> > i replaced the transform_v16, matmul4 and matmul34 with simd instruction > functions and there was no performance increase in quake2 or q3test. > any good reason to have simd support anyway ? :( > > -- > ralf willenbacher ([EMAIL PROTECTED]) > as stated by others, it will get more and more difficult to get impressive speedups in Q2 or Q3. I think you should have a look to the viewperf benchmark (http://www.spec.org/gpc/) suite that all card makers with geometry acceleration use to compare their boards. Viewperf give a good idea about the behavior of triangle intensive CAO applications and really depends on triangle processing. I know that because we got good speedups (see http://www.lri.fr/~alex/PMesa) when we were trying to paralellize geometry transformations on multiprocessors. Testing SSE (or 3DNow!) there may give more encouraging results. The benchmarks mostly used in the viewperf is ProCDRS but a lot of memory is needed. If you don't have enough memory, use CDRS. I don't know which transformation functions are used there, profiling should give the answer. By the way, Keith I think you mentionned a way to profile a library without recompiling the calling application (compaling the library with -pg alone doesn't work), but I can't find where. How do you do that ? Alexis -- ******************************************************* Alexis VARTANIAN - tel: (33) 1 69 15 64 07 Equipe Architectures Paralleles, LRI Batiment 490, Universite Paris-Sud, 91405 Orsay, France ******************************************************* _______________________________________________ Mesa-dev maillist - [EMAIL PROTECTED] http://lists.mesa3d.org/mailman/listinfo/mesa-dev
