ViewVC Help
View File | Revision Log | Show Annotations | View Changeset | Root Listing
root/OpenMD/trunk/src/parallel/ForceMatrixDecomposition.cpp
(Generate patch)

Comparing branches/development/src/parallel/ForceMatrixDecomposition.cpp (file contents):
Revision 1590 by gezelter, Mon Jul 11 01:39:49 2011 UTC vs.
Revision 1756 by gezelter, Mon Jun 18 18:23:20 2012 UTC

# Line 36 | Line 36
36   * [1]  Meineke, et al., J. Comp. Chem. 26, 252-271 (2005).            
37   * [2]  Fennell & Gezelter, J. Chem. Phys. 124, 234104 (2006).          
38   * [3]  Sun, Lin & Gezelter, J. Chem. Phys. 128, 24107 (2008).          
39 < * [4]  Vardeman & Gezelter, in progress (2009).                        
39 > * [4]  Kuang & Gezelter,  J. Chem. Phys. 133, 164101 (2010).
40 > * [5]  Vardeman, Stocker & Gezelter, J. Chem. Theory Comput. 7, 834 (2011).
41   */
42   #include "parallel/ForceMatrixDecomposition.hpp"
43   #include "math/SquareMatrix3.hpp"
# Line 47 | Line 48 | namespace OpenMD {
48   using namespace std;
49   namespace OpenMD {
50  
51 +  ForceMatrixDecomposition::ForceMatrixDecomposition(SimInfo* info, InteractionManager* iMan) : ForceDecomposition(info, iMan) {
52 +
53 +    // In a parallel computation, row and colum scans must visit all
54 +    // surrounding cells (not just the 14 upper triangular blocks that
55 +    // are used when the processor can see all pairs)
56 + #ifdef IS_MPI
57 +    cellOffsets_.clear();
58 +    cellOffsets_.push_back( Vector3i(-1,-1,-1) );
59 +    cellOffsets_.push_back( Vector3i( 0,-1,-1) );
60 +    cellOffsets_.push_back( Vector3i( 1,-1,-1) );                          
61 +    cellOffsets_.push_back( Vector3i(-1, 0,-1) );
62 +    cellOffsets_.push_back( Vector3i( 0, 0,-1) );
63 +    cellOffsets_.push_back( Vector3i( 1, 0,-1) );
64 +    cellOffsets_.push_back( Vector3i(-1, 1,-1) );
65 +    cellOffsets_.push_back( Vector3i( 0, 1,-1) );      
66 +    cellOffsets_.push_back( Vector3i( 1, 1,-1) );
67 +    cellOffsets_.push_back( Vector3i(-1,-1, 0) );
68 +    cellOffsets_.push_back( Vector3i( 0,-1, 0) );
69 +    cellOffsets_.push_back( Vector3i( 1,-1, 0) );
70 +    cellOffsets_.push_back( Vector3i(-1, 0, 0) );      
71 +    cellOffsets_.push_back( Vector3i( 0, 0, 0) );
72 +    cellOffsets_.push_back( Vector3i( 1, 0, 0) );
73 +    cellOffsets_.push_back( Vector3i(-1, 1, 0) );
74 +    cellOffsets_.push_back( Vector3i( 0, 1, 0) );
75 +    cellOffsets_.push_back( Vector3i( 1, 1, 0) );
76 +    cellOffsets_.push_back( Vector3i(-1,-1, 1) );
77 +    cellOffsets_.push_back( Vector3i( 0,-1, 1) );
78 +    cellOffsets_.push_back( Vector3i( 1,-1, 1) );
79 +    cellOffsets_.push_back( Vector3i(-1, 0, 1) );
80 +    cellOffsets_.push_back( Vector3i( 0, 0, 1) );
81 +    cellOffsets_.push_back( Vector3i( 1, 0, 1) );
82 +    cellOffsets_.push_back( Vector3i(-1, 1, 1) );
83 +    cellOffsets_.push_back( Vector3i( 0, 1, 1) );
84 +    cellOffsets_.push_back( Vector3i( 1, 1, 1) );
85 + #endif    
86 +  }
87 +
88 +
89    /**
90     * distributeInitialData is essentially a copy of the older fortran
91     * SimulationSetup
92     */
54 –  
93    void ForceMatrixDecomposition::distributeInitialData() {
94      snap_ = sman_->getCurrentSnapshot();
95      storageLayout_ = sman_->getStorageLayout();
96      ff_ = info_->getForceField();
97      nLocal_ = snap_->getNumberOfAtoms();
98 <    
98 >  
99      nGroups_ = info_->getNLocalCutoffGroups();
100      // gather the information for atomtype IDs (atids):
