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 1593 by gezelter, Fri Jul 15 21:35:14 2011 UTC vs.
trunk/src/parallel/ForceMatrixDecomposition.cpp (file contents), Revision 1782 by gezelter, Wed Aug 22 02:28:28 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 53 | Line 54 | namespace OpenMD {
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_.push_back( Vector3i(-1, 0, 0) );
57 <    cellOffsets_.push_back( Vector3i(-1,-1, 0) );
58 <    cellOffsets_.push_back( Vector3i( 0,-1, 0) );
59 <    cellOffsets_.push_back( Vector3i( 1,-1, 0) );
60 <    cellOffsets_.push_back( Vector3i( 0, 0,-1) );
61 <    cellOffsets_.push_back( Vector3i(-1, 0, 1) );
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) );
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) );
66 –    cellOffsets_.push_back( Vector3i( 1, 1,-1) );
67 –    cellOffsets_.push_back( Vector3i( 0, 1,-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  
# Line 79 | Line 95 | namespace OpenMD {
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 93 | 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      MPI::Intracomm row = rowComm.getComm();
# Line 129 | 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      
# Line 149 | Line 176 | namespace OpenMD {
176      pot_row.resize(nAtomsInRow_);
177      pot_col.resize(nAtomsInCol_);
178  
179 +    expot_row.resize(nAtomsInRow_);
180 +    expot_col.resize(nAtomsInCol_);
181 +
182      AtomRowToGlobal.resize(nAtomsInRow_);
183      AtomColToGlobal.resize(nAtomsInCol_);
184      AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
185      AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
156 –
157 –    cerr << "Atoms in Local:\n";
158 –    for (int i = 0; i < AtomLocalToGlobal.size(); i++) {
159 –      cerr << "i =\t" << i << "\t localAt =\t" << AtomLocalToGlobal[i] << "\n";
160 –    }
161 –    cerr << "Atoms in Row:\n";
162 –    for (int i = 0; i < AtomRowToGlobal.size(); i++) {
163 –      cerr << "i =\t" << i << "\t rowAt =\t" << AtomRowToGlobal[i] << "\n";
164 –    }
165 –    cerr << "Atoms in Col:\n";
166 –    for (int i = 0; i < AtomColToGlobal.size(); i++) {
167 –      cerr << "i =\t" << i << "\t colAt =\t" << AtomColToGlobal[i] << "\n";
168 –    }
186  
187      cgRowToGlobal.resize(nGroupsInRow_);
188      cgColToGlobal.resize(nGroupsInCol_);
189      cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
190      cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
191  
175 –    cerr << "Gruops in Local:\n";
176 –    for (int i = 0; i < cgLocalToGlobal.size(); i++) {
177 –      cerr << "i =\t" << i << "\t localCG =\t" << cgLocalToGlobal[i] << "\n";
178 –    }
179 –    cerr << "Groups in Row:\n";
180 –    for (int i = 0; i < cgRowToGlobal.size(); i++) {
181 –      cerr << "i =\t" << i << "\t rowCG =\t" << cgRowToGlobal[i] << "\n";
182 –    }
183 –    cerr << "Groups in Col:\n";
184 –    for (int i = 0; i < cgColToGlobal.size(); i++) {
185 –      cerr << "i =\t" << i << "\t colCG =\t" << cgColToGlobal[i] << "\n";
186 –    }
187 –
188 –
192      massFactorsRow.resize(nAtomsInRow_);
193      massFactorsCol.resize(nAtomsInCol_);
194      AtomPlanRealRow->gather(massFactors, massFactorsRow);
# Line 245 | Line 248 | namespace OpenMD {
248        }      
249      }
250  
251 < #endif
249 <
250 <    // allocate memory for the parallel objects
251 <    atypesLocal.resize(nLocal_);
252 <
253 <    for (int i = 0; i < nLocal_; i++)
254 <      atypesLocal[i] = ff_->getAtomType(idents[i]);
255 <
256 <    groupList_.clear();
257 <    groupList_.resize(nGroups_);
258 <    for (int i = 0; i < nGroups_; i++) {
259 <      int gid = cgLocalToGlobal[i];
260 <      for (int j = 0; j < nLocal_; j++) {
261 <        int aid = AtomLocalToGlobal[j];
262 <        if (globalGroupMembership[aid] == gid) {
263 <          groupList_[i].push_back(j);
264 <        }
265 <      }      
266 <    }
267 <
251 > #else
252      excludesForAtom.clear();
253      excludesForAtom.resize(nLocal_);
254      toposForAtom.clear();
# Line 297 | Line 281 | namespace OpenMD {
281          }
282        }      
283      }
284 <    
284 > #endif
285 >
286 >    // allocate memory for the parallel objects
287 >    atypesLocal.resize(nLocal_);
