ViewVC Help

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

root/OpenMD/branches/development/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 1798 by gezelter, Thu Sep 13 14:10:11 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();
171	>	for (int i = 0; i < nAtomsInRow_; i++)
172	>	atypesRow[i] = ff_->getAtomType(identsRow[i]);
173	>	for (int i = 0; i < nAtomsInCol_; i++)
174	>	atypesCol[i] = ff_->getAtomType(identsCol[i]);
175	>
176	>	pot_row.resize(nAtomsInRow_);
177	>	pot_col.resize(nAtomsInCol_);
178	>
179	>	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	<	// First time through, construct column stride.
287	<	if (communicators.size() == 0)
288	<	{
289	<	int nColumnsMax = (int) round(sqrt((float) nProcs));
290	<	for (int i = 0; i < nProcs; ++i)
291	<	{
292	<	if (nProcs%i==0) nColumns=i;
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	<	int nRows = nProcs/nColumns;
305	<	myRank_ = (int) worldRank%nColumns;
304	>
305	>	createGtypeCutoffMap();
306	>
307		}
308	<	else
309	<	{
310	<	myRank_ = myRank/nColumns;
311	<	}
312	<	MPI::Comm newComm = MPI:COMM_WORLD.Split(myRank_,0);
313	<
314	<	isColumn_ = false;
315	<
316	<	}
308	>
309	>	void ForceMatrixDecomposition::createGtypeCutoffMap() {
310	>
311	>	RealType tol = 1e-6;
312	>	largestRcut_ = 0.0;
313	>	int atid;
314	>	set<AtomType*> atypes = info_->getSimulatedAtomTypes();
315	>
316	>	map<int, RealType> atypeCutoff;
317	>
318	>	for (set<AtomType*>::iterator at = atypes.begin();
319	>	at != atypes.end(); ++at){
320	>	atid = (*at)->getIdent();
321	>	if (userChoseCutoff_)
322	>	atypeCutoff[atid] = userCutoff_;
323	>	else
324	>	atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at);
325	>	}
326	>
327	>	vector<RealType> gTypeCutoffs;
328	>	// first we do a single loop over the cutoff groups to find the
329	>	// largest cutoff for any atypes present in this group.
330	>	#ifdef IS_MPI
331	>	vector<RealType> groupCutoffRow(nGroupsInRow_, 0.0);
332	>	groupRowToGtype.resize(nGroupsInRow_);
333	>	for (int cg1 = 0; cg1 < nGroupsInRow_; cg1++) {
334	>	vector<int> atomListRow = getAtomsInGroupRow(cg1);
335	>	for (vector<int>::iterator ia = atomListRow.begin();
336	>	ia != atomListRow.end(); ++ia) {
337	>	int atom1 = (*ia);
338	>	atid = identsRow[atom1];
339	>	if (atypeCutoff[atid] > groupCutoffRow[cg1]) {
340	>	groupCutoffRow[cg1] = atypeCutoff[atid];
341	>	}
342	>	}
343
344	<	ForceDecomposition::gather(sendbuf, receivebuf){
345	<	communicators(myIndex_).Allgatherv();
346	<	}
344	>	bool gTypeFound = false;
345	>	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
346	>	if (abs(groupCutoffRow[cg1] - gTypeCutoffs[gt]) < tol) {
347	>	groupRowToGtype[cg1] = gt;
348	>	gTypeFound = true;
349	>	}
350	>	}
351	>	if (!gTypeFound) {
352	>	gTypeCutoffs.push_back( groupCutoffRow[cg1] );
353	>	groupRowToGtype[cg1] = gTypeCutoffs.size() - 1;
354	>	}
355	>
356	>	}
357	>	vector<RealType> groupCutoffCol(nGroupsInCol_, 0.0);
358	>	groupColToGtype.resize(nGroupsInCol_);
359	>	for (int cg2 = 0; cg2 < nGroupsInCol_; cg2++) {
360	>	vector<int> atomListCol = getAtomsInGroupColumn(cg2);
361	>	for (vector<int>::iterator jb = atomListCol.begin();
362	>	jb != atomListCol.end(); ++jb) {
363	>	int atom2 = (*jb);
364	>	atid = identsCol[atom2];
365	>	if (atypeCutoff[atid] > groupCutoffCol[cg2]) {
366	>	groupCutoffCol[cg2] = atypeCutoff[atid];
367	>	}
368	>	}
369	>	bool gTypeFound = false;
370	>	for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
371	>	if (abs(groupCutoffCol[cg2] - gTypeCutoffs[gt]) < tol) {
372	>	groupColToGtype[cg2] = gt;
373	>	gTypeFound = true;
374	>	}
375	>	}
376	>	if (!gTypeFound) {
377	>	gTypeCutoffs.push_back( groupCutoffCol[cg2] );
378	>	groupColToGtype[cg2] = gTypeCutoffs.size() - 1;
379	>	}
380	>	}
381	>	#else
382
383	+	vector<RealType> groupCutoff(nGroups_, 0.0);
384	+	groupToGtype.resize(nGroups_);
385	+	for (int cg1 = 0; cg1 < nGroups_; cg1++) {
386	+	groupCutoff[cg1] = 0.0;
387	+	vector<int> atomList = getAtomsInGroupRow(cg1);
388	+	for (vector<int>::iterator ia = atomList.begin();
389	+	ia != atomList.end(); ++ia) {
390	+	int atom1 = (*ia);
391	+	atid = idents[atom1];
392	+	if (atypeCutoff[atid] > groupCutoff[cg1])
