Re: GSL 2.0 roadmap (one man's view)

[Gerard Jungman <jungman@lanl.gov>]

On Mon, 2009-09-07 at 14:57 +0100, Brian Gough wrote:
> 
> In practice it seems like any solution is going come down to saying
> people should use FFTW (or LAPACK) one way or another.  Either we
> could bundle it, or -- if we don't supply the existing GSL routines --
> they could be forced to install it.  Either way it doesn't seem like
> the result is actually much different from now.

What we type may not be much different. What we say could be quite
different. We need to catalog the use cases for these things. It
mainly comes down to inter-operability of data structures.

In the end, everybody's data structures amount to something plastered
over the top of double *, with some meta-information for layout and
life-cycle control. This includes things likes the row-major vs
column-major distinction, which we should not ignore any longer. 

But this rapidly turns into a discussion of gsl_vector and friends,
which we should save for later. The distinct point is that the
GSL stop-gap implementations are hindering us. I have seen
benchmark web pages, regarding ffts or linear algebra,
where GSL looks like a turd. Well... if the dunce cap fits...


> While getting packages via RPM/DEB is typically fine for desktop use,
> I think the ease of doing this in the wider world generally is
> underestimated:

Packages are the best thing to happen in computer
organization and configuration ever. Period. But anyway...


> Consider an example of targetting an embedded system for
> data-collection with an ARM cpu where the vendor provides only an
> older version of GCC, without fortran.  Based on my experience, this
> is a realistic scenario in industry.

We could have a poll. How many people reading this mailing list
are targeting a platform like this, and how many are sitting
in front of a linux box right now, trying to get their crap
to work, so that they can get on with life?

The fft and linear algebra code need not be sent down the
garbage chute. It will be there for fortran-less ARM prisoners.
But the rest of us should get something better by default.

--
G. Jungman