Subject: Re: [ublas] [bindings] New traits system
From: Rutger ter Borg (rutger_at_[hidden])
Date: 2010-01-20 06:23:23
Thomas Klimpel wrote:
>
> "b" should be "Nx1" for column-major and "1xN" for row-major. The reason
> for this is explained under the "What's the deal with the RHS in the
> row-major factorization/solves?" item in the ATLAS FAQ:
>
> http://math-atlas.sourceforge.net/faq.html#RowSolve
>
> Regards,
> Thomas
I've re-read the text on the ATLAS site a couple of times, how confusing ...
As I understand it, its not necessarily about Nx1 or 1xN, but its all about
what values should be passed about B, and a its bit about the storage
requirements of B, too.
With this being the case, I would like to propose that we offer our users a
consistent interface, and have the bindings solve issues with backends such
as that of atlas. Let's take to our advantage that we have all necessary
knowledge during the gesv binding to pass the right values to atlas, and be
correct. Also, let's make sure the only needed code-change with a
recompilation against a true lapack backend is in the row_major/column_major
setting of the matrices.
I'll try to come up with the right trait functions for this case.
Cheers,
Rutger