101      idents = info_->getIdentArray();
# Line 71 | Line 109 | namespace OpenMD {
109      PairList* oneTwo = info_->getOneTwoInteractions();
110      PairList* oneThree = info_->getOneThreeInteractions();
111      PairList* oneFour = info_->getOneFourInteractions();
112 <
112 >    
113 >    if (needVelocities_)
114 >      snap_->cgData.setStorageLayout(DataStorage::dslPosition |
115 >                                     DataStorage::dslVelocity);
116 >    else
117 >      snap_->cgData.setStorageLayout(DataStorage::dslPosition);
118 >    
119   #ifdef IS_MPI
120  
121 <    AtomCommIntRow = new Communicator<Row,int>(nLocal_);
122 <    AtomCommRealRow = new Communicator<Row,RealType>(nLocal_);
79 <    AtomCommVectorRow = new Communicator<Row,Vector3d>(nLocal_);
80 <    AtomCommMatrixRow = new Communicator<Row,Mat3x3d>(nLocal_);
81 <    AtomCommPotRow = new Communicator<Row,potVec>(nLocal_);
121 >    MPI::Intracomm row = rowComm.getComm();
122 >    MPI::Intracomm col = colComm.getComm();
123  
124 <    AtomCommIntColumn = new Communicator<Column,int>(nLocal_);
125 <    AtomCommRealColumn = new Communicator<Column,RealType>(nLocal_);
126 <    AtomCommVectorColumn = new Communicator<Column,Vector3d>(nLocal_);
127 <    AtomCommMatrixColumn = new Communicator<Column,Mat3x3d>(nLocal_);
128 <    AtomCommPotColumn = new Communicator<Column,potVec>(nLocal_);
124 >    AtomPlanIntRow = new Plan<int>(row, nLocal_);
125 >    AtomPlanRealRow = new Plan<RealType>(row, nLocal_);
126 >    AtomPlanVectorRow = new Plan<Vector3d>(row, nLocal_);
127 >    AtomPlanMatrixRow = new Plan<Mat3x3d>(row, nLocal_);
128 >    AtomPlanPotRow = new Plan<potVec>(row, nLocal_);
129  
130 <    cgCommIntRow = new Communicator<Row,int>(nGroups_);
131 <    cgCommVectorRow = new Communicator<Row,Vector3d>(nGroups_);
132 <    cgCommIntColumn = new Communicator<Column,int>(nGroups_);
133 <    cgCommVectorColumn = new Communicator<Column,Vector3d>(nGroups_);
130 >    AtomPlanIntColumn = new Plan<int>(col, nLocal_);
131 >    AtomPlanRealColumn = new Plan<RealType>(col, nLocal_);
132 >    AtomPlanVectorColumn = new Plan<Vector3d>(col, nLocal_);
133 >    AtomPlanMatrixColumn = new Plan<Mat3x3d>(col, nLocal_);
134 >    AtomPlanPotColumn = new Plan<potVec>(col, nLocal_);
135  
136 <    nAtomsInRow_ = AtomCommIntRow->getSize();
137 <    nAtomsInCol_ = AtomCommIntColumn->getSize();
138 <    nGroupsInRow_ = cgCommIntRow->getSize();
139 <    nGroupsInCol_ = cgCommIntColumn->getSize();
136 >    cgPlanIntRow = new Plan<int>(row, nGroups_);
137 >    cgPlanVectorRow = new Plan<Vector3d>(row, nGroups_);
138 >    cgPlanIntColumn = new Plan<int>(col, nGroups_);
139 >    cgPlanVectorColumn = new Plan<Vector3d>(col, nGroups_);
140  
141 +    nAtomsInRow_ = AtomPlanIntRow->getSize();
142 +    nAtomsInCol_ = AtomPlanIntColumn->getSize();
143 +    nGroupsInRow_ = cgPlanIntRow->getSize();
144 +    nGroupsInCol_ = cgPlanIntColumn->getSize();
145 +
146      // Modify the data storage objects with the correct layouts and sizes:
147      atomRowData.resize(nAtomsInRow_);
148      atomRowData.setStorageLayout(storageLayout_);
# Line 104 | Line 151 | namespace OpenMD {
151      cgRowData.resize(nGroupsInRow_);
152      cgRowData.setStorageLayout(DataStorage::dslPosition);
153      cgColData.resize(nGroupsInCol_);
154 <    cgColData.setStorageLayout(DataStorage::dslPosition);
155 <        
154 >    if (needVelocities_)
155 >      // we only need column velocities if we need them.
