MAD performance

Hi!

I want to a little discuss this topic - MAD performance.

1. The reason is - usage of MAD in embedded systems. In this case memory
   usage and cpu cycles consumption are become critical issues. Rewriting
   of 7-8 functions on asm can provide good compromise between software
   complexity and decoder performance. Yes it makes code platform
   specific - but I hope it is acceptable way for embedded systems.

2. some current results. Using such way I've ported MAD to 2 dsp
   platform: TI tms320vc55xx and 3DSP sp3r5m.
   Memory:
   Simple libmab + minimad application can be fitted in less than 16 000
   DSP instructions and all data memory usage: constant tables +
   buffers + stack takes less than 64 kBytes.
   Cycles:
   there are functions with fixed "cycles requirements" - independent
   of input bitsream, slightly depended and highly.
   good example of first type is frame synthesis - mad_synth_frame()
   and last type - III_huffdecode();
   now in average for these dsp synthesis takes ~100-120k cycles

3. interest. And finally I wanna ask only one thing:
   Is it possible to improve III_huffdecode() function in meaning of
   cycles usage?
   because right now average frame size of input mp3 bitstream is
   300-400 bytes (128 kbps / 44 frames ) - and to decode this 350 bytes
   III_huffdecode() spends ~250k+  cycles! this is real bottleneck in
   calculation process. And at higher bitrates cycles consumption just
   growing up.
   
   What ways can you propose to overcome it?
   for example scenario:
   - "unpack necessary tables" : ~50k cycles
   - "look up" decoding process: less than ... cycles

   any proposal are highly welcome.
-- 
Best regards,
 Grigory                          mailto:[email protected]