[petsc-dev] Enabling single precision for MUMPS and SUPERLU

[Barry Smith]

  Stefano,

    Thanks for doing this.  A couple of comments:

       - probably would be better if the change for each package was a separate change set/patch. (I guessing they are not)

       -  I don't like this:

+/* calls to MUMPS */
+#if defined(PETSC_USE_COMPLEX)
+#if defined(PETSC_USE_REAL_SINGLE)
+#define PetscMUMPS_c cmumps_c
+#else
+#define PetscMUMPS_c zmumps_c
+#endif
+#else
+#if defined(PETSC_USE_REAL_SINGLE)
+#define PetscMUMPS_c smumps_c
+#else
+#define PetscMUMPS_c dmumps_c
+#endif
+#endif

            I understand why you are doing this (since xmumps_c appears in multiple places in the code and you don't want the ugly #ifdef at each of those places) but essentially you are doing what the mumps people should have done, providing a polymorphic interface to their solver. By doing this you are making the PETSc--mumps interface less transparent and introducing the complexity of additional macros. I think it would be better to use the old style we had (where the #ifdef appear each time as needed in the code).

    Barry



 On Sep 21, 2012, at 9:41 AM, Stefano Zampini <stefano.zampini at gmail.com> wrote:

> Sorry. I forgot the attachment
> 
> 2012/9/21 Stefano Zampini <stefano.zampini at gmail.com>
> Yesterday it was the first time I try to build a single-precision version of PETSc with some external packages.
> I noticed that SuperLU and MUMPS, which are currently fully compiled (i.e. with scalar types s,d,c and z) by the configure, cannot be used when requiring single precision at configure time. So I decided to wrote a patch to enable their use when PETSC_USE_REAL_SINGLE is defined.
> 
> Attached please find a patch. Is there any reason for which you did not already provide this functionality?
> I tested the patch on BlueGene/Q and it works for scalar types real. Currently, I don't have a complex version of PETSc installed. Can someone test it?
> 
> I also noticed that UMFPACK (which itself does not have support for single precision), is accepted by the configure by requiring --with-precision=single --download-umfpack=1. Is it normal? It breaks the final linking of the executables.  
> 
> -- 
> Stefano
> 
> 
> 
> -- 
> Stefano
> <patch_mumps_superlu_single.diff>