-/* Copyright (c) 2013-2017. The SimGrid Team.
+/* Copyright (c) 2013-2019. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
-#include "../colls_private.h"
-/* IMPLEMENTED BY PITCH PATARASUK
- Non-topoloty-specific (however, number of cores/node need to be changed)
+#include "../colls_private.hpp"
+/* IMPLEMENTED BY PITCH PATARASUK
+ Non-topoloty-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
+ It uses 2-layer communication: binomial for intra-communication
and rdb for inter-communication*/
*/
namespace simgrid{
namespace smpi{
-int Coll_allreduce_smp_rdb::allreduce(void *send_buf, void *recv_buf, int count,
+int Coll_allreduce_smp_rdb::allreduce(const void *send_buf, void *recv_buf, int count,
MPI_Datatype dtype, MPI_Op op,
MPI_Comm comm)
{
int comm_size, rank;
- void *tmp_buf;
int tag = COLL_TAG_ALLREDUCE;
int mask, src, dst;
MPI_Status status;
MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
#else
MPI_User_function *uop;
- struct MPIR_OP *op_ptr;
+ MPIR_OP *op_ptr;
op_ptr = MPIR_ToPointer(op);
uop = op_ptr->op;
#endif
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;
} /* end binomial reduce intra-communication */
- /* start rdb (recursive doubling) all-reduce inter-communication
+ /* start rdb (recursive doubling) all-reduce inter-communication
between each SMP nodes : each node only have one process that can communicate
to other nodes */
if (intra_rank == 0) {
newrank = inter_rank - rem;
}
- /* example inter-communication RDB rank change algorithm
+ /* example inter-communication RDB rank change algorithm
0,4,8,12..36 <= true rank (assume 4 core per SMP)
0123 4567 89 <= inter_rank
1 3 4567 89 (1,3 got data from 0,2 : 0,2 will be idle until the end)
- 0 1 4567 89
+ 0 1 4567 89
0 1 2345 67 => newrank
*/
}
}
- /* non pof2 case
- left-over processes (all even ranks: < 2 * rem) get the result
+ /* non pof2 case
+ left-over processes (all even ranks: < 2 * rem) get the result
*/
if (inter_rank < 2 * rem) {
if (inter_rank % 2) {