ViewVC Help

View File | Revision Log | Show Annotations | View Changeset | Root Listing

root/OpenMD/trunk/src/parallel/ForceMatrixDecomposition.cpp

Comparing branches/development/src/parallel/ForceMatrixDecomposition.cpp (file contents):
Revision 1554 by gezelter, Sat Apr 30 02:54:02 2011 UTC vs.
Revision 1760 by gezelter, Thu Jun 21 19:26:46 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"
44		#include "nonbonded/NonBondedInteraction.hpp"
45		#include "brains/SnapshotManager.hpp"
46	+	#include "brains/PairList.hpp"
47
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		*/
53	–
93		void ForceMatrixDecomposition::distributeInitialData() {
94		snap_ = sman_->getCurrentSnapshot();
95		storageLayout_ = sman_->getStorageLayout();
96	<	#ifdef IS_MPI
97	<	int nLocal = snap_->getNumberOfAtoms();
98	<	int nGroups = snap_->getNumberOfCutoffGroups();
96	>	ff_ = info_->getForceField();
97	>	nLocal_ = snap_->getNumberOfAtoms();
98	>
99	>	nGroups_ = info_->getNLocalCutoffGroups();
100	>	// gather the information for atomtype IDs (atids):
101	>	idents = info_->getIdentArray();
102	>	AtomLocalToGlobal = info_->getGlobalAtomIndices();
103	>	cgLocalToGlobal = info_->getGlobalGroupIndices();
104	>	vector<int> globalGroupMembership = info_->getGlobalGroupMembership();
105	>
106	>	massFactors = info_->getMassFactors();
107	>
108	>	PairList* excludes = info_->getExcludedInteractions();
109	>	PairList* oneTwo = info_->getOneTwoInteractions();
110	>	PairList* oneThree = info_->getOneThreeInteractions();
111	>	PairList* oneFour = info_->getOneFourInteractions();
112
113	<	AtomCommIntRow = new Communicator<Row,int>(nLocal);
114	<	AtomCommRealRow = new Communicator<Row,RealType>(nLocal);
115	<	AtomCommVectorRow = new Communicator<Row,Vector3d>(nLocal);
116	<	AtomCommMatrixRow = new Communicator<Row,Mat3x3d>(nLocal);
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	>	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);
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	<	int nAtomsInRow = AtomCommIntRow->getSize();
137	<	int nAtomsInCol = AtomCommIntColumn->getSize();
138	<	int nGroupsInRow = cgCommIntRow->getSize();
139	<	int 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);
147	>	atomRowData.resize(nAtomsInRow_);
148		atomRowData.setStorageLayout(storageLayout_);
149	<	atomColData.resize(nAtomsInCol);
149	>	atomColData.resize(nAtomsInCol_);
150		atomColData.setStorageLayout(storageLayout_);
151	<	cgRowData.resize(nGroupsInRow);
151	>	cgRowData.resize(nGroupsInRow_);
152		cgRowData.setStorageLayout(DataStorage::dslPosition);
153	<	cgColData.resize(nGroupsInCol);
154	<	cgColData.setStorageLayout(DataStorage::dslPosition);
153	>	cgColData.resize(nGroupsInCol_);
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	<	vector<vector<RealType> > pot_row(N_INTERACTION_FAMILIES,
165	<	vector<RealType> (nAtomsInRow, 0.0));
166	<	vector<vector<RealType> > pot_col(N_INTERACTION_FAMILIES,
167	<	vector<RealType> (nAtomsInCol, 0.0));
164	>	AtomPlanIntRow->gather(idents, identsRow);
165	>	AtomPlanIntColumn->gather(idents, identsCol);
166	>
167	>	// allocate memory for the parallel objects
168	>	atypesRow.resize(nAtomsInRow_);
169	>	atypesCol.resize(nAtomsInCol_);
170
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	<	vector<RealType> pot_local(N_INTERACTION_FAMILIES, 0.0);
176	>	pot_row.resize(nAtomsInRow_);
177	>	pot_col.resize(nAtomsInCol_);
178	>
179	>	expot_row.resize(nAtomsInRow_);
180	>	expot_col.resize(nAtomsInCol_);
181	>
182	>	AtomRowToGlobal.resize(nAtomsInRow_);
183	>	AtomColToGlobal.resize(nAtomsInCol_);
184	>	AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
185	>	AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
186	>
187	>	cgRowToGlobal.resize(nGroupsInRow_);
188	>	cgColToGlobal.resize(nGroupsInCol_);
189	>	cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
190	>	cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
191	>
192	>	massFactorsRow.resize(nAtomsInRow_);
193	>	massFactorsCol.resize(nAtomsInCol_);
194	>	AtomPlanRealRow->gather(massFactors, massFactorsRow);
195	>	AtomPlanRealColumn->gather(massFactors, massFactorsCol);
196	>
197	>	groupListRow_.clear();
198	>	groupListRow_.resize(nGroupsInRow_);
199	>	for (int i = 0; i < nGroupsInRow_; i++) {
200	>	int gid = cgRowToGlobal[i];
201	>	for (int j = 0; j < nAtomsInRow_; j++) {
202	>	int aid = AtomRowToGlobal[j];
203	>	if (globalGroupMembership[aid] == gid)
204	>	groupListRow_[i].push_back(j);
205	>	}
206	>	}
207	>
208	>	groupListCol_.clear();
209	>	groupListCol_.resize(nGroupsInCol_);
210	>	for (int i = 0; i < nGroupsInCol_; i++) {
211	>	int gid = cgColToGlobal[i];
212	>	for (int j = 0; j < nAtomsInCol_; j++) {
213	>	int aid = AtomColToGlobal[j];
214	>	if (globalGroupMembership[aid] == gid)
215	>	groupListCol_[i].push_back(j);
216	>	}
217	>	}
218	>
219	>	excludesForAtom.clear();
220	>	excludesForAtom.resize(nAtomsInRow_);
221	>	toposForAtom.clear();
222	>	toposForAtom.resize(nAtomsInRow_);
223	>	topoDist.clear();
224	>	topoDist.resize(nAtomsInRow_);
225	>	for (int i = 0; i < nAtomsInRow_; i++) {
226	>	int iglob = AtomRowToGlobal[i];
227	>
228	>	for (int j = 0; j < nAtomsInCol_; j++) {
229	>	int jglob = AtomColToGlobal[j];
230	>
231	>	if (excludes->hasPair(iglob, jglob))
232	>	excludesForAtom[i].push_back(j);
233	>
234	>	if (oneTwo->hasPair(iglob, jglob)) {
235	>	toposForAtom[i].push_back(j);
236	>	topoDist[i].push_back(1);
237	>	} else {
238	>	if (oneThree->hasPair(iglob, jglob)) {
239	>	toposForAtom[i].push_back(j);
240	>	topoDist[i].push_back(2);
241	>	} else {
242	>	if (oneFour->hasPair(iglob, jglob)) {
243	>	toposForAtom[i].push_back(j);
244	>	topoDist[i].push_back(3);
245	>	}
246	>	}
247	>	}
248	>	}
249	>	}
250	>
251	>	#else
252	>	excludesForAtom.clear();
253	>	excludesForAtom.resize(nLocal_);
254	>	toposForAtom.clear();
255	>	toposForAtom.resize(nLocal_);
256	>	topoDist.clear();
257	>	topoDist.resize(nLocal_);
258	>
