[FFmpeg-soc] [soc]: r1489 - rv40/rv34.c

Kostya kostya.shishkov at gmail.com
Sat Dec 1 19:16:37 CET 2007


On Sat, Dec 01, 2007 at 06:05:43PM +0100, Michael Niedermayer wrote:
> On Sat, Dec 01, 2007 at 05:23:38PM +0200, Kostya wrote:
> > On Sat, Dec 01, 2007 at 03:01:53PM +0100, Michael Niedermayer wrote:
> > > On Sat, Dec 01, 2007 at 02:47:01PM +0100, Michael Niedermayer wrote:
> > > > On Sat, Dec 01, 2007 at 03:53:15PM +0200, Kostya wrote:
> > > > > On Sat, Dec 01, 2007 at 01:57:29PM +0100, kostya wrote:
> > > > > > Author: kostya
> > > > > > Date: Sat Dec  1 13:57:28 2007
> > > > > > New Revision: 1489
> > > > > > 
> > > > > > Log:
> > > > > > More generic MC function
> > > > > > 
> > > > > > Modified:
> > > > > >    rv40/rv34.c
> > > > > 
> > > > > As a side effect I think it will be easier to add RV30
> > > > > motion compensation (it uses thirdpel MC)
> > > > 
> > > > for a h.264 draft with 1/3 1/2 and 1/6 pel MC see:
> > > > 
> > > > http://ftp3.itu.int/av-arch/video-site/0005_Osa/q15j08d0.doc
> > > > 
> > > > (note just a random pick iam sure many of the early drafts contained 1/3 pel)
> > > 
> > > and here is one with just 1/3 pel MC
> > > http://ftp3.itu.int/av-arch/video-site/9908_Ber/q15h36d2.doc
> > 
> > maybe upcoming RV5 will use 1/5pel MC
> 
> only if some h.26* draft uses that :)

Then they will leap straight to RV6 
 
> > 
> > as for RV30 I know it uses filter [-1 12 6 -1] for luma and [5 3] for chroma
> 
> from the draft above:
> 1 Fractional pixel accuracy
> 
> 
> In the present coding method, 1/3 pixel accuracy is used to define motion
> vectors.  In the figure below, capital letters indicate pixel positions
> whereas lower case letters represent interpolated 1/3 pixel positions.
> 
>   A     B     C     D
> 
> 
>   E     F a b G     H
>         c d e
>         f g h
>   I     J     K     L
> 
> 
>   M     N     O     P
> 
> For interpolation of a,b,c,d,e,f,g filters with one dimensional impulse
> response (1,  (-1 12 6 -1)/16 and (-1 6 12 -1)/16) are used in horizontal
> and vertical directions as appropriate.  To interpolate h a 'stronger'
> filter with impulse response (0 6 9 1)/16 is used.

Hmm, I will check what RV30 uses for that case.
Thank you for pointing to that doc.

[...]
> For fractional pixel interpolation for chroma prediction, the two tap
> filters (5 3)/8 and (3 5)/8 are used as appropriate for all fractional
> pixel positions (a - h) above.

No surprises here at least. 
 
> [...]
> -- 
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB




More information about the FFmpeg-soc mailing list