RE: fftw on cygwin

classic Classic list List threaded Threaded
4 messages Options
Reply | Threaded
Open this post in threaded view
|

RE: fftw on cygwin

James R. Phillips-2
--- "Billings, Paul"

> FFTW bundled would be a huge improvement for the casual user or more
> importantly, the guy that is evaluating if he can abandon the Matlab
> "maintenance" fees.

Is it possible to compile a generic version that would work on any ia32
processor?  Or will it by necessity use processor instruction set extensions in
it's optimization process?

> Packaging ATLAS as well would be fantastic.  I've seen some of the list
> traffic about the difficulties in doing this.  Sure would be nice if it
> could work, but I don't think it likely.
>

The Debian approach to packaging Atlas looked promising, but turned out to be
too difficult to mimic. I include an option to build Atlas from source in the
cygwin lapack package.

JRP

Reply | Threaded
Open this post in threaded view
|

RE: fftw on cygwin

Paul Billings
I believe a generic version of FFTW would suffice.  In contrast to ATLAS
which does extensive benchmarking at make/compile time, FFTW does run-time
measurements (or can load previously saved measurements or "wisdom").
Whatever instruction set that cygwin is generally targeting should be fine,
although I suppose you could have two versions: one for "barebones" and one
for SSE/SSE2 or whatever is appropriate (I forget the differences).  The
user would simply check one or the other in the cygwin setup application.

Paul

> -----Original Message-----
> From: James R. Phillips [mailto:[hidden email]]
> Sent: Tuesday, August 09, 2005 6:21 PM
> To: [hidden email]
> Subject: RE: fftw on cygwin
>
>
> --- "Billings, Paul"
>
> > FFTW bundled would be a huge improvement for the casual user or more
> > importantly, the guy that is evaluating if he can abandon the Matlab
> > "maintenance" fees.
>
> Is it possible to compile a generic version that would work on any ia32
> processor?  Or will it by necessity use processor instruction set
> extensions in
> it's optimization process?
>
> > Packaging ATLAS as well would be fantastic.  I've seen some of the list
> > traffic about the difficulties in doing this.  Sure would be nice if it
> > could work, but I don't think it likely.
> >
>
> The Debian approach to packaging Atlas looked promising, but
> turned out to be
> too difficult to mimic. I include an option to build Atlas from
> source in the
> cygwin lapack package.
>
> JRP

Reply | Threaded
Open this post in threaded view
|

RE: fftw on cygwin

James R. Phillips-2


--- Paul Billings wrote:

> I believe a generic version of FFTW would suffice.  In contrast to ATLAS
> which does extensive benchmarking at make/compile time, FFTW does run-time
> measurements (or can load previously saved measurements or "wisdom").
> Whatever instruction set that cygwin is generally targeting should be fine,
> although I suppose you could have two versions: one for "barebones" and one
> for SSE/SSE2 or whatever is appropriate (I forget the differences).  The
> user would simply check one or the other in the cygwin setup application.
>
> Paul

The FFTW web page indicates that a package built with sse2 optimization (the
most commonly available) would still run on a non-sse2 machine, so that is a
strong possibility.

I am experimenting with the source today, trying to build a shared
double-precision library on cygwin.  The build system uses libtool, which I am
fairly unfamiliar with, so this is a bit trial and error.

Would a double-precision library be sufficient for most uses, including for
linking with octave?

The debian octave package also includes hdf5 libs as a dependency.  This is
support for a portable scientific data format.  Do you have use for this on
cygwin?


Reply | Threaded
Open this post in threaded view
|

RE: fftw on cygwin

Paul Billings
Double precision FFTW is what I use with octave.  

FWIW, I personally do not use hdf5 or know anyone who does.

Paul


> -----Original Message-----
> From: James R. Phillips [mailto:[hidden email]]
> Sent: Wednesday, August 10, 2005 11:02 AM
> To: [hidden email]
> Subject: RE: fftw on cygwin
[...]

> Would a double-precision library be sufficient for most uses,
> including for
> linking with octave?
>
> The debian octave package also includes hdf5 libs as a
> dependency.  This is
> support for a portable scientific data format.  Do you have use
> for this on
> cygwin?
>