[ViewVC] Diff of: OpenMD/trunk/src/parallel/ForceMatrixDecomposition.cpp

Comparing branches/development/src/parallel/ForceMatrixDecomposition.cpp (file contents):
Revision 1582 by gezelter, Tue Jun 14 20:41:44 2011 UTC vs.
Revision 1755 by gezelter, Thu Jun 14 01:58:35 2012 UTC

#	Line 36 \| Line 36
36		* [1] Meineke, et al., J. Comp. Chem. 26, 252-271 (2005).
37		* [2] Fennell & Gezelter, J. Chem. Phys. 124, 234104 (2006).
38		* [3] Sun, Lin & Gezelter, J. Chem. Phys. 128, 24107 (2008).
39	<	* [4] Vardeman & Gezelter, in progress (2009).
39	>	* [4] Kuang & Gezelter, J. Chem. Phys. 133, 164101 (2010).
40	>	* [5] Vardeman, Stocker & Gezelter, J. Chem. Theory Comput. 7, 834 (2011).
41		*/
42		#include "parallel/ForceMatrixDecomposition.hpp"
43		#include "math/SquareMatrix3.hpp"
#	Line 47 \| Line 48 \| namespace OpenMD {
48		using namespace std;
49		namespace OpenMD {
50
51	+	ForceMatrixDecomposition::ForceMatrixDecomposition(SimInfo* info, InteractionManager* iMan) : ForceDecomposition(info, iMan) {
52	+
53	+	// In a parallel computation, row and colum scans must visit all
54	+	// surrounding cells (not just the 14 upper triangular blocks that
55	+	// are used when the processor can see all pairs)
56	+	#ifdef IS_MPI
57	+	cellOffsets_.clear();
58	+	cellOffsets_.push_back( Vector3i(-1,-1,-1) );
59	+	cellOffsets_.push_back( Vector3i( 0,-1,-1) );
60	+	cellOffsets_.push_back( Vector3i( 1,-1,-1) );
61	+	cellOffsets_.push_back( Vector3i(-1, 0,-1) );
62	+	cellOffsets_.push_back( Vector3i( 0, 0,-1) );
63	+	cellOffsets_.push_back( Vector3i( 1, 0,-1) );
64	+	cellOffsets_.push_back( Vector3i(-1, 1,-1) );
65	+	cellOffsets_.push_back( Vector3i( 0, 1,-1) );
66	+	cellOffsets_.push_back( Vector3i( 1, 1,-1) );
67	+	cellOffsets_.push_back( Vector3i(-1,-1, 0) );
68	+	cellOffsets_.push_back( Vector3i( 0,-1, 0) );
69	+	cellOffsets_.push_back( Vector3i( 1,-1, 0) );
70	+	cellOffsets_.push_back( Vector3i(-1, 0, 0) );
71	+	cellOffsets_.push_back( Vector3i( 0, 0, 0) );
72	+	cellOffsets_.push_back( Vector3i( 1, 0, 0) );
73	+	cellOffsets_.push_back( Vector3i(-1, 1, 0) );
74	+	cellOffsets_.push_back( Vector3i( 0, 1, 0) );
75	+	cellOffsets_.push_back( Vector3i( 1, 1, 0) );
76	+	cellOffsets_.push_back( Vector3i(-1,-1, 1) );
77	+	cellOffsets_.push_back( Vector3i( 0,-1, 1) );
78	+	cellOffsets_.push_back( Vector3i( 1,-1, 1) );
79	+	cellOffsets_.push_back( Vector3i(-1, 0, 1) );
80	+	cellOffsets_.push_back( Vector3i( 0, 0, 1) );
81	+	cellOffsets_.push_back( Vector3i( 1, 0, 1) );
82	+	cellOffsets_.push_back( Vector3i(-1, 1, 1) );
83	+	cellOffsets_.push_back( Vector3i( 0, 1, 1) );
84	+	cellOffsets_.push_back( Vector3i( 1, 1, 1) );
85	+	#endif
86	+	}
87	+
88	+
89		/**
90		* distributeInitialData is essentially a copy of the older fortran
91		* SimulationSetup
92		*/
54	–
93		void ForceMatrixDecomposition::distributeInitialData() {
94		snap_ = sman_->getCurrentSnapshot();
95		storageLayout_ = sman_->getStorageLayout();
96		ff_ = info_->getForceField();
97		nLocal_ = snap_->getNumberOfAtoms();
98	<
98	>
99		nGroups_ = info_->getNLocalCutoffGroups();
62	–	cerr << "in dId, nGroups = " << nGroups_ << "\n";
100		// gather the information for atomtype IDs (atids):
101	<	identsLocal = info_->getIdentArray();
101	>	idents = info_->getIdentArray();
102		AtomLocalToGlobal = info_->getGlobalAtomIndices();
103		cgLocalToGlobal = info_->getGlobalGroupIndices();
104		vector<int> globalGroupMembership = info_->getGlobalGroupMembership();
105	+
106		massFactors = info_->getMassFactors();
69	–	PairList excludes = info_->getExcludedInteractions();
70	–	PairList oneTwo = info_->getOneTwoInteractions();
71	–	PairList oneThree = info_->getOneThreeInteractions();
72	–	PairList oneFour = info_->getOneFourInteractions();
107
108	+	PairList* excludes = info_->getExcludedInteractions();
109	+	PairList* oneTwo = info_->getOneTwoInteractions();
110	+	PairList* oneThree = info_->getOneThreeInteractions();
111	+	PairList* oneFour = info_->getOneFourInteractions();
112	+
113	+	if (needVelocities_)
114	+	snap_->cgData.setStorageLayout(DataStorage::dslPosition \|
115	+	DataStorage::dslVelocity);
116	+	else
117	+	snap_->cgData.setStorageLayout(DataStorage::dslPosition);
118	+
119		#ifdef IS_MPI
120
121	<	AtomCommIntRow = new Communicator<Row,int>(nLocal_);
122	<	AtomCommRealRow = new Communicator<Row,RealType>(nLocal_);
78	<	AtomCommVectorRow = new Communicator<Row,Vector3d>(nLocal_);
79	<	AtomCommMatrixRow = new Communicator<Row,Mat3x3d>(nLocal_);
80	<	AtomCommPotRow = new Communicator<Row,potVec>(nLocal_);
121	>	MPI::Intracomm row = rowComm.getComm();
122	>	MPI::Intracomm col = colComm.getComm();
123
124	<	AtomCommIntColumn = new Communicator<Column,int>(nLocal_);
125	<	AtomCommRealColumn = new Communicator<Column,RealType>(nLocal_);
126	<	AtomCommVectorColumn = new Communicator<Column,Vector3d>(nLocal_);
127	<	AtomCommMatrixColumn = new Communicator<Column,Mat3x3d>(nLocal_);
128	<	AtomCommPotColumn = new Communicator<Column,potVec>(nLocal_);
124	>	AtomPlanIntRow = new Plan<int>(row, nLocal_);
125	>	AtomPlanRealRow = new Plan<RealType>(row, nLocal_);
126	>	AtomPlanVectorRow = new Plan<Vector3d>(row, nLocal_);
127	>	AtomPlanMatrixRow = new Plan<Mat3x3d>(row, nLocal_);
128	>	AtomPlanPotRow = new Plan<potVec>(row, nLocal_);
129
130	<	cgCommIntRow = new Communicator<Row,int>(nGroups_);
131	<	cgCommVectorRow = new Communicator<Row,Vector3d>(nGroups_);
132	<	cgCommIntColumn = new Communicator<Column,int>(nGroups_);
133	<	cgCommVectorColumn = new Communicator<Column,Vector3d>(nGroups_);
130	>	AtomPlanIntColumn = new Plan<int>(col, nLocal_);
131	>	AtomPlanRealColumn = new Plan<RealType>(col, nLocal_);
132	>	AtomPlanVectorColumn = new Plan<Vector3d>(col, nLocal_);
133	>	AtomPlanMatrixColumn = new Plan<Mat3x3d>(col, nLocal_);
134	>	AtomPlanPotColumn = new Plan<potVec>(col, nLocal_);
135
136	<	nAtomsInRow_ = AtomCommIntRow->getSize();
137	<	nAtomsInCol_ = AtomCommIntColumn->getSize();
138	<	nGroupsInRow_ = cgCommIntRow->getSize();
139	<	nGroupsInCol_ = cgCommIntColumn->getSize();
136	>	cgPlanIntRow = new Plan<int>(row, nGroups_);
137	>	cgPlanVectorRow = new Plan<Vector3d>(row, nGroups_);
138	>	cgPlanIntColumn = new Plan<int>(col, nGroups_);
139	>	cgPlanVectorColumn = new Plan<Vector3d>(col, nGroups_);
140
141	+	nAtomsInRow_ = AtomPlanIntRow->getSize();
142	+	nAtomsInCol_ = AtomPlanIntColumn->getSize();
143	+	nGroupsInRow_ = cgPlanIntRow->getSize();
144	+	nGroupsInCol_ = cgPlanIntColumn->getSize();
145	+
146		// Modify the data storage objects with the correct layouts and sizes:
147		atomRowData.resize(nAtomsInRow_);
148		atomRowData.setStorageLayout(storageLayout_);
#	Line 103 \| Line 151 \| namespace OpenMD {
151		cgRowData.resize(nGroupsInRow_);
152		cgRowData.setStorageLayout(DataStorage::dslPosition);
153		cgColData.resize(nGroupsInCol_);
154	<	cgColData.setStorageLayout(DataStorage::dslPosition);
155	<
154	>	if (needVelocities_)
155	>	// we only need column velocities if we need them.
