[1646] | 1 | #ifdef _usingEP |
---|
[1603] | 2 | /*! |
---|
| 3 | \file ep_reduce.cpp |
---|
| 4 | \since 2 may 2016 |
---|
| 5 | |
---|
| 6 | \brief Definitions of MPI collective function: MPI_Reduce, MPI_Allreduce |
---|
| 7 | */ |
---|
| 8 | |
---|
| 9 | #include "ep_lib.hpp" |
---|
| 10 | #include <mpi.h> |
---|
| 11 | #include "ep_declaration.hpp" |
---|
| 12 | #include "ep_mpi.hpp" |
---|
| 13 | |
---|
| 14 | using namespace std; |
---|
| 15 | |
---|
| 16 | |
---|
| 17 | namespace ep_lib { |
---|
| 18 | |
---|
| 19 | template<typename T> |
---|
| 20 | T max_op(T a, T b) |
---|
| 21 | { |
---|
| 22 | return max(a,b); |
---|
| 23 | } |
---|
| 24 | |
---|
| 25 | template<typename T> |
---|
| 26 | T min_op(T a, T b) |
---|
| 27 | { |
---|
| 28 | return min(a,b); |
---|
| 29 | } |
---|
| 30 | |
---|
| 31 | template<typename T> |
---|
| 32 | T lor_op(T a, T b) |
---|
| 33 | { |
---|
| 34 | return a||b; |
---|
| 35 | } |
---|
| 36 | |
---|
| 37 | template<typename T> |
---|
| 38 | void reduce_max(const T * buffer, T* recvbuf, int count) |
---|
| 39 | { |
---|
| 40 | transform(buffer, buffer+count, recvbuf, recvbuf, max_op<T>); |
---|
| 41 | } |
---|
| 42 | |
---|
| 43 | template<typename T> |
---|
| 44 | void reduce_min(const T * buffer, T* recvbuf, int count) |
---|
| 45 | { |
---|
| 46 | transform(buffer, buffer+count, recvbuf, recvbuf, min_op<T>); |
---|
| 47 | } |
---|
| 48 | |
---|
| 49 | template<typename T> |
---|
| 50 | void reduce_sum(const T * buffer, T* recvbuf, int count) |
---|
| 51 | { |
---|
| 52 | transform(buffer, buffer+count, recvbuf, recvbuf, std::plus<T>()); |
---|
| 53 | } |
---|
| 54 | |
---|
| 55 | template<typename T> |
---|
| 56 | void reduce_lor(const T * buffer, T* recvbuf, int count) |
---|
| 57 | { |
---|
| 58 | transform(buffer, buffer+count, recvbuf, recvbuf, lor_op<T>); |
---|
| 59 | } |
---|
| 60 | |
---|
| 61 | int MPI_Reduce_local(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int local_root, MPI_Comm comm) |
---|
| 62 | { |
---|
| 63 | assert(valid_type(datatype)); |
---|
| 64 | assert(valid_op(op)); |
---|
| 65 | |
---|
| 66 | ::MPI_Aint datasize, lb; |
---|
| 67 | ::MPI_Type_get_extent(to_mpi_type(datatype), &lb, &datasize); |
---|
| 68 | |
---|
| 69 | int ep_rank = comm->ep_comm_ptr->size_rank_info[0].first; |
---|
| 70 | int ep_rank_loc = comm->ep_comm_ptr->size_rank_info[1].first; |
---|
| 71 | int num_ep = comm->ep_comm_ptr->size_rank_info[1].second; |
---|
| 72 | |
---|
| 73 | #pragma omp critical (_reduce) |
---|
| 74 | comm->my_buffer->void_buffer[ep_rank_loc] = const_cast< void* >(sendbuf); |
---|
| 75 | |
---|
| 76 | MPI_Barrier_local(comm); |
---|
| 77 | |
---|
| 78 | if(ep_rank_loc == local_root) |
---|
| 79 | { |
---|
| 80 | |
