| 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 | // Michael Hussnaetter, 2017-2018 |
| 12 | // University of Erlangen-Nuremberg, Germany |
| 13 | // michael.hussnaetter -at- fau.de |
| 14 | // |
| 15 | // This file is part of the Lattice Boltzmann Benchmark Kernels (LbmBenchKernels). |
| 16 | // |
| 17 | // LbmBenchKernels is free software: you can redistribute it and/or modify |
| 18 | // it under the terms of the GNU General Public License as published by |
| 19 | // the Free Software Foundation, either version 3 of the License, or |
| 20 | // (at your option) any later version. |
| 21 | // |
| 22 | // LbmBenchKernels is distributed in the hope that it will be useful, |
| 23 | // but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 24 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 25 | // GNU General Public License for more details. |
| 26 | // |
| 27 | // You should have received a copy of the GNU General Public License |
| 28 | // along with LbmBenchKernels. If not, see <http://www.gnu.org/licenses/>. |
| 29 | // |
| 30 | // -------------------------------------------------------------------------- |
| 31 | #include "BenchKernelD3Q19ListAaPvGatherCommon.h" |
| 32 | |
| 33 | #include "Memory.h" |
| 34 | #include "Vector.h" |
| 35 | #include "Vtk.h" |
| 36 | |
| 37 | #include <math.h> |
| 38 | |
| 39 | #ifdef _OPENMP |
| 40 | #include <omp.h> |
| 41 | #endif |
| 42 | |
| 43 | #define PAGE_4K 4096 |
| 44 | |
| 45 | #if ALLOC_ADJ_LIST_IN_HBM == 1 |
| 46 | #define ADJ_LIST_ALLOCATOR HbwAllocAligned |
| 47 | #define ADJ_LIST_FREE HbwFree |
| 48 | #else |
| 49 | #define ADJ_LIST_ALLOCATOR MemAllocAligned |
| 50 | #define ADJ_LIST_FREE MemFree |
| 51 | #endif |
| 52 | |
| 53 | #if ALLOC_PDF_IN_HBM == 1 |
| 54 | #define PDF_ALLOCATOR HbwAllocAligned |
| 55 | #define PDF_FREE HbwFree |
| 56 | #else |
| 57 | #define PDF_ALLOCATOR MemAllocAligned |
| 58 | #define PDF_FREE MemFree |
| 59 | #endif |
| 60 | |
| 61 | // Forward definition. |
| 62 | void FNAME(D3Q19ListAaPvGatherKernel)(LatticeDesc * ld, struct KernelData_ * kd, CaseData * cd); |
| 63 | |
| 64 | |
| 65 | // ----------------------------------------------------------------------- |
| 66 | // Functions which are used as callback by the kernel to read or write |
| 67 | // PDFs and nodes. |
| 68 | |
| 69 | |
| 70 | static void FNAME(BCGetPdf)(KernelData * kd, int x, int y, int z, int dir, PdfT * pdf) |
| 71 | { |
| 72 | Assert(kd != NULL); |
| 73 | Assert(kd->PdfsActive != NULL); |
| 74 | Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]); |
| 75 | Assert(pdf != NULL); |
| 76 | |
| 77 | Assert(x >= 0); Assert(y >= 0); Assert(z >= 0); |
| 78 | Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]); |
| 79 | Assert(dir >= 0); Assert(dir < N_D3Q19); |
| 80 | |
| 81 | KernelDataList * kdl = KDL(kd); |
| 82 | uint32_t * adjList = kdl->AdjList; |
| 83 | |
| 84 | if (kdl->Iteration % 2 == 0) { |
| 85 | // Pdfs are stored inverse, local PDFs are located in remote nodes |
| 86 | |
| 87 | // getting node index |
| 88 | uint32_t index = kdl->Grid[L_INDEX_4(kd->Dims, x, y, z)]; |
| 89 | |
| 90 | if (dir != D3Q19_C) { |
| 91 | #define ADJ_LIST(dir) adjList[(index - (index % VSIZE)) * N_D3Q19_IDX + (dir * VSIZE) + (index % VSIZE)] |
| 92 | *pdf = kd->PdfsActive[ADJ_LIST(D3Q19_INV[dir])]; |
| 93 | #undef ADJ_LIST |
| 94 | } |
| 95 | else { |
| 96 | *pdf = kd->PdfsActive[P_INDEX_3(kdl->nCells, index, dir)]; |
| 97 | } |
| 98 | |
| 99 | } |
| 100 | else { |
| 101 | *pdf = kd->PdfsActive[P_INDEX_5(kdl, x, y, z, dir)]; |
| 102 | } |
| 103 | |
| 104 | return; |
| 105 | } |
| 106 | |
| 107 | static void FNAME(BCSetPdf)(KernelData * kd, int x, int y, int z, int dir, PdfT pdf) |
| 108 | { |
| 109 | Assert(kd != NULL); |
| 110 | Assert(kd->PdfsActive != NULL); |
| 111 | Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]); |
| 112 | Assert(x >= 0); Assert(y >= 0); Assert(z >= 0); |
| 113 | Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]); |
| 114 | Assert(dir >= 0); Assert(dir < N_D3Q19); |
| 115 | |
| 116 | if (isnan(pdf)) { |
| 117 | printf("ERROR: setting nan %d %d %d %d %s\n", x, y, z, dir, D3Q19_NAMES[dir]); |
| 118 | DEBUG_BREAK_POINT(); |
| 119 | exit(1); |
| 120 | } |
| 121 | |
| 122 | KernelDataList * kdl = KDL(kd); |
