[1381] | 1 | #include "ep_lib.hpp" |
---|
| 2 | #include <mpi.h> |
---|
| 3 | #include "ep_declaration.hpp" |
---|
| 4 | #include "ep_mpi.hpp" |
---|
| 5 | |
---|
| 6 | using namespace std; |
---|
| 7 | |
---|
| 8 | namespace ep_lib |
---|
| 9 | { |
---|
| 10 | int MPI_Intercomm_create(MPI_Comm local_comm, int local_leader, MPI_Comm peer_comm, int remote_leader, int tag, MPI_Comm *newintercomm) |
---|
| 11 | { |
---|
[1500] | 12 | assert(local_comm->is_ep); |
---|
[1381] | 13 | |
---|
| 14 | int ep_rank, ep_rank_loc, mpi_rank; |
---|
| 15 | int ep_size, num_ep, mpi_size; |
---|
| 16 | |
---|
[1500] | 17 | ep_rank = local_comm->ep_comm_ptr->size_rank_info[0].first; |
---|
| 18 | ep_rank_loc = local_comm->ep_comm_ptr->size_rank_info[1].first; |
---|
| 19 | mpi_rank = local_comm->ep_comm_ptr->size_rank_info[2].first; |
---|
| 20 | ep_size = local_comm->ep_comm_ptr->size_rank_info[0].second; |
---|
| 21 | num_ep = local_comm->ep_comm_ptr->size_rank_info[1].second; |
---|
| 22 | mpi_size = local_comm->ep_comm_ptr->size_rank_info[2].second; |
---|
[1381] | 23 | |
---|
| 24 | |
---|
| 25 | MPI_Barrier(local_comm); |
---|
| 26 | |
---|
[1515] | 27 | int leader_ranks_in_peer[3]; // local_leader_rank_in_peer |
---|
| 28 | // remote_leader_rank_in_peer |
---|
| 29 | // size of peer |
---|
[1499] | 30 | |
---|
[1515] | 31 | if(ep_rank == local_leader) |
---|
| 32 | { |
---|
| 33 | MPI_Comm_rank(peer_comm, &leader_ranks_in_peer[0]); |
---|
| 34 | leader_ranks_in_peer[1] = remote_leader; |
---|
| 35 | MPI_Comm_size(peer_comm, &leader_ranks_in_peer[2]); |
---|
| 36 | } |
---|
[1499] | 37 | |
---|
[1515] | 38 | MPI_Bcast(leader_ranks_in_peer, 3, MPI_INT, local_leader, local_comm); |
---|
| 39 | |
---|
| 40 | if(leader_ranks_in_peer[0] != leader_ranks_in_peer[2]) |
---|
| 41 | { |
---|
| 42 | Debug("calling MPI_Intercomm_create_kernel\n"); |
---|
| 43 | return MPI_Intercomm_create_kernel(local_comm, local_leader, peer_comm, remote_leader, tag, newintercomm); |
---|
| 44 | } |
---|
| 45 | |
---|
| 46 | else |
---|
| 47 | { |
---|
| 48 | if(leader_ranks_in_peer[2] == 1) |
---|
| 49 | { |
---|
| 50 | Debug("calling MPI_Intercomm_create_unique\n"); |
---|
| 51 | return MPI_Intercomm_create_unique_leader(local_comm, local_leader, peer_comm, remote_leader, tag, newintercomm); |
---|
| 52 | |
---|
| 53 | } |
---|
| 54 | else |
---|
| 55 | { |
---|
| 56 | Debug("calling MPI_Intercomm_create_world\n"); |
---|
| 57 | return MPI_Intercomm_create_from_world(local_comm, local_leader, peer_comm, remote_leader, tag, newintercomm); |
---|
| 58 | } |
---|
| 59 | |
---|
| 60 | } |
---|
| 61 | |
---|
| 62 | |
---|
| 63 | |
---|
| 64 | |
---|
| 65 | |
---|
[1381] | 66 | int leader_ranks[6]; //! 0: rank in world, 1: mpi_size, 2: rank_in_peer. |
---|
| 67 | //! 3, 4, 5 : remote |
---|
| 68 | |
---|
| 69 | bool is_decider = false; |
---|
| 70 | |
---|
| 71 | |
---|
| 72 | if(ep_rank == local_leader) |
---|
| 73 | { |
---|