156 >      cgColData.setStorageLayout(DataStorage::dslPosition |
157 >                                 DataStorage::dslVelocity);
158 >    else    
159 >      cgColData.setStorageLayout(DataStorage::dslPosition);
160 >      
161      identsRow.resize(nAtomsInRow_);
162      identsCol.resize(nAtomsInCol_);
163      
164 <    AtomCommIntRow->gather(idents, identsRow);
165 <    AtomCommIntColumn->gather(idents, identsCol);
164 >    AtomPlanIntRow->gather(idents, identsRow);
165 >    AtomPlanIntColumn->gather(idents, identsCol);
166      
167      // allocate memory for the parallel objects
168 +    atypesRow.resize(nAtomsInRow_);
169 +    atypesCol.resize(nAtomsInCol_);
170 +
171 +    for (int i = 0; i < nAtomsInRow_; i++)
172 +      atypesRow[i] = ff_->getAtomType(identsRow[i]);
173 +    for (int i = 0; i < nAtomsInCol_; i++)
174 +      atypesCol[i] = ff_->getAtomType(identsCol[i]);        
175 +
176 +    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 <    pot_row.resize(nAtomsInRow_);
192 <    pot_col.resize(nAtomsInCol_);
191 >    AtomPlanRealRow->gather(massFactors, massFactorsRow);
192 >    AtomPlanRealColumn->gather(massFactors, massFactorsCol);
193  
125 –    AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
126 –    AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
127 –    
128 –    cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
129 –    cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
130 –
131 –    AtomCommRealRow->gather(massFactors, massFactorsRow);
132 –    AtomCommRealColumn->gather(massFactors, massFactorsCol);
133 –
194      groupListRow_.clear();
195      groupListRow_.resize(nGroupsInRow_);
196      for (int i = 0; i < nGroupsInRow_; i++) {
# Line 185 | Line 245 | namespace OpenMD {
245        }      
246      }
247  
248 < #endif
189 <
190 <    groupList_.clear();
191 <    groupList_.resize(nGroups_);
192 <    for (int i = 0; i < nGroups_; i++) {
193 <      int gid = cgLocalToGlobal[i];
194 <      for (int j = 0; j < nLocal_; j++) {
195 <        int aid = AtomLocalToGlobal[j];
196 <        if (globalGroupMembership[aid] == gid) {
197 <          groupList_[i].push_back(j);
198 <        }
199 <      }      
200 <    }
201 <
248 > #else
249      excludesForAtom.clear();
250      excludesForAtom.resize(nLocal_);
251      toposForAtom.clear();
# Line 231 | Line 278 | namespace OpenMD {
278          }
279        }      
280      }
281 <    
281 > #endif
282 >
283 >    // allocate memory for the parallel objects
284 >    atypesLocal.resize(nLocal_);
285 >
286 >    for (int i = 0; i < nLocal_; i++)
287 >      atypesLocal[i] = ff_->getAtomType(idents[i]);
288 >
289 >    groupList_.clear();
290 >    groupList_.resize(nGroups_);
291 >    for (int i = 0; i < nGroups_; i++) {
292 >      int gid = cgLocalToGlobal[i];
293 >      for (int j = 0; j < nLocal_; j++) {
294 >        int aid = AtomLocalToGlobal[j];
295 >        if (globalGroupMembership[aid] == gid) {
296 >          groupList_[i].push_back(j);
297 >        }
298 >      }      
299 >    }
300 >
301 >
302      createGtypeCutoffMap();
303  
304    }
# Line 239 | Line 306 | namespace OpenMD {
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();
# Line 249 | Line 318 | namespace OpenMD {
318        atid = (*at)->getIdent();
319        if (userChoseCutoff_)
320          atypeCutoff[atid] = userCutoff_;
321 <      else
321 >      else
322          atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at);
323      }
324 <
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.
