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 1688 by gezelter, Wed Mar 14 17:56:01 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 > #endif
529 >    // even in parallel, we need to zero out the local arrays:
530 >
531      if (storageLayout_ & DataStorage::dslParticlePot) {      
532        fill(snap_->atomData.particlePot.begin(),
533             snap_->atomData.particlePot.end(), 0.0);
# Line 485 | Line 549 | namespace OpenMD {
549        fill(snap_->atomData.skippedCharge.begin(),
550             snap_->atomData.skippedCharge.end(), 0.0);
551      }
488 – #endif
552      
553    }
554  
# Line 496 | Line 559 | namespace OpenMD {
559   #ifdef IS_MPI
560      
561      // gather up the atomic positions
562 <    AtomCommVectorRow->gather(snap_->atomData.position,
562 >    AtomPlanVectorRow->gather(snap_->atomData.position,
563                                atomRowData.position);
564 <    AtomCommVectorColumn->gather(snap_->atomData.position,
564 >    AtomPlanVectorColumn->gather(snap_->atomData.position,
565                                   atomColData.position);
566      
567      // gather up the cutoff group positions
568 <    cgCommVectorRow->gather(snap_->cgData.position,
568 >
569 >    cgPlanVectorRow->gather(snap_->cgData.position,
570                              cgRowData.position);
571 <    cgCommVectorColumn->gather(snap_->cgData.position,
571 >
572 >    cgPlanVectorColumn->gather(snap_->cgData.position,
573                                 cgColData.position);
574 +
575      
576      // if needed, gather the atomic rotation matrices
577      if (storageLayout_ & DataStorage::dslAmat) {
578 <      AtomCommMatrixRow->gather(snap_->atomData.aMat,
578 >      AtomPlanMatrixRow->gather(snap_->atomData.aMat,
579                                  atomRowData.aMat);
580 <      AtomCommMatrixColumn->gather(snap_->atomData.aMat,
580 >      AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
581                                     atomColData.aMat);
582      }
583      
584      // if needed, gather the atomic eletrostatic frames
585      if (storageLayout_ & DataStorage::dslElectroFrame) {
586 <      AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
586 >      AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
587                                  atomRowData.electroFrame);
588 <      AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
588 >      AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
589                                     atomColData.electroFrame);
590      }
591 +
592   #endif      
593    }
594    
# Line 535 | Line 602 | namespace OpenMD {
602      
603      if (storageLayout_ & DataStorage::dslDensity) {
604        
605 <      AtomCommRealRow->scatter(atomRowData.density,
605 >      AtomPlanRealRow->scatter(atomRowData.density,
606                                 snap_->atomData.density);
607        
608        int n = snap_->atomData.density.size();
609        vector<RealType> rho_tmp(n, 0.0);
610 <      AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
610 >      AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
611        for (int i = 0; i < n; i++)
612          snap_->atomData.density[i] += rho_tmp[i];
613      }
# Line 556 | Line 623 | namespace OpenMD {
623      storageLayout_ = sman_->getStorageLayout();
624   #ifdef IS_MPI
625      if (storageLayout_ & DataStorage::dslFunctional) {
626 <      AtomCommRealRow->gather(snap_->atomData.functional,
626 >      AtomPlanRealRow->gather(snap_->atomData.functional,
627                                atomRowData.functional);
628 <      AtomCommRealColumn->gather(snap_->atomData.functional,
628 >      AtomPlanRealColumn->gather(snap_->atomData.functional,
629                                   atomColData.functional);
630      }
631      
632      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
633 <      AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
633 >      AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
634                                atomRowData.functionalDerivative);
635 <      AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
635 >      AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
636                                   atomColData.functionalDerivative);
637      }
638   #endif
# Line 579 | Line 646 | namespace OpenMD {
646      int n = snap_->atomData.force.size();
647      vector<Vector3d> frc_tmp(n, V3Zero);
648      
649 <    AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
649 >    AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
650      for (int i = 0; i < n; i++) {
651        snap_->atomData.force[i] += frc_tmp[i];
652        frc_tmp[i] = 0.0;
653      }
654      
655 <    AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
656 <    for (int i = 0; i < n; i++)
655 >    AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
656 >    for (int i = 0; i < n; i++) {
657        snap_->atomData.force[i] += frc_tmp[i];
658 <    
659 <    
658 >    }
659 >        
660      if (storageLayout_ & DataStorage::dslTorque) {
661  
662        int nt = snap_->atomData.torque.size();
663        vector<Vector3d> trq_tmp(nt, V3Zero);
664  
665 <      AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
665 >      AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
666        for (int i = 0; i < nt; i++) {
667          snap_->atomData.torque[i] += trq_tmp[i];
668          trq_tmp[i] = 0.0;
669        }
670        
671 <      AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
671 >      AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
672        for (int i = 0; i < nt; i++)
673          snap_->atomData.torque[i] += trq_tmp[i];
674      }
# Line 611 | Line 678 | namespace OpenMD {
678        int ns = snap_->atomData.skippedCharge.size();
679        vector<RealType> skch_tmp(ns, 0.0);
680  
