[LbmBenchmarkKernelsPublic.git] / src / BenchKernelD3Q19ListCommon.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 "Padding.h"

#include <math.h>


// Forward definition.
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, struct KernelData_ * kd, CaseData * cd);


// -----------------------------------------------------------------------
// Functions which are used as callback by the kernel to read or write
// PDFs and nodes.

static void FNAME(BCGetPdf)(KernelData * kd, int x, int y, int z, int dir, PdfT * pdf)
{
	Assert(kd != NULL);
	Assert(kd->PdfsActive != NULL);
	Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]);
	Assert(pdf != NULL);

	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);
	Assert(dir >= 0); Assert(dir < N_D3Q19);

#if 0
	*pdf = kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)];
#else
#ifdef PROP_MODEL_PUSH
	*pdf = kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)];
#elif PROP_MODEL_PULL


	// The relevant PDFs here are the ones, which will get streamed in later
	// during propagation. So we must return the *remote* PDFs.
	uint32_t nodeIndex = KDL(kd)->Grid[L_INDEX_4(kd->Dims, x, y, z)];

	if (dir != D3Q19_C) {

		uint32_t adjListIndex = nodeIndex * N_D3Q19_IDX;

		*pdf = kd->PdfsActive[KDL(kd)->AdjList[adjListIndex + dir]];
	}
	else {
		*pdf = kd->PdfsActive[P_INDEX_3(KDL(kd)->nCells, nodeIndex, dir)];

	}
#endif
#endif

	return;
}

static void FNAME(BCSetPdf)(KernelData * kd, int x, int y, int z, int dir, PdfT pdf)
{
	Assert(kd != NULL);
	Assert(kd->PdfsActive != NULL);
	Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]);
	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);
	Assert(dir >= 0); Assert(dir < N_D3Q19);

#if 0
	if (isnan(pdf)) {
		printf("ERROR: setting nan %d %d %d %d %s\n", x, y, z, dir, D3Q19_NAMES[dir]);
		DEBUG_BREAK_POINT();
		exit(1);
	}
#endif

#if 0
	kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)] = pdf;
#else
#ifdef PROP_MODEL_PUSH
	kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)] = pdf;
#elif PROP_MODEL_PULL

	// The relevant PDFs here are the ones, which will get streamed in later
	// during propagation. So we must set this *remote* PDFs.
	uint32_t nodeIndex = KDL(kd)->Grid[L_INDEX_4(kd->Dims, x, y, z)];

	if (dir != D3Q19_C) {

		uint32_t adjListIndex = nodeIndex * N_D3Q19_IDX;

		kd->PdfsActive[KDL(kd)->AdjList[adjListIndex + dir]] = pdf;
	}
	else {
		kd->PdfsActive[P_INDEX_3(KDL(kd)->nCells, nodeIndex, dir)] = pdf;

	}
#endif
#endif

	return;
}


static void GetNode(KernelData * kd, int x, int y, int z, PdfT * pdfs)
{
	Assert(kd != NULL);
	Assert(kd->PdfsActive != NULL);
	Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]);
	Assert(pdfs != NULL);
	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);

	PdfT sum = 0.0;

	// TODO: pull scheme?
	#define I(x, y, z, dir) P_INDEX_5(KDL(kd), (x), (y), (z), (dir))
	#define X(name, idx, idxinv, _x, _y, _z)	pdfs[idx] = kd->PdfsActive[I(x, y, z, idx)]; sum += pdfs[idx];
	D3Q19_LIST
	#undef X
	#undef I

	// if (sum < 0.0) {
	// 		printf("%d %d %d negative density \n", x, y, z);
	// 		exit(1);
	// }

#if 0
	for (int d = 0; d < 19; ++d) {
		if(isnan(pdfs[d]) || isinf(pdfs[d])) {
			printf("%d %d %d %d nan! get node\n", x, y, z, d);
						for (int d2 = 0; d2 < 19; ++d2) {
							printf("%d: %e\n", d2, pdfs[d2]);
						}
			exit(1);
		}
	}
#endif
	return;
}


static void SetNode(KernelData * kd, int x, int y, int z, PdfT * pdfs)
{
	Assert(kd != NULL);
	Assert(kd->PdfsActive != NULL);
	Assert(kd->PdfsActive == kd->Pdfs[0] || kd->PdfsActive == kd->Pdfs[1]);
	Assert(pdfs != NULL);

	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);

#if 0
	for (int d = 0; d < 19; ++d) {
		if(isnan(pdfs[d])) {
			printf("%d %d %d %d nan! get node\n", x, y, z, d);
						for (int d2 = 0; d2 < 19; ++d2) {
							printf("%d: %e\n", d2, pdfs[d2]);
						}
			exit(1);
		}
	}
#endif

	// TODO: pull scheme?
	#define I(x, y, z, dir) P_INDEX_5(KDL(kd), (x), (y), (z), (dir))
	#define X(name, idx, idxinv, _x, _y, _z)	kd->PdfsActive[I(x, y, z, idx)] = pdfs[idx];
	D3Q19_LIST
	#undef X
	#undef I

	return;
}

static void ParameterUsage()
{
	printf("Kernel parameters:\n");
	printf("  [-blk <n>] [-blk-[xyz] <n>]\n");
#ifdef DATA_LAYOUT_SOA
	printf("  [-pad auto|modulus_1+offset_1(,modulus_n+offset_n)*]\n");
#endif

	return;
}

static void ParseParameters(Parameters * params, int * blk, PadInfo ** padInfo)
{
	Assert(blk != NULL);

	blk[0] = 0; blk[1] = 0; blk[2] = 0;
	*padInfo = NULL;

