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 1613 by gezelter, Thu Aug 18 20:18:19 2011 UTC vs.
Revision 1756 by gezelter, Mon Jun 18 18:23:20 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 94 | 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 108 | 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 144 | 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 247 | Line 259 | namespace OpenMD {
259        for (int j = 0; j < nLocal_; j++) {
260          int jglob = AtomLocalToGlobal[j];
261  
262 <        if (excludes->hasPair(iglob, jglob))          
262 >        if (excludes->hasPair(iglob, jglob))
263            excludesForAtom[i].push_back(j);              
264          
253 –        
265          if (oneTwo->hasPair(iglob, jglob)) {
266            toposForAtom[i].push_back(j);
267            topoDist[i].push_back(1);
# Line 450 | Line 461 | namespace OpenMD {
461      }
462    }
463  
453 –
464    groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
465      int i, j;  
466   #ifdef IS_MPI
# Line 525 | Line 535 | namespace OpenMD {
535             atomColData.skippedCharge.end(), 0.0);
536      }
537  
538 +    if (storageLayout_ & DataStorage::dslFlucQForce) {      
539 +      fill(atomRowData.flucQFrc.begin(),
540 +           atomRowData.flucQFrc.end(), 0.0);
541 +      fill(atomColData.flucQFrc.begin(),
542 +           atomColData.flucQFrc.end(), 0.0);
543 +    }
544 +
545 +    if (storageLayout_ & DataStorage::dslElectricField) {    
546 +      fill(atomRowData.electricField.begin(),
547 +           atomRowData.electricField.end(), V3Zero);
548 +      fill(atomColData.electricField.begin(),
549 +           atomColData.electricField.end(), V3Zero);
550 +    }
551 +
552 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
553 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
554 +           0.0);
555 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
556 +           0.0);
557 +    }
558 +
559   #endif
560      // even in parallel, we need to zero out the local arrays:
561  
# Line 537 | Line 568 | namespace OpenMD {
568        fill(snap_->atomData.density.begin(),
569             snap_->atomData.density.end(), 0.0);
570      }
571 +
572      if (storageLayout_ & DataStorage::dslFunctional) {
573        fill(snap_->atomData.functional.begin(),
574             snap_->atomData.functional.end(), 0.0);
575      }
576 +
577      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
578        fill(snap_->atomData.functionalDerivative.begin(),
579             snap_->atomData.functionalDerivative.end(), 0.0);
580      }
581 +
582      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
583        fill(snap_->atomData.skippedCharge.begin(),
584             snap_->atomData.skippedCharge.end(), 0.0);
585      }
586 <    
586 >
587 >    if (storageLayout_ & DataStorage::dslElectricField) {      
588 >      fill(snap_->atomData.electricField.begin(),
589 >           snap_->atomData.electricField.end(), V3Zero);
590 >    }
591    }
592  
593  
# Line 572 | Line 610 | namespace OpenMD {
610      cgPlanVectorColumn->gather(snap_->cgData.position,
611                                 cgColData.position);
612  
613 +
614 +
615 +    if (needVelocities_) {
616 +      // gather up the atomic velocities
617 +      AtomPlanVectorColumn->gather(snap_->atomData.velocity,
618 +                                   atomColData.velocity);
619 +      
620 +      cgPlanVectorColumn->gather(snap_->cgData.velocity,
621 +                                 cgColData.velocity);
622 +    }
623 +
624      
625      // if needed, gather the atomic rotation matrices
626      if (storageLayout_ & DataStorage::dslAmat) {
# Line 589 | Line 638 | namespace OpenMD {
638                                     atomColData.electroFrame);
639      }
640  
641 +    // if needed, gather the atomic fluctuating charge values
642 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
643 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
644 +                              atomRowData.flucQPos);
645 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
646 +                                 atomColData.flucQPos);
647 +    }
648 +
649   #endif      
650    }
651    
# Line 610 | Line 667 | namespace OpenMD {
667        AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
