1 // --------------------------------------------------------------------------
4 // Markus Wittmann, 2016-2017
5 // RRZE, University of Erlangen-Nuremberg, Germany
6 // markus.wittmann -at- fau.de or hpc -at- rrze.fau.de
9 // LSS, University of Erlangen-Nuremberg, Germany
11 // This file is part of the Lattice Boltzmann Benchmark Kernels (LbmBenchKernels).
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.
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.
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/>.
26 // --------------------------------------------------------------------------
27 #include "BenchKernelD3Q19AaCommon.h"
40 void FNAME(D3Q19AaKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData * cd)
43 Assert(kernelData != NULL);
46 Assert(cd->Omega > F(0.0));
47 Assert(cd->Omega < F(2.0));
49 KernelData * kd = (KernelData *)kernelData;
56 int * gDims = kd->GlobalDims;
58 int oX = kd->Offsets[0];
59 int oY = kd->Offsets[1];
60 int oZ = kd->Offsets[2];
62 KernelDataAa * kda = KDA(kd);
69 PdfT omega = cd->Omega;
70 PdfT omegaEven = omega;
71 PdfT magicParam = F(1.0) / F(12.0);
72 // 1/4: best stability;
73 // 1/12: removes third-order advection error (best advection);
74 // 1/6: removes fourth-order diffusion error (best diffusion);
75 // 3/16: exact location of bounce back for poiseuille flow
77 PdfT omegaOdd = F(1.0)/( F(0.5) + magicParam/(F(1.0)/omega - F(0.5)) );
79 PdfT evenPart = F(0.0);
80 PdfT oddPart = F(0.0);
81 PdfT dir_indep_trm = F(0.0);
83 PdfT w_0 = F(1.0) / F(3.0);
84 PdfT w_1 = F(1.0) / F(18.0);
85 PdfT w_2 = F(1.0) / F(36.0);
87 PdfT w_1_x3 = w_1 * F(3.0); PdfT w_1_nine_half = w_1 * F(9.0)/F(2.0); PdfT w_1_indep = F(0.0);
88 PdfT w_2_x3 = w_2 * F(3.0); PdfT w_2_nine_half = w_2 * F(9.0)/F(2.0); PdfT w_2_indep = F(0.0);
93 // Declare pdf_N, pdf_E, pdf_S, pdf_W, ...
94 #define X(name, idx, idxinv, x, y, z) PdfT JOIN(pdf_,name);
98 PdfT * src = kd->Pdfs[0];
100 int maxIterations = cd->MaxIterations;
104 kd->PdfsActive = src;
105 VtkWrite(ld, kd, cd, -1);
110 kd->PdfsActive = src;
111 KernelStatistics(kd, ld, cd, 0);
118 nThreads = omp_get_max_threads();
121 X_KERNEL_START(kernelData);
123 for (int iter = 0; iter < maxIterations; iter += 2) {
125 // --------------------------------------------------------------------
128 X_LIKWID_START("aa-even");
133 #pragma omp parallel for default(none) \
134 shared(gDims,src, w_0, w_1, w_2, omegaEven, omegaOdd, \
135 w_1_x3, w_2_x3, w_1_nine_half, w_2_nine_half, cd, \
136 oX, oY, oZ, nX, nY, nZ, blk, nThreads, ld) \
137 private(ux, uy, uz, ui, dens, dir_indep_trm, \
139 pdf_N, pdf_E, pdf_S, pdf_W, \
140 pdf_NE, pdf_SE, pdf_SW, pdf_NW, \
141 pdf_T, pdf_TN, pdf_TE, pdf_TS, pdf_TW, \
142 pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW, \
143 evenPart, oddPart, w_1_indep, w_2_indep)
146 for (int i = 0; i < nThreads; ++i) {
148 int threadStartX = nX / nThreads * i;
149 int threadEndX = nX / nThreads * (i + 1);
151 if (nX % nThreads > 0) {
152 if (nX % nThreads > i) {
157 threadStartX += nX % nThreads;
158 threadEndX += nX % nThreads;
