Hi --
--- In "Gonzalo A. de la Vega"
<> wrote:
>
> --- In "kurmannthomas" <kurmannthomas@>
wrote:
> >
> > Hello
> > Im using FFTW3 to transform 9(3 Sensors 3 Axis) arrays with real
> > (double) data to frequency spectrum. The arrays all have the
same
> > length so im only planning one plan using:
> >
> > p = fftw_plan_dft_r2c_1d(N,in,out,FFTW_FORWARD)
> > and then executing on all 9 arrays (the data in in just gets
> > swapped). The results seem to be correct (verified with Matlab
which
> > also uses FFTW3). Now the problem is im getting a conversion
time for
> > a 4096 point FFT of about 4 seconds...this is horribly
slow...even
> > without EABI or Crunch i would have expected the TS-7260 is alot
> > faster.
There is also a way to 'plan' many transforms, if the data for all 9
sets is available. How are you executing the plan, fft_execute(p) I
guess? Also, be sure to use the provided fftw-malloc, to be sure
alignment is right.
Also, how long is your 'plan' step? If you are always working on
the same size & type of data arrays, you can 'plan' ahead of time
and save the plan on disk (see 'wisdom').
> >Now i am not experienced enough with Linux to "install" EABI
> > or Crunch (unless someone has a tar that i can just extract ;))
Now i
>
> You need a EABI or Crunch enabled kernel, so in order to use a tar
> package, you should use bootloader to boot the kernel. You will
> probably end being an experienced user after this, because you'll
need
> to build a couple of things.
> I recommend you build a kernel with FastFPE point,this is not
ANSI,
> but is supposed to be much faster than the standard.
You need a kernel that is compiled to support eabi, and a
distribution that supports eabi also. TS has recently started
supplying these, so at some point soon it will be quite easy,
although needing a reflash rather than just un-tarring.
> > am searching for a integer FFT which should be alot faster...Has
> > anyone else seen these extremly slow times when using FFTW3?
Maybe im
Well, I answered your post in June, recommending a fixed-point FFT
(not the same as an integer), if you could not use eabi and/or
crunch. Sorry, I don't have any pointers handy.
> > doing something wrong?
> >
> > Im going to check with another floating point FFT next week to
see if
> > its an FFTW3 problem, or really the TS-7260. My goal is to be
able to
> > calculate (up to) 9 FFTs with (up to) 16384 points in less than
1 or
> > 2 seconds...a DSP would have no problem with this...
To give you an idea of what is possible, I have found a factor of
12-15 speed up in general FP intensive code using softfloat and
eabi, compared to oabi. With FFTs, others have reported a speedup
of 20.
And just recently, Martin Guy has been working hard getting crunch
instructions working, and reports a speedup of 2.3 to 5.4
MFLOPS with the fftw benchmark, when using a crunch-enabled
compiler. Martin has written it up quite well; See
http://tech.groups.yahoo.com/group/ts-7000/message/8706
For comparison, what do you see for fftw MFLOPs benchmark on your
system?
In any case, you have maybe a factor of 40 speedup available in the
best case, so perhaps you could do (9) 4096 point FFTs in a second,
but almost certainly not (9) 16384 point FFTs in a second. If you
really need that much speed, you probably need a DSP.
Regards, ........ Charlie
Yahoo! Groups Links
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/ts-7000/
<*> Your email settings:
Individual Email | Traditional
<*> To change settings online go to:
http://groups.yahoo.com/group/ts-7000/join
(Yahoo! ID required)
<*> To change settings via email:
<*> To unsubscribe from this group, send an email to:
<*> Your use of Yahoo! Groups is subject to:
http://docs.yahoo.com/info/terms/
|