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 1588 by gezelter, Sat Jul 9 15:05:59 2011 UTC vs.
Revision 1721 by gezelter, Thu May 24 14:17:42 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();
# Line 74 | Line 112 | namespace OpenMD {
112  
113   #ifdef IS_MPI
114  
115 <    AtomCommIntRow = new Communicator<Row,int>(nLocal_);
116 <    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_);
115 >    MPI::Intracomm row = rowComm.getComm();
116 >    MPI::Intracomm col = colComm.getComm();
117  
118 <    AtomCommIntColumn = new Communicator<Column,int>(nLocal_);
119 <    AtomCommRealColumn = new Communicator<Column,RealType>(nLocal_);
120 <    AtomCommVectorColumn = new Communicator<Column,Vector3d>(nLocal_);
121 <    AtomCommMatrixColumn = new Communicator<Column,Mat3x3d>(nLocal_);
122 <    AtomCommPotColumn = new Communicator<Column,potVec>(nLocal_);
118 >    AtomPlanIntRow = new Plan<int>(row, nLocal_);
119 >    AtomPlanRealRow = new Plan<RealType>(row, nLocal_);
120 >    AtomPlanVectorRow = new Plan<Vector3d>(row, nLocal_);
121 >    AtomPlanMatrixRow = new Plan<Mat3x3d>(row, nLocal_);
122 >    AtomPlanPotRow = new Plan<potVec>(row, nLocal_);
123  
124 <    cgCommIntRow = new Communicator<Row,int>(nGroups_);
125 <    cgCommVectorRow = new Communicator<Row,Vector3d>(nGroups_);
126 <    cgCommIntColumn = new Communicator<Column,int>(nGroups_);
127 <    cgCommVectorColumn = new Communicator<Column,Vector3d>(nGroups_);
124 >    AtomPlanIntColumn = new Plan<int>(col, nLocal_);
125 >    AtomPlanRealColumn = new Plan<RealType>(col, nLocal_);
126 >    AtomPlanVectorColumn = new Plan<Vector3d>(col, nLocal_);
127 >    AtomPlanMatrixColumn = new Plan<Mat3x3d>(col, nLocal_);
128 >    AtomPlanPotColumn = new Plan<potVec>(col, nLocal_);
129  
130 <    nAtomsInRow_ = AtomCommIntRow->getSize();
131 <    nAtomsInCol_ = AtomCommIntColumn->getSize();
132 <    nGroupsInRow_ = cgCommIntRow->getSize();
133 <    nGroupsInCol_ = cgCommIntColumn->getSize();
130 >    cgPlanIntRow = new Plan<int>(row, nGroups_);
131 >    cgPlanVectorRow = new Plan<Vector3d>(row, nGroups_);
132 >    cgPlanIntColumn = new Plan<int>(col, nGroups_);
133 >    cgPlanVectorColumn = new Plan<Vector3d>(col, nGroups_);
134  
135 +    nAtomsInRow_ = AtomPlanIntRow->getSize();
136 +    nAtomsInCol_ = AtomPlanIntColumn->getSize();
137 +    nGroupsInRow_ = cgPlanIntRow->getSize();
138 +    nGroupsInCol_ = cgPlanIntColumn->getSize();
139 +
140      // Modify the data storage objects with the correct layouts and sizes:
141      atomRowData.resize(nAtomsInRow_);
142      atomRowData.setStorageLayout(storageLayout_);
# Line 109 | Line 150 | namespace OpenMD {
150      identsRow.resize(nAtomsInRow_);
151      identsCol.resize(nAtomsInCol_);
152      
153 <    AtomCommIntRow->gather(idents, identsRow);
154 <    AtomCommIntColumn->gather(idents, identsCol);
153 >    AtomPlanIntRow->gather(idents, identsRow);
154 >    AtomPlanIntColumn->gather(idents, identsCol);
155      
156 <    vector<int>::iterator it;
157 <    for (it = AtomLocalToGlobal.begin(); it != AtomLocalToGlobal.end(); ++it) {
158 <      cerr << "my AtomLocalToGlobal = " << (*it) << "\n";
118 <    }
119 <    AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
120 <    AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
121 <    
122 <    cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
123 <    cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
156 >    // allocate memory for the parallel objects
157 >    atypesRow.resize(nAtomsInRow_);
