| 1 | /* |
| 2 | * SpaceDistribution.hpp |
| 3 | * |
| 4 | * Created on: Dec 3, 2016 |
| 5 | * Author: i-bird |
| 6 | */ |
| 7 | |
| 8 | #ifndef SRC_DECOMPOSITION_DISTRIBUTION_SPACEDISTRIBUTION_HPP_ |
| 9 | #define SRC_DECOMPOSITION_DISTRIBUTION_SPACEDISTRIBUTION_HPP_ |
| 10 | |
| 11 | #include "util/mathutil.hpp" |
| 12 | #include "NN/CellList/CellDecomposer.hpp" |
| 13 | #include "Grid/grid_key_dx_iterator_hilbert.hpp" |
| 14 | |
| 15 | /*! \brief Class that distribute sub-sub-domains across processors using an hilbert curve |
| 16 | * to divide the space |
| 17 | * |
| 18 | * ### Initialize a Cartesian graph and decompose |
| 19 | * \snippet Distribution_unit_tests.hpp Initialize a Space Cartesian graph and decompose |
| 20 | * |
| 21 | * |
| 22 | */ |
| 23 | template<unsigned int dim, typename T> |
| 24 | class SpaceDistribution |
| 25 | { |
| 26 | //! Vcluster |
| 27 | Vcluster<> & v_cl; |
| 28 | |
| 29 | //! Structure that store the cartesian grid information |
| 30 | grid_sm<dim, void> gr; |
| 31 | |
| 32 | //! rectangular domain to decompose |
| 33 | Box<dim, T> domain; |
| 34 | |
| 35 | //! Global sub-sub-domain graph |
| 36 | Graph_CSR<nm_v<dim>, nm_e> gp; |
| 37 | |
| 38 | |
| 39 | public: |
| 40 | |
| 41 | /*! Constructor |
| 42 | * |
| 43 | * \param v_cl Vcluster to use as communication object in this class |
| 44 | */ |
| 45 | SpaceDistribution(Vcluster<> & v_cl) |
| 46 | :v_cl(v_cl) |
| 47 | { |
| 48 | } |
| 49 | |
| 50 | /*! Copy constructor |
| 51 | * |
| 52 | * \param pm Distribution to copy |
| 53 | * |
| 54 | */ |
| 55 | SpaceDistribution(const ParMetisDistribution<dim,T> & pm) |
| 56 | :v_cl(pm.v_cl) |
| 57 | { |
| 58 | this->operator=(pm); |
| 59 | } |
| 60 | |
| 61 | /*! Copy constructor |
| 62 | * |
| 63 | * \param pm Distribution to copy |
| 64 | * |
| 65 | */ |
| 66 | SpaceDistribution(SpaceDistribution<dim,T> && pm) |
| 67 | :v_cl(pm.v_cl) |
| 68 | { |
| 69 | this->operator=(pm); |
| 70 | } |
| 71 | |
| 72 | /*! \brief Create the Cartesian graph |
| 73 | * |
| 74 | * \param grid info |
| 75 | * \param dom domain |
| 76 | */ |
| 77 | void createCartGraph(grid_sm<dim, void> & grid, Box<dim, T> dom) |
| 78 | { |
| 79 | size_t bc[dim]; |
| 80 | |
| 81 | for (size_t i = 0 ; i < dim ; i++) |
| 82 | bc[i] = NON_PERIODIC; |
| 83 | |
| 84 | // Set grid and domain |
| 85 | gr = grid; |
| 86 | domain = dom; |
| 87 | |
| 88 | // Create a cartesian grid graph |
| 89 | CartesianGraphFactory<dim, Graph_CSR<nm_v<dim>, nm_e>> g_factory_part; |
| 90 | gp = g_factory_part.template construct<NO_EDGE, nm_v_id, T, dim - 1, 0>(gr.getSize(), domain, bc); |
| 91 | |
| 92 | // Init to 0.0 axis z (to fix in graphFactory) |
| 93 | if (dim < 3) |
| 94 | { |
| 95 | for (size_t i = 0; i < gp.getNVertex(); i++) |
| 96 | gp.vertex(i).template get<nm_v_x>()[2] = 0.0; |
| 97 | } |
| 98 | for (size_t i = 0; i < gp.getNVertex(); i++) |
| 99 | gp.vertex(i).template get<nm_v_global_id>() = i; |
| 100 | |
| 101 | } |
| 102 | |
