#include "../colls_private.hpp"
/* IMPLEMENTED BY PITCH PATARASUK
- Non-topoloty-specific (however, number of cores/node need to be changed)
+ Non-topology-specific (however, number of cores/node need to be changed)
all-reduce operation designed for smp clusters
It uses 2-layer communication: binomial for intra-communication
and rdb for inter-communication*/
-
/* ** NOTE **
Use -DMPICH2 if this code does not compile.
MPICH1 code also work on MPICH2 on our cluster and the performance are similar.
MPI_Comm comm)
{
int comm_size, rank;
- void *tmp_buf;
int tag = COLL_TAG_ALLREDUCE;
int mask, src, dst;
MPI_Status status;
rank = comm->rank();
MPI_Aint extent;
extent = dtype->get_extent();
- tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
+ unsigned char* tmp_buf = smpi_get_tmp_sendbuffer(count * extent);
/* compute intra and inter ranking */
int intra_rank, inter_rank;