158 >    atypesCol.resize(nAtomsInCol_);
159  
160 <    AtomCommRealRow->gather(massFactors, massFactorsRow);
161 <    AtomCommRealColumn->gather(massFactors, massFactorsCol);
160 >    for (int i = 0; i < nAtomsInRow_; i++)
161 >      atypesRow[i] = ff_->getAtomType(identsRow[i]);
162 >    for (int i = 0; i < nAtomsInCol_; i++)
163 >      atypesCol[i] = ff_->getAtomType(identsCol[i]);        
164  
165 +    pot_row.resize(nAtomsInRow_);
166 +    pot_col.resize(nAtomsInCol_);
167 +
168 +    AtomRowToGlobal.resize(nAtomsInRow_);
169 +    AtomColToGlobal.resize(nAtomsInCol_);
170 +    AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
171 +    AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
172 +
173 +    cgRowToGlobal.resize(nGroupsInRow_);
174 +    cgColToGlobal.resize(nGroupsInCol_);
175 +    cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
176 +    cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
177 +
178 +    massFactorsRow.resize(nAtomsInRow_);
179 +    massFactorsCol.resize(nAtomsInCol_);
180 +    AtomPlanRealRow->gather(massFactors, massFactorsRow);
181 +    AtomPlanRealColumn->gather(massFactors, massFactorsCol);
182 +
183      groupListRow_.clear();
184      groupListRow_.resize(nGroupsInRow_);
185      for (int i = 0; i < nGroupsInRow_; i++) {
# Line 179 | Line 234 | namespace OpenMD {
234        }      
235      }
236  
237 < #endif
183 <
184 <    groupList_.clear();
185 <    groupList_.resize(nGroups_);
186 <    for (int i = 0; i < nGroups_; i++) {
187 <      int gid = cgLocalToGlobal[i];
188 <      for (int j = 0; j < nLocal_; j++) {
189 <        int aid = AtomLocalToGlobal[j];
190 <        if (globalGroupMembership[aid] == gid) {
191 <          groupList_[i].push_back(j);
192 <        }
193 <      }      
194 <    }
195 <
237 > #else
238      excludesForAtom.clear();
239      excludesForAtom.resize(nLocal_);
240      toposForAtom.clear();
# Line 225 | Line 267 | namespace OpenMD {
267          }
268        }      
269      }
270 <    
270 > #endif
271 >
272 >    // allocate memory for the parallel objects
273 >    atypesLocal.resize(nLocal_);
274 >
275 >    for (int i = 0; i < nLocal_; i++)
276 >      atypesLocal[i] = ff_->getAtomType(idents[i]);
277 >
278 >    groupList_.clear();
279 >    groupList_.resize(nGroups_);
280 >    for (int i = 0; i < nGroups_; i++) {
281 >      int gid = cgLocalToGlobal[i];
282 >      for (int j = 0; j < nLocal_; j++) {
283 >        int aid = AtomLocalToGlobal[j];
284 >        if (globalGroupMembership[aid] == gid) {
285 >          groupList_[i].push_back(j);
286 >        }
287 >      }      
288 >    }
289 >
290 >
291      createGtypeCutoffMap();
292  
293    }
# Line 233 | Line 295 | namespace OpenMD {
295    void ForceMatrixDecomposition::createGtypeCutoffMap() {
296      
297      RealType tol = 1e-6;
298 +    largestRcut_ = 0.0;
299      RealType rc;
300      int atid;
301      set<AtomType*> atypes = info_->getSimulatedAtomTypes();
302 +    
303      map<int, RealType> atypeCutoff;
304        
305      for (set<AtomType*>::iterator at = atypes.begin();
# Line 243 | Line 307 | namespace OpenMD {
307        atid = (*at)->getIdent();
308        if (userChoseCutoff_)
309          atypeCutoff[atid] = userCutoff_;
310 <      else
310 >      else
311          atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at);
312      }
313 <
313 >    
314      vector<RealType> gTypeCutoffs;
315      // first we do a single loop over the cutoff groups to find the
316      // largest cutoff for any atypes present in this group.
