Re: [Fwd: Re: rfftw slower than fftw for "bad" size arrays]

[John W. Eaton]

On 29-Jan-2004, Dmitri A. Sergatskov <address@hidden> wrote:

| David Bateman wrote:
| 
| > I had a quick look at want is involved in converting to FFTW 3.0.1 and it
| > seems pretty easy to implement, as long as we stick to FFTW_ESTIMATE, and
|                                               ^^^^^^^^^^^^^^^^^^^^^^^^^^
| Perhaps it should be a fallback solution. It seems it should be able to
| check if "magic" file exists and use FFTW_ESTIMATE if it does not.
| One issue here is whether octave should drop its own fft stuff and
| require fftw installed. If it should carry its own copy of fftw
| then what shall we do with "wisdom" with binary distributions?

I like the idea of having fftpack available as a backup in case fftw
is not installed.  Since fftpack is small, I don't think it is much of
a problem to keep it.  I'd rather not force people to go get a complex
package for a core function like fft.  But if we are going to drop
fftpack and require fftw, then we have to make fft and related
functions optional in Octave.  As much as possible, I want people to
be able to download the Octave source tar file (just one file) and be
able to unpack it and run configure and make and generate a binary
that will at least run without having to chase around the web for
other packages.  Currently Octave's configure script will fail if
readline is not available, but we could fix that to just issue a
warning at the end of the configure run instead of having it be fatal.

| Since generating wisdom (http://fftw.org/fftw-wisdom.1)
| fftw-wisdom -v -c -o wisdom
| will take "many hours" it probably should be done as a post-install
| step.

We have the same problem for ATLAS, and the binary distributions tend
to build a sub-optimal version that works reasonbly well for a wide
range of systems.  If people want better performance, they can always
rebuild the ATLAS library or compute the proper parameters for fftw.

| Anyway, shall we go back to octave-maint.?

Yes, please.  I'm copying this message to the list.

jwe