668        for (int i = 0; i < n; i++)
669          snap_->atomData.density[i] += rho_tmp[i];
670 +    }
671 +
672 +    if (storageLayout_ & DataStorage::dslElectricField) {
673 +      
674 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
675 +                                 snap_->atomData.electricField);
676 +      
677 +      int n = snap_->atomData.electricField.size();
678 +      vector<Vector3d> field_tmp(n, V3Zero);
679 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
680 +      for (int i = 0; i < n; i++)
681 +        snap_->atomData.electricField[i] += field_tmp[i];
682      }
683   #endif
684    }
# Line 690 | Line 759 | namespace OpenMD {
759              
760      }
761      
762 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
763 +
764 +      int nq = snap_->atomData.flucQFrc.size();
765 +      vector<RealType> fqfrc_tmp(nq, 0.0);
766 +
767 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
768 +      for (int i = 0; i < nq; i++) {
769 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
770 +        fqfrc_tmp[i] = 0.0;
771 +      }
772 +      
773 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
774 +      for (int i = 0; i < nq; i++)
775 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
776 +            
777 +    }
778 +
779      nLocal_ = snap_->getNumberOfAtoms();
780  
781      vector<potVec> pot_temp(nLocal_,
# Line 701 | Line 787 | namespace OpenMD {
787  
788      for (int ii = 0;  ii < pot_temp.size(); ii++ )
789        pairwisePot += pot_temp[ii];
790 <    
790 >        
791 >    if (storageLayout_ & DataStorage::dslParticlePot) {
792 >      // This is the pairwise contribution to the particle pot.  The
793 >      // embedding contribution is added in each of the low level
794 >      // non-bonded routines.  In single processor, this is done in
795 >      // unpackInteractionData, not in collectData.
796 >      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
797 >        for (int i = 0; i < nLocal_; i++) {
798 >          // factor of two is because the total potential terms are divided
799 >          // by 2 in parallel due to row/ column scatter      
800 >          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
801 >        }
802 >      }
803 >    }
804 >
805      fill(pot_temp.begin(), pot_temp.end(),
806           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
807        
# Line 709 | Line 809 | namespace OpenMD {
809      
810      for (int ii = 0;  ii < pot_temp.size(); ii++ )
811        pairwisePot += pot_temp[ii];    
812 +
813 +    if (storageLayout_ & DataStorage::dslParticlePot) {
814 +      // This is the pairwise contribution to the particle pot.  The
815 +      // embedding contribution is added in each of the low level
816 +      // non-bonded routines.  In single processor, this is done in
817 +      // unpackInteractionData, not in collectData.
818 +      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
819 +        for (int i = 0; i < nLocal_; i++) {
820 +          // factor of two is because the total potential terms are divided
821 +          // by 2 in parallel due to row/ column scatter      
822 +          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
823 +        }
824 +      }
825 +    }
826      
827 +    if (storageLayout_ & DataStorage::dslParticlePot) {
828 +      int npp = snap_->atomData.particlePot.size();
829 +      vector<RealType> ppot_temp(npp, 0.0);
830 +
831 +      // This is the direct or embedding contribution to the particle
832 +      // pot.
833 +      
834 +      AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
835 +      for (int i = 0; i < npp; i++) {
836 +        snap_->atomData.particlePot[i] += ppot_temp[i];
837 +      }
838 +
839 +      fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
840 +      
841 +      AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
842 +      for (int i = 0; i < npp; i++) {
843 +        snap_->atomData.particlePot[i] += ppot_temp[i];
844 +      }
845 +    }
846 +
847      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
848        RealType ploc1 = pairwisePot[ii];
849        RealType ploc2 = 0.0;
850        MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
851        pairwisePot[ii] = ploc2;
852      }
853 +
854 +    // Here be dragons.