# Line 306 | Line 370 | namespace OpenMD {
370      vector<RealType> groupCutoff(nGroups_, 0.0);
371      groupToGtype.resize(nGroups_);
372      for (int cg1 = 0; cg1 < nGroups_; cg1++) {
309 –
373        groupCutoff[cg1] = 0.0;
374        vector<int> atomList = getAtomsInGroupRow(cg1);
312 –
375        for (vector<int>::iterator ia = atomList.begin();
376             ia != atomList.end(); ++ia) {            
377          int atom1 = (*ia);
378          atid = idents[atom1];
379 <        if (atypeCutoff[atid] > groupCutoff[cg1]) {
379 >        if (atypeCutoff[atid] > groupCutoff[cg1])
380            groupCutoff[cg1] = atypeCutoff[atid];
319 –        }
381        }
382 <
382 >      
383        bool gTypeFound = false;
384        for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
385          if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
# Line 326 | Line 387 | namespace OpenMD {
387            gTypeFound = true;
388          }
389        }
390 <      if (!gTypeFound) {
390 >      if (!gTypeFound) {      
391          gTypeCutoffs.push_back( groupCutoff[cg1] );
392          groupToGtype[cg1] = gTypeCutoffs.size() - 1;
393        }      
# Line 335 | Line 396 | namespace OpenMD {
396  
397      // Now we find the maximum group cutoff value present in the simulation
398  
399 <    RealType groupMax = *max_element(gTypeCutoffs.begin(), gTypeCutoffs.end());
399 >    RealType groupMax = *max_element(gTypeCutoffs.begin(),
400 >                                     gTypeCutoffs.end());
401  
402   #ifdef IS_MPI
403 <    MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE, MPI::MAX);
403 >    MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE,
404 >                              MPI::MAX);
405   #endif
406      
407      RealType tradRcut = groupMax;
# Line 368 | Line 431 | namespace OpenMD {
431  
432          pair<int,int> key = make_pair(i,j);
433          gTypeCutoffMap[key].first = thisRcut;
371 –
434          if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
373 –
435          gTypeCutoffMap[key].second = thisRcut*thisRcut;
375 –        
436          gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
377 –
437          // sanity check
438          
439          if (userChoseCutoff_) {
# Line 434 | Line 493 | namespace OpenMD {
493           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));  
494  
495      if (storageLayout_ & DataStorage::dslParticlePot) {    
496 <      fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(), 0.0);
497 <      fill(atomColData.particlePot.begin(), atomColData.particlePot.end(), 0.0);
496 >      fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(),
497 >           0.0);
498 >      fill(atomColData.particlePot.begin(), atomColData.particlePot.end(),
499 >           0.0);
500      }
501  
502      if (storageLayout_ & DataStorage::dslDensity) {      
# Line 444 | Line 505 | namespace OpenMD {
505      }
506  
507      if (storageLayout_ & DataStorage::dslFunctional) {  
508 <      fill(atomRowData.functional.begin(), atomRowData.functional.end(), 0.0);
509 <      fill(atomColData.functional.begin(), atomColData.functional.end(), 0.0);
508 >      fill(atomRowData.functional.begin(), atomRowData.functional.end(),
509 >           0.0);
510 >      fill(atomColData.functional.begin(), atomColData.functional.end(),
511 >           0.0);
512      }
513  
514      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
# Line 462 | Line 525 | namespace OpenMD {
525             atomColData.skippedCharge.end(), 0.0);
526      }
527  
528 < #else
529 <    
528 >    if (storageLayout_ & DataStorage::dslFlucQForce) {      
529 >      fill(atomRowData.flucQFrc.begin(),
530 >           atomRowData.flucQFrc.end(), 0.0);
531 >      fill(atomColData.flucQFrc.begin(),
532 >           atomColData.flucQFrc.end(), 0.0);
533 >    }
534 >
535 >    if (storageLayout_ & DataStorage::dslElectricField) {    
536 >      fill(atomRowData.electricField.begin(),
