In servalan.mailinglist.netbsd-current-users you write:

>On Sun, Jun 20, 1999 at 07:03:18AM +0900, Izumi Tsutsui wrote:

>> Yes, it seems doing many FFTs.

>Hmm...  I was always under the assumption that the FFT was by definition
>an optimized discrete Fourier transform, and thus no floating point was
>used.

You assume incorrectly. :-) "Discrete" in the DFT/FFT means that the frequency
and time spaces are discrete, not continuous, i.e. that you have a signal
sampled at fixed intervals and are interested in its frequency content at
fixed intervals in frequency space.  The signal values themselves are almost
always floating-point.