Re: [row/column majority] gslclapack

["Robert G. Brown" <rgb at phy dot duke dot edu>]

On Thu, 30 Mar 2006, Z F wrote:

> Row/column majority question comes from the times when
> God was REAL and Infinity was INTEGER (FORTRAN). But even
> at that time, some people, I hope, knew that the question of
> row/column majority is irrelevant. Think about it!
>
> So think.... Sorry, but I will use capital letters...
>
> COLUMN-MAJOR or ROW-MAJOR ??? IT IS ALL IN YOUR HEAD!!!!
> IT IS NOT REAL AND DOES NOT EXIST!!! THE MATRIX IS STORED
> AS A CONTINUOUS OBJECT AND ADDRESSING IS DONE VIA a FORUMLA
> LIKE THIS
>
> index = j*Ni+i!!!
>
> SO GIVE THE MAJORITY A REST!!!
>
> Once again, sorry if I offended anyone, but someone had to point it
> out.

Yes, but CPUs really HATE doing arithmetic with a string of objects
separated by a large offset.  They much prefer to do arithmetic with a
string of objects with unit offset.  That way caches, prefetch and so on
work.  They indicate their disapproval by running really really slowly.

In other words, it is good if "j" isn't the rapidly varying index in
your example of offset arithmetic.

   rgb

> Lazar
>
> --- James Bergstra <james.bergstra@umontreal.ca> wrote:
>
> > On Wed, Mar 29, 2006 at 07:04:39PM +0100, J.J.Green wrote:
> > > I think the problem with calling fortran from C is that
> > > it is compiler, OS and (I think) architecture dependant.
> > > CLAPACK solves the portability problem, but as you have
> > > noticed, not the arcane interface which puts most poeple
> > > off using it.
> >
> > Exactly the motivation for a gsl wrapper (or extension)!  A
> > gsl module would serve both as a gentle introduction to
> > using lapack, as well as a hook to place normal
> > documentation.
> >
> > It seems to me that the largest barrier to a lapack wrapper
> > is that gsl_matrix is an implicitly row-major structure, so
> > that in general, arguments must be out-of-place transposed
> > before a lapack call.  Although this is not significant in
> > terms of either O(n) time or space,  the overhead is more
> > noticeable in the case of large numbers of calls with small
> > arguments.
> >
> > Anyway, it seems like the concensus is that lapack is
> > annoying to use but good in theory, once it's working.  If
> > someone (eg. me) writes an extension that makes (at least
> > part of lapack) less annoying to use in practice, share with
> > the list?
> >
> >
> > James
> >
> > PS.  My extension would probably introduce a column-major
> > version of gsl_matrix.
> >
> > --
> > james bergstra
> > http://www-etud.iro.umontreal.ca/~bergstrj
>
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com

--
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb@phy.duke.edu