	#define ARG_IS(param)                   (!strcmp(params->KernelArgs[i], param))
	#define NEXT_ARG_PRESENT() \
		do { \
			if (i + 1 >= params->nKernelArgs) { \
				printf("ERROR: argument %s requires a parameter.\n", params->KernelArgs[i]); \
				exit(1); \
			} \
		} while (0)


	for (int i = 0; i < params->nKernelArgs; ++i) {
		if (ARG_IS("-blk") || ARG_IS("--blk")) {
			NEXT_ARG_PRESENT();

			int tmp = strtol(params->KernelArgs[++i], NULL, 0);

			if (tmp < 0) {
				printf("ERROR: blocking parameter must be >= 0.\n");
				exit(1);
			}

			blk[0] = blk[1] = blk[2] = tmp;
		}
		else if (ARG_IS("-blk-x") || ARG_IS("--blk-x")) {
			NEXT_ARG_PRESENT();

			int tmp = strtol(params->KernelArgs[++i], NULL, 0);

			if (tmp < 0) {
				printf("ERROR: blocking parameter must be >= 0.\n");
				exit(1);
			}

			blk[0] = tmp;
		}
		else if (ARG_IS("-blk-y") || ARG_IS("--blk-y")) {
			NEXT_ARG_PRESENT();

			int tmp = strtol(params->KernelArgs[++i], NULL, 0);

			if (tmp < 0) {
				printf("ERROR: blocking parameter must be >= 0.\n");
				exit(1);
			}

			blk[1] = tmp;
		}
		else if (ARG_IS("-blk-z") || ARG_IS("--blk-z")) {
			NEXT_ARG_PRESENT();

			int tmp = strtol(params->KernelArgs[++i], NULL, 0);

			if (tmp < 0) {
				printf("ERROR: blocking parameter must be >= 0.\n");
				exit(1);
			}

			blk[2] = tmp;
		}
		else if (ARG_IS("-h") || ARG_IS("-help") || ARG_IS("--help")) {
			ParameterUsage();
			exit(1);
		}
#ifdef DATA_LAYOUT_SOA
		else if (ARG_IS("-pad") || ARG_IS("--pad")) {
			NEXT_ARG_PRESENT();

			*padInfo = PadInfoFromStr(params->KernelArgs[++i]);
		}
#endif
		else {
			printf("ERROR: unknown kernel parameter.\n");
			ParameterUsage();
			exit(1);
		}
	}

	#undef ARG_IS
	#undef NEXT_ARG_PRESENT

	return;
}

void FNAME(D3Q19ListInit)(LatticeDesc * ld, KernelData ** kernelData, Parameters * params)
{
	KernelData * kd;
	KernelDataList * kdl;
	MemAlloc((void **)&kdl, sizeof(KernelDataList));

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

#ifdef DEBUG
	kd->Pdfs[0] = NULL;
	kd->Pdfs[1] = NULL;
	kd->PdfsActive = NULL;
	kd->DstPdfs = NULL;
	kd->SrcPdfs = NULL;
	kd->Dims[0] = -1;
	kd->Dims[1] = -1;
	kd->Dims[2] = -1;
	kd->GlobalDims[0] = -1;
	kd->GlobalDims[1] = -1;
	kd->GlobalDims[2] = -1;
	kd->Offsets[0] = -1;
	kd->Offsets[1] = -1;
	kd->Offsets[2] = -1;

	kd->ObstIndices = NULL;
	kd->nObstIndices = -1;
	kd->BounceBackPdfsSrc = NULL;
	kd->BounceBackPdfsDst = NULL;
	kd->nBounceBackPdfs = -1;

	kdl->AdjList = NULL;
	kdl->Coords = NULL;
	kdl->Grid = NULL;
	kdl->nCells = -1;
	kdl->nFluid = -1;
#endif

	int blk[3] = { 0 };
	PadInfo * padInfo = NULL;

	ParseParameters(params, blk, &padInfo);


	// Ajust the dimensions according to padding, if used.
	kd->Dims[0] = kd->GlobalDims[0] = ld->Dims[0];
	kd->Dims[1] = kd->GlobalDims[1] = ld->Dims[1];
	kd->Dims[2] = kd->GlobalDims[2] = ld->Dims[2];

	int * lDims = ld->Dims;

	int lX = lDims[0];
	int lY = lDims[1];
	int lZ = lDims[2];

	int nTotalCells = lX * lY * lZ;
	int nCells = ld->nFluid;
	int nFluid = ld->nFluid;

#ifdef DATA_LAYOUT_SOA
	{
		nCells = PadCellsAndReport(nCells, sizeof(PdfT), &padInfo);
		PadInfoFree(padInfo); padInfo = NULL;
	}
#endif

	kdl->nCells = nCells;
	kdl->nFluid = nFluid;

	PdfT * pdfs[2];

	if (blk[0] == 0) blk[0] = lX;
	if (blk[1] == 0) blk[1] = lY;
	if (blk[2] == 0) blk[2] = lZ;

	printf("# blocking               x: %3d y: %3d z: %3d\n", blk[0], blk[1], blk[2]);

	printf("# allocating data for %d fluid LB nodes with padding (%lu bytes = %f MiB for both lattices)\n",
		nCells, 2 * sizeof(PdfT) * nCells * N_D3Q19,
		2 * sizeof(PdfT) * nCells * N_D3Q19 / 1024.0 / 1024.0);

	MemAlloc((void **)&pdfs[0], sizeof(PdfT) * nCells * N_D3Q19);
	MemAlloc((void **)&pdfs[1], sizeof(PdfT) * nCells * N_D3Q19);

	kd->Pdfs[0] = pdfs[0];
	kd->Pdfs[1] = pdfs[1];

