[1037] | 1 | /*! |
---|
| 2 | \file ep_create.cpp |
---|
| 3 | \since 2 may 2016 |
---|
| 4 | |
---|
| 5 | \brief Definitions of MPI endpoint function: MPI_Comm_create_endpoints |
---|
| 6 | */ |
---|
| 7 | |
---|
| 8 | #include "ep_lib.hpp" |
---|
| 9 | #include <mpi.h> |
---|
| 10 | #include "ep_declaration.hpp" |
---|
| 11 | |
---|
| 12 | using namespace std; |
---|
| 13 | |
---|
| 14 | |
---|
| 15 | |
---|
| 16 | namespace ep_lib { |
---|
| 17 | |
---|
| 18 | /*! |
---|
| 19 | Dynamic creation of endpoints for each MPI process. |
---|
| 20 | The output of this function is an array of communicator handles of length num_ep, where each handle |
---|
| 21 | corresponds to a nez local randk in the output communicator. |
---|
| 22 | Once created, endpoints behave as MPI processes. |
---|
| 23 | \param [in] mpi_comm Parent MPI communicator. |
---|
| 24 | \param [in] num_ep Number of endpoints per process. |
---|
| 25 | \param [out] info Information of the EP creation. |
---|
| 26 | \param [out] out_comm_hdls Handles of EP communicators. |
---|
| 27 | */ |
---|
| 28 | #ifdef _intelmpi |
---|
| 29 | int MPI_Comm_create_endpoints(int base_comm_ptr, int num_ep, MPI_Info info, MPI_Comm *& out_comm_hdls) |
---|
| 30 | { |
---|
| 31 | int base_rank; |
---|
| 32 | int base_size; |
---|
| 33 | |
---|
| 34 | ::MPI_Comm mpi_base_comm = static_cast< ::MPI_Comm> (base_comm_ptr); |
---|
| 35 | |
---|
| 36 | ::MPI_Comm_size(mpi_base_comm, &base_size); // ep_lib::mpi_comm_size |
---|
| 37 | ::MPI_Comm_rank(mpi_base_comm, &base_rank); // ep_lib::mpi_comm_rank |
---|
| 38 | // parent_comm can also be endpoints communicators |
---|
| 39 | |
---|
| 40 | std::vector<int> recv_num_ep(base_size); |
---|
| 41 | |
---|
| 42 | out_comm_hdls = new MPI_Comm[num_ep]; |
---|
| 43 | |
---|
| 44 | for (int idx = 0; idx < num_ep; ++idx) |
---|
| 45 | { |
---|
| 46 | out_comm_hdls[idx].is_ep = true; |
---|
| 47 | out_comm_hdls[idx].is_intercomm = false; |
---|
| 48 | out_comm_hdls[idx].ep_comm_ptr = new ep_communicator; |
---|
| 49 | out_comm_hdls[idx].mpi_comm = base_comm_ptr; |
---|
| 50 | out_comm_hdls[idx].ep_comm_ptr->comm_list = out_comm_hdls; |
---|
| 51 | out_comm_hdls[idx].ep_comm_ptr->comm_label = 0; |
---|
| 52 | } |
---|
| 53 | |
---|
| 54 | ::MPI_Allgather(&num_ep, 1, MPI_INT_STD, &recv_num_ep[0], 1, MPI_INT_STD, mpi_base_comm); |
---|
| 55 | |
---|
| 56 | |
---|
| 57 | int sum = 0; // representing total ep number of process with smaller rank |
---|
| 58 | for (int i = 0; i < base_rank; ++i) {sum += recv_num_ep[i]; } |
---|
| 59 | |
---|
| 60 | int ep_size = std::accumulate(recv_num_ep.begin(), recv_num_ep.end(), 0); |
---|
| 61 | |
---|
| 62 | out_comm_hdls[0].ep_barrier = new OMPbarrier(num_ep); |
---|
| 63 | |
---|
| 64 | out_comm_hdls[0].my_buffer = new BUFFER; |
---|
| 65 | out_comm_hdls[0].my_buffer->buf_double = new double[BUFFER_SIZE]; |
---|
| 66 | out_comm_hdls[0].my_buffer->buf_float = new float[BUFFER_SIZE]; |
---|
| 67 | out_comm_hdls[0].my_buffer->buf_int = new int[BUFFER_SIZE]; |
---|
| 68 | out_comm_hdls[0].my_buffer->buf_long = new long[BUFFER_SIZE]; |
