src/BenchKernelD3Q19ListAa.c

   1 // --------------------------------------------------------------------------
   2 //
   3 // Copyright
   4 //   Markus Wittmann, 2016-2017
   5 //   RRZE, University of Erlangen-Nuremberg, Germany
   6 //   markus.wittmann -at- fau.de or hpc -at- rrze.fau.de
   7 //
   8 //   Viktor Haag, 2016
   9 //   LSS, University of Erlangen-Nuremberg, Germany
  10 //
  11 //  This file is part of the Lattice Boltzmann Benchmark Kernels (LbmBenchKernels).
  12 //
  13 //  LbmBenchKernels is free software: you can redistribute it and/or modify
  14 //  it under the terms of the GNU General Public License as published by
  15 //  the Free Software Foundation, either version 3 of the License, or
  16 //  (at your option) any later version.
  17 //
  18 //  LbmBenchKernels is distributed in the hope that it will be useful,
  19 //  but WITHOUT ANY WARRANTY; without even the implied warranty of
  20 //  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21 //  GNU General Public License for more details.
  22 //
  23 //  You should have received a copy of the GNU General Public License
  24 //  along with LbmBenchKernels.  If not, see <http://www.gnu.org/licenses/>.
  25 //
  26 // --------------------------------------------------------------------------
  27 #include "BenchKernelD3Q19ListAaCommon.h"
  28
  29 #include "Memory.h"
  30 #include "Vtk.h"
  31 #include "LikwidIf.h"
  32
  33 #include <inttypes.h>
  34 #include <math.h>
  35
  36 void FNAME(D3Q19ListAaKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData * cd)
  37 {
  38         Assert(ld != NULL);
  39         Assert(kernelData != NULL);
  40         Assert(cd != NULL);
  41
  42         Assert(cd->Omega > 0.0);
  43         Assert(cd->Omega < 2.0);
  44
  45         KernelData * kd = (KernelData *)kernelData;
  46         KernelDataList * kdl = (KernelDataList *)kernelData;
  47
  48         PdfT omega = cd->Omega;
  49         PdfT omegaEven = omega;
  50         //  1/4: best stability;
  51         // 1/12: removes third-order advection error (best advection);
  52         //  1/6: removes fourth-order diffusion error (best diffusion);
  53         // 3/16: exact location of bounce back for poiseuille flow
  54         PdfT magicParam = 1.0/12.0;
  55         PdfT omegaOdd = 1.0/( 0.5 + magicParam/(1.0/omega - 0.5) );
  56
  57         PdfT evenPart = 0.0;
  58         PdfT oddPart = 0.0;
  59         PdfT dir_indep_trm = 0.0;
  60
  61         const PdfT w_0 = 1.0 /  3.0;
  62         const PdfT w_1 = 1.0 / 18.0;
  63         const PdfT w_2 = 1.0 / 36.0;
  64
  65         const PdfT w_1_x3 = w_1 * 3.0;  const PdfT w_1_nine_half = w_1 * 9.0/2.0;       PdfT w_1_indep = 0.0;
  66         const PdfT w_2_x3 = w_2 * 3.0;  const PdfT w_2_nine_half = w_2 * 9.0/2.0;       PdfT w_2_indep = 0.0;
  67
  68         PdfT ui;
  69
  70         PdfT ux, uy, uz;
  71         PdfT dens;
  72
  73         // Declare pdf_N, pdf_E, pdf_S, pdf_W, ...