	// Initialize PDFs with some (arbitrary) data for correct NUMA placement.
	// Here we touch only the fluid nodes as this loop is OpenMP parallel and
	// we want the same scheduling as in the kernel.
	#ifdef _OPENMP
		#pragma omp parallel for
	#endif
	for (int i = 0; i < nFluid; ++i) { for(int d = 0; d < N_D3Q19; ++d) {
		pdfs[0][P_INDEX_3(nCells, i, d)] = 1.0;
		pdfs[1][P_INDEX_3(nCells, i, d)] = 1.0;
	} }

	// Initialize all PDFs to some standard value.
	for (int i = 0; i < nFluid; ++i) { for(int d = 0; d < N_D3Q19; ++d) {
		pdfs[0][P_INDEX_3(nCells, i, d)] = 0.0;
		pdfs[1][P_INDEX_3(nCells, i, d)] = 0.0;
	} }

	// ----------------------------------------------------------------------
	// create grid which will hold the index numbers of the fluid nodes

	uint32_t * grid;

	if (MemAlloc((void **)&grid, nTotalCells * sizeof(uint32_t))) {
		printf("ERROR: allocating grid for numbering failed: %lu bytes.\n", nTotalCells * sizeof(uint32_t));
		exit(1);
	}
	kdl->Grid = grid;

	int latticeIndex;

#ifdef DEBUG
	for(int z = 0; z < lZ; ++z) {
		for(int y = 0; y < lY; ++y) {
			for(int x = 0; x < lX; ++x) {

				latticeIndex = L_INDEX_4(ld->Dims, x, y, z);

				grid[latticeIndex] = ~0;
			}
		}
	}
#endif

	// ----------------------------------------------------------------------
	// generate numbering over grid

	uint32_t * coords;

	if (MemAlloc((void **)&coords, nFluid * sizeof(uint32_t) * 3)) {
		printf("ERROR: allocating coords array failed: %lu bytes.\n", nFluid * sizeof(uint32_t) * 3);
		exit(1);
	}

	kdl->Coords = coords;

	// Index for the PDF nodes can start at 0 as we distinguish solid and fluid nodes
	// through the ld->Lattice array.
	int counter = 0;

	// Blocking is implemented via setup of the adjacency list. The kernel later will
	// walk through the lattice blocked automatically.
	for (int bX = 0; bX < lX; bX += blk[0]) {
	for (int bY = 0; bY < lY; bY += blk[1]) {
	for (int bZ = 0; bZ < lZ; bZ += blk[2]) {

		int eX = MIN(bX + blk[0], lX);
		int eY = MIN(bY + blk[1], lY);
		int eZ = MIN(bZ + blk[2], lZ);

		for (int x = bX; x < eX; ++x) {
		for (int y = bY; y < eY; ++y) {
		for (int z = bZ; z < eZ; ++z) {

			latticeIndex = L_INDEX_4(lDims, x, y, z);

			if (ld->Lattice[latticeIndex] != LAT_CELL_OBSTACLE) {
				grid[latticeIndex] = counter;

				coords[C_INDEX_X(counter)] = x;
				coords[C_INDEX_Y(counter)] = y;
				coords[C_INDEX_Z(counter)] = z;

				++counter;
			}
		} } }
	} } }

	Verify(counter == nFluid);

	uint32_t * adjList;

	// AdjList only requires 18 instead of 19 entries per node, as
	// the center PDF needs no addressing.
	if (MemAlloc((void **)&adjList, nFluid * sizeof(uint32_t) * N_D3Q19_IDX)) {
		printf("ERROR: allocating adjList array failed: %lu bytes.\n", nFluid * sizeof(uint32_t) * N_D3Q19_IDX);
		exit(1);
	}

	kdl->AdjList = adjList;

	int x, y, z;

	uint32_t neighborIndex;
	uint32_t dstIndex;

	int nx, ny, nz, px, py, pz;

	// Loop over all fluid nodes and compute the indices to the neighboring
	// PDFs for configured data layout (AoS/SoA).
	#ifdef _OPENMP
		#pragma omp parallel for
	#endif
	for (int index = 0; index < nFluid; ++index) {
		for (int d = 0; d < N_D3Q19_IDX; ++d) {
			adjList[index * N_D3Q19_IDX + d] = -1;
		}
	}

	// #ifdef _OPENMP  --> add line continuation
	// 	#pragma omp parallel for default(none)
	// 		shared(nFluid, nCells, coords, D3Q19_INV, D3Q19_X, D3Q19_Y, D3Q19_Z,
	// 				stderr,
	// 				lDims, grid, ld, lX, lY, lZ, adjList)
	// 		private(x, y, z, nx, ny, nz, neighborIndex, dstIndex)
	// #endif
	for (int index = 0; index < nFluid; ++index) {
		x = coords[C_INDEX_X(index)];
		y = coords[C_INDEX_Y(index)];
		z = coords[C_INDEX_Z(index)];

		Assert(x >= 0 && x < lX);
		Assert(y >= 0 && y < lY);
		Assert(z >= 0 && z < lZ);

		Assert(ld->Lattice[L_INDEX_4(lDims, x, y, z)] != LAT_CELL_OBSTACLE);

