[Clam-devel] Configuring Spectrum Ports

[Xavier Amatriain]

Pau Arumí wrote:
> But for this concrete problem, i'd rather do the much needed "one 
> spectrum representation per class" refactoring. I'm willing to help on 
> that. Is this refactoring going in your direction?
I am with you on this one and, yes, I am working in this direction.

My profiles have shown that in an SMS transformation almost 10% of the 
time is spent in spectrum conversions
(and that's after all my work in optimizations and using the 
CLAM_Optimize flag). As I am now, the only way
around this issue is restricting the use of spectrum representations to 
only the most efficient one and performing the
conversion in the minimum amount of loops.

For instance, in the FFT we first compute the FFT, then loop through the 
results to convert to complex format, and
then loop again to convert to mag/phase.

I will give more details as soon as I have something worth sharing 
because my many efforts have been quite
deceiving so far (just got an overall 15% improvement after many "hacks").