[petsc-dev] Matlab implementation of PC

[Barry Smith]

  Dave,

   I started work last night on this and found the same problem.

   But I have a solution that I implemented. It seems to work, so far only tested with MPICH. You can pull, then run configure, make as usual with MPICH (or you can try other MPIs). Note that the PETSc Matlab interface has not changed, the PETSC_COMM_SELF is put in automatically for you.

   Barry

On Dec 10, 2010, at 6:47 AM, Dave May wrote:

> Hey Barry,
>  Looking through the docs associated with loadlibrary, calllib etc.,
> it wasn't obvious to me how to  extract a global variable
> (PETSC_COMM_SELF) from a c library.
> 
> One possible solution may be to write a function
>  PetscErrorCode PetscGetCommSelf(MPI_Comm *comm)
> which simply sets *comm = PETSC_COMM_SELF. If this function was part
> of libpetsc, then it could be accessed using calllib. Since it's a
> trivial getter which possibly doesn't belong in the libpetsc, it could
> be wrapped up with a standalone mex function. Does anyone know a
> better way?
> 
> Am I missing something obvious related to how to pull out COMM_SELF from petsc?
> 
> Cheers,
>  Dave
> 
> 
> On 9 December 2010 23:04, Barry Smith <bsmith at mcs.anl.gov> wrote:
>> 
>> On Dec 9, 2010, at 3:52 PM, Dave May wrote:
>> 
>>> Ok, after digging a little bit more, I realise I can configure the pc
>>> in a command line like
>>> manner using this syntax '-pc_type','lu'.
>>> 
>>> What I was hoping to use was ML and hypre, however it appears neither
>>> of these libraries can be built with --with-mpi=0, which seems to be a
>>> requirement when calling petsc from matlab.
>>> Is there a straight forward way to enable hypre and ml to be used when
>>> --with-mpi=0?
>> 
>> No. I won't recommend monkeying with that.
>> 
>> A much better way to do it would be to "improve" the Matlab interface to support running under MPI; first step of course just sequential (one process MPI).
>> 
>> This would be done by putting a PETSC_COMM_SELF field into PetscObject as a nonconstant property and then have PetscInitalize.m pull the PETSC_COMM_SELF up from C and put it into PetscObject. Then calls to, for example, KSPCreate() from Matlab would automatically pass that property in (currently they just pass 0 in, which works for MPI Uni).
>> 
>>   Barry
>> 
>>> 
>>> Cheers,
>>>  Dave
>>> 
>>> 
>>> 
>>> 
>>> On 9 December 2010 22:26, Dave May <dave.mayhem23 at gmail.com> wrote:
>>>> Hi Barry,
>>>>  Is it possible that you could possible add PetscPC.m?
>>>> I work with a bunch of matlab users solving Stokes on staggered grid who
>>>> rely on umfpack. I'd love to be able to get them to try ML or hypre via PETSc.
>>>> 
>>>> Cheers,
>>>>  Dave
>>>> 
>> 
>>