		// Loop over all directions except the center one.
		for(int d = 0; d < N_D3Q19 - 1; ++d) {
			Assert(d != D3Q19_C);
#ifdef PROP_MODEL_PUSH
			nx = x + D3Q19_X[d];
			ny = y + D3Q19_Y[d];
			nz = z + D3Q19_Z[d];
#elif PROP_MODEL_PULL
			nx = x - D3Q19_X[d];
			ny = y - D3Q19_Y[d];
			nz = z - D3Q19_Z[d];
#else
			#error No implementation for this PROP_MODEL_NAME.
#endif
			// If the neighbor is outside the latcie in X direction and we have a
			// periodic boundary then we need to wrap around.
			if ( 	((nx < 0 || nx >= lX) && ld->PeriodicX) ||
					((ny < 0 || ny >= lY) && ld->PeriodicY) ||
					((nz < 0 || nz >= lZ) && ld->PeriodicZ)
																){
				// x periodic

				if (nx < 0) {
					px = lX - 1;
				}
				else if (nx >= lX) {
					px = 0;
				} else {
					px = nx;
				}
				// y periodic
				if (ny < 0) {
					py = lY - 1;
				}
				else if (ny >= lY) {
					py = 0;
				} else {
					py = ny;
				}

				// z periodic
				if (nz < 0) {
					pz = lZ - 1;
				}
				else if (nz >= lZ) {
					pz = 0;
				} else {
					pz = nz;
				}

				if (ld->Lattice[L_INDEX_4(lDims, px, py, pz)] == LAT_CELL_OBSTACLE) {
					dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]);
				}
				else {
					neighborIndex = grid[L_INDEX_4(lDims, px, py, pz)];

					AssertMsg(neighborIndex != ~0, "Neighbor has no Index. (%d %d %d) direction %s (%d)\n", px, py, pz, D3Q19_NAMES[d], d);

					dstIndex = P_INDEX_3(nCells, neighborIndex, d);
				}
			}
			else if (nx < 0 || ny < 0 || nz < 0 || nx >= lX || ny >= lY || nz >= lZ) {
				dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]);
			}
			else if (ld->Lattice[L_INDEX_4(lDims, nx, ny, nz)] == LAT_CELL_OBSTACLE) {
				dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]);
			}
			else {
				neighborIndex = grid[L_INDEX_4(lDims, nx, ny, nz)];

				Assert(neighborIndex != ~0);

				dstIndex = P_INDEX_3(nCells, neighborIndex, d);
			}

			Assert(dstIndex >= 0);
			Assert(dstIndex < nCells * N_D3Q19);

			adjList[index * N_D3Q19_IDX + d] = dstIndex;
		}
	}


	// Fill remaining KernelData structures
	kd->GetNode = GetNode;
	kd->SetNode = SetNode;

	kd->BoundaryConditionsGetPdf = FNAME(BCGetPdf);
	kd->BoundaryConditionsSetPdf = FNAME(BCSetPdf);

	kd->Kernel = FNAME(D3Q19ListKernel);

	kd->DstPdfs = NULL;
	kd->PdfsActive = kd->Pdfs[0];

	return;
}

void FNAME(D3Q19ListDeinit)(LatticeDesc * ld, KernelData ** kernelData)
{
	KernelDataList ** kdl = (KernelDataList **)kernelData;

	MemFree((void **)&((*kernelData)->Pdfs[0]));
	MemFree((void **)&((*kernelData)->Pdfs[1]));

	MemFree((void **)&((*kdl)->AdjList));
	MemFree((void **)&((*kdl)->Coords));
	MemFree((void **)&((*kdl)->Grid));

	MemFree((void **)kernelData);