  74         #define X(name, idx, idxinv, x, y, z)   PdfT JOIN(pdf_,name);
  75         D3Q19_LIST
  76         #undef X
  77
  78         PdfT * src = kd->Pdfs[0];
  79
  80         int maxIterations = cd->MaxIterations;
  81         int nFluid = kdl->nFluid;
  82         int nCells = kdl->nCells;
  83
  84         uint32_t adjListIndex;
  85         uint32_t * adjList = kdl->AdjList;
  86
  87
  88
  89         #ifdef VTK_OUTPUT
  90                 if (cd->VtkOutput) {
  91                         kd->PdfsActive = src;
  92                         VtkWrite(ld, kd, cd, -1);
  93                 }
  94         #endif
  95
  96         #ifdef STATISTICS
  97                 kd->PdfsActive = src;
  98                 KernelStatistics(kd, ld, cd, 0);
  99         #endif
 100
 101         // TODO: outer openmp parallel
 102         for(int iter = 0; iter < maxIterations; iter += 2) {
 103
 104
 105                 // --------------------------------------------------------------------
 106                 // even time step
 107
 108                 X_LIKWID_START("list-aa-even");
 109
 110         #ifdef _OPENMP
 111                 #pragma omp parallel for default(none) \
 112                                 shared(nFluid, nCells, kd, kdl, adjList, omegaOdd, omegaEven, src) \
 113                                 private(ux, uy, uz, dens, adjListIndex, evenPart, oddPart, dir_indep_trm, w_1_indep, w_2_indep, ui,\
 114                                         pdf_C, \
 115                                         pdf_N, pdf_E, pdf_S, pdf_W, \
 116                                         pdf_NE, pdf_SE, pdf_SW, pdf_NW, \
 117                                         pdf_T, pdf_TN, pdf_TE, pdf_TS, pdf_TW, \
 118                                         pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW)
 119         #endif
 120                 for (int index = 0; index < nFluid; ++index) {
 121
 122
 123                         #define I(index, dir)   P_INDEX_3((nCells), (index), (dir))
 124
 125                         // Load PDFs of local cell: pdf_N = src[I(x, y, z, D3Q19_N)]; ...
 126                         #define X(name, idx, idxinv, _x, _y, _z)        JOIN(pdf_,name) = src[I(index, idx)];
 127                         D3Q19_LIST
 128                         #undef X
 129
 130 // #define LID_DRIVEN_CAVITY
 131
 132 #ifdef LID_DRIVEN_CAVITY
 133                         int nX = kd->Dims[0];
 134                         int nY = kd->Dims[1];
 135                         int nZ = kd->Dims[2];
 136
 137                         int x = kdl->Coords[C_INDEX_X(index)];
 138                         int y = kdl->Coords[C_INDEX_Y(index)];
 139                         int z = kdl->Coords[C_INDEX_Z(index)];
 140
 141                         if (z == nZ - 4 && x > 3 && x < (nX - 4) && y > 3 && y < (nY - 4)) {
 142                                 ux = 0.1 * 0.577;
 143                                 uy = 0.0;
 144                                 uz = 0.0;
 145                         } else {
 146 #endif
 147                                 ux = pdf_E + pdf_NE + pdf_SE + pdf_TE + pdf_BE -
 148                                          pdf_W - pdf_NW - pdf_SW - pdf_TW - pdf_BW;
 149                                 uy = pdf_N + pdf_NE + pdf_NW + pdf_TN + pdf_BN -
 150                                          pdf_S - pdf_SE - pdf_SW - pdf_TS - pdf_BS;
 151                                 uz = pdf_T + pdf_TE + pdf_TW + pdf_TN + pdf_TS -
 152                                          pdf_B - pdf_BE - pdf_BW - pdf_BN - pdf_BS;
 153 #ifdef LID_DRIVEN_CAVITY
 154                         }
 155 #endif
 156
 157                         dens = pdf_C +
 158                                    pdf_N  + pdf_E  + pdf_S  + pdf_W  +
 159                                    pdf_NE + pdf_SE + pdf_SW + pdf_NW +
 160                                    pdf_T  + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
 161                                    pdf_B  + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
 162
 163                         dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz)*3.0/2.0;
 164
 165                         // direction: w_0
 166                         src[I(index, D3Q19_C)             ]  = pdf_C - omegaEven*(pdf_C - w_0*dir_indep_trm);
 167
 168                         // direction: w_1
 169                         w_1_indep = w_1*dir_indep_trm;
 170
 171                         ui = uy;
 172                         evenPart = omegaEven*( 0.5*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
 173                         oddPart = omegaOdd*(0.5*(pdf_N - pdf_S) - ui*w_1_x3 );
 174                         src[I(index, D3Q19_S)]  = pdf_N - evenPart - oddPart;
 175                         src[I(index, D3Q19_N)]  = pdf_S - evenPart + oddPart;
 176
 177                         ui = ux;
