On Mon, 6 May 2002, Zdenek Kabelac wrote:
> > I tryed it with direct rendering enabled and using memcpy optimized for
> > mem2agp copy (copy blocks, aligned to cache size) and result is faster.
> > But main speed problem is in decoder and because we have no source from
> > windows codec we have to use alternative decoders, which are much more
> > faster than windows codecs (eg. ffmpeg, divx5).
>
> Well - could you be more specific in which are aviplay is copying memory ?
> (i.e. Win codecs are decompressing directly to the specified buffer -
> no memcpy - ffmpeg need stride memcpy version - which using mmx code
> - and this might be improved for PIII or better CPU by prefetching data -
> but I could hardly image it would gain some speed by calling general
> accelerated memcpy routine.
oops you dont uderstand me. I think, optimized codecs (not only memcpy,
but mainly optimized idct and other critical parts). We can't do anything
with codecs and codecs are mostly slower part when playing film.
>
> > But if you are not able to use direct rendering, the optimized memcpy can
>
> you should be always able to use direct mode unless you want
> to rescale RGB (which is btw made in the most stupid way in avifile and
> there are much better places for optimalization here which will
> gain 400% speed boost in this are)
>
Can we try to get yuv2rgb and rescale code from mplayer ?
Imho their code is well optimized.
> > bring small diference between 110% and 97% CPU usage.
> > Not all users have CPU under 1GHz and accelerated Graphics card.
>
> Well unlike Arpi who will spent hours with optimalizations for
> Pentim S3 Trio64 users I've considered myself that I've no time
> to suppert users who don't have $30 to buy even some old TNT2 card....
>
OK
you are right.
> > (i know, that using mplayer is better. But why not add it, if we can.
> > If you want i can try it and implement optimized memcpy into avifile).
>
> There would be much better idea to implement the mmx pipeline -
> In reality we actually do not need fast memcpy at all...
>
Yes the best idea is to have these basic functions optimized for all
aplications (glibc :-) and put our time in more important problems.
what do you mean mmx pipeline ? (or in which part of code ?)
MOJE
--
Tomas Konir
Brno
ICQ 25849167
_______________________________________________
Avifile mailing list
[EMAIL PROTECTED]
http://prak.org/mailman/listinfo/avifile