| 123 | uint32_t * adjList = kdl->AdjList; |
| 124 | |
| 125 | if (kdl->Iteration % 2 == 0) { |
| 126 | // Pdfs are stored inverse, local PDFs are located in remote nodes |
| 127 | |
| 128 | // getting node index |
| 129 | uint32_t index = kdl->Grid[L_INDEX_4(kd->Dims, x, y, z)]; |
| 130 | |
| 131 | if (dir != D3Q19_C) { |
| 132 | #define ADJ_LIST(dir) adjList[(index - (index % VSIZE)) * N_D3Q19_IDX + (dir * VSIZE) + (index % VSIZE)] |
| 133 | kd->PdfsActive[ADJ_LIST(D3Q19_INV[dir])] = pdf; |
| 134 | #undef ADJ_LIST |
| 135 | } else { |
| 136 | kd->PdfsActive[P_INDEX_3(kdl->nCells, index, dir)] = pdf; |
| 137 | } |
| 138 | |
| 139 | } else { |
| 140 | kd->PdfsActive[P_INDEX_5(kdl, x, y, z, dir)] = pdf; |
| 141 | } |
| 142 | |
| 143 | return; |
| 144 | } |
| 145 | |
| 146 | |
| 147 | static void GetNode(KernelData * kd, int x, int y, int z, PdfT * pdfs) |
| 148 | { |
| 149 | Assert(kd != NULL); |
| 150 | Assert(kd->PdfsActive != NULL); |
| 151 | Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]); |
| 152 | Assert(pdfs != NULL); |
| 153 | Assert(x >= 0); Assert(y >= 0); Assert(z >= 0); |
| 154 | Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]); |
| 155 | |
| 156 | KernelDataList * kdl = KDL(kd); |
| 157 | uint32_t * adjList = kdl->AdjList; |
| 158 | |
| 159 | if(kdl->Iteration % 2 == 0){ |
| 160 | |
| 161 | uint32_t index = kdl->Grid[L_INDEX_4(kdl->kd.Dims, x, y, z)]; |
| 162 | |
| 163 | // Load PDFs of local cell: pdf_N = src[adjList[adjListIndex + D3Q19_S]]; ... |
| 164 | #define ADJ_LIST(dir) adjList[(index - (index % VSIZE)) * N_D3Q19_IDX + (dir * VSIZE) + (index % VSIZE)] |
| 165 | #define X(name, idx, idxinv, _x, _y, _z) pdfs[idx] = kd->PdfsActive[ADJ_LIST(idxinv)]; |
| 166 | D3Q19_LIST_WO_C |
| 167 | #undef X |
| 168 | #undef ADJ_LIST |
| 169 | pdfs[D3Q19_C] = kd->PdfsActive[P_INDEX_3(kdl->nCells, index, D3Q19_C)]; |
| 170 | |
| 171 | } else { |
| 172 | |
| 173 | #define I(x, y, z, dir) P_INDEX_5(KDL(kd), (x), (y), (z), (dir)) |
| 174 | #define X(name, idx, idxinv, _x, _y, _z) pdfs[idx] = kd->PdfsActive[I(x, y, z, idx)]; |
| 175 | D3Q19_LIST |
| 176 | #undef X |
| 177 | #undef I |
| 178 | |
| 179 | } |
| 180 | |
| 181 | for (int d = 0; d < 19; ++d) { |
| 182 | if(isnan(pdfs[d]) || isinf(pdfs[d])) { |
| 183 | printf("%d %d %d %d nan! get node\n", x, y, z, d); |
| 184 | for (int d2 = 0; d2 < 19; ++d2) { |
| 185 | printf("%d: %e\n", d2, pdfs[d2]); |
| 186 | } |
| 187 | exit(1); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | return; |
| 192 | } |
| 193 | |
| 194 | |
| 195 | static void SetNode(KernelData * kd, int x, int y, int z, PdfT * pdfs) |
| 196 | { |
| 197 | Assert(kd != NULL); |
| 198 | Assert(kd->PdfsActive != NULL); |
| 199 | Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]); |
| 200 | Assert(pdfs != NULL); |
| 201 | |
| 202 | Assert(x >= 0); Assert(y >= 0); Assert(z >= 0); |
| 203 | Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]); |
| 204 | |
| 205 | for (int d = 0; d < 19; ++d) { |
| 206 | if(isnan(pdfs[d])) { |
| 207 | printf("%d %d %d %d nan! get node\n", x, y, z, d); |
| 208 | for (int d2 = 0; d2 < 19; ++d2) { |
| 209 | printf("%d: %e\n", d2, pdfs[d2]); |
| 210 | } |
| 211 | exit(1); |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | KernelDataList * kdl = KDL(kd); |
| 216 | uint32_t * adjList = kdl->AdjList; |
| 217 | |
| 218 | if(kdl->Iteration % 2 == 0){ |
| 219 | |
| 220 | uint32_t index = kdl->Grid[L_INDEX_4(kdl->kd.Dims, x, y, z)]; |
| 221 | |
| 222 | // Load PDFs of local cell: pdf_N = src[adjList[adjListIndex + D3Q19_S]]; ... |
| 223 | kd->PdfsActive[P_INDEX_3(kdl->nCells, index, D3Q19_C)] = pdfs[D3Q19_C]; |
| 224 | |
| 225 | #define ADJ_LIST(dir) adjList[(index - (index % VSIZE)) * N_D3Q19_IDX + (dir * VSIZE) + (index % VSIZE)] |
| 226 | #define X(name, idx, idxinv, _x, _y, _z) kd->PdfsActive[ADJ_LIST(idxinv)] = pdfs[idx]; |
| 227 | D3Q19_LIST_WO_C |
| 228 | #undef X |
| 229 | #undef ADJ_LIST |
| 230 | |
| 231 | } else { |
| 232 | |
| 233 | #define I(x, y, z, dir) P_INDEX_5(KDL(kd), (x), (y), (z), (dir)) |
| 234 | #define X(name, idx, idxinv, _x, _y, _z) kd->PdfsActive[I(x, y, z, idx)] = pdfs[idx]; |
| 235 | D3Q19_LIST |
| 236 | #undef X |
| 237 | #undef I |
| 238 | |
| 239 | } |
| 240 | |
| 241 | return; |
