Hello We have sucessfully (we think) ported our multiprocessor version of Mesa on the mesa-3.1beta1kw3. If you want to have a look : http://www.lri.fr/~alex/PMesa/Mesa-3.1beta1toPMesa-0.5beta3.tar.gz (more infos available at http://www.lri.fr/~alex/PMesa/) the most difficult issue were to change the code used for vertex shared between 2 VBs Features : *speedup measured up to 1.85 *you need to activate multiprocessing with setenv PMESA (or export PMESA=). If you don't declare PMESA, you have a single processor version with performance very close to mesa-3.1beta1kw3. The next step will be dynamic intelligent switching betw multi or non multi *seems to work correctly with mosts samples and demos Problems to be solved : *because of thread synchro speed, good perf are available only with big VB (VBSIZE=1920). This is a problem with 3d accelerators with small internal buffers. What I suggest is either to have a different thread that talks to the card (preemtive multitasking) or to interleaved to filling of a VB with the driver code on another VB (cooperative multitasking) *a few visual bugs left like stains on teapots (shading is not done correctly on copied vertices) *vertex arrays are not yet multithreaded. Alexis and Julien -- ******************************************************* Alexis VARTANIAN - tel: (33) 1 69 15 64 07 Equipe Architectures Paralleles, LRI Batiment 490, Universite Paris-Sud, 91405 Orsay, France *******************************************************
