lapackers

[James Demmel <demmel@cs.berkeley.edu>]

Yozo, thanks for the comparison of mpack with your planned high-precision
version of LAPACK.

I'm amazed that Nakata seems to rewritten so much of the library by hand 
in C++
(albeit missing the "hard parts", eigensolvers etc.)

Given that we are engaged in thinking about future LAPACK versions,
we certainly want to think about how to building them in a way that would
let us incorporate various precisions and perhaps more general types
(eg finite fields have been requested for some routines where they make 
sense).
Including into the test code, as you point out.

And it is really important to have it under mBSD license (which does not
prevent someone from linking it to some arithmetic package licensed
under GPL, but that is their decision).

Thanks,
Jim



Yozo Hida wrote:
> On Tue, Jan 27, 2009 at 09:49:17PM -0800, James Demmel wrote:
>   
> > Yozo: How does this differ from what you had planned to build, beyond
> > the choice of extra precise floating point library?
> >     
>
> The end goal is similar: Lapack routines with high precision.  Some
> differences from what I have in mind:
>
>   + I have a native Fortran 95 (or 2003) interface and implementation.
>     Nakata's Mpack is strictly C++, looks like it is re-written by hand.
>     I plan to use Fortran parsing techniques to generate source from the
>     official LAPACK fortran sources.  There are several advantages for
>     this:
>
>       - This way bug fixes in original LAPACK can be semi-automatically
>         incorporated into the multiple precision package.
>
>       - New routines that are added to LAPACK would be converted without
>         too much effort.  Currently mpack is missing some routines,
>         (e.g., csysv, all of eigen/svd routines), probably due to lack
>         of man-power :-)
>
>       - LAPACK testing code will also be semi-automatically incorporated.
>         Mpack currently only has rudimentary testing code.
>
>       - Obviously, a Fortran interface.  There is a surprising number of
>         people using Fortran high precision, as I found out through
>         QD/ARPREC support.
>
>   + A bit more flexibility in the high precision package, e.g.,
>     double-double, quad-double, and ARPREC.  But this should be
>     fairly easy to incorporate into Nakata's C++ implementation
>     (since C++ does have various overloading syntax making this easier).
>
>   + Does license have any bearing on this?  I myself don't care too much
>     about GPL vs. BSD.
>
> I may have more comments once I play around with Mpack a bit more.
>
> -Yozo
>
>