| 242 | } |
| 243 | |
| 244 | static void ParameterUsage() |
| 245 | { |
| 246 | printf("Kernel parameters:\n"); |
| 247 | printf(" [-blk <n>] [-blk-[xyz] <n>]\n"); |
| 248 | |
| 249 | return; |
| 250 | } |
| 251 | |
| 252 | static void ParseParameters(Parameters * params, int * blk) |
| 253 | { |
| 254 | Assert(blk != NULL); |
| 255 | |
| 256 | blk[0] = 0; blk[1] = 0; blk[2] = 0; |
| 257 | |
| 258 | #define ARG_IS(param) (!strcmp(params->KernelArgs[i], param)) |
| 259 | #define NEXT_ARG_PRESENT() \ |
| 260 | do { \ |
| 261 | if (i + 1 >= params->nKernelArgs) { \ |
| 262 | printf("ERROR: argument %s requires a parameter.\n", params->KernelArgs[i]); \ |
| 263 | exit(1); \ |
| 264 | } \ |
| 265 | } while (0) |
| 266 | |
| 267 | |
| 268 | for (int i = 0; i < params->nKernelArgs; ++i) { |
| 269 | if (ARG_IS("-blk") || ARG_IS("--blk")) { |
| 270 | NEXT_ARG_PRESENT(); |
| 271 | |
| 272 | int tmp = strtol(params->KernelArgs[++i], NULL, 0); |
| 273 | |
| 274 | if (tmp <= 0) { |
| 275 | printf("ERROR: blocking parameter must be > 0.\n"); |
| 276 | exit(1); |
| 277 | } |
| 278 | |
| 279 | blk[0] = blk[1] = blk[2] = tmp; |
| 280 | } |
| 281 | else if (ARG_IS("-blk-x") || ARG_IS("--blk-x")) { |
| 282 | NEXT_ARG_PRESENT(); |
| 283 | |
| 284 | int tmp = strtol(params->KernelArgs[++i], NULL, 0); |
| 285 | |
| 286 | if (tmp <= 0) { |
| 287 | printf("ERROR: blocking parameter must be > 0.\n"); |
| 288 | exit(1); |
| 289 | } |
| 290 | |
| 291 | blk[0] = tmp; |
| 292 | } |
| 293 | else if (ARG_IS("-blk-y") || ARG_IS("--blk-y")) { |
| 294 | NEXT_ARG_PRESENT(); |
| 295 | |
| 296 | int tmp = strtol(params->KernelArgs[++i], NULL, 0); |
| 297 | |
| 298 | if (tmp <= 0) { |
| 299 | printf("ERROR: blocking parameter must be > 0.\n"); |
| 300 | exit(1); |
| 301 | } |
| 302 | |
| 303 | blk[1] = tmp; |
| 304 | } |
| 305 | else if (ARG_IS("-blk-z") || ARG_IS("--blk-z")) { |
| 306 | NEXT_ARG_PRESENT(); |
| 307 | |
| 308 | int tmp = strtol(params->KernelArgs[++i], NULL, 0); |
| 309 | |
| 310 | if (tmp <= 0) { |
| 311 | printf("ERROR: blocking parameter must be > 0.\n"); |
| 312 | exit(1); |
| 313 | } |
| 314 | |
| 315 | blk[2] = tmp; |
| 316 | } |
| 317 | else if (ARG_IS("-h") || ARG_IS("-help") || ARG_IS("--help")) { |
| 318 | ParameterUsage(); |
| 319 | exit(1); |
| 320 | } |
| 321 | else { |
| 322 | printf("ERROR: unknown kernel parameter.\n"); |
| 323 | ParameterUsage(); |
| 324 | exit(1); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | #undef ARG_IS |
| 329 | #undef NEXT_ARG_PRESENT |
| 330 | |
| 331 | return; |
| 332 | } |
| 333 | |
| 334 | static void SetupConsecNodes(LatticeDesc * ld, KernelDataListRia * kdlr, int nThreads) |
| 335 | { |
| 336 | Assert(ld != NULL); |
| 337 | Assert(kdlr != NULL); |
| 338 | Assert(nThreads > 0); |
| 339 | |
| 340 | uint32_t * adjList = kdlr->kdl.AdjList; |
| 341 | |
| 342 | uint32_t nConsecNodes = 0; |
| 343 | uint32_t consecIndex = 0; |
| 344 | |
| 345 | int nFluid = kdlr->kdl.nFluid; |
| 346 | |
| 347 | uint32_t * consecThreadIndices = (uint32_t *)malloc(sizeof(uint32_t) * (nThreads + 1)); |
| 348 | |
| 349 | int nNodesPerThread = nFluid / nThreads; |
| 350 | |
| 351 | for (int i = 0; i < nThreads; ++i) { |
| 352 | consecThreadIndices[i] = i * nNodesPerThread + MinI(i, nFluid % nThreads); |
| 353 | } |
| 354 | consecThreadIndices[nThreads] = nFluid; |
| 355 | |
| 356 | int indexThread = 1; |
| 357 | int similarPatterns = 0; |
| 358 | int wasLastChunkThreadBoundary = 0; |
| 359 | // We execute following code two times. |
| 360 | // - The first time to get the count of how many entries we need for the |
| 361 | // consecNodes array. |
| 362 | // - The second time to fill the array. |
| 363 | |
| 364 | // Loop over adjacency list of all nodes. |
| 365 | // Compare if adjacent nodes share the same access pattern. |
| 366 | for (int fluidBaseIndex = VSIZE; fluidBaseIndex < nFluid; fluidBaseIndex += VSIZE) { |
| 367 | |
| 368 | int hasSimilarAccessPattern = 1; |
| 369 | |
| 370 | // Loop over all directions except the center one. |
| 371 | for(int d = 0; d < N_D3Q19 - 1; ++d) { |
| 372 | Assert(d != D3Q19_C); |
| 373 | |
| 374 | // check if cache line itself has consecutive memory access pattern |