156	>	cgColData.setStorageLayout(DataStorage::dslPosition \|
157	>	DataStorage::dslVelocity);
158	>	else
159	>	cgColData.setStorageLayout(DataStorage::dslPosition);
160	>
161		identsRow.resize(nAtomsInRow_);
162		identsCol.resize(nAtomsInCol_);
163
164	<	AtomCommIntRow->gather(identsLocal, identsRow);
165	<	AtomCommIntColumn->gather(identsLocal, identsCol);
164	>	AtomPlanIntRow->gather(idents, identsRow);
165	>	AtomPlanIntColumn->gather(idents, identsCol);
166
167	<	AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
168	<	AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
169	<
117	<	cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
118	<	cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
167	>	// allocate memory for the parallel objects
168	>	atypesRow.resize(nAtomsInRow_);
169	>	atypesCol.resize(nAtomsInCol_);
170
171	<	AtomCommRealRow->gather(massFactors, massFactorsRow);
172	<	AtomCommRealColumn->gather(massFactors, massFactorsCol);
171	>	for (int i = 0; i < nAtomsInRow_; i++)
172	>	atypesRow[i] = ff_->getAtomType(identsRow[i]);
173	>	for (int i = 0; i < nAtomsInCol_; i++)
174	>	atypesCol[i] = ff_->getAtomType(identsCol[i]);
175
176	+	pot_row.resize(nAtomsInRow_);
177	+	pot_col.resize(nAtomsInCol_);
178	+
179	+	AtomRowToGlobal.resize(nAtomsInRow_);
180	+	AtomColToGlobal.resize(nAtomsInCol_);
181	+	AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
182	+	AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
183	+
184	+	cgRowToGlobal.resize(nGroupsInRow_);
185	+	cgColToGlobal.resize(nGroupsInCol_);
186	+	cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
187	+	cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
188	+
189	+	massFactorsRow.resize(nAtomsInRow_);
190	+	massFactorsCol.resize(nAtomsInCol_);
191	+	AtomPlanRealRow->gather(massFactors, massFactorsRow);
192	+	AtomPlanRealColumn->gather(massFactors, massFactorsCol);
193	+
194		groupListRow_.clear();
195		groupListRow_.resize(nGroupsInRow_);
196		for (int i = 0; i < nGroupsInRow_; i++) {
#	Line 142 \| Line 213 \| namespace OpenMD {
213		}
214		}
215
216	<	skipsForAtom.clear();
217	<	skipsForAtom.resize(nAtomsInRow_);
216	>	excludesForAtom.clear();
217	>	excludesForAtom.resize(nAtomsInRow_);
218		toposForAtom.clear();
219		toposForAtom.resize(nAtomsInRow_);
220		topoDist.clear();
#	Line 154 \| Line 225 \| namespace OpenMD {
225		for (int j = 0; j < nAtomsInCol_; j++) {
226		int jglob = AtomColToGlobal[j];
227
228	<	if (excludes.hasPair(iglob, jglob))
229	<	skipsForAtom[i].push_back(j);
228	>	if (excludes->hasPair(iglob, jglob))
229	>	excludesForAtom[i].push_back(j);
230
231	<	if (oneTwo.hasPair(iglob, jglob)) {
231	>	if (oneTwo->hasPair(iglob, jglob)) {
232		toposForAtom[i].push_back(j);
233		topoDist[i].push_back(1);
234		} else {
235	<	if (oneThree.hasPair(iglob, jglob)) {
235	>	if (oneThree->hasPair(iglob, jglob)) {
236		toposForAtom[i].push_back(j);
237		topoDist[i].push_back(2);
238		} else {
239	<	if (oneFour.hasPair(iglob, jglob)) {
239	>	if (oneFour->hasPair(iglob, jglob)) {
240		toposForAtom[i].push_back(j);
241		topoDist[i].push_back(3);
242		}
#	Line 174 \| Line 245 \| namespace OpenMD {
245		}
246		}
247
248	<	#endif
249	<
250	<	groupList_.clear();
180	<	groupList_.resize(nGroups_);
181	<	for (int i = 0; i < nGroups_; i++) {
182	<	int gid = cgLocalToGlobal[i];
183	<	for (int j = 0; j < nLocal_; j++) {
184	<	int aid = AtomLocalToGlobal[j];
185	<	if (globalGroupMembership[aid] == gid) {
186	<	groupList_[i].push_back(j);
187	<	}
188	<	}
189	<	}
190	<
191	<	skipsForAtom.clear();
192	<	skipsForAtom.resize(nLocal_);
248	>	#else
249	>	excludesForAtom.clear();
250	>	excludesForAtom.resize(nLocal_);
251		toposForAtom.clear();
252		toposForAtom.resize(nLocal_);
253		topoDist.clear();
#	Line 201 \| Line 259 \| namespace OpenMD {
259		for (int j = 0; j < nLocal_; j++) {
260		int jglob = AtomLocalToGlobal[j];
261
262	<	if (excludes.hasPair(iglob, jglob))
263	<	skipsForAtom[i].push_back(j);
262	>	if (excludes->hasPair(iglob, jglob))
263	>	excludesForAtom[i].push_back(j);
264
265	<	if (oneTwo.hasPair(iglob, jglob)) {
265	>	if (oneTwo->hasPair(iglob, jglob)) {
266		toposForAtom[i].push_back(j);
267		topoDist[i].push_back(1);
268		} else {
269	<	if (oneThree.hasPair(iglob, jglob)) {
269	>	if (oneThree->hasPair(iglob, jglob)) {
270		toposForAtom[i].push_back(j);
271		topoDist[i].push_back(2);
272		} else {
273	<	if (oneFour.hasPair(iglob, jglob)) {
273	>	if (oneFour->hasPair(iglob, jglob)) {
274		toposForAtom[i].push_back(j);
275		topoDist[i].push_back(3);
276		}
#	Line 220 \| Line 278 \| namespace OpenMD {
278		}
279		}
280		}
281	<
281	>	#endif
282	>
283	>	// allocate memory for the parallel objects
284	>	atypesLocal.resize(nLocal_);
285	>
286	>	for (int i = 0; i < nLocal_; i++)
287	>	atypesLocal[i] = ff_->getAtomType(idents[i]);
288	>
289	>	groupList_.clear();
290	>	groupList_.resize(nGroups_);
291	>	for (int i = 0; i < nGroups_; i++) {
292	>	int gid = cgLocalToGlobal[i];
293	>	for (int j = 0; j < nLocal_; j++) {
294	>	int aid = AtomLocalToGlobal[j];
295	>	if (globalGroupMembership[aid] == gid) {
296	>	groupList_[i].push_back(j);
297	>	}
298	>	}
299	>	}
300	>
301	>
302		createGtypeCutoffMap();
303	+
304		}
305
306		void ForceMatrixDecomposition::createGtypeCutoffMap() {
307	<
307	>
308		RealType tol = 1e-6;
309	+	largestRcut_ = 0.0;
310		RealType rc;
311		int atid;
312		set<AtomType*> atypes = info_->getSimulatedAtomTypes();
313	<	vector<RealType> atypeCutoff;
314	<	atypeCutoff.resize( atypes.size() );
315	<
313	>
314	>	map<int, RealType> atypeCutoff;
315	>
316		for (set<AtomType*>::iterator at = atypes.begin();
317		at != atypes.end(); ++at){
238	–	rc = interactionMan_->getSuggestedCutoffRadius(*at);
318		atid = (*at)->getIdent();
319	<	atypeCutoff[atid] = rc;
319	>	if (userChoseCutoff_)
320	>	atypeCutoff[atid] = userCutoff_;
321	>	else
322	>	atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at);
323		}
324	<
324	>
325		vector<RealType> gTypeCutoffs;
244	–
326		// first we do a single loop over the cutoff groups to find the
327		// largest cutoff for any atypes present in this group.
