[Speex-dev] Speex on TI C6x, Problem with TI C5x Patch

[Jean-Marc Valin]

> About 42 MIPs peak for encode/decode, complexity 1, 8kbps, no preprocessor 
> or jitter buffer or echo canceller (yet).

Good. About the preprocessor, jitter buffer and echo canceller, I should
warn you that they haven't been converted to fixed point yet.

> No, they have always produced the same results.  The C54x seems to take 
> about 5x as many cycles to do the task.  I can see that 32x32 multiplies are 
> being done by library calls rather than inline, so I will play with this a 
> bit more.  Others have used C54x successfully, so significant improvement is 
> possible.

I have made sure that Speex does not require a 32x32 multiplier at all.
Only 16x16 with result in 32. Most of that are actually MACs, so this
should be very efficient on the C54x. Even for the C55x, I assume that
the current code generates a 32x32 multiplication followed by an add
instead of just a MAC, so I'm sure there's a lot to gain on the C55 too.

	Jean-Marc

-- 
Jean-Marc Valin <Jean-Marc.Valin at USherbrooke.ca>
Universite de Sherbrooke