288 >
289 >    for (int i = 0; i < nLocal_; i++)
290 >      atypesLocal[i] = ff_->getAtomType(idents[i]);
291 >
292 >    groupList_.clear();
293 >    groupList_.resize(nGroups_);
294 >    for (int i = 0; i < nGroups_; i++) {
295 >      int gid = cgLocalToGlobal[i];
296 >      for (int j = 0; j < nLocal_; j++) {
297 >        int aid = AtomLocalToGlobal[j];
298 >        if (globalGroupMembership[aid] == gid) {
299 >          groupList_[i].push_back(j);
300 >        }
301 >      }      
302 >    }
303 >
304 >
305      createGtypeCutoffMap();
306  
307    }
# Line 306 | Line 310 | namespace OpenMD {
310      
311      RealType tol = 1e-6;
312      largestRcut_ = 0.0;
309 –    RealType rc;
313      int atid;
314      set<AtomType*> atypes = info_->getSimulatedAtomTypes();
315      
# Line 391 | Line 394 | namespace OpenMD {
394        }
395        
396        bool gTypeFound = false;
397 <      for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
397 >      for (unsigned int gt = 0; gt < gTypeCutoffs.size(); gt++) {
398          if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
399            groupToGtype[cg1] = gt;
400            gTypeFound = true;
# Line 416 | Line 419 | namespace OpenMD {
419      
420      RealType tradRcut = groupMax;
421  
422 <    for (int i = 0; i < gTypeCutoffs.size();  i++) {
423 <      for (int j = 0; j < gTypeCutoffs.size();  j++) {      
422 >    for (unsigned int i = 0; i < gTypeCutoffs.size();  i++) {
423 >      for (unsigned int j = 0; j < gTypeCutoffs.size();  j++) {      
424          RealType thisRcut;
425          switch(cutoffPolicy_) {
426          case TRADITIONAL:
# Line 460 | Line 463 | namespace OpenMD {
463      }
464    }
465  
463 –
466    groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
467      int i, j;  
468   #ifdef IS_MPI
# Line 474 | Line 476 | namespace OpenMD {
476    }
477  
478    int ForceMatrixDecomposition::getTopologicalDistance(int atom1, int atom2) {
479 <    for (int j = 0; j < toposForAtom[atom1].size(); j++) {
479 >    for (unsigned int j = 0; j < toposForAtom[atom1].size(); j++) {
480        if (toposForAtom[atom1][j] == atom2)
481          return topoDist[atom1][j];
482      }
# Line 484 | Line 486 | namespace OpenMD {
486    void ForceMatrixDecomposition::zeroWorkArrays() {
487      pairwisePot = 0.0;
488      embeddingPot = 0.0;
489 +    excludedPot = 0.0;
490 +    excludedSelfPot = 0.0;
491  
492   #ifdef IS_MPI
493      if (storageLayout_ & DataStorage::dslForce) {
# Line 502 | Line 506 | namespace OpenMD {
506      fill(pot_col.begin(), pot_col.end(),
507           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));  
508  
509 +    fill(expot_row.begin(), expot_row.end(),
510 +         Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
511 +
512 +    fill(expot_col.begin(), expot_col.end(),
513 +         Vector<RealType, N_INTERACTION_FAMILIES> (0.0));  
514 +
515      if (storageLayout_ & DataStorage::dslParticlePot) {    
516        fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(),
517             0.0);
# Line 535 | Line 545 | namespace OpenMD {
545             atomColData.skippedCharge.end(), 0.0);
546      }
547  
548 +    if (storageLayout_ & DataStorage::dslFlucQForce) {      
549 +      fill(atomRowData.flucQFrc.begin(),
550 +           atomRowData.flucQFrc.end(), 0.0);
551 +      fill(atomColData.flucQFrc.begin(),
552 +           atomColData.flucQFrc.end(), 0.0);
553 +    }
554 +
555 +    if (storageLayout_ & DataStorage::dslElectricField) {    
556 +      fill(atomRowData.electricField.begin(),
557 +           atomRowData.electricField.end(), V3Zero);
558 +      fill(atomColData.electricField.begin(),
559 +           atomColData.electricField.end(), V3Zero);
560 +    }
561 +
562 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
563 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
564 +           0.0);
565 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
566 +           0.0);
567 +    }
568 +
569   #endif
570      // even in parallel, we need to zero out the local arrays:
571  
# Line 547 | Line 578 | namespace OpenMD {
578        fill(snap_->atomData.density.begin(),
