Fortran pair types

[jeffhammond]

Problem

The current list of enumerated pair types is rather limiting, particularly in Fortran. The Fortran pair types are designed for Fortran 77, which does not support anything like a C struct. Fortran 90+ supports user-defined types. Given that the MPI standard has never been compliant with Fortran 77, and we now have explicit support for Fortran 2008, we should leverage Fortran 90 features to improve pair types (which are used in MPI_MAXLOC and MPI_MINLOC reductions, of course).

This is a future-proof solution to the challenge of enumerating all the potentially valid pair types, which was proposed in #18. This or something similar is essential if and when we add large-count support to MPI (see https://github.com/mpiwg-large-count/large-count-issues/issues for details).

Proposal

We can create a pair type constructor along the lines of MPI_Type_create_f90_real...

MPI_Type_create_f90_pair(valtype, loctype, newtype, ierror)
    TYPE(MPI_Datatype), INTENT(IN) ::  valtype
    TYPE(MPI_Datatype), INTENT(IN) ::  loctype
    TYPE(MPI_Datatype), INTENT(OUT) ::  newtype
    INTEGER, OPTIONAL, INTENT(OUT) ::  ierror

...that returns to the user the predefined defined associated with the following Fortran type:

type pairtype
   <VALUE TYPE> :: value
   <LOCATION TYPE> :: location
end type pairtype

Note that this is a special case of MPI_Type_create_struct that only returns MPI_SUCCESS when the resulting type is valid for use with MPI_MAXLOC and MPI_MINLOC.

Changes to the Text

TODO

Impact on Implementations

This is a special case of MPI_Type_create_struct so the implementation is quite simple. Implementers merely need to determine what pair types are valid and only support those.

Having a separate function for this makes it easier to do error checking for the user.

Impact on Users

This function allows Fortran users to use modern language features instead of the awkward hacks MPI_2REAL, MPI_2DOUBLE, and MPI_2INTEGER. It also allows Fortran users to perform M**LOC reductions over more types, such as when the array values are INTEGER*8 and the locations are INTEGER, which is often INTEGER*4.

References

• Longer types for use with MPI_MINLOC and MPI_MAXLOC #18
• 16-bit floating-point support for C/C++ #65 (comment)

[tonyskjellum]

@jeffhammond , do you plan to work on this for the December timeframe; we need a champion to get this ready for reading. Otherwise, who else could we encourage to write it up?

Thanks,
Tony

[jeffhammond]

I seriously doubt I will have time for this for December. My priority is #23.

[tonyskjellum]

Ok thank you ; I’ll focus on big Mpi —- and see if we can find a fortranner for this one

[bosilca]

Predefined MPI_Op only apply to predefined MPI_Datatype. This dynamic creation of datatypes would require to also extend the definition of predefined ops and datatypes.