1 #include "colls_private.h"
7 int smpi_coll_tuned_allgather_loosely_lr(void *sbuf, int scount,
8 MPI_Datatype stype, void *rbuf,
9 int rcount, MPI_Datatype rtype,
13 int tag = COLL_TAG_ALLGATHER;
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 comm_size = smpi_comm_size(comm);
20 int num_core = simcall_host_get_core(SIMIX_host_self());
21 // do we use the default one or the number of cores in the platform ?
22 // if the number of cores is one, the platform may be simulated with 1 node = 1 core
23 if (num_core == 1) num_core = NUM_CORE;
25 if(comm_size%num_core)
26 THROWF(arg_error,0, "allgather loosely lr algorithm can't be used with non multiple of NUM_CORE=%d number of processes ! ",num_core);
28 rank = smpi_comm_rank(comm);
29 MPI_Aint rextent, sextent;
30 rextent = smpi_datatype_get_extent(rtype);
31 sextent = smpi_datatype_get_extent(stype);
32 MPI_Request inter_rrequest;
33 MPI_Request rrequest_array[128];
34 MPI_Request srequest_array[128];
35 MPI_Request inter_srequest_array[128];
38 int rrequest_count = 0;
39 int srequest_count = 0;
40 int inter_srequest_count = 0;
44 intra_rank = rank % num_core;
45 inter_rank = rank / num_core;
46 inter_comm_size = (comm_size + num_core - 1) / num_core;
47 intra_comm_size = num_core;
51 //copy corresponding message from sbuf to rbuf
52 recv_offset = rank * rextent * rcount;
53 smpi_mpi_sendrecv(sbuf, scount, stype, rank, tag,
54 (char *)rbuf + recv_offset, rcount, rtype, rank, tag, comm, &status);
57 int inter_send_offset, inter_recv_offset;
61 inter_srequest_count = 0;
63 for (i = 0; i < inter_comm_size; i++) {
65 // inter_communication
67 inter_dst = (rank + intra_comm_size) % comm_size;
68 inter_src = (rank - intra_comm_size + comm_size) % comm_size;
71 ((inter_rank - 1 - i +
72 inter_comm_size) % inter_comm_size) * intra_comm_size + intra_rank;
75 inter_comm_size) % inter_comm_size) * intra_comm_size + intra_rank;
77 inter_send_offset = dst_seg * sextent * scount;
78 inter_recv_offset = src_seg * rextent * rcount;
80 for (j = 0; j < intra_comm_size; j++) {
82 // inter communication
83 if (intra_rank == j) {
84 if (i != inter_comm_size - 1) {
86 inter_rrequest = smpi_mpi_irecv((char *)rbuf + inter_recv_offset, rcount, rtype,
87 inter_src, tag, comm);
88 inter_srequest_array[inter_srequest_count++] = smpi_mpi_isend((char *)rbuf + inter_send_offset, scount, stype,
89 inter_dst, tag, comm);
93 src = inter_rank * intra_comm_size + j;
94 dst = inter_rank * intra_comm_size + j;
98 inter_comm_size) % inter_comm_size) * intra_comm_size + j;
101 inter_comm_size) % inter_comm_size) * intra_comm_size + intra_rank;
103 send_offset = dst_seg * sextent * scount;
104 recv_offset = src_seg * rextent * rcount;
107 if (j != intra_rank) {
109 rrequest_array[rrequest_count++] = smpi_mpi_irecv((char *)rbuf + recv_offset, rcount, rtype, src, tag, comm);
110 srequest_array[srequest_count++] = smpi_mpi_isend((char *)rbuf + send_offset, scount, stype, dst, tag, comm);
116 // wait for inter communication to finish for these rounds (# of round equals num_core)
117 if (i != inter_comm_size - 1) {
118 smpi_mpi_wait(&inter_rrequest, &status);
123 smpi_mpi_waitall(rrequest_count, rrequest_array, MPI_STATUSES_IGNORE);
124 smpi_mpi_waitall(srequest_count, srequest_array, MPI_STATUSES_IGNORE);
125 smpi_mpi_waitall(inter_srequest_count, inter_srequest_array, MPI_STATUSES_IGNORE);