537 >           atomRowData.electricField.end(), V3Zero);
538 >      fill(atomColData.electricField.begin(),
539 >           atomColData.electricField.end(), V3Zero);
540 >    }
541 >
542 >    if (storageLayout_ & DataStorage::dslFlucQForce) {    
543 >      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
544 >           0.0);
545 >      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
546 >           0.0);
547 >    }
548 >
549 > #endif
550 >    // even in parallel, we need to zero out the local arrays:
551 >
552      if (storageLayout_ & DataStorage::dslParticlePot) {      
553        fill(snap_->atomData.particlePot.begin(),
554             snap_->atomData.particlePot.end(), 0.0);
# Line 473 | Line 558 | namespace OpenMD {
558        fill(snap_->atomData.density.begin(),
559             snap_->atomData.density.end(), 0.0);
560      }
561 +
562      if (storageLayout_ & DataStorage::dslFunctional) {
563        fill(snap_->atomData.functional.begin(),
564             snap_->atomData.functional.end(), 0.0);
565      }
566 +
567      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
568        fill(snap_->atomData.functionalDerivative.begin(),
569             snap_->atomData.functionalDerivative.end(), 0.0);
570      }
571 +
572      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
573        fill(snap_->atomData.skippedCharge.begin(),
574             snap_->atomData.skippedCharge.end(), 0.0);
575      }
576 < #endif
577 <    
576 >
577 >    if (storageLayout_ & DataStorage::dslElectricField) {      
578 >      fill(snap_->atomData.electricField.begin(),
579 >           snap_->atomData.electricField.end(), V3Zero);
580 >    }
581    }
582  
583  
# Line 496 | Line 587 | namespace OpenMD {
587   #ifdef IS_MPI
588      
589      // gather up the atomic positions
590 <    AtomCommVectorRow->gather(snap_->atomData.position,
590 >    AtomPlanVectorRow->gather(snap_->atomData.position,
591                                atomRowData.position);
592 <    AtomCommVectorColumn->gather(snap_->atomData.position,
592 >    AtomPlanVectorColumn->gather(snap_->atomData.position,
593                                   atomColData.position);
594      
595      // gather up the cutoff group positions
596 <    cgCommVectorRow->gather(snap_->cgData.position,
596 >
597 >    cgPlanVectorRow->gather(snap_->cgData.position,
598                              cgRowData.position);
599 <    cgCommVectorColumn->gather(snap_->cgData.position,
599 >
600 >    cgPlanVectorColumn->gather(snap_->cgData.position,
601                                 cgColData.position);
602 +
603      
604      // if needed, gather the atomic rotation matrices
605      if (storageLayout_ & DataStorage::dslAmat) {
606 <      AtomCommMatrixRow->gather(snap_->atomData.aMat,
606 >      AtomPlanMatrixRow->gather(snap_->atomData.aMat,
607                                  atomRowData.aMat);
608 <      AtomCommMatrixColumn->gather(snap_->atomData.aMat,
608 >      AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
609                                     atomColData.aMat);
610      }
611      
612      // if needed, gather the atomic eletrostatic frames
613      if (storageLayout_ & DataStorage::dslElectroFrame) {
614 <      AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
614 >      AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
615                                  atomRowData.electroFrame);
616 <      AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
616 >      AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
617                                     atomColData.electroFrame);
618 +    }
619 +
620 +    // if needed, gather the atomic fluctuating charge values
621 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
622 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
623 +                              atomRowData.flucQPos);
624 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
625 +                                 atomColData.flucQPos);
