7 int smpi_coll_tuned_allgather_loosely_lr(void *sbuf, int scount,
8 MPI_Datatype stype, void *rbuf,
9 int rcount, MPI_Datatype rtype,
14 int i, j, send_offset, recv_offset;
15 int intra_rank, inter_rank, inter_comm_size, intra_comm_size;
16 int inter_dst, inter_src;
18 MPI_Comm_size(comm, &comm_size);
19 MPI_Comm_rank(comm, &rank);
20 MPI_Aint rextent, sextent;
21 MPI_Type_extent(rtype, &rextent);
22 MPI_Type_extent(stype, &sextent);
23 MPI_Request inter_rrequest;
24 MPI_Request rrequest_array[128];
25 MPI_Request srequest_array[128];
26 MPI_Request inter_srequest_array[128];
29 int rrequest_count = 0;
30 int srequest_count = 0;
31 int inter_srequest_count = 0;
35 intra_rank = rank % NUM_CORE;
36 inter_rank = rank / NUM_CORE;
37 inter_comm_size = (comm_size + NUM_CORE - 1) / NUM_CORE;
38 intra_comm_size = NUM_CORE;
42 //copy corresponding message from sbuf to rbuf
43 recv_offset = rank * rextent * rcount;
44 MPI_Sendrecv(sbuf, scount, stype, rank, tag,
45 (char *)rbuf + recv_offset, rcount, rtype, rank, tag, comm, &status);
48 int inter_send_offset, inter_recv_offset;
52 inter_srequest_count = 0;
54 for (i = 0; i < inter_comm_size; i++) {
56 // inter_communication
58 inter_dst = (rank + intra_comm_size) % comm_size;
59 inter_src = (rank - intra_comm_size + comm_size) % comm_size;
62 ((inter_rank - 1 - i +
63 inter_comm_size) % inter_comm_size) * intra_comm_size + intra_rank;
66 inter_comm_size) % inter_comm_size) * intra_comm_size + intra_rank;
68 inter_send_offset = dst_seg * sextent * scount;
69 inter_recv_offset = src_seg * rextent * rcount;
71 for (j = 0; j < intra_comm_size; j++) {
73 // inter communication
74 if (intra_rank == j) {
75 if (i != inter_comm_size - 1) {
77 MPI_Irecv((char *)rbuf + inter_recv_offset, rcount, rtype, inter_src, tag,
78 comm, &inter_rrequest);
79 MPI_Isend((char *)rbuf + inter_send_offset, scount, stype, inter_dst, tag,
80 comm, &inter_srequest_array[inter_srequest_count++]);
85 src = inter_rank * intra_comm_size + j;
86 dst = inter_rank * intra_comm_size + j;
90 inter_comm_size) % inter_comm_size) * intra_comm_size + j;
93 inter_comm_size) % inter_comm_size) * intra_comm_size + intra_rank;
95 send_offset = dst_seg * sextent * scount;
96 recv_offset = src_seg * rextent * rcount;
99 if (j != intra_rank) {
101 MPI_Irecv((char *)rbuf + recv_offset, rcount, rtype, src, tag, comm,
102 &rrequest_array[rrequest_count++]);
103 MPI_Isend((char *)rbuf + send_offset, scount, stype, dst, tag, comm,
104 &srequest_array[srequest_count++]);
110 // wait for inter communication to finish for these rounds (# of round equals NUM_CORE)
111 if (i != inter_comm_size - 1) {
112 MPI_Wait(&inter_rrequest, &status);
117 MPI_Waitall(rrequest_count, rrequest_array, MPI_STATUSES_IGNORE);
118 MPI_Waitall(srequest_count, srequest_array, MPI_STATUSES_IGNORE);
119 MPI_Waitall(inter_srequest_count, inter_srequest_array, MPI_STATUSES_IGNORE);