2 /*****************************************************************************
4 * Function: alltoall_ring_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 - i -> j -> j + i.
18 Light barriers are inserted between communications in different
23 ****************************************************************************/
25 smpi_coll_tuned_alltoall_ring_light_barrier(void * send_buff, int send_count,
26 MPI_Datatype send_type, void * recv_buff,
27 int recv_count, MPI_Datatype recv_type,
30 MPI_Aint send_chunk, recv_chunk;
32 int i, src, dst, rank, num_procs, next_dst, next_src;
33 int tag = 1, success = 1; /*, failure = 0;*/
35 char send_sync = 'a', recv_sync = 'b';
36 char * send_ptr = (char *) send_buff;
37 char * recv_ptr = (char *) recv_buff;
39 MPI_Comm_rank(comm, &rank);
40 MPI_Comm_size(comm, &num_procs);
41 MPI_Type_extent(send_type, &send_chunk);
42 MPI_Type_extent(recv_type, &recv_chunk);
44 send_chunk *= send_count;
45 recv_chunk *= recv_count;
47 MPI_Sendrecv(send_ptr + rank * send_chunk, send_count, send_type, rank, tag,
48 recv_ptr + rank * recv_chunk, recv_count, recv_type, rank, tag,
51 for (i = 1; i < num_procs; i++)
53 src = (rank - i + num_procs) % num_procs;
54 dst = (rank + i) % num_procs;
56 MPI_Sendrecv(send_ptr + dst * send_chunk, send_count, send_type,
57 dst, tag, recv_ptr + src * recv_chunk, recv_count,
58 recv_type, src, tag, comm, &s);
60 if ((i + 1) < num_procs)
62 next_src = (rank - (i + 1) + num_procs) % num_procs;
63 next_dst = (rank + (i + 1) + num_procs) % num_procs;
64 MPI_Sendrecv(&send_sync, 1, MPI_CHAR, next_src, tag,
65 &recv_sync, 1, MPI_CHAR, next_dst, tag,