[XviD-devel] Re: xvid_fdct_sse + qpel ASM

[suxen_drol]

On 20 Jan 2003 16:17:56 +0100 skal <skal@planet-d.net> wrote:

> 
> 	Re,
> 
> 	Ok, ok, i was curious, so I've quickly hacked
> 	the MMX 'punpck' soup mentioned in my previous
> 	mail. Here's what's to add in fdct_sse.asm
> 	for the MMX fdct. It's 10% slower than SSE,
> 	(but still 30% faster than current XVID one,
> 	it seems;)
> 
> 
> 	bye!

thanks skal. you really know your stuff. its been fun having you about!

ive committed your new fdct to cvs (along with michaeln's c/nasm
simple_idct). the code is not enabled by default, as ive only performed
mmx testing. it works well. accuracy is almost identical to the current
ap-922 derived fdct.

to enable skal's fdct / simple idct, edit xvid.c:xvid_init(), and add
	fdct = xvid_fdct_mmx;	or 	fdct = fdct_fdct_xmm;
	idct = simple_idct_c;	or	idct = simple_idct_mmx2;

btw. do not use simple_idct_mmx; it expects permutated data.

cya!
-- pete;