Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!apple!gem.mps.ohio-state.edu!ginosko!uunet!crdgw1!ge-dab!sunny!harrison
From: harrison@sunny.DAB.GE.COM (Gregory Harrison)
Newsgroups: comp.dsp
Subject: Re: FFT for integer data
Message-ID: <2382@ge-dab.GE.COM>
Date: 11 Oct 89 19:59:16 GMT
References: <2620@pur-phy>
Sender: news@ge-dab.GE.COM
Reply-To: harrison@sunny.UUCP (Gregory Harrison)
Organization: GE Simulation & Control Systems Dept., Daytona Beach, FL
Lines: 22
Distribution:USA

In article <2620@pur-phy> murphy@newton.physics.purdue.edu.UUCP (William J. Murphy) writes:
>
>2.)  Though this shows my naivety about the details of the FFT, Is there
>     any acceleration available for strictly integer data input versus
>     real input to the FFT routine? i.e. is there an optimization available
>     for integer data?  I assume that regardless of input you will return
>     a real complex spectrum instead of an integer complex spectrum?
>
Due to the nature of integer processing on a computer (native units), 
operating on integers will normally be much faster than operating
on floating-point numbers, especially without a floating point
coprocessor chip.  An integer FFT will return integer spectral data.
An integer FFT will lose approximately 1/2 bit of precision for each
level of butterflys in the algorithm.  Thus a 1024 point integer FFT
will lose approximately 5 bits of precision in the conversion from
time-domain to frequency domain.

Thus floating point processing (which does not have the precision loss
problem in this manner) is preferable for larger FFTs.   

Greg Harrison
My opinions are not intended to reflect those of GE.