579             snap_->atomData.density.end(), 0.0);
580      }
581 +
582      if (storageLayout_ & DataStorage::dslFunctional) {
583        fill(snap_->atomData.functional.begin(),
584             snap_->atomData.functional.end(), 0.0);
585      }
586 +
587      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
588        fill(snap_->atomData.functionalDerivative.begin(),
589             snap_->atomData.functionalDerivative.end(), 0.0);
590      }
591 +
592      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
593        fill(snap_->atomData.skippedCharge.begin(),
594             snap_->atomData.skippedCharge.end(), 0.0);
595      }
596 <    
596 >
597 >    if (storageLayout_ & DataStorage::dslElectricField) {      
598 >      fill(snap_->atomData.electricField.begin(),
599 >           snap_->atomData.electricField.end(), V3Zero);
600 >    }
601    }
602  
603  
# Line 576 | Line 614 | namespace OpenMD {
614      
615      // gather up the cutoff group positions
616  
579 –    cerr  << "before gather\n";
580 –    for (int i = 0; i < snap_->cgData.position.size(); i++) {
581 –      cerr << "cgpos = " << snap_->cgData.position[i] << "\n";
582 –    }
583 –
617      cgPlanVectorRow->gather(snap_->cgData.position,
618                              cgRowData.position);
619  
587 –    cerr  << "after gather\n";
588 –    for (int i = 0; i < cgRowData.position.size(); i++) {
589 –      cerr << "cgRpos = " << cgRowData.position[i] << "\n";
590 –    }
591 –
620      cgPlanVectorColumn->gather(snap_->cgData.position,
621                                 cgColData.position);
622 <    for (int i = 0; i < cgColData.position.size(); i++) {
623 <      cerr << "cgCpos = " << cgColData.position[i] << "\n";
622 >
623 >
624 >
625 >    if (needVelocities_) {
626 >      // gather up the atomic velocities
627 >      AtomPlanVectorColumn->gather(snap_->atomData.velocity,
628 >                                   atomColData.velocity);
629 >      
630 >      cgPlanVectorColumn->gather(snap_->cgData.velocity,
631 >                                 cgColData.velocity);
632      }
633  
634      
# Line 612 | Line 648 | namespace OpenMD {
648                                     atomColData.electroFrame);
649      }
650  
651 +    // if needed, gather the atomic fluctuating charge values
652 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
653 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
654 +                              atomRowData.flucQPos);
655 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
656 +                                 atomColData.flucQPos);
657 +    }
658 +
659   #endif      
660    }
661    
# Line 634 | Line 678 | namespace OpenMD {
678        for (int i = 0; i < n; i++)
679          snap_->atomData.density[i] += rho_tmp[i];
680      }
681 +
682 +    if (storageLayout_ & DataStorage::dslElectricField) {
683 +      
684 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
685 +                                 snap_->atomData.electricField);
686 +      
687 +      int n = snap_->atomData.electricField.size();
688 +      vector<Vector3d> field_tmp(n, V3Zero);
689 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
690 +      for (int i = 0; i < n; i++)
691 +        snap_->atomData.electricField[i] += field_tmp[i];
692 +    }
693   #endif
694    }
695  
# Line 708 | Line 764 | namespace OpenMD {
764        }
765        
766        AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
767 <      for (int i = 0; i < ns; i++)
767 >      for (int i = 0; i < ns; i++)
768          snap_->atomData.skippedCharge[i] += skch_tmp[i];
769 +            
770      }
771      
772 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
773 +
774 +      int nq = snap_->atomData.flucQFrc.size();
775 +      vector<RealType> fqfrc_tmp(nq, 0.0);
776 +
777 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
778 +      for (int i = 0; i < nq; i++) {
779 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
780 +        fqfrc_tmp[i] = 0.0;
781 +      }
782 +      
783 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
784 +      for (int i = 0; i < nq; i++)
785 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
786 +            
787 +    }