	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 "Padding.h"
10988083 MW	32
	33	#include <math.h>
	34
	35
	36	// Forward definition.
	37	void FNAME(D3Q19ListKernel)(LatticeDesc * ld, struct KernelData_ * kd, CaseData * cd);
	38
	39
	40
	41
	42	// -----------------------------------------------------------------------
	43	// Functions which are used as callback by the kernel to read or write
	44	// PDFs and nodes.
	45
	46	static void FNAME(BCGetPdf)(KernelData * kd, int x, int y, int z, int dir, PdfT * pdf)
	47	{
	48	Assert(kd != NULL);
	49	Assert(kd->PdfsActive != NULL);
	50	Assert(kd->PdfsActive == kd->Pdfs[0] \|\| kd->PdfsActive == kd->Pdfs[1]);
	51	Assert(pdf != NULL);
	52
	53	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
	54	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);
	55	Assert(dir >= 0); Assert(dir < N_D3Q19);
	56
	57	#if 0
	58	*pdf = kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)];
	59	#else
	60	#ifdef PROP_MODEL_PUSH
	61	*pdf = kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)];
	62	#elif PROP_MODEL_PULL
	63
	64
	65	// The relevant PDFs here are the ones, which will get streamed in later
	66	// during propagation. So we must return the remote PDFs.
	67	uint32_t nodeIndex = KDL(kd)->Grid[L_INDEX_4(kd->Dims, x, y, z)];
	68
	69	if (dir != D3Q19_C) {
	70
	71	uint32_t adjListIndex = nodeIndex * N_D3Q19_IDX;
	72
	73	*pdf = kd->PdfsActive[KDL(kd)->AdjList[adjListIndex + dir]];
	74	}
	75	else {
	76	*pdf = kd->PdfsActive[P_INDEX_3(KDL(kd)->nCells, nodeIndex, dir)];
	77
	78	}
	79	#endif
	80	#endif
	81
	82	return;
	83	}
	84
	85	static void FNAME(BCSetPdf)(KernelData * kd, int x, int y, int z, int dir, PdfT pdf)
	86	{
	87	Assert(kd != NULL);
	88	Assert(kd->PdfsActive != NULL);
	89	Assert(kd->PdfsActive == kd->Pdfs[0] \|\| kd->PdfsActive == kd->Pdfs[1]);
	90	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
	91	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);
	92	Assert(dir >= 0); Assert(dir < N_D3Q19);
	93
	94	#if 0
	95	if (isnan(pdf)) {
96	printf("ERROR: setting nan %d %d %d %d %s\n", x, y, z, dir, D3Q19_NAMES[dir]);
97	DEBUG_BREAK_POINT();
98	exit(1);
99	}
100	#endif
101
102	#if 0
103	kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)] = pdf;
104	#else
105	#ifdef PROP_MODEL_PUSH
106	kd->PdfsActive[P_INDEX_5(KDL(kd), x, y, z, dir)] = pdf;
107	#elif PROP_MODEL_PULL
108
109	// The relevant PDFs here are the ones, which will get streamed in later
110	// during propagation. So we must set this remote PDFs.
111	uint32_t nodeIndex = KDL(kd)->Grid[L_INDEX_4(kd->Dims, x, y, z)];
112
113	if (dir != D3Q19_C) {
114
115	uint32_t adjListIndex = nodeIndex * N_D3Q19_IDX;
116
117	kd->PdfsActive[KDL(kd)->AdjList[adjListIndex + dir]] = pdf;
118	}
119	else {
120	kd->PdfsActive[P_INDEX_3(KDL(kd)->nCells, nodeIndex, dir)] = pdf;
121
122	}
123	#endif
124	#endif
125
126	return;
127	}
128
129
130	static void GetNode(KernelData * kd, int x, int y, int z, PdfT * pdfs)
131	{
132	Assert(kd != NULL);
133	Assert(kd->PdfsActive != NULL);
134	Assert(kd->PdfsActive == kd->Pdfs[0] \|\| kd->PdfsActive == kd->Pdfs[1]);
135	Assert(pdfs != NULL);
136	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
137	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);
138
139	PdfT sum = 0.0;
140
141	// TODO: pull scheme?
142	#define I(x, y, z, dir) P_INDEX_5(KDL(kd), (x), (y), (z), (dir))
143	#define X(name, idx, idxinv, _x, _y, _z) pdfs[idx] = kd->PdfsActive[I(x, y, z, idx)]; sum += pdfs[idx];
144	D3Q19_LIST
145	#undef X
146	#undef I
147
148	// if (sum < 0.0) {
149	// printf("%d %d %d negative density \n", x, y, z);
150	// exit(1);
151	// }
152
153	#if 0
154	for (int d = 0; d < 19; ++d) {
155	if(isnan(pdfs[d]) \|\| isinf(pdfs[d])) {
156	printf("%d %d %d %d nan! get node\n", x, y, z, d);
157	for (int d2 = 0; d2 < 19; ++d2) {
158	printf("%d: %e\n", d2, pdfs[d2]);
159	}
160	exit(1);
161	}
162	}
163	#endif
164	return;
165	}
166
167
168	static void SetNode(KernelData * kd, int x, int y, int z, PdfT * pdfs)
169	{
170	Assert(kd != NULL);
171	Assert(kd->PdfsActive != NULL);
172	Assert(kd->PdfsActive == kd->Pdfs[0] \|\| kd->PdfsActive == kd->Pdfs[1]);
173	Assert(pdfs != NULL);
174
175	Assert(x >= 0); Assert(y >= 0); Assert(z >= 0);
176	Assert(x < kd->Dims[0]); Assert(y < kd->Dims[1]); Assert(z < kd->Dims[2]);
