#!/bin/tcsh
cat << EOS > pgma.f
      program main
      implicit none
      include 'mpif.h'
      integer lm, ln
      integer i, j
      integer rank, size, ierr
      parameter (lm=10,ln=4)
      integer iarray_of_sizes(2)
      integer iarray_of_subsizes(2)
      integer iarray_of_starts(2)
      integer icoord(2), nproc(2), ndims
      integer ifh
      integer offset
      integer status(MPI_STATUS_SIZE)
      real Alocal(lm,ln)
      double precision t0, t1
C
      call MPI_INIT( ierr )
      call MPI_COMM_RANK( MPI_COMM_WORLD, rank, ierr )
      call MPI_COMM_SIZE( MPI_COMM_WORLD, size, ierr )

      if (size .ne. 4) then 
         print *, 'This example requires 6 processes'
         call MPI_ABORT( MPI_COMM_WORLD, 1, ierr )
      endif
      
      nproc(1) = 2
      nproc(2) = 2

      icoord(1) = rank / nproc(2)
      icoord(2) = mod(rank,nproc(2))

      iarray_of_sizes(1)    = lm * nproc(1)
      iarray_of_sizes(2)    = ln * nproc(2)
      iarray_of_subsizes(1) = lm
      iarray_of_subsizes(2) = ln
C
C Note that starts are 0-origin, even in Fortran
      iarray_of_starts(1) = icoord(1) * iarray_of_subsizes(1)
      iarray_of_starts(2) = icoord(2) * iarray_of_subsizes(2)
      ndims = 2
C
C Initialize Alocal with the order values in Global numbering
      do 10 j=1,ln
         do 10 i=1,lm
            Alocal(i,j) = i + iarray_of_starts(1) + (j - 1 +
     &           iarray_of_starts(2)) * iarray_of_sizes(1)
 10      continue
C
C Do the file operations
C
      call MPI_FILE_OPEN(MPI_COMM_WORLD, 'testa', MPI_MODE_CREATE +
     &     MPI_MODE_RDWR, MPI_INFO_NULL, ifh, ierr) 
      call MPI_BARRIER( MPI_COMM_WORLD, ierr )
      t0 = MPI_WTIME()
      do 20 j=1, ln
         offset = iarray_of_starts(1) + (j-1 + iarray_of_starts(2))
     &        * iarray_of_sizes(1)
         offset = offset * 4
         call MPI_FILE_SEEK( ifh, offset, MPI_SEEK_SET, ierr )
         call MPI_FILE_WRITE_ALL( ifh, Alocal(1,j), lm,
     &        MPI_REAL, status, ierr ) 
 20   continue
      call MPI_BARRIER( MPI_COMM_WORLD, ierr )
      t1 = MPI_WTIME() - t0
      if (rank .eq. 0) then
         print *, 'Write Aglobal in ', t1, ' seconds'
      endif

      call MPI_FILE_CLOSE(ifh, ierr)

      call MPI_FINALIZE( ierr )
      end
EOS
#
mpif90 pgma.f
mpirun -np 4 a.out