| 375 | for(int inChunkIndex = 0; (inChunkIndex < VSIZE) && ((fluidBaseIndex + inChunkIndex) < nFluid); ++inChunkIndex){ |
| 376 | int index = fluidBaseIndex + inChunkIndex; |
| 377 | |
| 378 | Assert(index < nFluid); |
| 379 | |
| 380 | #define ADJ_LIST(idx, dir) adjList[((idx) - ((idx) % VSIZE)) * N_D3Q19_IDX + ((dir) * VSIZE) + ((idx) % VSIZE)] |
| 381 | //if (adjList[index * N_D3Q19_IDX + d] != adjList[(index - 1) * N_D3Q19_IDX + d] + 1) |
| 382 | if (ADJ_LIST(index, d) != ADJ_LIST(index-VSIZE, d) + VSIZE) { |
| 383 | //printf("different @: ADJ_LST(%d,%d)=%d != %d=ADJ_LST(%d, %d) + VSIZE\n", index, d, ADJ_LIST(index,d), ADJ_LIST(index-VSIZE,d) + VSIZE, index-VSIZE, d); |
| 384 | // Different access pattern. |
| 385 | hasSimilarAccessPattern = 0; |
| 386 | break; |
| 387 | } |
| 388 | #undef ADJ_LIST |
| 389 | } |
| 390 | |
| 391 | if(!hasSimilarAccessPattern){ |
| 392 | break; //exit from nested loop |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | long threadBoundaryIndex = consecThreadIndices[indexThread]; |
| 397 | if (fluidBaseIndex <= threadBoundaryIndex && |
| 398 | threadBoundaryIndex < fluidBaseIndex + VSIZE) { |
| 399 | // Current chunk contains thread boundary. |
| 400 | // These chunks are treated by scalar peel and reminder loops |
| 401 | // in kernel of every thread to ensure VSIZE aligned access to |
| 402 | // adjacency list |
| 403 | |
| 404 | // final cells of current thread |
| 405 | ++consecIndex; |
| 406 | |
| 407 | // first cells of next thread |
| 408 | ++indexThread; |
| 409 | ++consecIndex; |
| 410 | |
| 411 | wasLastChunkThreadBoundary = 1; |
| 412 | } |
| 413 | else { |
| 414 | // We are not at a thread boundary |
| 415 | if (hasSimilarAccessPattern && !wasLastChunkThreadBoundary){ |
| 416 | ++similarPatterns; |
| 417 | } |
| 418 | else { |
| 419 | ++consecIndex; |
| 420 | } |
| 421 | |
| 422 | wasLastChunkThreadBoundary = 0; |
| 423 | |
| 424 | /* |
| 425 | if (!hasSimilarAccessPattern) { |
| 426 | ++consecIndex; |
| 427 | } |
| 428 | else { |
| 429 | ++similarPatterns; |
| 430 | } |
| 431 | */ |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | if (nFluid > 0) { |
| 436 | nConsecNodes = consecIndex + 1; |
| 437 | } |
| 438 | |
| 439 | uint32_t * consecNodes; |
| 440 | MemAlloc((void **)&consecNodes, sizeof(uint32_t) * nConsecNodes); |
| 441 | |
| 442 | unsigned long consecNodesByte = (nConsecNodes) * sizeof(uint32_t); |
| 443 | |
| 444 | printf("# Consec. Nodes Array Allocation:\n"); |
| 445 | printf("# similar patterns\t\t%d\n", similarPatterns); |
| 446 | printf("# elements: \t\t%d\n", nConsecNodes); |
| 447 | printf("# size: \t\t%e MiB\n", consecNodesByte / 1024.0 / 1024.0); |
| 448 | printf("# alignment: \t\t%d b\n", PAGE_4K); |
| 449 | |
| 450 | if (MemAllocAligned((void **)&consecNodesByte, consecNodesByte, PAGE_4K)) { |
| 451 | printf("ERROR: allocating consecNodes array with MemAllocAligned failed: %lu bytes.\n", consecNodesByte); |
| 452 | exit(1); |
| 453 | } |
| 454 | else { |
| 455 | printf("# allocator: \t\t\tMemAllocAligned()\n"); |
| 456 | } |
| 457 | |
| 458 | consecIndex = 0; |
| 459 | |
| 460 | if (nFluid > 0) { |
| 461 | consecNodes[consecIndex] = VSIZE; |
| 462 | } |
| 463 | |
| 464 | indexThread = 1; |
| 465 | consecThreadIndices[0] = 0; |
| 466 | |
| 467 | //add first chunk manually to enable backward check for consecutive pattern |
| 468 | consecNodes[consecIndex] = VSIZE; |
| 469 | |
| 470 | wasLastChunkThreadBoundary = 0; |
| 471 | |
| 472 | // Loop over adjacency list of all nodes. |
| 473 | // Compare if access pattern does not change on chunk level |
| 474 | // Since gather instructions are used, access pattern may not be consecutive |
| 475 | for (int fluidBaseIndex = VSIZE; fluidBaseIndex < nFluid; fluidBaseIndex += VSIZE) { |
| 476 | |
| 477 | int hasSimilarAccessPattern = 1; |
| 478 | |
| 479 | // Loop over all directions except the center one. |
| 480 | for(int d = 0; d < N_D3Q19 - 1; ++d) { |
| 481 | Assert(d != D3Q19_C); |
| 482 | |
| 483 | // check if cache line itself has consecutive memory access pattern |
| 484 | for(int inChunkIndex = 0; (inChunkIndex < VSIZE) && ((fluidBaseIndex + inChunkIndex) < nFluid); ++inChunkIndex){ |
| 485 | int index = fluidBaseIndex + inChunkIndex; |
| 486 | |
| 487 | Assert(index < nFluid); |
| 488 | |