177
178	#if 0
179	for (int d = 0; d < 19; ++d) {
180	if(isnan(pdfs[d])) {
181	printf("%d %d %d %d nan! get node\n", x, y, z, d);
182	for (int d2 = 0; d2 < 19; ++d2) {
183	printf("%d: %e\n", d2, pdfs[d2]);
184	}
185	exit(1);
186	}
187	}
188	#endif
189
190	// TODO: pull scheme?
191	#define I(x, y, z, dir) P_INDEX_5(KDL(kd), (x), (y), (z), (dir))
192	#define X(name, idx, idxinv, _x, _y, _z) kd->PdfsActive[I(x, y, z, idx)] = pdfs[idx];
193	D3Q19_LIST
194	#undef X
195	#undef I
196
197	return;
198	}
199
200	static void ParameterUsage()
201	{
202	printf("Kernel parameters:\n");
203	printf(" [-blk <n>] [-blk-[xyz] <n>]\n");
e3f82424 MW	204	#ifdef DATA_LAYOUT_SOA
	205	printf(" [-pad auto\|modulus_1+offset_1(,modulus_n+offset_n)*]\n");
	206	#endif
10988083 MW	207
	208	return;
	209	}
	210
e3f82424	211	static void ParseParameters(Parameters * params, int * blk, PadInfo ** padInfo)
10988083 MW	212	{
	213	Assert(blk != NULL);
	214
	215	blk[0] = 0; blk[1] = 0; blk[2] = 0;
e3f82424	216	*padInfo = NULL;
10988083 MW	217
	218	#define ARG_IS(param) (!strcmp(params->KernelArgs[i], param))
	219	#define NEXT_ARG_PRESENT() \
	220	do { \
	221	if (i + 1 >= params->nKernelArgs) { \
	222	printf("ERROR: argument %s requires a parameter.\n", params->KernelArgs[i]); \
	223	exit(1); \
	224	} \
	225	} while (0)
	226
	227
	228	for (int i = 0; i < params->nKernelArgs; ++i) {
	229	if (ARG_IS("-blk") \|\| ARG_IS("--blk")) {
	230	NEXT_ARG_PRESENT();
	231
	232	int tmp = strtol(params->KernelArgs[++i], NULL, 0);
	233
e3f82424 MW	234	if (tmp < 0) {
e3f82424 MW	235	printf("ERROR: blocking parameter must be >= 0.\n");
10988083 MW	236	exit(1);
	237	}
	238
	239	blk[0] = blk[1] = blk[2] = tmp;
	240	}
	241	else if (ARG_IS("-blk-x") \|\| ARG_IS("--blk-x")) {
	242	NEXT_ARG_PRESENT();
	243
	244	int tmp = strtol(params->KernelArgs[++i], NULL, 0);
	245
e3f82424 MW	246	if (tmp < 0) {
e3f82424 MW	247	printf("ERROR: blocking parameter must be >= 0.\n");
10988083 MW	248	exit(1);
	249	}
	250
	251	blk[0] = tmp;
	252	}
	253	else if (ARG_IS("-blk-y") \|\| ARG_IS("--blk-y")) {
	254	NEXT_ARG_PRESENT();
	255
	256	int tmp = strtol(params->KernelArgs[++i], NULL, 0);
	257
e3f82424 MW	258	if (tmp < 0) {
e3f82424 MW	259	printf("ERROR: blocking parameter must be >= 0.\n");
10988083 MW	260	exit(1);
	261	}
	262
	263	blk[1] = tmp;
	264	}
	265	else if (ARG_IS("-blk-z") \|\| ARG_IS("--blk-z")) {
	266	NEXT_ARG_PRESENT();
	267
	268	int tmp = strtol(params->KernelArgs[++i], NULL, 0);
	269
e3f82424 MW	270	if (tmp < 0) {
e3f82424 MW	271	printf("ERROR: blocking parameter must be >= 0.\n");
10988083 MW	272	exit(1);
	273	}
	274
	275	blk[2] = tmp;
	276	}
	277	else if (ARG_IS("-h") \|\| ARG_IS("-help") \|\| ARG_IS("--help")) {
	278	ParameterUsage();
	279	exit(1);
	280	}
e3f82424 MW	281	#ifdef DATA_LAYOUT_SOA
	282	else if (ARG_IS("-pad") \|\| ARG_IS("--pad")) {
	283	NEXT_ARG_PRESENT();
	284
	285	*padInfo = PadInfoFromStr(params->KernelArgs[++i]);
	286	}
	287	#endif
10988083 MW	288	else {
	289	printf("ERROR: unknown kernel parameter.\n");
	290	ParameterUsage();
	291	exit(1);
	292	}
	293	}
	294
	295	#undef ARG_IS
	296	#undef NEXT_ARG_PRESENT
	297
	298	return;
	299	}
	300
	301	void FNAME(D3Q19ListInit)(LatticeDesc * ld, KernelData ** kernelData, Parameters * params)
	302	{
	303	KernelData * kd;
	304	KernelDataList * kdl;
	305	MemAlloc((void **)&kdl, sizeof(KernelDataList));
	306
	307	kd = (KernelData *)kdl;
	308	*kernelData = kd;
	309
	310	#ifdef DEBUG
	311	kd->Pdfs[0] = NULL;
	312	kd->Pdfs[1] = NULL;
	313	kd->PdfsActive = NULL;
	314	kd->DstPdfs = NULL;
	315	kd->SrcPdfs = NULL;
	316	kd->Dims[0] = -1;
	317	kd->Dims[1] = -1;
	318	kd->Dims[2] = -1;
	319	kd->GlobalDims[0] = -1;
	320	kd->GlobalDims[1] = -1;
	321	kd->GlobalDims[2] = -1;
	322	kd->Offsets[0] = -1;
	323	kd->Offsets[1] = -1;
	324	kd->Offsets[2] = -1;
	325
	326	kd->ObstIndices = NULL;
	327	kd->nObstIndices = -1;
	328	kd->BounceBackPdfsSrc = NULL;
	329	kd->BounceBackPdfsDst = NULL;
	330	kd->nBounceBackPdfs = -1;
	331
	332	kdl->AdjList = NULL;
	333	kdl->Coords = NULL;
	334	kdl->Grid = NULL;
	335	kdl->nCells = -1;
	336	kdl->nFluid = -1;
	337	#endif
	338
e3f82424 MW	339	int blk[3] = { 0 };
	340	PadInfo * padInfo = NULL;
	341
	342	ParseParameters(params, blk, &padInfo);
	343
	344
10988083 MW	345	// Ajust the dimensions according to padding, if used.
