect?

If you made an apitrace of the GL calls 
(https://github.com/apitrace/apitrace) I might be able to take a quick 
look.

-Brian