[LbmBenchmarkKernelsPublic.git] / src / BenchKernelD3Q19List.c

// --------------------------------------------------------------------------
//
// Copyright
//   Markus Wittmann, 2016-2017
//   RRZE, University of Erlangen-Nuremberg, Germany
//   markus.wittmann -at- fau.de or hpc -at- rrze.fau.de
//
//   Viktor Haag, 2016
//   LSS, University of Erlangen-Nuremberg, Germany
//
//  This file is part of the Lattice Boltzmann Benchmark Kernels (LbmBenchKernels).
//
//  LbmBenchKernels is free software: you can redistribute it and/or modify
//  it under the terms of the GNU General Public License as published by
//  the Free Software Foundation, either version 3 of the License, or
//  (at your option) any later version.
//
//  LbmBenchKernels is distributed in the hope that it will be useful,
//  but WITHOUT ANY WARRANTY; without even the implied warranty of
//  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
//  GNU General Public License for more details.
//
//  You should have received a copy of the GNU General Public License
//  along with LbmBenchKernels.  If not, see <http://www.gnu.org/licenses/>.
//
// --------------------------------------------------------------------------
#include "BenchKernelD3Q19ListCommon.h"

#include "Memory.h"
#include "Vtk.h"
#include "LikwidIf.h"

#include <inttypes.h>
#include <math.h>


void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData * cd)
{
	Assert(ld != NULL);
	Assert(kernelData != NULL);
	Assert(cd != NULL);

	Assert(cd->Omega > 0.0);
	Assert(cd->Omega < 2.0);

	KernelData * kd = (KernelData *)kernelData;
	KernelDataList * kdl = (KernelDataList *)kernelData;

	PdfT omega = cd->Omega;
	PdfT omegaEven = omega;
// 	PdfT omegaOdd = 8.0*((2.0-omegaEven)/(8.0-omegaEven)); //"standard" trt odd relaxation parameter
	PdfT magicParam = 1.0/12.0;	// 1/4: best stability;	1/12: removes third-order advection error (best advection);	1/6: removes fourth-order diffusion error (best diffusion);	3/16: exact location of bounce back for poiseuille flow
	PdfT omegaOdd = 1.0/( 0.5 + magicParam/(1.0/omega - 0.5) );

	PdfT evenPart = 0.0;
	PdfT oddPart = 0.0;
	PdfT dir_indep_trm = 0.0;

	PdfT w_0 = 1.0 /  3.0;
	PdfT w_1 = 1.0 / 18.0;
	PdfT w_2 = 1.0 / 36.0;

	PdfT w_1_x3 = w_1 * 3.0;	PdfT w_1_nine_half = w_1 * 9.0/2.0;	PdfT w_1_indep = 0.0;
	PdfT w_2_x3 = w_2 * 3.0;	PdfT w_2_nine_half = w_2 * 9.0/2.0;	PdfT w_2_indep = 0.0;

	PdfT ux, uy, uz, ui;
	PdfT dens;

	// Declare pdf_N, pdf_E, pdf_S, pdf_W, ...
	#define X(name, idx, idxinv, x, y, z)	PdfT JOIN(pdf_,name);
	D3Q19_LIST
	#undef X

	PdfT * src = kd->Pdfs[0];
	PdfT * dst = kd->Pdfs[1];
	PdfT * tmp;

	int maxIterations = cd->MaxIterations;
	int nFluid = kdl->nFluid;
	int nCells = kdl->nCells;

	uint32_t adjListIndex;
	uint32_t * adjList = kdl->AdjList;

	#ifdef VTK_OUTPUT
		if (cd->VtkOutput) {
			kd->PdfsActive = src;
			VtkWrite(ld, kd, cd, 0);
		}
	#endif

	#ifdef STATISTICS
		kd->PdfsActive = src;
		KernelStatistics(kd, ld, cd, 0);
	#endif

	// TODO: outer openmp parallel
	for(int iter = 0; iter < maxIterations; ++iter) {


	X_LIKWID_START("list-os");

