[XviD-devel] Fwd: is a bug?

[Michael Militzer]

Hi Radek,

I think the (0,0) case the guy describes is just a special case and I
think his proposed patch is not correct (as I wrote him already in private
mail). I think his point is the following:

_If_ the prediction value is the zero vector and this vector is chosen as
the best predictor from our initial predictor set and therefore becomes the
center of the diamond in our regular search, it is possible that we're
performing the same diamond (around the (0,0) predictor value) again in
extsearch.

Now the correct fix for this (imho) would be to store the motion vector
that was used as center of the first diamond in our regular search and
then check in extsearch whether this former diamond center vector is 
identical to our prediction vector. If yes, then skip the first diamond 
search in extsearch. This check could/should replace the current check
against backupMV (if I'm not mistaken ;-))...

Michael

Quoting Radek Czyz <syskin at ihug.com.au>:

> Michael Militzer wrote:
> > Hi all,
> > 
> > I received this 'bug report' today by private mail. Actually, it's not
> > really a bug because it just describes a potential situation where we 
> > might do a diamond search twice but without any bad side effects (apart
> > from the speed loss).
> > 
> > I believe the report is correct and it might indeed be possible to 
> > slightly speed up extsearch by adding some more checks. The question is
> > just whether the additional checks would really make things faster or
> > whether the described case(s) are so rare that no speed-up is achieved.
> > However, the (0,0) case should be rather frequent...
> 
> Hi Michael,
> 
> it's being discussed here: 
> http://www.xvid.org/modules.php?op=modload&name=phpBB2&file=viewtopic&t=2254
> 
> In short, the guy is not exactly right in this particular case - but in 
> general yes, extsearch might be a big waste of time because there is no 
> checks whether we're doing the same thing twice. This is contrary to 
> non-extsearch search, which avoids double-checks beautifully...
> 
> Radek
> 
> PS. SearchP takes so little time at defaults, that even good speedups 
> are hardly worth it :/
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel
>