162 for (int bX = oX + threadStartX; bX < threadEndX + oX; bX += blk[0]) {
163 for (int bY = oY; bY < nY + oY; bY += blk[1]) {
164 for (int bZ = oZ; bZ < nZ + oZ; bZ += blk[2]) {
166 int eX = MIN(bX + blk[0], threadEndX + oX);
167 int eY = MIN(bY + blk[1], nY + oY);
168 int eZ = MIN(bZ + blk[2], nZ + oZ);
170 // printf("%d: %d-%d %d-%d %d-%d %d - %d\n", omp_get_thread_num(), bZ, eZ, bY, eY, bX, eX, threadStartX, threadEndX);
172 for (int x = bX; x < eX; ++x) {
173 for (int y = bY; y < eY; ++y) {
174 #ifdef INTEL_OPT_DIRECTIVES
176 #pragma vector always
179 for (int z = bZ; z < eZ; ++z) { // LOOP aa-even
182 if (ld->Lattice[L_INDEX_4(ld->Dims, x - oX, y - oY, z - oZ)] == LAT_CELL_OBSTACLE) {
186 #define I(x, y, z, dir) P_INDEX_5(gDims, (x), (y), (z), (dir))
189 // Load PDFs of local cell: pdf_N = src[I(x, y, z, D3Q19_N)]; ...
190 #define X(name, idx, idxinv, _x, _y, _z) JOIN(pdf_,name) = src[I(x, y, z, idx)];
194 // #define LID_DRIVEN_CAVITY
196 #ifdef LID_DRIVEN_CAVITY
198 if (z == nZ - 4 + oZ && x > 3 + oX && x < (nX - 4 + oX) && y > 3 + oY && y < (nY - 4 + oY)) {
199 ux = F(0.1) * F(0.5)77;
205 ux = pdf_E + pdf_NE + pdf_SE + pdf_TE + pdf_BE -
206 pdf_W - pdf_NW - pdf_SW - pdf_TW - pdf_BW;
207 uy = pdf_N + pdf_NE + pdf_NW + pdf_TN + pdf_BN -
208 pdf_S - pdf_SE - pdf_SW - pdf_TS - pdf_BS;
209 uz = pdf_T + pdf_TE + pdf_TW + pdf_TN + pdf_TS -
210 pdf_B - pdf_BE - pdf_BW - pdf_BN - pdf_BS;
211 #ifdef LID_DRIVEN_CAVITY
216 pdf_N + pdf_E + pdf_S + pdf_W +
217 pdf_NE + pdf_SE + pdf_SW + pdf_NW +
218 pdf_T + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
219 pdf_B + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
221 dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz)*F(3.0)/F(2.0);
224 src[I(x, y, z, D3Q19_C)] = pdf_C - omegaEven*(pdf_C - w_0*dir_indep_trm);
227 w_1_indep = w_1*dir_indep_trm;
230 evenPart = omegaEven*( F(0.5)*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
231 oddPart = omegaOdd*(F(0.5)*(pdf_N - pdf_S) - ui*w_1_x3 );
232 src[I(x, y, z, D3Q19_S)] = pdf_N - evenPart - oddPart;
233 src[I(x, y, z, D3Q19_N)] = pdf_S - evenPart + oddPart;
236 evenPart = omegaEven*( F(0.5)*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
237 oddPart = omegaOdd*(F(0.5)*(pdf_E - pdf_W) - ui*w_1_x3 );
238 src[I(x, y, z, D3Q19_W)] = pdf_E - evenPart - oddPart;
239 src[I(x, y, z, D3Q19_E)] = pdf_W - evenPart + oddPart;
242 evenPart = omegaEven*( F(0.5)*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
243 oddPart = omegaOdd*(F(0.5)*(pdf_T - pdf_B) - ui*w_1_x3 );
244 src[I(x, y, z, D3Q19_B)] = pdf_T - evenPart - oddPart;
245 src[I(x, y, z, D3Q19_T)] = pdf_B - evenPart + oddPart;
248 w_2_indep = w_2*dir_indep_trm;
251 evenPart = omegaEven*( F(0.5)*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
252 oddPart = omegaOdd*(F(0.5)*(pdf_NW - pdf_SE) - ui*w_2_x3 );
253 src[I(x, y, z, D3Q19_SE)] = pdf_NW - evenPart - oddPart;
254 src[I(x, y, z, D3Q19_NW)] = pdf_SE - evenPart + oddPart;
257 evenPart = omegaEven*( F(0.5)*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
258 oddPart = omegaOdd*(F(0.5)*(pdf_NE - pdf_SW) - ui*w_2_x3 );
