[XviD-devel] asm

[Michael Militzer]

Hi,

> On Fri, 20 Dec 2002 01:25:07 +1030 Radek Czyz <radoslaw@syskin.cjb.net>
wrote:
>
> > Hello,
> >
> > Recent asm optimizations made me look at the current state of sad16v
> > function.
> > I can see that the only fully asm-ed sad16v is xmm-ed one. It's fine
> > with me (although I wouldn't mind having it athlon-optimized as well),
> > but for every non xmm-ed architecture sad16v will be pretty slow.
>
> xmm == athlon.

I guess he's referring to the Athlon optimized code I commited 2 days ago
which was contributed to us by Jaan Kalda. This code is still xmm code but
has been especially optimized for the Athlon pipeline and might therefore
not work that well on P3s. xvid_bench shows that the Athlon optimized
functions are really faster than their normal xmm equivalents, however in
real encoding I measured no noticable difference...

> > Please keep in mind that sad16v is the base of current P-frame motion
> > search. It's actually the only sad function used there (apart from
> > sad8, but sad8 is used much less).
>
> looking into it...

we already have a sad16v_mmx function for a long time now. Looking into cvs,
it tells me that a guy called Isibaar has commited sad16v_mmx code on Sep 24
;-)) So what's the problem with it?

bye,
Michael