ViewVC Help

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

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

Comparing:
branches/development/src/parallel/ForceDecomposition.cpp (file contents), Revision 1538 by chuckv, Tue Jan 11 18:58:12 2011 UTC vs.
branches/development/src/parallel/ForceMatrixDecomposition.cpp (file contents), Revision 1755 by gezelter, Thu Jun 14 01:58:35 2012 UTC

#	Line 1 | Line 1
1	<	/**
2	<	* @file ForceDecomposition.cpp
3	<	* @author Charles Vardeman <cvardema.at.nd.edu>
4	<	* @date 08/18/2010
5	<	* @time 11:56am
6	<	* @version 1.0
1	>	/*
2	>	* Copyright (c) 2005 The University of Notre Dame. All Rights Reserved.
3		*
8	–	* @section LICENSE
9	–	* Copyright (c) 2010 The University of Notre Dame. All Rights Reserved.
10	–	*
4		* The University of Notre Dame grants you ("Licensee") a
5		* non-exclusive, royalty free, license to use, modify and
6		* redistribute this software in source and binary code form, provided
#	Line 43 | 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	<	/*  -*- c++ -*-  */
54	<	#include "config.h"
55	<	#include <stdlib.h>
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	<	#include <mpi.h>
58	<	#endif
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
58	–	#include <iostream>
59	–	#include <vector>
60	–	#include <algorithm>
61	–	#include <cmath>
62	–	#include "parallel/ForceDecomposition.hpp"
88
89	+	/**
90	+	* distributeInitialData is essentially a copy of the older fortran
91	+	* SimulationSetup
92	+	*/
93	+	void ForceMatrixDecomposition::distributeInitialData() {
94	+	snap_ = sman_->getCurrentSnapshot();
95	+	storageLayout_ = sman_->getStorageLayout();
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	<	using namespace std;
66	<	using namespace OpenMD;
106	>	massFactors = info_->getMassFactors();
107
108	<	//__static
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	<	static vector<MPI:Comm> communictors;
121	<	#endif
120	>
121	>	MPI::Intracomm row = rowComm.getComm();
122	>	MPI::Intracomm col = colComm.getComm();
123
124	<	//____ MPITypeTraits
125	<	template<typename T>
126	<	struct MPITypeTraits;
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	<	#ifdef IS_MPI
131	<	template<>
132	<	struct MPITypeTraits<RealType> {
133	<	static const MPI::Datatype datatype;
134	<	};
82	<	const MPI_Datatype MPITypeTraits<RealType>::datatype = MY_MPI_REAL;
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	<	template<>
137	<	struct MPITypeTraits<int> {
138	<	static const MPI::Datatype datatype;
139	<	};
88	<	const MPI::Datatype MPITypeTraits<int>::datatype = MPI_INT;
89	<	#endif
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	<	/**
142	<	* Constructor for ForceDecomposition Parallel Decomposition Method
143	<	* Will try to construct a symmetric grid of processors. Ideally, the
144	<	* number of processors will be a square ex: 4, 9, 16, 25.
95	<	*
96	<	*/
141	>	nAtomsInRow_ = AtomPlanIntRow->getSize();
142	>	nAtomsInCol_ = AtomPlanIntColumn->getSize();
143	>	nGroupsInRow_ = cgPlanIntRow->getSize();
144	>	nGroupsInCol_ = cgPlanIntColumn->getSize();
145
146	<	ForceDecomposition::ForceDecomposition() {
146	>	// Modify the data storage objects with the correct layouts and sizes:
147	>	atomRowData.resize(nAtomsInRow_);
148	>	atomRowData.setStorageLayout(storageLayout_);
149	>	atomColData.resize(nAtomsInCol_);
150	>	atomColData.setStorageLayout(storageLayout_);
151	>	cgRowData.resize(nGroupsInRow_);
152	>	cgRowData.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	>	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	<	#ifdef IS_MPI
172	<	int nProcs = MPI::COMM_WORLD.Get_size();
173	<	int worldRank = MPI::COMM_WORLD.Get_rank();
174	<	#endif
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	<	// First time through, construct column stride.
