Dimitri Komatitsch <[email protected]> writes:

Dear all,

The (only) three things to do to make mpif.h conform to recent standards (F90, F95, F2003) would be:

- change all character* declarations to character(len=...)

- change all real*8 to either double precision or real(kind=8) (both are OK)

- change all integer*8 to integer(kind=8)

Well, mpif.h is supposed to be backward-compatible to F77. Upon checking the code, I see that MPICH offers the feature of tolerating -i8 and -r8 options. Open MPI uses 'double precision' for MPI_WTIME, which would be promoted to 16 bytes under -r8, so it won't behave correctly if the user expects a real*8 (perhaps specified as real(kind=8) [1]). The MPI standard says the prototype is DOUBLE PRECISION MPI_WTIME() so I would say that MPICH is non-conforming by using REAL*8, though the standard does not make any statements about what changes when compiled with -r8. I haven't checked to see how MPI data types are mapped in case of -r8. For example, MPI_REAL is defined to a constant in mpif.h, and the MPI library (which was compiled in advance) would not be aware of whether user code was being built with -r8. Is the user expected to use MPI_REAL4 and MPI_REAL8 explicitly in this case? Is -r8 and -i8 tested somewhere in MPICH? What exactly is the user expected to do in order to use this feature? I see this comment in src/binding/f90/mpi_sizeofs.f90.in: ! If reals and doubles have been forced to the same size (e.g., with ! -i8 -r8 to compilers like g95), then the compiler may refuse to ! allow interfaces that use real and double precision (failing to ! determine which one is intended) [1] Note that real(kind=8) is not guaranteed by the Fortran standard to be 8 bytes. For example, g77 and the Salford F95 compilers denote 4-bytes reals with real(kind=1), and 8-byte reals with real(kind=2). http://www.silverfrost.com/manuals/salfordftn95.pdf