[XviD-devel] [CVS commit] devapi4 -- Dmitry iDCT SSE2 code back

[Edouard Gomez]

Michael Militzer (michael at xvid.org) wrote:
> 1)    Skal's   fdct    is   much    more   accurate    than   Dmitry's
> implementation. Don't know how speed is in comparison, but I'd go with
> Skal's code.

Fine
 
> 2) Do _NOT_  make any idct different from Walken  the default idct for
> any platform (especially not for  a popular platform): this would give
> idct mismatch then. It  doesn't matter whether the idct implemetations
> are IEEE1180 compliant or not: there is still a mismatch (which can be
> very visible). The  only way we can be sure that  all XviD content can
> at least  be played back correctly  with XviD itself is  by only using
> Walken idct in XviD  and nothing else. If we want to  make use of SSE2
> for the idct  and gain some minor speed  improvement, we should rather
> try to create an SSE2 version of  the Walken idct. Refer to my mail to
> the team list for further details about the idct mismatch problem...

I won't do any change w/o your permission in SSE2 functions for a simple
reason: i can't test it.

well  bochs  isn't  really a  solution.   i've  still  to find  a  guest
distro[1] that let  me transfer files between the host  OS and the guest
OS), and i even ignore if all SSE2 opcodes are actually implemented.

[1] i  need a  distro that  is small (because  installing a  huge distro
would   take  hours   in  bochs)   and   ships  usual   shell  stuff   +
gcc+nasm+ssh(d). The problem  is that i don't find  such a distribution,
all source based distros are  simply overkill (sourcemage and gentoo) as
the first thing they ask me to do is compiling sshd (too long)

-- 
Edouard Gomez