What is the proper way to handle MPI communicators in Fortran?

Question

I read that is recommended to use the MPI module rather than include mpif.h file. However, I get the following error

Error: There is no specific subroutine for the generic ‘mpi_comm_split’

when I run this program

program hello_world
  use mpi_f08
  implicit none
  ! include 'mpif.h'
  integer :: ierr, num_procs, my_id,newcomm
  integer :: color,key

  call MPI_INIT ( ierr )
  color =1; key=0
  call MPI_COMM_RANK (MPI_COMM_WORLD, my_id, ierr)
  call MPI_COMM_SIZE (MPI_COMM_WORLD, num_procs, ierr)

  call MPI_Comm_split(MPI_COMM_WORLD, color,key,newcomm, ierr)

  call MPI_FINALIZE ( ierr )

end

The error disappears if I include 'mpif.h' instead of using the MPI module. Why is that?

Simitar

Never tried it, but it looks like the type to use for f08 interface uses type(MPI_comm) for newcomm... – Lioness

Related to what Gilles has said what happens if you simple "use mpi"? What compiler and version and what mpi and version might be useful as well. – Caudex

The error disappears if I include 'mpif.h' instead of using the MPI module. Why is that? Because, by using the mpi module you are enabling the compiler to check that routine formal and actual arguments match at compile time. The include approach lets all sorts of nonsense happen at run time. – Teletype

Answer 1

The use mpi_f08 interface introduces different wrapper types for the different MPI handle objects. While in mpif.h and the use mpi interfaces all handles are simply INTEGERs, in the use mpi_f08 interface there are TYPE(MPI_Comm), TYPE(MPI_File), etc. This allows the compiler to perform checks for things like passing a communicator handle where a file handle is expected.

This is a breaking change on the source level as code must be rewritten, e.g.,

INTEGER :: newcomm

becomes

TYPE(MPI_Comm) :: newcomm

On the binary level there is no change since all those MPI_Xyz types are simply an INTEGER wrapped in a TYPE specifier, which makes them layout compatible. Old Fortran code can still exchange MPI handles with modern Fortran code and vice versa - the INTEGER handle value can be set or extracted via newcomm%MPI_VAL.

Ruelas

Thank you! I am getting another error when using MPI_types in reduce function such as call MPI_Allreduce( x,y , N, MPI_DOUBLE, MPI_SUM, newcomm, ierr ) that Symbol ‘mpi_double’ at (1) has no IMPLICIT type. How should this be modified? using type(MPI_DOUBLE) doesn't help. – Simitar

MPI_DOUBLE corresponds to the C/C++ double type. It must not be used in Fortran code. Use MPI_REAL or whatever the type of x and y is (MPI_DOUBLE_PRECISION, etc.) – Ruelas

However if I have a real(c_double) I would hope a corresponding MPI type is available... – Eartha

Isn't real(c_double) only meant to be used when interfacing to C/C++ code? – Ruelas

Not really@HristoIliev but it is one way to get the number of bytes consistent, so for interfacing with c it is nice. On Intel then (KIND=8) works, but if and when one wants to use gfortran then it can be easier to replace that with c_double which is more portable by design. Whether there are other implications I am not sure, but it works in pure fortran find-n-dandy. – Hydrolyze