[petsc-dev] PETSc with precision=longdouble

[Barry Smith]

On Mar 9, 2011, at 4:54 PM, Lisandro Dalcin wrote:

> Barry, what's the status of this?:
> http://lists.mcs.anl.gov/pipermail/petsc-dev/2005-November/000086.html
> 
> I configured with --download-c-blas-lapack, but it seems that the
> argument is plain ignored, configure.log does not provide any clue
> about what's going on, configure success, BLAS/LAPACK config vars are
> plain empty, PETSc builds fine, but at make test I get linker errors
> about missing BLAS symbols... BTW, I've manually downloaded f2c
> blas-lapack tarball, but I cannot figure out how you managed to get a
> "long double" build...

   Unfortunately Satish threw away that tarball of cblas/lapack and replaced it with one generated by the lapack boys which cannot build with long double.

> 
> Do you really got it working?

   Yes but it won't work anymore.
 
   To top it off long double in gcc uses 128 bits but actually only uses the 80 bit floating point units so it is not worth it. Plus the MPICH long double operations like MPI_SUM didn't work properly.

   But, Jed pointed out that gcc has a new way of doing 128 bit floating point. You may see some pushes into PETSc-dev in the next few days as I try to get it working. Keep your eyes open.

   Barry

  BTW: this is most definitely a petsc-dev question and not a personal question to me.



> 
> 
> -- 
> Lisandro Dalcin
> ---------------
> CIMEC (INTEC/CONICET-UNL)
> Predio CONICET-Santa Fe
> Colectora RN 168 Km 472, Paraje El Pozo
> 3000 Santa Fe, Argentina
> Tel: +54-342-4511594 (ext 1011)
> Tel/Fax: +54-342-4511169