Hello!
A few replies.
1. What do you mean by "what kind of power should I take into account"?
Is it the power of the signal?
2. Algorithm, not algorythm. Well, if it's for music, there can be also rythms...
Why do you want 33 bands (I mean, why not 18? Why not 49?).
3. Digital processing of signals, by Maurice Bellanger.
4. Slicing your FFT is a good idea. But here are some issues:
4.1. If you want to cut your signal in sub signals, then it will be very tricky
to reassemble them after FFT.
4.2. If you want to cut your signal in FFT passes (it's possible, I did it), then
you have to do one pass at a time. So using for instance 9 processor for a
9 stage FFT (512 points) can be done, but in order to process one stage,
you need the full results of the previous stage. And on top of that, you have
to pass the results of one stage to the next processor memory.
The only possible advantage is if you do some continuous processing.
- Process Sample1.pass1 in processor 1
- Move data from processor 1 to processor 2, then process in parallel
Sampler.pass1 on processor1 and sample1.pass2 on processor2
- Move data from processor2 to processor 3, then from processor1 to processor2
then process in parallel sample3.pass1, sampl2.pass2, sampl1.pass3
Etc. If you do 9 passes (in the case of 512 points), then it will be a headache
and on top of that, I wouldn't be surprised if you get something slower.
Basically, what I understand is that you want to perform fast convolution,
right?
Dora.