[petsc-dev] [Facets-devel] dealing with MPIUNi

[John Cary]

UEDGE needs to be built three ways"

1. Real parallel, which is done with petsc
2. Serial with petsc, and then one has to use mpiuni.  The only place 
where I
see use of mpif.h is
       subroutine jacstnlout

       implicit none
c_mpi      include 'mpif.h'

in bbb/odesetup.m.  It is possible that this could be completely
commented out.  (c_mpi = "" if one is using petsc, i.e., whether
with MPI or pseudo MPI.)  If it can, then the problem may be solved.

(The .m files are LLNL-FES paleo-Fortran.)

3. Serial without petsc (the real legacy code).

While I am not sure of what is necessary, given the condition of
the code and the need for migration, we had to support
   Finding MPI (pseudo/uni or real) for any use of PETSc
   Not having any mpif.h if not using PETSc.

At one point it seemed that the PETSc-serial build still required
an include of mpif.h.  This might be fixable, not sure.

I am not sure where this leaves us, but we can probably make anything
work eventually.

John

PS - I recognize that many folks just "always build with MPI".  We do
not.  We like to support "pure serial" -- not to impose that any mpi
be installed if they want only a serial build of any of our code.



On 2/26/10 1:27 PM, Satish Balay wrote:
> On Fri, 26 Feb 2010, Barry Smith wrote:
>
>
>> On Feb 26, 2010, at 2:56 PM, Satish Balay wrote:
>>
>>
>>> On Fri, 26 Feb 2010, Barry Smith wrote:
>>>
>>>
>>>> On Feb 26, 2010, at 2:43 PM, Satish Balay wrote:
>>>>
>>>>
>>>>> Looks like uedge build tools look for mpiuni spearately and adds in
>>>>> -Impiuni.
>>>>>
>>>>> But I think its good to keep -Impiuni in petsc makefiles for
>>>>> regular users who have mpi.h [or mpif.h] in their non-petsc soures.
>>>>>
>>>>>
>>>> But doesn't this mean that the default PETSc build now with single PETSc
>>>> library is broken for Facets?
>>>>
>>> For some reason it needs mpif.h explicitly - so its looking for it.
>>> [and adding -Ipath].
>>>
>>   How come it cannot just use the output from getincludedirs to have the search
>> path to find mpif.h
>>
>
>
>>   That is, isn't the best approach one where it can use MPIUni, but
>> uses it automatically without needing any mention of MPIUNI in the
>> FACETS build system at all? Or does it already work that way and we
>> are all set?
>>
> facets has an autoconf macro to detect PETSc [multiple versions,
> different install types, with prefix - or not etc..] There might be
> scope to simplify this code.
>
> As of now - it appears to work [with petsc-dev except for the makefile
> change].
>
> Satish
> _______________________________________________
> Facets-devel mailing list
> Facets-devel at ice.txcorp.com
> https://ice.txcorp.com/mailman/listinfo/facets-devel
> Search this and related lists: https://ice.txcorp.com/cgi-bin/facets
>
>