DSP Stuff

[Harvey Devoe Thornburg]

> 
> >Beware, there are two EZ-Kit's: one with the ADSP2181, which is about $100.  
> >This is a fixed point DSP. Then there's another one with the floating point  
> >Sharc DSP, but it is more expensive. The floating point capabilities should  
> >make it easier to program, but I have no experience with that.
> >

Floating point is really the most natural for audio, because our 
perception of frequencies and amplitudes is logarithmic, and also 
because the most efficient attempts to simulate floating point
with long ints in fixed point instructions almost always generates 
unreadable code (reliance on shifts etc.) for ex. to divide by 255 
(common in graphics and image processing) on a fixed point 
machine it is easier use the 
relation 1/255 = 1/256 + 1/256^2 + 1/256^3 + ... and use shifts and
accumulates.  I have seen, but thankfully never had to write, this kind
of stuff in audio applications.  Do you really want to be thinking 
about this stuff (hint: the compiler never helps you) or thinking 
about the algorithm?

One thing might be disturbing with floating point is the generation 
of underflow errors. let's say you're using a one-pole filter for an 
envelope generator (like for a 303 model) and there's a long time 
between gates, then sooner or later you will underflow. 
some processors (Pentium) react very poorly to this, causing much 
slowdown (or maybe this is due to Microsoft DirectSound standard?)  

Traditionally the audio industry has used Motorola 56K's but many new
companies are transitioning to ADSP21061 or 21160 (600 mflops) sharc.
The main issue seems to be floating point, the availablity of
C development environments, and parallelization (hardware support
for multiprocessing stuff like semaphores). I think Digidesign still 
uses 56K (which isn't much consolation for the $$$$$ for protools systems)!
There is some effort underway to port Csound to the sharc but I am
not up on this.

Something which interests me is the TI stuff like TMS320C60 floating
point series.  Performance is supposed to be in excess of 1000 mflops.
But the evaluation kit costs $1500 not to mention 3rd party development
tools.  I've been considering it (just to be different and to work
with a more powerful machine) but I don't know many in industry who
are adopting TI.

--Harvey