Compared to JNTFFT, ACM's unidimensional FFT is slower.
Benchmark and results coming soon.
Report was based on a pre-3.0 snapshot.
All is fine in the official 3.0 release.
Ok, that was a quick one : my tests had been done last december with a snapshot version of Commons Math.
Now with the official 3.0 release I'm finally getting performance numbers similar to JNTFFT.
So I'll close this ticket, thank you guys.