[XviD-devel] recent mmx patch (xvid_bench results)
jmtest at voila.fr
jmtest at voila.fr
Wed Apr 14 09:20:44 CEST 2004
So the conclusion is: send my mem_transfer to trash, it's the best place for it.
Sorry for this one.
The H263 rewrite should be better (and speed improvement visible in the bench).
But I missed that cmov is only available since pentium pro... I will rewrite this part.
Since RC4 we have an exhaustive loop in test_quant(). This goes through all the quant,
all the input values and compare the C output with the asm output.
I hope this will avoid the case of bad functions that pass the bench...
Regards,
Jean-Marc
> Message du 13/04/04 à 23h05
> De : Edouard Gomez <ed.gomez at free.fr>
> A : xvid-devel at xvid.org
> Copie à :
> Objet : Re: [XviD-devel] recent mmx patch (xvid_bench results)
>
> suxen_drol (suxen_drol at hotmail.com) wrote:
> > err.. i havent had time to weed through the transfer code.
>
> The error was that the new code was doing the sub operations on bytes,
> to save cycles, but that isn't possible because of overflows, and signed
> 8 to 16 conversions using a simple unpack operation.
>
> > however the the crash was casused by some conditional moves. the cmov
> > instructions is generally supported by cpus with MMXEXT. my cpu which
> > only has MMX.
> >
> > * quant_h263_intra_mmx has cmovs. this should be renamed to _xmm and
> > version developed without the cmovs called '_mmx'.
> >
> > * dequant_h263_intra_mmx has cmovs. we already have an _xmm version, and
> > i assume it is faster than the _mmx (somebody may want to check).
> > therefore, the cmovs should be replaced.
>
> Ok doki, we'll just place a test somewhere to avoid that sigill problem.
>
> --
> Edouard Gomez
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel
>
------------------------------------------
Faites un voeu et puis Voila ! www.voila.fr
More information about the XviD-devel
mailing list