[Speex-dev] Speex on a Blackfin DSP

[Jean-Marc Valin]

> Which we are constantly working on. We do have some targetted pieces of C 
> code, that we noticed could be done better, and are in process of working on.

David Rowe actually had some details about optimising Speex here:
http://www.rowetel.com/blog/?p=6

Some of the issues we found:
- Enabling write back helps a lot
- There *probably* is I1 cache trashing because we don't use the I1 SRAM
part

In terms of gcc some of the things that *probably* have the most impact:
- No hardware loops
- No conditional moves
- No MAC instructions
- No parallel loads

The last two mean that an inner product ends up looking like:
load
load
mul
add
instead of
mac | load | load

All that is based on observations from last year. Things might have
changed since then.

	Jean-Marc