626      }
627 +
628   #endif      
629    }
630    
# Line 535 | Line 638 | namespace OpenMD {
638      
639      if (storageLayout_ & DataStorage::dslDensity) {
640        
641 <      AtomCommRealRow->scatter(atomRowData.density,
641 >      AtomPlanRealRow->scatter(atomRowData.density,
642                                 snap_->atomData.density);
643        
644        int n = snap_->atomData.density.size();
645        vector<RealType> rho_tmp(n, 0.0);
646 <      AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
646 >      AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
647        for (int i = 0; i < n; i++)
648          snap_->atomData.density[i] += rho_tmp[i];
649      }
650 +
651 +    if (storageLayout_ & DataStorage::dslElectricField) {
652 +      
653 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
654 +                                 snap_->atomData.electricField);
655 +      
656 +      int n = snap_->atomData.electricField.size();
657 +      vector<Vector3d> field_tmp(n, V3Zero);
658 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
659 +      for (int i = 0; i < n; i++)
660 +        snap_->atomData.electricField[i] += field_tmp[i];
661 +    }
662   #endif
663    }
664  
# Line 556 | Line 671 | namespace OpenMD {
671      storageLayout_ = sman_->getStorageLayout();
672   #ifdef IS_MPI
673      if (storageLayout_ & DataStorage::dslFunctional) {
674 <      AtomCommRealRow->gather(snap_->atomData.functional,
674 >      AtomPlanRealRow->gather(snap_->atomData.functional,
675                                atomRowData.functional);
676 <      AtomCommRealColumn->gather(snap_->atomData.functional,
676 >      AtomPlanRealColumn->gather(snap_->atomData.functional,
677                                   atomColData.functional);
678      }
679      
680      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
681 <      AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
681 >      AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
682                                atomRowData.functionalDerivative);
683 <      AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
683 >      AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
684                                   atomColData.functionalDerivative);
685      }
686   #endif
# Line 579 | Line 694 | namespace OpenMD {
694      int n = snap_->atomData.force.size();
695      vector<Vector3d> frc_tmp(n, V3Zero);
696      
697 <    AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
697 >    AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
698      for (int i = 0; i < n; i++) {
699        snap_->atomData.force[i] += frc_tmp[i];
700        frc_tmp[i] = 0.0;
701      }
702      
703 <    AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
704 <    for (int i = 0; i < n; i++)
703 >    AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
704 >    for (int i = 0; i < n; i++) {
705        snap_->atomData.force[i] += frc_tmp[i];
706 <    
707 <    
706 >    }
707 >        
708      if (storageLayout_ & DataStorage::dslTorque) {
709  
710        int nt = snap_->atomData.torque.size();
711        vector<Vector3d> trq_tmp(nt, V3Zero);
712  
713 <      AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
713 >      AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
714        for (int i = 0; i < nt; i++) {
715          snap_->atomData.torque[i] += trq_tmp[i];
716          trq_tmp[i] = 0.0;
717        }
718        
719 <      AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
719 >      AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
720        for (int i = 0; i < nt; i++)
721          snap_->atomData.torque[i] += trq_tmp[i];
722      }
# Line 611 | Line 726 | namespace OpenMD {
726        int ns = snap_->atomData.skippedCharge.size();
727        vector<RealType> skch_tmp(ns, 0.0);
728  
729 <      AtomCommRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
729 >      AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
