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 1589 by gezelter, Sun Jul 10 16:05:34 2011 UTC vs.
Revision 1706 by gezelter, Fri Apr 27 20:44:16 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      // allocate memory for the parallel objects
157 +    atypesRow.resize(nAtomsInRow_);
158 +    atypesCol.resize(nAtomsInCol_);
159 +
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 <    pot_row.resize(nAtomsInRow_);
181 <    pot_col.resize(nAtomsInCol_);
124 <
125 <    AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
126 <    AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
127 <    
128 <    cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
129 <    cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
180 >    AtomPlanRealRow->gather(massFactors, massFactorsRow);
181 >    AtomPlanRealColumn->gather(massFactors, massFactorsCol);
182  
131 –    AtomCommRealRow->gather(massFactors, massFactorsRow);
132 –    AtomCommRealColumn->gather(massFactors, massFactorsCol);
133 –
183      groupListRow_.clear();
184      groupListRow_.resize(nGroupsInRow_);
185      for (int i = 0; i < nGroupsInRow_; i++) {
# Line 185 | Line 234 | namespace OpenMD {
234        }      
235      }
236  
237 < #endif
189 <
190 <    groupList_.clear();
191 <    groupList_.resize(nGroups_);
192 <    for (int i = 0; i < nGroups_; i++) {
193 <      int gid = cgLocalToGlobal[i];
194 <      for (int j = 0; j < nLocal_; j++) {
195 <        int aid = AtomLocalToGlobal[j];
196 <        if (globalGroupMembership[aid] == gid) {
197 <          groupList_[i].push_back(j);
198 <        }
199 <      }      
200 <    }
201 <
237 > #else
238      excludesForAtom.clear();
239      excludesForAtom.resize(nLocal_);
240      toposForAtom.clear();
# Line 231 | 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 239 | 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 249 | 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 312 | Line 370 | namespace OpenMD {
370      vector<RealType> groupCutoff(nGroups_, 0.0);
371      groupToGtype.resize(nGroups_);
372      for (int cg1 = 0; cg1 < nGroups_; cg1++) {
315 –
373        groupCutoff[cg1] = 0.0;
374        vector<int> atomList = getAtomsInGroupRow(cg1);
318 –
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];
325 –        }
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 332 | 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 341 | 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 374 | Line 431 | namespace OpenMD {
431  
432          pair<int,int> key = make_pair(i,j);
433          gTypeCutoffMap[key].first = thisRcut;
377 –
434          if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
379 –
435          gTypeCutoffMap[key].second = thisRcut*thisRcut;
381 –        
436          gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
383 –
437          // sanity check
438          
439          if (userChoseCutoff_) {
# Line 440 | 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 450 | 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 468 | 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 479 | Line 537 | namespace OpenMD {
537        fill(snap_->atomData.density.begin(),
538             snap_->atomData.density.end(), 0.0);
539      }
540 +
541      if (storageLayout_ & DataStorage::dslFunctional) {
542        fill(snap_->atomData.functional.begin(),
543             snap_->atomData.functional.end(), 0.0);
544      }
545 +
546      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
547        fill(snap_->atomData.functionalDerivative.begin(),
548             snap_->atomData.functionalDerivative.end(), 0.0);
549      }
550 +
551      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
552        fill(snap_->atomData.skippedCharge.begin(),
553             snap_->atomData.skippedCharge.end(), 0.0);
554      }
494 – #endif
495 –    
555    }
556  
557  
# Line 502 | Line 561 | namespace OpenMD {
561   #ifdef IS_MPI
562      
563      // gather up the atomic positions
564 <    AtomCommVectorRow->gather(snap_->atomData.position,
564 >    AtomPlanVectorRow->gather(snap_->atomData.position,
565                                atomRowData.position);
566 <    AtomCommVectorColumn->gather(snap_->atomData.position,
566 >    AtomPlanVectorColumn->gather(snap_->atomData.position,
567                                   atomColData.position);
568      
569      // gather up the cutoff group positions
570 <    cgCommVectorRow->gather(snap_->cgData.position,
570 >
571 >    cgPlanVectorRow->gather(snap_->cgData.position,
572                              cgRowData.position);
573 <    cgCommVectorColumn->gather(snap_->cgData.position,
573 >
574 >    cgPlanVectorColumn->gather(snap_->cgData.position,
575                                 cgColData.position);
576 +
577      
578      // if needed, gather the atomic rotation matrices
579      if (storageLayout_ & DataStorage::dslAmat) {
580 <      AtomCommMatrixRow->gather(snap_->atomData.aMat,
580 >      AtomPlanMatrixRow->gather(snap_->atomData.aMat,
581                                  atomRowData.aMat);
582 <      AtomCommMatrixColumn->gather(snap_->atomData.aMat,
582 >      AtomPlanMatrixColumn->gather(snap_->atomData.aMat,
583                                     atomColData.aMat);
584      }
585      