393	+	groupCutoff[cg1] = atypeCutoff[atid];
394	+	}
395	+
396	+	bool gTypeFound = false;
397	+	for (unsigned int gt = 0; gt < gTypeCutoffs.size(); gt++) {
398	+	if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
399	+	groupToGtype[cg1] = gt;
400	+	gTypeFound = true;
401	+	}
402	+	}
403	+	if (!gTypeFound) {
404	+	gTypeCutoffs.push_back( groupCutoff[cg1] );
405	+	groupToGtype[cg1] = gTypeCutoffs.size() - 1;
406	+	}
407	+	}
408	+	#endif
409
410	+	// Now we find the maximum group cutoff value present in the simulation
411
412	<	ForceDecomposition::scatter(sbuffer, rbuffer){
413	<	communicators(myIndex_).Reduce_scatter(sbuffer, recevbuf. recvcounts, MPI::DOUBLE, MPI::SUM);
414	<	}
412	>	RealType groupMax = *max_element(gTypeCutoffs.begin(),
413	>	gTypeCutoffs.end());
414	>
415	>	#ifdef IS_MPI
416	>	MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE,
417	>	MPI::MAX);
418	>	#endif
419	>
420	>	RealType tradRcut = groupMax;
421	>
422	>	for (unsigned int i = 0; i < gTypeCutoffs.size();  i++) {
423	>	for (unsigned int j = 0; j < gTypeCutoffs.size();  j++) {
424	>	RealType thisRcut;
425	>	switch(cutoffPolicy_) {
426	>	case TRADITIONAL:
427	>	thisRcut = tradRcut;
428	>	break;
429	>	case MIX:
430	>	thisRcut = 0.5 * (gTypeCutoffs[i] + gTypeCutoffs[j]);
431	>	break;
432	>	case MAX:
433	>	thisRcut = max(gTypeCutoffs[i], gTypeCutoffs[j]);
434	>	break;
435	>	default:
436	>	sprintf(painCave.errMsg,
437	>	"ForceMatrixDecomposition::createGtypeCutoffMap "
438	>	"hit an unknown cutoff policy!\n");
439	>	painCave.severity = OPENMD_ERROR;
440	>	painCave.isFatal = 1;
441	>	simError();
442	>	break;
443	>	}
444	>
445	>	pair<int,int> key = make_pair(i,j);
446	>	gTypeCutoffMap[key].first = thisRcut;
447	>	if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
448	>	gTypeCutoffMap[key].second = thisRcut*thisRcut;
449	>	gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
450	>	// sanity check
451	>
452	>	if (userChoseCutoff_) {
453	>	if (abs(gTypeCutoffMap[key].first - userCutoff_) > 0.0001) {
454	>	sprintf(painCave.errMsg,
455	>	"ForceMatrixDecomposition::createGtypeCutoffMap "
456	>	"user-specified rCut (%lf) does not match computed group Cutoff\n", userCutoff_);
457	>	painCave.severity = OPENMD_ERROR;
458	>	painCave.isFatal = 1;
459	>	simError();
460	>	}
461	>	}
462	>	}
463	>	}
464	>	}
465	>
466	>	groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
467	>	int i, j;
468	>	#ifdef IS_MPI
469	>	i = groupRowToGtype[cg1];
470	>	j = groupColToGtype[cg2];
471	>	#else
472	>	i = groupToGtype[cg1];
473	>	j = groupToGtype[cg2];
474	>	#endif
475	>	return gTypeCutoffMap[make_pair(i,j)];
476	>	}
477	>
478	>	int ForceMatrixDecomposition::getTopologicalDistance(int atom1, int atom2) {
479	>	for (unsigned int j = 0; j < toposForAtom[atom1].size(); j++) {
480	>	if (toposForAtom[atom1][j] == atom2)
481	>	return topoDist[atom1][j];
482	>	}
483	>	return 0;
484	>	}
485	>
486	>	void ForceMatrixDecomposition::zeroWorkArrays() {
487	>	pairwisePot = 0.0;
488	>	embeddingPot = 0.0;
489	>	excludedPot = 0.0;
490	>	excludedSelfPot = 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	>	#endif
563	>	// even in parallel, we need to zero out the local arrays:
564	>
565	>	if (storageLayout_ & DataStorage::dslParticlePot) {
566	>	fill(snap_->atomData.particlePot.begin(),
567	>	snap_->atomData.particlePot.end(), 0.0);
568	>	}
569	>
570	>	if (storageLayout_ & DataStorage::dslDensity) {
571	>	fill(snap_->atomData.density.begin(),
572	>	snap_->atomData.density.end(), 0.0);
573	>	}
574	>
575	>	if (storageLayout_ & DataStorage::dslFunctional) {
576	>	fill(snap_->atomData.functional.begin(),
577	>	snap_->atomData.functional.end(), 0.0);
578	>	}
579	>
580	>	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
581	>	fill(snap_->atomData.functionalDerivative.begin(),
582	>	snap_->atomData.functionalDerivative.end(), 0.0);
583	>	}
584	>
585	>	if (storageLayout_ & DataStorage::dslSkippedCharge) {
586	>	fill(snap_->atomData.skippedCharge.begin(),
587	>	snap_->atomData.skippedCharge.end(), 0.0);
588	>	}
589	>
590	>	if (storageLayout_ & DataStorage::dslElectricField) {
591	>	fill(snap_->atomData.electricField.begin(),