788 +
789      nLocal_ = snap_->getNumberOfAtoms();
790  
791      vector<potVec> pot_temp(nLocal_,
792                              Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
793 +    vector<potVec> expot_temp(nLocal_,
794 +                              Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
795  
796      // scatter/gather pot_row into the members of my column
797            
798      AtomPlanPotRow->scatter(pot_row, pot_temp);
799 +    AtomPlanPotRow->scatter(expot_row, expot_temp);
800  
801 <    for (int ii = 0;  ii < pot_temp.size(); ii++ )
801 >    for (int ii = 0;  ii < pot_temp.size(); ii++ )
802        pairwisePot += pot_temp[ii];
803 <    
803 >
804 >    for (int ii = 0;  ii < expot_temp.size(); ii++ )
805 >      excludedPot += expot_temp[ii];
806 >        
807 >    if (storageLayout_ & DataStorage::dslParticlePot) {
808 >      // This is the pairwise contribution to the particle pot.  The
809 >      // embedding contribution is added in each of the low level
810 >      // non-bonded routines.  In single processor, this is done in
811 >      // unpackInteractionData, not in collectData.
812 >      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
813 >        for (int i = 0; i < nLocal_; i++) {
814 >          // factor of two is because the total potential terms are divided
815 >          // by 2 in parallel due to row/ column scatter      
816 >          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
817 >        }
818 >      }
819 >    }
820 >
821      fill(pot_temp.begin(), pot_temp.end(),
822           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
823 +    fill(expot_temp.begin(), expot_temp.end(),
824 +         Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
825        
826      AtomPlanPotColumn->scatter(pot_col, pot_temp);    
827 +    AtomPlanPotColumn->scatter(expot_col, expot_temp);    
828      
829      for (int ii = 0;  ii < pot_temp.size(); ii++ )
830        pairwisePot += pot_temp[ii];    
734 – #endif
831  
832 <    cerr << "pairwisePot = " <<  pairwisePot << "\n";
832 >    for (int ii = 0;  ii < expot_temp.size(); ii++ )
833 >      excludedPot += expot_temp[ii];    
834 >
835 >    if (storageLayout_ & DataStorage::dslParticlePot) {
836 >      // This is the pairwise contribution to the particle pot.  The
837 >      // embedding contribution is added in each of the low level
838 >      // non-bonded routines.  In single processor, this is done in
839 >      // unpackInteractionData, not in collectData.
840 >      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
841 >        for (int i = 0; i < nLocal_; i++) {
842 >          // factor of two is because the total potential terms are divided
843 >          // by 2 in parallel due to row/ column scatter      
844 >          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
845 >        }
846 >      }
847 >    }
848 >    
849 >    if (storageLayout_ & DataStorage::dslParticlePot) {
850 >      int npp = snap_->atomData.particlePot.size();
851 >      vector<RealType> ppot_temp(npp, 0.0);
852 >
853 >      // This is the direct or embedding contribution to the particle
854 >      // pot.
855 >      
856 >      AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
857 >      for (int i = 0; i < npp; i++) {
858 >        snap_->atomData.particlePot[i] += ppot_temp[i];
859 >      }
860 >
861 >      fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
862 >      
863 >      AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
864 >      for (int i = 0; i < npp; i++) {
865 >        snap_->atomData.particlePot[i] += ppot_temp[i];
866 >      }
867 >    }
868 >
869 >    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
870 >      RealType ploc1 = pairwisePot[ii];
871 >      RealType ploc2 = 0.0;
872 >      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
873 >      pairwisePot[ii] = ploc2;
874 >    }
875 >
876 >    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
877 >      RealType ploc1 = excludedPot[ii];
878 >      RealType ploc2 = 0.0;
879 >      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
880 >      excludedPot[ii] = ploc2;
881 >    }
882 >
883 >    // Here be dragons.