	#ifdef _OPENMP
		#pragma omp parallel for default(none) \
				shared(nFluid, nCells, kd, kdl, adjList, src, dst, w_0, w_1, w_2, omegaEven, omegaOdd, \
				w_1_x3, w_2_x3, w_1_nine_half, w_2_nine_half, cd) \
				private(ux, uy, uz, ui, dens, dir_indep_trm, adjListIndex, \
					pdf_C, \
				  	pdf_N, pdf_E, pdf_S, pdf_W, \
				  	pdf_NE, pdf_SE, pdf_SW, pdf_NW, \
				  	pdf_T, pdf_TN, pdf_TE, pdf_TS, pdf_TW, \
				  	pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW, \
				  	evenPart, oddPart, w_1_indep, w_2_indep)
	#endif
	for (int index = 0; index < nFluid; ++index) {

			#define I(index, dir)	P_INDEX_3((nCells), (index), (dir))

#ifdef PROP_MODEL_PUSH

			// Load PDFs of local cell: pdf_N = src[I(x, y, z, D3Q19_N)]; ...
			#define X(name, idx, idxinv, _x, _y, _z)	JOIN(pdf_,name) = src[I(index, idx)];
			D3Q19_LIST
			#undef X

#elif PROP_MODEL_PULL

			adjListIndex = index * N_D3Q19_IDX;

			pdf_C = src[P_INDEX_3(nCells, index, D3Q19_C)];

			// Load PDFs of local cell: pdf_N = src[adjList[adjListIndex + D3Q19_N]]; ...
			#define X(name, idx, idxinv, _x, _y, _z)	JOIN(pdf_,name) = src[adjList[adjListIndex + idx]];
			D3Q19_LIST_WO_C
			#undef X

#else
	#error No implementation for PROP_MODEL_NAME.
#endif

// #define LID_DRIVEN_CAVITY

#ifdef LID_DRIVEN_CAVITY
			int nX = kd->Dims[0];
			int nY = kd->Dims[1];
			int nZ = kd->Dims[2];

			int x = kdl->Coords[C_INDEX_X(index)];
			int y = kdl->Coords[C_INDEX_Y(index)];
			int z = kdl->Coords[C_INDEX_Z(index)];

			if (z == nZ - 4 && x > 3 && x < (nX - 4) && y > 3 && y < (nY - 4)) {
				ux = 0.1 * 0.577;
				uy = 0.0;
				uz = 0.0;
			} else {
#endif
				ux = pdf_E + pdf_NE + pdf_SE + pdf_TE + pdf_BE -
					 pdf_W - pdf_NW - pdf_SW - pdf_TW - pdf_BW;
				uy = pdf_N + pdf_NE + pdf_NW + pdf_TN + pdf_BN -
					 pdf_S - pdf_SE - pdf_SW - pdf_TS - pdf_BS;
				uz = pdf_T + pdf_TE + pdf_TW + pdf_TN + pdf_TS -
					 pdf_B - pdf_BE - pdf_BW - pdf_BN - pdf_BS;
#ifdef LID_DRIVEN_CAVITY
			}
#endif

			dens = pdf_C +
				   pdf_N  + pdf_E  + pdf_S  + pdf_W  +
				   pdf_NE + pdf_SE + pdf_SW + pdf_NW +
				   pdf_T  + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
				   pdf_B  + pdf_BN + pdf_BE + pdf_BS + pdf_BW;

			dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz)*3.0/2.0;

#ifdef PROP_MODEL_PUSH

			adjListIndex = index * N_D3Q19_IDX;

			// direction: w_0
			dst[I(index, D3Q19_C)             ]  = pdf_C - omegaEven*(pdf_C - w_0*dir_indep_trm);

			// direction: w_1
			w_1_indep = w_1*dir_indep_trm;

			ui = uy;
			evenPart = omegaEven*( 0.5*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
			oddPart = omegaOdd*(0.5*(pdf_N - pdf_S) - ui*w_1_x3 );
			dst[adjList[adjListIndex + D3Q19_N]]  = pdf_N - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_S]]  = pdf_S - evenPart + oddPart;

			ui = ux;
			evenPart = omegaEven*( 0.5*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
			oddPart = omegaOdd*(0.5*(pdf_E - pdf_W) - ui*w_1_x3 );
			dst[adjList[adjListIndex + D3Q19_E]]  = pdf_E - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_W]]  = pdf_W - evenPart + oddPart;

			ui = uz;
			evenPart = omegaEven*( 0.5*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
			oddPart = omegaOdd*(0.5*(pdf_T - pdf_B) - ui*w_1_x3 );
			dst[adjList[adjListIndex + D3Q19_T]]  = pdf_T - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_B]]  = pdf_B - evenPart + oddPart;

