This is the mail archive of the gsl-discuss@sourceware.org mailing list for the GSL project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: gsl container designs


On Wed, 2010-01-06 at 22:29 -0500, Robert G. Brown wrote:

> However, some of my beowulfish friends who work(ed)
> for compiler companies argued that one of Fortran's longstanding
> advantages in numerical code efficiency is simply because of the fact
> that matrices in Fortran are written in stone and so compiler writers
> can optimize the hell out of them.  C pointers, OTOH -- well,

I think the main difference is that fortran can assume no aliasing,
which allows some extra optimization that C cannot do. The new 
'restrict' keyword is supposed to help with this, but we'll see.


> I've written code to do so many times -- Numerical Recipes provides the
> basic idea in its matrix-packing routines.  Simply allocate e.g.
> 
> double **..*m,*v;

I understand the mechanics of it. I'm not sure what my concern was;
I thought about this a few months ago and decided there was some
semantic problem. But it looks like the **..*m has exactly the
same life-cycle and maintenance as *v, so I'm not sure what I
was thinking. It might come back to me.


> I can probably dig this code out and post it here (or post a link to it)
> if you want to look at it.  I think I went up to 9 or 10 dimensions (way
> more than I needed) and did a lot of things via brute force to make the
> purely mechanical recursion obvious; I planned an eventual rewrite into
> a more compact form that I never got around to.

Sure. Feel free to post it. I like to look at everything and
cherry-pick the best parts from other people's stuff. Makes
things a lot easier.


--
G. Jungman



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]