592	>	snap_->atomData.electricField.end(), V3Zero);
593	>	}
594	>	}
595	>
596	>
597	>	void ForceMatrixDecomposition::distributeData()  {
598	>	snap_ = sman_->getCurrentSnapshot();
599	>	storageLayout_ = sman_->getStorageLayout();
600	>	#ifdef IS_MPI
601	>
602	>	// gather up the atomic positions
603	>	AtomPlanVectorRow->gather(snap_->atomData.position,
604	>	atomRowData.position);
605	>	AtomPlanVectorColumn->gather(snap_->atomData.position,
606	>	atomColData.position);
607	>
608	>	// gather up the cutoff group positions
609	>
610	>	cgPlanVectorRow->gather(snap_->cgData.position,
611	>	cgRowData.position);
612	>
613	>	cgPlanVectorColumn->gather(snap_->cgData.position,
614	>	cgColData.position);
615	>
616	>
617	>
618	>	if (needVelocities_) {
619	>	// gather up the atomic velocities
620	>	AtomPlanVectorColumn->gather(snap_->atomData.velocity,
621	>	atomColData.velocity);
622	>
623	>	cgPlanVectorColumn->gather(snap_->cgData.velocity,
624	>	cgColData.velocity);
625	>	}
626	>
627	>
628	>	// if needed, gather the atomic rotation matrices
629	>	if (storageLayout_ & DataStorage::dslAmat) {
630	>	AtomPlanMatrixRow->gather(snap_->atomData.aMat,
631	>	atomRowData.aMat);
632	>	AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
633	>	atomColData.aMat);
634	>	}
635	>
636	>	// if needed, gather the atomic eletrostatic information
637	>	if (storageLayout_ & DataStorage::dslDipole) {
638	>	AtomPlanVectorRow->gather(snap_->atomData.dipole,
639	>	atomRowData.dipole);
640	>	AtomPlanVectorColumn->gather(snap_->atomData.dipole,
641	>	atomColData.dipole);
642	>	}
643	>
644	>	if (storageLayout_ & DataStorage::dslQuadrupole) {
645	>	AtomPlanMatrixRow->gather(snap_->atomData.quadrupole,
646	>	atomRowData.quadrupole);
647	>	AtomPlanMatrixColumn->gather(snap_->atomData.quadrupole,
648	>	atomColData.quadrupole);
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();
668	>	#ifdef IS_MPI
669	>
670	>	if (storageLayout_ & DataStorage::dslDensity) {
671	>
672	>	AtomPlanRealRow->scatter(atomRowData.density,
673	>	snap_->atomData.density);
674	>
675	>	int n = snap_->atomData.density.size();
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	>
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	>	AtomPlanRealRow->gather(snap_->atomData.functional,
706	>	atomRowData.functional);
707	>	AtomPlanRealColumn->gather(snap_->atomData.functional,
708	>	atomColData.functional);
709	>	}
710	>
711	>	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
712	>	AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
713	>	atomRowData.functionalDerivative);
714	>	AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
715	>	atomColData.functionalDerivative);
716	>	}
717	>	#endif
718	>	}
719	>
720	>
721	>	void ForceMatrixDecomposition::collectData() {
722	>	snap_ = sman_->getCurrentSnapshot();
723	>	storageLayout_ = sman_->getStorageLayout();
724	>	#ifdef IS_MPI
725	>	int n = snap_->atomData.force.size();
726	>	vector<Vector3d> frc_tmp(n, V3Zero);
727	>
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	>	AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
735	>	for (int i = 0; i < n; i++) {
736	>	snap_->atomData.force[i] += frc_tmp[i];
737	>	}
738	>
739	>	if (storageLayout_ & DataStorage::dslTorque) {
740	>
741	>	int nt = snap_->atomData.torque.size();
742	>	vector<Vector3d> trq_tmp(nt, V3Zero);
743	>
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	>	AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
751	>	for (int i = 0; i < nt; i++)
752	>	snap_->atomData.torque[i] += trq_tmp[i];
753	>	}
754	>
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	>	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	>	for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
911	>	RealType ploc1 = excludedSelfPot[ii];
912	>	RealType ploc2 = 0.0;
913	>	MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
914	>	excludedSelfPot[ii] = ploc2;
915	>	}
916	>	#endif
917	>
918	>	}
919	>
920	>
921	>
922	>	int ForceMatrixDecomposition::getNAtomsInRow() {
923	>	#ifdef IS_MPI
924	>	return nAtomsInRow_;
925	>	#else
926	>	return nLocal_;
927	>	#endif
928	>	}
929	>
930	>	/**
931	>	* returns the list of atoms belonging to this group.
