1 #include "colls_private.h"
2 /*****************************************************************************
4 * Function: alltoall_ring_mpi_barrier
9 send_buff: send input buffer
10 send_count: number of elements to send
11 send_type: data type of elements being sent
12 recv_buff: receive output buffer
13 recv_count: number of elements to received
14 recv_type: data type of elements being received
17 * Descrp: Function works in P - 1 steps. In step i, node j - i -> j -> j + i.
18 MPI barriers are added between each two phases.
22 ****************************************************************************/
24 smpi_coll_tuned_alltoallv_ring_mpi_barrier(void *send_buff, int *send_counts, int *send_disps,
25 MPI_Datatype send_type,
26 void *recv_buff, int *recv_counts, int *recv_disps,
27 MPI_Datatype recv_type, MPI_Comm comm)
30 MPI_Aint send_chunk, recv_chunk;
31 int i, src, dst, rank, num_procs;
32 int tag = COLL_TAG_ALLTOALLV;
34 char *send_ptr = (char *) send_buff;
35 char *recv_ptr = (char *) recv_buff;
37 rank = smpi_comm_rank(comm);
38 num_procs = smpi_comm_size(comm);
39 send_chunk = smpi_datatype_get_extent(send_type);
40 recv_chunk = smpi_datatype_get_extent(recv_type);
42 for (i = 0; i < num_procs; i++) {
43 src = (rank - i + num_procs) % num_procs;
44 dst = (rank + i) % num_procs;
46 smpi_mpi_barrier(comm);
47 smpi_mpi_sendrecv(send_ptr + send_disps[dst] * send_chunk, send_counts[dst], send_type, dst,
48 tag, recv_ptr + recv_disps[src] * recv_chunk, recv_counts[src], recv_type,