			// direction: w_2
			w_2_indep = w_2*dir_indep_trm;

			ui = -ux + uy;
			evenPart = omegaEven*( 0.5*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_NW - pdf_SE) - ui*w_2_x3 );
			dst[adjList[adjListIndex + D3Q19_NW]] = pdf_NW - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_SE]] = pdf_SE - evenPart + oddPart;

			ui = ux + uy;
			evenPart = omegaEven*( 0.5*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_NE - pdf_SW) - ui*w_2_x3 );
			dst[adjList[adjListIndex + D3Q19_NE]] = pdf_NE - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_SW]] = pdf_SW - evenPart + oddPart;

			ui = -ux + uz;
			evenPart = omegaEven*( 0.5*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TW - pdf_BE) - ui*w_2_x3 );
			dst[adjList[adjListIndex + D3Q19_TW]] = pdf_TW - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_BE]] = pdf_BE - evenPart + oddPart;

			ui = ux + uz;
			evenPart = omegaEven*( 0.5*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TE - pdf_BW) - ui*w_2_x3 );
			dst[adjList[adjListIndex + D3Q19_TE]] = pdf_TE - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_BW]] = pdf_BW - evenPart + oddPart;

			ui = -uy + uz;
			evenPart = omegaEven*( 0.5*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TS - pdf_BN) - ui*w_2_x3 );
			dst[adjList[adjListIndex + D3Q19_TS]] = pdf_TS - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_BN]] = pdf_BN - evenPart + oddPart;

			ui = uy + uz;
			evenPart = omegaEven*( 0.5*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TN - pdf_BS) - ui*w_2_x3 );
			dst[adjList[adjListIndex + D3Q19_TN]] = pdf_TN - evenPart - oddPart;
			dst[adjList[adjListIndex + D3Q19_BS]] = pdf_BS - evenPart + oddPart;

#elif PROP_MODEL_PULL

			// direction: w_0
			dst[I(index, D3Q19_C )]  = pdf_C - omegaEven*(pdf_C - w_0*dir_indep_trm);

			// direction: w_1
			w_1_indep = w_1*dir_indep_trm;

			ui = uy;
			evenPart = omegaEven*( 0.5*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
			oddPart = omegaOdd*(0.5*(pdf_N - pdf_S) - ui*w_1_x3 );
			dst[I(index, D3Q19_N )]  = pdf_N - evenPart - oddPart;
			dst[I(index, D3Q19_S )]  = pdf_S - evenPart + oddPart;

			ui = ux;
			evenPart = omegaEven*( 0.5*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
			oddPart = omegaOdd*(0.5*(pdf_E - pdf_W) - ui*w_1_x3 );
			dst[I(index, D3Q19_E )]  = pdf_E - evenPart - oddPart;
			dst[I(index, D3Q19_W )]  = pdf_W - evenPart + oddPart;

			ui = uz;
			evenPart = omegaEven*( 0.5*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
			oddPart = omegaOdd*(0.5*(pdf_T - pdf_B) - ui*w_1_x3 );
			dst[I(index, D3Q19_T )]  = pdf_T - evenPart - oddPart;
			dst[I(index, D3Q19_B )]  = pdf_B - evenPart + oddPart;

