From owner-help-octave at bevo dot che dot wisc dot edu  Wed Jan 24 02:46:34 1996
Subject: fsolve speedup ... ?!?
From: Ferdinand Schinagl <ferdi at maxwell dot hlphys dot uni-linz dot ac dot at>
To: help-octave at che dot utexas dot edu
Date: Wed, 24 Jan 1996 09:48:39 +0100


Hello,

This is just some kind of brainstorming !!!!!!!!!!!!!!!!!!!!!

  I had a closer look at the sources of octave, especially the minpack
stuff from libcruft. It contains code which fsolve basically relies on.
If I'm right fsolve only uses HYBRD.F (HYBRD1.F). These are routines
which calculate the jacobian of the (equation)system themselves. The
others namely HYBRJ.F (HYBRJ1.F) make use of an externally defined
function (fortran subroutine) to calculate or better to say to determine
the jacobian matrix! I think that the latter routines would decrease
calculation time enormously, especially when trying to solve large systems
of complex equations, simply because of the fact that the
forward-difference method used in HYBRD.F needs a lot more function
evaluations in comparison !?!
  The problem for me now is that I don't know how to modify
(if nobody minds?) the sources myself because I'm not
familiar with programing C++. The only thing I can do for know would
be to propose some kind of modified interface for the fsolve function.
  The other thing is that I've asked our local experts from the center of
supercomputing at our university, if they could try to port octave on
to a CONVEX SPP1200/XA-24 system (24 parallel hppa-7200 processor machine).
It's so that the binary distribution of octave works fine on this platform
but only seqentially :( I don't know if someone has any experience in such
porting - any support would be useful)

Greetings

Ferdinand