	346	kd->Dims[0] = kd->GlobalDims[0] = ld->Dims[0];
	347	kd->Dims[1] = kd->GlobalDims[1] = ld->Dims[1];
	348	kd->Dims[2] = kd->GlobalDims[2] = ld->Dims[2];
	349
	350	int * lDims = ld->Dims;
	351
	352	int lX = lDims[0];
	353	int lY = lDims[1];
	354	int lZ = lDims[2];
	355
	356	int nTotalCells = lX * lY * lZ;
e3f82424	357	int nCells = ld->nFluid;
10988083 MW	358	int nFluid = ld->nFluid;
10988083 MW	359
e3f82424 MW	360	#ifdef DATA_LAYOUT_SOA
	361	{
	362	nCells = PadCellsAndReport(nCells, sizeof(PdfT), &padInfo);
	363	PadInfoFree(padInfo); padInfo = NULL;
	364	}
	365	#endif
	366
10988083 MW	367	kdl->nCells = nCells;
	368	kdl->nFluid = nFluid;
	369
	370	PdfT * pdfs[2];
	371
10988083 MW	372	if (blk[0] == 0) blk[0] = lX;
	373	if (blk[1] == 0) blk[1] = lY;
	374	if (blk[2] == 0) blk[2] = lZ;
	375
	376	printf("# blocking x: %3d y: %3d z: %3d\n", blk[0], blk[1], blk[2]);
	377
	378	printf("# allocating data for %d fluid LB nodes with padding (%lu bytes = %f MiB for both lattices)\n",
	379	nCells, 2 * sizeof(PdfT) * nCells * N_D3Q19,
	380	2 * sizeof(PdfT) * nCells * N_D3Q19 / 1024.0 / 1024.0);
	381
	382	MemAlloc((void *)&pdfs[0], sizeof(PdfT) nCells * N_D3Q19);
	383	MemAlloc((void *)&pdfs[1], sizeof(PdfT) nCells * N_D3Q19);
	384
	385	kd->Pdfs[0] = pdfs[0];
	386	kd->Pdfs[1] = pdfs[1];
	387
	388	// Initialize PDFs with some (arbitrary) data for correct NUMA placement.
	389	// Here we touch only the fluid nodes as this loop is OpenMP parallel and
	390	// we want the same scheduling as in the kernel.
	391	#ifdef _OPENMP
	392	#pragma omp parallel for
	393	#endif
	394	for (int i = 0; i < nFluid; ++i) { for(int d = 0; d < N_D3Q19; ++d) {
	395	pdfs[0][P_INDEX_3(nCells, i, d)] = 1.0;
	396	pdfs[1][P_INDEX_3(nCells, i, d)] = 1.0;
	397	} }
	398
	399	// Initialize all PDFs to some standard value.
	400	for (int i = 0; i < nFluid; ++i) { for(int d = 0; d < N_D3Q19; ++d) {
	401	pdfs[0][P_INDEX_3(nCells, i, d)] = 0.0;
	402	pdfs[1][P_INDEX_3(nCells, i, d)] = 0.0;
	403	} }
	404
	405	// ----------------------------------------------------------------------
	406	// create grid which will hold the index numbers of the fluid nodes
	407
	408	uint32_t * grid;
	409
	410	if (MemAlloc((void *)&grid, nTotalCells sizeof(uint32_t))) {
	411	printf("ERROR: allocating grid for numbering failed: %lu bytes.\n", nTotalCells * sizeof(uint32_t));
	412	exit(1);
	413	}
	414	kdl->Grid = grid;
	415
	416	int latticeIndex;
	417
	418	#ifdef DEBUG
	419	for(int z = 0; z < lZ; ++z) {
	420	for(int y = 0; y < lY; ++y) {
	421	for(int x = 0; x < lX; ++x) {
	422
	423	latticeIndex = L_INDEX_4(ld->Dims, x, y, z);
	424
	425	grid[latticeIndex] = ~0;
	426	}
	427	}
	428	}
	429	#endif
	430
	431	// ----------------------------------------------------------------------
	432	// generate numbering over grid
	433
	434	uint32_t * coords;
	435
436	if (MemAlloc((void *)&coords, nFluid sizeof(uint32_t) * 3)) {
437	printf("ERROR: allocating coords array failed: %lu bytes.\n", nFluid * sizeof(uint32_t) * 3);
438	exit(1);
439	}
440
441	kdl->Coords = coords;
442
443	// Index for the PDF nodes can start at 0 as we distinguish solid and fluid nodes
444	// through the ld->Lattice array.
445	int counter = 0;
446
447	// Blocking is implemented via setup of the adjacency list. The kernel later will
448	// walk through the lattice blocked automatically.
10988083	449	for (int bX = 0; bX < lX; bX += blk[0]) {
e3f82424 MW	450	for (int bY = 0; bY < lY; bY += blk[1]) {
e3f82424 MW	451	for (int bZ = 0; bZ < lZ; bZ += blk[2]) {
10988083 MW	452
	453	int eX = MIN(bX + blk[0], lX);
	454	int eY = MIN(bY + blk[1], lY);
	455	int eZ = MIN(bZ + blk[2], lZ);
	456
10988083	457	for (int x = bX; x < eX; ++x) {
e3f82424 MW	458	for (int y = bY; y < eY; ++y) {
e3f82424 MW	459	for (int z = bZ; z < eZ; ++z) {
10988083 MW	460
	461	latticeIndex = L_INDEX_4(lDims, x, y, z);
	462
	463	if (ld->Lattice[latticeIndex] != LAT_CELL_OBSTACLE) {
	464	grid[latticeIndex] = counter;
	465
	466	coords[C_INDEX_X(counter)] = x;
	467	coords[C_INDEX_Y(counter)] = y;
	468	coords[C_INDEX_Z(counter)] = z;
	469
	470	++counter;
	471	}
	472	} } }
	473	} } }
	474
	475	Verify(counter == nFluid);
	476
	477	uint32_t * adjList;
	478
	479	// AdjList only requires 18 instead of 19 entries per node, as
	480	// the center PDF needs no addressing.
