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 1591 by gezelter, Tue Jul 12 15:25:07 2011 UTC vs.
Revision 1755 by gezelter, Thu Jun 14 01:58:35 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_);
# Line 126 | Line 178 | namespace OpenMD {
178  
179      AtomRowToGlobal.resize(nAtomsInRow_);
180      AtomColToGlobal.resize(nAtomsInCol_);
181 <    AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
182 <    AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
183 <    
181 >    AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
182 >    AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
183 >
184      cgRowToGlobal.resize(nGroupsInRow_);
185      cgColToGlobal.resize(nGroupsInCol_);
186 <    cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
187 <    cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
186 >    cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
187 >    cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
188  
189      massFactorsRow.resize(nAtomsInRow_);
190      massFactorsCol.resize(nAtomsInCol_);
191 <    AtomCommRealRow->gather(massFactors, massFactorsRow);
192 <    AtomCommRealColumn->gather(massFactors, massFactorsCol);
191 >    AtomPlanRealRow->gather(massFactors, massFactorsRow);
192 >    AtomPlanRealColumn->gather(massFactors, massFactorsCol);
193  
194      groupListRow_.clear();
195      groupListRow_.resize(nGroupsInRow_);
# Line 193 | Line 245 | namespace OpenMD {
245        }      
246      }
247  
248 < #endif
197 <
198 <    // allocate memory for the parallel objects
199 <    atypesLocal.resize(nLocal_);
200 <
201 <    for (int i = 0; i < nLocal_; i++)
202 <      atypesLocal[i] = ff_->getAtomType(idents[i]);
203 <
204 <    groupList_.clear();
205 <    groupList_.resize(nGroups_);
206 <    for (int i = 0; i < nGroups_; i++) {
207 <      int gid = cgLocalToGlobal[i];
208 <      for (int j = 0; j < nLocal_; j++) {
209 <        int aid = AtomLocalToGlobal[j];
210 <        if (globalGroupMembership[aid] == gid) {
211 <          groupList_[i].push_back(j);
212 <        }
213 <      }      
214 <    }
215 <
248 > #else
249      excludesForAtom.clear();
250      excludesForAtom.resize(nLocal_);
251      toposForAtom.clear();
# Line 245 | 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 253 | 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 263 | 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 326 | Line 381 | namespace OpenMD {
381      vector<RealType> groupCutoff(nGroups_, 0.0);
382      groupToGtype.resize(nGroups_);
383      for (int cg1 = 0; cg1 < nGroups_; cg1++) {
329 –
384        groupCutoff[cg1] = 0.0;
385        vector<int> atomList = getAtomsInGroupRow(cg1);
332 –
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];
339 –        }
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 346 | 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 390 | Line 442 | namespace OpenMD {
442  
443          pair<int,int> key = make_pair(i,j);
444          gTypeCutoffMap[key].first = thisRcut;
393 –
445          if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
395 –
446          gTypeCutoffMap[key].second = thisRcut*thisRcut;
397 –        
447          gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
399 –
448          // sanity check
449          
450          if (userChoseCutoff_) {
# Line 488 | Line 536 | namespace OpenMD {
536             atomColData.skippedCharge.end(), 0.0);
537      }
538  
539 +    if (storageLayout_ & DataStorage::dslFlucQForce) {      
540 +      fill(atomRowData.flucQFrc.begin(),
541 +           atomRowData.flucQFrc.end(), 0.0);
542 +      fill(atomColData.flucQFrc.begin(),
543 +           atomColData.flucQFrc.end(), 0.0);
544 +    }
545 +
546 +    if (storageLayout_ & DataStorage::dslElectricField) {    
547 +      fill(atomRowData.electricField.begin(),
548 +           atomRowData.electricField.end(), V3Zero);
549 +      fill(atomColData.electricField.begin(),
550 +           atomColData.electricField.end(), V3Zero);
551 +    }
552 +
553 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
554 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
555 +           0.0);
556 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
557 +           0.0);
558 +    }
559 +
560   #endif
561      // even in parallel, we need to zero out the local arrays:
562  
# Line 500 | Line 569 | namespace OpenMD {
569        fill(snap_->atomData.density.begin(),
570             snap_->atomData.density.end(), 0.0);
571      }
572 +
573      if (storageLayout_ & DataStorage::dslFunctional) {
574        fill(snap_->atomData.functional.begin(),