730        for (int i = 0; i < ns; i++) {
731 <        snap_->atomData.skippedCharge[i] = skch_tmp[i];
731 >        snap_->atomData.skippedCharge[i] += skch_tmp[i];
732          skch_tmp[i] = 0.0;
733        }
734        
735 <      AtomCommRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
736 <      for (int i = 0; i < ns; i++)
735 >      AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
736 >      for (int i = 0; i < ns; i++)
737          snap_->atomData.skippedCharge[i] += skch_tmp[i];
738 +            
739      }
740      
741 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
742 +
743 +      int nq = snap_->atomData.flucQFrc.size();
744 +      vector<RealType> fqfrc_tmp(nq, 0.0);
745 +
746 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
747 +      for (int i = 0; i < nq; i++) {
748 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
749 +        fqfrc_tmp[i] = 0.0;
750 +      }
751 +      
752 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
753 +      for (int i = 0; i < nq; i++)
754 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
755 +            
756 +    }
757 +
758      nLocal_ = snap_->getNumberOfAtoms();
759  
760      vector<potVec> pot_temp(nLocal_,
# Line 629 | Line 762 | namespace OpenMD {
762  
763      // scatter/gather pot_row into the members of my column
764            
765 <    AtomCommPotRow->scatter(pot_row, pot_temp);
765 >    AtomPlanPotRow->scatter(pot_row, pot_temp);
766  
767      for (int ii = 0;  ii < pot_temp.size(); ii++ )
768        pairwisePot += pot_temp[ii];
# Line 637 | Line 770 | namespace OpenMD {
770      fill(pot_temp.begin(), pot_temp.end(),
771           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
772        
773 <    AtomCommPotColumn->scatter(pot_col, pot_temp);    
773 >    AtomPlanPotColumn->scatter(pot_col, pot_temp);    
774      
775      for (int ii = 0;  ii < pot_temp.size(); ii++ )
776        pairwisePot += pot_temp[ii];    
777 +    
778 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
779 +      RealType ploc1 = pairwisePot[ii];
780 +      RealType ploc2 = 0.0;
781 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
782 +      pairwisePot[ii] = ploc2;
783 +    }
784 +
785 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
786 +      RealType ploc1 = embeddingPot[ii];
787 +      RealType ploc2 = 0.0;
788 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
789 +      embeddingPot[ii] = ploc2;
790 +    }
791 +
792   #endif
793  
794    }
# Line 753 | Line 901 | namespace OpenMD {
901     */
902    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
903      int unique_id_1, unique_id_2;
904 <
904 >        
905   #ifdef IS_MPI
906      // in MPI, we have to look up the unique IDs for each atom
907      unique_id_1 = AtomRowToGlobal[atom1];
908      unique_id_2 = AtomColToGlobal[atom2];
909 + #else
910 +    unique_id_1 = AtomLocalToGlobal[atom1];
911 +    unique_id_2 = AtomLocalToGlobal[atom2];
912 + #endif  
913  
762 –    // this situation should only arise in MPI simulations
914      if (unique_id_1 == unique_id_2) return true;
915 <    
915 >
916 > #ifdef IS_MPI
917      // this prevents us from doing the pair on multiple processors
918      if (unique_id_1 < unique_id_2) {
919        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
920      } else {
921 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
921 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
922      }
923   #endif
924 +    
925      return false;
926    }
927  
# Line 782 | Line 935 | namespace OpenMD {
935     * field) must still be handled for these pairs.
936     */
937    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
938 <    int unique_id_2;
938 >
939 >    // excludesForAtom was constructed to use row/column indices in the MPI
940 >    // version, and to use local IDs in the non-MPI version:
941      
787 – #ifdef IS_MPI
788 –    // in MPI, we have to look up the unique IDs for the row atom.