---|
| 81 | memcpy(recvbuf, comm->my_buffer->void_buffer[0], datasize * count); |
---|
| 82 | |
---|
| 83 | if(op == MPI_MAX) |
---|
| 84 | { |
---|
| 85 | if(datatype == MPI_INT) |
---|
| 86 | { |
---|
| 87 | assert(datasize == sizeof(int)); |
---|
| 88 | for(int i=1; i<num_ep; i++) |
---|
| 89 | reduce_max<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
| 90 | } |
---|
| 91 | |
---|
| 92 | else if(datatype == MPI_FLOAT) |
---|
| 93 | { |
---|
| 94 | assert(datasize == sizeof(float)); |
---|
| 95 | for(int i=1; i<num_ep; i++) |
---|
| 96 | reduce_max<float>(static_cast<float*>(comm->my_buffer->void_buffer[i]), static_cast<float*>(recvbuf), count); |
---|
| 97 | } |
---|
| 98 | |
---|
| 99 | else if(datatype == MPI_DOUBLE) |
---|
| 100 | { |
---|
| 101 | assert(datasize == sizeof(double)); |
---|
| 102 | for(int i=1; i<num_ep; i++) |
---|
| 103 | reduce_max<double>(static_cast<double*>(comm->my_buffer->void_buffer[i]), static_cast<double*>(recvbuf), count); |
---|
| 104 | } |
---|
| 105 | |
---|
| 106 | else if(datatype == MPI_CHAR) |
---|
| 107 | { |
---|
| 108 | assert(datasize == sizeof(char)); |
---|
| 109 | for(int i=1; i<num_ep; i++) |
---|
| 110 | reduce_max<char>(static_cast<char*>(comm->my_buffer->void_buffer[i]), static_cast<char*>(recvbuf), count); |
---|
| 111 | } |
---|
| 112 | |
---|
| 113 | else if(datatype == MPI_LONG) |
---|
| 114 | { |
---|
| 115 | assert(datasize == sizeof(long)); |
---|
| 116 | for(int i=1; i<num_ep; i++) |
---|
| 117 | reduce_max<long>(static_cast<long*>(comm->my_buffer->void_buffer[i]), static_cast<long*>(recvbuf), count); |
---|
| 118 | } |
---|
| 119 | |
---|
| 120 | else if(datatype == MPI_UNSIGNED_LONG) |
---|
| 121 | { |
---|
| 122 | assert(datasize == sizeof(unsigned long)); |
---|
| 123 | for(int i=1; i<num_ep; i++) |
---|
| 124 | reduce_max<unsigned long>(static_cast<unsigned long*>(comm->my_buffer->void_buffer[i]), static_cast<unsigned long*>(recvbuf), count); |
---|
| 125 | } |
---|
| 126 | |
---|
| 127 | else if(datatype == MPI_LONG_LONG_INT) |
---|
| 128 | { |
---|
| 129 | assert(datasize == sizeof(long long)); |
---|
| 130 | for(int i=1; i<num_ep; i++) |
---|
| 131 | reduce_max<long long>(static_cast<long long*>(comm->my_buffer->void_buffer[i]), static_cast<long long*>(recvbuf), count); |
---|
| 132 | } |
---|
| 133 | |
---|
| 134 | else |
---|
| 135 | { |
---|
| 136 | printf("datatype Error in ep_reduce : INT, FLOAT, DOUBLE, CHAR, LONG, UNSIGNED_LONG, LONG_LONG_INT\n"); |
---|
| 137 | MPI_Abort(comm, 0); |
---|
| 138 | } |
---|
| 139 | |
---|
| 140 | } |
---|
| 141 | |
---|
| 142 | else if(op == MPI_MIN) |
---|
| 143 | { |
---|
| 144 | if(datatype ==MPI_INT) |
---|
| 145 | { |
---|
| 146 | assert(datasize == sizeof(int)); |