575             snap_->atomData.functional.end(), 0.0);
576      }
577 +
578      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
579        fill(snap_->atomData.functionalDerivative.begin(),
580             snap_->atomData.functionalDerivative.end(), 0.0);
581      }
582 +
583      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
584        fill(snap_->atomData.skippedCharge.begin(),
585             snap_->atomData.skippedCharge.end(), 0.0);
586      }
587 <    
587 >
588 >    if (storageLayout_ & DataStorage::dslElectricField) {      
589 >      fill(snap_->atomData.electricField.begin(),
590 >           snap_->atomData.electricField.end(), V3Zero);
591 >    }
592    }
593  
594  
# Line 522 | Line 598 | namespace OpenMD {
598   #ifdef IS_MPI
599      
600      // gather up the atomic positions
601 <    AtomCommVectorRow->gather(snap_->atomData.position,
601 >    AtomPlanVectorRow->gather(snap_->atomData.position,
602                                atomRowData.position);
603 <    AtomCommVectorColumn->gather(snap_->atomData.position,
603 >    AtomPlanVectorColumn->gather(snap_->atomData.position,
604                                   atomColData.position);
605      
606      // gather up the cutoff group positions
607 <    cgCommVectorRow->gather(snap_->cgData.position,
607 >
608 >    cgPlanVectorRow->gather(snap_->cgData.position,
609                              cgRowData.position);
610 <    cgCommVectorColumn->gather(snap_->cgData.position,
610 >
611 >    cgPlanVectorColumn->gather(snap_->cgData.position,
612                                 cgColData.position);
613 +
614 +
615 +
616 +    if (needVelocities_) {
617 +      // gather up the atomic velocities
618 +      AtomPlanVectorColumn->gather(snap_->atomData.velocity,
619 +                                   atomColData.velocity);
620 +      
621 +      cgPlanVectorColumn->gather(snap_->cgData.velocity,
622 +                                 cgColData.velocity);
623 +    }
624 +
625      
626      // if needed, gather the atomic rotation matrices
627      if (storageLayout_ & DataStorage::dslAmat) {
628 <      AtomCommMatrixRow->gather(snap_->atomData.aMat,
628 >      AtomPlanMatrixRow->gather(snap_->atomData.aMat,
629                                  atomRowData.aMat);
630 <      AtomCommMatrixColumn->gather(snap_->atomData.aMat,
630 >      AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
631                                     atomColData.aMat);
632      }
633      
634      // if needed, gather the atomic eletrostatic frames
635      if (storageLayout_ & DataStorage::dslElectroFrame) {
636 <      AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
636 >      AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
637                                  atomRowData.electroFrame);
638 <      AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
638 >      AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
639                                     atomColData.electroFrame);
640      }
641  
642 +    // if needed, gather the atomic fluctuating charge values
643 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
644 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
645 +                              atomRowData.flucQPos);
646 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
647 +                                 atomColData.flucQPos);
648 +    }
649 +
650   #endif      
651    }
652    
# Line 562 | Line 660 | namespace OpenMD {
660      
661      if (storageLayout_ & DataStorage::dslDensity) {
662        
663 <      AtomCommRealRow->scatter(atomRowData.density,
663 >      AtomPlanRealRow->scatter(atomRowData.density,
664                                 snap_->atomData.density);
665        
666        int n = snap_->atomData.density.size();
667        vector<RealType> rho_tmp(n, 0.0);
668 <      AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
668 >      AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
669        for (int i = 0; i < n; i++)
670          snap_->atomData.density[i] += rho_tmp[i];
671      }
672 +
673 +    if (storageLayout_ & DataStorage::dslElectricField) {
674 +      
675 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
676 +                                 snap_->atomData.electricField);
677 +      
678 +      int n = snap_->atomData.electricField.size();
679 +      vector<Vector3d> field_tmp(n, V3Zero);
680 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
681 +      for (int i = 0; i < n; i++)
682 +        snap_->atomData.electricField[i] += field_tmp[i];
683 +    }
684   #endif
685    }