884 >    MPI::Intracomm col = colComm.getComm();
885 >
886 >    col.Allreduce(MPI::IN_PLACE,
887 >                  &snap_->frameData.conductiveHeatFlux[0], 3,
888 >                  MPI::REALTYPE, MPI::SUM);
889 >
890 >
891 > #endif
892 >
893    }
894  
895 +  /**
896 +   * Collects information obtained during the post-pair (and embedding
897 +   * functional) loops onto local data structures.
898 +   */
899 +  void ForceMatrixDecomposition::collectSelfData() {
900 +    snap_ = sman_->getCurrentSnapshot();
901 +    storageLayout_ = sman_->getStorageLayout();
902 +
903 + #ifdef IS_MPI
904 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
905 +      RealType ploc1 = embeddingPot[ii];
906 +      RealType ploc2 = 0.0;
907 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
908 +      embeddingPot[ii] = ploc2;
909 +    }    
910 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
911 +      RealType ploc1 = excludedSelfPot[ii];
912 +      RealType ploc2 = 0.0;
913 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
914 +      excludedSelfPot[ii] = ploc2;
915 +    }    
916 + #endif
917 +    
918 +  }
919 +
920 +
921 +
922    int ForceMatrixDecomposition::getNAtomsInRow() {  
923   #ifdef IS_MPI
924      return nAtomsInRow_;
# Line 768 | Line 951 | namespace OpenMD {
951      
952   #ifdef IS_MPI
953      d = cgColData.position[cg2] - cgRowData.position[cg1];
771 –    cerr << "cg1 = " << cg1 << "\tcg1p = " << cgRowData.position[cg1] << "\n";
772 –    cerr << "cg2 = " << cg2 << "\tcg2p = " << cgColData.position[cg2] << "\n";
954   #else
955      d = snap_->cgData.position[cg2] - snap_->cgData.position[cg1];
775 –    cerr << "cg1 = " << cg1 << "\tcg1p = " << snap_->cgData.position[cg1] << "\n";
776 –    cerr << "cg2 = " << cg2 << "\tcg2p = " << snap_->cgData.position[cg2] << "\n";
956   #endif
957      
958      snap_->wrapVector(d);
959      return d;    
960    }
961  
962 +  Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
963 + #ifdef IS_MPI
964 +    return cgColData.velocity[cg2];
965 + #else
966 +    return snap_->cgData.velocity[cg2];
967 + #endif
968 +  }
969  
970 +  Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
971 + #ifdef IS_MPI
972 +    return atomColData.velocity[atom2];
973 + #else
974 +    return snap_->atomData.velocity[atom2];
975 + #endif
976 +  }
977 +
978 +
979    Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
980  
981      Vector3d d;
# Line 846 | Line 1041 | namespace OpenMD {
1041     * We need to exclude some overcounted interactions that result from
1042     * the parallel decomposition.
1043     */
1044 <  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
1045 <    int unique_id_1, unique_id_2;
1046 <    
852 <
853 <    cerr << "sap with atom1, atom2 =\t" << atom1 << "\t" << atom2 << "\n";
1044 >  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) {
1045 >    int unique_id_1, unique_id_2, group1, group2;
1046 >        
1047   #ifdef IS_MPI
1048      // in MPI, we have to look up the unique IDs for each atom
1049      unique_id_1 = AtomRowToGlobal[atom1];
1050      unique_id_2 = AtomColToGlobal[atom2];
1051 +    group1 = cgRowToGlobal[cg1];
1052 +    group2 = cgColToGlobal[cg2];
1053 + #else
1054 +    unique_id_1 = AtomLocalToGlobal[atom1];
1055 +    unique_id_2 = AtomLocalToGlobal[atom2];
1056 +    group1 = cgLocalToGlobal[cg1];
1057 +    group2 = cgLocalToGlobal[cg2];
1058 + #endif  
1059  
859 –    cerr << "sap with uid1, uid2 =\t" << unique_id_1 << "\t" << unique_id_2 << "\n";
860 –    // this situation should only arise in MPI simulations
1060      if (unique_id_1 == unique_id_2) return true;
1061 <    
1061 >
1062 > #ifdef IS_MPI
1063      // this prevents us from doing the pair on multiple processors
1064      if (unique_id_1 < unique_id_2) {
1065        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1066      } else {
1067 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1067 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1068 >    }
1069 > #endif    
1070 >
1071 > #ifndef IS_MPI
1072 >    if (group1 == group2) {
1073 >      if (unique_id_1 < unique_id_2) return true;
1074      }
1075   #endif
1076 +    
1077      return false;
1078    }
1079  
# Line 880 | Line 1087 | namespace OpenMD {
1087     * field) must still be handled for these pairs.