177	<	if (communicators.size() == 0)
178	<	{
179	<	int nColumnsMax = (int) round(sqrt((float) nProcs));
180	<	for (int i = 0; i < nProcs; ++i)
181	<	{
182	<	if (nProcs%i==0) nColumns=i;
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++) {
197	>	int gid = cgRowToGlobal[i];
198	>	for (int j = 0; j < nAtomsInRow_; j++) {
199	>	int aid = AtomRowToGlobal[j];
200	>	if (globalGroupMembership[aid] == gid)
201	>	groupListRow_[i].push_back(j);
202	>	}
203		}
204
205	<	int nRows = nProcs/nColumns;
206	<	myRank_ = (int) worldRank%nColumns;
207	<	}
208	<	else
209	<	{
210	<	myRank_ = myRank/nColumns;
211	<	}
212	<	MPI::Comm newComm = MPI:COMM_WORLD.Split(myRank_,0);
213	<
214	<	isColumn_ = false;
124	<
125	<	}
205	>	groupListCol_.clear();
206	>	groupListCol_.resize(nGroupsInCol_);
207	>	for (int i = 0; i < nGroupsInCol_; i++) {
208	>	int gid = cgColToGlobal[i];
209	>	for (int j = 0; j < nAtomsInCol_; j++) {
210	>	int aid = AtomColToGlobal[j];
211	>	if (globalGroupMembership[aid] == gid)
212	>	groupListCol_[i].push_back(j);
213	>	}
214	>	}
215
216	<	ForceDecomposition::gather(sendbuf, receivebuf){
217	<	communicators(myIndex_).Allgatherv();
218	<	}
216	>	excludesForAtom.clear();
217	>	excludesForAtom.resize(nAtomsInRow_);
218	>	toposForAtom.clear();
219	>	toposForAtom.resize(nAtomsInRow_);
220	>	topoDist.clear();
221	>	topoDist.resize(nAtomsInRow_);
222	>	for (int i = 0; i < nAtomsInRow_; i++) {
223	>	int iglob = AtomRowToGlobal[i];
224
225	+	for (int j = 0; j < nAtomsInCol_; j++) {
226	+	int jglob = AtomColToGlobal[j];
227
228	+	if (excludes->hasPair(iglob, jglob))
229	+	excludesForAtom[i].push_back(j);
230	+
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)) {
236	+	toposForAtom[i].push_back(j);
237	+	topoDist[i].push_back(2);
238	+	} else {
239	+	if (oneFour->hasPair(iglob, jglob)) {
240	+	toposForAtom[i].push_back(j);
241	+	topoDist[i].push_back(3);
242	+	}
243	+	}
244	+	}
245	+	}
246	+	}
247
248	<	ForceDecomposition::scatter(sbuffer, rbuffer){
249	<	communicators(myIndex_).Reduce_scatter(sbuffer, recevbuf. recvcounts, MPI::DOUBLE, MPI::SUM);
250	<	}
248	>	#else
249	>	excludesForAtom.clear();
250	>	excludesForAtom.resize(nLocal_);
251	>	toposForAtom.clear();
252	>	toposForAtom.resize(nLocal_);
253	>	topoDist.clear();
254	>	topoDist.resize(nLocal_);
255
256	+	for (int i = 0; i < nLocal_; i++) {
257	+	int iglob = AtomLocalToGlobal[i];
258
259	+	for (int j = 0; j < nLocal_; j++) {
260	+	int jglob = AtomLocalToGlobal[j];
261	+
262	+	if (excludes->hasPair(iglob, jglob))
263	+	excludesForAtom[i].push_back(j);
264	+
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)) {
270	+	toposForAtom[i].push_back(j);
271	+	topoDist[i].push_back(2);
272	+	} else {
273	+	if (oneFour->hasPair(iglob, jglob)) {
274	+	toposForAtom[i].push_back(j);
275	+	topoDist[i].push_back(3);
276	+	}
277	+	}
278	+	}
279	+	}
280	+	}
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	+
308	+	RealType tol = 1e-6;
309	+	largestRcut_ = 0.0;
310	+	RealType rc;
311	+	int atid;
312	+	set<AtomType*> atypes = info_->getSimulatedAtomTypes();
313	+
314	+	map<int, RealType> atypeCutoff;
315	+
316	+	for (set<AtomType*>::iterator at = atypes.begin();
317	+	at != atypes.end(); ++at){
318	+	atid = (*at)->getIdent();
319	+	if (userChoseCutoff_)
320	+	atypeCutoff[atid] = userCutoff_;
321	+	else
322	+	atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at);
323	+	}
324	+
325	+	vector<RealType> gTypeCutoffs;
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