259	>	for (int i = 0; i < nLocal_; i++) {
260	>	int iglob = AtomLocalToGlobal[i];
261	>
262	>	for (int j = 0; j < nLocal_; j++) {
263	>	int jglob = AtomLocalToGlobal[j];
264	>
265	>	if (excludes->hasPair(iglob, jglob))
266	>	excludesForAtom[i].push_back(j);
267	>
268	>	if (oneTwo->hasPair(iglob, jglob)) {
269	>	toposForAtom[i].push_back(j);
270	>	topoDist[i].push_back(1);
271	>	} else {
272	>	if (oneThree->hasPair(iglob, jglob)) {
273	>	toposForAtom[i].push_back(j);
274	>	topoDist[i].push_back(2);
275	>	} else {
276	>	if (oneFour->hasPair(iglob, jglob)) {
277	>	toposForAtom[i].push_back(j);
278	>	topoDist[i].push_back(3);
279	>	}
280	>	}
281	>	}
282	>	}
283	>	}
284	>	#endif
285	>
286	>	// allocate memory for the parallel objects
287	>	atypesLocal.resize(nLocal_);
288	>
289	>	for (int i = 0; i < nLocal_; i++)
290	>	atypesLocal[i] = ff_->getAtomType(idents[i]);
291	>
292	>	groupList_.clear();
293	>	groupList_.resize(nGroups_);
294	>	for (int i = 0; i < nGroups_; i++) {
295	>	int gid = cgLocalToGlobal[i];
296	>	for (int j = 0; j < nLocal_; j++) {
297	>	int aid = AtomLocalToGlobal[j];
298	>	if (globalGroupMembership[aid] == gid) {
299	>	groupList_[i].push_back(j);
300	>	}
301	>	}
302	>	}
303	>
304	>
305	>	createGtypeCutoffMap();
306	>
307	>	}
308	>
309	>	void ForceMatrixDecomposition::createGtypeCutoffMap() {
310
311	<	// gather the information for atomtype IDs (atids):
312	<	vector<int> identsLocal = info_->getIdentArray();
313	<	identsRow.reserve(nAtomsInRow);
314	<	identsCol.reserve(nAtomsInCol);
311	>	RealType tol = 1e-6;
312	>	largestRcut_ = 0.0;
313	>	RealType rc;
314	>	int atid;
315	>	set<AtomType*> atypes = info_->getSimulatedAtomTypes();
316
317	<	AtomCommIntRow->gather(identsLocal, identsRow);
318	<	AtomCommIntColumn->gather(identsLocal, identsCol);
317	>	map<int, RealType> atypeCutoff;
318	>
319	>	for (set<AtomType*>::iterator at = atypes.begin();
320	>	at != atypes.end(); ++at){
321	>	atid = (*at)->getIdent();
322	>	if (userChoseCutoff_)
323	>	atypeCutoff[atid] = userCutoff_;
324	>	else
325	>	atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at);
326	>	}
327
328	<	AtomLocalToGlobal = info_->getGlobalAtomIndices();
329	<	AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
330	<	AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
331	<
332	<	cgLocalToGlobal = info_->getGlobalGroupIndices();
333	<	cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
334	<	cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
328	>	vector<RealType> gTypeCutoffs;
329	>	// first we do a single loop over the cutoff groups to find the
330	>	// largest cutoff for any atypes present in this group.
331	>	#ifdef IS_MPI
332	>	vector<RealType> groupCutoffRow(nGroupsInRow_, 0.0);
333	>	groupRowToGtype.resize(nGroupsInRow_);
334	>	for (int cg1 = 0; cg1 < nGroupsInRow_; cg1++) {
335	>	vector<int> atomListRow = getAtomsInGroupRow(cg1);
336	>	for (vector<int>::iterator ia = atomListRow.begin();
337	>	ia != atomListRow.end(); ++ia) {
338	>	int atom1 = (*ia);
339	>	atid = identsRow[atom1];
340	>	if (atypeCutoff[atid] > groupCutoffRow[cg1]) {
341	>	groupCutoffRow[cg1] = atypeCutoff[atid];
342	>	}
343	>	}
344
345	<	// still need:
346	<	// topoDist
347	<	// exclude
345	>	bool gTypeFound = false;
346	>	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
347	>	if (abs(groupCutoffRow[cg1] - gTypeCutoffs[gt]) < tol) {
348	>	groupRowToGtype[cg1] = gt;
349	>	gTypeFound = true;
350	>	}
351	>	}
352	>	if (!gTypeFound) {
353	>	gTypeCutoffs.push_back( groupCutoffRow[cg1] );
354	>	groupRowToGtype[cg1] = gTypeCutoffs.size() - 1;
355	>	}
356	>
357	>	}
358	>	vector<RealType> groupCutoffCol(nGroupsInCol_, 0.0);
359	>	groupColToGtype.resize(nGroupsInCol_);
360	>	for (int cg2 = 0; cg2 < nGroupsInCol_; cg2++) {
361	>	vector<int> atomListCol = getAtomsInGroupColumn(cg2);
362	>	for (vector<int>::iterator jb = atomListCol.begin();
363	>	jb != atomListCol.end(); ++jb) {
364	>	int atom2 = (*jb);
365	>	atid = identsCol[atom2];
366	>	if (atypeCutoff[atid] > groupCutoffCol[cg2]) {
367	>	groupCutoffCol[cg2] = atypeCutoff[atid];
368	>	}
369	>	}
370	>	bool gTypeFound = false;
371	>	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
372	>	if (abs(groupCutoffCol[cg2] - gTypeCutoffs[gt]) < tol) {
373	>	groupColToGtype[cg2] = gt;
374	>	gTypeFound = true;
375	>	}
376	>	}
377	>	if (!gTypeFound) {
378	>	gTypeCutoffs.push_back( groupCutoffCol[cg2] );
379	>	groupColToGtype[cg2] = gTypeCutoffs.size() - 1;
380	>	}
381	>	}
382	>	#else
383	>
384	>	vector<RealType> groupCutoff(nGroups_, 0.0);
385	>	groupToGtype.resize(nGroups_);
386	>	for (int cg1 = 0; cg1 < nGroups_; cg1++) {
387	>	groupCutoff[cg1] = 0.0;
388	>	vector<int> atomList = getAtomsInGroupRow(cg1);
389	>	for (vector<int>::iterator ia = atomList.begin();
390	>	ia != atomList.end(); ++ia) {
391	>	int atom1 = (*ia);
392	>	atid = idents[atom1];
393	>	if (atypeCutoff[atid] > groupCutoff[cg1])
394	>	groupCutoff[cg1] = atypeCutoff[atid];
395	>	}
396	>
397	>	bool gTypeFound = false;
398	>	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
399	>	if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
400	>	groupToGtype[cg1] = gt;
401	>	gTypeFound = true;
402	>	}
403	>	}
404	>	if (!gTypeFound) {
405	>	gTypeCutoffs.push_back( groupCutoff[cg1] );
406	>	groupToGtype[cg1] = gTypeCutoffs.size() - 1;
407	>	}
408	>	}
409		#endif
410	+
411	+	// Now we find the maximum group cutoff value present in the simulation
412	+
413	+	RealType groupMax = *max_element(gTypeCutoffs.begin(),
414	+	gTypeCutoffs.end());
415	+
416	+	#ifdef IS_MPI
417	+	MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE,
418	+	MPI::MAX);
419	+	#endif
420	+
421	+	RealType tradRcut = groupMax;
422	+
423	+	for (int i = 0; i < gTypeCutoffs.size();  i++) {
424	+	for (int j = 0; j < gTypeCutoffs.size();  j++) {
425	+	RealType thisRcut;
426	+	switch(cutoffPolicy_) {
427	+	case TRADITIONAL:
428	+	thisRcut = tradRcut;
429	+	break;
430	+	case MIX:
431	+	thisRcut = 0.5 * (gTypeCutoffs[i] + gTypeCutoffs[j]);