1088     */
1089    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1090 <    int unique_id_2;
1091 < #ifdef IS_MPI
1092 <    // in MPI, we have to look up the unique IDs for the row atom.
886 <    unique_id_2 = AtomColToGlobal[atom2];
887 < #else
888 <    // in the normal loop, the atom numbers are unique
889 <    unique_id_2 = atom2;
890 < #endif
1090 >
1091 >    // excludesForAtom was constructed to use row/column indices in the MPI
1092 >    // version, and to use local IDs in the non-MPI version:
1093      
1094      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1095           i != excludesForAtom[atom1].end(); ++i) {
1096 <      if ( (*i) == unique_id_2 ) return true;
1096 >      if ( (*i) == atom2 ) return true;
1097      }
1098  
1099      return false;
# Line 965 | Line 1167 | namespace OpenMD {
1167        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1168      }
1169  
1170 < #else
1170 >    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1171 >      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1172 >      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1173 >    }
1174  
1175 + #else
1176 +    
1177      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
971 –    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
972 –    //                         ff_->getAtomType(idents[atom2]) );
1178  
1179      if (storageLayout_ & DataStorage::dslAmat) {
1180        idat.A1 = &(snap_->atomData.aMat[atom1]);
# Line 1010 | Line 1215 | namespace OpenMD {
1215        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1216        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1217      }
1218 +
1219 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1220 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1221 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1222 +    }
1223 +
1224   #endif
1225    }
1226  
1227    
1228    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1229   #ifdef IS_MPI
1230 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1231 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1230 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1231 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1232 >    expot_row[atom1] += RealType(0.5) *  *(idat.excludedPot);
1233 >    expot_col[atom2] += RealType(0.5) *  *(idat.excludedPot);
1234  
1235      atomRowData.force[atom1] += *(idat.f1);
1236      atomColData.force[atom2] -= *(idat.f1);
1237 +
1238 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1239 +      atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1240 +      atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1241 +    }
1242 +
1243 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1244 +      atomRowData.electricField[atom1] += *(idat.eField1);
1245 +      atomColData.electricField[atom2] += *(idat.eField2);
1246 +    }
1247 +
1248   #else
1249      pairwisePot += *(idat.pot);
1250 +    excludedPot += *(idat.excludedPot);
1251  
1252      snap_->atomData.force[atom1] += *(idat.f1);
1253      snap_->atomData.force[atom2] -= *(idat.f1);
1254 +
1255 +    if (idat.doParticlePot) {
1256 +      // This is the pairwise contribution to the particle pot.  The
1257 +      // embedding contribution is added in each of the low level
1258 +      // non-bonded routines.  In parallel, this calculation is done
1259 +      // in collectData, not in unpackInteractionData.
1260 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1261 +      snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1262 +    }
1263 +    
1264 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1265 +      snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1266 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1267 +    }
1268 +
1269 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1270 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1271 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1272 +    }
1273 +
1274   #endif
1275      
1276    }
# Line 1094 | Line 1339 | namespace OpenMD {
1339          for (int j = 0; j < 3; j++) {
1340            scaled[j] -= roundMe(scaled[j]);
1341            scaled[j] += 0.5;
1342 +          // Handle the special case when an object is exactly on the
1343 +          // boundary (a scaled coordinate of 1.0 is the same as
1344 +          // scaled coordinate of 0.0)
1345 +          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
1346          }
1347          
1348          // find xyz-indices of cell that cutoffGroup is in.
# Line 1118 | Line 1367 | namespace OpenMD {
1367          for (int j = 0; j < 3; j++) {
1368            scaled[j] -= roundMe(scaled[j]);
1369            scaled[j] += 0.5;
1370 +          // Handle the special case when an object is exactly on the
1371 +          // boundary (a scaled coordinate of 1.0 is the same as
1372 +          // scaled coordinate of 0.0)
1373 +          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
1374          }
1375          
1376          // find xyz-indices of cell that cutoffGroup is in.