686  
# Line 583 | Line 693 | namespace OpenMD {
693      storageLayout_ = sman_->getStorageLayout();
694   #ifdef IS_MPI
695      if (storageLayout_ & DataStorage::dslFunctional) {
696 <      AtomCommRealRow->gather(snap_->atomData.functional,
696 >      AtomPlanRealRow->gather(snap_->atomData.functional,
697                                atomRowData.functional);
698 <      AtomCommRealColumn->gather(snap_->atomData.functional,
698 >      AtomPlanRealColumn->gather(snap_->atomData.functional,
699                                   atomColData.functional);
700      }
701      
702      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
703 <      AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
703 >      AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
704                                atomRowData.functionalDerivative);
705 <      AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
705 >      AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
706                                   atomColData.functionalDerivative);
707      }
708   #endif
# Line 606 | Line 716 | namespace OpenMD {
716      int n = snap_->atomData.force.size();
717      vector<Vector3d> frc_tmp(n, V3Zero);
718      
719 <    AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
719 >    AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
720      for (int i = 0; i < n; i++) {
721        snap_->atomData.force[i] += frc_tmp[i];
722        frc_tmp[i] = 0.0;
723      }
724      
725 <    AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
726 <    for (int i = 0; i < n; i++)
725 >    AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
726 >    for (int i = 0; i < n; i++) {
727        snap_->atomData.force[i] += frc_tmp[i];
728 +    }
729          
730      if (storageLayout_ & DataStorage::dslTorque) {
731  
732        int nt = snap_->atomData.torque.size();
733        vector<Vector3d> trq_tmp(nt, V3Zero);
734  
735 <      AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
735 >      AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
736        for (int i = 0; i < nt; i++) {
737          snap_->atomData.torque[i] += trq_tmp[i];
738          trq_tmp[i] = 0.0;
739        }
740        
741 <      AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
741 >      AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
742        for (int i = 0; i < nt; i++)
743          snap_->atomData.torque[i] += trq_tmp[i];
744      }
# Line 637 | Line 748 | namespace OpenMD {
748        int ns = snap_->atomData.skippedCharge.size();
749        vector<RealType> skch_tmp(ns, 0.0);
750  
751 <      AtomCommRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
751 >      AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
752        for (int i = 0; i < ns; i++) {
753          snap_->atomData.skippedCharge[i] += skch_tmp[i];
754          skch_tmp[i] = 0.0;
755        }
756        
757 <      AtomCommRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
758 <      for (int i = 0; i < ns; i++)
757 >      AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
758 >      for (int i = 0; i < ns; i++)
759          snap_->atomData.skippedCharge[i] += skch_tmp[i];
760 +            
761      }
762      
763 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
764 +
765 +      int nq = snap_->atomData.flucQFrc.size();
766 +      vector<RealType> fqfrc_tmp(nq, 0.0);
767 +
768 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
769 +      for (int i = 0; i < nq; i++) {
770 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
771 +        fqfrc_tmp[i] = 0.0;
772 +      }
773 +      
774 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
775 +      for (int i = 0; i < nq; i++)
776 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
777 +            
778 +    }
779 +
780      nLocal_ = snap_->getNumberOfAtoms();
781  
782      vector<potVec> pot_temp(nLocal_,
# Line 655 | Line 784 | namespace OpenMD {
784  
785      // scatter/gather pot_row into the members of my column
786            
787 <    AtomCommPotRow->scatter(pot_row, pot_temp);
787 >    AtomPlanPotRow->scatter(pot_row, pot_temp);
788  
789      for (int ii = 0;  ii < pot_temp.size(); ii++ )
790        pairwisePot += pot_temp[ii];
791 <    
791 >        
792 >    if (storageLayout_ & DataStorage::dslParticlePot) {
793 >      // This is the pairwise contribution to the particle pot.  The
794 >      // embedding contribution is added in each of the low level
795 >      // non-bonded routines.  In single processor, this is done in
796 >      // unpackInteractionData, not in collectData.