932	>	*/
933	>	vector<int> ForceMatrixDecomposition::getAtomsInGroupRow(int cg1){
934	>	#ifdef IS_MPI
935	>	return groupListRow_[cg1];
936	>	#else
937	>	return groupList_[cg1];
938	>	#endif
939	>	}
940	>
941	>	vector<int> ForceMatrixDecomposition::getAtomsInGroupColumn(int cg2){
942	>	#ifdef IS_MPI
943	>	return groupListCol_[cg2];
944	>	#else
945	>	return groupList_[cg2];
946	>	#endif
947	>	}
948	>
949	>	Vector3d ForceMatrixDecomposition::getIntergroupVector(int cg1, int cg2){
950	>	Vector3d d;
951	>
952	>	#ifdef IS_MPI
953	>	d = cgColData.position[cg2] - cgRowData.position[cg1];
954	>	#else
955	>	d = snap_->cgData.position[cg2] - snap_->cgData.position[cg1];
956	>	#endif
957	>
958	>	snap_->wrapVector(d);
959	>	return d;
960	>	}
961	>
962	>	Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
963	>	#ifdef IS_MPI
964	>	return cgColData.velocity[cg2];
965	>	#else
966	>	return snap_->cgData.velocity[cg2];
967	>	#endif
968	>	}
969	>
970	>	Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
971	>	#ifdef IS_MPI
972	>	return atomColData.velocity[atom2];
973	>	#else
974	>	return snap_->atomData.velocity[atom2];
975	>	#endif
976	>	}
977	>
978	>
979	>	Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
980	>
981	>	Vector3d d;
982	>
983	>	#ifdef IS_MPI
984	>	d = cgRowData.position[cg1] - atomRowData.position[atom1];
985	>	#else
986	>	d = snap_->cgData.position[cg1] - snap_->atomData.position[atom1];
987	>	#endif
988	>
989	>	snap_->wrapVector(d);
990	>	return d;
991	>	}
992	>
993	>	Vector3d ForceMatrixDecomposition::getAtomToGroupVectorColumn(int atom2, int cg2){
994	>	Vector3d d;
995	>
996	>	#ifdef IS_MPI
997	>	d = cgColData.position[cg2] - atomColData.position[atom2];
998	>	#else
999	>	d = snap_->cgData.position[cg2] - snap_->atomData.position[atom2];
1000	>	#endif
1001	>
1002	>	snap_->wrapVector(d);
1003	>	return d;
1004	>	}
1005	>
1006	>	RealType ForceMatrixDecomposition::getMassFactorRow(int atom1) {
1007	>	#ifdef IS_MPI
1008	>	return massFactorsRow[atom1];
1009	>	#else
1010	>	return massFactors[atom1];
1011	>	#endif
1012	>	}
1013	>
1014	>	RealType ForceMatrixDecomposition::getMassFactorColumn(int atom2) {
1015	>	#ifdef IS_MPI
1016	>	return massFactorsCol[atom2];
1017	>	#else
1018	>	return massFactors[atom2];
1019	>	#endif
1020	>
1021	>	}
1022	>
1023	>	Vector3d ForceMatrixDecomposition::getInteratomicVector(int atom1, int atom2){
1024	>	Vector3d d;
1025	>
1026	>	#ifdef IS_MPI
1027	>	d = atomColData.position[atom2] - atomRowData.position[atom1];
1028	>	#else
1029	>	d = snap_->atomData.position[atom2] - snap_->atomData.position[atom1];
1030	>	#endif
1031	>
1032	>	snap_->wrapVector(d);
1033	>	return d;
1034	>	}
1035	>
1036	>	vector<int> ForceMatrixDecomposition::getExcludesForAtom(int atom1) {
1037	>	return excludesForAtom[atom1];
1038	>	}
1039	>
1040	>	/**
1041	>	* We need to exclude some overcounted interactions that result from
1042	>	* the parallel decomposition.
1043	>	*/
1044	>	bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) {
1045	>	int unique_id_1, unique_id_2;
1046	>
1047	>	#ifdef IS_MPI
1048	>	// in MPI, we have to look up the unique IDs for each atom
1049	>	unique_id_1 = AtomRowToGlobal[atom1];
1050	>	unique_id_2 = AtomColToGlobal[atom2];
1051	>	// group1 = cgRowToGlobal[cg1];
1052	>	// group2 = cgColToGlobal[cg2];
1053	>	#else
1054	>	unique_id_1 = AtomLocalToGlobal[atom1];
1055	>	unique_id_2 = AtomLocalToGlobal[atom2];
1056	>	int group1 = cgLocalToGlobal[cg1];
1057	>	int group2 = cgLocalToGlobal[cg2];
1058	>	#endif
1059	>
1060	>	if (unique_id_1 == unique_id_2) return true;
1061	>
1062	>	#ifdef IS_MPI
1063	>	// this prevents us from doing the pair on multiple processors
1064	>	if (unique_id_1 < unique_id_2) {
1065	>	if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1066	>	} else {
1067	>	if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1068	>	}
1069	>	#endif
1070	>
1071	>	#ifndef IS_MPI
1072	>	if (group1 == group2) {
1073	>	if (unique_id_1 < unique_id_2) return true;
1074	>	}
1075	>	#endif
1076	>
1077	>	return false;
1078	>	}
1079	>
1080	>	/**
1081	>	* We need to handle the interactions for atoms who are involved in
1082	>	* the same rigid body as well as some short range interactions
1083	>	* (bonds, bends, torsions) differently from other interactions.