# Line 312 | Line 381 | namespace OpenMD {
381      vector<RealType> groupCutoff(nGroups_, 0.0);
382      groupToGtype.resize(nGroups_);
383      for (int cg1 = 0; cg1 < nGroups_; cg1++) {
315 –
384        groupCutoff[cg1] = 0.0;
385        vector<int> atomList = getAtomsInGroupRow(cg1);
318 –
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]) {
390 >        if (atypeCutoff[atid] > groupCutoff[cg1])
391            groupCutoff[cg1] = atypeCutoff[atid];
325 –        }
392        }
393 <
393 >      
394        bool gTypeFound = false;
395        for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
396          if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
# Line 332 | Line 398 | namespace OpenMD {
398            gTypeFound = true;
399          }
400        }
401 <      if (!gTypeFound) {
401 >      if (!gTypeFound) {      
402          gTypeCutoffs.push_back( groupCutoff[cg1] );
403          groupToGtype[cg1] = gTypeCutoffs.size() - 1;
404        }      
# Line 376 | Line 442 | namespace OpenMD {
442  
443          pair<int,int> key = make_pair(i,j);
444          gTypeCutoffMap[key].first = thisRcut;
379 –
445          if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
381 –
446          gTypeCutoffMap[key].second = thisRcut*thisRcut;
383 –        
447          gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
385 –
448          // sanity check
449          
450          if (userChoseCutoff_) {
# Line 399 | Line 461 | namespace OpenMD {
461      }
462    }
463  
402 –
464    groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
465      int i, j;  
466   #ifdef IS_MPI
# Line 474 | Line 535 | namespace OpenMD {
535             atomColData.skippedCharge.end(), 0.0);
536      }
537  
538 +    if (storageLayout_ & DataStorage::dslFlucQForce) {      
539 +      fill(atomRowData.flucQFrc.begin(),
540 +           atomRowData.flucQFrc.end(), 0.0);
541 +      fill(atomColData.flucQFrc.begin(),
542 +           atomColData.flucQFrc.end(), 0.0);
543 +    }
544 +
545 +    if (storageLayout_ & DataStorage::dslElectricField) {    
546 +      fill(atomRowData.electricField.begin(),
547 +           atomRowData.electricField.end(), V3Zero);
548 +      fill(atomColData.electricField.begin(),
549 +           atomColData.electricField.end(), V3Zero);
550 +    }
551 +
552 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
553 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
554 +           0.0);
555 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
556 +           0.0);
557 +    }
558 +
559   #endif
560      // even in parallel, we need to zero out the local arrays:
561  
# Line 486 | Line 568 | namespace OpenMD {
568        fill(snap_->atomData.density.begin(),
569             snap_->atomData.density.end(), 0.0);
570      }
571 +
572      if (storageLayout_ & DataStorage::dslFunctional) {
573        fill(snap_->atomData.functional.begin(),
574             snap_->atomData.functional.end(), 0.0);
575      }
576 +
577      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
578        fill(snap_->atomData.functionalDerivative.begin(),
579             snap_->atomData.functionalDerivative.end(), 0.0);
580      }
581 +
582      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
583        fill(snap_->atomData.skippedCharge.begin(),
584             snap_->atomData.skippedCharge.end(), 0.0);
585      }
586 <    
586 >
587 >    if (storageLayout_ & DataStorage::dslElectricField) {      
588 >      fill(snap_->atomData.electricField.begin(),
589 >           snap_->atomData.electricField.end(), V3Zero);
590 >    }
591    }
592  
593  
# Line 508 | Line 597 | namespace OpenMD {
597   #ifdef IS_MPI
598      
599      // gather up the atomic positions
600 <    AtomCommVectorRow->gather(snap_->atomData.position,
600 >    AtomPlanVectorRow->gather(snap_->atomData.position,
601                                atomRowData.position);
602 <    AtomCommVectorColumn->gather(snap_->atomData.position,
602 >    AtomPlanVectorColumn->gather(snap_->atomData.position,
603                                   atomColData.position);
604      
605      // gather up the cutoff group positions
606 <    cgCommVectorRow->gather(snap_->cgData.position,
606 >
607 >    cgPlanVectorRow->gather(snap_->cgData.position,
608                              cgRowData.position);
609 <    cgCommVectorColumn->gather(snap_->cgData.position,
609 >
610 >    cgPlanVectorColumn->gather(snap_->cgData.position,
611                                 cgColData.position);
612 +
613 +
614 +
615 +    if (needVelocities_) {
616 +      // gather up the atomic velocities
617 +      AtomPlanVectorColumn->gather(snap_->atomData.velocity,
618 +                                   atomColData.velocity);
619 +      
620 +      cgPlanVectorColumn->gather(snap_->cgData.velocity,
621 +                                 cgColData.velocity);
622 +    }
623 +
624      
625      // if needed, gather the atomic rotation matrices
626      if (storageLayout_ & DataStorage::dslAmat) {
627 <      AtomCommMatrixRow->gather(snap_->atomData.aMat,
627 >      AtomPlanMatrixRow->gather(snap_->atomData.aMat,
628                                  atomRowData.aMat);
629 <      AtomCommMatrixColumn->gather(snap_->atomData.aMat,
629 >      AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
630                                     atomColData.aMat);
631      }
632      
633      // if needed, gather the atomic eletrostatic frames
634      if (storageLayout_ & DataStorage::dslElectroFrame) {
635 <      AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
635 >      AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
636                                  atomRowData.electroFrame);
637 <      AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
637 >      AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
638                                     atomColData.electroFrame);
639      }
640  
641 +    // if needed, gather the atomic fluctuating charge values
642 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