| 103 | /*! \brief Get the current graph (main) |
| 104 | * |
| 105 | */ |
| 106 | Graph_CSR<nm_v<dim>, nm_e> & getGraph() |
| 107 | { |
| 108 | return gp; |
| 109 | } |
| 110 | |
| 111 | /*! \brief Create the decomposition |
| 112 | * |
| 113 | */ |
| 114 | void decompose() |
| 115 | { |
| 116 | // Get the number of processing units |
| 117 | size_t Np = v_cl.getProcessingUnits(); |
| 118 | |
| 119 | // Calculate the best number of sub-domains for each |
| 120 | // processor |
| 121 | size_t N_tot = gr.size(); |
| 122 | size_t N_best_each = N_tot / Np; |
| 123 | size_t N_rest = N_tot % Np; |
| 124 | |
| 125 | openfpm::vector<size_t> accu(Np); |
| 126 | accu.get(0) = N_best_each + ((0 < N_rest)?1:0); |
| 127 | for (size_t i = 1 ; i < Np ; i++) |
| 128 | accu.get(i) = accu.get(i-1) + N_best_each + ((i < N_rest)?1:0); |
| 129 | |
| 130 | // Get the maximum along dimensions and take the smallest n number |
| 131 | // such that 2^n < m. n it will be order of the hilbert curve |
| 132 | |
| 133 | size_t max = 0; |
| 134 | |
| 135 | for (size_t i = 0; i < dim ; i++) |
| 136 | { |
| 137 | if (max < gr.size(i)) |
| 138 | max = gr.size(i); |
| 139 | } |
| 140 | |
| 141 | // Get the order of the hilbert-curve |
| 142 | size_t order = openfpm::math::log2_64(max); |
| 143 | if (1ul << order < max) |
| 144 | order += 1; |
| 145 | |
| 146 | size_t n = 1 << order; |
| 147 | |
| 148 | // Create the CellDecomoser |
| 149 | |
| 150 | CellDecomposer_sm<dim,T> cd_sm; |
| 151 | cd_sm.setDimensions(domain, gr.getSize(), 0); |
| 152 | |
| 153 | // create the hilbert curve |
| 154 | |
| 155 | //hilbert curve iterator |
| 156 | grid_key_dx_iterator_hilbert<dim> h_it(order); |
| 157 | |
| 158 | T spacing[dim]; |
| 159 | |
| 160 | // Calculate the hilbert curve spacing |
| 161 | for (size_t i = 0 ; i < dim ; i++) |
| 162 | spacing[i] = (domain.getHigh(i) - domain.getLow(i)) / n; |
| 163 | |
| 164 | // Small grid to detect already assigned sub-sub-domains |
| 165 | grid_cpu<dim,aggregate<long int>> g(gr.getSize()); |
| 166 | g.setMemory(); |
| 167 | |
| 168 | // Reset the grid to -1 |
| 169 | grid_key_dx_iterator<dim> it(gr); |
| 170 | while (it.isNext()) |
| 171 | { |
| 172 | auto key = it.get(); |
| 173 | |
| 174 | g.template get<0>(key) = -1; |
| 175 | |
| 176 | ++it; |
| 177 | } |
| 178 | |
| 179 | // Go along the hilbert-curve and divide the space |
| 180 | |
| 181 | size_t proc_d = 0; |
| 182 | size_t ele_d = 0; |
| 183 | while (h_it.isNext()) |
| 184 | { |
| 185 | auto key = h_it.get(); |
| 186 | |
| 187 | // Point p |
| 188 | Point<dim,T> p; |
| 189 | |
| 190 | for (size_t i = 0 ; i < dim ; i++) |
| 191 | p.get(i) = key.get(i) * spacing[i] + spacing[i] / 2; |
| 192 | |
| 193 | grid_key_dx<dim> sp = cd_sm.getCellGrid(p); |
| 194 | |
| 195 | if (g.template get<0>(sp) == -1) |
| 196 | { |
| 197 | g.template get<0>(sp) = proc_d; |
| 198 | ele_d++; |
| 199 | |
| 200 | if (ele_d >= accu.get(proc_d)) |
| 201 | proc_d++; |
| 202 | } |
| 203 | |
| 204 | ++h_it; |
| 205 | } |