			// direction: w_2
			w_2_indep = w_2*dir_indep_trm;

			ui = -ux + uy;
			evenPart = omegaEven*( 0.5*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_NW - pdf_SE) - ui*w_2_x3 );
			dst[I(index, D3Q19_NW)] = pdf_NW - evenPart - oddPart;
			dst[I(index, D3Q19_SE)] = pdf_SE - evenPart + oddPart;

			ui = ux + uy;
			evenPart = omegaEven*( 0.5*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_NE - pdf_SW) - ui*w_2_x3 );
			dst[I(index, D3Q19_NE)] = pdf_NE - evenPart - oddPart;
			dst[I(index, D3Q19_SW)] = pdf_SW - evenPart + oddPart;

			ui = -ux + uz;
			evenPart = omegaEven*( 0.5*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TW - pdf_BE) - ui*w_2_x3 );
			dst[I(index, D3Q19_TW)] = pdf_TW - evenPart - oddPart;
			dst[I(index, D3Q19_BE)] = pdf_BE - evenPart + oddPart;

			ui = ux + uz;
			evenPart = omegaEven*( 0.5*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TE - pdf_BW) - ui*w_2_x3 );
			dst[I(index, D3Q19_TE)] = pdf_TE - evenPart - oddPart;
			dst[I(index, D3Q19_BW)] = pdf_BW - evenPart + oddPart;

			ui = -uy + uz;
			evenPart = omegaEven*( 0.5*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TS - pdf_BN) - ui*w_2_x3 );
			dst[I(index, D3Q19_TS)] = pdf_TS - evenPart - oddPart;
			dst[I(index, D3Q19_BN)] = pdf_BN - evenPart + oddPart;

			ui = uy + uz;
			evenPart = omegaEven*( 0.5*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
			oddPart = omegaOdd*(0.5*(pdf_TN - pdf_BS) - ui*w_2_x3 );
			dst[I(index, D3Q19_TN)] = pdf_TN - evenPart - oddPart;
			dst[I(index, D3Q19_BS)] = pdf_BS - evenPart + oddPart;

#endif
			#undef I
		} // loop over fluid nodes

		X_LIKWID_STOP("list-os");

		#ifdef VERIFICATION
			kd->PdfsActive = dst;
			KernelAddBodyForce(kd, ld, cd);
		#endif

		#ifdef VTK_OUTPUT
			if (cd->VtkOutput && (iter % cd->VtkModulus) == 0) {
				kd->PdfsActive = dst;
				VtkWrite(ld, kd, cd, iter);
			}
		#endif

		#ifdef STATISTICS
			kd->PdfsActive = dst;
			KernelStatistics(kd, ld, cd, iter);
		#endif

		// swap grids
		tmp = src;
		src = dst;
		dst = tmp;

	} // for (int iter = 0; ...

#ifdef VTK_OUTPUT
	if (cd->VtkOutput) {
		kd->PdfsActive = src;
		VtkWrite(ld, kd, cd, maxIterations);
	}
#endif

#ifdef STATISTICS
	kd->PdfsActive = src;
	KernelStatistics(kd, ld, cd, maxIterations);
#endif

