[XviD-devel] RD stuff and more

[Michael Militzer]

Hi again,

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

> Michael Militzer wrote:
>  > Well, that's it basically. If you need more
> > explanations, don't hesitate to ask...
> 
> Allrighty, I got it and (after a long struggle lol) confirmed that this 
> is exactly what the code does :)
> 
> The refinement is more aggresive at not checking diagonal positions than 
> I expected :) but it's probably really fast. I've finished my version of 
> it, I hope it will work - it combines halfpel and qpel refinement for 
> extra speed and doesn't require any extra checkcandidate code (or, even 
> worse, any new checkcandidates).
> 
> I wonder what is faster and gives better quality - full sad-based 
> refinement, or fast RD-based one :) If everything works, I'll have the 
> answer tommorow.

Hm, I'd expect fast RD-based to perform better than full SAD refinement,
but also still to be slower. But I guess this would be a result you could
very well live with, right? ;-)

> I think I'll also create a fast version of normal halfpel refinement - 
> should be useful :)

yes, but: since the halfpel refinement (whether in qpel or normal bilinear
mode) is not done block-based but on interpolated images, it's already
pretty fast. A fast refinement scheme can 'only' help to avoid some SADs
but doesn't speed up the interpolation (at least not how things are
currently done in XviD). So the speed gain from a fast half-pel refinement
will be most likely not be too impressive, so only useful for someone who
needs maximum speed. For qpel refinement, the fast qpel scheme gives a
significant speed-up at very low quality loss - so therefore I'd even
recommend it to the highly quality-oriented user ;-)

Michael