| 206 | |
| 207 | // Fill from the grid to the graph |
| 208 | |
| 209 | // Reset the grid to -1 |
| 210 | grid_key_dx_iterator<dim> it2(gr); |
| 211 | while (it2.isNext()) |
| 212 | { |
| 213 | auto key = it2.get(); |
| 214 | |
| 215 | gp.template vertex_p<nm_v_proc_id>(gr.LinId(key)) = g.template get<0>(key); |
| 216 | |
| 217 | ++it2; |
| 218 | } |
| 219 | |
| 220 | return; |
| 221 | } |
| 222 | |
| 223 | /*! \brief Refine current decomposition |
| 224 | * |
| 225 | * Has no effect in this case |
| 226 | * |
| 227 | */ |
| 228 | void refine() |
| 229 | { |
| 230 | std::cout << __FILE__ << ":" << __LINE__ << " You are trying to dynamicaly balance a fixed decomposition, this operation has no effect" << std::endl; |
| 231 | } |
| 232 | |
| 233 | /*! \brief Compute the unbalance of the processor compared to the optimal balance |
| 234 | * |
| 235 | * \return the unbalance from the optimal one 0.01 mean 1% |
| 236 | */ |
| 237 | float getUnbalance() |
| 238 | { |
| 239 | return gr.size() % v_cl.getProcessingUnits(); |
| 240 | } |
| 241 | |
| 242 | /*! \brief function that return the position of the vertex in the space |
| 243 | * |
| 244 | * \param id vertex id |
| 245 | * \param pos vector that will contain x, y, z |
| 246 | * |
| 247 | */ |
| 248 | void getSubSubDomainPosition(size_t id, T (&pos)[dim]) |
| 249 | { |
| 250 | #ifdef SE_CLASS1 |
| 251 | if (id >= gp.getNVertex()) |
| 252 | std::cerr << __FILE__ << ":" << __LINE__ << "Such vertex doesn't exist (id = " << id << ", " << "total size = " << gp.getNVertex() << ")\n" ; |
| 253 | #endif |
| 254 | |
| 255 | // Copy the geometrical informations inside the pos vector |
| 256 | pos[0] = gp.vertex(id).template get<nm_v_x>()[0]; |
| 257 | pos[1] = gp.vertex(id).template get<nm_v_x>()[1]; |
| 258 | if (dim == 3) |
| 259 | pos[2] = gp.vertex(id).template get<nm_v_x>()[2]; |
| 260 | } |
| 261 | |
| 262 | /*! \brief Function that set the weight of the vertex |
| 263 | * |
| 264 | * \param id vertex id |
| 265 | * \param weight to give to the vertex |
| 266 | * |
| 267 | */ |
| 268 | inline void setComputationCost(size_t id, size_t weight) |
| 269 | { |
| 270 | std::cout << __FILE__ << ":" << __LINE__ << " You are trying to set the computation cost on a fixed decomposition, this operation has no effect" << std::endl; |
| 271 | } |
| 272 | |
| 273 | /*! \brief Checks if weights are used on the vertices |
| 274 | * |
| 275 | * \return true if weights are used in the decomposition |
| 276 | */ |
| 277 | bool weightsAreUsed() |
| 278 | { |
| 279 | return false; |
| 280 | } |
| 281 | |
| 282 | /*! \brief function that get the weight of the vertex |
| 283 | * |
| 284 | * \param id vertex id |
| 285 | * |
| 286 | * \return the weight of the vertex |
| 287 | * |
| 288 | */ |
| 289 | size_t getSubSubDomainComputationCost(size_t id) |
| 290 | { |
| 291 | return 1.0; |
| 292 | } |
| 293 | |
| 294 | /*! \brief Compute the processor load counting the total weights of its vertices |
| 295 | * |