1084	>	* We'll still visit the pairwise routines, but with a flag that
1085	>	* tells those routines to exclude the pair from direct long range
1086	>	* interactions.  Some indirect interactions (notably reaction
1087	>	* field) must still be handled for these pairs.
1088	>	*/
1089	>	bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1090	>
1091	>	// excludesForAtom was constructed to use row/column indices in the MPI
1092	>	// version, and to use local IDs in the non-MPI version:
1093	>
1094	>	for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1095	>	i != excludesForAtom[atom1].end(); ++i) {
1096	>	if ( (*i) == atom2 ) return true;
1097	>	}
1098	>
1099	>	return false;
1100	>	}
1101	>
1102	>
1103	>	void ForceMatrixDecomposition::addForceToAtomRow(int atom1, Vector3d fg){
1104	>	#ifdef IS_MPI
1105	>	atomRowData.force[atom1] += fg;
1106	>	#else
1107	>	snap_->atomData.force[atom1] += fg;
1108	>	#endif
1109	>	}
1110	>
1111	>	void ForceMatrixDecomposition::addForceToAtomColumn(int atom2, Vector3d fg){
1112	>	#ifdef IS_MPI
1113	>	atomColData.force[atom2] += fg;
1114	>	#else
1115	>	snap_->atomData.force[atom2] += fg;
1116	>	#endif
1117	>	}
1118	>
1119	>	// filling interaction blocks with pointers
1120	>	void ForceMatrixDecomposition::fillInteractionData(InteractionData &idat,
1121	>	int atom1, int atom2) {
1122	>
1123	>	idat.excluded = excludeAtomPair(atom1, atom2);
1124	>
1125	>	#ifdef IS_MPI
1126	>	idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
1127	>	//idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
1128	>	//                         ff_->getAtomType(identsCol[atom2]) );
1129	>
1130	>	if (storageLayout_ & DataStorage::dslAmat) {
1131	>	idat.A1 = &(atomRowData.aMat[atom1]);
1132	>	idat.A2 = &(atomColData.aMat[atom2]);
1133	>	}
1134	>
1135	>	if (storageLayout_ & DataStorage::dslTorque) {
1136	>	idat.t1 = &(atomRowData.torque[atom1]);
1137	>	idat.t2 = &(atomColData.torque[atom2]);
1138	>	}
1139	>
1140	>	if (storageLayout_ & DataStorage::dslDipole) {
1141	>	idat.dipole1 = &(atomRowData.dipole[atom1]);
1142	>	idat.dipole2 = &(atomColData.dipole[atom2]);
1143	>	}
1144	>
1145	>	if (storageLayout_ & DataStorage::dslQuadrupole) {
1146	>	idat.quadrupole1 = &(atomRowData.quadrupole[atom1]);
1147	>	idat.quadrupole2 = &(atomColData.quadrupole[atom2]);
1148	>	}
1149	>
1150	>	if (storageLayout_ & DataStorage::dslDensity) {
1151	>	idat.rho1 = &(atomRowData.density[atom1]);
1152	>	idat.rho2 = &(atomColData.density[atom2]);
1153	>	}
1154	>
1155	>	if (storageLayout_ & DataStorage::dslFunctional) {
1156	>	idat.frho1 = &(atomRowData.functional[atom1]);
1157	>	idat.frho2 = &(atomColData.functional[atom2]);
1158	>	}
1159	>
1160	>	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
1161	>	idat.dfrho1 = &(atomRowData.functionalDerivative[atom1]);
1162	>	idat.dfrho2 = &(atomColData.functionalDerivative[atom2]);
1163	>	}
1164
1165	+	if (storageLayout_ & DataStorage::dslParticlePot) {
1166	+	idat.particlePot1 = &(atomRowData.particlePot[atom1]);
1167	+	idat.particlePot2 = &(atomColData.particlePot[atom2]);
1168	+	}
1169
1170	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1171	+	idat.skippedCharge1 = &(atomRowData.skippedCharge[atom1]);
1172	+	idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1173	+	}
1174	+
1175	+	if (storageLayout_ & DataStorage::dslFlucQPosition) {
1176	+	idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1177	+	idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1178	+	}
1179	+
1180	+	#else
1181	+
1182	+	idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1183	+
1184	+	if (storageLayout_ & DataStorage::dslAmat) {
1185	+	idat.A1 = &(snap_->atomData.aMat[atom1]);
1186	+	idat.A2 = &(snap_->atomData.aMat[atom2]);
1187	+	}
1188	+
1189	+	if (storageLayout_ & DataStorage::dslTorque) {
1190	+	idat.t1 = &(snap_->atomData.torque[atom1]);
1191	+	idat.t2 = &(snap_->atomData.torque[atom2]);
1192	+	}
1193	+
1194	+	if (storageLayout_ & DataStorage::dslDipole) {
