Re: [tlug] [OT] Japan and ?? credit cards

["Stephen J. Turnbull" <stephen@example.com>]

Edward Middleton writes:

 > Fare enought, it was a deal breaker for me because every calculation
 > required a different size matrix.  After using both I am glad I went
 > with fortran90 on a sparc because it was about 100x faster then the
 > equivalent in g77 and produced more accurate results.

Oh, I didn't realize you were serious about the "77".  I don't know
what the guy on DJGPP was doing, I was just pointing out that FORTRAN
(even ancient specs) is still alive and well.

 > I don't know, most numerical calculation break down to a few basic
 > matrix calculations which have optimized libraries for the specific
 > hardware.

Sure, but there are several of those for FORTRAN, as well.  I'm not a
numerical specialist, but what I see at the conferences is (a)
prototypes which handle toy problems written in whatever language you
like, (b) optimized codes written in FORTRAN for use in highly
specialized applications, and (c) ports to other languages (including
R/S, Matlab, etc) if and when popular demand justifies it.

 > It is obviously possible to go further then this in optimize
 > specific algorithms for specific hardware, and there has been some
 > interesting haskell research into automatically doing this for
 > video card graphics processors and ocaml for producing hardware
 > optimized FFT

Sure, but we were talking about statistical and financial
applications, I thought.  Those tend to be much less parallelizable;
very often the computational burden is computing fixed points (which
even today is most efficiently done by iteration of a contraction
mapping), numerical integrations, or optimizations, typically repeated
with small parameter tweaks hundreds of thousands of times.