328		#ifdef IS_MPI
#	Line 299 \| Line 380 \| namespace OpenMD {
380
381		vector<RealType> groupCutoff(nGroups_, 0.0);
382		groupToGtype.resize(nGroups_);
302	–
303	–	cerr << "nGroups = " << nGroups_ << "\n";
383		for (int cg1 = 0; cg1 < nGroups_; cg1++) {
305	–
384		groupCutoff[cg1] = 0.0;
385		vector<int> atomList = getAtomsInGroupRow(cg1);
308	–
386		for (vector<int>::iterator ia = atomList.begin();
387		ia != atomList.end(); ++ia) {
388		int atom1 = (*ia);
389	<	atid = identsLocal[atom1];
390	<	if (atypeCutoff[atid] > groupCutoff[cg1]) {
389	>	atid = idents[atom1];
390	>	if (atypeCutoff[atid] > groupCutoff[cg1])
391		groupCutoff[cg1] = atypeCutoff[atid];
315	–	}
392		}
393	<
393	>
394		bool gTypeFound = false;
395		for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
396		if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
#	Line 322 \| Line 398 \| namespace OpenMD {
398		gTypeFound = true;
399		}
400		}
401	<	if (!gTypeFound) {
401	>	if (!gTypeFound) {
402		gTypeCutoffs.push_back( groupCutoff[cg1] );
403		groupToGtype[cg1] = gTypeCutoffs.size() - 1;
404		}
405		}
406		#endif
407
332	–	cerr << "gTypeCutoffs.size() = " << gTypeCutoffs.size() << "\n";
408		// Now we find the maximum group cutoff value present in the simulation
409
410	<	RealType groupMax = *max_element(gTypeCutoffs.begin(), gTypeCutoffs.end());
410	>	RealType groupMax = *max_element(gTypeCutoffs.begin(),
411	>	gTypeCutoffs.end());
412
413		#ifdef IS_MPI
414	<	MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE, MPI::MAX);
414	>	MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE,
415	>	MPI::MAX);
416		#endif
417
418		RealType tradRcut = groupMax;
#	Line 365 \| Line 442 \| namespace OpenMD {
442
443		pair<int,int> key = make_pair(i,j);
444		gTypeCutoffMap[key].first = thisRcut;
368	–
445		if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
370	–
446		gTypeCutoffMap[key].second = thisRcut*thisRcut;
372	–
447		gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
374	–
448		// sanity check
449
450		if (userChoseCutoff_) {
451		if (abs(gTypeCutoffMap[key].first - userCutoff_) > 0.0001) {
452		sprintf(painCave.errMsg,
453		"ForceMatrixDecomposition::createGtypeCutoffMap "
454	<	"user-specified rCut does not match computed group Cutoff\n");
454	>	"user-specified rCut (%lf) does not match computed group Cutoff\n", userCutoff_);
455		painCave.severity = OPENMD_ERROR;
456		painCave.isFatal = 1;
457		simError();
#	Line 410 \| Line 483 \| namespace OpenMD {
483		}
484
485		void ForceMatrixDecomposition::zeroWorkArrays() {
486	+	pairwisePot = 0.0;
487	+	embeddingPot = 0.0;
488
414	–	for (int j = 0; j < N_INTERACTION_FAMILIES; j++) {
415	–	longRangePot_[j] = 0.0;
416	–	}
417	–
489		#ifdef IS_MPI
490		if (storageLayout_ & DataStorage::dslForce) {
491		fill(atomRowData.force.begin(), atomRowData.force.end(), V3Zero);
#	Line 430 \| Line 501 \| namespace OpenMD {
501		Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
502
503		fill(pot_col.begin(), pot_col.end(),
504	<	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
434	<
435	<	pot_local = Vector<RealType, N_INTERACTION_FAMILIES>(0.0);
504	>	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
505
506		if (storageLayout_ & DataStorage::dslParticlePot) {
507	<	fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(), 0.0);
508	<	fill(atomColData.particlePot.begin(), atomColData.particlePot.end(), 0.0);
507	>	fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(),
508	>	0.0);
509	>	fill(atomColData.particlePot.begin(), atomColData.particlePot.end(),
510	>	0.0);
511		}
512
513		if (storageLayout_ & DataStorage::dslDensity) {
#	Line 445 \| Line 516 \| namespace OpenMD {
516		}
517
518		if (storageLayout_ & DataStorage::dslFunctional) {
519	<	fill(atomRowData.functional.begin(), atomRowData.functional.end(), 0.0);
520	<	fill(atomColData.functional.begin(), atomColData.functional.end(), 0.0);
519	>	fill(atomRowData.functional.begin(), atomRowData.functional.end(),
520	>	0.0);
521	>	fill(atomColData.functional.begin(), atomColData.functional.end(),
522	>	0.0);
523		}
524
525		if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
#	Line 456 \| Line 529 \| namespace OpenMD {
529		atomColData.functionalDerivative.end(), 0.0);
530		}
531
532	<	#else
533	<
532	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
533	>	fill(atomRowData.skippedCharge.begin(),
534	>	atomRowData.skippedCharge.end(), 0.0);
535	>	fill(atomColData.skippedCharge.begin(),
536	>	atomColData.skippedCharge.end(), 0.0);
537	>	}
538	>
539	>	if (storageLayout_ & DataStorage::dslFlucQForce) {
540	>	fill(atomRowData.flucQFrc.begin(),
541	>	atomRowData.flucQFrc.end(), 0.0);
542	>	fill(atomColData.flucQFrc.begin(),
543	>	atomColData.flucQFrc.end(), 0.0);
544	>	}
545	>
546	>	if (storageLayout_ & DataStorage::dslElectricField) {
547	>	fill(atomRowData.electricField.begin(),
548	>	atomRowData.electricField.end(), V3Zero);
549	>	fill(atomColData.electricField.begin(),
550	>	atomColData.electricField.end(), V3Zero);
551	>	}
552	>
553	>	if (storageLayout_ & DataStorage::dslFlucQForce) {
554	>	fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
555	>	0.0);
556	>	fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
557	>	0.0);
558	>	}
559	>
560	>	#endif
561	>	// even in parallel, we need to zero out the local arrays:
562	>
563		if (storageLayout_ & DataStorage::dslParticlePot) {
564		fill(snap_->atomData.particlePot.begin(),
565		snap_->atomData.particlePot.end(), 0.0);
#	Line 467 \| Line 569 \| namespace OpenMD {
569		fill(snap_->atomData.density.begin(),
570		snap_->atomData.density.end(), 0.0);
571		}
572	+
573		if (storageLayout_ & DataStorage::dslFunctional) {
574		fill(snap_->atomData.functional.begin(),
575		snap_->atomData.functional.end(), 0.0);
576		}
577	+
578		if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
579		fill(snap_->atomData.functionalDerivative.begin(),
580		snap_->atomData.functionalDerivative.end(), 0.0);
581		}
582	<	#endif
583	<
582	>
583	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
584	>	fill(snap_->atomData.skippedCharge.begin(),
585	>	snap_->atomData.skippedCharge.end(), 0.0);
586	>	}
587	>
588	>	if (storageLayout_ & DataStorage::dslElectricField) {
589	>	fill(snap_->atomData.electricField.begin(),
590	>	snap_->atomData.electricField.end(), V3Zero);
591	>	}
592		}
593
594
#	Line 486 \| Line 598 \| namespace OpenMD {
598		#ifdef IS_MPI
599
600		// gather up the atomic positions
601	<	AtomCommVectorRow->gather(snap_->atomData.position,
601	>	AtomPlanVectorRow->gather(snap_->atomData.position,
602		atomRowData.position);
603	<	AtomCommVectorColumn->gather(snap_->atomData.position,
603	>	AtomPlanVectorColumn->gather(snap_->atomData.position,
604		atomColData.position);
605
606		// gather up the cutoff group positions
