2 /*****************************************************************************
4 * Function: alltoall_pair_light_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 exchanges data
18 with node i ^ j. Light barriers are inserted between
19 communications in different phases.
23 ****************************************************************************/
25 smpi_coll_tuned_alltoall_pair_light_barrier(void *send_buff, int send_count,
26 MPI_Datatype send_type,
27 void *recv_buff, int recv_count,
28 MPI_Datatype recv_type,
31 MPI_Aint send_chunk, recv_chunk;
33 int i, src, dst, rank, num_procs, next_partner;
34 int tag = 1, success = 1; /*, failure = 0; */
36 char send_sync = 'a', recv_sync = 'b';
37 char *send_ptr = (char *) send_buff;
38 char *recv_ptr = (char *) recv_buff;
40 MPI_Comm_rank(comm, &rank);
41 MPI_Comm_size(comm, &num_procs);
42 MPI_Type_extent(send_type, &send_chunk);
43 MPI_Type_extent(recv_type, &recv_chunk);
45 send_chunk *= send_count;
46 recv_chunk *= recv_count;
48 MPI_Sendrecv(send_ptr + rank * send_chunk, send_count, send_type, rank, tag,
49 recv_ptr + rank * recv_chunk, recv_count, recv_type, rank, tag,
52 for (i = 1; i < num_procs; i++) {
55 MPI_Sendrecv(send_ptr + dst * send_chunk, send_count, send_type,
56 dst, tag, recv_ptr + src * recv_chunk, recv_count,
57 recv_type, src, tag, comm, &s);
59 if ((i + 1) < num_procs) {
60 next_partner = rank ^ (i + 1);
61 MPI_Sendrecv(&send_sync, 1, MPI_CHAR, next_partner, tag,
62 &recv_sync, 1, MPI_CHAR, next_partner, tag, comm, &s);