432	+	break;
433	+	case MAX:
434	+	thisRcut = max(gTypeCutoffs[i], gTypeCutoffs[j]);
435	+	break;
436	+	default:
437	+	sprintf(painCave.errMsg,
438	+	"ForceMatrixDecomposition::createGtypeCutoffMap "
439	+	"hit an unknown cutoff policy!\n");
440	+	painCave.severity = OPENMD_ERROR;
441	+	painCave.isFatal = 1;
442	+	simError();
443	+	break;
444	+	}
445	+
446	+	pair<int,int> key = make_pair(i,j);
447	+	gTypeCutoffMap[key].first = thisRcut;
448	+	if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
449	+	gTypeCutoffMap[key].second = thisRcut*thisRcut;
450	+	gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
451	+	// sanity check
452	+
453	+	if (userChoseCutoff_) {
454	+	if (abs(gTypeCutoffMap[key].first - userCutoff_) > 0.0001) {
455	+	sprintf(painCave.errMsg,
456	+	"ForceMatrixDecomposition::createGtypeCutoffMap "
457	+	"user-specified rCut (%lf) does not match computed group Cutoff\n", userCutoff_);
458	+	painCave.severity = OPENMD_ERROR;
459	+	painCave.isFatal = 1;
460	+	simError();
461	+	}
462	+	}
463	+	}
464	+	}
465		}
466	+
467	+	groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
468	+	int i, j;
469	+	#ifdef IS_MPI
470	+	i = groupRowToGtype[cg1];
471	+	j = groupColToGtype[cg2];
472	+	#else
473	+	i = groupToGtype[cg1];
474	+	j = groupToGtype[cg2];
475	+	#endif
476	+	return gTypeCutoffMap[make_pair(i,j)];
477	+	}
478	+
479	+	int ForceMatrixDecomposition::getTopologicalDistance(int atom1, int atom2) {
480	+	for (int j = 0; j < toposForAtom[atom1].size(); j++) {
481	+	if (toposForAtom[atom1][j] == atom2)
482	+	return topoDist[atom1][j];
483	+	}
484	+	return 0;
485	+	}
486	+
487	+	void ForceMatrixDecomposition::zeroWorkArrays() {
488	+	pairwisePot = 0.0;
489	+	embeddingPot = 0.0;
490	+	excludedPot = 0.0;
491	+
492	+	#ifdef IS_MPI
493	+	if (storageLayout_ & DataStorage::dslForce) {
494	+	fill(atomRowData.force.begin(), atomRowData.force.end(), V3Zero);
495	+	fill(atomColData.force.begin(), atomColData.force.end(), V3Zero);
496	+	}
497	+
498	+	if (storageLayout_ & DataStorage::dslTorque) {
499	+	fill(atomRowData.torque.begin(), atomRowData.torque.end(), V3Zero);
500	+	fill(atomColData.torque.begin(), atomColData.torque.end(), V3Zero);
501	+	}
502
503	+	fill(pot_row.begin(), pot_row.end(),
504	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
505
506	+	fill(pot_col.begin(), pot_col.end(),
507	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
508
509	+	fill(expot_row.begin(), expot_row.end(),
510	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
511	+
512	+	fill(expot_col.begin(), expot_col.end(),
513	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
514	+
515	+	if (storageLayout_ & DataStorage::dslParticlePot) {
516	+	fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(),
517	+	0.0);
518	+	fill(atomColData.particlePot.begin(), atomColData.particlePot.end(),
519	+	0.0);
520	+	}
521	+
522	+	if (storageLayout_ & DataStorage::dslDensity) {
523	+	fill(atomRowData.density.begin(), atomRowData.density.end(), 0.0);
524	+	fill(atomColData.density.begin(), atomColData.density.end(), 0.0);
525	+	}
526	+
527	+	if (storageLayout_ & DataStorage::dslFunctional) {
528	+	fill(atomRowData.functional.begin(), atomRowData.functional.end(),
529	+	0.0);
530	+	fill(atomColData.functional.begin(), atomColData.functional.end(),
531	+	0.0);
532	+	}
533	+
534	+	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
535	+	fill(atomRowData.functionalDerivative.begin(),
536	+	atomRowData.functionalDerivative.end(), 0.0);
537	+	fill(atomColData.functionalDerivative.begin(),
538	+	atomColData.functionalDerivative.end(), 0.0);
539	+	}
540	+
541	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
542	+	fill(atomRowData.skippedCharge.begin(),
543	+	atomRowData.skippedCharge.end(), 0.0);
544	+	fill(atomColData.skippedCharge.begin(),
545	+	atomColData.skippedCharge.end(), 0.0);
546	+	}
547	+
548	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
549	+	fill(atomRowData.flucQFrc.begin(),
550	+	atomRowData.flucQFrc.end(), 0.0);
551	+	fill(atomColData.flucQFrc.begin(),
552	+	atomColData.flucQFrc.end(), 0.0);
553	+	}
554	+
555	+	if (storageLayout_ & DataStorage::dslElectricField) {
556	+	fill(atomRowData.electricField.begin(),
557	+	atomRowData.electricField.end(), V3Zero);
558	+	fill(atomColData.electricField.begin(),
559	+	atomColData.electricField.end(), V3Zero);
560	+	}
561	+
562	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
563	+	fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
564	+	0.0);
565	+	fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
566	+	0.0);
567	+	}
568	+
569	+	#endif
570	+	// even in parallel, we need to zero out the local arrays:
571	+
572	+	if (storageLayout_ & DataStorage::dslParticlePot) {
573	+	fill(snap_->atomData.particlePot.begin(),
574	+	snap_->atomData.particlePot.end(), 0.0);
575	+	}
576	+
577	+	if (storageLayout_ & DataStorage::dslDensity) {
578	+	fill(snap_->atomData.density.begin(),
579	+	snap_->atomData.density.end(), 0.0);
580	+	}
581	+
582	+	if (storageLayout_ & DataStorage::dslFunctional) {
583	+	fill(snap_->atomData.functional.begin(),
584	+	snap_->atomData.functional.end(), 0.0);
585	+	}
586	+
587	+	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
588	+	fill(snap_->atomData.functionalDerivative.begin(),
589	+	snap_->atomData.functionalDerivative.end(), 0.0);
590	+	}
591	+
592	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
593	+	fill(snap_->atomData.skippedCharge.begin(),
594	+	snap_->atomData.skippedCharge.end(), 0.0);
595	+	}
596	+
597	+	if (storageLayout_ & DataStorage::dslElectricField) {
598	+	fill(snap_->atomData.electricField.begin(),
599	+	snap_->atomData.electricField.end(), V3Zero);
600	+	}
601	+	}
602	+
603	+
604		void ForceMatrixDecomposition::distributeData()  {
605		snap_ = sman_->getCurrentSnapshot();
606		storageLayout_ = sman_->getStorageLayout();
607		#ifdef IS_MPI
608
609		// gather up the atomic positions