607	<	cgCommVectorRow->gather(snap_->cgData.position,
607	>
608	>	cgPlanVectorRow->gather(snap_->cgData.position,
609		cgRowData.position);
610	<	cgCommVectorColumn->gather(snap_->cgData.position,
610	>
611	>	cgPlanVectorColumn->gather(snap_->cgData.position,
612		cgColData.position);
613	+
614	+
615	+
616	+	if (needVelocities_) {
617	+	// gather up the atomic velocities
618	+	AtomPlanVectorColumn->gather(snap_->atomData.velocity,
619	+	atomColData.velocity);
620	+
621	+	cgPlanVectorColumn->gather(snap_->cgData.velocity,
622	+	cgColData.velocity);
623	+	}
624	+
625
626		// if needed, gather the atomic rotation matrices
627		if (storageLayout_ & DataStorage::dslAmat) {
628	<	AtomCommMatrixRow->gather(snap_->atomData.aMat,
628	>	AtomPlanMatrixRow->gather(snap_->atomData.aMat,
629		atomRowData.aMat);
630	<	AtomCommMatrixColumn->gather(snap_->atomData.aMat,
630	>	AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
631		atomColData.aMat);
632		}
633
634		// if needed, gather the atomic eletrostatic frames
635		if (storageLayout_ & DataStorage::dslElectroFrame) {
636	<	AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
636	>	AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
637		atomRowData.electroFrame);
638	<	AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
638	>	AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
639		atomColData.electroFrame);
640		}
641	+
642	+	// if needed, gather the atomic fluctuating charge values
643	+	if (storageLayout_ & DataStorage::dslFlucQPosition) {
644	+	AtomPlanRealRow->gather(snap_->atomData.flucQPos,
645	+	atomRowData.flucQPos);
646	+	AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
647	+	atomColData.flucQPos);
648	+	}
649	+
650		#endif
651		}
652
#	Line 525 \| Line 660 \| namespace OpenMD {
660
661		if (storageLayout_ & DataStorage::dslDensity) {
662
663	<	AtomCommRealRow->scatter(atomRowData.density,
663	>	AtomPlanRealRow->scatter(atomRowData.density,
664		snap_->atomData.density);
665
666		int n = snap_->atomData.density.size();
667		vector<RealType> rho_tmp(n, 0.0);
668	<	AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
668	>	AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
669		for (int i = 0; i < n; i++)
670		snap_->atomData.density[i] += rho_tmp[i];
671		}
672	+
673	+	if (storageLayout_ & DataStorage::dslElectricField) {
674	+
675	+	AtomPlanVectorRow->scatter(atomRowData.electricField,
676	+	snap_->atomData.electricField);
677	+
678	+	int n = snap_->atomData.electricField.size();
679	+	vector<Vector3d> field_tmp(n, V3Zero);
680	+	AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
681	+	for (int i = 0; i < n; i++)
682	+	snap_->atomData.electricField[i] += field_tmp[i];
683	+	}
684		#endif
685		}
686
#	Line 546 \| Line 693 \| namespace OpenMD {
693		storageLayout_ = sman_->getStorageLayout();
694		#ifdef IS_MPI
695		if (storageLayout_ & DataStorage::dslFunctional) {
696	<	AtomCommRealRow->gather(snap_->atomData.functional,
696	>	AtomPlanRealRow->gather(snap_->atomData.functional,
697		atomRowData.functional);
698	<	AtomCommRealColumn->gather(snap_->atomData.functional,
698	>	AtomPlanRealColumn->gather(snap_->atomData.functional,
699		atomColData.functional);
700		}
701
702		if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
703	<	AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
703	>	AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
704		atomRowData.functionalDerivative);
705	<	AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
705	>	AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
706		atomColData.functionalDerivative);
707		}
708		#endif
#	Line 569 \| Line 716 \| namespace OpenMD {
716		int n = snap_->atomData.force.size();
717		vector<Vector3d> frc_tmp(n, V3Zero);
718
719	<	AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
719	>	AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
720		for (int i = 0; i < n; i++) {
721		snap_->atomData.force[i] += frc_tmp[i];
722		frc_tmp[i] = 0.0;
723		}
724
725	<	AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
726	<	for (int i = 0; i < n; i++)
725	>	AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
726	>	for (int i = 0; i < n; i++) {
727		snap_->atomData.force[i] += frc_tmp[i];
728	<
729	<
728	>	}
729	>
730		if (storageLayout_ & DataStorage::dslTorque) {
731
732	<	int nt = snap_->atomData.force.size();
732	>	int nt = snap_->atomData.torque.size();
733		vector<Vector3d> trq_tmp(nt, V3Zero);
734
735	<	AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
736	<	for (int i = 0; i < n; i++) {
735	>	AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
736	>	for (int i = 0; i < nt; i++) {
737		snap_->atomData.torque[i] += trq_tmp[i];
738		trq_tmp[i] = 0.0;
739		}
740
741	<	AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
742	<	for (int i = 0; i < n; i++)
741	>	AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
742	>	for (int i = 0; i < nt; i++)
743		snap_->atomData.torque[i] += trq_tmp[i];
744		}
745	+
746	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
747	+
748	+	int ns = snap_->atomData.skippedCharge.size();
749	+	vector<RealType> skch_tmp(ns, 0.0);
750	+
751	+	AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
752	+	for (int i = 0; i < ns; i++) {
753	+	snap_->atomData.skippedCharge[i] += skch_tmp[i];
754	+	skch_tmp[i] = 0.0;
755	+	}
756	+
757	+	AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
758	+	for (int i = 0; i < ns; i++)
759	+	snap_->atomData.skippedCharge[i] += skch_tmp[i];
760	+
761	+	}
762
763	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
764	+
765	+	int nq = snap_->atomData.flucQFrc.size();
766	+	vector<RealType> fqfrc_tmp(nq, 0.0);
767	+
768	+	AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
769	+	for (int i = 0; i < nq; i++) {
770	+	snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
771	+	fqfrc_tmp[i] = 0.0;
772	+	}
773	+
774	+	AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
775	+	for (int i = 0; i < nq; i++)
776	+	snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
777	+
778	+	}
779	+
780		nLocal_ = snap_->getNumberOfAtoms();
781
782		vector<potVec> pot_temp(nLocal_,
#	Line 603 \| Line 784 \| namespace OpenMD {
784
785		// scatter/gather pot_row into the members of my column
786
787	<	AtomCommPotRow->scatter(pot_row, pot_temp);
787	>	AtomPlanPotRow->scatter(pot_row, pot_temp);
788
789		for (int ii = 0; ii < pot_temp.size(); ii++ )
790	<	pot_local += pot_temp[ii];
791	<
790	>	pairwisePot += pot_temp[ii];
791	>
792	>	if (storageLayout_ & DataStorage::dslParticlePot) {
793	>	// This is the pairwise contribution to the particle pot. The
794	>	// embedding contribution is added in each of the low level
795	>	// non-bonded routines. In single processor, this is done in
796	>	// unpackInteractionData, not in collectData.
