Hi Andre,

For those interested, please find attached a patch for 0.11.4b which includes some optimisation of the layer3 requantisation function.

Thanks!

Its slightly rough around the edges at this stage, but the brief highlights are:

• stripped x86 madplay binary reduced in size from 109k to 77k

This is an impressive accomplishment; it's healthy to be reminded that more than 30% of MAD consists of data.

• peformance improves by approx. 3 percent (although my previous

imdct_s() optimisation is included as well which helps towards this).

I don't seem to be getting as much of an improvement... on my Celeron 500, I'm only seeing a difference of about 0.3 MHz (28.0 vs 28.3). That's with --disable-debugging --enable-speed.

On my StrongARM 220, the difference is about 0.7 MHz (52.9 vs 53.6) with just --disable-debugging.

• audio quality should hopefully be improved ??

Unfortunately, a quick compliance check shows otherwise:

RMS Level 4.302e-03 Max Diff 4.002e-02 RMS: NOT compliant

although it still sounds pretty good.

Also, I got the following while decoding the compliance test data:

Oops, cube root error: -8388608

Still, I think you may be on to something good.

Cheers, -rob