586      // if needed, gather the atomic eletrostatic frames
587      if (storageLayout_ & DataStorage::dslElectroFrame) {
588 <      AtomCommMatrixRow->gather(snap_->atomData.electroFrame,
588 >      AtomPlanMatrixRow->gather(snap_->atomData.electroFrame,
589                                  atomRowData.electroFrame);
590 <      AtomCommMatrixColumn->gather(snap_->atomData.electroFrame,
590 >      AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame,
591                                     atomColData.electroFrame);
592      }
593 +
594   #endif      
595    }
596    
# Line 541 | Line 604 | namespace OpenMD {
604      
605      if (storageLayout_ & DataStorage::dslDensity) {
606        
607 <      AtomCommRealRow->scatter(atomRowData.density,
607 >      AtomPlanRealRow->scatter(atomRowData.density,
608                                 snap_->atomData.density);
609        
610        int n = snap_->atomData.density.size();
611        vector<RealType> rho_tmp(n, 0.0);
612 <      AtomCommRealColumn->scatter(atomColData.density, rho_tmp);
612 >      AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
613        for (int i = 0; i < n; i++)
614          snap_->atomData.density[i] += rho_tmp[i];
615      }
# Line 562 | Line 625 | namespace OpenMD {
625      storageLayout_ = sman_->getStorageLayout();
626   #ifdef IS_MPI
627      if (storageLayout_ & DataStorage::dslFunctional) {
628 <      AtomCommRealRow->gather(snap_->atomData.functional,
628 >      AtomPlanRealRow->gather(snap_->atomData.functional,
629                                atomRowData.functional);
630 <      AtomCommRealColumn->gather(snap_->atomData.functional,
630 >      AtomPlanRealColumn->gather(snap_->atomData.functional,
631                                   atomColData.functional);
632      }
633      
634      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {
635 <      AtomCommRealRow->gather(snap_->atomData.functionalDerivative,
635 >      AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
636                                atomRowData.functionalDerivative);
637 <      AtomCommRealColumn->gather(snap_->atomData.functionalDerivative,
637 >      AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
638                                   atomColData.functionalDerivative);
639      }
640   #endif
# Line 585 | Line 648 | namespace OpenMD {
648      int n = snap_->atomData.force.size();
649      vector<Vector3d> frc_tmp(n, V3Zero);
650      
651 <    AtomCommVectorRow->scatter(atomRowData.force, frc_tmp);
651 >    AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp);
652      for (int i = 0; i < n; i++) {
653        snap_->atomData.force[i] += frc_tmp[i];
654        frc_tmp[i] = 0.0;
655      }
656      
657 <    AtomCommVectorColumn->scatter(atomColData.force, frc_tmp);
658 <    for (int i = 0; i < n; i++)
657 >    AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
658 >    for (int i = 0; i < n; i++) {
659        snap_->atomData.force[i] += frc_tmp[i];
660 <    
661 <    
660 >    }
661 >        
662      if (storageLayout_ & DataStorage::dslTorque) {
663  
664        int nt = snap_->atomData.torque.size();
665        vector<Vector3d> trq_tmp(nt, V3Zero);
666  
667 <      AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp);
667 >      AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp);
668        for (int i = 0; i < nt; i++) {
669          snap_->atomData.torque[i] += trq_tmp[i];
670          trq_tmp[i] = 0.0;
671        }
672        
673 <      AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp);
673 >      AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
674        for (int i = 0; i < nt; i++)
675          snap_->atomData.torque[i] += trq_tmp[i];
676      }
# Line 617 | Line 680 | namespace OpenMD {
680        int ns = snap_->atomData.skippedCharge.size();
681        vector<RealType> skch_tmp(ns, 0.0);
682  
683 <      AtomCommRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
683 >      AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp);
684        for (int i = 0; i < ns; i++) {
685 <        snap_->atomData.skippedCharge[i] = skch_tmp[i];
685 >        snap_->atomData.skippedCharge[i] += skch_tmp[i];
686          skch_tmp[i] = 0.0;
687        }
688        
689 <      AtomCommRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
690 <      for (int i = 0; i < ns; i++)
689 >      AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
690 >      for (int i = 0; i < ns; i++)
691          snap_->atomData.skippedCharge[i] += skch_tmp[i];
692 +            
693      }
694      
695      nLocal_ = snap_->getNumberOfAtoms();
# Line 635 | Line 699 | namespace OpenMD {
699  
700      // scatter/gather pot_row into the members of my column
701            
702 <    AtomCommPotRow->scatter(pot_row, pot_temp);
702 >    AtomPlanPotRow->scatter(pot_row, pot_temp);
703  
704      for (int ii = 0;  ii < pot_temp.size(); ii++ )
705        pairwisePot += pot_temp[ii];
# Line 643 | Line 707 | namespace OpenMD {
707      fill(pot_temp.begin(), pot_temp.end(),
708           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
709        
710 <    AtomCommPotColumn->scatter(pot_col, pot_temp);    
710 >    AtomPlanPotColumn->scatter(pot_col, pot_temp);    
711      
712      for (int ii = 0;  ii < pot_temp.size(); ii++ )
