[MPlayer-dev-eng] TODO

[Michael Niedermayer]

Hi

On Tuesday 12 February 2002 18:16, Michael Niedermayer wrote:
> On Tuesday 12 February 2002 17:25, Arpi wrote:
[...]
> > >
> > > the current fastmemcpy is perhaps not optimal for mem -> agp/pci copies
> > > because it assumes that read/writes go over the same bus, so it simply
> > > reads a few kb (they are in the L1 cache after that) and then reads
> > > them again and stores them (the trick is from amd's manual not from me
> > > ...) if both source &  destination are in main mem then the bus will be
> > > 100% busy cuz the data is read from L1 cache on the read/write pass and
> > > the accesses will be sequential instead of alternating between 2
> > > locations so its faster but if the destination is on an agp card than
> > > the agp bus will be idle 50% of the time and the mem bus will be idle
> > > the other 50% ... pretty bad allthough i didnt benchmark the mem->agp
> > > variant so its just theory ...
> >
> > hmm. try it, i wanna see a 2 times faster agp_memcpy() :)
>
> simply comparing the asm memcpy version in vo_dga against fastmemcpy
> ./mplayer -nosound
> ~michael/videos/Lorenna_McKennit-Mummers_Dance-Mononoke_Hime-gabucino.avi
> -vo dga -benchmark
>
> fastmemcpy:
> BENCHMARKs: V:   8.666s VO:   2.516s A:   0.000s Sys:   4.344s =   15.526s
> BENCHMARK%: V: 55.8162% VO: 16.2078% A:  0.0000% Sys: 27.9760% = 100.0000%
>
> vo_dga asm:
> BENCHMARKs: V:   8.726s VO:   2.019s A:   0.000s Sys:   4.372s =   15.117s
> BENCHMARK%: V: 57.7245% VO: 13.3552% A:  0.0000% Sys: 28.9203% = 100.0000%
> [...]
>
> ill write a agp_memcpy ...
done, heres the benchmark for my p3 at 500

./mplayer -nosound 
~michael/videos/Lorenna_McKennit-Mummers_Dance-Mononoke_Hime-gabucino.avi -vo 
dga -benchmark
rep_movsl (the asm code from vo_dga.c)
BENCHMARKs: V:   9.189s VO:   2.134s A:   0.000s Sys:   4.992s =   16.316s
BENCHMARKs: V:   9.232s VO:   2.162s A:   0.000s Sys:   4.625s =   16.019s

fastmemcpy
BENCHMARKs: V:   9.415s VO:   2.836s A:   0.000s Sys:   4.631s =   16.883s
BENCHMARKs: V:   9.213s VO:   2.726s A:   0.000s Sys:   4.700s =   16.639s

mem2agpcpy
BENCHMARKs: V:   9.189s VO:   2.033s A:   0.000s Sys:   4.654s =   15.877s
BENCHMARKs: V:   9.272s VO:   2.055s A:   0.000s Sys:   4.563s =   15.890s

[...]
Michael