259 src[I(x, y, z, D3Q19_SW)] = pdf_NE - evenPart - oddPart;
260 src[I(x, y, z, D3Q19_NE)] = pdf_SW - evenPart + oddPart;
263 evenPart = omegaEven*( F(0.5)*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
264 oddPart = omegaOdd*(F(0.5)*(pdf_TW - pdf_BE) - ui*w_2_x3 );
265 src[I(x, y, z, D3Q19_BE)] = pdf_TW - evenPart - oddPart;
266 src[I(x, y, z, D3Q19_TW)] = pdf_BE - evenPart + oddPart;
269 evenPart = omegaEven*( F(0.5)*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
270 oddPart = omegaOdd*(F(0.5)*(pdf_TE - pdf_BW) - ui*w_2_x3 );
271 src[I(x, y, z, D3Q19_BW)] = pdf_TE - evenPart - oddPart;
272 src[I(x, y, z, D3Q19_TE)] = pdf_BW - evenPart + oddPart;
275 evenPart = omegaEven*( F(0.5)*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
276 oddPart = omegaOdd*(F(0.5)*(pdf_TS - pdf_BN) - ui*w_2_x3 );
277 src[I(x, y, z, D3Q19_BN)] = pdf_TS - evenPart - oddPart;
278 src[I(x, y, z, D3Q19_TS)] = pdf_BN - evenPart + oddPart;
281 evenPart = omegaEven*( F(0.5)*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
282 oddPart = omegaOdd*(F(0.5)*(pdf_TN - pdf_BS) - ui*w_2_x3 );
283 src[I(x, y, z, D3Q19_BS)] = pdf_TN - evenPart - oddPart;
284 src[I(x, y, z, D3Q19_TN)] = pdf_BS - evenPart + oddPart;
287 } } } // z, y, x (from inner to outer)
288 } } } // z, y, x (from inner to outer)
290 } // loop over threads
294 X_LIKWID_STOP("aa-even");
297 kd->PdfsActive = src;
298 KernelStatistics(kd, ld, cd, iter);
301 // Fixup bounce back PDFs.
303 #pragma omp parallel for default(none) \
306 #ifdef INTEL_OPT_DIRECTIVES
309 for (int i = 0; i < kd->nBounceBackPdfs; ++i) {
310 src[kd->BounceBackPdfsSrc[i]] = src[kd->BounceBackPdfsDst[i]];
313 // save current iteration
314 kda->Iteration = iter;
317 kd->PdfsActive = src;
318 KernelAddBodyForce(kd, ld, cd);
322 if (cd->VtkOutput && (iter % cd->VtkModulus) == 0) {
323 kd->PdfsActive = src;
324 VtkWrite(ld, kd, cd, iter);
329 kd->PdfsActive = src;
330 KernelStatistics(kd, ld, cd, iter);
333 // --------------------------------------------------------------------
337 X_LIKWID_START("aa-odd");
342 #pragma omp parallel for default(none) \
343 shared(gDims,src, w_0, w_1, w_2, omegaEven, omegaOdd, \
344 w_1_x3, w_2_x3, w_1_nine_half, w_2_nine_half, cd, \
345 oX, oY, oZ, nX, nY, nZ, blk, nThreads) \
346 private(ux, uy, uz, ui, dens, dir_indep_trm, \
348 pdf_N, pdf_E, pdf_S, pdf_W, \
349 pdf_NE, pdf_SE, pdf_SW, pdf_NW, \
350 pdf_T, pdf_TN, pdf_TE, pdf_TS, pdf_TW, \
351 pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW, \
352 evenPart, oddPart, w_1_indep, w_2_indep)
355 for (int i = 0; i < nThreads; ++i) {
357 int threadStartX = nX / nThreads * i;
358 int threadEndX = nX / nThreads * (i + 1);
360 if (nX % nThreads > 0) {
361 if (nX % nThreads > i) {
366 threadStartX += nX % nThreads;
367 threadEndX += nX % nThreads;
371 for (int bX = oX + threadStartX; bX < threadEndX + oX; bX += blk[0]) {
372 for (int bY = oY; bY < nY + oY; bY += blk[1]) {
373 for (int bZ = oZ; bZ < nZ + oZ; bZ += blk[2]) {
375 // Must do everything here, else it would break collapse.