| 489 | #define ADJ_LIST(idx, dir) adjList[((idx) - ((idx) % VSIZE)) * N_D3Q19_IDX + ((dir) * VSIZE) + ((idx) % VSIZE)] |
| 490 | //if (adjList[index * N_D3Q19_IDX + d] != adjList[(index - 1) * N_D3Q19_IDX + d] + 1) |
| 491 | if (ADJ_LIST(index, d) != ADJ_LIST(index-VSIZE, d) + VSIZE) { |
| 492 | // Different access pattern. |
| 493 | hasSimilarAccessPattern = 0; |
| 494 | break; |
| 495 | } |
| 496 | #undef ADJ_LIST |
| 497 | } |
| 498 | |
| 499 | if(!hasSimilarAccessPattern){ |
| 500 | break; //exit from nested loop |
| 501 | } |
| 502 | } |
| 503 | |
| 504 | long threadBoundaryIndex = consecThreadIndices[indexThread]; |
| 505 | if (fluidBaseIndex <= threadBoundaryIndex && |
| 506 | threadBoundaryIndex < fluidBaseIndex + VSIZE) { |
| 507 | // Current chunk contains thread boundary. |
| 508 | // These chunks are treated by scalar peel and reminder loops |
| 509 | // in kernel of every thread to ensure VSIZE aligned access to |
| 510 | // adjacency list |
| 511 | |
| 512 | // final cells of current thread |
| 513 | ++consecIndex; |
| 514 | //consecThreadIndices[indexThread] = consecIndex; |
| 515 | consecNodes[consecIndex] = threadBoundaryIndex - fluidBaseIndex; |
| 516 | |
| 517 | |
| 518 | // first cells of next thread |
| 519 | ++consecIndex; |
| 520 | consecThreadIndices[indexThread] = consecIndex; |
| 521 | consecNodes[consecIndex] = (fluidBaseIndex + VSIZE) - threadBoundaryIndex; |
| 522 | ++indexThread; |
| 523 | |
| 524 | wasLastChunkThreadBoundary = 1; |
| 525 | |
| 526 | } |
| 527 | else { |
| 528 | // We are not at a thread boundary |
| 529 | if (hasSimilarAccessPattern && !wasLastChunkThreadBoundary){ |
| 530 | Assert(consecIndex < nConsecNodes); |
| 531 | consecNodes[consecIndex] += VSIZE; |
| 532 | } |
| 533 | else { |
| 534 | ++consecIndex; |
| 535 | Assert(consecIndex < nConsecNodes); |
| 536 | consecNodes[consecIndex] = VSIZE; |
| 537 | } |
| 538 | |
| 539 | /* |
| 540 | if (!hasSimilarAccessPattern) { |
| 541 | ++consecIndex; |
| 542 | Assert(consecIndex < nConsecNodes); |
| 543 | consecNodes[consecIndex] = VSIZE; |
| 544 | } |
| 545 | else { |
| 546 | Assert(consecIndex < nConsecNodes); |
| 547 | consecNodes[consecIndex] += VSIZE; |
| 548 | } |
| 549 | */ |
| 550 | wasLastChunkThreadBoundary = 0; |
| 551 | |
| 552 | } |
| 553 | } |
| 554 | |
| 555 | /* |
| 556 | printf("consecNodes:\n"); |
| 557 | for(int i = 0; i < nConsecNodes + 5; ++i){ |
| 558 | printf("%d ", consecNodes[i]); |
| 559 | } |
| 560 | printf("\n"); |
| 561 | */ |
| 562 | /* |
| 563 | printf("consecThreadIndices:\n"); |
| 564 | for(int i = 0; i < nThreads + 5; ++i){ |
| 565 | printf("%d ", consecThreadIndices[i]); |
| 566 | } |
| 567 | printf("\n"); |
| 568 | */ |
| 569 | |
| 570 | kdlr->ConsecNodes = consecNodes; |
| 571 | kdlr->nConsecNodes = nConsecNodes; |
| 572 | |
| 573 | kdlr->ConsecThreadIndices = consecThreadIndices; |
| 574 | kdlr->nConsecThreadIndices = nThreads; |
| 575 | |
| 576 | double loopBalanceEven = 2.0 * 19 * sizeof(PdfT); |
| 577 | //N_D3Q19 - 1: C lookup not required, +1: transfer of consecValue |
| 578 | double loopBalanceOdd = 2.0 * 19 * sizeof(PdfT) + ((double)nConsecNodes *((N_D3Q19 - 1) * VSIZE + 1)) / nFluid * sizeof(int); |
| 579 | double loopBalance = (loopBalanceEven + loopBalanceOdd) / 2.0; |
| 580 | |
| 581 | kdlr->kdl.kd.LoopBalance = loopBalance; |
| 582 | |
| 583 | printf("# loop balance:\n"); |
| 584 | printf("# even timestep: \t\t%.2f B/FLUP\n", loopBalanceEven); |
| 585 | printf("# odd timestep: \t\t%.2f B/FLUP\n", loopBalanceOdd); |
| 586 | printf("# average: \t\t%.2f B/FLUP\n", loopBalance); |
| 587 | |
| 588 | return; |
| 589 | } |
| 590 | |
| 591 | void FNAME(D3Q19ListAaPvGatherInit)(LatticeDesc * ld, KernelData ** kernelData, Parameters * params) |
| 592 | { |
| 593 | KernelData * kd; |
| 594 | KernelDataList * kdl; |
| 595 | KernelDataListRia * kdlr; |
| 596 | MemAlloc((void **)&kdlr, sizeof(KernelDataListRia)); |
| 597 | |
| 598 | kd = (KernelData *)kdlr; |
| 599 | kdl = KDL(kdlr); |
| 600 | |
| 601 | *kernelData = kd; |
| 602 | |
| 603 | #ifdef DEBUG |
| 604 | kd->Pdfs[0] = NULL; |
| 605 | kd->Pdfs[1] = NULL; |
| 606 | kd->PdfsActive = NULL; |
| 607 | kd->DstPdfs = NULL; |
| 608 | kd->SrcPdfs = NULL; |
| 609 | kd->Dims[0] = -1; |
| 610 | kd->Dims[1] = -1; |
