[FFmpeg-devel] [PATCH] VP8 MMX optimizations (MC and IDCT dc_add)

[Frank Barchard]

>On Thu, Jun 24, 2010 at 6:33 PM, Jason Garrett-Glaser <
darkshikari at gmail.com> wrote:

> > Now with 8x8 intra pred modes and non-broken line endings.  Did I
> > mention this makes h264 faster too?
> >
> > Dark Shikari
> >
>
> And one more SSSE3 function, because pshufb is amazing.
>

yes, but 4 cycles on atom.

On Tue, Jun 22, 2010 at 3:29 PM, Michael Niedermayer <michaelni at gmx.at>
 wrote:
>> +    punpcklbw mm2, mm6
>> +
>>> +.nextrow
>> +    ; first tap
>> +    pshufw    mm3, mm7, 0x0                ; splat first coeff

> are you sure all these pshufw are faster than reading them from a table?

Ronald's pshufw follow unpacks, so the tables would be twice as big.
On Atom, cache and memory are dirt slow, so keeping the tables small makes
sense.