1195	+	idat.dipole1 = &(snap_->atomData.dipole[atom1]);
1196	+	idat.dipole2 = &(snap_->atomData.dipole[atom2]);
1197	+	}
1198	+
1199	+	if (storageLayout_ & DataStorage::dslQuadrupole) {
1200	+	idat.quadrupole1 = &(snap_->atomData.quadrupole[atom1]);
1201	+	idat.quadrupole2 = &(snap_->atomData.quadrupole[atom2]);
1202	+	}
1203	+
1204	+	if (storageLayout_ & DataStorage::dslDensity) {
1205	+	idat.rho1 = &(snap_->atomData.density[atom1]);
1206	+	idat.rho2 = &(snap_->atomData.density[atom2]);
1207	+	}
1208	+
1209	+	if (storageLayout_ & DataStorage::dslFunctional) {
1210	+	idat.frho1 = &(snap_->atomData.functional[atom1]);
1211	+	idat.frho2 = &(snap_->atomData.functional[atom2]);
1212	+	}
1213	+
1214	+	if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
1215	+	idat.dfrho1 = &(snap_->atomData.functionalDerivative[atom1]);
1216	+	idat.dfrho2 = &(snap_->atomData.functionalDerivative[atom2]);
1217	+	}
1218	+
1219	+	if (storageLayout_ & DataStorage::dslParticlePot) {
1220	+	idat.particlePot1 = &(snap_->atomData.particlePot[atom1]);
1221	+	idat.particlePot2 = &(snap_->atomData.particlePot[atom2]);
1222	+	}
1223	+
1224	+	if (storageLayout_ & DataStorage::dslSkippedCharge) {
1225	+	idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1226	+	idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1227	+	}
1228	+
1229	+	if (storageLayout_ & DataStorage::dslFlucQPosition) {
1230	+	idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1231	+	idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1232	+	}
1233	+
1234	+	#endif
1235	+	}
1236	+
1237	+
1238	+	void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {
1239	+	#ifdef IS_MPI
1240	+	pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1241	+	pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1242	+	expot_row[atom1] += RealType(0.5) *  *(idat.excludedPot);
1243	+	expot_col[atom2] += RealType(0.5) *  *(idat.excludedPot);
1244	+
1245	+	atomRowData.force[atom1] += *(idat.f1);
1246	+	atomColData.force[atom2] -= *(idat.f1);
1247	+
1248	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
1249	+	atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1250	+	atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1251	+	}
1252	+
1253	+	if (storageLayout_ & DataStorage::dslElectricField) {
1254	+	atomRowData.electricField[atom1] += *(idat.eField1);
1255	+	atomColData.electricField[atom2] += *(idat.eField2);
1256	+	}
1257	+
1258	+	#else
1259	+	pairwisePot += *(idat.pot);
1260	+	excludedPot += *(idat.excludedPot);
1261	+
1262	+	snap_->atomData.force[atom1] += *(idat.f1);
1263	+	snap_->atomData.force[atom2] -= *(idat.f1);
1264	+
1265	+	if (idat.doParticlePot) {
1266	+	// This is the pairwise contribution to the particle pot.  The
1267	+	// embedding contribution is added in each of the low level
1268	+	// non-bonded routines.  In parallel, this calculation is done
1269	+	// in collectData, not in unpackInteractionData.
1270	+	snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1271	+	snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1272	+	}
1273	+
1274	+	if (storageLayout_ & DataStorage::dslFlucQForce) {
1275	+	snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1276	+	snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1277	+	}
1278	+
1279	+	if (storageLayout_ & DataStorage::dslElectricField) {
1280	+	snap_->atomData.electricField[atom1] += *(idat.eField1);
1281	+	snap_->atomData.electricField[atom2] += *(idat.eField2);
1282	+	}
1283	+
1284	+	#endif
1285	+
1286	+	}
1287	+
1288	+	/*
1289	+	* buildNeighborList
1290	+	*
1291	+	* first element of pair is row-indexed CutoffGroup
1292	+	* second element of pair is column-indexed CutoffGroup
1293	+	*/
1294	+	vector<pair<int, int> > ForceMatrixDecomposition::buildNeighborList() {
1295	+
1296	+	vector<pair<int, int> > neighborList;
1297	+	groupCutoffs cuts;
1298	+	bool doAllPairs = false;
1299	+
1300	+	#ifdef IS_MPI
1301	+	cellListRow_.clear();
1302	+	cellListCol_.clear();