610	<	AtomCommVectorRow->gather(snap_->atomData.position,
610	>	AtomPlanVectorRow->gather(snap_->atomData.position,
611		atomRowData.position);
612	<	AtomCommVectorColumn->gather(snap_->atomData.position,
612	>	AtomPlanVectorColumn->gather(snap_->atomData.position,
613		atomColData.position);
614
615		// gather up the cutoff group positions
616	<	cgCommVectorRow->gather(snap_->cgData.position,
616	>
617	>	cgPlanVectorRow->gather(snap_->cgData.position,
618		cgRowData.position);
619	<	cgCommVectorColumn->gather(snap_->cgData.position,
619	>
620	>	cgPlanVectorColumn->gather(snap_->cgData.position,
621		cgColData.position);
622	+
623	+
624	+
625	+	if (needVelocities_) {
626	+	// gather up the atomic velocities
627	+	AtomPlanVectorColumn->gather(snap_->atomData.velocity,
628	+	atomColData.velocity);
629	+
630	+	cgPlanVectorColumn->gather(snap_->cgData.velocity,
631	+	cgColData.velocity);
632	+	}
633	+
634
635		// if needed, gather the atomic rotation matrices
636		if (storageLayout_ & DataStorage::dslAmat) {
637	<	AtomCommMatrixRow->gather(snap_->atomData.aMat,
637	>	AtomPlanMatrixRow->gather(snap_->atomData.aMat,
638		atomRowData.aMat);
639	<	AtomCommMatrixColumn->gather(snap_->atomData.aMat,
639	>	AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
640		atomColData.aMat);
641		}
642
643		// if needed, gather the atomic eletrostatic frames
644		if (storageLayout_ & DataStorage::dslElectroFrame) {
645	<	AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
645	>	AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
646		atomRowData.electroFrame);
647	<	AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
647	>	AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
648		atomColData.electroFrame);
649		}
650	+
651	+	// if needed, gather the atomic fluctuating charge values
652	+	if (storageLayout_ & DataStorage::dslFlucQPosition) {
653	+	AtomPlanRealRow->gather(snap_->atomData.flucQPos,
654	+	atomRowData.flucQPos);
655	+	AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
656	+	atomColData.flucQPos);
657	+	}
658	+
659		#endif
660		}
661
662	+	/* collects information obtained during the pre-pair loop onto local
663	+	* data structures.
664	+	*/
665		void ForceMatrixDecomposition::collectIntermediateData() {
666		snap_ = sman_->getCurrentSnapshot();
667		storageLayout_ = sman_->getStorageLayout();
#	Line 162 | Line 669 | namespace OpenMD {
669
670		if (storageLayout_ & DataStorage::dslDensity) {
671
672	<	AtomCommRealRow->scatter(atomRowData.density,
672	>	AtomPlanRealRow->scatter(atomRowData.density,
673		snap_->atomData.density);
674
675		int n = snap_->atomData.density.size();
676	<	std::vector<RealType> rho_tmp(n, 0.0);
677	<	AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
676	>	vector<RealType> rho_tmp(n, 0.0);
677	>	AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
678		for (int i = 0; i < n; i++)
679		snap_->atomData.density[i] += rho_tmp[i];
680		}
681	+
682	+	if (storageLayout_ & DataStorage::dslElectricField) {
683	+
684	+	AtomPlanVectorRow->scatter(atomRowData.electricField,
685	+	snap_->atomData.electricField);
686	+
687	+	int n = snap_->atomData.electricField.size();
688	+	vector<Vector3d> field_tmp(n, V3Zero);
689	+	AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
690	+	for (int i = 0; i < n; i++)
691	+	snap_->atomData.electricField[i] += field_tmp[i];
692	+	}
693		#endif
694		}
695	<
695	>
696	>	/*
697	>	* redistributes information obtained during the pre-pair loop out to
698	>	* row and column-indexed data structures
699	>	*/
700		void ForceMatrixDecomposition::distributeIntermediateData() {
701		snap_ = sman_->getCurrentSnapshot();
702		storageLayout_ = sman_->getStorageLayout();
703		#ifdef IS_MPI
704		if (storageLayout_ & DataStorage::dslFunctional) {
705	<	AtomCommRealRow->gather(snap_->atomData.functional,
705	>	AtomPlanRealRow->gather(snap_->atomData.functional,
706		atomRowData.functional);
707	<	AtomCommRealColumn->gather(snap_->atomData.functional,
707	>	AtomPlanRealColumn->gather(snap_->atomData.functional,
708		atomColData.functional);
709		}
710
711		if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
712	<	AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
712	>	AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
713		atomRowData.functionalDerivative);
714	<	AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
714	>	AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
715		atomColData.functionalDerivative);
716		}
717		#endif
#	Line 202 | Line 725 | namespace OpenMD {
725		int n = snap_->atomData.force.size();
726		vector<Vector3d> frc_tmp(n, V3Zero);
727
728	<	AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
728	>	AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
729		for (int i = 0; i < n; i++) {
730		snap_->atomData.force[i] += frc_tmp[i];
731		frc_tmp[i] = 0.0;
732		}
733
734	<	AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
735	<	for (int i = 0; i < n; i++)
734	>	AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
735	>	for (int i = 0; i < n; i++) {
736		snap_->atomData.force[i] += frc_tmp[i];
737	<
738	<
737	>	}
738	>
739		if (storageLayout_ & DataStorage::dslTorque) {
740
741	<	int nt = snap_->atomData.force.size();
741	>	int nt = snap_->atomData.torque.size();
742		vector<Vector3d> trq_tmp(nt, V3Zero);
743
744	<	AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
745	<	for (int i = 0; i < n; i++) {
744	>	AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
745	>	for (int i = 0; i < nt; i++) {
746		snap_->atomData.torque[i] += trq_tmp[i];
747		trq_tmp[i] = 0.0;
748		}
749
750	<	AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
751	<	for (int i = 0; i < n; i++)
750	>	AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
751	>	for (int i = 0; i < nt; i++)
752		snap_->atomData.torque[i] += trq_tmp[i];
753		}
231	–
232	–	int nLocal = snap_->getNumberOfAtoms();
754
755	<	vector<vector<RealType> > pot_temp(N_INTERACTION_FAMILIES,
