allnor@tele.ntnu.no (Rune Allnor) wrote in message news:<f56893ae.0307050903.7598ce29@posting.google.com>...
> Jerry Avins <jya@ieee.org> wrote in message news:<3F06DD76.2AFBE9D7@ieee.org>...
> > Rune Allnor wrote:
> > >
> > > ... I found the error. The code
> > > we had got from the teacher used the symbol "l" (lowercase L)
> > > as an internal indexing variable. Even worse, it was used only once
> > > in the code, in a setting like
> >
> > Help me avoid sleepless nights. If the only appearance of the offending
> > variable was in
> > >
> > > for k=l to m do
> > >
> > What was the starting value of k when the line was executed?
>
> Yep, you're right. l had to be initialized somewhere. As far as I remember
> it was initialized at the very start at the procedure.
Having saved your good night sleep, I'll try to secure your afternoon
nap as well: "l" was also updated at the end of the foor loop.
Rune
Reply by Erik de Castro Lopo●July 5, 20032003-07-05
Matthew Donadio wrote:
>
> It is more likely than you think.
I agree.
What about using the rather interesting result that the inverse
fft can be calulated for A, by (Matlab/Octave notation):
0) B = imag (A) + i * real (A)
1) C = fft (B) / N
2) D = imag (C) + i * real (C)
where D is the ifft of A.
If both this methods and the other method of calculaing the
inverse fft works and their results agree in the frequency
domain then it pretty unlikely that its wrong.
Erik
--
+-----------------------------------------------------------+
Erik de Castro Lopo nospam@mega-nerd.com (Yes it's valid)
+-----------------------------------------------------------+
"In civilian equipment, such as computers, the number
of components alone makes miniaturization essential if
the computer is to be housed in a reasonable-sized
building." Electronics Oct. 1, 1957, p. 178
Reply by Matthew Donadio●July 5, 20032003-07-05
On Fri, 04 Jul 2003 22:58:00 -0400, robert bristow-johnson wrote:
> one way i would recommend testing an FFT is to take any arbitrary input, FFT
> it, then inverse FFT, and finally subtract the original input from the
> result. it's unlikely that both the FFT and iFFT will err in a way to undo
> each other's mistakes.
It is more likely than you think. When I was working on some FFT code I
was able to introduce a few errors that the above method did not pick up.
My other post in this thread has the reference for a better method.
basically, you test some of the properties of the FFT to verify
correctness.
--
Matthew Donadio (m.p.donadio@ieee.org)
Reply by Matthew Donadio●July 5, 20032003-07-05
On Fri, 04 Jul 2003 17:23:36 -0700, Bob wrote:
> However how does one test a complex FFT to make sure it is working? Is
> there a datafile that could be generated and what would the output
> result look like?
Check out this paper:
Funda Ergun, "Testing multivariate linear functions: Overcoming the
generator bottleneck, Proc. 27th ACM Symposium on the Theory of Computing,
407-416 (1995).
This is the method that the FFTW package uses.
There are a lot of seemingly good tests that will fail to produce failures
for subtle errors. The above method tests the linearity, impulse
response, and shift properties of the DFT.
--
Matthew Donadio (m.p.donadio@ieee.org)
Reply by Rune Allnor●July 5, 20032003-07-05
Jerry Avins <jya@ieee.org> wrote in message news:<3F06DD76.2AFBE9D7@ieee.org>...
> Rune Allnor wrote:
> >
> > ... I found the error. The code
> > we had got from the teacher used the symbol "l" (lowercase L)
> > as an internal indexing variable. Even worse, it was used only once
> > in the code, in a setting like
>
> Help me avoid sleepless nights. If the only appearance of the offending
> variable was in
> >
> > for k=l to m do
> >
> What was the starting value of k when the line was executed?
Yep, you're right. l had to be initialized somewhere. As far as I remember
it was initialized at the very start at the procedure.
> ...
> >
> > Yep, Jerry and Randy, even if the ADCs in these applications sample
> > real-valued data, after the temporal transform I regard the spatial
> > data in each frequency vector as "complex data", no different than
> > if they should be provided by a complex-valued ADC.
>
> After the transform, that's what it is. The storage needed for N points
> tells the whole story.
>
> > < running for cover... >
>
> Chicken!
Better safe than sorry...
Rune
>
> Jerry
Reply by Jerry Avins●July 5, 20032003-07-05
Rune Allnor wrote:
>
> ... I found the error. The code
> we had got from the teacher used the symbol "l" (lowercase L)
> as an internal indexing variable. Even worse, it was used only once
> in the code, in a setting like
Help me avoid sleepless nights. If the only appearance of the offending
variable was in
>
> for k=l to m do
>
What was the starting value of k when the line was executed?
...
>
> Yep, Jerry and Randy, even if the ADCs in these applications sample
> real-valued data, after the temporal transform I regard the spatial
> data in each frequency vector as "complex data", no different than
> if they should be provided by a complex-valued ADC.
After the transform, that's what it is. The storage needed for N points
tells the whole story.
> < running for cover... >
Chicken!
> Rune
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply by Rune Allnor●July 5, 20032003-07-05
stenasc@yahoo.com (Bob) wrote in message news:<20540d3a.0307041523.7f672cd4@posting.google.com>...
> Hi Folks,
>
> A few weeks ago, I posted a query on FFT scaling.
> After your kind replys, I got it working. I was
> able to get the real only FFT to work. By this
> I mean, there was no imaginary data input.