681 <      AtomCommRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
681 >      AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
682        for (int i = 0; i < ns; i++) {
683 <        snap_->atomData.skippedCharge[i] = skch_tmp[i];
683 >        snap_->atomData.skippedCharge[i] += skch_tmp[i];
684          skch_tmp[i] = 0.0;
685        }
686        
687 <      AtomCommRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
688 <      for (int i = 0; i < ns; i++)
687 >      AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
688 >      for (int i = 0; i < ns; i++)
689          snap_->atomData.skippedCharge[i] += skch_tmp[i];
690 +            
691      }
692      
693      nLocal_ = snap_->getNumberOfAtoms();
# Line 629 | Line 697 | namespace OpenMD {
697  
698      // scatter/gather pot_row into the members of my column
699            
700 <    AtomCommPotRow->scatter(pot_row, pot_temp);
700 >    AtomPlanPotRow->scatter(pot_row, pot_temp);
701  
702      for (int ii = 0;  ii < pot_temp.size(); ii++ )
703        pairwisePot += pot_temp[ii];
# Line 637 | Line 705 | namespace OpenMD {
705      fill(pot_temp.begin(), pot_temp.end(),
706           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
707        
708 <    AtomCommPotColumn->scatter(pot_col, pot_temp);    
708 >    AtomPlanPotColumn->scatter(pot_col, pot_temp);    
709      
710      for (int ii = 0;  ii < pot_temp.size(); ii++ )
711        pairwisePot += pot_temp[ii];    
712 +    
713 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
714 +      RealType ploc1 = pairwisePot[ii];
715 +      RealType ploc2 = 0.0;
716 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
717 +      pairwisePot[ii] = ploc2;
718 +    }
719 +
720 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
721 +      RealType ploc1 = embeddingPot[ii];
722 +      RealType ploc2 = 0.0;
723 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
724 +      embeddingPot[ii] = ploc2;
725 +    }
726 +
727   #endif
728  
729    }
# Line 753 | Line 836 | namespace OpenMD {
836     */
837    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
838      int unique_id_1, unique_id_2;
839 <
839 >        
840   #ifdef IS_MPI
841      // in MPI, we have to look up the unique IDs for each atom
842      unique_id_1 = AtomRowToGlobal[atom1];
843      unique_id_2 = AtomColToGlobal[atom2];
844 + #else
845 +    unique_id_1 = AtomLocalToGlobal[atom1];
846 +    unique_id_2 = AtomLocalToGlobal[atom2];
847 + #endif  
848  
762 –    // this situation should only arise in MPI simulations
849      if (unique_id_1 == unique_id_2) return true;
850 <    
850 >
851 > #ifdef IS_MPI
852      // this prevents us from doing the pair on multiple processors
853      if (unique_id_1 < unique_id_2) {
854        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
855      } else {
856 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
856 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
857      }
858   #endif
859 +    
860      return false;
861    }
862  
# Line 782 | Line 870 | namespace OpenMD {
870     * field) must still be handled for these pairs.
871     */
872    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
873 <    int unique_id_2;
874 <    
875 < #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
873 >
874 >    // excludesForAtom was constructed to use row/column indices in the MPI
875 >    // version, and to use local IDs in the non-MPI version:
876      
877      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
878           i != excludesForAtom[atom1].end(); ++i) {
879 <      if ( (*i) == unique_id_2 ) return true;
879 >      if ( (*i) == atom2 ) return true;
880      }
881  
882      return false;
# Line 824 | Line 906 | namespace OpenMD {
906      idat.excluded = excludeAtomPair(atom1, atom2);
907    
908   #ifdef IS_MPI
909 +    idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
910 +    //idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
911 +    //                         ff_->getAtomType(identsCol[atom2]) );
912      
828 –    idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
829 –                             ff_->getAtomType(identsCol[atom2]) );
830 –    
913      if (storageLayout_ & DataStorage::dslAmat) {
914        idat.A1 = &(atomRowData.aMat[atom1]);
915        idat.A2 = &(atomColData.aMat[atom2]);
# Line 869 | Line 951 | namespace OpenMD {
951      }
952  
953   #else
954 +    
955  
956 <    idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
957 <                             ff_->getAtomType(idents[atom2]) );
956 >    // cerr << "atoms = " << atom1 << " " << atom2 << "\n";
957 >    // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n";
958 >    // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n";
959  
960 +    idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
961 +    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
962 +    //                         ff_->getAtomType(idents[atom2]) );
963 +
964      if (storageLayout_ & DataStorage::dslAmat) {
965        idat.A1 = &(snap_->atomData.aMat[atom1]);
966        idat.A2 = &(snap_->atomData.aMat[atom2]);
# Line 918 | Line 1006 | namespace OpenMD {
1006    
1007    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1008   #ifdef IS_MPI
1009 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1010 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1009 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1010 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1011  
1012      atomRowData.force[atom1] += *(idat.f1);
1013      atomColData.force[atom2] -= *(idat.f1);
# Line 1009 | Line 1097 | namespace OpenMD {
1097          // add this cutoff group to the list of groups in this cell;
1098          cellListRow_[cellIndex].push_back(i);
1099        }
1012 –      
1100        for (int i = 0; i < nGroupsInCol_; i++) {
1101          rs = cgColData.position[i];
