I think the bad result of llvm can be explained because of the lack of muladd support currently. Unigine 3.0 has a lot of geometry and i suspect vertex shader being almost twice bigger than they are in tgsi case does not help.
Fwiw with an hd 6950 I have the same performance in unigine 3 high, medium texture, no ssao (it seems to use indirect addressing) with llvm backend as high, high texture, no ssao with fglrx under Windows. Its not a fair comparaison but I think 3.8 kernel may provide the necessary boost to cope up with fglrx. Anyways I have some fps peak at 60fps that does not show up with tgsi backend that I also have with fglrx that makes me think llvm backend generates rather efficient code, but i always cherry pick muladd patches.
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
