[MPI3 Fortran] Start discussing new Fortran binding

[Jeff Squyres]

On Feb 19, 2008, at 10:27 AM, Craig Rasmussen wrote:

> Don't know if there are implementations where the MPI handle is not
> an integer of size of an address.  Would your suggestion work if it
> were?

Actually, does it matter?

In the MPI spec, we can just mandate this part:

type, BIND(C) :: MPI_HANDLE_TYPE
  private
....something....
  end type MPI_HANDLE_TYPE

So that MPI_HANDLE_TYPE is required to be available in Fortran.  What  
exactly it contains and how it is implemented is up to each MPI  
implementation.  They can make it an equivalent of a C pointer, or a C  
int, or whatever they need.

Let's take two obvious concrete examples:

1. Open MPI defines C MPI communicators as

   typedef struct ompi_comm_t * MPI_Comm;

As such, Open MPI's F03 interface for communicators would be:

type, BIND(C) :: MPI_COMMUNICATOR
  private
  type(C_PTR) :: handle
  end type MPI_COMMUNICATOR

2. MPICH2 defines C MPI communicators as ints.  Hence, MPICH2's F03  
interface for communicators would be:

type, BIND(C) :: MPI_COMMUNICATOR
  private
  type(C_INT) :: handle
  end type MPI_COMMUNICATOR

Is that right?

If so, my only concern is that the compiler can automatically  
translate between the F03 and C values without the need for a  
translation layer provided by the MPI implementation.  That is, a  
Fortran function of MPI_SEND that takes an MPI_COMMUNICATOR type  
parameter (as described above) can be automatically translated to the  
appropriate MPI_Comm type in C.  That is one of the goals of this  
interface, after all.

The reason that I mention this is because I'm a C guy, not a Fortran  
guy, and the above "type, BIND(C)..." stuff looks analogous to a  
struct to me.  If it is, and if that binding contains more than one  
member, are we ok on the "compiler will automatically translate it"  
front as long as there is only one member in the binding?

An excellent topic for the teleconference!  :-)

-- 
Jeff Squyres
Cisco Systems