789 –    unique_id_2 = AtomColToGlobal[atom2];
790 – #else
791 –    // in the normal loop, the atom numbers are unique
792 –    unique_id_2 = atom2;
793 – #endif
794 –    
942      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
943           i != excludesForAtom[atom1].end(); ++i) {
944 <      if ( (*i) == unique_id_2 ) return true;
944 >      if ( (*i) == atom2 ) return true;
945      }
946  
947      return false;
# Line 824 | Line 971 | namespace OpenMD {
971      idat.excluded = excludeAtomPair(atom1, atom2);
972    
973   #ifdef IS_MPI
974 +    idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
975 +    //idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
976 +    //                         ff_->getAtomType(identsCol[atom2]) );
977      
828 –    idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
829 –                             ff_->getAtomType(identsCol[atom2]) );
830 –    
978      if (storageLayout_ & DataStorage::dslAmat) {
979        idat.A1 = &(atomRowData.aMat[atom1]);
980        idat.A2 = &(atomColData.aMat[atom2]);
# Line 866 | Line 1013 | namespace OpenMD {
1013      if (storageLayout_ & DataStorage::dslSkippedCharge) {              
1014        idat.skippedCharge1 = &(atomRowData.skippedCharge[atom1]);
1015        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1016 +    }
1017 +
1018 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1019 +      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1020 +      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1021      }
1022  
1023   #else
1024 +    
1025  
1026 <    idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
1027 <                             ff_->getAtomType(idents[atom2]) );
1026 >    // cerr << "atoms = " << atom1 << " " << atom2 << "\n";
1027 >    // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n";
1028 >    // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n";
1029  
1030 +    idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1031 +    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
1032 +    //                         ff_->getAtomType(idents[atom2]) );
1033 +
1034      if (storageLayout_ & DataStorage::dslAmat) {
1035        idat.A1 = &(snap_->atomData.aMat[atom1]);
1036        idat.A2 = &(snap_->atomData.aMat[atom2]);
# Line 912 | Line 1070 | namespace OpenMD {
1070        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1071        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1072      }
1073 +
1074 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1075 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1076 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1077 +    }
1078 +
1079   #endif
1080    }
1081  
1082    
1083    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1084   #ifdef IS_MPI
1085 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1086 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1085 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1086 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1087  
1088      atomRowData.force[atom1] += *(idat.f1);
1089      atomColData.force[atom2] -= *(idat.f1);
1090 +
1091 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1092 +      atomRowData.flucQFrc[atom1] += *(idat.dVdFQ1);
1093 +      atomColData.flucQFrc[atom2] += *(idat.dVdFQ2);
1094 +    }
1095 +
1096 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1097 +      atomRowData.electricField[atom1] += *(idat.eField1);
1098 +      atomColData.electricField[atom2] += *(idat.eField2);
1099 +    }
1100 +
1101 +    // should particle pot be done here also?
1102   #else
1103      pairwisePot += *(idat.pot);
1104  
1105      snap_->atomData.force[atom1] += *(idat.f1);
1106      snap_->atomData.force[atom2] -= *(idat.f1);
1107 +
1108 +    if (idat.doParticlePot) {
1109 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1110 +      snap_->atomData.particlePot[atom2] -= *(idat.vpair) * *(idat.sw);
1111 +    }
1112 +    
1113 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1114 +      snap_->atomData.flucQFrc[atom1] += *(idat.dVdFQ1);
1115 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1116 +    }
1117 +
1118 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1119 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1120 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1121 +    }
1122 +
1123   #endif
1124      
1125    }
# Line 1009 | Line 1201 | namespace OpenMD {
1201          // add this cutoff group to the list of groups in this cell;
1202          cellListRow_[cellIndex].push_back(i);
1203        }
1012 –      
1204        for (int i = 0; i < nGroupsInCol_; i++) {
1205          rs = cgColData.position[i];
1206          
# Line 1034 | Line 1225 | namespace OpenMD {
1225          // add this cutoff group to the list of groups in this cell;
1226          cellListCol_[cellIndex].push_back(i);
1227        }
1228 +    
1229   #else
1230        for (int i = 0; i < nGroups_; i++) {
1231          rs = snap_->cgData.position[i];
# Line 1054 | Line 1246 | namespace OpenMD {
1246          whichCell.z() = nCells_.z() * scaled.z();
1247          
1248          // find single index of this cell:
1249 <        cellIndex = Vlinear(whichCell, nCells_);      
1249 >        cellIndex = Vlinear(whichCell, nCells_);