 178                         evenPart = omegaEven*( 0.5*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
 179                         oddPart = omegaOdd*(0.5*(pdf_E - pdf_W) - ui*w_1_x3 );
 180                         src[I(index, D3Q19_W)]  = pdf_E - evenPart - oddPart;
 181                         src[I(index, D3Q19_E)]  = pdf_W - evenPart + oddPart;
 182
 183                         ui = uz;
 184                         evenPart = omegaEven*( 0.5*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
 185                         oddPart = omegaOdd*(0.5*(pdf_T - pdf_B) - ui*w_1_x3 );
 186                         src[I(index, D3Q19_B)]  = pdf_T - evenPart - oddPart;
 187                         src[I(index, D3Q19_T)]  = pdf_B - evenPart + oddPart;
 188
 189                         // direction: w_2
 190                         w_2_indep = w_2*dir_indep_trm;
 191
 192                         ui = -ux + uy;
 193                         evenPart = omegaEven*( 0.5*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
 194                         oddPart = omegaOdd*(0.5*(pdf_NW - pdf_SE) - ui*w_2_x3 );
 195                         src[I(index, D3Q19_SE)] = pdf_NW - evenPart - oddPart;
 196                         src[I(index, D3Q19_NW)] = pdf_SE - evenPart + oddPart;
 197
 198                         ui = ux + uy;
 199                         evenPart = omegaEven*( 0.5*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
 200                         oddPart = omegaOdd*(0.5*(pdf_NE - pdf_SW) - ui*w_2_x3 );
 201                         src[I(index, D3Q19_SW)] = pdf_NE - evenPart - oddPart;
 202                         src[I(index, D3Q19_NE)] = pdf_SW - evenPart + oddPart;
 203
 204                         ui = -ux + uz;
 205                         evenPart = omegaEven*( 0.5*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
 206                         oddPart = omegaOdd*(0.5*(pdf_TW - pdf_BE) - ui*w_2_x3 );
 207                         src[I(index, D3Q19_BE)] = pdf_TW - evenPart - oddPart;
 208                         src[I(index, D3Q19_TW)] = pdf_BE - evenPart + oddPart;
 209
 210                         ui = ux + uz;
 211                         evenPart = omegaEven*( 0.5*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
 212                         oddPart = omegaOdd*(0.5*(pdf_TE - pdf_BW) - ui*w_2_x3 );
 213                         src[I(index, D3Q19_BW)] = pdf_TE - evenPart - oddPart;
 214                         src[I(index, D3Q19_TE)] = pdf_BW - evenPart + oddPart;
 215
 216                         ui = -uy + uz;
 217                         evenPart = omegaEven*( 0.5*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
 218                         oddPart = omegaOdd*(0.5*(pdf_TS - pdf_BN) - ui*w_2_x3 );
 219                         src[I(index, D3Q19_BN)] = pdf_TS - evenPart - oddPart;
 220                         src[I(index, D3Q19_TS)] = pdf_BN - evenPart + oddPart;
 221
 222                         ui = uy + uz;
 223                         evenPart = omegaEven*( 0.5*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
 224                         oddPart = omegaOdd*(0.5*(pdf_TN - pdf_BS) - ui*w_2_x3 );
 225                         src[I(index, D3Q19_BS)] = pdf_TN - evenPart - oddPart;
 226                         src[I(index, D3Q19_TN)] = pdf_BS - evenPart + oddPart;
 227
 228                 } // loop over fluid nodes
 229
 230                 X_LIKWID_STOP("list-aa-even");
 231
 232                 // save current iteration
 233                 kdl->Iteration = iter;
 234                 #ifdef VERIFICATION
 235                         kd->PdfsActive = src;
 236                         KernelAddBodyForce(kd, ld, cd);
 237                 #endif
 238
 239                 // --------------------------------------------------------------------
 240                 // odd time step
 241
 242                 X_LIKWID_START("list-aa-odd");
 243
 244 #ifdef _OPENMP
 245                 #pragma omp parallel for default(none) \
 246                                 shared(nFluid, nCells, kd, kdl, adjList, omegaOdd, omegaEven, src) \
 247                                 private(ux, uy, uz, dens, adjListIndex, evenPart, oddPart, dir_indep_trm, w_1_indep, w_2_indep, ui,\
 248                                         pdf_C, \
 249                                         pdf_N, pdf_E, pdf_S, pdf_W, \
 250                                         pdf_NE, pdf_SE, pdf_SW, pdf_NW, \
 251                                         pdf_T, pdf_TN, pdf_TE, pdf_TS, pdf_TW, \
 252                                         pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW)
 253 #endif
 254                 for (int index = 0; index < nFluid; ++index) {
 255
 256
 257                         adjListIndex = index * N_D3Q19_IDX;
 258
 259                         // Load PDFs of local cell: pdf_N = src[adjList[adjListIndex + D3Q19_S]]; ...
