Re: [ublas] leading dimension for lapack bindings

Subject: Re: [ublas] leading dimension for lapack bindings
From: petros (pmamales_at_[hidden])
Date: 2012-03-02 16:22:44

.. and also,
..boost/numeric/bindings/detail/property_map.hpp(30): error C2039: 
'property_map' : is not a member of 
'boost::numeric::bindings::detail::adaptor_access<T>'
The compiler is right, there is no such type in the adaptor_access..(just 
poured it from the sandbox)
TIA,
Petros

-----Original Message----- 
From: petros
Sent: Friday, March 02, 2012 3:57 PM
To: ublas mailing list
Subject: Re: [ublas] leading dimension for lapack bindings

Thank you Thomas.
Please allow for a couple more questions.
Want to use the bindings with mkl. Can I ? I thought I saw people
on the web claiming they could - no problem, they said.
Assuming this to be the case, can I also do it for the 64-bit builds?
If I do not use CLAPACK does this mean that I use the fortran interface ?
I have been assured from intel that mkl and clapack have the same
signatures.

The C API of mkl does not provide with the const qualifiers. Neither does
the CLAPACK.h - at least the
one I just checked upon in the web.
Can I hope that the const qualifiers I see in the bindings are a result of
some extra effort put
by the designers of the library (vary likely, some boost type "magic" ? ).
Btw, is there a way to overwrite the include <CLAPACK.h> ? (or just create
my dummy one, caling  mkl.h  ) ?

Finally, to answer your question, for mkl to work optimally and securely,
certain constraints on the
leading dimension need to be met - more than smple data allignment. This
forces me to define my matrix as
a range of a ublas matrix that I hold under the proverbial rug. Hence my
bind (no pun or maybe a little ;-))

Thank you very much for your help, which is invaluable given the scarcity of
documentation.
Apologies for imposing on a Fri afternoon,
Best Regards,
Petros

-----Original Message----- 
From: Thomas Klimpel
Sent: Friday, March 02, 2012 3:39 PM
To: ublas mailing list
Subject: Re: [ublas] leading dimension for lapack bindings

petros wrote:
> I want to use the gesv function of numeric bindings for a matrix_range 
> (i.e. a submatrix).
> LAPACK expects this to be the number of rows (column_major layout) of the 
> original matrix.
> Is the bindings library “smart” enough to figure it out ?

Of course it is smart enough.

> Alternatively, if I want to use the bindings calls with derivative classes 
> of the boost
> matrix  how can I overwrite this “standard” behavior (i.e. leading 
> dimension)?

I get the impression that you ask how to write a binding for your own matrix
type. That's quite easy, but hopefully won't be required...

Regards,
Thomas
_______________________________________________
ublas mailing list
ublas_at_[hidden]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: pmamales_at_[hidden]

_______________________________________________
ublas mailing list
ublas_at_[hidden]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: pmamales_at_[hidden]