# Line 1131 | Line 1384 | namespace OpenMD {
1384          // add this cutoff group to the list of groups in this cell;
1385          cellListCol_[cellIndex].push_back(i);
1386        }
1387 +    
1388   #else
1389        for (int i = 0; i < nGroups_; i++) {
1390          rs = snap_->cgData.position[i];
# Line 1143 | Line 1397 | namespace OpenMD {
1397          for (int j = 0; j < 3; j++) {
1398            scaled[j] -= roundMe(scaled[j]);
1399            scaled[j] += 0.5;
1400 +          // Handle the special case when an object is exactly on the
1401 +          // boundary (a scaled coordinate of 1.0 is the same as
1402 +          // scaled coordinate of 0.0)
1403 +          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
1404          }
1405          
1406          // find xyz-indices of cell that cutoffGroup is in.
# Line 1156 | Line 1414 | namespace OpenMD {
1414          // add this cutoff group to the list of groups in this cell;
1415          cellList_[cellIndex].push_back(i);
1416        }
1417 +
1418   #endif
1419  
1420        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1168 | Line 1427 | namespace OpenMD {
1427                   os != cellOffsets_.end(); ++os) {
1428                
1429                Vector3i m2v = m1v + (*os);
1430 <              
1430 >            
1431 >
1432                if (m2v.x() >= nCells_.x()) {
1433                  m2v.x() = 0;          
1434                } else if (m2v.x() < 0) {
# Line 1186 | Line 1446 | namespace OpenMD {
1446                } else if (m2v.z() < 0) {
1447                  m2v.z() = nCells_.z() - 1;
1448                }
1449 <              
1449 >
1450                int m2 = Vlinear (m2v, nCells_);
1451                
1452   #ifdef IS_MPI
# Line 1195 | Line 1455 | namespace OpenMD {
1455                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1456                       j2 != cellListCol_[m2].end(); ++j2) {
1457                    
1458 <                  // In parallel, we need to visit *all* pairs of row &
1459 <                  // column indicies and will truncate later on.
1458 >                  // In parallel, we need to visit *all* pairs of row
1459 >                  // & column indicies and will divide labor in the
1460 >                  // force evaluation later.
1461                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1462                    snap_->wrapVector(dr);
1463                    cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1206 | Line 1467 | namespace OpenMD {
1467                  }
1468                }
1469   #else
1209 –              
1470                for (vector<int>::iterator j1 = cellList_[m1].begin();
1471                     j1 != cellList_[m1].end(); ++j1) {
1472                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1473                       j2 != cellList_[m2].end(); ++j2) {
1474 <                  
1474 >    
1475                    // Always do this if we're in different cells or if
1476 <                  // we're in the same cell and the global index of the
1477 <                  // j2 cutoff group is less than the j1 cutoff group
1478 <                  
1479 <                  if (m2 != m1 || (*j2) < (*j1)) {
1476 >                  // we're in the same cell and the global index of
1477 >                  // the j2 cutoff group is greater than or equal to
1478 >                  // the j1 cutoff group.  Note that Rappaport's code
1479 >                  // has a "less than" conditional here, but that
1480 >                  // deals with atom-by-atom computation.  OpenMD
1481 >                  // allows atoms within a single cutoff group to
1482 >                  // interact with each other.
1483 >
1484 >
1485 >
1486 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1487 >
1488                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1489                      snap_->wrapVector(dr);
1490                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1235 | Line 1503 | namespace OpenMD {
1503        // branch to do all cutoff group pairs
1504   #ifdef IS_MPI
1505        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1506 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1506 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1507            dr = cgColData.position[j2] - cgRowData.position[j1];
1508            snap_->wrapVector(dr);
1509            cuts = getGroupCutoffs( j1, j2 );
# Line 1243 | Line 1511 | namespace OpenMD {
1511              neighborList.push_back(make_pair(j1, j2));
1512            }
1513          }
1514 <      }
1514 >      }      
1515   #else
1516 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1517 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1516 >      // include all groups here.
1517 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1518 >        // include self group interactions j2 == j1
1519 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1520            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1521            snap_->wrapVector(dr);
1522            cuts = getGroupCutoffs( j1, j2 );
1523            if (dr.lengthSquare() < cuts.third) {
1524              neighborList.push_back(make_pair(j1, j2));
1525            }
1526 <        }
1527 <      }        
1526 >        }    
1527 >      }
1528   #endif
1529      }
1530        

Diff Legend

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