---|
| 69 | out_comm_hdls[0].my_buffer->buf_ulong = new unsigned long[BUFFER_SIZE]; |
---|
| 70 | out_comm_hdls[0].my_buffer->buf_char = new char[BUFFER_SIZE]; |
---|
| 71 | |
---|
| 72 | out_comm_hdls[0].rank_map = new RANK_MAP; |
---|
| 73 | out_comm_hdls[0].rank_map->resize(ep_size); |
---|
| 74 | |
---|
| 75 | |
---|
| 76 | for (int i = 1; i < num_ep; i++) |
---|
| 77 | { |
---|
| 78 | out_comm_hdls[i].ep_barrier = out_comm_hdls[0].ep_barrier; |
---|
| 79 | out_comm_hdls[i].my_buffer = out_comm_hdls[0].my_buffer; |
---|
| 80 | out_comm_hdls[i].rank_map = out_comm_hdls[0].rank_map; |
---|
| 81 | } |
---|
| 82 | |
---|
| 83 | |
---|
| 84 | for (int i = 0; i < num_ep; i++) |
---|
| 85 | { |
---|
| 86 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[0] = std::make_pair(sum+i, ep_size); |
---|
| 87 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[1] = std::make_pair(i, num_ep); |
---|
| 88 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[2] = std::make_pair(base_rank, base_size); |
---|
| 89 | |
---|
| 90 | out_comm_hdls[i].ep_comm_ptr->message_queue = new Message_list; |
---|
| 91 | } |
---|
| 92 | |
---|
| 93 | |
---|
| 94 | int ind = 0; |
---|
| 95 | |
---|
| 96 | for(int i=0; i<base_size; i++) |
---|
| 97 | { |
---|
| 98 | for(int j=0; j<recv_num_ep[i]; j++) |
---|
| 99 | { |
---|
| 100 | out_comm_hdls[0].rank_map->at(ind) = make_pair(j, i); |
---|
| 101 | ind++; |
---|
| 102 | } |
---|
| 103 | } |
---|
| 104 | |
---|
| 105 | printf("ep_lib::MPI_Comm_create_endpoints() OK from int\n"); |
---|
| 106 | |
---|
| 107 | return 0; |
---|
| 108 | |
---|
| 109 | } //MPI_Comm_create_endpoints |
---|
| 110 | |
---|
| 111 | #elif _openmpi |
---|
| 112 | int MPI_Comm_create_endpoints(void* base_comm_ptr, int num_ep, MPI_Info info, MPI_Comm *& out_comm_hdls) |
---|
| 113 | { |
---|
| 114 | |
---|
| 115 | int base_rank; |
---|
| 116 | int base_size; |
---|
| 117 | |
---|
| 118 | ::MPI_Comm mpi_base_comm = static_cast< ::MPI_Comm> (base_comm_ptr); |
---|
| 119 | |
---|
| 120 | ::MPI_Comm_size(mpi_base_comm, &base_size); // ep_lib::mpi_comm_size |
---|
| 121 | ::MPI_Comm_rank(mpi_base_comm, &base_rank); // ep_lib::mpi_comm_rank |
---|
| 122 | // parent_comm can also be endpoints communicators ? |
---|
| 123 | std::vector<int> recv_num_ep(base_size); |
---|
| 124 | |
---|
| 125 | out_comm_hdls = new MPI_Comm[num_ep]; |
---|
| 126 | |
---|
| 127 | for (int idx = 0; idx < num_ep; ++idx) |
---|
| 128 | { |
---|
| 129 | out_comm_hdls[idx].is_ep = true; |
---|
| 130 | out_comm_hdls[idx].is_intercomm = false; |
---|
| 131 | out_comm_hdls[idx].ep_comm_ptr = new ep_communicator; |
---|
| 132 | out_comm_hdls[idx].mpi_comm = base_comm_ptr; |
---|
| 133 | out_comm_hdls[idx].ep_comm_ptr->comm_list = out_comm_hdls; |
---|
| 134 | out_comm_hdls[idx].ep_comm_ptr->comm_label = 0; |
---|
| 135 | } |
---|
| 136 | |
---|
| 137 | ::MPI_Allgather(&num_ep, 1, MPI_INT_STD, &recv_num_ep[0], 1, MPI_INT_STD, mpi_base_comm); |
---|
| 138 | |
---|
| 139 | int sum = 0; // representing total ep number of process with smaller rank |
---|