797	>	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
798	>	for (int i = 0; i < nLocal_; i++) {
799	>	// factor of two is because the total potential terms are divided
800	>	// by 2 in parallel due to row/ column scatter
801	>	snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
802	>	}
803	>	}
804	>	}
805	>
806		fill(pot_temp.begin(), pot_temp.end(),
807		Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
808
809	<	AtomCommPotColumn->scatter(pot_col, pot_temp);
809	>	AtomPlanPotColumn->scatter(pot_col, pot_temp);
810
811		for (int ii = 0; ii < pot_temp.size(); ii++ )
812	<	pot_local += pot_temp[ii];
812	>	pairwisePot += pot_temp[ii];
813	>
814	>	if (storageLayout_ & DataStorage::dslParticlePot) {
815	>	// This is the pairwise contribution to the particle pot. The
816	>	// embedding contribution is added in each of the low level
817	>	// non-bonded routines. In single processor, this is done in
818	>	// unpackInteractionData, not in collectData.
819	>	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
820	>	for (int i = 0; i < nLocal_; i++) {
821	>	// factor of two is because the total potential terms are divided
822	>	// by 2 in parallel due to row/ column scatter
823	>	snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
824	>	}
825	>	}
826	>	}
827	>
828	>	if (storageLayout_ & DataStorage::dslParticlePot) {
829	>	int npp = snap_->atomData.particlePot.size();
830	>	vector<RealType> ppot_temp(npp, 0.0);
831	>
832	>	// This is the direct or embedding contribution to the particle
833	>	// pot.
834	>
835	>	AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
836	>	for (int i = 0; i < npp; i++) {
837	>	snap_->atomData.particlePot[i] += ppot_temp[i];
838	>	}
839	>
840	>	fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
841	>
842	>	AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
843	>	for (int i = 0; i < npp; i++) {
844	>	snap_->atomData.particlePot[i] += ppot_temp[i];
845	>	}
846	>	}
847	>
848	>	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
849	>	RealType ploc1 = pairwisePot[ii];
850	>	RealType ploc2 = 0.0;
851	>	MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
852	>	pairwisePot[ii] = ploc2;
853	>	}
854	>
855	>	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
856	>	RealType ploc1 = embeddingPot[ii];
857	>	RealType ploc2 = 0.0;
858	>	MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
859	>	embeddingPot[ii] = ploc2;
860	>	}
861
862	+	// Here be dragons.
863	+	MPI::Intracomm col = colComm.getComm();
864	+
865	+	col.Allreduce(MPI::IN_PLACE,
866	+	&snap_->frameData.conductiveHeatFlux[0], 3,
867	+	MPI::REALTYPE, MPI::SUM);
868	+
869	+
870		#endif
871	+
872		}
873
874		int ForceMatrixDecomposition::getNAtomsInRow() {
#	Line 659 \| Line 911 \| namespace OpenMD {
911		return d;
912		}
913
914	+	Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
915	+	#ifdef IS_MPI
916	+	return cgColData.velocity[cg2];
917	+	#else
918	+	return snap_->cgData.velocity[cg2];
919	+	#endif
920	+	}
921
922	+	Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
923	+	#ifdef IS_MPI
924	+	return atomColData.velocity[atom2];
925	+	#else
926	+	return snap_->atomData.velocity[atom2];
927	+	#endif
928	+	}
929	+
930	+
931		Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
932
933		Vector3d d;
#	Line 717 \| Line 985 \| namespace OpenMD {
985		return d;
986		}
987
988	<	vector<int> ForceMatrixDecomposition::getSkipsForAtom(int atom1) {
989	<	return skipsForAtom[atom1];
988	>	vector<int> ForceMatrixDecomposition::getExcludesForAtom(int atom1) {
989	>	return excludesForAtom[atom1];
990		}
991
992		/**
993	<	* There are a number of reasons to skip a pair or a
726	<	* particle. Mostly we do this to exclude atoms who are involved in
727	<	* short range interactions (bonds, bends, torsions), but we also
728	<	* need to exclude some overcounted interactions that result from
993	>	* We need to exclude some overcounted interactions that result from
994		* the parallel decomposition.
995		*/
996		bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
997		int unique_id_1, unique_id_2;
998	<
998	>
999		#ifdef IS_MPI
1000		// in MPI, we have to look up the unique IDs for each atom
1001		unique_id_1 = AtomRowToGlobal[atom1];
1002		unique_id_2 = AtomColToGlobal[atom2];
1003	+	#else
1004	+	unique_id_1 = AtomLocalToGlobal[atom1];
1005	+	unique_id_2 = AtomLocalToGlobal[atom2];
1006	+	#endif
1007
739	–	// this situation should only arise in MPI simulations
1008		if (unique_id_1 == unique_id_2) return true;
1009	<
1009	>
1010	>	#ifdef IS_MPI
1011		// this prevents us from doing the pair on multiple processors
1012		if (unique_id_1 < unique_id_2) {
1013		if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1014		} else {
1015	<	if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1015	>	if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1016		}
748	–	#else
749	–	// in the normal loop, the atom numbers are unique
750	–	unique_id_1 = atom1;
751	–	unique_id_2 = atom2;
1017		#endif
1018
1019	<	for (vector<int>::iterator i = skipsForAtom[atom1].begin();
1020	<	i != skipsForAtom[atom1].end(); ++i) {
756	<	if ( (*i) == unique_id_2 ) return true;
757	<	}
1019	>	return false;
1020	>	}
1021
1022	+	/**
1023	+	* We need to handle the interactions for atoms who are involved in
1024	+	* the same rigid body as well as some short range interactions
1025	+	* (bonds, bends, torsions) differently from other interactions.
1026	+	* We'll still visit the pairwise routines, but with a flag that
1027	+	* tells those routines to exclude the pair from direct long range
1028	+	* interactions. Some indirect interactions (notably reaction
1029	+	* field) must still be handled for these pairs.