---|
| 147 | for(int i=1; i<num_ep; i++) |
---|
| 148 | reduce_min<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
| 149 | } |
---|
| 150 | |
---|
| 151 | else if(datatype == MPI_FLOAT) |
---|
| 152 | { |
---|
| 153 | assert(datasize == sizeof(float)); |
---|
| 154 | for(int i=1; i<num_ep; i++) |
---|
| 155 | reduce_min<float>(static_cast<float*>(comm->my_buffer->void_buffer[i]), static_cast<float*>(recvbuf), count); |
---|
| 156 | } |
---|
| 157 | |
---|
| 158 | else if(datatype == MPI_DOUBLE) |
---|
| 159 | { |
---|
| 160 | assert(datasize == sizeof(double)); |
---|
| 161 | for(int i=1; i<num_ep; i++) |
---|
| 162 | reduce_min<double>(static_cast<double*>(comm->my_buffer->void_buffer[i]), static_cast<double*>(recvbuf), count); |
---|
| 163 | } |
---|
| 164 | |
---|
| 165 | else if(datatype == MPI_CHAR) |
---|
| 166 | { |
---|
| 167 | assert(datasize == sizeof(char)); |
---|
| 168 | for(int i=1; i<num_ep; i++) |
---|
| 169 | reduce_min<char>(static_cast<char*>(comm->my_buffer->void_buffer[i]), static_cast<char*>(recvbuf), count); |
---|
| 170 | } |
---|
| 171 | |
---|
| 172 | else if(datatype == MPI_LONG) |
---|
| 173 | { |
---|
| 174 | assert(datasize == sizeof(long)); |
---|
| 175 | for(int i=1; i<num_ep; i++) |
---|
| 176 | reduce_min<long>(static_cast<long*>(comm->my_buffer->void_buffer[i]), static_cast<long*>(recvbuf), count); |
---|
| 177 | } |
---|
| 178 | |
---|
| 179 | else if(datatype == MPI_UNSIGNED_LONG) |
---|
| 180 | { |
---|
| 181 | assert(datasize == sizeof(unsigned long)); |
---|
| 182 | for(int i=1; i<num_ep; i++) |
---|
| 183 | reduce_min<unsigned long>(static_cast<unsigned long*>(comm->my_buffer->void_buffer[i]), static_cast<unsigned long*>(recvbuf), count); |
---|
| 184 | } |
---|
| 185 | |
---|
| 186 | else if(datatype == MPI_LONG_LONG_INT) |
---|
| 187 | { |
---|
| 188 | assert(datasize == sizeof(long long)); |
---|
| 189 | for(int i=1; i<num_ep; i++) |
---|
| 190 | reduce_min<long long>(static_cast<long long*>(comm->my_buffer->void_buffer[i]), static_cast<long long*>(recvbuf), count); |
---|
| 191 | } |
---|
| 192 | |
---|
| 193 | else |
---|
| 194 | { |
---|
| 195 | printf("datatype Error in ep_reduce : INT, FLOAT, DOUBLE, CHAR, LONG, UNSIGNED_LONG, LONG_LONG_INT\n"); |
---|
| 196 | MPI_Abort(comm, 0); |
---|
| 197 | } |
---|
| 198 | |
---|
| 199 | } |
---|
| 200 | |
---|
| 201 | |
---|
| 202 | else if(op == MPI_SUM) |
---|
| 203 | { |
---|
| 204 | if(datatype==MPI_INT) |
---|
| 205 | { |
---|
| 206 | assert(datasize == sizeof(int)); |
---|
| 207 | for(int i=1; i<num_ep; i++) |
---|
| 208 | reduce_sum<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
| 209 | } |
---|
| 210 | |
---|
| 211 | else if(datatype == MPI_FLOAT) |
---|
| 212 | { |
---|