643 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
644 +                              atomRowData.flucQPos);
645 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
646 +                                 atomColData.flucQPos);
647 +    }
648 +
649   #endif      
650    }
651    
# Line 548 | Line 659 | namespace OpenMD {
659      
660      if (storageLayout_ & DataStorage::dslDensity) {
661        
662 <      AtomCommRealRow->scatter(atomRowData.density,
662 >      AtomPlanRealRow->scatter(atomRowData.density,
663                                 snap_->atomData.density);
664        
665        int n = snap_->atomData.density.size();
666        vector<RealType> rho_tmp(n, 0.0);
667 <      AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
667 >      AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
668        for (int i = 0; i < n; i++)
669          snap_->atomData.density[i] += rho_tmp[i];
670      }
671 +
672 +    if (storageLayout_ & DataStorage::dslElectricField) {
673 +      
674 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
675 +                                 snap_->atomData.electricField);
676 +      
677 +      int n = snap_->atomData.electricField.size();
678 +      vector<Vector3d> field_tmp(n, V3Zero);
679 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
680 +      for (int i = 0; i < n; i++)
681 +        snap_->atomData.electricField[i] += field_tmp[i];
682 +    }
683   #endif
684    }
685  
# Line 569 | Line 692 | namespace OpenMD {
692      storageLayout_ = sman_->getStorageLayout();
693   #ifdef IS_MPI
694      if (storageLayout_ & DataStorage::dslFunctional) {
695 <      AtomCommRealRow->gather(snap_->atomData.functional,
695 >      AtomPlanRealRow->gather(snap_->atomData.functional,
696                                atomRowData.functional);
697 <      AtomCommRealColumn->gather(snap_->atomData.functional,
697 >      AtomPlanRealColumn->gather(snap_->atomData.functional,
698                                   atomColData.functional);
699      }
700      
701      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
702 <      AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
702 >      AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
703                                atomRowData.functionalDerivative);
704 <      AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
704 >      AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
705                                   atomColData.functionalDerivative);
706      }
707   #endif
# Line 592 | Line 715 | namespace OpenMD {
715      int n = snap_->atomData.force.size();
716      vector<Vector3d> frc_tmp(n, V3Zero);
717      
718 <    AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
718 >    AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
719      for (int i = 0; i < n; i++) {
720        snap_->atomData.force[i] += frc_tmp[i];
721        frc_tmp[i] = 0.0;
722      }
723      
724 <    AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
725 <    for (int i = 0; i < n; i++)
724 >    AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
725 >    for (int i = 0; i < n; i++) {
726        snap_->atomData.force[i] += frc_tmp[i];
727 +    }
728          
729      if (storageLayout_ & DataStorage::dslTorque) {
730  
731        int nt = snap_->atomData.torque.size();
732        vector<Vector3d> trq_tmp(nt, V3Zero);
733  
734 <      AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
734 >      AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
735        for (int i = 0; i < nt; i++) {
736          snap_->atomData.torque[i] += trq_tmp[i];
737          trq_tmp[i] = 0.0;
738        }
739        
740 <      AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
740 >      AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
741        for (int i = 0; i < nt; i++)
742          snap_->atomData.torque[i] += trq_tmp[i];
743      }
# Line 623 | Line 747 | namespace OpenMD {
747        int ns = snap_->atomData.skippedCharge.size();
748        vector<RealType> skch_tmp(ns, 0.0);
749  
750 <      AtomCommRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
750 >      AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
751        for (int i = 0; i < ns; i++) {
752          snap_->atomData.skippedCharge[i] += skch_tmp[i];
753          skch_tmp[i] = 0.0;
754        }
755        
756 <      AtomCommRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
757 <      for (int i = 0; i < ns; i++)
756 >      AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
757 >      for (int i = 0; i < ns; i++)
758          snap_->atomData.skippedCharge[i] += skch_tmp[i];
759 +            
760      }
761      
762 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
763 +
764 +      int nq = snap_->atomData.flucQFrc.size();
765 +      vector<RealType> fqfrc_tmp(nq, 0.0);
766 +
767 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
768 +      for (int i = 0; i < nq; i++) {
769 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
770 +        fqfrc_tmp[i] = 0.0;
771 +      }
772 +      
773 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
774 +      for (int i = 0; i < nq; i++)
775 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
776 +            
777 +    }
778 +
779      nLocal_ = snap_->getNumberOfAtoms();
780  
781      vector<potVec> pot_temp(nLocal_,
# Line 641 | Line 783 | namespace OpenMD {
783  
784      // scatter/gather pot_row into the members of my column
785            
786 <    AtomCommPotRow->scatter(pot_row, pot_temp);
786 >    AtomPlanPotRow->scatter(pot_row, pot_temp);
787  
788      for (int ii = 0;  ii < pot_temp.size(); ii++ )
789        pairwisePot += pot_temp[ii];
790 <    
790 >        
791 >    if (storageLayout_ & DataStorage::dslParticlePot) {
792 >      // This is the pairwise contribution to the particle pot.  The
793 >      // embedding contribution is added in each of the low level
794 >      // non-bonded routines.  In single processor, this is done in
795 >      // unpackInteractionData, not in collectData.