1030	+	*/
1031	+	bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1032	+
1033	+	// excludesForAtom was constructed to use row/column indices in the MPI
1034	+	// version, and to use local IDs in the non-MPI version:
1035	+
1036	+	for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1037	+	i != excludesForAtom[atom1].end(); ++i) {
1038	+	if ( (*i) == atom2 ) return true;
1039	+	}
1040	+
1041	+	return false;
1042		}
1043
1044
#	Line 777 \| Line 1060 \| namespace OpenMD {
1060
1061		// filling interaction blocks with pointers
1062		void ForceMatrixDecomposition::fillInteractionData(InteractionData &idat,
1063	<	int atom1, int atom2) {
1063	>	int atom1, int atom2) {
1064	>
1065	>	idat.excluded = excludeAtomPair(atom1, atom2);
1066	>
1067		#ifdef IS_MPI
1068	+	idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
1069	+	//idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
1070	+	// ff_->getAtomType(identsCol[atom2]) );
1071
783	–	idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
784	–	ff_->getAtomType(identsCol[atom2]) );
785	–
1072		if (storageLayout_ & DataStorage::dslAmat) {
1073		idat.A1 = &(atomRowData.aMat[atom1]);
1074		idat.A2 = &(atomColData.aMat[atom2]);
#	Line 818 \| Line 1104 \| namespace OpenMD {
1104		idat.particlePot2 = &(atomColData.particlePot[atom2]);
1105		}
1106
1107	<	#else
1107	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1108	>	idat.skippedCharge1 = &(atomRowData.skippedCharge[atom1]);
1109	>	idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1110	>	}
1111
1112	<	idat.atypes = make_pair( ff_->getAtomType(identsLocal[atom1]),
1113	<	ff_->getAtomType(identsLocal[atom2]) );
1112	>	if (storageLayout_ & DataStorage::dslFlucQPosition) {
1113	>	idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1114	>	idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1115	>	}
1116
1117	+	#else
1118	+
1119	+	idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1120	+
1121		if (storageLayout_ & DataStorage::dslAmat) {
1122		idat.A1 = &(snap_->atomData.aMat[atom1]);
1123		idat.A2 = &(snap_->atomData.aMat[atom2]);
#	Line 838 \| Line 1133 \| namespace OpenMD {
1133		idat.t2 = &(snap_->atomData.torque[atom2]);
1134		}
1135
1136	<	if (storageLayout_ & DataStorage::dslDensity) {
1136	>	if (storageLayout_ & DataStorage::dslDensity) {
1137		idat.rho1 = &(snap_->atomData.density[atom1]);
1138		idat.rho2 = &(snap_->atomData.density[atom2]);
1139		}
#	Line 858 \| Line 1153 \| namespace OpenMD {
1153		idat.particlePot2 = &(snap_->atomData.particlePot[atom2]);
1154		}
1155
1156	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1157	+	idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1158	+	idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1159	+	}
1160	+
1161	+	if (storageLayout_ & DataStorage::dslFlucQPosition) {
1162	+	idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1163	+	idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1164	+	}
1165	+
1166		#endif
1167		}
1168
1169
1170		void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {
1171		#ifdef IS_MPI
1172	<	pot_row[atom1] += 0.5 * *(idat.pot);
1173	<	pot_col[atom2] += 0.5 * *(idat.pot);
1172	>	pot_row[atom1] += RealType(0.5) * *(idat.pot);
1173	>	pot_col[atom2] += RealType(0.5) * *(idat.pot);
1174
1175		atomRowData.force[atom1] += *(idat.f1);
1176		atomColData.force[atom2] -= *(idat.f1);
872	–	#else
873	–	longRangePot_ += *(idat.pot);
874	–
875	–	snap_->atomData.force[atom1] += *(idat.f1);
876	–	snap_->atomData.force[atom2] -= *(idat.f1);
877	–	#endif
1177
1178	<	}
1178	>	if (storageLayout_ & DataStorage::dslFlucQForce) {
1179	>	atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1180	>	atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1181	>	}
1182
1183	+	if (storageLayout_ & DataStorage::dslElectricField) {
1184	+	atomRowData.electricField[atom1] += *(idat.eField1);
1185	+	atomColData.electricField[atom2] += *(idat.eField2);
1186	+	}
1187
1188	<	void ForceMatrixDecomposition::fillSkipData(InteractionData &idat,
1189	<	int atom1, int atom2) {
884	<	#ifdef IS_MPI
885	<	idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
886	<	ff_->getAtomType(identsCol[atom2]) );
1188	>	#else
1189	>	pairwisePot += *(idat.pot);
1190
1191	<	if (storageLayout_ & DataStorage::dslElectroFrame) {
1192	<	idat.eFrame1 = &(atomRowData.electroFrame[atom1]);
1193	<	idat.eFrame2 = &(atomColData.electroFrame[atom2]);
1191	>	snap_->atomData.force[atom1] += *(idat.f1);
1192	>	snap_->atomData.force[atom2] -= *(idat.f1);
1193	>
1194	>	if (idat.doParticlePot) {
1195	>	// This is the pairwise contribution to the particle pot. The
1196	>	// embedding contribution is added in each of the low level
1197	>	// non-bonded routines. In parallel, this calculation is done
1198	>	// in collectData, not in unpackInteractionData.
1199	>	snap_->atomData.particlePot[atom1] += (idat.vpair) *(idat.sw);
1200	>	snap_->atomData.particlePot[atom2] += (idat.vpair) *(idat.sw);
1201		}
1202	<	if (storageLayout_ & DataStorage::dslTorque) {
1203	<	idat.t1 = &(atomRowData.torque[atom1]);
1204	<	idat.t2 = &(atomColData.torque[atom2]);
1202	>
1203	>	if (storageLayout_ & DataStorage::dslFlucQForce) {
1204	>	snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1205	>	snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1206		}
896	–	#else
897	–	idat.atypes = make_pair( ff_->getAtomType(identsLocal[atom1]),
898	–	ff_->getAtomType(identsLocal[atom2]) );
1207
1208	<	if (storageLayout_ & DataStorage::dslElectroFrame) {
1209	<	idat.eFrame1 = &(snap_->atomData.electroFrame[atom1]);
1210	<	idat.eFrame2 = &(snap_->atomData.electroFrame[atom2]);
1208	>	if (storageLayout_ & DataStorage::dslElectricField) {
1209	>	snap_->atomData.electricField[atom1] += *(idat.eField1);
1210	>	snap_->atomData.electricField[atom2] += *(idat.eField2);
1211		}
1212	<	if (storageLayout_ & DataStorage::dslTorque) {
1213	<	idat.t1 = &(snap_->atomData.torque[atom1]);
1214	<	idat.t2 = &(snap_->atomData.torque[atom2]);
907	<	}
908	<	#endif
1212	>
1213	>	#endif
1214	>
1215		}
1216
1217		/*
#	Line 918 \| Line 1224 \| namespace OpenMD {
1224
1225		vector<pair<int, int> > neighborList;
1226		groupCutoffs cuts;
1227	+	bool doAllPairs = false;
1228	+
1229		#ifdef IS_MPI
1230		cellListRow_.clear();
1231		cellListCol_.clear();
#	Line 937 \| Line 1245 \| namespace OpenMD {
1245		nCells_.y() = (int) ( Hy.length() )/ rList_;
1246		nCells_.z() = (int) ( Hz.length() )/ rList_;
1247
1248	+	// handle small boxes where the cell offsets can end up repeating cells
1249	+
1250	+	if (nCells_.x() < 3) doAllPairs = true;
1251	+	if (nCells_.y() < 3) doAllPairs = true;
1252	+	if (nCells_.z() < 3) doAllPairs = true;
1253	+
1254		Mat3x3d invHmat = snap_->getInvHmat();
1255		Vector3d rs, scaled, dr;
1256		Vector3i whichCell;
#	Line 950 \| Line 1264 \| namespace OpenMD {
1264		cellList_.resize(nCtot);
1265		#endif
1266
1267	+	if (!doAllPairs) {
1268		#ifdef IS_MPI
954	–	for (int i = 0; i < nGroupsInRow_; i++) {
955	–	rs = cgRowData.position[i];
1269
1270	<	// scaled positions relative to the box vectors
1271	<	scaled = invHmat * rs;
1272	<
1273	<	// wrap the vector back into the unit box by subtracting integer box
1274	<	// numbers
1275	<	for (int j = 0; j < 3; j++) {
1276	<	scaled[j] -= roundMe(scaled[j]);
1277	<	scaled[j] += 0.5;
1270	>	for (int i = 0; i < nGroupsInRow_; i++) {
1271	>	rs = cgRowData.position[i];
1272	>
1273	>	// scaled positions relative to the box vectors
1274	>	scaled = invHmat * rs;
1275	>
1276	>	// wrap the vector back into the unit box by subtracting integer box
1277	>	// numbers
1278	>	for (int j = 0; j < 3; j++) {
1279	>	scaled[j] -= roundMe(scaled[j]);
1280	>	scaled[j] += 0.5;
1281	>	}
1282	>
1283	>	// find xyz-indices of cell that cutoffGroup is in.