| 213 | assert(datasize == sizeof(float)); |
---|
| 214 | for(int i=1; i<num_ep; i++) |
---|
| 215 | reduce_sum<float>(static_cast<float*>(comm->my_buffer->void_buffer[i]), static_cast<float*>(recvbuf), count); |
---|
| 216 | } |
---|
| 217 | |
---|
| 218 | else if(datatype == MPI_DOUBLE) |
---|
| 219 | { |
---|
| 220 | assert(datasize == sizeof(double)); |
---|
| 221 | for(int i=1; i<num_ep; i++) |
---|
| 222 | reduce_sum<double>(static_cast<double*>(comm->my_buffer->void_buffer[i]), static_cast<double*>(recvbuf), count); |
---|
| 223 | } |
---|
| 224 | |
---|
| 225 | else if(datatype == MPI_CHAR) |
---|
| 226 | { |
---|
| 227 | assert(datasize == sizeof(char)); |
---|
| 228 | for(int i=1; i<num_ep; i++) |
---|
| 229 | reduce_sum<char>(static_cast<char*>(comm->my_buffer->void_buffer[i]), static_cast<char*>(recvbuf), count); |
---|
| 230 | } |
---|
| 231 | |
---|
| 232 | else if(datatype == MPI_LONG) |
---|
| 233 | { |
---|
| 234 | assert(datasize == sizeof(long)); |
---|
| 235 | for(int i=1; i<num_ep; i++) |
---|
| 236 | reduce_sum<long>(static_cast<long*>(comm->my_buffer->void_buffer[i]), static_cast<long*>(recvbuf), count); |
---|
| 237 | } |
---|
| 238 | |
---|
| 239 | else if(datatype ==MPI_UNSIGNED_LONG) |
---|
| 240 | { |
---|
| 241 | assert(datasize == sizeof(unsigned long)); |
---|
| 242 | for(int i=1; i<num_ep; i++) |
---|
| 243 | reduce_sum<unsigned long>(static_cast<unsigned long*>(comm->my_buffer->void_buffer[i]), static_cast<unsigned long*>(recvbuf), count); |
---|
| 244 | } |
---|
| 245 | |
---|
| 246 | else if(datatype ==MPI_LONG_LONG_INT) |
---|
| 247 | { |
---|
| 248 | assert(datasize == sizeof(long long)); |
---|
| 249 | for(int i=1; i<num_ep; i++) |
---|
| 250 | reduce_sum<long long>(static_cast<long long*>(comm->my_buffer->void_buffer[i]), static_cast<long long*>(recvbuf), count); |
---|
| 251 | } |
---|
| 252 | |
---|
| 253 | else |
---|
| 254 | { |
---|
| 255 | printf("datatype Error in ep_reduce : INT, FLOAT, DOUBLE, CHAR, LONG, UNSIGNED_LONG, LONG_LONG_INT\n"); |
---|
| 256 | MPI_Abort(comm, 0); |
---|
| 257 | } |
---|
| 258 | |
---|
| 259 | } |
---|
| 260 | |
---|
| 261 | else if(op == MPI_LOR) |
---|
| 262 | { |
---|
| 263 | if(datatype != MPI_INT) |
---|
| 264 | printf("datatype Error, must be MPI_INT\n"); |
---|
| 265 | else |
---|
| 266 | { |
---|
| 267 | assert(datasize == sizeof(int)); |
---|
| 268 | for(int i=1; i<num_ep; i++) |
---|
| 269 | reduce_lor<int>(static_cast<int*>(comm->my_buffer->void_buffer[i]), static_cast<int*>(recvbuf), count); |
---|
| 270 | } |
---|
| 271 | } |
---|
| 272 | |
---|
| 273 | else |
---|
| 274 | { |
---|
| 275 | printf("op type Error in ep_reduce : MPI_MAX, MPI_MIN, MPI_SUM, MPI_LOR\n"); |
---|
| 276 | MPI_Abort(comm, 0); |