 260                         pdf_C = src[P_INDEX_3(nCells, index, D3Q19_C)];
 261
 262                         #define X(name, idx, idxinv, _x, _y, _z)        JOIN(pdf_,name) = src[adjList[adjListIndex + idxinv]];
 263                         D3Q19_LIST_WO_C
 264                         #undef X
 265
 266 #ifdef LID_DRIVEN_CAVITY
 267                         int nX = kd->Dims[0];
 268                         int nY = kd->Dims[1];
 269                         int nZ = kd->Dims[2];
 270
 271                         int x = kdl->Coords[C_INDEX_X(index)];
 272                         int y = kdl->Coords[C_INDEX_Y(index)];
 273                         int z = kdl->Coords[C_INDEX_Z(index)];
 274
 275                         if (z == nZ - 4 && x > 3 && x < (nX - 4) && y > 3 && y < (nY - 4)) {
 276                                 ux = 0.1 * 0.577;
 277                                 uy = 0.0;
 278                                 uz = 0.0;
 279                         } else {
 280 #endif
 281                                 ux = pdf_E + pdf_NE + pdf_SE + pdf_TE + pdf_BE -
 282                                          pdf_W - pdf_NW - pdf_SW - pdf_TW - pdf_BW;
 283                                 uy = pdf_N + pdf_NE + pdf_NW + pdf_TN + pdf_BN -
 284                                          pdf_S - pdf_SE - pdf_SW - pdf_TS - pdf_BS;
 285                                 uz = pdf_T + pdf_TE + pdf_TW + pdf_TN + pdf_TS -
 286                                          pdf_B - pdf_BE - pdf_BW - pdf_BN - pdf_BS;
 287 #ifdef LID_DRIVEN_CAVITY
 288                         }
 289 #endif
 290
 291                         dens = pdf_C +
 292                                    pdf_N  + pdf_E  + pdf_S  + pdf_W  +
 293                                    pdf_NE + pdf_SE + pdf_SW + pdf_NW +
 294                                    pdf_T  + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
 295                                    pdf_B  + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
 296
 297                         dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz)*3.0/2.0;
 298
 299                         adjListIndex = index * N_D3Q19_IDX;
 300
 301                         // direction: w_0
 302                         src[I(index, D3Q19_C)             ]  = pdf_C - omegaEven*(pdf_C - w_0*dir_indep_trm);
 303
 304                         // direction: w_1
 305                         w_1_indep = w_1*dir_indep_trm;
 306
 307                         ui = uy;
 308                         evenPart = omegaEven*( 0.5*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
 309                         oddPart = omegaOdd*(0.5*(pdf_N - pdf_S) - ui*w_1_x3 );
 310                         src[adjList[adjListIndex + D3Q19_N]]  = pdf_N - evenPart - oddPart;
 311                         src[adjList[adjListIndex + D3Q19_S]]  = pdf_S - evenPart + oddPart;
 312
 313                         ui = ux;
 314                         evenPart = omegaEven*( 0.5*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
 315                         oddPart = omegaOdd*(0.5*(pdf_E - pdf_W) - ui*w_1_x3 );
 316                         src[adjList[adjListIndex + D3Q19_E]]  = pdf_E - evenPart - oddPart;
 317                         src[adjList[adjListIndex + D3Q19_W]]  = pdf_W - evenPart + oddPart;
 318
 319                         ui = uz;
 320                         evenPart = omegaEven*( 0.5*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
 321                         oddPart = omegaOdd*(0.5*(pdf_T - pdf_B) - ui*w_1_x3 );
 322                         src[adjList[adjListIndex + D3Q19_T]]  = pdf_T - evenPart - oddPart;
 323                         src[adjList[adjListIndex + D3Q19_B]]  = pdf_B - evenPart + oddPart;
 324
 325                         // direction: w_2
 326                         w_2_indep = w_2*dir_indep_trm;
 327
 328                         ui = -ux + uy;