	return;
}
Commit	Line	Data
10988083 MW	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 "BenchKernelD3Q19ListCommon.h"
	28
	29	#include "Memory.h"
	30	#include "Vtk.h"
e3f82424	31	#include "LikwidIf.h"
10988083 MW	32
	33	#include <inttypes.h>
	34	#include <math.h>
	35
	36
	37	void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData * cd)
	38	{
	39	Assert(ld != NULL);
	40	Assert(kernelData != NULL);
	41	Assert(cd != NULL);
	42
	43	Assert(cd->Omega > 0.0);
	44	Assert(cd->Omega < 2.0);
	45
	46	KernelData * kd = (KernelData *)kernelData;
	47	KernelDataList * kdl = (KernelDataList *)kernelData;
	48
	49	PdfT omega = cd->Omega;
	50	PdfT omegaEven = omega;
	51	// PdfT omegaOdd = 8.0*((2.0-omegaEven)/(8.0-omegaEven)); //"standard" trt odd relaxation parameter
	52	PdfT magicParam = 1.0/12.0; // 1/4: best stability; 1/12: removes third-order advection error (best advection); 1/6: removes fourth-order diffusion error (best diffusion); 3/16: exact location of bounce back for poiseuille flow
	53	PdfT omegaOdd = 1.0/( 0.5 + magicParam/(1.0/omega - 0.5) );
	54
	55	PdfT evenPart = 0.0;
	56	PdfT oddPart = 0.0;
	57	PdfT dir_indep_trm = 0.0;
	58
	59	PdfT w_0 = 1.0 / 3.0;
	60	PdfT w_1 = 1.0 / 18.0;
	61	PdfT w_2 = 1.0 / 36.0;
	62
	63	PdfT w_1_x3 = w_1 * 3.0; PdfT w_1_nine_half = w_1 * 9.0/2.0; PdfT w_1_indep = 0.0;
	64	PdfT w_2_x3 = w_2 * 3.0; PdfT w_2_nine_half = w_2 * 9.0/2.0; PdfT w_2_indep = 0.0;
	65
	66	PdfT ux, uy, uz, ui;
	67	PdfT dens;
	68
	69	// Declare pdf_N, pdf_E, pdf_S, pdf_W, ...
	70	#define X(name, idx, idxinv, x, y, z) PdfT JOIN(pdf_,name);
	71	D3Q19_LIST
	72	#undef X
	73
	74	PdfT * src = kd->Pdfs[0];
	75	PdfT * dst = kd->Pdfs[1];
	76	PdfT * tmp;
	77
	78	int maxIterations = cd->MaxIterations;
	79	int nFluid = kdl->nFluid;
	80	int nCells = kdl->nCells;
	81
	82	uint32_t adjListIndex;
	83	uint32_t * adjList = kdl->AdjList;
	84
	85	#ifdef VTK_OUTPUT
	86	if (cd->VtkOutput) {
	87	kd->PdfsActive = src;
	88	VtkWrite(ld, kd, cd, 0);
	89	}
	90	#endif
	91
	92	#ifdef STATISTICS
	93	kd->PdfsActive = src;
	94	KernelStatistics(kd, ld, cd, 0);
	95	#endif
96
97	// TODO: outer openmp parallel
98	for(int iter = 0; iter < maxIterations; ++iter) {
99
100
e3f82424	101	X_LIKWID_START("list-os");
10988083 MW	102
	103	#ifdef _OPENMP
	104	#pragma omp parallel for default(none) \
	105	shared(nFluid, nCells, kd, kdl, adjList, src, dst, w_0, w_1, w_2, omegaEven, omegaOdd, \
	106	w_1_x3, w_2_x3, w_1_nine_half, w_2_nine_half, cd) \
	107	private(ux, uy, uz, ui, dens, dir_indep_trm, adjListIndex, \
	108	pdf_C, \
	109	pdf_N, pdf_E, pdf_S, pdf_W, \
	110	pdf_NE, pdf_SE, pdf_SW, pdf_NW, \
	111	pdf_T, pdf_TN, pdf_TE, pdf_TS, pdf_TW, \
	112	pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW, \
	113	evenPart, oddPart, w_1_indep, w_2_indep)
	114	#endif
	115	for (int index = 0; index < nFluid; ++index) {
	116