| 611 | kd->Dims[2] = -1; |
| 612 | kd->GlobalDims[0] = -1; |
| 613 | kd->GlobalDims[1] = -1; |
| 614 | kd->GlobalDims[2] = -1; |
| 615 | kd->Offsets[0] = -1; |
| 616 | kd->Offsets[1] = -1; |
| 617 | kd->Offsets[2] = -1; |
| 618 | |
| 619 | kd->ObstIndices = NULL; |
| 620 | kd->nObstIndices = -1; |
| 621 | kd->BounceBackPdfsSrc = NULL; |
| 622 | kd->BounceBackPdfsDst = NULL; |
| 623 | kd->nBounceBackPdfs = -1; |
| 624 | |
| 625 | kdl->AdjList = NULL; |
| 626 | kdl->Coords = NULL; |
| 627 | kdl->Grid = NULL; |
| 628 | kdl->nCells = -1; |
| 629 | kdl->nFluid = -1; |
| 630 | |
| 631 | kdlr->ConsecNodes = NULL; |
| 632 | kdlr->nConsecNodes = 0; |
| 633 | kdlr->ConsecThreadIndices = NULL; |
| 634 | kdlr->nConsecThreadIndices = 0; |
| 635 | #endif |
| 636 | |
| 637 | // Ajust the dimensions according to padding, if used. |
| 638 | kd->Dims[0] = kd->GlobalDims[0] = ld->Dims[0]; |
| 639 | kd->Dims[1] = kd->GlobalDims[1] = ld->Dims[1]; |
| 640 | kd->Dims[2] = kd->GlobalDims[2] = ld->Dims[2]; |
| 641 | |
| 642 | int * lDims = ld->Dims; |
| 643 | |
| 644 | int lX = lDims[0]; |
| 645 | int lY = lDims[1]; |
| 646 | int lZ = lDims[2]; |
| 647 | |
| 648 | int nTotalCells = lX * lY * lZ; |
| 649 | int nCells = ld->nFluid; // TODO: + padding |
| 650 | int nFluid = ld->nFluid; |
| 651 | |
| 652 | // TODO: check nCells/nFluid do not exceed 2^31. This actually has to be |
| 653 | // done during lattice setup. |
| 654 | kdl->nCells = nCells; |
| 655 | kdl->nFluid = nFluid; |
| 656 | |
| 657 | PdfT * pdfs[2]; |
| 658 | |
| 659 | int blk[3] = { 0 }; |
| 660 | |
| 661 | ParseParameters(params, blk); |
| 662 | |
| 663 | if (blk[0] == 0) blk[0] = lX; |
| 664 | if (blk[1] == 0) blk[1] = lY; |
| 665 | if (blk[2] == 0) blk[2] = lZ; |
| 666 | |
| 667 | printf("# blocking: \t\tx: %3d y: %3d z: %3d\n", blk[0], blk[1], blk[2]); |
| 668 | |
| 669 | unsigned long latByte = nCells * sizeof(PdfT) * N_D3Q19; |
| 670 | unsigned long latFluidByte = nFluid * sizeof(PdfT) * N_D3Q19; |
| 671 | unsigned long latPadByte = (nCells - nFluid) * sizeof(PdfT) * N_D3Q19; |
| 672 | |
| 673 | printf("# Lattice Array Allocation:\n"); |
| 674 | printf("# lattice size: \t\t%e MiB\n", latByte / 1024.0 / 1024.0); |
| 675 | printf("# fluid lattice size:\t\t%e MiB\n", latFluidByte / 1024.0 / 1024.0); |
| 676 | printf("# lattice padding: \t\t%e MiB\n", latPadByte / 1024.0 / 1024.0); |
| 677 | |
| 678 | |
| 679 | printf("# alignment: \t\t%d b\n", PAGE_4K); |
| 680 | |
| 681 | if (PDF_ALLOCATOR((void **)&pdfs[0], latFluidByte, PAGE_4K)) { |
| 682 | printf("ERROR: allocating PDF array with %s() failed: %lu bytes.\n", STRINGIFY(PDF_ALLOCATOR), latFluidByte); |
| 683 | exit(1); |
| 684 | } |
| 685 | else { |
| 686 | printf("# allocator: \t\t\t%s()\n", STRINGIFY(PDF_ALLOCATOR)); |
| 687 | } |
| 688 | |
| 689 | kd->Pdfs[0] = pdfs[0]; |
| 690 | |
| 691 | // Initialize PDFs with some (arbitrary) data for correct NUMA placement. |
| 692 | // Here we touch only the fluid nodes as this loop is OpenMP parallel and |
| 693 | // we want the same scheduling as in the kernel. |
| 694 | #ifdef _OPENMP |
| 695 | #pragma omp parallel for |
| 696 | #endif |
| 697 | for (int i = 0; i < nFluid; ++i) { for(int d = 0; d < N_D3Q19; ++d) { |
| 698 | pdfs[0][P_INDEX_3(nCells, i, d)] = 1.0; |
| 699 | } } |
| 700 | |
| 701 | // Initialize all PDFs to some standard value. |
| 702 | for (int i = 0; i < nFluid; ++i) { for(int d = 0; d < N_D3Q19; ++d) { |
| 703 | pdfs[0][P_INDEX_3(nCells, i, d)] = 0.0; |
| 704 | } } |
| 705 | |
| 706 | // ---------------------------------------------------------------------- |
| 707 | // create grid which will hold the index numbers of the fluid nodes |
| 708 | |
| 709 | uint32_t * grid; |
| 710 | |
| 711 | if (MemAlloc((void **)&grid, nTotalCells * sizeof(uint32_t))) { |
| 712 | printf("ERROR: allocating grid for numbering failed: %lu bytes.\n", nTotalCells * sizeof(uint32_t)); |
| 713 | exit(1); |
| 714 | } |
| 715 | kdl->Grid = grid; |
| 716 | |
| 717 | int latticeIndex; |
| 718 | |
| 719 | #ifdef DEBUG |
| 720 | for(int z = 0; z < lZ; ++z) { |
| 721 | for(int y = 0; y < lY; ++y) { |
| 722 | for(int x = 0; x < lX; ++x) { |
| 723 | |
| 724 | latticeIndex = L_INDEX_4(ld->Dims, x, y, z); |
| 725 | |
