[FFmpeg-devel] [PATCH] aacenc: optimize cost cache.

[Reimar Döffinger]

On Wed, Mar 02, 2016 at 07:19:08PM +0000, Rostislav Pehlivanov wrote:
> > -        entry->cb = cb;
> > -        entry->rtz = rtz;
> > +                                       cb, 1.0f, INFINITY, &entry->bits,
> > &entry->energy, 0);
> > +        *cache_state = cb;
> >      }
> > -    if (bits)
> > -        *bits = entry->bits;
> > -    if (energy)
> > -        *energy = entry->energy;
> > +    *bits = entry->bits & 0xffffff;
> > +    *energy = entry->energy;
> >      return entry->rd;
> >  }
> >
> >
> The whole point of the function was to be a universal caching system for
> every single scalefactor band which could be used in other parts of the
> code (e.g. intensity stereo, mid/side, prediction, pns, etc.). Granted,
> caching hasn't been implemented there yet. But by reducing the number of
> arguments and assuming everything uses the same value for lambda and uplim
> removes any possibility of the caching system being used anywhere outside.

Which is a reason why such design considerations really should be
in a comment near the relevant code, because otherwise the code as-is
doesn't make any sense and looks pointlessly obfuscated to anyone
reading it.
That said, those changes are mostly a left-over from experiments
as I couldn't figure out why the code is so slow.
I can look into it, but it might be worth for someone to look
more closely into how this implementation can be made more
CPU-cache-efficient as I only optimized the cost of clearing
but not the CPU cache hit rates when we actually have a match.