796 >      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
797 >        for (int i = 0; i < nLocal_; i++) {
798 >          // factor of two is because the total potential terms are divided
799 >          // by 2 in parallel due to row/ column scatter      
800 >          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
801 >        }
802 >      }
803 >    }
804 >
805      fill(pot_temp.begin(), pot_temp.end(),
806           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
807        
808 <    AtomCommPotColumn->scatter(pot_col, pot_temp);    
808 >    AtomPlanPotColumn->scatter(pot_col, pot_temp);    
809      
810      for (int ii = 0;  ii < pot_temp.size(); ii++ )
811        pairwisePot += pot_temp[ii];    
812 +
813 +    if (storageLayout_ & DataStorage::dslParticlePot) {
814 +      // This is the pairwise contribution to the particle pot.  The
815 +      // embedding contribution is added in each of the low level
816 +      // non-bonded routines.  In single processor, this is done in
817 +      // unpackInteractionData, not in collectData.
818 +      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
819 +        for (int i = 0; i < nLocal_; i++) {
820 +          // factor of two is because the total potential terms are divided
821 +          // by 2 in parallel due to row/ column scatter      
822 +          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
823 +        }
824 +      }
825 +    }
826 +    
827 +    if (storageLayout_ & DataStorage::dslParticlePot) {
828 +      int npp = snap_->atomData.particlePot.size();
829 +      vector<RealType> ppot_temp(npp, 0.0);
830 +
831 +      // This is the direct or embedding contribution to the particle
832 +      // pot.
833 +      
834 +      AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
835 +      for (int i = 0; i < npp; i++) {
836 +        snap_->atomData.particlePot[i] += ppot_temp[i];
837 +      }
838 +
839 +      fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
840 +      
841 +      AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
842 +      for (int i = 0; i < npp; i++) {
843 +        snap_->atomData.particlePot[i] += ppot_temp[i];
844 +      }
845 +    }
846 +
847 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
848 +      RealType ploc1 = pairwisePot[ii];
849 +      RealType ploc2 = 0.0;
850 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
851 +      pairwisePot[ii] = ploc2;
852 +    }
853 +
854 +    // Here be dragons.
855 +    MPI::Intracomm col = colComm.getComm();
856 +
857 +    col.Allreduce(MPI::IN_PLACE,
858 +                  &snap_->frameData.conductiveHeatFlux[0], 3,
859 +                  MPI::REALTYPE, MPI::SUM);
860 +
861 +
862   #endif
863  
864    }
865  
866 +  /**
867 +   * Collects information obtained during the post-pair (and embedding
868 +   * functional) loops onto local data structures.
869 +   */
870 +  void ForceMatrixDecomposition::collectSelfData() {
871 +    snap_ = sman_->getCurrentSnapshot();
872 +    storageLayout_ = sman_->getStorageLayout();
873 +
874 + #ifdef IS_MPI
875 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
876 +      RealType ploc1 = embeddingPot[ii];
877 +      RealType ploc2 = 0.0;
878 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
879 +      embeddingPot[ii] = ploc2;
880 +    }    
881 + #endif
882 +    
883 +  }
884 +
885 +
886 +
887    int ForceMatrixDecomposition::getNAtomsInRow() {  
888   #ifdef IS_MPI
889      return nAtomsInRow_;
# Line 697 | Line 924 | namespace OpenMD {
924      return d;    
925    }
926  
927 +  Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
928 + #ifdef IS_MPI
929 +    return cgColData.velocity[cg2];
930 + #else
931 +    return snap_->cgData.velocity[cg2];
932 + #endif
933 +  }
934  
935 +  Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
936 + #ifdef IS_MPI
937 +    return atomColData.velocity[atom2];
938 + #else
939 +    return snap_->atomData.velocity[atom2];
940 + #endif
941 +  }
942 +
943 +
944    Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
945  
946      Vector3d d;
# Line 763 | Line 1006 | namespace OpenMD {
1006     * We need to exclude some overcounted interactions that result from
1007     * the parallel decomposition.