756	<	vector<RealType> (nLocal, 0.0));
755	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
756	>
757	>	int ns = snap_->atomData.skippedCharge.size();
758	>	vector<RealType> skch_tmp(ns, 0.0);
759	>
760	>	AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
761	>	for (int i = 0; i < ns; i++) {
762	>	snap_->atomData.skippedCharge[i] += skch_tmp[i];
763	>	skch_tmp[i] = 0.0;
764	>	}
765	>
766	>	AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
767	>	for (int i = 0; i < ns; i++)
768	>	snap_->atomData.skippedCharge[i] += skch_tmp[i];
769	>
770	>	}
771
772	<	for (int i = 0; i < N_INTERACTION_FAMILIES; i++) {
773	<	AtomCommRealRow->scatter(pot_row[i], pot_temp[i]);
774	<	for (int ii = 0;  ii < pot_temp[i].size(); ii++ ) {
775	<	pot_local[i] += pot_temp[i][ii];
772	>	if (storageLayout_ & DataStorage::dslFlucQForce) {
773	>
774	>	int nq = snap_->atomData.flucQFrc.size();
775	>	vector<RealType> fqfrc_tmp(nq, 0.0);
776	>
777	>	AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
778	>	for (int i = 0; i < nq; i++) {
779	>	snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
780	>	fqfrc_tmp[i] = 0.0;
781		}
782	+
783	+	AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
784	+	for (int i = 0; i < nq; i++)
785	+	snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
786	+
787		}
788	+
789	+	nLocal_ = snap_->getNumberOfAtoms();
790	+
791	+	vector<potVec> pot_temp(nLocal_,
792	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
793	+	vector<potVec> expot_temp(nLocal_,
794	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
795	+
796	+	// scatter/gather pot_row into the members of my column
797	+
798	+	AtomPlanPotRow->scatter(pot_row, pot_temp);
799	+	AtomPlanPotRow->scatter(expot_row, expot_temp);
800	+
801	+	for (int ii = 0;  ii < pot_temp.size(); ii++ )
802	+	pairwisePot += pot_temp[ii];
803	+
804	+	for (int ii = 0;  ii < expot_temp.size(); ii++ )
805	+	excludedPot += expot_temp[ii];
806	+
807	+	if (storageLayout_ & DataStorage::dslParticlePot) {
808	+	// This is the pairwise contribution to the particle pot.  The
809	+	// embedding contribution is added in each of the low level
810	+	// non-bonded routines.  In single processor, this is done in
811	+	// unpackInteractionData, not in collectData.
812	+	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
813	+	for (int i = 0; i < nLocal_; i++) {
814	+	// factor of two is because the total potential terms are divided
815	+	// by 2 in parallel due to row/ column scatter
816	+	snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
817	+	}
818	+	}
819	+	}
820	+
821	+	fill(pot_temp.begin(), pot_temp.end(),
822	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
823	+	fill(expot_temp.begin(), expot_temp.end(),
824	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
825	+
826	+	AtomPlanPotColumn->scatter(pot_col, pot_temp);
827	+	AtomPlanPotColumn->scatter(expot_col, expot_temp);
828	+
829	+	for (int ii = 0;  ii < pot_temp.size(); ii++ )
830	+	pairwisePot += pot_temp[ii];
831	+
832	+	for (int ii = 0;  ii < expot_temp.size(); ii++ )
833	+	excludedPot += expot_temp[ii];
834	+
835	+	if (storageLayout_ & DataStorage::dslParticlePot) {
836	+	// This is the pairwise contribution to the particle pot.  The
837	+	// embedding contribution is added in each of the low level
838	+	// non-bonded routines.  In single processor, this is done in
839	+	// unpackInteractionData, not in collectData.
840	+	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
841	+	for (int i = 0; i < nLocal_; i++) {
842	+	// factor of two is because the total potential terms are divided
843	+	// by 2 in parallel due to row/ column scatter
844	+	snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
845	+	}
846	+	}
847	+	}
848	+
849	+	if (storageLayout_ & DataStorage::dslParticlePot) {
850	+	int npp = snap_->atomData.particlePot.size();
851	+	vector<RealType> ppot_temp(npp, 0.0);
852	+
853	+	// This is the direct or embedding contribution to the particle
854	+	// pot.
855	+
856	+	AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
857	+	for (int i = 0; i < npp; i++) {
858	+	snap_->atomData.particlePot[i] += ppot_temp[i];
859	+	}
860	+
861	+	fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
862	+
863	+	AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
864	+	for (int i = 0; i < npp; i++) {
865	+	snap_->atomData.particlePot[i] += ppot_temp[i];
866	+	}
867	+	}
868	+
869	+	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
870	+	RealType ploc1 = pairwisePot[ii];
871	+	RealType ploc2 = 0.0;
872	+	MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
873	+	pairwisePot[ii] = ploc2;
874	+	}
875	+
876	+	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
877	+	RealType ploc1 = excludedPot[ii];
878	+	RealType ploc2 = 0.0;
879	+	MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
880	+	excludedPot[ii] = ploc2;
881	+	}
882	+
883	+	// Here be dragons.
884	+	MPI::Intracomm col = colComm.getComm();
885	+
886	+	col.Allreduce(MPI::IN_PLACE,
887	+	&snap_->frameData.conductiveHeatFlux[0], 3,
888	+	MPI::REALTYPE, MPI::SUM);
889	+
890	+
891	+	#endif
892	+
893	+	}
894	+
895	+	/**
896	+	* Collects information obtained during the post-pair (and embedding
897	+	* functional) loops onto local data structures.
898	+	*/
899	+	void ForceMatrixDecomposition::collectSelfData() {
900	+	snap_ = sman_->getCurrentSnapshot();
901	+	storageLayout_ = sman_->getStorageLayout();
902	+
903	+	#ifdef IS_MPI
904	+	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
905	+	RealType ploc1 = embeddingPot[ii];
906	+	RealType ploc2 = 0.0;
907	+	MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
908	+	embeddingPot[ii] = ploc2;
909	+	}
910	+	#endif
911	+
912	+	}
913	+
914	+
915	+
916	+	int ForceMatrixDecomposition::getNAtomsInRow() {
917	+	#ifdef IS_MPI
918	+	return nAtomsInRow_;
919	+	#else
920	+	return nLocal_;
921	+	#endif
922	+	}
923	+
924	+	/**
925	+	* returns the list of atoms belonging to this group.
