dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Jan 04, 2010 at 10:24:01PM +0100, Garth N. Wells wrote:
>
>
> Anders Logg wrote:
> > On Sun, Jan 03, 2010 at 03:38:19PM +0100, Garth N. Wells wrote:
> >> It would be useful to have an efficient and simple data structure for
> >> small and large dense matrices with lots of functionality. We have
> >> uBLAS, but I'm not convinced of its efficiency for small matrices, and
> >> it doesn't interface to LAPACK, which is why LAPACKMatrix was added.
> >>
> >> There are a few libraries around that provide expression templates (for
> >> nice syntax) and LAPACK/ATLAS interfaces, see for example
> >>
> >>     http://arma.sourceforge.net/
> >>
> >> I think that wrapping such a library would be very useful and could be
> >> better that having LAPACKMatrix, etc.
> >>
> >> Garth
> >
> > Agree. I was thinking of wrapping a suitable library before adding
> > LAPACKMatrix, but it was too much effort at the time.
> >
>
> Even less effort for your solvers would be to use Armadillo directly.
> It's available as a package through apt-get.

Yes, we should use the solvers also. That's one of the major points to
get all LAPACK functionality for free, not just bits of it.

> > Incidentally, I also looked at Armadillo as the main candidate for
> > wrapping.
> >
>
> I converted the fenics-plasticity solve from uBLASDenseMatrix to
> Armadillo yesterday, and got a total solve time speed up of 100%. This
> includes assembly, etc, so I expect the code that uses dense matrices is
> an order of magnitude faster. From some simple tests, matrix inversion
> in Armadillo was about 35 times faster that uBLASDenseMatrix, and the
> syntax is nicer than uBLAS. Addition, etc, was also much faster with
> Armadillo.

Impressive.

--
Anders

Attachment: signature.asc
Description: Digital signature