1008     */
1009 <  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
1010 <    int unique_id_1, unique_id_2;
1011 <
1009 >  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) {
1010 >    int unique_id_1, unique_id_2, group1, group2;
1011 >        
1012   #ifdef IS_MPI
1013      // in MPI, we have to look up the unique IDs for each atom
1014      unique_id_1 = AtomRowToGlobal[atom1];
1015      unique_id_2 = AtomColToGlobal[atom2];
1016 +    group1 = cgRowToGlobal[cg1];
1017 +    group2 = cgColToGlobal[cg2];
1018 + #else
1019 +    unique_id_1 = AtomLocalToGlobal[atom1];
1020 +    unique_id_2 = AtomLocalToGlobal[atom2];
1021 +    group1 = cgLocalToGlobal[cg1];
1022 +    group2 = cgLocalToGlobal[cg2];
1023 + #endif  
1024  
774 –    // this situation should only arise in MPI simulations
1025      if (unique_id_1 == unique_id_2) return true;
1026 <    
1026 >
1027 > #ifdef IS_MPI
1028      // this prevents us from doing the pair on multiple processors
1029      if (unique_id_1 < unique_id_2) {
1030        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1031      } else {
1032 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1032 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1033      }
1034 + #endif    
1035 +
1036 + #ifndef IS_MPI
1037 +    if (group1 == group2) {
1038 +      if (unique_id_1 < unique_id_2) return true;
1039 +    }
1040   #endif
1041 +    
1042      return false;
1043    }
1044  
# Line 794 | Line 1052 | namespace OpenMD {
1052     * field) must still be handled for these pairs.
1053     */
1054    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1055 <    int unique_id_2;
1055 >
1056 >    // excludesForAtom was constructed to use row/column indices in the MPI
1057 >    // version, and to use local IDs in the non-MPI version:
1058      
799 – #ifdef IS_MPI
800 –    // in MPI, we have to look up the unique IDs for the row atom.
801 –    unique_id_2 = AtomColToGlobal[atom2];
802 – #else
803 –    // in the normal loop, the atom numbers are unique
804 –    unique_id_2 = atom2;
805 – #endif
806 –    
1059      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1060           i != excludesForAtom[atom1].end(); ++i) {
1061 <      if ( (*i) == unique_id_2 ) return true;
1061 >      if ( (*i) == atom2 ) return true;
1062      }
1063  
1064      return false;
# Line 836 | Line 1088 | namespace OpenMD {
1088      idat.excluded = excludeAtomPair(atom1, atom2);
1089    
1090   #ifdef IS_MPI
1091 +    idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
1092 +    //idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
1093 +    //                         ff_->getAtomType(identsCol[atom2]) );
1094      
840 –    idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
841 –                             ff_->getAtomType(identsCol[atom2]) );
842 –    
1095      if (storageLayout_ & DataStorage::dslAmat) {
1096        idat.A1 = &(atomRowData.aMat[atom1]);
1097        idat.A2 = &(atomColData.aMat[atom2]);
# Line 880 | Line 1132 | namespace OpenMD {
1132        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1133      }
1134  
1135 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1136 +      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1137 +      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1138 +    }
1139 +
1140   #else
1141 +    
1142 +    idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1143  
885 –    idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
886 –                             ff_->getAtomType(idents[atom2]) );
887 –
1144      if (storageLayout_ & DataStorage::dslAmat) {
1145        idat.A1 = &(snap_->atomData.aMat[atom1]);
1146        idat.A2 = &(snap_->atomData.aMat[atom2]);
# Line 924 | Line 1180 | namespace OpenMD {
1180        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1181        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1182      }
1183 +
1184 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1185 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1186 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1187 +    }
1188 +
1189   #endif
1190    }
1191  
1192    
1193    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1194   #ifdef IS_MPI
1195 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1196 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1195 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1196 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1197  
1198      atomRowData.force[atom1] += *(idat.f1);
1199      atomColData.force[atom2] -= *(idat.f1);
1200 +
1201 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1202 +      atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1203 +      atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1204 +    }
1205 +
1206 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1207 +      atomRowData.electricField[atom1] += *(idat.eField1);
1208 +      atomColData.electricField[atom2] += *(idat.eField2);
1209 +    }
1210 +
1211   #else
1212      pairwisePot += *(idat.pot);
1213  
1214      snap_->atomData.force[atom1] += *(idat.f1);
1215      snap_->atomData.force[atom2] -= *(idat.f1);
1216 +
1217 +    if (idat.doParticlePot) {
1218 +      // This is the pairwise contribution to the particle pot.  The
1219 +      // embedding contribution is added in each of the low level
1220 +      // non-bonded routines.  In parallel, this calculation is done
1221 +      // in collectData, not in unpackInteractionData.