1102          
# Line 1034 | Line 1121 | namespace OpenMD {
1121          // add this cutoff group to the list of groups in this cell;
1122          cellListCol_[cellIndex].push_back(i);
1123        }
1124 +    
1125   #else
1126        for (int i = 0; i < nGroups_; i++) {
1127          rs = snap_->cgData.position[i];
# Line 1054 | Line 1142 | namespace OpenMD {
1142          whichCell.z() = nCells_.z() * scaled.z();
1143          
1144          // find single index of this cell:
1145 <        cellIndex = Vlinear(whichCell, nCells_);      
1145 >        cellIndex = Vlinear(whichCell, nCells_);
1146          
1147          // add this cutoff group to the list of groups in this cell;
1148          cellList_[cellIndex].push_back(i);
1149        }
1150 +
1151   #endif
1152  
1153        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1071 | Line 1160 | namespace OpenMD {
1160                   os != cellOffsets_.end(); ++os) {
1161                
1162                Vector3i m2v = m1v + (*os);
1163 <              
1163 >            
1164 >
1165                if (m2v.x() >= nCells_.x()) {
1166                  m2v.x() = 0;          
1167                } else if (m2v.x() < 0) {
# Line 1089 | Line 1179 | namespace OpenMD {
1179                } else if (m2v.z() < 0) {
1180                  m2v.z() = nCells_.z() - 1;
1181                }
1182 <              
1182 >
1183                int m2 = Vlinear (m2v, nCells_);
1184                
1185   #ifdef IS_MPI
# Line 1098 | Line 1188 | namespace OpenMD {
1188                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1189                       j2 != cellListCol_[m2].end(); ++j2) {
1190                    
1191 <                  // Always do this if we're in different cells or if
1192 <                  // we're in the same cell and the global index of the
1193 <                  // j2 cutoff group is less than the j1 cutoff group
1194 <                  
1195 <                  if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) {
1196 <                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1197 <                    snap_->wrapVector(dr);
1198 <                    cuts = getGroupCutoffs( (*j1), (*j2) );
1199 <                    if (dr.lengthSquare() < cuts.third) {
1110 <                      neighborList.push_back(make_pair((*j1), (*j2)));
1111 <                    }
1112 <                  }
1191 >                  // In parallel, we need to visit *all* pairs of row
1192 >                  // & column indicies and will divide labor in the
1193 >                  // force evaluation later.
1194 >                  dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1195 >                  snap_->wrapVector(dr);
1196 >                  cuts = getGroupCutoffs( (*j1), (*j2) );
1197 >                  if (dr.lengthSquare() < cuts.third) {
1198 >                    neighborList.push_back(make_pair((*j1), (*j2)));
1199 >                  }                  
1200                  }
1201                }
1202   #else
1116 –              
1203                for (vector<int>::iterator j1 = cellList_[m1].begin();
1204                     j1 != cellList_[m1].end(); ++j1) {
1205                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1206                       j2 != cellList_[m2].end(); ++j2) {
1207 <                  
1207 >    
1208                    // Always do this if we're in different cells or if
1209 <                  // we're in the same cell and the global index of the
1210 <                  // j2 cutoff group is less than the j1 cutoff group
1211 <                  
1212 <                  if (m2 != m1 || (*j2) < (*j1)) {
1209 >                  // we're in the same cell and the global index of
1210 >                  // the j2 cutoff group is greater than or equal to
1211 >                  // the j1 cutoff group.  Note that Rappaport's code
1212 >                  // has a "less than" conditional here, but that
1213 >                  // deals with atom-by-atom computation.  OpenMD
1214 >                  // allows atoms within a single cutoff group to
1215 >                  // interact with each other.
1216 >
1217 >
1218 >
1219 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1220 >
1221                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1222                      snap_->wrapVector(dr);
1223                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1142 | Line 1236 | namespace OpenMD {
1236        // branch to do all cutoff group pairs
1237   #ifdef IS_MPI
1238        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1239 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1239 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1240            dr = cgColData.position[j2] - cgRowData.position[j1];
1241            snap_->wrapVector(dr);
1242            cuts = getGroupCutoffs( j1, j2 );
# Line 1150 | Line 1244 | namespace OpenMD {
1244              neighborList.push_back(make_pair(j1, j2));
1245            }
1246          }
1247 <      }
1247 >      }      
1248   #else
1249 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1250 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1249 >      // include all groups here.
1250 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1251 >        // include self group interactions j2 == j1
1252 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1253            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1254            snap_->wrapVector(dr);
1255            cuts = getGroupCutoffs( j1, j2 );
1256            if (dr.lengthSquare() < cuts.third) {
1257              neighborList.push_back(make_pair(j1, j2));
1258            }
1259 <        }
1260 <      }        
1259 >        }    
1260 >      }
1261   #endif
1262      }
1263        

Diff Legend

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