Support tile side size beyond INT_MAX
In case m
or ld
is very large, m*elemsize
ld*elemsize
may be beyond INT_MAX
and thus MPI_Type_vector
would truncate the value when using an mpi implementation where the count
and stride
parameters are only 32bit ints.
cti_allocate_datatype_node
should thus use MPI_Type_vector_c
when available.