376 int eZ = MIN(bZ + blk[2], nZ + oZ);
377 int eY = MIN(bY + blk[1], nY + oY);
378 int eX = MIN(bX + blk[0], threadEndX + oX);
380 for (int x = bX; x < eX; ++x) {
381 for (int y = bY; y < eY; ++y) {
382 #ifdef INTEL_OPT_DIRECTIVES
384 #pragma vector always
387 for (int z = bZ; z < eZ; ++z) { // LOOP aa-odd
389 #define I(x, y, z, dir) P_INDEX_5(gDims, (x), (y), (z), (dir))
391 // Load PDFs of local cell: pdf_N = src[I(x, y, z, D3Q19_N)]; ...
392 #define X(name, idx, idxinv, _x, _y, _z) JOIN(pdf_,name) = src[I(x - _x, y - _y, z - _z, idxinv)];
397 // #define LID_DRIVEN_CAVITY
399 #ifdef LID_DRIVEN_CAVITY
401 if (z == nZ - 4 + oZ && x > 3 + oX && x < (nX - 4 + oX) && y > 3 + oY && y < (nY - 4 + oY)) {
402 ux = F(0.1) * F(0.5)77;
408 ux = pdf_E + pdf_NE + pdf_SE + pdf_TE + pdf_BE -
409 pdf_W - pdf_NW - pdf_SW - pdf_TW - pdf_BW;
410 uy = pdf_N + pdf_NE + pdf_NW + pdf_TN + pdf_BN -
411 pdf_S - pdf_SE - pdf_SW - pdf_TS - pdf_BS;
412 uz = pdf_T + pdf_TE + pdf_TW + pdf_TN + pdf_TS -
413 pdf_B - pdf_BE - pdf_BW - pdf_BN - pdf_BS;
414 #ifdef LID_DRIVEN_CAVITY
419 pdf_N + pdf_E + pdf_S + pdf_W +
420 pdf_NE + pdf_SE + pdf_SW + pdf_NW +
421 pdf_T + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
422 pdf_B + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
424 dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz)*F(3.0)/F(2.0);
427 src[I(x, y, z, D3Q19_C)] = pdf_C - omegaEven*(pdf_C - w_0*dir_indep_trm);
430 w_1_indep = w_1*dir_indep_trm;
433 evenPart = omegaEven*( F(0.5)*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
434 oddPart = omegaOdd*(F(0.5)*(pdf_N - pdf_S) - ui*w_1_x3 );
435 src[I(x, y + 1, z, D3Q19_N)] = pdf_N - evenPart - oddPart;
436 src[I(x, y - 1, z, D3Q19_S)] = pdf_S - evenPart + oddPart;
439 evenPart = omegaEven*( F(0.5)*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
440 oddPart = omegaOdd*(F(0.5)*(pdf_E - pdf_W) - ui*w_1_x3 );
441 src[I(x + 1, y, z, D3Q19_E)] = pdf_E - evenPart - oddPart;
442 src[I(x - 1, y, z, D3Q19_W)] = pdf_W - evenPart + oddPart;
445 evenPart = omegaEven*( F(0.5)*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
446 oddPart = omegaOdd*(F(0.5)*(pdf_T - pdf_B) - ui*w_1_x3 );
447 src[I(x, y, z + 1, D3Q19_T)] = pdf_T - evenPart - oddPart;
448 src[I(x, y, z - 1, D3Q19_B)] = pdf_B - evenPart + oddPart;
451 w_2_indep = w_2*dir_indep_trm;
454 evenPart = omegaEven*( F(0.5)*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
455 oddPart = omegaOdd*(F(0.5)*(pdf_NW - pdf_SE) - ui*w_2_x3 );