1222 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1223 +      snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1224 +    }
1225 +    
1226 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1227 +      snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1228 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1229 +    }
1230 +
1231 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1232 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1233 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1234 +    }
1235 +
1236   #endif
1237      
1238    }
# Line 1021 | Line 1314 | namespace OpenMD {
1314          // add this cutoff group to the list of groups in this cell;
1315          cellListRow_[cellIndex].push_back(i);
1316        }
1024 –      
1317        for (int i = 0; i < nGroupsInCol_; i++) {
1318          rs = cgColData.position[i];
1319          
# Line 1046 | Line 1338 | namespace OpenMD {
1338          // add this cutoff group to the list of groups in this cell;
1339          cellListCol_[cellIndex].push_back(i);
1340        }
1341 +    
1342   #else
1343        for (int i = 0; i < nGroups_; i++) {
1344          rs = snap_->cgData.position[i];
# Line 1066 | Line 1359 | namespace OpenMD {
1359          whichCell.z() = nCells_.z() * scaled.z();
1360          
1361          // find single index of this cell:
1362 <        cellIndex = Vlinear(whichCell, nCells_);      
1362 >        cellIndex = Vlinear(whichCell, nCells_);
1363          
1364          // add this cutoff group to the list of groups in this cell;
1365          cellList_[cellIndex].push_back(i);
1366        }
1367 +
1368   #endif
1369  
1370        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1083 | Line 1377 | namespace OpenMD {
1377                   os != cellOffsets_.end(); ++os) {
1378                
1379                Vector3i m2v = m1v + (*os);
1380 <              
1380 >            
1381 >
1382                if (m2v.x() >= nCells_.x()) {
1383                  m2v.x() = 0;          
1384                } else if (m2v.x() < 0) {
# Line 1101 | Line 1396 | namespace OpenMD {
1396                } else if (m2v.z() < 0) {
1397                  m2v.z() = nCells_.z() - 1;
1398                }
1399 <              
1399 >
1400                int m2 = Vlinear (m2v, nCells_);
1401                
1402   #ifdef IS_MPI
# Line 1110 | Line 1405 | namespace OpenMD {
1405                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1406                       j2 != cellListCol_[m2].end(); ++j2) {
1407                    
1408 <                  // Always do this if we're in different cells or if
1409 <                  // we're in the same cell and the global index of the
1410 <                  // j2 cutoff group is less than the j1 cutoff group
1411 <                  
1412 <                  if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) {
1413 <                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1414 <                    snap_->wrapVector(dr);
1415 <                    cuts = getGroupCutoffs( (*j1), (*j2) );
1416 <                    if (dr.lengthSquare() < cuts.third) {
1122 <                      neighborList.push_back(make_pair((*j1), (*j2)));
1123 <                    }
1124 <                  }
1408 >                  // In parallel, we need to visit *all* pairs of row
1409 >                  // & column indicies and will divide labor in the
1410 >                  // force evaluation later.
1411 >                  dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1412 >                  snap_->wrapVector(dr);
1413 >                  cuts = getGroupCutoffs( (*j1), (*j2) );
1414 >                  if (dr.lengthSquare() < cuts.third) {
1415 >                    neighborList.push_back(make_pair((*j1), (*j2)));
1416 >                  }                  
1417                  }
1418                }
1419   #else
1128 –              
1420                for (vector<int>::iterator j1 = cellList_[m1].begin();
1421                     j1 != cellList_[m1].end(); ++j1) {
1422                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1423                       j2 != cellList_[m2].end(); ++j2) {
1424 <                  
1424 >    
1425                    // Always do this if we're in different cells or if
1426 <                  // we're in the same cell and the global index of the
1427 <                  // j2 cutoff group is less than the j1 cutoff group
1428 <                  
1429 <                  if (m2 != m1 || (*j2) < (*j1)) {
1426 >                  // we're in the same cell and the global index of
1427 >                  // the j2 cutoff group is greater than or equal to
1428 >                  // the j1 cutoff group.  Note that Rappaport's code
1429 >                  // has a "less than" conditional here, but that
1430 >                  // deals with atom-by-atom computation.  OpenMD
1431 >                  // allows atoms within a single cutoff group to
1432 >                  // interact with each other.
1433 >
1434 >
1435 >
1436 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1437 >
1438                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1439                      snap_->wrapVector(dr);
1440                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1154 | Line 1453 | namespace OpenMD {
1453        // branch to do all cutoff group pairs
1454   #ifdef IS_MPI
1455        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1456 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1456 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1457            dr = cgColData.position[j2] - cgRowData.position[j1];
1458            snap_->wrapVector(dr);
1459            cuts = getGroupCutoffs( j1, j2 );
# Line 1162 | Line 1461 | namespace OpenMD {
1461              neighborList.push_back(make_pair(j1, j2));
1462            }
1463          }
1464 <      }
1464 >      }      
1465   #else
1466 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1467 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1466 >      // include all groups here.
1467 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1468 >        // include self group interactions j2 == j1
1469 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1470            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1471            snap_->wrapVector(dr);
1472            cuts = getGroupCutoffs( j1, j2 );
1473            if (dr.lengthSquare() < cuts.third) {
1474              neighborList.push_back(make_pair(j1, j2));
1475            }
1476 <        }
1477 <      }        
1476 >        }    
1477 >      }
1478   #endif
1479      }
1480        

Diff Legend

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