1250          
1251          // add this cutoff group to the list of groups in this cell;
1252          cellList_[cellIndex].push_back(i);
1253        }
1254 +
1255   #endif
1256  
1257        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1071 | Line 1264 | namespace OpenMD {
1264                   os != cellOffsets_.end(); ++os) {
1265                
1266                Vector3i m2v = m1v + (*os);
1267 <              
1267 >            
1268 >
1269                if (m2v.x() >= nCells_.x()) {
1270                  m2v.x() = 0;          
1271                } else if (m2v.x() < 0) {
# Line 1089 | Line 1283 | namespace OpenMD {
1283                } else if (m2v.z() < 0) {
1284                  m2v.z() = nCells_.z() - 1;
1285                }
1286 <              
1286 >
1287                int m2 = Vlinear (m2v, nCells_);
1288                
1289   #ifdef IS_MPI
# Line 1098 | Line 1292 | namespace OpenMD {
1292                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1293                       j2 != cellListCol_[m2].end(); ++j2) {
1294                    
1295 <                  // Always do this if we're in different cells or if
1296 <                  // we're in the same cell and the global index of the
1297 <                  // j2 cutoff group is less than the j1 cutoff group
1298 <                  
1299 <                  if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) {
1300 <                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1301 <                    snap_->wrapVector(dr);
1302 <                    cuts = getGroupCutoffs( (*j1), (*j2) );
1303 <                    if (dr.lengthSquare() < cuts.third) {
1110 <                      neighborList.push_back(make_pair((*j1), (*j2)));
1111 <                    }
1112 <                  }
1295 >                  // In parallel, we need to visit *all* pairs of row
1296 >                  // & column indicies and will divide labor in the
1297 >                  // force evaluation later.
1298 >                  dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1299 >                  snap_->wrapVector(dr);
1300 >                  cuts = getGroupCutoffs( (*j1), (*j2) );
1301 >                  if (dr.lengthSquare() < cuts.third) {
1302 >                    neighborList.push_back(make_pair((*j1), (*j2)));
1303 >                  }                  
1304                  }
1305                }
1306   #else
1116 –              
1307                for (vector<int>::iterator j1 = cellList_[m1].begin();
1308                     j1 != cellList_[m1].end(); ++j1) {
1309                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1310                       j2 != cellList_[m2].end(); ++j2) {
1311 <                  
1311 >    
1312                    // Always do this if we're in different cells or if
1313 <                  // we're in the same cell and the global index of the
1314 <                  // j2 cutoff group is less than the j1 cutoff group
1315 <                  
1316 <                  if (m2 != m1 || (*j2) < (*j1)) {
1313 >                  // we're in the same cell and the global index of
1314 >                  // the j2 cutoff group is greater than or equal to
1315 >                  // the j1 cutoff group.  Note that Rappaport's code
1316 >                  // has a "less than" conditional here, but that
1317 >                  // deals with atom-by-atom computation.  OpenMD
1318 >                  // allows atoms within a single cutoff group to
1319 >                  // interact with each other.
1320 >
1321 >
1322 >
1323 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1324 >
1325                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1326                      snap_->wrapVector(dr);
1327                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1142 | Line 1340 | namespace OpenMD {
1340        // branch to do all cutoff group pairs
1341   #ifdef IS_MPI
1342        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1343 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1343 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1344            dr = cgColData.position[j2] - cgRowData.position[j1];
1345            snap_->wrapVector(dr);
1346            cuts = getGroupCutoffs( j1, j2 );
# Line 1150 | Line 1348 | namespace OpenMD {
1348              neighborList.push_back(make_pair(j1, j2));
1349            }
1350          }
1351 <      }
1351 >      }      
1352   #else
1353 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1354 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1353 >      // include all groups here.
1354 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1355 >        // include self group interactions j2 == j1
1356 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1357            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1358            snap_->wrapVector(dr);
1359            cuts = getGroupCutoffs( j1, j2 );
1360            if (dr.lengthSquare() < cuts.third) {
1361              neighborList.push_back(make_pair(j1, j2));
1362            }
1363 <        }
1364 <      }        
1363 >        }    
1364 >      }
1365   #endif
1366      }
1367        

Diff Legend

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