The Cyber-Spy.Com Usenet Archive Feeds Directly
From The Open And Publicly Available Newsgroup
This Group And Thousands Of Others Are Available
On Most IS NNTP News Servers On Port 119.
Cyber-Spy.Com Is NOT Responsible For Any Topic,
Opinions Or Content Posted To This Or Any Other
Newsgroup. This Web Archive Of The Newsgroup And
Posts Are For Informational Purposes Only.
From: Chuck Simmons
Organization: You jest.
X-Mailer: Mozilla 4.61 [en] (X11; U; Linux 2.0.33 i586)
Subject: Re: fft in hc11
Date: Mon, 14 Oct 2002 16:21:25 GMT
NNTP-Posting-Date: Mon, 14 Oct 2002 09:21:25 PDT
Ken Smith wrote:
> In article <3DA67B43.C5E3AD2F@earthlink.net>,
> Robert Baer wrote:
> >Ken Smith wrote:
> >> If you know the input frequency exactly you know the harmonics. There is
> >> no need to make the DFT produce anything but the amplitudes at the
> >> frequencies you care about. Doing a 1 million point FFT is a lot harder
> >> than making 10 or 20 channels of DFT for 1 million points.
> >> --
> >> --
> >> firstname.lastname@example.org forging knowledge
> > Million point FFT harder???
> > As simple as 1000 point FFT; i work with multi-million point
> >algorithms, and faster than almost anybody else in the field.
> > On speed, once i saw an ad by TI on some "hardware" 1024 point FFT
> >that stated its speed, and my *software* had it beat!
> Yes a Million point FFT is "harder" for the following reasons:
> Every pass through the outer loop of the FFT loses you about 1 LSB of the
> mantissa of the numbers involved. If you do a 10 result DFT you can use
> a shorter floating point format. Remember this was a microcontroller not
> a DSP he was using.
> Doing a 10 answer DFT takes on the order of 10*N operations. Doing a
> million point FFT takes on the order N*log(N) operations. Long before you
> get to the million point version, the 10 anwer version beats the FFT.
> If you do an FFT you need to keep all of the numbers in memory. For the
> 10 answer DFT you just need to keep 10 answers and a few other values in
> memory. You can process the values as they go by.
Precisely. Always fit the algorithm to the problem. One size does not
20 years ago I had a problem where the micro had to approximate a DFT
for one frequency. The multiply instruction was so slow that I decided
it could not be used. I investigated alternate transformation kernels
which might avoid any multiply operations in computing the one frequency
I needed (I needed both amplitude and phase). I hit on a modified Walsh
kernel since the Walsh transformation only requires multiplication by +1
and -1. This worked very well and the product development group which
picked up the technology shipped product with this very simple DFT
approximation as part of key power up calibration of the device.
... The times have been,
That, when the brains were out,
the man would die. ... Macbeth
Chuck Simmons email@example.com
Go Back To The Cyber-Spy.Com
Usenet Web Archive Index Of
The sci.electronics.design Newsgroup