797 >      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
798 >        for (int i = 0; i < nLocal_; i++) {
799 >          // factor of two is because the total potential terms are divided
800 >          // by 2 in parallel due to row/ column scatter      
801 >          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
802 >        }
803 >      }
804 >    }
805 >
806      fill(pot_temp.begin(), pot_temp.end(),
807           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
808        
809 <    AtomCommPotColumn->scatter(pot_col, pot_temp);    
809 >    AtomPlanPotColumn->scatter(pot_col, pot_temp);    
810      
811      for (int ii = 0;  ii < pot_temp.size(); ii++ )
812        pairwisePot += pot_temp[ii];    
813 +
814 +    if (storageLayout_ & DataStorage::dslParticlePot) {
815 +      // This is the pairwise contribution to the particle pot.  The
816 +      // embedding contribution is added in each of the low level
817 +      // non-bonded routines.  In single processor, this is done in
818 +      // unpackInteractionData, not in collectData.
819 +      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
820 +        for (int i = 0; i < nLocal_; i++) {
821 +          // factor of two is because the total potential terms are divided
822 +          // by 2 in parallel due to row/ column scatter      
823 +          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
824 +        }
825 +      }
826 +    }
827 +    
828 +    if (storageLayout_ & DataStorage::dslParticlePot) {
829 +      int npp = snap_->atomData.particlePot.size();
830 +      vector<RealType> ppot_temp(npp, 0.0);
831 +
832 +      // This is the direct or embedding contribution to the particle
833 +      // pot.
834 +      
835 +      AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
836 +      for (int i = 0; i < npp; i++) {
837 +        snap_->atomData.particlePot[i] += ppot_temp[i];
838 +      }
839 +
840 +      fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
841 +      
842 +      AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
843 +      for (int i = 0; i < npp; i++) {
844 +        snap_->atomData.particlePot[i] += ppot_temp[i];
845 +      }
846 +    }
847 +
848 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
849 +      RealType ploc1 = pairwisePot[ii];
850 +      RealType ploc2 = 0.0;
851 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
852 +      pairwisePot[ii] = ploc2;
853 +    }
854 +
855 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
856 +      RealType ploc1 = embeddingPot[ii];
857 +      RealType ploc2 = 0.0;
858 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
859 +      embeddingPot[ii] = ploc2;
860 +    }
861 +    
862 +    // Here be dragons.
863 +    MPI::Intracomm col = colComm.getComm();
864 +
865 +    col.Allreduce(MPI::IN_PLACE,
866 +                  &snap_->frameData.conductiveHeatFlux[0], 3,
867 +                  MPI::REALTYPE, MPI::SUM);
868 +
869 +
870   #endif
871  
872    }
# Line 711 | Line 911 | namespace OpenMD {
911      return d;    
912    }
913  
914 +  Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
915 + #ifdef IS_MPI
916 +    return cgColData.velocity[cg2];
917 + #else
918 +    return snap_->cgData.velocity[cg2];
919 + #endif
920 +  }
921  
922 +  Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
923 + #ifdef IS_MPI
924 +    return atomColData.velocity[atom2];
925 + #else
926 +    return snap_->atomData.velocity[atom2];
927 + #endif
928 +  }
929 +
930 +
931    Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
932  
933      Vector3d d;
# Line 779 | Line 995 | namespace OpenMD {
995     */
996    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
997      int unique_id_1, unique_id_2;
998 <
998 >        
999   #ifdef IS_MPI
1000      // in MPI, we have to look up the unique IDs for each atom
1001      unique_id_1 = AtomRowToGlobal[atom1];
1002      unique_id_2 = AtomColToGlobal[atom2];
1003 + #else
1004 +    unique_id_1 = AtomLocalToGlobal[atom1];
1005 +    unique_id_2 = AtomLocalToGlobal[atom2];
1006 + #endif  
1007  
788 –    // this situation should only arise in MPI simulations
1008      if (unique_id_1 == unique_id_2) return true;
1009 <    
1009 >
1010 > #ifdef IS_MPI
1011      // this prevents us from doing the pair on multiple processors
1012      if (unique_id_1 < unique_id_2) {
1013        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1014      } else {
1015 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1015 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1016      }
1017   #endif
1018 +    
1019      return false;
1020    }
1021  
# Line 808 | Line 1029 | namespace OpenMD {
1029     * field) must still be handled for these pairs.
