PGF90-S-0155-Could not resolve generic procedure mpi_scatter

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

I’m currently wondering what I’m doing wrong. It’s been a while since I’ve done some MPI programming and so I decided to start out with a simple example of MPI_SCATTER to make sure I remembered what I’m doing. To wit, I wrote this rather awful program:

program scattertest

   use mpi

   implicit none

   integer, parameter :: m = 4
   integer, parameter :: np = 6

   integer :: a(m,np), b(m,np)
   integer :: alocal(m/2,np), blocal(m/2,np)
   integer :: ierr, total, rank
   integer :: msize, i, j, k

   call mpi_init(ierr)
   call mpi_comm_size(MPI_COMM_WORLD, total, ierr)
   call mpi_comm_rank(MPI_COMM_WORLD, rank, ierr)

   write (*, *) "I am process ", rank, " of ", total

   if (rank == 0) then
      do i=1,m
         a(i,:)=i
      end do
   end if

   write (*,*) a(1,:)

   call mpi_scatter(a,2*6,MPI_INTEGER,&
                    alocal,2*6,MPI_INTEGER,&
                    0,MPI_COMM_WORLD,ierr)

   write (*,*) alocal(1,:)

   call mpi_finalize(ierr)
end program scattertest

I’m only trying to run this on 2 processors at the moment, thus I’ve sort of hardwired in some constants. However, when I try to compile this:

> mpif90 scattertest.f90
PGF90-S-0155-Could not resolve generic procedure mpi_scatter (scattertest.f90: 29)
  0 inform,   0 warnings,   1 severes, 0 fatal for scattertest

I’m not too sure how to interpret that error. (Oh, and I suppose if I have the MPI code completely wrong, speak up there as well.)

2

Answering myself, it looks like the reason is the strong-typing of Fortran 90. If one instead uses “include ‘mpif.h’” rather than “use mpi”, the behavior I’m used to returns.

3

Hi Matt,

The type mismatch is because you’re using multidimensional arrays which MPI_SCATTER is expecting a single dimension. Either you need to reshape the array or not use mpi.mod. With “mpi.h”, you’re using F77 calling conventions so only a pointer is passed to MPI_SCATTER.

  • Mat