1303	+	#else
1304	+	cellList_.clear();
1305	+	#endif
1306	+
1307	+	RealType rList_ = (largestRcut_ + skinThickness_);
1308	+	RealType rl2 = rList_ * rList_;
1309	+	Snapshot* snap_ = sman_->getCurrentSnapshot();
1310	+	Mat3x3d Hmat = snap_->getHmat();
1311	+	Vector3d Hx = Hmat.getColumn(0);
1312	+	Vector3d Hy = Hmat.getColumn(1);
1313	+	Vector3d Hz = Hmat.getColumn(2);
1314	+
1315	+	nCells_.x() = (int) ( Hx.length() )/ rList_;
1316	+	nCells_.y() = (int) ( Hy.length() )/ rList_;
1317	+	nCells_.z() = (int) ( Hz.length() )/ rList_;
1318	+
1319	+	// handle small boxes where the cell offsets can end up repeating cells
1320	+
1321	+	if (nCells_.x() < 3) doAllPairs = true;
1322	+	if (nCells_.y() < 3) doAllPairs = true;
1323	+	if (nCells_.z() < 3) doAllPairs = true;
1324	+
1325	+	Mat3x3d invHmat = snap_->getInvHmat();
1326	+	Vector3d rs, scaled, dr;
1327	+	Vector3i whichCell;
1328	+	int cellIndex;
1329	+	int nCtot = nCells_.x() * nCells_.y() * nCells_.z();
1330	+
1331	+	#ifdef IS_MPI
1332	+	cellListRow_.resize(nCtot);
1333	+	cellListCol_.resize(nCtot);
1334	+	#else
1335	+	cellList_.resize(nCtot);
1336	+	#endif
1337	+
1338	+	if (!doAllPairs) {
1339	+	#ifdef IS_MPI
1340	+
1341	+	for (int i = 0; i < nGroupsInRow_; i++) {
1342	+	rs = cgRowData.position[i];
1343	+
1344	+	// scaled positions relative to the box vectors
1345	+	scaled = invHmat * rs;
1346	+
1347	+	// wrap the vector back into the unit box by subtracting integer box
1348	+	// numbers
1349	+	for (int j = 0; j < 3; j++) {
1350	+	scaled[j] -= roundMe(scaled[j]);
1351	+	scaled[j] += 0.5;
1352	+	// Handle the special case when an object is exactly on the
1353	+	// boundary (a scaled coordinate of 1.0 is the same as
1354	+	// scaled coordinate of 0.0)
1355	+	if (scaled[j] >= 1.0) scaled[j] -= 1.0;
1356	+	}
1357	+
1358	+	// find xyz-indices of cell that cutoffGroup is in.
1359	+	whichCell.x() = nCells_.x() * scaled.x();
1360	+	whichCell.y() = nCells_.y() * scaled.y();
1361	+	whichCell.z() = nCells_.z() * scaled.z();
1362	+
1363	+	// find single index of this cell:
1364	+	cellIndex = Vlinear(whichCell, nCells_);
1365	+
1366	+	// add this cutoff group to the list of groups in this cell;
1367	+	cellListRow_[cellIndex].push_back(i);
1368	+	}
1369	+	for (int i = 0; i < nGroupsInCol_; i++) {
1370	+	rs = cgColData.position[i];
1371	+
1372	+	// scaled positions relative to the box vectors
1373	+	scaled = invHmat * rs;
1374	+
1375	+	// wrap the vector back into the unit box by subtracting integer box
1376	+	// numbers
1377	+	for (int j = 0; j < 3; j++) {
1378	+	scaled[j] -= roundMe(scaled[j]);
1379	+	scaled[j] += 0.5;
1380	+	// Handle the special case when an object is exactly on the
1381	+	// boundary (a scaled coordinate of 1.0 is the same as
1382	+	// scaled coordinate of 0.0)
1383	+	if (scaled[j] >= 1.0) scaled[j] -= 1.0;
1384	+	}
1385	+
1386	+	// find xyz-indices of cell that cutoffGroup is in.
1387	+	whichCell.x() = nCells_.x() * scaled.x();
1388	+	whichCell.y() = nCells_.y() * scaled.y();
1389	+	whichCell.z() = nCells_.z() * scaled.z();
1390	+
1391	+	// find single index of this cell:
1392	+	cellIndex = Vlinear(whichCell, nCells_);
1393	+
1394	+	// add this cutoff group to the list of groups in this cell;
1395	+	cellListCol_[cellIndex].push_back(i);
1396	+	}
1397	+
1398	+	#else
1399	+	for (int i = 0; i < nGroups_; i++) {
1400	+	rs = snap_->cgData.position[i];
1401	+
1402	+	// scaled positions relative to the box vectors
1403	+	scaled = invHmat * rs;
1404	+
1405	+	// wrap the vector back into the unit box by subtracting integer box
1406	+	// numbers
1407	+	for (int j = 0; j < 3; j++) {
1408	+	scaled[j] -= roundMe(scaled[j]);
1409	+	scaled[j] += 0.5;
1410	+	// Handle the special case when an object is exactly on the
1411	+	// boundary (a scaled coordinate of 1.0 is the same as
1412	+	// scaled coordinate of 0.0)
1413	+	if (scaled[j] >= 1.0) scaled[j] -= 1.0;
1414	+	}
1415	+
1416	+	// find xyz-indices of cell that cutoffGroup is in.