| 726 | grid[latticeIndex] = ~0; |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | #endif |
| 731 | |
| 732 | // ---------------------------------------------------------------------- |
| 733 | // generate numbering over grid |
| 734 | |
| 735 | uint32_t * coords; |
| 736 | |
| 737 | if (MemAlloc((void **)&coords, nFluid * sizeof(uint32_t) * 3)) { |
| 738 | printf("ERROR: allocating coords array failed: %lu bytes.\n", nFluid * sizeof(uint32_t) * 3); |
| 739 | exit(1); |
| 740 | } |
| 741 | |
| 742 | kdl->Coords = coords; |
| 743 | |
| 744 | // Index for the PDF nodes can start at 0 as we distinguish solid and fluid nodes |
| 745 | // through the ld->Lattice array. |
| 746 | int counter = 0; |
| 747 | |
| 748 | // Blocking is implemented via setup of the adjacency list. The kernel later will |
| 749 | // walk through the lattice blocked automatically. |
| 750 | for (int bZ = 0; bZ < lZ; bZ += blk[2]) { |
| 751 | for (int bY = 0; bY < lY; bY += blk[1]) { |
| 752 | for (int bX = 0; bX < lX; bX += blk[0]) { |
| 753 | |
| 754 | int eX = MIN(bX + blk[0], lX); |
| 755 | int eY = MIN(bY + blk[1], lY); |
| 756 | int eZ = MIN(bZ + blk[2], lZ); |
| 757 | |
| 758 | |
| 759 | for (int z = bZ; z < eZ; ++z) { |
| 760 | for (int y = bY; y < eY; ++y) { |
| 761 | for (int x = bX; x < eX; ++x) { |
| 762 | |
| 763 | latticeIndex = L_INDEX_4(lDims, x, y, z); |
| 764 | |
| 765 | if (ld->Lattice[latticeIndex] != LAT_CELL_OBSTACLE) { |
| 766 | grid[latticeIndex] = counter; |
| 767 | |
| 768 | coords[C_INDEX_X(counter)] = x; |
| 769 | coords[C_INDEX_Y(counter)] = y; |
| 770 | coords[C_INDEX_Z(counter)] = z; |
| 771 | |
| 772 | ++counter; |
| 773 | } |
| 774 | } } } |
| 775 | } } } |
| 776 | |
| 777 | Verify(counter == nFluid); |
| 778 | uint32_t * adjList; |
| 779 | |
| 780 | // AoSoA addressing for adjList needs padding for (nFluid % VSIZE) != 0 |
| 781 | unsigned long adjListBytes = nFluid * sizeof(int) * N_D3Q19_IDX; |
| 782 | |
| 783 | printf("# Adjacency List Allocation:\n"); |
| 784 | printf("# size: \t\t%e MiB\n", adjListBytes / 1024.0 / 1024.0); |
| 785 | printf("# alignment: \t\t%d b\n", PAGE_4K); |
| 786 | |
| 787 | // AdjList only requires 18 instead of 19 entries per node, as |
| 788 | // the center PDF needs no addressing. |
| 789 | if (ADJ_LIST_ALLOCATOR((void **)&adjList, adjListBytes, PAGE_4K)) { |
| 790 | printf("ERROR: allocating adjList array with %s() failed: %lu bytes.\n", STRINGIFY(ADJ_LIST_ALLOCATOR), adjListBytes); |
| 791 | exit(1); |
| 792 | } |
| 793 | else { |
| 794 | printf("# allocator: \t\t\t%s()\n", STRINGIFY(ADJ_LIST_ALLOCATOR)); |
| 795 | } |
| 796 | |
| 797 | kdl->AdjList = adjList; |
| 798 | |
| 799 | int x, y, z; |
| 800 | |
| 801 | uint32_t neighborIndex; |
| 802 | uint32_t dstIndex; |
| 803 | |
| 804 | int nx, ny, nz, px, py, pz; |
| 805 | |
| 806 | // Loop over all fluid nodes and compute the indices to the neighboring |
| 807 | // PDFs for configured data layout (AoS/SoA). |
| 808 | // Parallelized loop to ensure correct NUMA placement. |
| 809 | // #ifdef _OPENMP --> add line continuation |
| 810 | // #pragma omp parallel for default(none) |
| 811 | // shared(nFluid, nCells, coords, D3Q19_INV, D3Q19_X, D3Q19_Y, D3Q19_Z, |
| 812 | // stderr, |
| 813 | // lDims, grid, ld, lX, lY, lZ, adjList) |
| 814 | // private(x, y, z, nx, ny, nz, neighborIndex, dstIndex) |
| 815 | // #endif |
| 816 | for (int fluidBaseIndex = 0; fluidBaseIndex < nFluid; fluidBaseIndex+=VSIZE) { |
| 817 | |
| 818 | |
| 819 | // Loop over all directions except the center one. |
| 820 | for(int d = 0; d < N_D3Q19 - 1; ++d) { |
| 821 | Assert(d != D3Q19_C); |
| 822 | |
| 823 | for(int inChunkIndex = 0; (inChunkIndex < VSIZE) && ((fluidBaseIndex + inChunkIndex) < nFluid); ++inChunkIndex){ |
| 824 | int index = fluidBaseIndex + inChunkIndex; |
| 825 | |
| 826 | Assert(index < nFluid); |
| 827 | |
| 828 | x = coords[C_INDEX_X(index)]; |
| 829 | y = coords[C_INDEX_Y(index)]; |
| 830 | z = coords[C_INDEX_Z(index)]; |
| 831 | |
| 832 | Assert(x >= 0 && x < lX); |
| 833 | Assert(y >= 0 && y < lY); |
| 834 | Assert(z >= 0 && z < lZ); |
| 835 | |
| 836 | Assert(ld->Lattice[L_INDEX_4(lDims, x, y, z)] != LAT_CELL_OBSTACLE); |
| 837 | |
| 838 | #ifdef PROP_MODEL_PUSH |
| 839 | nx = x + D3Q19_X[d]; |
| 840 | ny = y + D3Q19_Y[d]; |