| 296 | * \return the computational load of the processor graph |
| 297 | */ |
| 298 | size_t getProcessorLoad() |
| 299 | { |
| 300 | // Get the number of processing units |
| 301 | size_t Np = v_cl.getProcessingUnits(); |
| 302 | |
| 303 | // Calculate the best number of sub-domains for each |
| 304 | // processor |
| 305 | size_t N_tot = gr.size(); |
| 306 | size_t N_best_each = N_tot / Np; |
| 307 | size_t N_rest = N_tot % Np; |
| 308 | |
| 309 | if (v_cl.getProcessUnitID() < N_rest) |
| 310 | N_best_each += 1; |
| 311 | |
| 312 | return N_best_each; |
| 313 | } |
| 314 | |
| 315 | /*! \brief Set migration cost of the vertex id |
| 316 | * |
| 317 | * \param id of the vertex to update |
| 318 | * \param migration cost of the migration |
| 319 | */ |
| 320 | void setMigrationCost(size_t id, size_t migration) |
| 321 | { |
| 322 | } |
| 323 | |
| 324 | /*! \brief Set communication cost of the edge id |
| 325 | * |
| 326 | * \param v_id Id of the source vertex of the edge |
| 327 | * \param e i child of the vertex |
| 328 | * \param communication Communication value |
| 329 | */ |
| 330 | void setCommunicationCost(size_t v_id, size_t e, size_t communication) |
| 331 | { |
| 332 | } |
| 333 | |
| 334 | /*! \brief Returns total number of sub-sub-domains in the distribution graph |
| 335 | * |
| 336 | * \return number of sub-sub-domain |
| 337 | * |
| 338 | */ |
| 339 | size_t getNSubSubDomains() |
| 340 | { |
| 341 | return gp.getNVertex(); |
| 342 | } |
| 343 | |
| 344 | /*! \brief Returns total number of neighbors of the sub-sub-domain id |
| 345 | * |
| 346 | * \param id id of the sub-sub-domain |
| 347 | */ |
| 348 | size_t getNSubSubDomainNeighbors(size_t id) |
| 349 | { |
| 350 | return gp.getNChilds(id); |
| 351 | } |
| 352 | |
| 353 | /*! \brief Print the current distribution and save it to VTK file |
| 354 | * |
| 355 | * \param file filename |
| 356 | * |
| 357 | */ |
| 358 | void write(const std::string & file) |
| 359 | { |
| 360 | VTKWriter<Graph_CSR<nm_v<dim>, nm_e>, VTK_GRAPH> gv2(gp); |
| 361 | gv2.write(std::to_string(v_cl.getProcessUnitID()) + "_" + file + ".vtk" ); |
| 362 | } |
| 363 | |
| 364 | const SpaceDistribution<dim,T> & operator=(const SpaceDistribution<dim,T> & dist) |
| 365 | { |
| 366 | gr = dist.gr; |
| 367 | domain = dist.domain; |
| 368 | gp = dist.gp; |
| 369 | |
| 370 | return *this; |
| 371 | } |
| 372 | |
| 373 | const SpaceDistribution<dim,T> & operator=(SpaceDistribution<dim,T> && dist) |
| 374 | { |
| 375 | v_cl = dist.v_cl; |
| 376 | gr = dist.gr; |
| 377 | domain = dist.domain; |
| 378 | gp.swap(dist.gp); |
| 379 | |
| 380 | return *this; |
| 381 | } |
| 382 | |
| 383 | /*! \brief It return the decomposition id |
| 384 | * |
| 385 | * It just return 0 |
| 386 | * |
| 387 | * \return 0 |
| 388 | * |
| 389 | */ |
| 390 | size_t get_ndec() |
| 391 | { |
| 392 | return 0; |
| 393 | } |
| 394 | }; |
| 395 | |
| 396 | |
| 397 | #endif /* SRC_DECOMPOSITION_DISTRIBUTION_SPACEDISTRIBUTION_HPP_ */ |
| 398 | |