[MPlayer-dev-eng] Cache optimized rotation.

[David Bateman]

On 4/10/07, Reimar Döffinger <Reimar.Doeffinger at stud.uni-karlsruhe.de>
wrote:
>
>
> Well, the simplest solution that works well enough is always preferred.
> The second patch can probably be also better optimized of there are
> streaming write instruction available that bypass the cache, the first
> version there would have more cache trashing due to the block.


I don't think the first solution should thrash the cache as the tile size
can be selected to fit entirely within one cache line


It also does not need tuning depending on cache size.
> But neither performance nor complexity of a final patch are clear to see
> in both patches, since at least in the first there are several
> optimization possibilities and both probably can be simplified quite a
> bit.
>
>
True that the tuning due to cache size might be an issue. At the moment I've
just limited the size of the tile to be smaller than the smallest cache line
that I know about. In fact it probably therefore makes more sense just to
make that choice permanant and remove all of teh comments about the tuning.
Ok, then given that I'll make up new versions of both trying to take into
account your comments

Cheers
D.