| 841 | nz = z + D3Q19_Z[d]; |
| 842 | |
| 843 | #elif PROP_MODEL_PULL |
| 844 | nx = x - D3Q19_X[d]; |
| 845 | ny = y - D3Q19_Y[d]; |
| 846 | nz = z - D3Q19_Z[d]; |
| 847 | #else |
| 848 | #error No implementation for this PROP_MODEL_NAME. |
| 849 | #endif |
| 850 | // If the neighbor is outside the latice in X direction and we have a |
| 851 | // periodic boundary then we need to wrap around. |
| 852 | if ( ((nx < 0 || nx >= lX) && ld->PeriodicX) || |
| 853 | ((ny < 0 || ny >= lY) && ld->PeriodicY) || |
| 854 | ((nz < 0 || nz >= lZ) && ld->PeriodicZ) |
| 855 | ){ |
| 856 | // x periodic |
| 857 | |
| 858 | if (nx < 0) { |
| 859 | px = lX - 1; |
| 860 | } |
| 861 | else if (nx >= lX) { |
| 862 | px = 0; |
| 863 | } else { |
| 864 | px = nx; |
| 865 | } |
| 866 | // y periodic |
| 867 | if (ny < 0) { |
| 868 | py = lY - 1; |
| 869 | } |
| 870 | else if (ny >= lY) { |
| 871 | py = 0; |
| 872 | } else { |
| 873 | py = ny; |
| 874 | } |
| 875 | |
| 876 | // z periodic |
| 877 | if (nz < 0) { |
| 878 | pz = lZ - 1; |
| 879 | } |
| 880 | else if (nz >= lZ) { |
| 881 | pz = 0; |
| 882 | } else { |
| 883 | pz = nz; |
| 884 | } |
| 885 | |
| 886 | if (ld->Lattice[L_INDEX_4(lDims, px, py, pz)] == LAT_CELL_OBSTACLE) { |
| 887 | dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]); |
| 888 | } |
| 889 | else { |
| 890 | neighborIndex = grid[L_INDEX_4(lDims, px, py, pz)]; |
| 891 | |
| 892 | AssertMsg(neighborIndex != ~0, "Neighbor has no Index. (%d %d %d) direction %s (%d)\n", px, py, pz, D3Q19_NAMES[d], d); |
| 893 | |
| 894 | dstIndex = P_INDEX_3(nCells, neighborIndex, d); |
| 895 | } |
| 896 | } |
| 897 | else if (nx < 0 || ny < 0 || nz < 0 || nx >= lX || ny >= lY || nz >= lZ) { |
| 898 | dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]); |
| 899 | } |
| 900 | else if (ld->Lattice[L_INDEX_4(lDims, nx, ny, nz)] == LAT_CELL_OBSTACLE) { |
| 901 | dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]); |
| 902 | } |
| 903 | else { |
| 904 | neighborIndex = grid[L_INDEX_4(lDims, nx, ny, nz)]; |
| 905 | |
| 906 | Assert(neighborIndex != ~0); |
| 907 | |
| 908 | dstIndex = P_INDEX_3(nCells, neighborIndex, d); |
| 909 | } |
| 910 | |
| 911 | Assert(dstIndex >= 0); |
| 912 | Assert(dstIndex < nCells * N_D3Q19); |
| 913 | |
| 914 | adjList[(index - (index % VSIZE)) * N_D3Q19_IDX + (d * VSIZE) + (index % VSIZE)] = dstIndex; |
| 915 | } |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | /* |
| 920 | printf("============\n"); |
| 921 | for(int baseIndex = 0; baseIndex < nFluid; baseIndex+=VSIZE){ |
| 922 | for(int i = 0; i < VSIZE; ++i){ |
| 923 | int index = baseIndex + i; |
| 924 | |
| 925 | printf("%d ", adjList[(index - (index % VSIZE)) * N_D3Q19_IDX + (0 * VSIZE) + (index % VSIZE)]); |
| 926 | } |
| 927 | printf("\n"); |
| 928 | } |
| 929 | printf("============\n"); |
| 930 | */ |
| 931 | |
| 932 | int nThreads = 1; |
| 933 | |
| 934 | #ifdef _OPENMP |
| 935 | nThreads = omp_get_max_threads(); |
| 936 | #endif |
| 937 | |
| 938 | SetupConsecNodes(ld, KDLR(kd), nThreads); |
| 939 | |
| 940 | // Fill remaining KernelData structures |
| 941 | kd->GetNode = GetNode; |
| 942 | kd->SetNode = SetNode; |
| 943 | |
| 944 | kd->BoundaryConditionsGetPdf = FNAME(BCGetPdf); |
| 945 | kd->BoundaryConditionsSetPdf = FNAME(BCSetPdf); |
| 946 | |
| 947 | kd->Kernel = FNAME(D3Q19ListAaPvGatherKernel); |
| 948 | |
| 949 | kd->DstPdfs = NULL; |
| 950 | kd->PdfsActive = kd->Pdfs[0]; |
| 951 | |
| 952 | return; |
| 953 | } |
| 954 | |
| 955 | void FNAME(D3Q19ListAaPvGatherDeinit)(LatticeDesc * ld, KernelData ** kernelData) |
| 956 | { |
| 957 | KernelDataListRia ** kdlr = (KernelDataListRia **)kernelData; |
| 958 | |
| 959 | MemFree((void **)&((*kdlr)->ConsecNodes)); |
| 960 | |
| 961 | if ((*kdlr)->ConsecThreadIndices != NULL) { |
| 962 | MemFree((void **)&((*kdlr)->ConsecThreadIndices)); |
| 963 | } |
| 964 | |
| 965 | KernelDataList ** kdl = (KernelDataList **)kernelData; |
| 966 | |
| 967 | ADJ_LIST_FREE((void **)&((*kdl)->AdjList)); |
| 968 | |
| 969 | MemFree((void **)&((*kdl)->Coords)); |
| 970 | MemFree((void **)&((*kdl)->Grid)); |
| 971 | |
| 972 | PDF_FREE((void **)&((*kernelData)->Pdfs[0])); |
| 973 | |
| 974 | MemFree((void **)kernelData); |
| 975 | return; |
| 976 | } |
| 977 | |