1030     */
1031    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1032 <    int unique_id_2;
1032 >
1033 >    // excludesForAtom was constructed to use row/column indices in the MPI
1034 >    // version, and to use local IDs in the non-MPI version:
1035      
813 – #ifdef IS_MPI
814 –    // in MPI, we have to look up the unique IDs for the row atom.
815 –    unique_id_2 = AtomColToGlobal[atom2];
816 – #else
817 –    // in the normal loop, the atom numbers are unique
818 –    unique_id_2 = atom2;
819 – #endif
820 –    
1036      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1037           i != excludesForAtom[atom1].end(); ++i) {
1038 <      if ( (*i) == unique_id_2 ) return true;
1038 >      if ( (*i) == atom2 ) return true;
1039      }
1040  
1041      return false;
# Line 894 | Line 1109 | namespace OpenMD {
1109        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1110      }
1111  
1112 < #else
1112 >    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1113 >      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1114 >      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1115 >    }
1116  
1117 + #else
1118 +    
1119      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
900 –    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
901 –    //                         ff_->getAtomType(idents[atom2]) );
1120  
1121      if (storageLayout_ & DataStorage::dslAmat) {
1122        idat.A1 = &(snap_->atomData.aMat[atom1]);
# Line 939 | Line 1157 | namespace OpenMD {
1157        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1158        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1159      }
1160 +
1161 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1162 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1163 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1164 +    }
1165 +
1166   #endif
1167    }
1168  
1169    
1170    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1171   #ifdef IS_MPI
1172 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1173 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1172 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1173 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1174  
1175      atomRowData.force[atom1] += *(idat.f1);
1176      atomColData.force[atom2] -= *(idat.f1);
1177 +
1178 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1179 +      atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1180 +      atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1181 +    }
1182 +
1183 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1184 +      atomRowData.electricField[atom1] += *(idat.eField1);
1185 +      atomColData.electricField[atom2] += *(idat.eField2);
1186 +    }
1187 +
1188   #else
1189      pairwisePot += *(idat.pot);
1190  
1191      snap_->atomData.force[atom1] += *(idat.f1);
1192      snap_->atomData.force[atom2] -= *(idat.f1);
1193 +
1194 +    if (idat.doParticlePot) {
1195 +      // This is the pairwise contribution to the particle pot.  The
1196 +      // embedding contribution is added in each of the low level
1197 +      // non-bonded routines.  In parallel, this calculation is done
1198 +      // in collectData, not in unpackInteractionData.
1199 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1200 +      snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1201 +    }
1202 +    
1203 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1204 +      snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1205 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1206 +    }
1207 +
1208 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1209 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1210 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1211 +    }
1212 +
1213   #endif
1214      
1215    }
# Line 1036 | Line 1291 | namespace OpenMD {
1291          // add this cutoff group to the list of groups in this cell;
1292          cellListRow_[cellIndex].push_back(i);
1293        }
1039 –      
1294        for (int i = 0; i < nGroupsInCol_; i++) {
1295          rs = cgColData.position[i];
1296          
# Line 1061 | Line 1315 | namespace OpenMD {
1315          // add this cutoff group to the list of groups in this cell;
1316          cellListCol_[cellIndex].push_back(i);
1317        }
1318 +    
1319   #else
1320        for (int i = 0; i < nGroups_; i++) {
1321          rs = snap_->cgData.position[i];
# Line 1081 | Line 1336 | namespace OpenMD {
1336          whichCell.z() = nCells_.z() * scaled.z();
1337          
1338          // find single index of this cell:
1339 <        cellIndex = Vlinear(whichCell, nCells_);      
1339 >        cellIndex = Vlinear(whichCell, nCells_);
1340          
1341          // add this cutoff group to the list of groups in this cell;
1342          cellList_[cellIndex].push_back(i);
1343        }
1344 +
1345   #endif
1346  
1347        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1098 | Line 1354 | namespace OpenMD {