1284	>	whichCell.x() = nCells_.x() * scaled.x();
1285	>	whichCell.y() = nCells_.y() * scaled.y();
1286	>	whichCell.z() = nCells_.z() * scaled.z();
1287	>
1288	>	// find single index of this cell:
1289	>	cellIndex = Vlinear(whichCell, nCells_);
1290	>
1291	>	// add this cutoff group to the list of groups in this cell;
1292	>	cellListRow_[cellIndex].push_back(i);
1293		}
1294	<
1295	<	// find xyz-indices of cell that cutoffGroup is in.
1296	<	whichCell.x() = nCells_.x() * scaled.x();
1297	<	whichCell.y() = nCells_.y() * scaled.y();
1298	<	whichCell.z() = nCells_.z() * scaled.z();
1299	<
1300	<	// find single index of this cell:
1301	<	cellIndex = Vlinear(whichCell, nCells_);
1302	<
1303	<	// add this cutoff group to the list of groups in this cell;
1304	<	cellListRow_[cellIndex].push_back(i);
1305	<	}
1306	<
1307	<	for (int i = 0; i < nGroupsInCol_; i++) {
1308	<	rs = cgColData.position[i];
1309	<
1310	<	// scaled positions relative to the box vectors
1311	<	scaled = invHmat * rs;
1312	<
1313	<	// wrap the vector back into the unit box by subtracting integer box
1314	<	// numbers
1315	<	for (int j = 0; j < 3; j++) {
1316	<	scaled[j] -= roundMe(scaled[j]);
989	<	scaled[j] += 0.5;
1294	>	for (int i = 0; i < nGroupsInCol_; i++) {
1295	>	rs = cgColData.position[i];
1296	>
1297	>	// scaled positions relative to the box vectors
1298	>	scaled = invHmat * rs;
1299	>
1300	>	// wrap the vector back into the unit box by subtracting integer box
1301	>	// numbers
1302	>	for (int j = 0; j < 3; j++) {
1303	>	scaled[j] -= roundMe(scaled[j]);
1304	>	scaled[j] += 0.5;
1305	>	}
1306	>
1307	>	// find xyz-indices of cell that cutoffGroup is in.
1308	>	whichCell.x() = nCells_.x() * scaled.x();
1309	>	whichCell.y() = nCells_.y() * scaled.y();
1310	>	whichCell.z() = nCells_.z() * scaled.z();
1311	>
1312	>	// find single index of this cell:
1313	>	cellIndex = Vlinear(whichCell, nCells_);
1314	>
1315	>	// add this cutoff group to the list of groups in this cell;
1316	>	cellListCol_[cellIndex].push_back(i);
1317		}
1318	<
992	<	// find xyz-indices of cell that cutoffGroup is in.
993	<	whichCell.x() = nCells_.x() * scaled.x();
994	<	whichCell.y() = nCells_.y() * scaled.y();
995	<	whichCell.z() = nCells_.z() * scaled.z();
996	<
997	<	// find single index of this cell:
998	<	cellIndex = Vlinear(whichCell, nCells_);
999	<
1000	<	// add this cutoff group to the list of groups in this cell;
1001	<	cellListCol_[cellIndex].push_back(i);
1002	<	}
1318	>
1319		#else
1320	<	for (int i = 0; i < nGroups_; i++) {
1321	<	rs = snap_->cgData.position[i];
1322	<
1323	<	// scaled positions relative to the box vectors
1324	<	scaled = invHmat * rs;
1325	<
1326	<	// wrap the vector back into the unit box by subtracting integer box
1327	<	// numbers
1328	<	for (int j = 0; j < 3; j++) {
1329	<	scaled[j] -= roundMe(scaled[j]);
1330	<	scaled[j] += 0.5;
1320	>	for (int i = 0; i < nGroups_; i++) {
1321	>	rs = snap_->cgData.position[i];
1322	>
1323	>	// scaled positions relative to the box vectors
1324	>	scaled = invHmat * rs;
1325	>
1326	>	// wrap the vector back into the unit box by subtracting integer box
1327	>	// numbers
1328	>	for (int j = 0; j < 3; j++) {
1329	>	scaled[j] -= roundMe(scaled[j]);
1330	>	scaled[j] += 0.5;
1331	>	}
1332	>
1333	>	// find xyz-indices of cell that cutoffGroup is in.
1334	>	whichCell.x() = nCells_.x() * scaled.x();
1335	>	whichCell.y() = nCells_.y() * scaled.y();
1336	>	whichCell.z() = nCells_.z() * scaled.z();
1337	>
1338	>	// find single index of this cell:
1339	>	cellIndex = Vlinear(whichCell, nCells_);
1340	>
1341	>	// add this cutoff group to the list of groups in this cell;
1342	>	cellList_[cellIndex].push_back(i);
1343		}
1344
1017	–	// find xyz-indices of cell that cutoffGroup is in.