	117	#define I(index, dir) P_INDEX_3((nCells), (index), (dir))
	118
	119	#ifdef PROP_MODEL_PUSH
	120
	121	// Load PDFs of local cell: pdf_N = src[I(x, y, z, D3Q19_N)]; ...
	122	#define X(name, idx, idxinv, _x, _y, _z) JOIN(pdf_,name) = src[I(index, idx)];
	123	D3Q19_LIST
	124	#undef X
	125
	126	#elif PROP_MODEL_PULL
	127
	128	adjListIndex = index * N_D3Q19_IDX;
	129
	130	pdf_C = src[P_INDEX_3(nCells, index, D3Q19_C)];
	131
	132	// Load PDFs of local cell: pdf_N = src[adjList[adjListIndex + D3Q19_N]]; ...
	133	#define X(name, idx, idxinv, _x, _y, _z) JOIN(pdf_,name) = src[adjList[adjListIndex + idx]];
	134	D3Q19_LIST_WO_C
	135	#undef X
	136
	137	#else
	138	#error No implementation for PROP_MODEL_NAME.
	139	#endif
	140
	141	// #define LID_DRIVEN_CAVITY
	142
	143	#ifdef LID_DRIVEN_CAVITY
	144	int nX = kd->Dims[0];
	145	int nY = kd->Dims[1];
	146	int nZ = kd->Dims[2];
	147
	148	int x = kdl->Coords[C_INDEX_X(index)];
	149	int y = kdl->Coords[C_INDEX_Y(index)];
	150	int z = kdl->Coords[C_INDEX_Z(index)];
	151
	152	if (z == nZ - 4 && x > 3 && x < (nX - 4) && y > 3 && y < (nY - 4)) {
	153	ux = 0.1 * 0.577;
	154	uy = 0.0;
	155	uz = 0.0;
	156	} else {
	157	#endif
	158	ux = pdf_E + pdf_NE + pdf_SE + pdf_TE + pdf_BE -
	159	pdf_W - pdf_NW - pdf_SW - pdf_TW - pdf_BW;
	160	uy = pdf_N + pdf_NE + pdf_NW + pdf_TN + pdf_BN -
	161	pdf_S - pdf_SE - pdf_SW - pdf_TS - pdf_BS;
	162	uz = pdf_T + pdf_TE + pdf_TW + pdf_TN + pdf_TS -
	163	pdf_B - pdf_BE - pdf_BW - pdf_BN - pdf_BS;
	164	#ifdef LID_DRIVEN_CAVITY
	165	}
166	#endif
167
168	dens = pdf_C +
169	pdf_N + pdf_E + pdf_S + pdf_W +
170	pdf_NE + pdf_SE + pdf_SW + pdf_NW +
171	pdf_T + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
172	pdf_B + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
173
174	dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz)*3.0/2.0;
175
176	#ifdef PROP_MODEL_PUSH
177
178	adjListIndex = index * N_D3Q19_IDX;
179
180	// direction: w_0
181	dst[I(index, D3Q19_C) ] = pdf_C - omegaEven(pdf_C - w_0dir_indep_trm);
182
183	// direction: w_1
184	w_1_indep = w_1*dir_indep_trm;
185
186	ui = uy;
187	evenPart = omegaEven( 0.5(pdf_N + pdf_S) - uiuiw_1_nine_half - w_1_indep );
188	oddPart = omegaOdd(0.5(pdf_N - pdf_S) - ui*w_1_x3 );
189	dst[adjList[adjListIndex + D3Q19_N]] = pdf_N - evenPart - oddPart;
190	dst[adjList[adjListIndex + D3Q19_S]] = pdf_S - evenPart + oddPart;
191
192	ui = ux;
193	evenPart = omegaEven( 0.5(pdf_E + pdf_W) - uiuiw_1_nine_half - w_1_indep );
194	oddPart = omegaOdd(0.5(pdf_E - pdf_W) - ui*w_1_x3 );
195	dst[adjList[adjListIndex + D3Q19_E]] = pdf_E - evenPart - oddPart;
196	dst[adjList[adjListIndex + D3Q19_W]] = pdf_W - evenPart + oddPart;
197
198	ui = uz;
199	evenPart = omegaEven( 0.5(pdf_T + pdf_B) - uiuiw_1_nine_half - w_1_indep );
200	oddPart = omegaOdd(0.5(pdf_T - pdf_B) - ui*w_1_x3 );
