3 /*****************************************************************************
5 * Function: alltoall_pair
10 send_buff: send input buffer
11 send_count: number of elements to send
12 send_type: data type of elements being sent
13 recv_buff: receive output buffer
14 recv_count: number of elements to received
15 recv_type: data type of elements being received
18 * Descrp: Function works when P is power of two. In each phase of P - 1
19 phases, nodes in pair communicate their data.
23 ****************************************************************************/
25 int alltoall_pair(void *send_buff, int send_count, MPI_Datatype send_type,
26 void *recv_buff, int recv_count, MPI_Datatype recv_type,
30 MPI_Aint send_chunk, recv_chunk;
34 int i, src, dst, rank, num_procs;
35 int tag = 1, success = 1, failure = 0, pof2 = 1;
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 MPI_Win_create(recv_buff, num_procs * recv_chunk * send_count, recv_chunk, 0,
47 send_chunk *= send_count;
48 recv_chunk *= recv_count;
50 MPI_Win_fence(assert, win);
51 for (i = 0; i < num_procs; i++) {
53 MPI_Put(send_ptr + dst * send_chunk, send_count, send_type, dst,
54 rank * send_chunk, send_count, send_type, win);
56 MPI_Win_fence(assert, win);
62 int smpi_coll_tuned_alltoall_pair(void *send_buff, int send_count,
63 MPI_Datatype send_type,
64 void *recv_buff, int recv_count,
65 MPI_Datatype recv_type, MPI_Comm comm)
68 MPI_Aint send_chunk, recv_chunk;
70 int i, src, dst, rank, num_procs;
71 int tag = 1, success = 1;
73 char *send_ptr = (char *) send_buff;
74 char *recv_ptr = (char *) recv_buff;
76 MPI_Comm_rank(comm, &rank);
77 MPI_Comm_size(comm, &num_procs);
78 MPI_Type_extent(send_type, &send_chunk);
79 MPI_Type_extent(recv_type, &recv_chunk);
81 send_chunk *= send_count;
82 recv_chunk *= recv_count;
84 for (i = 0; i < num_procs; i++) {
86 MPI_Sendrecv(send_ptr + dst * send_chunk, send_count, send_type, dst,
87 tag, recv_ptr + src * recv_chunk, recv_count, recv_type,