>
> However how does one test a complex FFT to make sure
> it is working? Is there a datafile that could
> be generated and what would the output result look
> like?
When I went to college, the DSP teacher handed out a pascal code
for complex-valued FFT. What we got was the source code printed
on a piece of paper. The teacher said that was the routine
everybody used who didn't write commercial code, so I had quite
high expectations when I sat down typing it into my own computer.
However, I didn't get anything sane-looking out when I tested the
routine. I tried to use one single sine located exactly on one
of the Fourier frequencies, but no such thing as a line spectrum.
I checked the code, I had written everything down as printed, I
asked the professor, but once he heard I couldn't find any typos
he couldn't help me, nor could he suggest where to look for
errors... quite frustrating. I left the code alone and went on
with other stuff.
After a few weeks I thought it was too bad not having a working
FFT routine, so I decided to give it another try before I attempted
coding the whole thing from scratch. I found the error. The code
we had got from the teacher used the symbol "l" (lowercase L)
as an internal indexing variable. Even worse, it was used only once
in the code, in a setting like
for k=l to m do
In the printed text it was impossible to distinguish from the
number 1, which I had written when I type copied the code. Once
that blunder was corrected, the code worked just fine. I have never
used lowercase L as variable name in any of my code ever since.
> Also...In what sort of application would a complex fft be used?
All over the place. First, with an FFT routine that accepts complex
data as input, you have the IFFT routine for free. The difference
between the FFT and IFFT is basically a scaling of the result.
You need to pay some attention to phase conjugates as well, but
that's details. The only reason I can see for writing specialized,
real-only FFTs, are high-performance real time systems.
In some types of applications the complex forward FFT is very useful.
I have worked with data that consist of time series recorded by
sensor arrays. Lots of algorithms that are useful for these kinds of
data can be formulated as 2D DFTs. While the 2D DFT can be implemented
as two 1D DFT pairs, time <-> frequency and aperture <-> wavenumber,
and the transform pairs commute (i.e. it makes no formal difference
what order one chooses to compute the time to frequency transform and
aperture to wavenumber transform), most algorithms are based on first
transforming the real-valued data from time to frequency, and then
compute some sort of transform of the now complex-valued spatial
data contained in each (frequency, space) vector.
Yep, Jerry and Randy, even if the ADCs in these applications sample
real-valued data, after the temporal transform I regard the spatial
data in each frequency vector as "complex data", no different than
if they should be provided by a complex-valued ADC.
< running for cover... >
Rune
> Thanks alot
> Bob
Reply by ●July 5, 20032003-07-05
robert bristow-johnson <rbj@surfglobal.net> wrote:
>one way i would recommend testing an FFT is to take any arbitrary input, FFT
>it, then inverse FFT, and finally subtract the original input from the
>result. it's unlikely that both the FFT and iFFT will err in a way to undo
>each other's mistakes.
Ah yeah, that's one I've done before too.
Robert
www.gldsp.com
( modify address for return email )
www.numbersusa.com
www.americanpatrol.com
Reply by ●July 5, 20032003-07-05
stenasc@yahoo.com (Bob) wrote:
>Hi Folks,
>
>A few weeks ago, I posted a query on FFT scaling.
>After your kind replys, I got it working. I was
>able to get the real only FFT to work. By this
>I mean, there was no imaginary data input.
>
>However how does one test a complex FFT to make sure
>it is working? Is there a datafile that could
>be generated and what would the output result look
>like?
I typically code my FFT's in C first, before implementing on a DSP.
Then, I either get a hold of a known data set input and output, and
confirm against that. Or else, you can input a clean sine wave (
fabricated in C ), so that it resides on a single bin ... and confirm
that it ends up there, and of the correct amplitude.
Once the C version works, the DSP assembly version can be confirmed
against it.
>
>Also...In what sort of application would a complex fft be used?
I know there are some applications that require both the real and
imaginary info. What comes to mind also is that the complex fft can
be used to perform 2 real fft's.
Regards,
Robert
www.gldsp.com
( modify address for return email )
www.numbersusa.com
www.americanpatrol.com
Reply by robert bristow-johnson●July 5, 20032003-07-05
In article BB2BA8E8.1EF6%rbj@surfglobal.net, robert bristow-johnson at
rbj@surfglobal.net wrote on 07/04/2003 21:58:
> In article 3F062329.31186F88@ieee.org, Randy Yates at yates@ieee.org wrote
> on 07/04/2003 21:00:
>
...
>>
>> One very simple test would be to generate a complex
>> sinusoid, x[nT] = r*e^{i*2*pi*f*n*T}, where T is 1/Fs
>> Fs is the sample rate, and f lies between -Fs/2 and +Fs/2.
>> If you use a rectangular window you should see a sinc-like
>> patter whose peak moves around according to f.
>
> except if the frequency of the complex sinusoid is mid-bin. i.e. f = n*Fs/N
> where N is the FFT size and -N/2 < n < N/2 .
i meant to add that you don't see the sinc but you'll see a single bin spike
at the nth bin.
>>> Also...In what sort of application would a complex fft be used?
>>
>> There is a way to perform an FFT on 2N point real points using an N-point
>> FFT.
>
> yeah, you can't work out the details but the odd samples go into the
> imaginary parts of the FFT input and then you have to do a little massaging
> on the output.
and here i meant to say that you *can* work out the details.
whatever.
r b-j