926	+	*/
927	+	vector<int> ForceMatrixDecomposition::getAtomsInGroupRow(int cg1){
928	+	#ifdef IS_MPI
929	+	return groupListRow_[cg1];
930	+	#else
931	+	return groupList_[cg1];
932		#endif
933		}
934
935	+	vector<int> ForceMatrixDecomposition::getAtomsInGroupColumn(int cg2){
936	+	#ifdef IS_MPI
937	+	return groupListCol_[cg2];
938	+	#else
939	+	return groupList_[cg2];
940	+	#endif
941	+	}
942
943		Vector3d ForceMatrixDecomposition::getIntergroupVector(int cg1, int cg2){
944		Vector3d d;
#	Line 257 | Line 953 | namespace OpenMD {
953		return d;
954		}
955
956	+	Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
957	+	#ifdef IS_MPI
958	+	return cgColData.velocity[cg2];
959	+	#else
960	+	return snap_->cgData.velocity[cg2];
961	+	#endif
962	+	}
963
964	+	Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
965	+	#ifdef IS_MPI
966	+	return atomColData.velocity[atom2];
967	+	#else
968	+	return snap_->atomData.velocity[atom2];
969	+	#endif
970	+	}
971	+
972	+
973		Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
974
975		Vector3d d;
#	Line 284 | Line 996 | namespace OpenMD {
996		snap_->wrapVector(d);
997		return d;
998		}
999	+
1000	+	RealType ForceMatrixDecomposition::getMassFactorRow(int atom1) {
1001	+	#ifdef IS_MPI
1002	+	return massFactorsRow[atom1];
1003	+	#else
1004	+	return massFactors[atom1];
1005	+	#endif
1006	+	}
1007	+
1008	+	RealType ForceMatrixDecomposition::getMassFactorColumn(int atom2) {
1009	+	#ifdef IS_MPI
1010	+	return massFactorsCol[atom2];
1011	+	#else
1012	+	return massFactors[atom2];
1013	+	#endif
1014	+
1015	+	}
1016
1017		Vector3d ForceMatrixDecomposition::getInteratomicVector(int atom1, int atom2){
1018		Vector3d d;
#	Line 298 | Line 1027 | namespace OpenMD {
1027		return d;
1028		}
1029
1030	+	vector<int> ForceMatrixDecomposition::getExcludesForAtom(int atom1) {
1031	+	return excludesForAtom[atom1];
1032	+	}
1033	+
1034	+	/**
1035	+	* We need to exclude some overcounted interactions that result from
1036	+	* the parallel decomposition.
1037	+	*/
1038	+	bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) {
1039	+	int unique_id_1, unique_id_2, group1, group2;
1040	+
1041	+	#ifdef IS_MPI
1042	+	// in MPI, we have to look up the unique IDs for each atom
1043	+	unique_id_1 = AtomRowToGlobal[atom1];
1044	+	unique_id_2 = AtomColToGlobal[atom2];
1045	+	group1 = cgRowToGlobal[cg1];
1046	+	group2 = cgColToGlobal[cg2];
1047	+	#else
1048	+	unique_id_1 = AtomLocalToGlobal[atom1];
1049	+	unique_id_2 = AtomLocalToGlobal[atom2];
1050	+	group1 = cgLocalToGlobal[cg1];
1051	+	group2 = cgLocalToGlobal[cg2];
1052	+	#endif
1053	+
1054	+	if (unique_id_1 == unique_id_2) return true;
1055	+
1056	+	#ifdef IS_MPI
1057	+	// this prevents us from doing the pair on multiple processors
1058	+	if (unique_id_1 < unique_id_2) {
1059	+	if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1060	+	} else {
1061	+	if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1062	+	}
1063	+	#endif
1064	+
1065	+	#ifndef IS_MPI
1066	+	if (group1 == group2) {
1067	+	if (unique_id_1 < unique_id_2) return true;
1068	+	}
1069	+	#endif
1070	+
1071	+	return false;
1072	+	}
1073	+
1074	+	/**
1075	+	* We need to handle the interactions for atoms who are involved in
1076	+	* the same rigid body as well as some short range interactions
1077	+	* (bonds, bends, torsions) differently from other interactions.
1078	+	* We'll still visit the pairwise routines, but with a flag that
1079	+	* tells those routines to exclude the pair from direct long range
1080	+	* interactions.  Some indirect interactions (notably reaction
1081	+	* field) must still be handled for these pairs.
1082	+	*/
1083	+	bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1084	+
1085	+	// excludesForAtom was constructed to use row/column indices in the MPI
1086	+	// version, and to use local IDs in the non-MPI version:
1087	+
1088	+	for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1089	+	i != excludesForAtom[atom1].end(); ++i) {
1090	+	if ( (*i) == atom2 ) return true;
1091	+	}
1092	+
1093	+	return false;
1094	+	}
1095	+
1096	+
1097		void ForceMatrixDecomposition::addForceToAtomRow(int atom1, Vector3d fg){
1098		#ifdef IS_MPI
1099		atomRowData.force[atom1] += fg;
#	Line 312 | Line 1108 | namespace OpenMD {
1108		#else
1109		snap_->atomData.force[atom2] += fg;
1110		#endif
315	–
1111		}
1112
1113		// filling interaction blocks with pointers
1114	<	InteractionData ForceMatrixDecomposition::fillInteractionData(int atom1, int atom2) {
1114	>	void ForceMatrixDecomposition::fillInteractionData(InteractionData &idat,
1115	>	int atom1, int atom2) {
1116
1117	<	InteractionData idat;
1117	>	idat.excluded = excludeAtomPair(atom1, atom2);
1118	>
1119		#ifdef IS_MPI
1120	+	idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
1121	+	//idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
1122	+	//                         ff_->getAtomType(identsCol[atom2]) );
1123	+
1124		if (storageLayout_ & DataStorage::dslAmat) {
1125		idat.A1 = &(atomRowData.aMat[atom1]);
1126		idat.A2 = &(atomColData.aMat[atom2]);
1127		}
1128	<
1128	>
1129		if (storageLayout_ & DataStorage::dslElectroFrame) {
1130		idat.eFrame1 = &(atomRowData.electroFrame[atom1]);
1131		idat.eFrame2 = &(atomColData.electroFrame[atom2]);
#	Line 340 | Line 1141 | namespace OpenMD {
1141		idat.rho2 = &(atomColData.density[atom2]);
1142		}
1143
1144	+	if (storageLayout_ & DataStorage::dslFunctional) {
1145	+	idat.frho1 = &(atomRowData.functional[atom1]);
1146	+	idat.frho2 = &(atomColData.functional[atom2]);
1147	+	}
1148	+
1149		if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
1150		idat.dfrho1 = &(atomRowData.functionalDerivative[atom1]);
1151		idat.dfrho2 = &(atomColData.functionalDerivative[atom2]);
1152		}
1153	<	#endif
1153	>
1154	>	if (storageLayout_ & DataStorage::dslParticlePot) {
1155	>	idat.particlePot1 = &(atomRowData.particlePot[atom1]);
1156	>	idat.particlePot2 = &(atomColData.particlePot[atom2]);
1157	>	}
1158	>
1159	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1160	>	idat.skippedCharge1 = &(atomRowData.skippedCharge[atom1]);
1161	>	idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1162	>	}
1163	>
1164	>	if (storageLayout_ & DataStorage::dslFlucQPosition) {
1165	>	idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1166	>	idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1167	>	}
1168	>
1169	>	#else
1170
1171	<	}
1172	<	InteractionData ForceMatrixDecomposition::fillSkipData(int atom1, int atom2){
1173	<	}
1174	<	SelfData ForceMatrixDecomposition::fillSelfData(int atom1) {
1171	>	idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1172	>