| 74 | MPI_Comm_rank(MPI_COMM_WORLD, &leader_ranks[0]); |
---|
| 75 | |
---|
| 76 | leader_ranks[1] = mpi_size; |
---|
| 77 | MPI_Comm_rank(peer_comm, &leader_ranks[2]); |
---|
| 78 | |
---|
| 79 | MPI_Request request[2]; |
---|
| 80 | MPI_Status status[2]; |
---|
| 81 | |
---|
| 82 | MPI_Isend(&leader_ranks[0], 3, MPI_INT, remote_leader, tag, peer_comm, &request[0]); |
---|
| 83 | MPI_Irecv(&leader_ranks[3], 3, MPI_INT, remote_leader, tag, peer_comm, &request[1]); |
---|
| 84 | |
---|
| 85 | MPI_Waitall(2, request, status); |
---|
| 86 | |
---|
| 87 | } |
---|
| 88 | |
---|
| 89 | |
---|
| 90 | MPI_Bcast(leader_ranks, 6, MPI_INT, local_leader, local_comm); |
---|
| 91 | |
---|
| 92 | |
---|
| 93 | MPI_Barrier(local_comm); |
---|
| 94 | |
---|
| 95 | |
---|
| 96 | if(leader_ranks[0] == leader_ranks[3]) |
---|
| 97 | { |
---|
| 98 | if( leader_ranks[1] * leader_ranks[4] == 1) |
---|
| 99 | { |
---|
[1511] | 100 | if(ep_rank == local_leader) printf("calling MPI_Intercomm_create_unique_leader\n"); |
---|
[1500] | 101 | local_comm->ep_comm_ptr->comm_label = -99; |
---|
[1381] | 102 | |
---|
| 103 | return MPI_Intercomm_create_unique_leader(local_comm, local_leader, peer_comm, remote_leader, tag, newintercomm); |
---|
| 104 | } |
---|
| 105 | else // leader_ranks[1] * leader_ranks[4] != 1 |
---|
| 106 | { |
---|
| 107 | // change leader |
---|
| 108 | int new_local_leader; |
---|
| 109 | |
---|
| 110 | if(leader_ranks[2] < leader_ranks[5]) |
---|
| 111 | { |
---|
| 112 | if(leader_ranks[1] > 1) //! change leader |
---|
| 113 | { |
---|
| 114 | // change leader |
---|
| 115 | is_decider = true; |
---|
[1503] | 116 | int target = local_comm->ep_rank_map->at(local_leader).second; |
---|
[1381] | 117 | { |
---|
| 118 | for(int i=0; i<ep_size; i++) |
---|
| 119 | { |
---|
[1503] | 120 | if(local_comm->ep_rank_map->at(i).second != target && local_comm->ep_rank_map->at(i).first == 0) |
---|
[1381] | 121 | { |
---|
| 122 | new_local_leader = i; |
---|
| 123 | break; |
---|
| 124 | } |
---|
| 125 | } |
---|
| 126 | } |
---|
| 127 | } |
---|
| 128 | else |
---|
| 129 | { |
---|
| 130 | new_local_leader = local_leader; |
---|
| 131 | } |
---|
| 132 | } |
---|
| 133 | else |
---|
| 134 | { |
---|
| 135 | if(leader_ranks[4] == 1) |
---|
| 136 | { |
---|
| 137 | // change leader |
---|
| 138 | is_decider = true; |
---|
[1503] | 139 | int target = local_comm->ep_rank_map->at(local_leader).second; |
---|
[1381] | 140 | { |
---|
| 141 | for(int i=0; i<ep_size; i++) |
---|
| 142 | { |
---|
[1503] | 143 | if(local_comm->ep_rank_map->at(i).second != target && local_comm->ep_rank_map->at(i).first == 0) |
---|
[1381] | 144 | { |
---|
| 145 | new_local_leader = i; |
---|
| 146 | break; |
---|
| 147 | } |
---|
| 148 | } |
---|
| 149 | } |
---|
| 150 | } |
---|
| 151 | else |
---|
| 152 | { |
---|
| 153 | new_local_leader = local_leader; |
---|
| 154 | } |
---|
| 155 | } |
---|
| 156 | |
---|
| 157 | |
---|
| 158 | int new_tag_in_world; |
