Fortran implementations of MPI Reductions.
This is a Finnish word that is roughly equivalent to what "reduction" means in MPI.
The intuitive English pronunciation is wrong. Listen to https://forvo.com/word/koota/.
We implement type-specific reductions with the length passed by value as size_t
because these will be called internally after the MPI implementation has
determined the type. We do not need two versions for large-count.
This is necessary because CFI types aren't what MPI operates on,
and Fortran types aren't necessarily C interoperable, so we cannot
use bind(C) if we declare the arrays with Fortran types.
If we do not comply with bind(C), then we can't specify the C linkage symbol name.
! call C_F_POINTER(cptr,fptr[,shape])
<type>, pointer, dimension(:) :: fiptr, fioptr
call C_F_POINTER(invec,fiptr,[len])
call C_F_POINTER(inoutvec,fioptr,[len])This is the list of reductions required by MPI:
| Op | Allowed Types |
|---|---|
MPI_MAX, MPI_MIN |
integer, floating point |
MPI_SUM, MPI_PROD |
integer, floating point, complex, |
MPI_LAND, MPI_LOR, MPI_LXOR |
logical |
MPI_BAND, MPI_BOR, MPI_BXOR |
integer, byte |