1018	–	whichCell.x() = nCells_.x() * scaled.x();
1019	–	whichCell.y() = nCells_.y() * scaled.y();
1020	–	whichCell.z() = nCells_.z() * scaled.z();
1021	–
1022	–	// find single index of this cell:
1023	–	cellIndex = Vlinear(whichCell, nCells_);
1024	–
1025	–	// add this cutoff group to the list of groups in this cell;
1026	–	cellList_[cellIndex].push_back(i);
1027	–	}
1345		#endif
1346
1347	<	for (int m1z = 0; m1z < nCells_.z(); m1z++) {
1348	<	for (int m1y = 0; m1y < nCells_.y(); m1y++) {
1349	<	for (int m1x = 0; m1x < nCells_.x(); m1x++) {
1350	<	Vector3i m1v(m1x, m1y, m1z);
1351	<	int m1 = Vlinear(m1v, nCells_);
1035	<
1036	<	for (vector<Vector3i>::iterator os = cellOffsets_.begin();
1037	<	os != cellOffsets_.end(); ++os) {
1347	>	for (int m1z = 0; m1z < nCells_.z(); m1z++) {
1348	>	for (int m1y = 0; m1y < nCells_.y(); m1y++) {
1349	>	for (int m1x = 0; m1x < nCells_.x(); m1x++) {
1350	>	Vector3i m1v(m1x, m1y, m1z);
1351	>	int m1 = Vlinear(m1v, nCells_);
1352
1353	<	Vector3i m2v = m1v + (*os);
1354	<
1355	<	if (m2v.x() >= nCells_.x()) {
1356	<	m2v.x() = 0;
1357	<	} else if (m2v.x() < 0) {
1044	<	m2v.x() = nCells_.x() - 1;
1045	<	}
1046	<
1047	<	if (m2v.y() >= nCells_.y()) {
1048	<	m2v.y() = 0;
1049	<	} else if (m2v.y() < 0) {
1050	<	m2v.y() = nCells_.y() - 1;
1051	<	}
1052	<
1053	<	if (m2v.z() >= nCells_.z()) {
1054	<	m2v.z() = 0;
1055	<	} else if (m2v.z() < 0) {
1056	<	m2v.z() = nCells_.z() - 1;
1057	<	}
1058	<
1059	<	int m2 = Vlinear (m2v, nCells_);
1353	>	for (vector<Vector3i>::iterator os = cellOffsets_.begin();
1354	>	os != cellOffsets_.end(); ++os) {
1355	>
1356	>	Vector3i m2v = m1v + (*os);
1357	>
1358
1359	<	#ifdef IS_MPI
1360	<	for (vector<int>::iterator j1 = cellListRow_[m1].begin();
1361	<	j1 != cellListRow_[m1].end(); ++j1) {
1362	<	for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1363	<	j2 != cellListCol_[m2].end(); ++j2) {
1364	<
1365	<	// Always do this if we're in different cells or if
1366	<	// we're in the same cell and the global index of the
1367	<	// j2 cutoff group is less than the j1 cutoff group
1359	>	if (m2v.x() >= nCells_.x()) {
1360	>	m2v.x() = 0;
1361	>	} else if (m2v.x() < 0) {
1362	>	m2v.x() = nCells_.x() - 1;
1363	>	}
1364	>
1365	>	if (m2v.y() >= nCells_.y()) {
1366	>	m2v.y() = 0;
1367	>	} else if (m2v.y() < 0) {
1368	>	m2v.y() = nCells_.y() - 1;
1369	>	}
1370	>
1371	>	if (m2v.z() >= nCells_.z()) {
1372	>	m2v.z() = 0;
1373	>	} else if (m2v.z() < 0) {
1374	>	m2v.z() = nCells_.z() - 1;
1375	>	}
1376
1377	<	if (m2 != m1 \|\| cgColToGlobal[(j2)] < cgRowToGlobal[(j1)]) {
1377	>	int m2 = Vlinear (m2v, nCells_);
1378	>
1379	>	#ifdef IS_MPI
1380	>	for (vector<int>::iterator j1 = cellListRow_[m1].begin();
1381	>	j1 != cellListRow_[m1].end(); ++j1) {
1382	>	for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1383	>	j2 != cellListCol_[m2].end(); ++j2) {
1384	>
1385	>	// In parallel, we need to visit all pairs of row
1386	>	// & column indicies and will divide labor in the
1387	>	// force evaluation later.
1388		dr = cgColData.position[(j2)] - cgRowData.position[(j1)];
1389		snap_->wrapVector(dr);
1390		cuts = getGroupCutoffs( (j1), (j2) );
1391		if (dr.lengthSquare() < cuts.third) {
1392		neighborList.push_back(make_pair((j1), (j2)));
1393	<	}
1393	>	}
1394		}
1395		}
1080	–	}
1396		#else
1397	+	for (vector<int>::iterator j1 = cellList_[m1].begin();
1398	+	j1 != cellList_[m1].end(); ++j1) {
1399	+	for (vector<int>::iterator j2 = cellList_[m2].begin();
1400	+	j2 != cellList_[m2].end(); ++j2) {
1401	+
1402	+	// Always do this if we're in different cells or if
1403	+	// we're in the same cell and the global index of
1404	+	// the j2 cutoff group is greater than or equal to
1405	+	// the j1 cutoff group. Note that Rappaport's code
1406	+	// has a "less than" conditional here, but that
1407	+	// deals with atom-by-atom computation. OpenMD
1408	+	// allows atoms within a single cutoff group to
1409	+	// interact with each other.
1410
1083	–	for (vector<int>::iterator j1 = cellList_[m1].begin();
1084	–	j1 != cellList_[m1].end(); ++j1) {
1085	–	for (vector<int>::iterator j2 = cellList_[m2].begin();
1086	–	j2 != cellList_[m2].end(); ++j2) {
1411
1088	–	// Always do this if we're in different cells or if
1089	–	// we're in the same cell and the global index of the
1090	–	// j2 cutoff group is less than the j1 cutoff group
1412
1413	<	if (m2 != m1 \|\| (j2) < (j1)) {
1414	<	dr = snap_->cgData.position[(j2)] - snap_->cgData.position[(j1)];
1415	<	snap_->wrapVector(dr);
1416	<	cuts = getGroupCutoffs( (j1), (j2) );
1417	<	if (dr.lengthSquare() < cuts.third) {
1418	<	neighborList.push_back(make_pair((j1), (j2)));
1413	>	if (m2 != m1 \|\| (j2) >= (j1) ) {
1414	>
1415	>	dr = snap_->cgData.position[(j2)] - snap_->cgData.position[(j1)];
1416	>	snap_->wrapVector(dr);
1417	>	cuts = getGroupCutoffs( (j1), (j2) );
1418	>	if (dr.lengthSquare() < cuts.third) {
1419	>	neighborList.push_back(make_pair((j1), (j2)));
1420	>	}
1421		}
1422		}
1423		}
1101	–	}
1424		#endif
1425	+	}
1426		}
1427		}
1428		}
1429	+	} else {
1430	+	// branch to do all cutoff group pairs
1431	+	#ifdef IS_MPI
1432	+	for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1433	+	for (int j2 = 0; j2 < nGroupsInCol_; j2++) {
1434	+	dr = cgColData.position[j2] - cgRowData.position[j1];
1435	+	snap_->wrapVector(dr);
1436	+	cuts = getGroupCutoffs( j1, j2 );
1437	+	if (dr.lengthSquare() < cuts.third) {
1438	+	neighborList.push_back(make_pair(j1, j2));
1439	+	}
1440	+	}
1441	+	}
1442	+	#else
1443	+	// include all groups here.
1444	+	for (int j1 = 0; j1 < nGroups_; j1++) {
1445	+	// include self group interactions j2 == j1
1446	+	for (int j2 = j1; j2 < nGroups_; j2++) {
1447	+	dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1448	+	snap_->wrapVector(dr);
1449	+	cuts = getGroupCutoffs( j1, j2 );
1450	+	if (dr.lengthSquare() < cuts.third) {
1451	+	neighborList.push_back(make_pair(j1, j2));
1452	+	}
1453	+	}
1454	+	}
1455	+	#endif
1456		}
1457	<
1457	>
1458		// save the local cutoff group positions for the check that is
1459		// done on each loop:
1460		saved_CG_positions_.clear();

Diff Legend

-–
+Removed lines
-+
+Added lines
-<
+Changed lines
->
+Changed lines

Comparing branches/development/src/parallel/ForceMatrixDecomposition.cpp (file contents): Revision 1582 by gezelter, Tue Jun 14 20:41:44 2011 UTC vs. Revision 1755 by gezelter, Thu Jun 14 01:58:35 2012 UTC

Diff Legend

Comparing branches/development/src/parallel/ForceMatrixDecomposition.cpp (file contents):
Revision 1582 by gezelter, Tue Jun 14 20:41:44 2011 UTC vs.
Revision 1755 by gezelter, Thu Jun 14 01:58:35 2012 UTC