MPICH error when using petsc-3.0.0-p4

[Satish Balay]

MPIU_SCALAR is defined to be MPI_DOUBLE_PRECISION

I guess MPI_REAL/MPI_REAL8 behavior is affected by -r8.

Satish

On Tue, 14 Apr 2009, Wee-Beng TAY wrote:

> Hi Satish,
> 
> Moreover, I found that as long as I use MPIU_SCALAR instead of MPI_REAL8,
> there's no problem. Btw, I'm using MPICH 1.25.
> 
> Thank you very much and have a nice day!
> 
> Yours sincerely,
> 
> Wee-Beng Tay
> 
> 
> 
> Satish Balay wrote:
> > (--with-scalar-type=[real,complex], --with-precision=[single,double])
> > 
> > Depending upon how petsc is configured - PetscScalar can be either
> > real4,real8,complex8. With most common usage - [i.e default build]
> > PetscScalar is real8.
> > 
> > Earlier you claimed: MPI_ISEND(v_ast) worked with MPI_REAL but not
> > MPI_REAL8. This sugested that v_ast was declared as real4 - not real8.
> > 
> > Satish
> > 
> > On Tue, 14 Apr 2009, Wee-Beng TAY wrote:
> > 
> >   
> > > Hi,
> > > 
> > > So when must I use PetscScalar, and when can I just use real(8)? This is
> > > because most of my variables are in real(8), except for the Mat, Vec,
> > > required
> > > in the solving of the linear eqn. Is there a rule? I initially thought
> > > they
> > > are the same and if I remember correctly, it seems to work fine in
> > > PETSc2.3.3.
> > > 
> > > My v_ast is now defined as:
> > > 
> > > PetscScalar, allocatable ::    u_ast(:,:), v_ast(:,:)
> > > 
> > > allocate (u_ast(0:size_x+1,jsta_ext:jend_ext), STAT=status(1));allocate
> > > (v_ast(0:size_x+1,jsta_ext:jend_ext), STAT=status(2))
> > > 
> > > if (status(1)/=0 .or. status(2)/=0) STOP "Cannot allocate memory"
> > > 
> > > Thank you very much and have a nice day!
> > > 
> > > Yours sincerely,
> > > 
> > > Wee-Beng Tay
> > > 
> > > 
> > > 
> > > Satish Balay wrote:
> > >     
> > > > And if v_ast() is used with PETSc - then it should be defined
> > > > 'PetscScalar'
> > > > 
> > > > and then you use MPI_ISEND(....,MPIU_SCALAR,...)
> > > > 
> > > > Satish
> > > > 
> > > > 
> > > > On Mon, 13 Apr 2009, Barry Smith wrote:
> > > > 
> > > >         
> > > > >   Where is your implicit none that all sane programmers begin their
> > > > > Fortran
> > > > > subroutines with?
> > > > > 
> > > > > 
> > > > > On Apr 13, 2009, at 11:11 AM, Wee-Beng TAY wrote:
> > > > > 
> > > > >             
> > > > > > Hi Satish,
> > > > > > 
> > > > > > Compiling and building now worked without error. However, when I
> > > > > > run, I
> > > > > > get
> > > > > > the error:
> > > > > > 
> > > > > > 0 - MPI_ISEND : Datatype is MPI_TYPE_NULL
> > > > > > [0]  Aborting program !
> > > > > > [0] Aborting program!
> > > > > > Error 323, process 0, host GOTCHAMA-E73BB3:
> > > > > > 
> > > > > > The problem lies in this subroutine:
> > > > > > 
> > > > > > subroutine v_ast_row_copy
> > > > > > 
> > > > > > #include "finclude/petsc.h"
> > > > > > #include "finclude/petscvec.h"
> > > > > > #include "finclude/petscmat.h"
> > > > > > #include "finclude/petscksp.h"
> > > > > > #include "finclude/petscpc.h"
> > > > > > #include "finclude/petscsys.h"
> > > > > > 
> > > > > > !to copy data of jend row to others
> > > > > > 
> > > > > > integer :: inext,iprev,istatus(MPI_STATUS_SIZE),irecv1,ierr,isend1
> > > > > > 
> > > > > > inext = myid + 1
> > > > > > iprev = myid - 1
> > > > > > 
> > > > > > if (myid == num_procs - 1) inext = MPI_PROC_NULL
> > > > > > 
> > > > > > if (myid == 0) iprev = MPI_PROC_NULL
> > > > > > 
> > > > > > CALL
> > > > > > MPI_ISEND(v_ast(1,jend),size_x,MPI_REAL8,inext,1,MPI_COMM_WORLD,isend1,ierr)
> > > > > > CALL
> > > > > > MPI_IRECV(v_ast(1,jsta-1),size_x,MPI_REAL8,iprev,1,MPI_COMM_WORLD,irecv1,ierr)
> > > > > > CALL MPI_WAIT(isend1, istatus, ierr)
> > > > > > CALL MPI_WAIT(irecv1, istatus, ierr)
> > > > > > 
> > > > > > end subroutine v_ast_row_copy
> > > > > > 
> > > > > > 
> > > > > > I copied this subroutine from the RS6000 mpi manual and it used to
> > > > > > work.
> > > > > > I
> > > > > > wonder if this is a MPI or PETSc problem? Strange because I already
> > > > > > specify
> > > > > > the type to be MPI_REAL8. However changing to MPI_REAL solves the
> > > > > > problem.
> > > > > > 
> > > > > > If this is a MPI problem, then you can just ignore it. I'll check it
> > > > > > in
> > > > > > some
> > > > > > MPI forum.
> > > > > > 
> > > > > > 
> > > > > > Thank you very much and have a nice day!
> > > > > > 
> > > > > > Yours sincerely,
> > > > > > 
> > > > > > Wee-Beng Tay
> > > > > >           
> > 
> >