855 +    MPI::Intracomm col = colComm.getComm();
856 +
857 +    col.Allreduce(MPI::IN_PLACE,
858 +                  &snap_->frameData.conductiveHeatFlux[0], 3,
859 +                  MPI::REALTYPE, MPI::SUM);
860  
861 +
862 + #endif
863 +
864 +  }
865 +
866 +  /**
867 +   * Collects information obtained during the post-pair (and embedding
868 +   * functional) loops onto local data structures.
869 +   */
870 +  void ForceMatrixDecomposition::collectSelfData() {
871 +    snap_ = sman_->getCurrentSnapshot();
872 +    storageLayout_ = sman_->getStorageLayout();
873 +
874 + #ifdef IS_MPI
875      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
876        RealType ploc1 = embeddingPot[ii];
877        RealType ploc2 = 0.0;
878        MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
879        embeddingPot[ii] = ploc2;
880 <    }
726 <
880 >    }    
881   #endif
882 <
882 >    
883    }
884  
885 +
886 +
887    int ForceMatrixDecomposition::getNAtomsInRow() {  
888   #ifdef IS_MPI
889      return nAtomsInRow_;
# Line 768 | Line 924 | namespace OpenMD {
924      return d;    
925    }
926  
927 +  Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
928 + #ifdef IS_MPI
929 +    return cgColData.velocity[cg2];
930 + #else
931 +    return snap_->cgData.velocity[cg2];
932 + #endif
933 +  }
934  
935 +  Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
936 + #ifdef IS_MPI
937 +    return atomColData.velocity[atom2];
938 + #else
939 +    return snap_->atomData.velocity[atom2];
940 + #endif
941 +  }
942 +
943 +
944    Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
945  
946      Vector3d d;
# Line 834 | Line 1006 | namespace OpenMD {
1006     * We need to exclude some overcounted interactions that result from
1007     * the parallel decomposition.
1008     */
1009 <  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
1010 <    int unique_id_1, unique_id_2;
1011 <    
1009 >  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) {
1010 >    int unique_id_1, unique_id_2, group1, group2;
1011 >        
1012   #ifdef IS_MPI
1013      // in MPI, we have to look up the unique IDs for each atom
1014      unique_id_1 = AtomRowToGlobal[atom1];
1015      unique_id_2 = AtomColToGlobal[atom2];
1016 +    group1 = cgRowToGlobal[cg1];
1017 +    group2 = cgColToGlobal[cg2];
1018 + #else
1019 +    unique_id_1 = AtomLocalToGlobal[atom1];
1020 +    unique_id_2 = AtomLocalToGlobal[atom2];
1021 +    group1 = cgLocalToGlobal[cg1];
1022 +    group2 = cgLocalToGlobal[cg2];
1023 + #endif  
1024  
845 –    // this situation should only arise in MPI simulations
1025      if (unique_id_1 == unique_id_2) return true;
1026 <    
1026 >
1027 > #ifdef IS_MPI
1028      // this prevents us from doing the pair on multiple processors
1029      if (unique_id_1 < unique_id_2) {
1030        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1031      } else {
1032 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1032 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1033      }
1034 + #endif    
1035 +
1036 + #ifndef IS_MPI
1037 +    if (group1 == group2) {
1038 +      if (unique_id_1 < unique_id_2) return true;
1039 +    }
1040   #endif
1041 +    
1042      return false;
1043    }
1044  
# Line 871 | Line 1058 | namespace OpenMD {
1058      
1059      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1060           i != excludesForAtom[atom1].end(); ++i) {
1061 <      if ( (*i) == atom2 )  return true;
1061 >      if ( (*i) == atom2 ) return true;
1062      }
1063  
1064      return false;
# Line 945 | Line 1132 | namespace OpenMD {
1132        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1133      }
1134  
1135 < #else
1135 >    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1136 >      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1137 >      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1138 >    }
1139  
1140 + #else
1141 +    
1142      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
951 –    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
952 –    //                         ff_->getAtomType(idents[atom2]) );
1143  
1144      if (storageLayout_ & DataStorage::dslAmat) {
1145        idat.A1 = &(snap_->atomData.aMat[atom1]);
# Line 990 | Line 1180 | namespace OpenMD {
1180        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1181        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1182      }
1183 +
1184 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1185 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1186 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1187 +    }
1188 +
1189   #endif
1190    }
1191  
1192    
1193    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1194   #ifdef IS_MPI