329	+	vector<RealType> groupCutoffRow(nGroupsInRow_, 0.0);
330	+	groupRowToGtype.resize(nGroupsInRow_);
331	+	for (int cg1 = 0; cg1 < nGroupsInRow_; cg1++) {
332	+	vector<int> atomListRow = getAtomsInGroupRow(cg1);
333	+	for (vector<int>::iterator ia = atomListRow.begin();
334	+	ia != atomListRow.end(); ++ia) {
335	+	int atom1 = (*ia);
336	+	atid = identsRow[atom1];
337	+	if (atypeCutoff[atid] > groupCutoffRow[cg1]) {
338	+	groupCutoffRow[cg1] = atypeCutoff[atid];
339	+	}
340	+	}
341	+
342	+	bool gTypeFound = false;
343	+	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
344	+	if (abs(groupCutoffRow[cg1] - gTypeCutoffs[gt]) < tol) {
345	+	groupRowToGtype[cg1] = gt;
346	+	gTypeFound = true;
347	+	}
348	+	}
349	+	if (!gTypeFound) {
350	+	gTypeCutoffs.push_back( groupCutoffRow[cg1] );
351	+	groupRowToGtype[cg1] = gTypeCutoffs.size() - 1;
352	+	}
353	+
354	+	}
355	+	vector<RealType> groupCutoffCol(nGroupsInCol_, 0.0);
356	+	groupColToGtype.resize(nGroupsInCol_);
357	+	for (int cg2 = 0; cg2 < nGroupsInCol_; cg2++) {
358	+	vector<int> atomListCol = getAtomsInGroupColumn(cg2);
359	+	for (vector<int>::iterator jb = atomListCol.begin();
360	+	jb != atomListCol.end(); ++jb) {
361	+	int atom2 = (*jb);
362	+	atid = identsCol[atom2];
363	+	if (atypeCutoff[atid] > groupCutoffCol[cg2]) {
364	+	groupCutoffCol[cg2] = atypeCutoff[atid];
365	+	}
366	+	}
367	+	bool gTypeFound = false;
368	+	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
369	+	if (abs(groupCutoffCol[cg2] - gTypeCutoffs[gt]) < tol) {
370	+	groupColToGtype[cg2] = gt;
371	+	gTypeFound = true;
372	+	}
373	+	}
374	+	if (!gTypeFound) {
375	+	gTypeCutoffs.push_back( groupCutoffCol[cg2] );
376	+	groupColToGtype[cg2] = gTypeCutoffs.size() - 1;
377	+	}
378	+	}
379	+	#else
380	+
381	+	vector<RealType> groupCutoff(nGroups_, 0.0);
382	+	groupToGtype.resize(nGroups_);
383	+	for (int cg1 = 0; cg1 < nGroups_; cg1++) {
384	+	groupCutoff[cg1] = 0.0;
385	+	vector<int> atomList = getAtomsInGroupRow(cg1);
386	+	for (vector<int>::iterator ia = atomList.begin();
387	+	ia != atomList.end(); ++ia) {
388	+	int atom1 = (*ia);
389	+	atid = idents[atom1];
390	+	if (atypeCutoff[atid] > groupCutoff[cg1])
391	+	groupCutoff[cg1] = atypeCutoff[atid];
392	+	}
393	+
394	+	bool gTypeFound = false;
395	+	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
396	+	if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
397	+	groupToGtype[cg1] = gt;
398	+	gTypeFound = true;
399	+	}
400	+	}
401	+	if (!gTypeFound) {
402	+	gTypeCutoffs.push_back( groupCutoff[cg1] );
403	+	groupToGtype[cg1] = gTypeCutoffs.size() - 1;
404	+	}
405	+	}
406	+	#endif
407	+
408	+	// Now we find the maximum group cutoff value present in the simulation
409	+
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,
415	+	MPI::MAX);
416	+	#endif
417	+
418	+	RealType tradRcut = groupMax;
419	+
420	+	for (int i = 0; i < gTypeCutoffs.size();  i++) {
421	+	for (int j = 0; j < gTypeCutoffs.size();  j++) {
422	+	RealType thisRcut;
423	+	switch(cutoffPolicy_) {
424	+	case TRADITIONAL:
425	+	thisRcut = tradRcut;
426	+	break;
427	+	case MIX:
428	+	thisRcut = 0.5 * (gTypeCutoffs[i] + gTypeCutoffs[j]);
429	+	break;
430	+	case MAX:
431	+	thisRcut = max(gTypeCutoffs[i], gTypeCutoffs[j]);
432	+	break;
433	+	default:
434	+	sprintf(painCave.errMsg,
435	+	"ForceMatrixDecomposition::createGtypeCutoffMap "
436	+	"hit an unknown cutoff policy!\n");
437	+	painCave.severity = OPENMD_ERROR;
438	+	painCave.isFatal = 1;
439	+	simError();
440	+	break;
441	+	}
442	+
443	+	pair<int,int> key = make_pair(i,j);