---|
| 159 | |
---|
| 160 | int leader_in_world[2]; |
---|
| 161 | |
---|
| 162 | |
---|
| 163 | if(is_decider) |
---|
| 164 | { |
---|
| 165 | if(ep_rank == new_local_leader) |
---|
| 166 | { |
---|
| 167 | new_tag_in_world = TAG++; |
---|
| 168 | } |
---|
| 169 | MPI_Bcast(&new_tag_in_world, 1, MPI_INT, new_local_leader, local_comm); |
---|
| 170 | if(ep_rank == local_leader) MPI_Send(&new_tag_in_world, 1, MPI_INT, remote_leader, tag, peer_comm); |
---|
| 171 | } |
---|
| 172 | else |
---|
| 173 | { |
---|
| 174 | if(ep_rank == local_leader) |
---|
| 175 | { |
---|
| 176 | MPI_Status status; |
---|
| 177 | MPI_Recv(&new_tag_in_world, 1, MPI_INT, remote_leader, tag, peer_comm, &status); |
---|
| 178 | } |
---|
| 179 | MPI_Bcast(&new_tag_in_world, 1, MPI_INT, new_local_leader, local_comm); |
---|
| 180 | } |
---|
| 181 | |
---|
| 182 | |
---|
| 183 | if(ep_rank == new_local_leader) |
---|
| 184 | { |
---|
[1500] | 185 | ::MPI_Comm_rank(to_mpi_comm(MPI_COMM_WORLD->mpi_comm), &leader_in_world[0]); |
---|
[1381] | 186 | } |
---|
| 187 | |
---|
| 188 | MPI_Bcast(&leader_in_world[0], 1, MPI_INT, new_local_leader, local_comm); |
---|
| 189 | |
---|
| 190 | |
---|
| 191 | if(ep_rank == local_leader) |
---|
| 192 | { |
---|
| 193 | MPI_Request request[2]; |
---|
| 194 | MPI_Status status[2]; |
---|
| 195 | |
---|
| 196 | MPI_Isend(&leader_in_world[0], 1, MPI_INT, remote_leader, tag, peer_comm, &request[0]); |
---|
| 197 | MPI_Irecv(&leader_in_world[1], 1, MPI_INT, remote_leader, tag, peer_comm, &request[1]); |
---|
| 198 | |
---|
| 199 | MPI_Waitall(2, request, status); |
---|
| 200 | } |
---|
| 201 | |
---|
| 202 | MPI_Bcast(&leader_in_world[1], 1, MPI_INT, local_leader, local_comm); |
---|
| 203 | |
---|
[1500] | 204 | local_comm->ep_comm_ptr->comm_label = tag; |
---|
[1381] | 205 | |
---|
[1511] | 206 | if(ep_rank == local_leader) printf("calling MPI_Intercomm_create_from_world\n"); |
---|
[1381] | 207 | |
---|
[1500] | 208 | return MPI_Intercomm_create_from_world(local_comm, new_local_leader, MPI_COMM_WORLD->mpi_comm, leader_in_world[1], new_tag_in_world, newintercomm); |
---|
[1381] | 209 | |
---|
| 210 | } |
---|
| 211 | } |
---|
| 212 | |
---|
[1515] | 213 | |
---|
[1381] | 214 | |
---|
| 215 | } |
---|
| 216 | |
---|
| 217 | int MPI_Comm_test_inter(MPI_Comm comm, int *flag) |
---|
| 218 | { |
---|
| 219 | *flag = false; |
---|
[1500] | 220 | if(comm->is_ep) |
---|
[1381] | 221 | { |
---|
[1500] | 222 | *flag = comm->is_intercomm; |
---|
[1381] | 223 | return 0; |
---|
| 224 | } |
---|
[1500] | 225 | else if(comm->mpi_comm != static_cast< ::MPI_Comm*>(MPI_COMM_NULL->mpi_comm)) |
---|
[1381] | 226 | { |
---|
[1500] | 227 | ::MPI_Comm mpi_comm = to_mpi_comm(comm->mpi_comm); |
---|
[1381] | 228 | |
---|
| 229 | ::MPI_Comm_test_inter(mpi_comm, flag); |
---|
| 230 | return 0; |
---|
| 231 | } |
---|
| 232 | return 0; |
---|
| 233 | } |
---|
| 234 | |
---|
| 235 | |
---|
| 236 | } |
---|