| 140 | for (int i = 0; i < base_rank; ++i) {sum += recv_num_ep[i]; } |
---|
| 141 | |
---|
| 142 | int ep_size = std::accumulate(recv_num_ep.begin(), recv_num_ep.end(), 0); |
---|
| 143 | |
---|
| 144 | out_comm_hdls[0].ep_barrier = new OMPbarrier(num_ep); |
---|
| 145 | |
---|
| 146 | out_comm_hdls[0].my_buffer = new BUFFER; |
---|
| 147 | out_comm_hdls[0].my_buffer->buf_double = new double[BUFFER_SIZE]; |
---|
| 148 | out_comm_hdls[0].my_buffer->buf_float = new float[BUFFER_SIZE]; |
---|
| 149 | out_comm_hdls[0].my_buffer->buf_int = new int[BUFFER_SIZE]; |
---|
| 150 | out_comm_hdls[0].my_buffer->buf_long = new long[BUFFER_SIZE]; |
---|
| 151 | out_comm_hdls[0].my_buffer->buf_ulong = new unsigned long[BUFFER_SIZE]; |
---|
| 152 | out_comm_hdls[0].my_buffer->buf_char = new char[BUFFER_SIZE]; |
---|
| 153 | |
---|
| 154 | out_comm_hdls[0].rank_map = new RANK_MAP; |
---|
| 155 | out_comm_hdls[0].rank_map->resize(ep_size); |
---|
| 156 | |
---|
| 157 | |
---|
| 158 | for (int i = 1; i < num_ep; i++) |
---|
| 159 | { |
---|
| 160 | out_comm_hdls[i].ep_barrier = out_comm_hdls[0].ep_barrier; |
---|
| 161 | out_comm_hdls[i].my_buffer = out_comm_hdls[0].my_buffer; |
---|
| 162 | out_comm_hdls[i].rank_map = out_comm_hdls[0].rank_map; |
---|
| 163 | } |
---|
| 164 | |
---|
| 165 | |
---|
| 166 | for (int i = 0; i < num_ep; i++) |
---|
| 167 | { |
---|
| 168 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[0] = std::make_pair(sum+i, ep_size); |
---|
| 169 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[1] = std::make_pair(i, num_ep); |
---|
| 170 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[2] = std::make_pair(base_rank, base_size); |
---|
| 171 | |
---|
| 172 | out_comm_hdls[i].ep_comm_ptr->message_queue = new Message_list; |
---|
| 173 | } |
---|
| 174 | |
---|
| 175 | |
---|
| 176 | int ind = 0; |
---|
| 177 | |
---|
| 178 | for(int i=0; i<base_size; i++) |
---|
| 179 | { |
---|
| 180 | for(int j=0; j<recv_num_ep[i]; j++) |
---|
| 181 | { |
---|
| 182 | out_comm_hdls[0].rank_map->at(ind) = make_pair(j, i); |
---|
| 183 | ind++; |
---|
| 184 | } |
---|
| 185 | } |
---|
| 186 | |
---|
| 187 | printf("ep_lib::MPI_Comm_create_endpoints() OK from void*\n"); |
---|
| 188 | |
---|
| 189 | return 0; |
---|
| 190 | |
---|
| 191 | } //MPI_Comm_create_endpoints |
---|
| 192 | |
---|
| 193 | #endif |
---|
| 194 | |
---|
| 195 | |
---|
| 196 | int MPI_Comm_create_endpoints(MPI_Comm base_comm, int num_ep, MPI_Info info, MPI_Comm *& out_comm_hdls) |
---|
| 197 | { |
---|
| 198 | |
---|
| 199 | int base_rank; |
---|
| 200 | int base_size; |
---|
| 201 | |
---|
| 202 | assert(!base_comm.is_ep); |
---|
| 203 | |
---|
| 204 | ::MPI_Comm mpi_base_comm = static_cast< ::MPI_Comm> (base_comm.mpi_comm); |
---|
| 205 | |
---|
| 206 | ::MPI_Comm_size(mpi_base_comm, &base_size); // ep_lib::mpi_comm_size |
---|
| 207 | ::MPI_Comm_rank(mpi_base_comm, &base_rank); // ep_lib::mpi_comm_rank |
---|
| 208 | // parent_comm can also be endpoints communicators |
---|
| 209 | |
---|
| 210 | std::vector<int> recv_num_ep(base_size); |
---|
| 211 | |
---|
| 212 | out_comm_hdls = new MPI_Comm[num_ep]; |