444	+	gTypeCutoffMap[key].first = thisRcut;
445	+	if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
446	+	gTypeCutoffMap[key].second = thisRcut*thisRcut;
447	+	gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
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 (%lf) does not match computed group Cutoff\n", userCutoff_);
455	+	painCave.severity = OPENMD_ERROR;
456	+	painCave.isFatal = 1;
457	+	simError();
458	+	}
459	+	}
460	+	}
461	+	}
462	+	}
463	+
464	+
465	+	groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
466	+	int i, j;
467	+	#ifdef IS_MPI
468	+	i = groupRowToGtype[cg1];
469	+	j = groupColToGtype[cg2];
470	+	#else
471	+	i = groupToGtype[cg1];
472	+	j = groupToGtype[cg2];
473	+	#endif
474	+	return gTypeCutoffMap[make_pair(i,j)];
475	+	}
476	+
477	+	int ForceMatrixDecomposition::getTopologicalDistance(int atom1, int atom2) {
478	+	for (int j = 0; j < toposForAtom[atom1].size(); j++) {
479	+	if (toposForAtom[atom1][j] == atom2)
480	+	return topoDist[atom1][j];
481	+	}
482	+	return 0;
483	+	}
484	+
485	+	void ForceMatrixDecomposition::zeroWorkArrays() {
486	+	pairwisePot = 0.0;
487	+	embeddingPot = 0.0;
488	+
489	+	#ifdef IS_MPI
490	+	if (storageLayout_ & DataStorage::dslForce) {
491	+	fill(atomRowData.force.begin(), atomRowData.force.end(), V3Zero);
492	+	fill(atomColData.force.begin(), atomColData.force.end(), V3Zero);
493	+	}
494	+
495	+	if (storageLayout_ & DataStorage::dslTorque) {
496	+	fill(atomRowData.torque.begin(), atomRowData.torque.end(), V3Zero);
497	+	fill(atomColData.torque.begin(), atomColData.torque.end(), V3Zero);
498	+	}
499	+
500	+	fill(pot_row.begin(), pot_row.end(),
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));
505	+
506	+	if (storageLayout_ & DataStorage::dslParticlePot) {
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) {
514	+	fill(atomRowData.density.begin(), atomRowData.density.end(), 0.0);
515	+	fill(atomColData.density.begin(), atomColData.density.end(), 0.0);
516	+	}
517	+
518	+	if (storageLayout_ & DataStorage::dslFunctional) {
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) {
526	+	fill(atomRowData.functionalDerivative.begin(),
527	+	atomRowData.functionalDerivative.end(), 0.0);
528	+	fill(atomColData.functionalDerivative.begin(),
529	+	atomColData.functionalDerivative.end(), 0.0);
530	+	}
531	+
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);
566	+	}
567	+
568	+	if (storageLayout_ & DataStorage::dslDensity) {
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	+
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	+
595	+	void ForceMatrixDecomposition::distributeData()  {
596	+	snap_ = sman_->getCurrentSnapshot();
597	+	storageLayout_ = sman_->getStorageLayout();
598	+	#ifdef IS_MPI
599	+
600	+	// gather up the atomic positions
601	+	AtomPlanVectorRow->gather(snap_->atomData.position,
602	+	atomRowData.position);
603	+	AtomPlanVectorColumn->gather(snap_->atomData.position,
604	+	atomColData.position);
605	+
606	+	// gather up the cutoff group positions
607	+
608	+	cgPlanVectorRow->gather(snap_->cgData.position,
609	+	cgRowData.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	+	AtomPlanMatrixRow->gather(snap_->atomData.aMat,
629	+	atomRowData.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	+	AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
637	+	atomRowData.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	+
653	+	/* collects information obtained during the pre-pair loop onto local
654	+	* data structures.
