[XviD-devel] SAD cache

[Christoph Lampert]

On 11 Mar 2003, skal wrote:
> > Do you have any numbers on its effect? Are there really that many 
> > researches? I guess only when checking several predictors they often are 
> > identical.
> 
> 	so far I've seen, it's 10-20% hits for 'regular' fast moving
> 	scenes. For slow moving ones (where the diamond/square search
> 	does not wander too far), it raises up to 20-30%. Conversely,
> 	 for 'hard' scenes, it drops to 5%.

Does that include double checking predictors? Or only during diamond? 
10-20% during diamond I would consider a _big_ problem. I could only 
explain this by many "diagonal" steps: 

123    when going  5->6->3 or similar, we re-check 2 in next step
456    and similar for all others diagonal paths. 
789

However, it might be better to modify DiamondSearch to reduce this (I
guess AdvDiamond already does by going directly to 3 and "remembering" 
that 2 and 6 had been checked in the last step). 

If it's 10-20% of _all_ checks, that's okay, because we check up to 8
predictors or so, many of them will coincide. We could only call SAD-cache
for predictor check and don't use a hash during diamond phase. 

gruel