6 int smpi_coll_tuned_allgather_SMP_simple(void *send_buf, int scount,
7 MPI_Datatype stype, void *recv_buf,
8 int rcount, MPI_Datatype rtype,
11 int src, dst, comm_size, rank;
12 MPI_Comm_size(comm, &comm_size);
13 MPI_Comm_rank(comm, &rank);
14 MPI_Aint rextent, sextent;
15 MPI_Type_extent(rtype, &rextent);
16 MPI_Type_extent(stype, &sextent);
19 int i, send_offset, recv_offset;
20 int intra_rank, inter_rank;
21 int num_core = NUM_CORE;
22 intra_rank = rank % num_core;
23 inter_rank = rank / num_core;
24 int inter_comm_size = (comm_size + num_core - 1) / num_core;
25 int num_core_in_current_smp = num_core;
27 // the last SMP node may have fewer number of running processes than all others
28 if (inter_rank == (inter_comm_size - 1)) {
29 num_core_in_current_smp = comm_size - (inter_rank * num_core);
32 recv_offset = rank * rextent * rcount;
33 MPI_Sendrecv(send_buf, scount, stype, rank, tag,
34 ((char *) recv_buf + recv_offset), rcount, rtype, rank, tag,
36 for (i = 1; i < num_core_in_current_smp; i++) {
39 (inter_rank * num_core) + (intra_rank + i) % (num_core_in_current_smp);
41 (inter_rank * num_core) + (intra_rank - i +
42 num_core_in_current_smp) %
43 (num_core_in_current_smp);
44 recv_offset = src * rextent * rcount;
46 MPI_Sendrecv(send_buf, scount, stype, dst, tag,
47 ((char *) recv_buf + recv_offset), rcount, rtype, src, tag,
52 // INTER-SMP-ALLGATHER
53 // Every root of each SMP node post INTER-Sendrecv, then do INTRA-Bcast for each receiving message
57 if (intra_rank == 0) {
58 MPI_Request *reqs, *req_ptr;
59 int num_req = (inter_comm_size - 1) * 2;
60 reqs = (MPI_Request *) malloc(num_req * sizeof(MPI_Request));
63 stat = (MPI_Status *) malloc(num_req * sizeof(MPI_Status));
65 for (i = 1; i < inter_comm_size; i++) {
67 //dst = ((inter_rank+i)%inter_comm_size) * num_core;
68 src = ((inter_rank - i + inter_comm_size) % inter_comm_size) * num_core;
69 //send_offset = (rank * sextent * scount);
70 recv_offset = (src * sextent * scount);
71 // MPI_Sendrecv((recv_buf+send_offset), (scount * num_core), stype, dst, tag,
72 // (recv_buf+recv_offset), (rcount * num_core), rtype, src, tag, comm, &status);
73 //MPIC_Isend((recv_buf+send_offset), (scount * num_core), stype, dst, tag, comm, req_ptr++);
74 MPI_Irecv(((char *) recv_buf + recv_offset), (rcount * num_core), rtype,
75 src, tag, comm, req_ptr++);
77 for (i = 1; i < inter_comm_size; i++) {
79 dst = ((inter_rank + i) % inter_comm_size) * num_core;
80 //src = ((inter_rank-i+inter_comm_size)%inter_comm_size) * num_core;
81 send_offset = (rank * sextent * scount);
82 //recv_offset = (src * sextent * scount);
83 // MPI_Sendrecv((recv_buf+send_offset), (scount * num_core), stype, dst, tag,
84 // (recv_buf+recv_offset), (rcount * num_core), rtype, src, tag, comm, &status);
85 MPI_Isend(((char *) recv_buf + send_offset), (scount * num_core), stype,
86 dst, tag, comm, req_ptr++);
87 //MPIC_Irecv((recv_buf+recv_offset), (rcount * num_core), rtype, src, tag, comm, req_ptr++);
89 MPI_Waitall(num_req, reqs, stat);
94 //INTRA-BCAST (use flat tree)
96 if (intra_rank == 0) {
97 for (i = 1; i < num_core_in_current_smp; i++) {
98 //printf("rank = %d, num = %d send to %d\n",rank, num_core_in_current_smp, (rank + i));
99 MPI_Send(recv_buf, (scount * comm_size), stype, (rank + i), tag, comm);
102 //printf("rank = %d recv from %d\n",rank, (inter_rank * num_core));
103 MPI_Recv(recv_buf, (rcount * comm_size), rtype, (inter_rank * num_core),