655	+	*/
656	+	void ForceMatrixDecomposition::collectIntermediateData() {
657	+	snap_ = sman_->getCurrentSnapshot();
658	+	storageLayout_ = sman_->getStorageLayout();
659	+	#ifdef IS_MPI
660	+
661	+	if (storageLayout_ & DataStorage::dslDensity) {
662	+
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	+	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	+
687	+	/*
688	+	* redistributes information obtained during the pre-pair loop out to
689	+	* row and column-indexed data structures
690	+	*/
691	+	void ForceMatrixDecomposition::distributeIntermediateData() {
692	+	snap_ = sman_->getCurrentSnapshot();
693	+	storageLayout_ = sman_->getStorageLayout();
694	+	#ifdef IS_MPI
695	+	if (storageLayout_ & DataStorage::dslFunctional) {
696	+	AtomPlanRealRow->gather(snap_->atomData.functional,
697	+	atomRowData.functional);
698	+	AtomPlanRealColumn->gather(snap_->atomData.functional,
699	+	atomColData.functional);
700	+	}
701	+
702	+	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
703	+	AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
704	+	atomRowData.functionalDerivative);
705	+	AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
706	+	atomColData.functionalDerivative);
707	+	}
708	+	#endif
709	+	}
710	+
711	+
712	+	void ForceMatrixDecomposition::collectData() {
713	+	snap_ = sman_->getCurrentSnapshot();
714	+	storageLayout_ = sman_->getStorageLayout();
715	+	#ifdef IS_MPI
716	+	int n = snap_->atomData.force.size();
717	+	vector<Vector3d> frc_tmp(n, V3Zero);
718	+
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	+	AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
726	+	for (int i = 0; i < n; i++) {
727	+	snap_->atomData.force[i] += frc_tmp[i];
728	+	}
729	+
730	+	if (storageLayout_ & DataStorage::dslTorque) {
731	+
732	+	int nt = snap_->atomData.torque.size();
733	+	vector<Vector3d> trq_tmp(nt, V3Zero);
734	+
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	+	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_,
783	+	Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
784	+
785	+	// scatter/gather pot_row into the members of my column
786	+
787	+	AtomPlanPotRow->scatter(pot_row, pot_temp);
788	+
789	+	for (int ii = 0;  ii < pot_temp.size(); ii++ )
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	+	AtomPlanPotColumn->scatter(pot_col, pot_temp);
810	+
811	+	for (int ii = 0;  ii < pot_temp.size(); 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() {
875	+	#ifdef IS_MPI
876	+	return nAtomsInRow_;
877	+	#else
878	+	return nLocal_;
879	+	#endif
880	+	}
881	+
882	+	/**
883	+	* returns the list of atoms belonging to this group.
884	+	*/
885	+	vector<int> ForceMatrixDecomposition::getAtomsInGroupRow(int cg1){
886	+	#ifdef IS_MPI
887	+	return groupListRow_[cg1];
888	+	#else
889	+	return groupList_[cg1];
890	+	#endif
891	+	}
892	+
893	+	vector<int> ForceMatrixDecomposition::getAtomsInGroupColumn(int cg2){
894	+	#ifdef IS_MPI
895	+	return groupListCol_[cg2];
896	+	#else
897	+	return groupList_[cg2];
898	+	#endif
899	+	}
900	+
901	+	Vector3d ForceMatrixDecomposition::getIntergroupVector(int cg1, int cg2){
902	+	Vector3d d;
903	+
904	+	#ifdef IS_MPI
905	+	d = cgColData.position[cg2] - cgRowData.position[cg1];
906	+	#else
907	+	d = snap_->cgData.position[cg2] - snap_->cgData.position[cg1];
908	+	#endif
909	+
910	+	snap_->wrapVector(d);
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;
934	+
935	+	#ifdef IS_MPI
936	+	d = cgRowData.position[cg1] - atomRowData.position[atom1];
937	+	#else
938	+	d = snap_->cgData.position[cg1] - snap_->atomData.position[atom1];
939	+	#endif
940	+
941	+	snap_->wrapVector(d);
942	+	return d;
943	+	}
944	+
945	+	Vector3d ForceMatrixDecomposition::getAtomToGroupVectorColumn(int atom2, int cg2){
946	+	Vector3d d;
947	+
948	+	#ifdef IS_MPI
949	+	d = cgColData.position[cg2] - atomColData.position[atom2];
