On Thu, 23 Sep 2010, Carlos R. Mafra wrote: > On Thu, 23 Sep 2010 at 18:40:40 +0400, Alexey I. Froloff wrote: > > On Thu, Sep 23, 2010 at 04:21:49PM +0200, Andreas Tscharner wrote: > > > > Inline assembly doesn't makes sense with modern compilers. > > Oh, I'm sorry, this was more common than I wanted to say. > > > > > I think the ffmpeg developers would not agree with you. > > I doubt wrlib was made to process 512kbps stream in realtime ;-) > > However, FFmpeg built with CMOV optimisation shows ~1% > > performance improvement. > > > > First of all, assembly optimisations are very platform-dependant. > > Next, for every processor you have to implement several > > realisations, for ix86 they are MMX, SSE, 3DNow!, SSE4, SSE4.2 > > and so on. Hand-made MMX (as in our case) almost useless on > > current hardware and compilers... > > Perhaps we should use the -march= kind of optimization that gcc > offers by doing some detection voodoo in the configure script.
please be careful to only use these kinds of flags if the architecture and the compiler both support it. there are not many more disappointing things in the arena when a software confidently thinks it will be compiled on a p-mmx with gcc, whereas i was actually doing it on sparc with sun c... just did a non-authoritative quick test, and both clang and sunc (on linux at least) seem to generate workable binaries, let's not ruin this. i know you are careful, i'd just like to stress this. -- [-] mkdir /nonexistent -- To unsubscribe, send mail to [email protected].