1173	>	if (storageLayout_ & DataStorage::dslAmat) {
1174	>	idat.A1 = &(snap_->atomData.aMat[atom1]);
1175	>	idat.A2 = &(snap_->atomData.aMat[atom2]);
1176	>	}
1177	>
1178	>	if (storageLayout_ & DataStorage::dslElectroFrame) {
1179	>	idat.eFrame1 = &(snap_->atomData.electroFrame[atom1]);
1180	>	idat.eFrame2 = &(snap_->atomData.electroFrame[atom2]);
1181	>	}
1182	>
1183	>	if (storageLayout_ & DataStorage::dslTorque) {
1184	>	idat.t1 = &(snap_->atomData.torque[atom1]);
1185	>	idat.t2 = &(snap_->atomData.torque[atom2]);
1186	>	}
1187	>
1188	>	if (storageLayout_ & DataStorage::dslDensity) {
1189	>	idat.rho1 = &(snap_->atomData.density[atom1]);
1190	>	idat.rho2 = &(snap_->atomData.density[atom2]);
1191	>	}
1192	>
1193	>	if (storageLayout_ & DataStorage::dslFunctional) {
1194	>	idat.frho1 = &(snap_->atomData.functional[atom1]);
1195	>	idat.frho2 = &(snap_->atomData.functional[atom2]);
1196	>	}
1197	>
1198	>	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
1199	>	idat.dfrho1 = &(snap_->atomData.functionalDerivative[atom1]);
1200	>	idat.dfrho2 = &(snap_->atomData.functionalDerivative[atom2]);
1201	>	}
1202	>
1203	>	if (storageLayout_ & DataStorage::dslParticlePot) {
1204	>	idat.particlePot1 = &(snap_->atomData.particlePot[atom1]);
1205	>	idat.particlePot2 = &(snap_->atomData.particlePot[atom2]);
1206	>	}
1207	>
1208	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1209	>	idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1210	>	idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1211	>	}
1212	>
1213	>	if (storageLayout_ & DataStorage::dslFlucQPosition) {
1214	>	idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1215	>	idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1216	>	}
1217	>
1218	>	#endif
1219		}
1220
1221
1222	+	void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {
1223	+	#ifdef IS_MPI
1224	+	pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1225	+	pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1226	+	expot_row[atom1] += RealType(0.5) *  *(idat.excludedPot);
1227	+	expot_col[atom2] += RealType(0.5) *  *(idat.excludedPot);
1228	+
1229	+	atomRowData.force[atom1] += *(idat.f1);
1230	+	atomColData.force[atom2] -= *(idat.f1);
1231	+
1232	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
1233	+	atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1234	+	atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1235	+	}
1236	+
1237	+	if (storageLayout_ & DataStorage::dslElectricField) {
1238	+	atomRowData.electricField[atom1] += *(idat.eField1);
1239	+	atomColData.electricField[atom2] += *(idat.eField2);
1240	+	}
1241	+
1242	+	#else
1243	+	pairwisePot += *(idat.pot);
1244	+	excludedPot += *(idat.excludedPot);
1245	+
1246	+	snap_->atomData.force[atom1] += *(idat.f1);
1247	+	snap_->atomData.force[atom2] -= *(idat.f1);
1248	+
1249	+	if (idat.doParticlePot) {
1250	+	// This is the pairwise contribution to the particle pot.  The
1251	+	// embedding contribution is added in each of the low level
1252	+	// non-bonded routines.  In parallel, this calculation is done
1253	+	// in collectData, not in unpackInteractionData.
1254	+	snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1255	+	snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1256	+	}
1257	+
1258	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
1259	+	snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1260	+	snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1261	+	}
1262	+
1263	+	if (storageLayout_ & DataStorage::dslElectricField) {
1264	+	snap_->atomData.electricField[atom1] += *(idat.eField1);
1265	+	snap_->atomData.electricField[atom2] += *(idat.eField2);
1266	+	}
1267	+
1268	+	#endif
1269	+
1270	+	}
1271	+
1272	+	/*
1273	+	* buildNeighborList
1274	+	*
1275	+	* first element of pair is row-indexed CutoffGroup
1276	+	* second element of pair is column-indexed CutoffGroup
1277	+	*/
1278	+	vector<pair<int, int> > ForceMatrixDecomposition::buildNeighborList() {
1279	+
1280	+	vector<pair<int, int> > neighborList;
1281	+	groupCutoffs cuts;
1282	+	bool doAllPairs = false;
1283	+
1284	+	#ifdef IS_MPI
1285	+	cellListRow_.clear();
1286	+	cellListCol_.clear();
1287	+	#else
1288	+	cellList_.clear();
1289	+	#endif
1290	+
1291	+	RealType rList_ = (largestRcut_ + skinThickness_);
1292	+	RealType rl2 = rList_ * rList_;
1293	+	Snapshot* snap_ = sman_->getCurrentSnapshot();
1294	+	Mat3x3d Hmat = snap_->getHmat();
1295	+	Vector3d Hx = Hmat.getColumn(0);
1296	+	Vector3d Hy = Hmat.getColumn(1);
1297	+	Vector3d Hz = Hmat.getColumn(2);
1298	+
1299	+	nCells_.x() = (int) ( Hx.length() )/ rList_;
1300	+	nCells_.y() = (int) ( Hy.length() )/ rList_;
1301	+	nCells_.z() = (int) ( Hz.length() )/ rList_;
1302	+
1303	+	// handle small boxes where the cell offsets can end up repeating cells
1304	+
1305	+	if (nCells_.x() < 3) doAllPairs = true;
1306	+	if (nCells_.y() < 3) doAllPairs = true;
1307	+	if (nCells_.z() < 3) doAllPairs = true;
1308	+
1309	+	Mat3x3d invHmat = snap_->getInvHmat();
1310	+	Vector3d rs, scaled, dr;
1311	+	Vector3i whichCell;
1312	+	int cellIndex;
1313	+	int nCtot = nCells_.x() * nCells_.y() * nCells_.z();
1314	+
1315	+	#ifdef IS_MPI
1316	+	cellListRow_.resize(nCtot);
1317	+	cellListCol_.resize(nCtot);
1318	+	#else
1319	+	cellList_.resize(nCtot);
1320	+	#endif
1321	+
1322	+	if (!doAllPairs) {
1323	+	#ifdef IS_MPI
1324	+
1325	+	for (int i = 0; i < nGroupsInRow_; i++) {
1326	+	rs = cgRowData.position[i];
1327	+
1328	+	// scaled positions relative to the box vectors
1329	+	scaled = invHmat * rs;
1330	+
1331	+	// wrap the vector back into the unit box by subtracting integer box
1332	+	// numbers
1333	+	for (int j = 0; j < 3; j++) {
1334	+	scaled[j] -= roundMe(scaled[j]);
1335	+	scaled[j] += 0.5;
1336	+	}
1337	+
1338	+	// find xyz-indices of cell that cutoffGroup is in.
1339	+	whichCell.x() = nCells_.x() * scaled.x();
1340	+	whichCell.y() = nCells_.y() * scaled.y();
1341	+	whichCell.z() = nCells_.z() * scaled.z();
1342	+
1343	+	// find single index of this cell:
1344	+	cellIndex = Vlinear(whichCell, nCells_);
1345	+
1346	+	// add this cutoff group to the list of groups in this cell;
1347	+	cellListRow_[cellIndex].push_back(i);
1348	+	}
1349	+	for (int i = 0; i < nGroupsInCol_; i++) {
1350	+	rs = cgColData.position[i];
1351	+
1352	+	// scaled positions relative to the box vectors
1353	+	scaled = invHmat * rs;
1354	+
1355	+	// wrap the vector back into the unit box by subtracting integer box
1356	+	// numbers
1357	+	for (int j = 0; j < 3; j++) {
1358	+	scaled[j] -= roundMe(scaled[j]);
1359	+	scaled[j] += 0.5;
1360	+	}
1361	+
1362	+	// find xyz-indices of cell that cutoffGroup is in.
