[FFmpeg-devel] [PATCH] Dsputilize some functions from APE decode 2/2 - SSE2

[Kostya]

On Thu, Jul 10, 2008 at 11:48:14AM +0200, Michael Niedermayer wrote:
> On Thu, Jul 10, 2008 at 11:16:01AM +0300, Kostya wrote:
> [...]
> > +static void add_int16_sse2(int16_t * v1, int16_t * v2, int order)
> > +{
> > +    x86_reg o = -(order << 1);
> > +    v1 += order;
> > +    v2 += order;
> > +    asm volatile(
> > +        "1:                       \n\t"
> 
> > +        "movdqu  (%1,%2), %%xmm0  \n\t"
> > +        "paddw   (%0,%2), %%xmm0  \n\t"
> > +        "movdqa  %%xmm0,  (%0,%2) \n\t"
> > +        "add     $16,     %2      \n\t"
> > +        "movdqu  (%1,%2), %%xmm0  \n\t"
> > +        "paddw   (%0,%2), %%xmm0  \n\t"
> > +        "movdqa  %%xmm0,  (%0,%2) \n\t"
> > +        "add     $16,     %2      \n\t"
> 
> is that faster than:
> "movdqu    (%1,%2), %%xmm0  \n\t"
> "paddw     (%0,%2), %%xmm0  \n\t"
> "movdqa  %%xmm0,    (%0,%2) \n\t"
> "movdqu  16(%1,%2), %%xmm0  \n\t"
> "paddw   16(%0,%2), %%xmm0  \n\t"
> "movdqa  %%xmm0,  16(%0,%2) \n\t"
> "add     $32,     %2      \n\t"
> 
> ?
 
It was the first thing I've tried. It was slower (on Core2).
 
> [...]
> -- 
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
> 
> I have often repented speaking, but never of holding my tongue.
> -- Xenocrates