LAPACK Archives

[Lapack] Divide and conquer diagonalization routines fail when high prec

For information and completeness of the mailing-list archive,
( )

The discussion has been held off-line between Ren-Cang Li and Javier 
Junquera. The problems of Javier Junquera are solved by using some patch 
routines from Ren-Cang Li for xlaed6.f.


On Fri, 3 Mar 2006, Javier Junquera wrote:

  Dear Lapack customer support:

  I think I have found a bug in the divide and conquer routines
supplied in the LAPACK.

  Given platform (AMD Opteron 64 bit, Intel Xeon 32 and 64-bit,
Mac, Pentium-IV,...), compiler (g95, pathscale, pgf90, ifort,...) and 
optimization flag (-O0, -O2), the diagonalization using
divide and conquer seems to work depending on the precision of the
eigenvalue problem to be solved.

  The problem appears both using the LAPACK routines directly supplied
by Lapack (in the example provided below, they are included
in the Libs directory). 
  For a medium precision it works nicely. Increasing the precision to high, 
it crashes independently of the compiler, architecture, compiler flags...

  Please, find enclosed a simplified program showing this bug.
After unzip and untar the package,
read the INSTALL file to know how to change the precision, compile and run 
the code.
You will also find a file (lapack.pdf) with a
summary of our results.

   Thank you very much for your help.

   Best regards,


Javier Junquera

Ciencias de la Tierra y Fisica de la Materia Condensada (CITIMAC)
Avenida de Los Castros s/n
Facultad de Ciencias
Universidad de Cantabria
E-39005 Santander, Cantabria

e-mail:   javier.junquera@Domain.Removed
Tel:      (+34) 942 20 15 16
Fax:      (+34) 942 20 14 02

<Prev in Thread] Current Thread [Next in Thread>

For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or