201	dst[adjList[adjListIndex + D3Q19_T]] = pdf_T - evenPart - oddPart;
202	dst[adjList[adjListIndex + D3Q19_B]] = pdf_B - evenPart + oddPart;
203
204	// direction: w_2
205	w_2_indep = w_2*dir_indep_trm;
206
207	ui = -ux + uy;
208	evenPart = omegaEven( 0.5(pdf_NW + pdf_SE) - uiuiw_2_nine_half - w_2_indep );
209	oddPart = omegaOdd(0.5(pdf_NW - pdf_SE) - ui*w_2_x3 );
210	dst[adjList[adjListIndex + D3Q19_NW]] = pdf_NW - evenPart - oddPart;
211	dst[adjList[adjListIndex + D3Q19_SE]] = pdf_SE - evenPart + oddPart;
212
213	ui = ux + uy;
214	evenPart = omegaEven( 0.5(pdf_NE + pdf_SW) - uiuiw_2_nine_half - w_2_indep );
215	oddPart = omegaOdd(0.5(pdf_NE - pdf_SW) - ui*w_2_x3 );
216	dst[adjList[adjListIndex + D3Q19_NE]] = pdf_NE - evenPart - oddPart;
217	dst[adjList[adjListIndex + D3Q19_SW]] = pdf_SW - evenPart + oddPart;
218
219	ui = -ux + uz;
220	evenPart = omegaEven( 0.5(pdf_TW + pdf_BE) - uiuiw_2_nine_half - w_2_indep );
221	oddPart = omegaOdd(0.5(pdf_TW - pdf_BE) - ui*w_2_x3 );
222	dst[adjList[adjListIndex + D3Q19_TW]] = pdf_TW - evenPart - oddPart;
223	dst[adjList[adjListIndex + D3Q19_BE]] = pdf_BE - evenPart + oddPart;
224
225	ui = ux + uz;
226	evenPart = omegaEven( 0.5(pdf_TE + pdf_BW) - uiuiw_2_nine_half - w_2_indep );
227	oddPart = omegaOdd(0.5(pdf_TE - pdf_BW) - ui*w_2_x3 );
228	dst[adjList[adjListIndex + D3Q19_TE]] = pdf_TE - evenPart - oddPart;
229	dst[adjList[adjListIndex + D3Q19_BW]] = pdf_BW - evenPart + oddPart;
230
231	ui = -uy + uz;
232	evenPart = omegaEven( 0.5(pdf_TS + pdf_BN) - uiuiw_2_nine_half - w_2_indep );
233	oddPart = omegaOdd(0.5(pdf_TS - pdf_BN) - ui*w_2_x3 );
234	dst[adjList[adjListIndex + D3Q19_TS]] = pdf_TS - evenPart - oddPart;
235	dst[adjList[adjListIndex + D3Q19_BN]] = pdf_BN - evenPart + oddPart;
236
237	ui = uy + uz;
238	evenPart = omegaEven( 0.5(pdf_TN + pdf_BS) - uiuiw_2_nine_half - w_2_indep );
239	oddPart = omegaOdd(0.5(pdf_TN - pdf_BS) - ui*w_2_x3 );
240	dst[adjList[adjListIndex + D3Q19_TN]] = pdf_TN - evenPart - oddPart;
241	dst[adjList[adjListIndex + D3Q19_BS]] = pdf_BS - evenPart + oddPart;
242
243	#elif PROP_MODEL_PULL
244
245	// direction: w_0
246	dst[I(index, D3Q19_C )] = pdf_C - omegaEven(pdf_C - w_0dir_indep_trm);
247
248	// direction: w_1
249	w_1_indep = w_1*dir_indep_trm;
250
251	ui = uy;
252	evenPart = omegaEven( 0.5(pdf_N + pdf_S) - uiuiw_1_nine_half - w_1_indep );
253	oddPart = omegaOdd(0.5(pdf_N - pdf_S) - ui*w_1_x3 );
254	dst[I(index, D3Q19_N )] = pdf_N - evenPart - oddPart;
255	dst[I(index, D3Q19_S )] = pdf_S - evenPart + oddPart;
256
257	ui = ux;
258	evenPart = omegaEven( 0.5(pdf_E + pdf_W) - uiuiw_1_nine_half - w_1_indep );
259	oddPart = omegaOdd(0.5(pdf_E - pdf_W) - ui*w_1_x3 );
260	dst[I(index, D3Q19_E )] = pdf_E - evenPart - oddPart;
261	dst[I(index, D3Q19_W )] = pdf_W - evenPart + oddPart;
262
263	ui = uz;
264	evenPart = omegaEven( 0.5(pdf_T + pdf_B) - uiuiw_1_nine_half - w_1_indep );