950	+	#else
951	+	d = snap_->cgData.position[cg2] - snap_->atomData.position[atom2];
952	+	#endif
953	+
954	+	snap_->wrapVector(d);
955	+	return d;
956	+	}
957	+
958	+	RealType ForceMatrixDecomposition::getMassFactorRow(int atom1) {
959	+	#ifdef IS_MPI
960	+	return massFactorsRow[atom1];
961	+	#else
962	+	return massFactors[atom1];
963	+	#endif
964	+	}
965	+
966	+	RealType ForceMatrixDecomposition::getMassFactorColumn(int atom2) {
967	+	#ifdef IS_MPI
968	+	return massFactorsCol[atom2];
969	+	#else
970	+	return massFactors[atom2];
971	+	#endif
972	+
973	+	}
974	+
975	+	Vector3d ForceMatrixDecomposition::getInteratomicVector(int atom1, int atom2){
976	+	Vector3d d;
977	+
978	+	#ifdef IS_MPI
979	+	d = atomColData.position[atom2] - atomRowData.position[atom1];
980	+	#else
981	+	d = snap_->atomData.position[atom2] - snap_->atomData.position[atom1];
982	+	#endif
983	+
984	+	snap_->wrapVector(d);
985	+	return d;
986	+	}
987	+
988	+	vector<int> ForceMatrixDecomposition::getExcludesForAtom(int atom1) {
989	+	return excludesForAtom[atom1];
990	+	}
991	+
992	+	/**
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	+
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	+
1008	+	if (unique_id_1 == unique_id_2) return true;
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;
1016	+	}
1017	+	#endif
1018	+
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	+
1045	+	void ForceMatrixDecomposition::addForceToAtomRow(int atom1, Vector3d fg){
1046	+	#ifdef IS_MPI
1047	+	atomRowData.force[atom1] += fg;
1048	+	#else
1049	+	snap_->atomData.force[atom1] += fg;
1050	+	#endif
1051	+	}
1052	+
1053	+	void ForceMatrixDecomposition::addForceToAtomColumn(int atom2, Vector3d fg){
1054	+	#ifdef IS_MPI
1055	+	atomColData.force[atom2] += fg;
1056	+	#else
1057	+	snap_->atomData.force[atom2] += fg;
1058	+	#endif
1059	+	}
1060	+
1061	+	// filling interaction blocks with pointers
1062	+	void ForceMatrixDecomposition::fillInteractionData(InteractionData &idat,
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	+
1072	+	if (storageLayout_ & DataStorage::dslAmat) {
1073	+	idat.A1 = &(atomRowData.aMat[atom1]);
1074	+	idat.A2 = &(atomColData.aMat[atom2]);
1075	+	}
1076	+
1077	+	if (storageLayout_ & DataStorage::dslElectroFrame) {
1078	+	idat.eFrame1 = &(atomRowData.electroFrame[atom1]);
1079	+	idat.eFrame2 = &(atomColData.electroFrame[atom2]);
1080	+	}
1081	+
1082	+	if (storageLayout_ & DataStorage::dslTorque) {
1083	+	idat.t1 = &(atomRowData.torque[atom1]);
1084	+	idat.t2 = &(atomColData.torque[atom2]);
1085	+	}
1086	+
1087	+	if (storageLayout_ & DataStorage::dslDensity) {
1088	+	idat.rho1 = &(atomRowData.density[atom1]);
1089	+	idat.rho2 = &(atomColData.density[atom2]);
1090	+	}
1091	+
1092	+	if (storageLayout_ & DataStorage::dslFunctional) {
1093	+	idat.frho1 = &(atomRowData.functional[atom1]);
1094	+	idat.frho2 = &(atomColData.functional[atom2]);
1095	+	}
1096	+
1097	+	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
1098	+	idat.dfrho1 = &(atomRowData.functionalDerivative[atom1]);
1099	+	idat.dfrho2 = &(atomColData.functionalDerivative[atom2]);
1100	+	}
1101	+
1102	+	if (storageLayout_ & DataStorage::dslParticlePot) {
1103	+	idat.particlePot1 = &(atomRowData.particlePot[atom1]);
1104	+	idat.particlePot2 = &(atomColData.particlePot[atom2]);
1105	+	}
1106	+
1107	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1108	+	idat.skippedCharge1 = &(atomRowData.skippedCharge[atom1]);
1109	+	idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1110	+	}
1111	+
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]);
1124	+	}
1125	+
1126	+	if (storageLayout_ & DataStorage::dslElectroFrame) {
1127	+	idat.eFrame1 = &(snap_->atomData.electroFrame[atom1]);
1128	+	idat.eFrame2 = &(snap_->atomData.electroFrame[atom2]);
1129	+	}
1130	+
1131	+	if (storageLayout_ & DataStorage::dslTorque) {
