Re: [PATCH] Add workaround for broken xGELSD workspace queries.

[David Bateman]

John W. Eaton wrote:
> On 12-Feb-2008, Jason Riedy wrote:
>
> | The xGELSD workspace queries are inconsistant with the logic inside
> | those routines in the short and fat case (m > n).  For that case,
> | calculate a workspace size large enough to enable the efficient
> | algorithm.  This may allocate more memory than necessary for small
> | cases.  The calculated alternate size is at least the documented
> | minimum workspace requirement, so this amount always will work even
> | if the routines change.
> | ---
> |    Sorry it's been so long.  I don't want to describe the LAPACK
> |    developer dynamic, but it's severely busted.  ergh.  I think
> |    this is a minimal change, and the ILAENV declaration should
> |    work on every Fortran compiler I know. -- Jason
>
> Just to avoid possible incompatible C<->Fortran calling conventions
> for Fortran funtions, we have always wrapped them in Fortran
> subroutines (see the files in libcruft/lapack-xtra).  I did that for
> ilaenv and checked in your change.
>
> Thanks,
>
> jwe
>
>   
This supercedes the xGELSD change I sent. However I have one
disagreement with this patch. The xGELSD.f functions use the condition

               IF( N.GE.MNTHR ) THEN

rather that (n > m) as Jason used in his patch to check for the
workaround. Where MNTHR is given by

            MNTHR = ILAENV( 6, 'ZGELSD', ' ', M, N, NRHS, -1 )

or 1.6 *m. Therefore I believe that the workaround should only apply
with the condition "if (1.6 * n > m)" instead of teh "if (n > m)" that
is currently used.

D.






-- 
David Bateman                                address@hidden
Motorola Labs - Paris                        +33 1 69 35 48 04 (Ph) 
Parc Les Algorithmes, Commune de St Aubin    +33 6 72 01 06 33 (Mob) 
91193 Gif-Sur-Yvette FRANCE                  +33 1 69 35 77 01 (Fax) 

The information contained in this communication has been classified as: 

[x] General Business Information 
[ ] Motorola Internal Use Only 
[ ] Motorola Confidential Proprietary