Michael Jerris wrote:
We have discussed this with David some. The biggest issue on the blackfin relates to floating point math. The main places that you will see this in freeswitch is the codecs, the resample code, and in a couple places in applications like conference. For the codecs, much of the work is already done (like for speex, David's changes are already worked into our tree) but would need some more work. For the applications like conference, it would need small or no modifications. The resample code is the only code that would needs some real work, I think David can comment better than any of us how much work that would take.
Mike
David has been converting some of spandsp to fixed point. Before he converted the DTMF code, that alone was consuming most of the blackfin :-). Now, of course, its a pretty minor activity. If you want Freeswitch to be able to go embedded, you need to have a fixed point build option for any of the DSP functions which are frequently used. I think most floating point outside the DSP functions is probably fairly harmless, though undesirable. In most cases it won't be executed often enough to matter. RTCP statistics calculations would be the exception (if you don't consider those signal processing). I wonder how far Digium have got with fixed-pointifying Asterisk for their Blackfin box? Regards, Steve