1132	+	idat.t1 = &(snap_->atomData.torque[atom1]);
1133	+	idat.t2 = &(snap_->atomData.torque[atom2]);
1134	+	}
1135	+
1136	+	if (storageLayout_ & DataStorage::dslDensity) {
1137	+	idat.rho1 = &(snap_->atomData.density[atom1]);
1138	+	idat.rho2 = &(snap_->atomData.density[atom2]);
1139	+	}
1140	+
1141	+	if (storageLayout_ & DataStorage::dslFunctional) {
1142	+	idat.frho1 = &(snap_->atomData.functional[atom1]);
1143	+	idat.frho2 = &(snap_->atomData.functional[atom2]);
1144	+	}
1145	+
1146	+	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
1147	+	idat.dfrho1 = &(snap_->atomData.functionalDerivative[atom1]);
1148	+	idat.dfrho2 = &(snap_->atomData.functionalDerivative[atom2]);
1149	+	}
1150	+
1151	+	if (storageLayout_ & DataStorage::dslParticlePot) {
1152	+	idat.particlePot1 = &(snap_->atomData.particlePot[atom1]);
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] += 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);
1177	+
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	+	#else
1189	+	pairwisePot += *(idat.pot);
1190	+
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	+
1203	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
1204	+	snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1205	+	snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1206	+	}
1207	+
1208	+	if (storageLayout_ & DataStorage::dslElectricField) {
1209	+	snap_->atomData.electricField[atom1] += *(idat.eField1);
1210	+	snap_->atomData.electricField[atom2] += *(idat.eField2);
1211	+	}
1212	+
1213	+	#endif
1214	+
1215	+	}
1216	+
1217	+	/*
1218	+	* buildNeighborList
1219	+	*
1220	+	* first element of pair is row-indexed CutoffGroup
1221	+	* second element of pair is column-indexed CutoffGroup
1222	+	*/
1223	+	vector<pair<int, int> > ForceMatrixDecomposition::buildNeighborList() {
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();
1232	+	#else
1233	+	cellList_.clear();
1234	+	#endif
1235	+
1236	+	RealType rList_ = (largestRcut_ + skinThickness_);
1237	+	RealType rl2 = rList_ * rList_;
1238	+	Snapshot* snap_ = sman_->getCurrentSnapshot();
1239	+	Mat3x3d Hmat = snap_->getHmat();
1240	+	Vector3d Hx = Hmat.getColumn(0);
1241	+	Vector3d Hy = Hmat.getColumn(1);
1242	+	Vector3d Hz = Hmat.getColumn(2);
1243	+
1244	+	nCells_.x() = (int) ( Hx.length() )/ rList_;
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;
1257	+	int cellIndex;
1258	+	int nCtot = nCells_.x() * nCells_.y() * nCells_.z();
1259	+
1260	+	#ifdef IS_MPI
1261	+	cellListRow_.resize(nCtot);
1262	+	cellListCol_.resize(nCtot);
1263	+	#else
1264	+	cellList_.resize(nCtot);
1265	+	#endif
1266	+
1267	+	if (!doAllPairs) {
1268	+	#ifdef IS_MPI
1269	+
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	+	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	+
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;
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	+
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_);
1352	+
1353	+	for (vector<Vector3i>::iterator os = cellOffsets_.begin();
1354	+	os != cellOffsets_.end(); ++os) {
1355	+
1356	+	Vector3i m2v = m1v + (*os);
1357	+
1358	+
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	+	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	+	}
1394	+	}
1395	+	}
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	+
1411	+
1412	+
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	+	}
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	+
1458	+	// save the local cutoff group positions for the check that is
1459	+	// done on each loop:
1460	+	saved_CG_positions_.clear();
1461	+	for (int i = 0; i < nGroups_; i++)
1462	+	saved_CG_positions_.push_back(snap_->cgData.position[i]);
1463	+
1464	+	return neighborList;
1465	+	}
1466	+	} //end namespace OpenMD

Diff Legend

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