713        pairwisePot += pot_temp[ii];    
714 +    
715 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
716 +      RealType ploc1 = pairwisePot[ii];
717 +      RealType ploc2 = 0.0;
718 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
719 +      pairwisePot[ii] = ploc2;
720 +    }
721 +
722 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
723 +      RealType ploc1 = embeddingPot[ii];
724 +      RealType ploc2 = 0.0;
725 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
726 +      embeddingPot[ii] = ploc2;
727 +    }
728 +
729   #endif
730  
731    }
# Line 759 | Line 838 | namespace OpenMD {
838     */
839    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
840      int unique_id_1, unique_id_2;
841 <
841 >        
842   #ifdef IS_MPI
843      // in MPI, we have to look up the unique IDs for each atom
844      unique_id_1 = AtomRowToGlobal[atom1];
845      unique_id_2 = AtomColToGlobal[atom2];
846 + #else
847 +    unique_id_1 = AtomLocalToGlobal[atom1];
848 +    unique_id_2 = AtomLocalToGlobal[atom2];
849 + #endif  
850  
768 –    // this situation should only arise in MPI simulations
851      if (unique_id_1 == unique_id_2) return true;
852 <    
852 >
853 > #ifdef IS_MPI
854      // this prevents us from doing the pair on multiple processors
855      if (unique_id_1 < unique_id_2) {
856        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
857      } else {
858 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
858 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
859      }
860   #endif
861 +    
862      return false;
863    }
864  
# Line 788 | Line 872 | namespace OpenMD {
872     * field) must still be handled for these pairs.
873     */
874    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
875 <    int unique_id_2;
876 <    
877 < #ifdef IS_MPI
794 <    // in MPI, we have to look up the unique IDs for the row atom.
795 <    unique_id_2 = AtomColToGlobal[atom2];
796 < #else
797 <    // in the normal loop, the atom numbers are unique
798 <    unique_id_2 = atom2;
799 < #endif
875 >
876 >    // excludesForAtom was constructed to use row/column indices in the MPI
877 >    // version, and to use local IDs in the non-MPI version:
878      
879      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
880           i != excludesForAtom[atom1].end(); ++i) {
881 <      if ( (*i) == unique_id_2 ) return true;
881 >      if ( (*i) == atom2 ) return true;
882      }
883  
884      return false;
# Line 830 | Line 908 | namespace OpenMD {
908      idat.excluded = excludeAtomPair(atom1, atom2);
909    
910   #ifdef IS_MPI
911 +    idat.atypes = make_pair( atypesRow[atom1], atypesCol[atom2]);
912 +    //idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
913 +    //                         ff_->getAtomType(identsCol[atom2]) );
914      
834 –    idat.atypes = make_pair( ff_->getAtomType(identsRow[atom1]),
835 –                             ff_->getAtomType(identsCol[atom2]) );
836 –    
915      if (storageLayout_ & DataStorage::dslAmat) {
916        idat.A1 = &(atomRowData.aMat[atom1]);
917        idat.A2 = &(atomColData.aMat[atom2]);
# Line 875 | Line 953 | namespace OpenMD {
953      }
954  
955   #else
956 +    
957  
958 <    idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
959 <                             ff_->getAtomType(idents[atom2]) );
958 >    // cerr << "atoms = " << atom1 << " " << atom2 << "\n";
959 >    // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n";
960 >    // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n";
961  
962 +    idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
963 +    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
964 +    //                         ff_->getAtomType(idents[atom2]) );
965 +
966      if (storageLayout_ & DataStorage::dslAmat) {
967        idat.A1 = &(snap_->atomData.aMat[atom1]);
968        idat.A2 = &(snap_->atomData.aMat[atom2]);
# Line 924 | Line 1008 | namespace OpenMD {
1008    
1009    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1010   #ifdef IS_MPI
1011 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1012 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1011 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1012 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1013  
1014      atomRowData.force[atom1] += *(idat.f1);
1015      atomColData.force[atom2] -= *(idat.f1);
# Line 1015 | Line 1099 | namespace OpenMD {
1099          // add this cutoff group to the list of groups in this cell;
1100          cellListRow_[cellIndex].push_back(i);
1101        }
1018 –      
1102        for (int i = 0; i < nGroupsInCol_; i++) {
1103          rs = cgColData.position[i];
1104          
# Line 1040 | Line 1123 | namespace OpenMD {
1123          // add this cutoff group to the list of groups in this cell;
1124          cellListCol_[cellIndex].push_back(i);
1125        }
1126 +    
1127   #else
1128        for (int i = 0; i < nGroups_; i++) {
1129          rs = snap_->cgData.position[i];
# Line 1060 | Line 1144 | namespace OpenMD {
1144          whichCell.z() = nCells_.z() * scaled.z();
1145          
1146          // find single index of this cell:
1147 <        cellIndex = Vlinear(whichCell, nCells_);      
1147 >        cellIndex = Vlinear(whichCell, nCells_);