1363	+	whichCell.x() = nCells_.x() * scaled.x();
1364	+	whichCell.y() = nCells_.y() * scaled.y();
1365	+	whichCell.z() = nCells_.z() * scaled.z();
1366	+
1367	+	// find single index of this cell:
1368	+	cellIndex = Vlinear(whichCell, nCells_);
1369	+
1370	+	// add this cutoff group to the list of groups in this cell;
1371	+	cellListCol_[cellIndex].push_back(i);
1372	+	}
1373	+
1374	+	#else
1375	+	for (int i = 0; i < nGroups_; i++) {
1376	+	rs = snap_->cgData.position[i];
1377	+
1378	+	// scaled positions relative to the box vectors
1379	+	scaled = invHmat * rs;
1380	+
1381	+	// wrap the vector back into the unit box by subtracting integer box
1382	+	// numbers
1383	+	for (int j = 0; j < 3; j++) {
1384	+	scaled[j] -= roundMe(scaled[j]);
1385	+	scaled[j] += 0.5;
1386	+	}
1387	+
1388	+	// find xyz-indices of cell that cutoffGroup is in.
1389	+	whichCell.x() = nCells_.x() * scaled.x();
1390	+	whichCell.y() = nCells_.y() * scaled.y();
1391	+	whichCell.z() = nCells_.z() * scaled.z();
1392	+
1393	+	// find single index of this cell:
1394	+	cellIndex = Vlinear(whichCell, nCells_);
1395	+
1396	+	// add this cutoff group to the list of groups in this cell;
1397	+	cellList_[cellIndex].push_back(i);
1398	+	}
1399	+
1400	+	#endif
1401	+
1402	+	for (int m1z = 0; m1z < nCells_.z(); m1z++) {
1403	+	for (int m1y = 0; m1y < nCells_.y(); m1y++) {
1404	+	for (int m1x = 0; m1x < nCells_.x(); m1x++) {
1405	+	Vector3i m1v(m1x, m1y, m1z);
1406	+	int m1 = Vlinear(m1v, nCells_);
1407	+
1408	+	for (vector<Vector3i>::iterator os = cellOffsets_.begin();
1409	+	os != cellOffsets_.end(); ++os) {
1410	+
1411	+	Vector3i m2v = m1v + (*os);
1412	+
1413	+
1414	+	if (m2v.x() >= nCells_.x()) {
1415	+	m2v.x() = 0;
1416	+	} else if (m2v.x() < 0) {
1417	+	m2v.x() = nCells_.x() - 1;
1418	+	}
1419	+
1420	+	if (m2v.y() >= nCells_.y()) {
1421	+	m2v.y() = 0;
1422	+	} else if (m2v.y() < 0) {
1423	+	m2v.y() = nCells_.y() - 1;
1424	+	}
1425	+
1426	+	if (m2v.z() >= nCells_.z()) {
1427	+	m2v.z() = 0;
1428	+	} else if (m2v.z() < 0) {
1429	+	m2v.z() = nCells_.z() - 1;
1430	+	}
1431	+
1432	+	int m2 = Vlinear (m2v, nCells_);
1433	+
1434	+	#ifdef IS_MPI
1435	+	for (vector<int>::iterator j1 = cellListRow_[m1].begin();
1436	+	j1 != cellListRow_[m1].end(); ++j1) {
1437	+	for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1438	+	j2 != cellListCol_[m2].end(); ++j2) {
1439	+
1440	+	// In parallel, we need to visit *all* pairs of row
1441	+	// & column indicies and will divide labor in the
1442	+	// force evaluation later.
1443	+	dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1444	+	snap_->wrapVector(dr);
1445	+	cuts = getGroupCutoffs( (*j1), (*j2) );
1446	+	if (dr.lengthSquare() < cuts.third) {
1447	+	neighborList.push_back(make_pair((*j1), (*j2)));
1448	+	}
1449	+	}
1450	+	}
1451	+	#else
1452	+	for (vector<int>::iterator j1 = cellList_[m1].begin();
1453	+	j1 != cellList_[m1].end(); ++j1) {
1454	+	for (vector<int>::iterator j2 = cellList_[m2].begin();
1455	+	j2 != cellList_[m2].end(); ++j2) {
1456	+
1457	+	// Always do this if we're in different cells or if
1458	+	// we're in the same cell and the global index of
1459	+	// the j2 cutoff group is greater than or equal to
1460	+	// the j1 cutoff group.  Note that Rappaport's code
1461	+	// has a "less than" conditional here, but that
1462	+	// deals with atom-by-atom computation.  OpenMD
1463	+	// allows atoms within a single cutoff group to
1464	+	// interact with each other.
1465	+
1466	+
1467	+
1468	+	if (m2 != m1 || (*j2) >= (*j1) ) {
1469	+
1470	+	dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1471	+	snap_->wrapVector(dr);
1472	+	cuts = getGroupCutoffs( (*j1), (*j2) );
1473	+	if (dr.lengthSquare() < cuts.third) {
1474	+	neighborList.push_back(make_pair((*j1), (*j2)));
1475	+	}
1476	+	}
1477	+	}
1478	+	}
1479	+	#endif
1480	+	}
1481	+	}
1482	+	}
1483	+	}
1484	+	} else {
1485	+	// branch to do all cutoff group pairs
1486	+	#ifdef IS_MPI
1487	+	for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1488	+	for (int j2 = 0; j2 < nGroupsInCol_; j2++) {
1489	+	dr = cgColData.position[j2] - cgRowData.position[j1];
1490	+	snap_->wrapVector(dr);
1491	+	cuts = getGroupCutoffs( j1, j2 );
1492	+	if (dr.lengthSquare() < cuts.third) {
1493	+	neighborList.push_back(make_pair(j1, j2));
1494	+	}
1495	+	}
1496	+	}
1497	+	#else
1498	+	// include all groups here.
1499	+	for (int j1 = 0; j1 < nGroups_; j1++) {
1500	+	// include self group interactions j2 == j1
1501	+	for (int j2 = j1; j2 < nGroups_; j2++) {
1502	+	dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1503	+	snap_->wrapVector(dr);
1504	+	cuts = getGroupCutoffs( j1, j2 );
1505	+	if (dr.lengthSquare() < cuts.third) {
1506	+	neighborList.push_back(make_pair(j1, j2));
1507	+	}
1508	+	}
1509	+	}
1510	+	#endif
1511	+	}
1512	+
1513	+	// save the local cutoff group positions for the check that is
1514	+	// done on each loop:
1515	+	saved_CG_positions_.clear();
1516	+	for (int i = 0; i < nGroups_; i++)
1517	+	saved_CG_positions_.push_back(snap_->cgData.position[i]);
1518	+
1519	+	return neighborList;
1520	+	}
1521		} //end namespace OpenMD

Diff Legend

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