1354                   os != cellOffsets_.end(); ++os) {
1355                
1356                Vector3i m2v = m1v + (*os);
1357 <              
1357 >            
1358 >
1359                if (m2v.x() >= nCells_.x()) {
1360                  m2v.x() = 0;          
1361                } else if (m2v.x() < 0) {
# Line 1116 | Line 1373 | namespace OpenMD {
1373                } else if (m2v.z() < 0) {
1374                  m2v.z() = nCells_.z() - 1;
1375                }
1376 <              
1376 >
1377                int m2 = Vlinear (m2v, nCells_);
1378                
1379   #ifdef IS_MPI
# Line 1125 | Line 1382 | namespace OpenMD {
1382                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1383                       j2 != cellListCol_[m2].end(); ++j2) {
1384                    
1385 <                  // Always do this if we're in different cells or if
1386 <                  // we're in the same cell and the global index of the
1387 <                  // j2 cutoff group is less than the j1 cutoff group
1388 <                  
1389 <                  if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) {
1390 <                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1391 <                    snap_->wrapVector(dr);
1392 <                    cuts = getGroupCutoffs( (*j1), (*j2) );
1393 <                    if (dr.lengthSquare() < cuts.third) {
1137 <                      neighborList.push_back(make_pair((*j1), (*j2)));
1138 <                    }
1139 <                  }
1385 >                  // In parallel, we need to visit *all* pairs of row
1386 >                  // & column indicies and will divide labor in the
1387 >                  // force evaluation later.
1388 >                  dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1389 >                  snap_->wrapVector(dr);
1390 >                  cuts = getGroupCutoffs( (*j1), (*j2) );
1391 >                  if (dr.lengthSquare() < cuts.third) {
1392 >                    neighborList.push_back(make_pair((*j1), (*j2)));
1393 >                  }                  
1394                  }
1395                }
1396   #else
1143 –              
1397                for (vector<int>::iterator j1 = cellList_[m1].begin();
1398                     j1 != cellList_[m1].end(); ++j1) {
1399                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1400                       j2 != cellList_[m2].end(); ++j2) {
1401 <                  
1401 >    
1402                    // Always do this if we're in different cells or if
1403 <                  // we're in the same cell and the global index of the
1404 <                  // j2 cutoff group is less than the j1 cutoff group
1405 <                  
1406 <                  if (m2 != m1 || (*j2) < (*j1)) {
1403 >                  // we're in the same cell and the global index of
1404 >                  // the j2 cutoff group is greater than or equal to
1405 >                  // the j1 cutoff group.  Note that Rappaport's code
1406 >                  // has a "less than" conditional here, but that
1407 >                  // deals with atom-by-atom computation.  OpenMD
1408 >                  // allows atoms within a single cutoff group to
1409 >                  // interact with each other.
1410 >
1411 >
1412 >
1413 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1414 >
1415                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1416                      snap_->wrapVector(dr);
1417                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1169 | Line 1430 | namespace OpenMD {
1430        // branch to do all cutoff group pairs
1431   #ifdef IS_MPI
1432        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1433 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1433 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1434            dr = cgColData.position[j2] - cgRowData.position[j1];
1435            snap_->wrapVector(dr);
1436            cuts = getGroupCutoffs( j1, j2 );
# Line 1177 | Line 1438 | namespace OpenMD {
1438              neighborList.push_back(make_pair(j1, j2));
1439            }
1440          }
1441 <      }
1441 >      }      
1442   #else
1443 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1444 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1443 >      // include all groups here.
1444 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1445 >        // include self group interactions j2 == j1
1446 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1447            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1448            snap_->wrapVector(dr);
1449            cuts = getGroupCutoffs( j1, j2 );
1450            if (dr.lengthSquare() < cuts.third) {
1451              neighborList.push_back(make_pair(j1, j2));
1452            }
1453 <        }
1454 <      }        
1453 >        }    
1454 >      }
1455   #endif
1456      }
1457        

Diff Legend

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