 329                         evenPart = omegaEven*( 0.5*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
 330                         oddPart = omegaOdd*(0.5*(pdf_NW - pdf_SE) - ui*w_2_x3 );
 331                         src[adjList[adjListIndex + D3Q19_NW]] = pdf_NW - evenPart - oddPart;
 332                         src[adjList[adjListIndex + D3Q19_SE]] = pdf_SE - evenPart + oddPart;
 333
 334                         ui = ux + uy;
 335                         evenPart = omegaEven*( 0.5*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
 336                         oddPart = omegaOdd*(0.5*(pdf_NE - pdf_SW) - ui*w_2_x3 );
 337                         src[adjList[adjListIndex + D3Q19_NE]] = pdf_NE - evenPart - oddPart;
 338                         src[adjList[adjListIndex + D3Q19_SW]] = pdf_SW - evenPart + oddPart;
 339
 340                         ui = -ux + uz;
 341                         evenPart = omegaEven*( 0.5*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
 342                         oddPart = omegaOdd*(0.5*(pdf_TW - pdf_BE) - ui*w_2_x3 );
 343                         src[adjList[adjListIndex + D3Q19_TW]] = pdf_TW - evenPart - oddPart;
 344                         src[adjList[adjListIndex + D3Q19_BE]] = pdf_BE - evenPart + oddPart;
 345
 346                         ui = ux + uz;
 347                         evenPart = omegaEven*( 0.5*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
 348                         oddPart = omegaOdd*(0.5*(pdf_TE - pdf_BW) - ui*w_2_x3 );
 349                         src[adjList[adjListIndex + D3Q19_TE]] = pdf_TE - evenPart - oddPart;
 350                         src[adjList[adjListIndex + D3Q19_BW]] = pdf_BW - evenPart + oddPart;
 351
 352                         ui = -uy + uz;
 353                         evenPart = omegaEven*( 0.5*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
 354                         oddPart = omegaOdd*(0.5*(pdf_TS - pdf_BN) - ui*w_2_x3 );
 355                         src[adjList[adjListIndex + D3Q19_TS]] = pdf_TS - evenPart - oddPart;
 356                         src[adjList[adjListIndex + D3Q19_BN]] = pdf_BN - evenPart + oddPart;
 357
 358                         ui = uy + uz;
 359                         evenPart = omegaEven*( 0.5*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
 360                         oddPart = omegaOdd*(0.5*(pdf_TN - pdf_BS) - ui*w_2_x3 );
 361                         src[adjList[adjListIndex + D3Q19_TN]] = pdf_TN - evenPart - oddPart;
 362                         src[adjList[adjListIndex + D3Q19_BS]] = pdf_BS - evenPart + oddPart;
 363
 364                         #undef I
 365                 } // loop over fluid nodes
 366
 367                 X_LIKWID_STOP("list-aa-odd");
 368
 369                 // save current iteration
 370                 kdl->Iteration = iter + 1;
 371
 372                 #ifdef VERIFICATION
 373                         kd->PdfsActive = src;
 374                         KernelAddBodyForce(kd, ld, cd);
 375                 #endif
 376
 377                 #ifdef VTK_OUTPUT
 378                         if (cd->VtkOutput && (iter % cd->VtkModulus) == 0) {
 379                                 kd->PdfsActive = src;
 380                                 VtkWrite(ld, kd, cd, iter);
 381                         }
 382                 #endif
 383
 384                 #ifdef STATISTICS
 385                         kd->PdfsActive = src;
 386                         KernelStatistics(kd, ld, cd, iter);
 387                 #endif
 388
 389
 390         } // for (int iter = 0; ...
 391
 392
 393 #ifdef VTK_OUTPUT
 394         if (cd->VtkOutput) {
 395                 kd->PdfsActive = src;
 396                 VtkWrite(ld, kd, cd, maxIterations);
 397         }
 398 #endif
 399
 400 #ifdef STATISTICS
 401         kd->PdfsActive = src;
 402         KernelStatistics(kd, ld, cd, maxIterations);
 403 #endif
 404
 405         return;
 406 }