Boost logo

Ublas :

From: Riccardo Rossi (rrossi_at_[hidden])
Date: 2006-08-01 06:20:07

Hello Again Gunter,
     thanks for the explication we will work a bit on it and hopefully
in september we will tell you of the results...

as a personal point of view i do not completely agree about the issue of
performance... in my opinion performance IS of paramount importance
particularly if ublas has all the ingredients to do it very well. We
will see if for this particular case we can do something fast...and of
course if we manage we will make you know



On Tue, 2006-08-01 at 10:10 +0200, Gunter Winkler wrote:
> Riccardo Rossi schrieb:
> > Hello Gunter,
> > i like your matrix implementation as it is reasonably simple and (as
> > i understand) it does not have the overhead of the two sizes :-)
> >
> > please find attached our own "Tiny Vector" that should be the the
> > perfect companion for it... it is a little more complicated but i guess
> > this depends on our little experience inside the ublas :-D
> >
> After reading a few lines of your code I thought your class is the same
> as ublas::bounded_vector. You have a dynamic size vector inside a fixed
> size storage array. IMHO a real fixed size vector should only allow
> exactly one size which is given by the template parameter.
> about the possible perfomance: If we had a fixed size matrix and a fixed
> size vector we could specialize all combinations of prod() to get
> maximal performance (but this is typically the goal of atlas and not
> ublas). I don't think we should double the work of the atlas team. Maybe
> there is an easy way to directly map the specialized prod(), axpy_prod()
> to some atlas core routines. The only thing we should do is to give the
> compiler a chance to optimize as good as possible. This can be done by
> creating lots of specializations (by a shell script) that are as simple
> as possible.
> about the iterators: I chose indexed iterators for simplicity. Maybe
> "pointer" interators give better performance.
> mfg
> Gunter
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]