456 src[I(x - 1, y + 1, z, D3Q19_NW)] = pdf_NW - evenPart - oddPart;
457 src[I(x + 1, y - 1, z, D3Q19_SE)] = pdf_SE - evenPart + oddPart;
460 evenPart = omegaEven*( F(0.5)*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
461 oddPart = omegaOdd*(F(0.5)*(pdf_NE - pdf_SW) - ui*w_2_x3 );
462 src[I(x + 1, y + 1, z, D3Q19_NE)] = pdf_NE - evenPart - oddPart;
463 src[I(x - 1, y - 1, z, D3Q19_SW)] = pdf_SW - evenPart + oddPart;
466 evenPart = omegaEven*( F(0.5)*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
467 oddPart = omegaOdd*(F(0.5)*(pdf_TW - pdf_BE) - ui*w_2_x3 );
468 src[I(x - 1, y, z + 1, D3Q19_TW)] = pdf_TW - evenPart - oddPart;
469 src[I(x + 1, y, z - 1, D3Q19_BE)] = pdf_BE - evenPart + oddPart;
472 evenPart = omegaEven*( F(0.5)*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
473 oddPart = omegaOdd*(F(0.5)*(pdf_TE - pdf_BW) - ui*w_2_x3 );
474 src[I(x + 1, y, z + 1, D3Q19_TE)] = pdf_TE - evenPart - oddPart;
475 src[I(x - 1, y, z - 1, D3Q19_BW)] = pdf_BW - evenPart + oddPart;
478 evenPart = omegaEven*( F(0.5)*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
479 oddPart = omegaOdd*(F(0.5)*(pdf_TS - pdf_BN) - ui*w_2_x3 );
480 src[I(x, y - 1, z + 1, D3Q19_TS)] = pdf_TS - evenPart - oddPart;
481 src[I(x, y + 1, z - 1, D3Q19_BN)] = pdf_BN - evenPart + oddPart;
484 evenPart = omegaEven*( F(0.5)*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
485 oddPart = omegaOdd*(F(0.5)*(pdf_TN - pdf_BS) - ui*w_2_x3 );
486 src[I(x, y + 1, z + 1, D3Q19_TN)] = pdf_TN - evenPart - oddPart;
487 src[I(x, y - 1, z - 1, D3Q19_BS)] = pdf_BS - evenPart + oddPart;
491 } } } // z, y, x (from inner to outer)
492 } } } // z, y, x (from inner to outer)
493 } // loop over threads
497 // Stop counters before bounce back. Else computing loop balance will be incorrect.
499 X_LIKWID_STOP("aa-odd");
501 // Fixup bounce back PDFs.
503 #pragma omp parallel for default(none) \
506 #ifdef INTEL_OPT_DIRECTIVES
509 for (int i = 0; i < kd->nBounceBackPdfs; ++i) {
510 src[kd->BounceBackPdfsDst[i]] = src[kd->BounceBackPdfsSrc[i]];
513 // save current iteration
514 kda->Iteration = iter + 1;
517 kd->PdfsActive = src;
518 KernelAddBodyForce(kd, ld, cd);
522 if (cd->VtkOutput && (iter + 1 % cd->VtkModulus) == 0) {
523 kd->PdfsActive = src;
524 VtkWrite(ld, kd, cd, iter + 1);
529 kd->PdfsActive = src;
530 KernelStatistics(kd, ld, cd, iter + 1);
534 } // for (int iter = 0; ...
536 X_KERNEL_END(kernelData);
541 kd->PdfsActive = src;
542 VtkWrite(ld, kd, cd, maxIterations);