---|
| 277 | } |
---|
| 278 | } |
---|
| 279 | |
---|
| 280 | MPI_Barrier_local(comm); |
---|
| 281 | |
---|
| 282 | } |
---|
| 283 | |
---|
| 284 | |
---|
| 285 | int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) |
---|
| 286 | { |
---|
| 287 | |
---|
| 288 | if(!comm->is_ep) return ::MPI_Reduce(sendbuf, recvbuf, count, to_mpi_type(datatype), to_mpi_op(op), root, to_mpi_comm(comm->mpi_comm)); |
---|
| 289 | if(comm->is_intercomm) return MPI_Reduce_intercomm(sendbuf, recvbuf, count, datatype, op, root, comm); |
---|
| 290 | |
---|
| 291 | |
---|
| 292 | |
---|
| 293 | int ep_rank = comm->ep_comm_ptr->size_rank_info[0].first; |
---|
| 294 | int ep_rank_loc = comm->ep_comm_ptr->size_rank_info[1].first; |
---|
| 295 | int mpi_rank = comm->ep_comm_ptr->size_rank_info[2].first; |
---|
| 296 | int ep_size = comm->ep_comm_ptr->size_rank_info[0].second; |
---|
| 297 | int num_ep = comm->ep_comm_ptr->size_rank_info[1].second; |
---|
| 298 | int mpi_size = comm->ep_comm_ptr->size_rank_info[2].second; |
---|
| 299 | |
---|
| 300 | int root_mpi_rank = comm->ep_rank_map->at(root).second; |
---|
| 301 | int root_ep_loc = comm->ep_rank_map->at(root).first; |
---|
| 302 | |
---|
| 303 | ::MPI_Aint datasize, lb; |
---|
| 304 | |
---|
| 305 | ::MPI_Type_get_extent(*(static_cast< ::MPI_Datatype*>(datatype)), &lb, &datasize); |
---|
| 306 | |
---|
| 307 | bool is_master = (ep_rank_loc==0 && mpi_rank != root_mpi_rank ) || ep_rank == root; |
---|
| 308 | bool is_root = ep_rank == root; |
---|
| 309 | |
---|
| 310 | void* local_recvbuf; |
---|
| 311 | |
---|
| 312 | if(is_master) |
---|
| 313 | { |
---|
| 314 | local_recvbuf = new void*[datasize * count]; |
---|
| 315 | } |
---|
| 316 | |
---|
| 317 | if(mpi_rank == root_mpi_rank) MPI_Reduce_local(sendbuf, local_recvbuf, count, datatype, op, root_ep_loc, comm); |
---|
| 318 | else MPI_Reduce_local(sendbuf, local_recvbuf, count, datatype, op, 0, comm); |
---|
| 319 | |
---|
| 320 | |
---|
| 321 | |
---|
| 322 | if(is_master) |
---|
| 323 | { |
---|
| 324 | ::MPI_Reduce(local_recvbuf, recvbuf, count, to_mpi_type(datatype), to_mpi_op(op), root_mpi_rank, to_mpi_comm(comm->mpi_comm)); |
---|
| 325 | |
---|
| 326 | } |
---|
| 327 | |
---|
| 328 | if(is_master) |
---|
| 329 | { |
---|
| 330 | delete[] local_recvbuf; |
---|
| 331 | } |
---|
| 332 | |
---|
| 333 | MPI_Barrier_local(comm); |
---|
| 334 | } |
---|
| 335 | |
---|
| 336 | |
---|
| 337 | int MPI_Reduce_intercomm(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) |
---|
| 338 | { |
---|
| 339 | printf("MPI_Reduce_intercomm not yet implemented\n"); |
---|
| 340 | MPI_Abort(comm, 0); |
---|
| 341 | } |
---|
| 342 | } |
---|
| 343 | |
---|
[1745] | 344 | #endif |
---|
| 345 | |
---|
| 346 | |
---|