265	oddPart = omegaOdd(0.5(pdf_T - pdf_B) - ui*w_1_x3 );
266	dst[I(index, D3Q19_T )] = pdf_T - evenPart - oddPart;
267	dst[I(index, D3Q19_B )] = pdf_B - evenPart + oddPart;
268
269	// direction: w_2
270	w_2_indep = w_2*dir_indep_trm;
271
272	ui = -ux + uy;
273	evenPart = omegaEven( 0.5(pdf_NW + pdf_SE) - uiuiw_2_nine_half - w_2_indep );
274	oddPart = omegaOdd(0.5(pdf_NW - pdf_SE) - ui*w_2_x3 );
275	dst[I(index, D3Q19_NW)] = pdf_NW - evenPart - oddPart;
276	dst[I(index, D3Q19_SE)] = pdf_SE - evenPart + oddPart;
277
278	ui = ux + uy;
279	evenPart = omegaEven( 0.5(pdf_NE + pdf_SW) - uiuiw_2_nine_half - w_2_indep );
280	oddPart = omegaOdd(0.5(pdf_NE - pdf_SW) - ui*w_2_x3 );
281	dst[I(index, D3Q19_NE)] = pdf_NE - evenPart - oddPart;
282	dst[I(index, D3Q19_SW)] = pdf_SW - evenPart + oddPart;
283
284	ui = -ux + uz;
285	evenPart = omegaEven( 0.5(pdf_TW + pdf_BE) - uiuiw_2_nine_half - w_2_indep );
286	oddPart = omegaOdd(0.5(pdf_TW - pdf_BE) - ui*w_2_x3 );
287	dst[I(index, D3Q19_TW)] = pdf_TW - evenPart - oddPart;
288	dst[I(index, D3Q19_BE)] = pdf_BE - evenPart + oddPart;
289
290	ui = ux + uz;
291	evenPart = omegaEven( 0.5(pdf_TE + pdf_BW) - uiuiw_2_nine_half - w_2_indep );
292	oddPart = omegaOdd(0.5(pdf_TE - pdf_BW) - ui*w_2_x3 );
293	dst[I(index, D3Q19_TE)] = pdf_TE - evenPart - oddPart;
294	dst[I(index, D3Q19_BW)] = pdf_BW - evenPart + oddPart;
295
296	ui = -uy + uz;
297	evenPart = omegaEven( 0.5(pdf_TS + pdf_BN) - uiuiw_2_nine_half - w_2_indep );
298	oddPart = omegaOdd(0.5(pdf_TS - pdf_BN) - ui*w_2_x3 );
299	dst[I(index, D3Q19_TS)] = pdf_TS - evenPart - oddPart;
300	dst[I(index, D3Q19_BN)] = pdf_BN - evenPart + oddPart;
301
302	ui = uy + uz;
303	evenPart = omegaEven( 0.5(pdf_TN + pdf_BS) - uiuiw_2_nine_half - w_2_indep );
304	oddPart = omegaOdd(0.5(pdf_TN - pdf_BS) - ui*w_2_x3 );
305	dst[I(index, D3Q19_TN)] = pdf_TN - evenPart - oddPart;
306	dst[I(index, D3Q19_BS)] = pdf_BS - evenPart + oddPart;
307
308	#endif
309	#undef I
310	} // loop over fluid nodes
311
e3f82424 MW	312	X_LIKWID_STOP("list-os");
e3f82424 MW	313
10988083 MW	314	#ifdef VERIFICATION
	315	kd->PdfsActive = dst;
	316	KernelAddBodyForce(kd, ld, cd);
	317	#endif
	318
	319	#ifdef VTK_OUTPUT
	320	if (cd->VtkOutput && (iter % cd->VtkModulus) == 0) {
	321	kd->PdfsActive = dst;
	322	VtkWrite(ld, kd, cd, iter);
	323	}
	324	#endif
	325
	326	#ifdef STATISTICS
	327	kd->PdfsActive = dst;
	328	KernelStatistics(kd, ld, cd, iter);
	329	#endif
	330
	331	// swap grids
	332	tmp = src;
	333	src = dst;
	334	dst = tmp;
	335
	336	} // for (int iter = 0; ...
	337
	338	#ifdef VTK_OUTPUT
	339	if (cd->VtkOutput) {
	340	kd->PdfsActive = src;
	341	VtkWrite(ld, kd, cd, maxIterations);
	342	}
	343	#endif
	344
	345	#ifdef STATISTICS
	346	kd->PdfsActive = src;
	347	KernelStatistics(kd, ld, cd, maxIterations);
	348	#endif
	349
	350	return;
	351	}