1417	+	whichCell.x() = nCells_.x() * scaled.x();
1418	+	whichCell.y() = nCells_.y() * scaled.y();
1419	+	whichCell.z() = nCells_.z() * scaled.z();
1420	+
1421	+	// find single index of this cell:
1422	+	cellIndex = Vlinear(whichCell, nCells_);
1423	+
1424	+	// add this cutoff group to the list of groups in this cell;
1425	+	cellList_[cellIndex].push_back(i);
1426	+	}
1427	+
1428	+	#endif
1429	+
1430	+	for (int m1z = 0; m1z < nCells_.z(); m1z++) {
1431	+	for (int m1y = 0; m1y < nCells_.y(); m1y++) {
1432	+	for (int m1x = 0; m1x < nCells_.x(); m1x++) {
1433	+	Vector3i m1v(m1x, m1y, m1z);
1434	+	int m1 = Vlinear(m1v, nCells_);
1435	+
1436	+	for (vector<Vector3i>::iterator os = cellOffsets_.begin();
1437	+	os != cellOffsets_.end(); ++os) {
1438	+
1439	+	Vector3i m2v = m1v + (*os);
1440	+
1441	+
1442	+	if (m2v.x() >= nCells_.x()) {
1443	+	m2v.x() = 0;
1444	+	} else if (m2v.x() < 0) {
1445	+	m2v.x() = nCells_.x() - 1;
1446	+	}
1447	+
1448	+	if (m2v.y() >= nCells_.y()) {
1449	+	m2v.y() = 0;
1450	+	} else if (m2v.y() < 0) {
1451	+	m2v.y() = nCells_.y() - 1;
1452	+	}
1453	+
1454	+	if (m2v.z() >= nCells_.z()) {
1455	+	m2v.z() = 0;
1456	+	} else if (m2v.z() < 0) {
1457	+	m2v.z() = nCells_.z() - 1;
1458	+	}
1459	+
1460	+	int m2 = Vlinear (m2v, nCells_);
1461	+
1462	+	#ifdef IS_MPI
1463	+	for (vector<int>::iterator j1 = cellListRow_[m1].begin();
1464	+	j1 != cellListRow_[m1].end(); ++j1) {
1465	+	for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1466	+	j2 != cellListCol_[m2].end(); ++j2) {
1467	+
1468	+	// In parallel, we need to visit *all* pairs of row
1469	+	// & column indicies and will divide labor in the
1470	+	// force evaluation later.
1471	+	dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1472	+	snap_->wrapVector(dr);
1473	+	cuts = getGroupCutoffs( (*j1), (*j2) );
1474	+	if (dr.lengthSquare() < cuts.third) {
1475	+	neighborList.push_back(make_pair((*j1), (*j2)));
1476	+	}
1477	+	}
1478	+	}
1479	+	#else
1480	+	for (vector<int>::iterator j1 = cellList_[m1].begin();
1481	+	j1 != cellList_[m1].end(); ++j1) {
1482	+	for (vector<int>::iterator j2 = cellList_[m2].begin();
1483	+	j2 != cellList_[m2].end(); ++j2) {
1484	+
1485	+	// Always do this if we're in different cells or if
1486	+	// we're in the same cell and the global index of
1487	+	// the j2 cutoff group is greater than or equal to
1488	+	// the j1 cutoff group.  Note that Rappaport's code
1489	+	// has a "less than" conditional here, but that
1490	+	// deals with atom-by-atom computation.  OpenMD
1491	+	// allows atoms within a single cutoff group to
1492	+	// interact with each other.
1493	+
1494	+
1495	+
1496	+	if (m2 != m1 || (*j2) >= (*j1) ) {
1497	+
1498	+	dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1499	+	snap_->wrapVector(dr);
1500	+	cuts = getGroupCutoffs( (*j1), (*j2) );
1501	+	if (dr.lengthSquare() < cuts.third) {
1502	+	neighborList.push_back(make_pair((*j1), (*j2)));
1503	+	}
1504	+	}
1505	+	}
1506	+	}
1507	+	#endif
1508	+	}
1509	+	}
1510	+	}
1511	+	}
1512	+	} else {
1513	+	// branch to do all cutoff group pairs
1514	+	#ifdef IS_MPI
1515	+	for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1516	+	for (int j2 = 0; j2 < nGroupsInCol_; j2++) {
1517	+	dr = cgColData.position[j2] - cgRowData.position[j1];
1518	+	snap_->wrapVector(dr);
1519	+	cuts = getGroupCutoffs( j1, j2 );
1520	+	if (dr.lengthSquare() < cuts.third) {
1521	+	neighborList.push_back(make_pair(j1, j2));
1522	+	}
1523	+	}
1524	+	}
1525	+	#else
1526	+	// include all groups here.
1527	+	for (int j1 = 0; j1 < nGroups_; j1++) {
1528	+	// include self group interactions j2 == j1
1529	+	for (int j2 = j1; j2 < nGroups_; j2++) {
1530	+	dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1531	+	snap_->wrapVector(dr);
1532	+	cuts = getGroupCutoffs( j1, j2 );
1533	+	if (dr.lengthSquare() < cuts.third) {
1534	+	neighborList.push_back(make_pair(j1, j2));
1535	+	}
1536	+	}
1537	+	}
1538	+	#endif
1539	+	}
1540	+
1541	+	// save the local cutoff group positions for the check that is
1542	+	// done on each loop:
1543	+	saved_CG_positions_.clear();
1544	+	for (int i = 0; i < nGroups_; i++)
1545	+	saved_CG_positions_.push_back(snap_->cgData.position[i]);
1546	+
1547	+	return neighborList;
1548	+	}
1549	+	} //end namespace OpenMD

Diff Legend

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