1195 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1196 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1195 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1196 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1197  
1198      atomRowData.force[atom1] += *(idat.f1);
1199      atomColData.force[atom2] -= *(idat.f1);
1200 +
1201 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1202 +      atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1203 +      atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1204 +    }
1205 +
1206 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1207 +      atomRowData.electricField[atom1] += *(idat.eField1);
1208 +      atomColData.electricField[atom2] += *(idat.eField2);
1209 +    }
1210 +
1211   #else
1212      pairwisePot += *(idat.pot);
1213  
1214      snap_->atomData.force[atom1] += *(idat.f1);
1215      snap_->atomData.force[atom2] -= *(idat.f1);
1216 +
1217 +    if (idat.doParticlePot) {
1218 +      // This is the pairwise contribution to the particle pot.  The
1219 +      // embedding contribution is added in each of the low level
1220 +      // non-bonded routines.  In parallel, this calculation is done
1221 +      // in collectData, not in unpackInteractionData.
1222 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1223 +      snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1224 +    }
1225 +    
1226 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1227 +      snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1228 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1229 +    }
1230 +
1231 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1232 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1233 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1234 +    }
1235 +
1236   #endif
1237      
1238    }
# Line 1190 | Line 1417 | namespace OpenMD {
1417                  }
1418                }
1419   #else
1193 –              
1420                for (vector<int>::iterator j1 = cellList_[m1].begin();
1421                     j1 != cellList_[m1].end(); ++j1) {
1422                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1423                       j2 != cellList_[m2].end(); ++j2) {
1424 <                  
1424 >    
1425                    // Always do this if we're in different cells or if
1426 <                  // we're in the same cell and the global index of the
1427 <                  // j2 cutoff group is less than the j1 cutoff group
1428 <                  
1429 <                  if (m2 != m1 || (*j2) < (*j1)) {
1426 >                  // we're in the same cell and the global index of
1427 >                  // the j2 cutoff group is greater than or equal to
1428 >                  // the j1 cutoff group.  Note that Rappaport's code
1429 >                  // has a "less than" conditional here, but that
1430 >                  // deals with atom-by-atom computation.  OpenMD
1431 >                  // allows atoms within a single cutoff group to
1432 >                  // interact with each other.
1433 >
1434 >
1435 >
1436 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1437 >
1438                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1439                      snap_->wrapVector(dr);
1440                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1219 | Line 1453 | namespace OpenMD {
1453        // branch to do all cutoff group pairs
1454   #ifdef IS_MPI
1455        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1456 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1456 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1457            dr = cgColData.position[j2] - cgRowData.position[j1];
1458            snap_->wrapVector(dr);
1459            cuts = getGroupCutoffs( j1, j2 );
# Line 1227 | Line 1461 | namespace OpenMD {
1461              neighborList.push_back(make_pair(j1, j2));
1462            }
1463          }
1464 <      }
1464 >      }      
1465   #else
1466 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1467 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1466 >      // include all groups here.
1467 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1468 >        // include self group interactions j2 == j1
1469 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1470            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1471            snap_->wrapVector(dr);
1472            cuts = getGroupCutoffs( j1, j2 );
1473            if (dr.lengthSquare() < cuts.third) {
1474              neighborList.push_back(make_pair(j1, j2));
1475            }
1476 <        }
1477 <      }        
1476 >        }    
1477 >      }
1478   #endif
1479      }
1480        

Diff Legend

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