	481	if (MemAlloc((void *)&adjList, nFluid sizeof(uint32_t) * N_D3Q19_IDX)) {
	482	printf("ERROR: allocating adjList array failed: %lu bytes.\n", nFluid * sizeof(uint32_t) * N_D3Q19_IDX);
	483	exit(1);
	484	}
	485
	486	kdl->AdjList = adjList;
	487
	488	int x, y, z;
	489
	490	uint32_t neighborIndex;
	491	uint32_t dstIndex;
	492
	493	int nx, ny, nz, px, py, pz;
	494
	495	// Loop over all fluid nodes and compute the indices to the neighboring
	496	// PDFs for configured data layout (AoS/SoA).
e3f82424 MW	497	#ifdef _OPENMP
	498	#pragma omp parallel for
	499	#endif
	500	for (int index = 0; index < nFluid; ++index) {
	501	for (int d = 0; d < N_D3Q19_IDX; ++d) {
	502	adjList[index * N_D3Q19_IDX + d] = -1;
	503	}
	504	}
	505
10988083 MW	506	// #ifdef _OPENMP --> add line continuation
	507	// #pragma omp parallel for default(none)
	508	// shared(nFluid, nCells, coords, D3Q19_INV, D3Q19_X, D3Q19_Y, D3Q19_Z,
	509	// stderr,
	510	// lDims, grid, ld, lX, lY, lZ, adjList)
	511	// private(x, y, z, nx, ny, nz, neighborIndex, dstIndex)
	512	// #endif
	513	for (int index = 0; index < nFluid; ++index) {
	514	x = coords[C_INDEX_X(index)];
	515	y = coords[C_INDEX_Y(index)];
	516	z = coords[C_INDEX_Z(index)];
	517
	518	Assert(x >= 0 && x < lX);
	519	Assert(y >= 0 && y < lY);
	520	Assert(z >= 0 && z < lZ);
	521
	522	Assert(ld->Lattice[L_INDEX_4(lDims, x, y, z)] != LAT_CELL_OBSTACLE);
	523
	524	// Loop over all directions except the center one.
	525	for(int d = 0; d < N_D3Q19 - 1; ++d) {
	526	Assert(d != D3Q19_C);
	527	#ifdef PROP_MODEL_PUSH
	528	nx = x + D3Q19_X[d];
	529	ny = y + D3Q19_Y[d];
	530	nz = z + D3Q19_Z[d];
	531	#elif PROP_MODEL_PULL
	532	nx = x - D3Q19_X[d];
	533	ny = y - D3Q19_Y[d];
	534	nz = z - D3Q19_Z[d];
	535	#else
	536	#error No implementation for this PROP_MODEL_NAME.
	537	#endif
	538	// If the neighbor is outside the latcie in X direction and we have a
	539	// periodic boundary then we need to wrap around.
	540	if ( ((nx < 0 \|\| nx >= lX) && ld->PeriodicX) \|\|
	541	((ny < 0 \|\| ny >= lY) && ld->PeriodicY) \|\|
	542	((nz < 0 \|\| nz >= lZ) && ld->PeriodicZ)
	543	){
	544	// x periodic
	545
	546	if (nx < 0) {
	547	px = lX - 1;
	548	}
	549	else if (nx >= lX) {
	550	px = 0;
	551	} else {
	552	px = nx;
	553	}
	554	// y periodic
	555	if (ny < 0) {
	556	py = lY - 1;
	557	}
	558	else if (ny >= lY) {
	559	py = 0;
	560	} else {
	561	py = ny;
	562	}
	563
	564	// z periodic
	565	if (nz < 0) {
	566	pz = lZ - 1;
	567	}
	568	else if (nz >= lZ) {
	569	pz = 0;
570	} else {
571	pz = nz;
572	}
573
574	if (ld->Lattice[L_INDEX_4(lDims, px, py, pz)] == LAT_CELL_OBSTACLE) {
575	dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]);
576	}
577	else {
578	neighborIndex = grid[L_INDEX_4(lDims, px, py, pz)];
579
580	AssertMsg(neighborIndex != ~0, "Neighbor has no Index. (%d %d %d) direction %s (%d)\n", px, py, pz, D3Q19_NAMES[d], d);
581
582	dstIndex = P_INDEX_3(nCells, neighborIndex, d);
583	}
584	}
585	else if (nx < 0 \|\| ny < 0 \|\| nz < 0 \|\| nx >= lX \|\| ny >= lY \|\| nz >= lZ) {
586	dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]);
587	}
588	else if (ld->Lattice[L_INDEX_4(lDims, nx, ny, nz)] == LAT_CELL_OBSTACLE) {
589	dstIndex = P_INDEX_3(nCells, index, D3Q19_INV[d]);
590	}
591	else {
592	neighborIndex = grid[L_INDEX_4(lDims, nx, ny, nz)];
593
594	Assert(neighborIndex != ~0);
595
596	dstIndex = P_INDEX_3(nCells, neighborIndex, d);
597	}
598
599	Assert(dstIndex >= 0);
600	Assert(dstIndex < nCells * N_D3Q19);
601
602	adjList[index * N_D3Q19_IDX + d] = dstIndex;
603	}
604	}
605
606
607	// Fill remaining KernelData structures
608	kd->GetNode = GetNode;
609	kd->SetNode = SetNode;
610
611	kd->BoundaryConditionsGetPdf = FNAME(BCGetPdf);
612	kd->BoundaryConditionsSetPdf = FNAME(BCSetPdf);
613
614	kd->Kernel = FNAME(D3Q19ListKernel);
615
616	kd->DstPdfs = NULL;
617	kd->PdfsActive = kd->Pdfs[0];
618
619	return;
620	}
621
622	void FNAME(D3Q19ListDeinit)(LatticeDesc * ld, KernelData ** kernelData)
623	{
624	KernelDataList kdl = (KernelDataList )kernelData;
625
626	MemFree((void *)&((kernelData)->Pdfs[0]));
627	MemFree((void *)&((kernelData)->Pdfs[1]));
628
629	MemFree((void *)&((kdl)->AdjList));
630	MemFree((void *)&((kdl)->Coords));
631	MemFree((void *)&((kdl)->Grid));
632
633	MemFree((void **)kernelData);
634
635	return;
636	}
637