1148          
1149          // add this cutoff group to the list of groups in this cell;
1150          cellList_[cellIndex].push_back(i);
1151        }
1152 +
1153   #endif
1154  
1155        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1077 | Line 1162 | namespace OpenMD {
1162                   os != cellOffsets_.end(); ++os) {
1163                
1164                Vector3i m2v = m1v + (*os);
1165 <              
1165 >            
1166 >
1167                if (m2v.x() >= nCells_.x()) {
1168                  m2v.x() = 0;          
1169                } else if (m2v.x() < 0) {
# Line 1095 | Line 1181 | namespace OpenMD {
1181                } else if (m2v.z() < 0) {
1182                  m2v.z() = nCells_.z() - 1;
1183                }
1184 <              
1184 >
1185                int m2 = Vlinear (m2v, nCells_);
1186                
1187   #ifdef IS_MPI
# Line 1104 | Line 1190 | namespace OpenMD {
1190                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1191                       j2 != cellListCol_[m2].end(); ++j2) {
1192                    
1193 <                  // Always do this if we're in different cells or if
1194 <                  // we're in the same cell and the global index of the
1195 <                  // j2 cutoff group is less than the j1 cutoff group
1196 <                  
1197 <                  if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) {
1198 <                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1199 <                    snap_->wrapVector(dr);
1200 <                    cuts = getGroupCutoffs( (*j1), (*j2) );
1201 <                    if (dr.lengthSquare() < cuts.third) {
1116 <                      neighborList.push_back(make_pair((*j1), (*j2)));
1117 <                    }
1118 <                  }
1193 >                  // In parallel, we need to visit *all* pairs of row
1194 >                  // & column indicies and will divide labor in the
1195 >                  // force evaluation later.
1196 >                  dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1197 >                  snap_->wrapVector(dr);
1198 >                  cuts = getGroupCutoffs( (*j1), (*j2) );
1199 >                  if (dr.lengthSquare() < cuts.third) {
1200 >                    neighborList.push_back(make_pair((*j1), (*j2)));
1201 >                  }                  
1202                  }
1203                }
1204   #else
1122 –              
1205                for (vector<int>::iterator j1 = cellList_[m1].begin();
1206                     j1 != cellList_[m1].end(); ++j1) {
1207                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1208                       j2 != cellList_[m2].end(); ++j2) {
1209 <                  
1209 >    
1210                    // Always do this if we're in different cells or if
1211 <                  // we're in the same cell and the global index of the
1212 <                  // j2 cutoff group is less than the j1 cutoff group
1213 <                  
1214 <                  if (m2 != m1 || (*j2) < (*j1)) {
1211 >                  // we're in the same cell and the global index of
1212 >                  // the j2 cutoff group is greater than or equal to
1213 >                  // the j1 cutoff group.  Note that Rappaport's code
1214 >                  // has a "less than" conditional here, but that
1215 >                  // deals with atom-by-atom computation.  OpenMD
1216 >                  // allows atoms within a single cutoff group to
1217 >                  // interact with each other.
1218 >
1219 >
1220 >
1221 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1222 >
1223                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1224                      snap_->wrapVector(dr);
1225                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1148 | Line 1238 | namespace OpenMD {
1238        // branch to do all cutoff group pairs
1239   #ifdef IS_MPI
1240        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1241 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1241 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1242            dr = cgColData.position[j2] - cgRowData.position[j1];
1243            snap_->wrapVector(dr);
1244            cuts = getGroupCutoffs( j1, j2 );
# Line 1156 | Line 1246 | namespace OpenMD {
1246              neighborList.push_back(make_pair(j1, j2));
1247            }
1248          }
1249 <      }
1249 >      }      
1250   #else
1251 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1252 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1251 >      // include all groups here.
1252 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1253 >        // include self group interactions j2 == j1
1254 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1255            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1256            snap_->wrapVector(dr);
1257            cuts = getGroupCutoffs( j1, j2 );
1258            if (dr.lengthSquare() < cuts.third) {
1259              neighborList.push_back(make_pair(j1, j2));
1260            }
1261 <        }
1262 <      }        
1261 >        }    
1262 >      }
1263   #endif
1264      }
1265        

Diff Legend

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