---|
| 213 | |
---|
| 214 | for (int idx = 0; idx < num_ep; ++idx) |
---|
| 215 | { |
---|
| 216 | out_comm_hdls[idx].is_ep = true; |
---|
| 217 | out_comm_hdls[idx].is_intercomm = false; |
---|
| 218 | out_comm_hdls[idx].ep_comm_ptr = new ep_communicator; |
---|
| 219 | out_comm_hdls[idx].mpi_comm = base_comm.mpi_comm; |
---|
| 220 | out_comm_hdls[idx].ep_comm_ptr->comm_list = out_comm_hdls; |
---|
| 221 | out_comm_hdls[idx].ep_comm_ptr->comm_label = 0; |
---|
| 222 | } |
---|
| 223 | |
---|
| 224 | ::MPI_Allgather(&num_ep, 1, MPI_INT_STD, &recv_num_ep[0], 1, MPI_INT_STD, mpi_base_comm); |
---|
| 225 | |
---|
| 226 | |
---|
| 227 | int sum = 0; // representing total ep number of process with smaller rank |
---|
| 228 | for (int i = 0; i < base_rank; ++i) {sum += recv_num_ep[i]; } |
---|
| 229 | |
---|
| 230 | int ep_size = std::accumulate(recv_num_ep.begin(), recv_num_ep.end(), 0); |
---|
| 231 | |
---|
| 232 | out_comm_hdls[0].ep_barrier = new OMPbarrier(num_ep); |
---|
| 233 | |
---|
| 234 | out_comm_hdls[0].my_buffer = new BUFFER; |
---|
| 235 | out_comm_hdls[0].my_buffer->buf_double = new double[BUFFER_SIZE]; |
---|
| 236 | out_comm_hdls[0].my_buffer->buf_float = new float[BUFFER_SIZE]; |
---|
| 237 | out_comm_hdls[0].my_buffer->buf_int = new int[BUFFER_SIZE]; |
---|
| 238 | out_comm_hdls[0].my_buffer->buf_long = new long[BUFFER_SIZE]; |
---|
| 239 | out_comm_hdls[0].my_buffer->buf_ulong = new unsigned long[BUFFER_SIZE]; |
---|
| 240 | out_comm_hdls[0].my_buffer->buf_char = new char[BUFFER_SIZE]; |
---|
| 241 | |
---|
| 242 | out_comm_hdls[0].rank_map = new RANK_MAP; |
---|
| 243 | out_comm_hdls[0].rank_map->resize(ep_size); |
---|
| 244 | |
---|
| 245 | |
---|
| 246 | for (int i = 1; i < num_ep; i++) |
---|
| 247 | { |
---|
| 248 | out_comm_hdls[i].ep_barrier = out_comm_hdls[0].ep_barrier; |
---|
| 249 | out_comm_hdls[i].my_buffer = out_comm_hdls[0].my_buffer; |
---|
| 250 | out_comm_hdls[i].rank_map = out_comm_hdls[0].rank_map; |
---|
| 251 | } |
---|
| 252 | |
---|
| 253 | |
---|
| 254 | for (int i = 0; i < num_ep; i++) |
---|
| 255 | { |
---|
| 256 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[0] = std::make_pair(sum+i, ep_size); |
---|
| 257 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[1] = std::make_pair(i, num_ep); |
---|
| 258 | out_comm_hdls[i].ep_comm_ptr->size_rank_info[2] = std::make_pair(base_rank, base_size); |
---|
| 259 | |
---|
| 260 | out_comm_hdls[i].ep_comm_ptr->message_queue = new Message_list; |
---|
| 261 | } |
---|
| 262 | |
---|
| 263 | |
---|
| 264 | int ind = 0; |
---|
| 265 | |
---|
| 266 | for(int i=0; i<base_size; i++) |
---|
| 267 | { |
---|
| 268 | for(int j=0; j<recv_num_ep[i]; j++) |
---|
| 269 | { |
---|
| 270 | out_comm_hdls[0].rank_map->at(ind) = make_pair(j, i); |
---|
| 271 | ind++; |
---|
| 272 | } |
---|
| 273 | } |
---|
| 274 | |
---|
| 275 | printf("ep_lib::MPI_Comm_create_endpoints() OK from MPI_Comm\n"); |
---|
| 276 | |
---|
| 277 | return 0; |
---|
| 278 | |
---|
| 279 | } //MPI_Comm_create_endpoints |
---|
| 280 | |
---|
| 281 | |
---|
| 282 | } //namespace ep_lib |
---|