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 1612 by gezelter, Fri Aug 12 19:59:56 2011 UTC vs.
Revision 1755 by gezelter, Thu Jun 14 01:58:35 2012 UTC

# Line 36 | Line 36
36   * [1]  Meineke, et al., J. Comp. Chem. 26, 252-271 (2005).            
37   * [2]  Fennell & Gezelter, J. Chem. Phys. 124, 234104 (2006).          
38   * [3]  Sun, Lin & Gezelter, J. Chem. Phys. 128, 24107 (2008).          
39 < * [4]  Vardeman & Gezelter, in progress (2009).                        
39 > * [4]  Kuang & Gezelter,  J. Chem. Phys. 133, 164101 (2010).
40 > * [5]  Vardeman, Stocker & Gezelter, J. Chem. Theory Comput. 7, 834 (2011).
41   */
42   #include "parallel/ForceMatrixDecomposition.hpp"
43   #include "math/SquareMatrix3.hpp"
# Line 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 233 | Line 245 | namespace OpenMD {
245        }      
246      }
247  
248 < #endif
237 <
238 <    // allocate memory for the parallel objects
239 <    atypesLocal.resize(nLocal_);
240 <
241 <    for (int i = 0; i < nLocal_; i++)
242 <      atypesLocal[i] = ff_->getAtomType(idents[i]);
243 <
244 <    groupList_.clear();
245 <    groupList_.resize(nGroups_);
246 <    for (int i = 0; i < nGroups_; i++) {
247 <      int gid = cgLocalToGlobal[i];
248 <      for (int j = 0; j < nLocal_; j++) {
249 <        int aid = AtomLocalToGlobal[j];
250 <        if (globalGroupMembership[aid] == gid) {
251 <          groupList_[i].push_back(j);
252 <        }
253 <      }      
254 <    }
255 <
248 > #else
249      excludesForAtom.clear();
250      excludesForAtom.resize(nLocal_);
251      toposForAtom.clear();
# Line 285 | Line 278 | namespace OpenMD {
278          }
279        }      
280      }
281 <    
281 > #endif
282 >
283 >    // allocate memory for the parallel objects
284 >    atypesLocal.resize(nLocal_);
285 >
286 >    for (int i = 0; i < nLocal_; i++)
287 >      atypesLocal[i] = ff_->getAtomType(idents[i]);
288 >
289 >    groupList_.clear();
290 >    groupList_.resize(nGroups_);
291 >    for (int i = 0; i < nGroups_; i++) {
292 >      int gid = cgLocalToGlobal[i];
293 >      for (int j = 0; j < nLocal_; j++) {
294 >        int aid = AtomLocalToGlobal[j];
295 >        if (globalGroupMembership[aid] == gid) {
296 >          groupList_[i].push_back(j);
297 >        }
298 >      }      
299 >    }
300 >
301 >
302      createGtypeCutoffMap();
303  
304    }
# Line 523 | Line 536 | namespace OpenMD {
536             atomColData.skippedCharge.end(), 0.0);
537      }
538  
539 +    if (storageLayout_ & DataStorage::dslFlucQForce) {      
540 +      fill(atomRowData.flucQFrc.begin(),
541 +           atomRowData.flucQFrc.end(), 0.0);
542 +      fill(atomColData.flucQFrc.begin(),
543 +           atomColData.flucQFrc.end(), 0.0);
544 +    }
545 +
546 +    if (storageLayout_ & DataStorage::dslElectricField) {    
547 +      fill(atomRowData.electricField.begin(),
548 +           atomRowData.electricField.end(), V3Zero);
549 +      fill(atomColData.electricField.begin(),
550 +           atomColData.electricField.end(), V3Zero);
551 +    }
552 +
553 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
554 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
555 +           0.0);
556 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
557 +           0.0);
558 +    }
559 +
560   #endif
561      // even in parallel, we need to zero out the local arrays:
562  
# Line 535 | Line 569 | namespace OpenMD {
569        fill(snap_->atomData.density.begin(),
570             snap_->atomData.density.end(), 0.0);
571      }
572 +
573      if (storageLayout_ & DataStorage::dslFunctional) {
574        fill(snap_->atomData.functional.begin(),
575             snap_->atomData.functional.end(), 0.0);
576      }
577 +
578      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
579        fill(snap_->atomData.functionalDerivative.begin(),
580             snap_->atomData.functionalDerivative.end(), 0.0);
581      }
582 +
583      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
584        fill(snap_->atomData.skippedCharge.begin(),
585             snap_->atomData.skippedCharge.end(), 0.0);
586      }
587 <    
587 >
588 >    if (storageLayout_ & DataStorage::dslElectricField) {      
589 >      fill(snap_->atomData.electricField.begin(),
590 >           snap_->atomData.electricField.end(), V3Zero);
591 >    }
592    }
593  
594  
# Line 570 | Line 611 | namespace OpenMD {
611      cgPlanVectorColumn->gather(snap_->cgData.position,
612                                 cgColData.position);
613  
614 +
615 +
616 +    if (needVelocities_) {
617 +      // gather up the atomic velocities
618 +      AtomPlanVectorColumn->gather(snap_->atomData.velocity,
619 +                                   atomColData.velocity);
620 +      
621 +      cgPlanVectorColumn->gather(snap_->cgData.velocity,
622 +                                 cgColData.velocity);
623 +    }
624 +
625      
626      // if needed, gather the atomic rotation matrices
627      if (storageLayout_ & DataStorage::dslAmat) {
# Line 587 | Line 639 | namespace OpenMD {
639                                     atomColData.electroFrame);
640      }
641  
642 +    // if needed, gather the atomic fluctuating charge values
643 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
644 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
645 +                              atomRowData.flucQPos);
646 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
647 +                                 atomColData.flucQPos);
648 +    }
649 +
650   #endif      
651    }
652    
# Line 609 | Line 669 | namespace OpenMD {
669        for (int i = 0; i < n; i++)
670          snap_->atomData.density[i] += rho_tmp[i];
671      }
672 +
673 +    if (storageLayout_ & DataStorage::dslElectricField) {
674 +      
675 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
676 +                                 snap_->atomData.electricField);
677 +      
678 +      int n = snap_->atomData.electricField.size();
679 +      vector<Vector3d> field_tmp(n, V3Zero);
680 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
681 +      for (int i = 0; i < n; i++)
682 +        snap_->atomData.electricField[i] += field_tmp[i];
683 +    }
684   #endif
685    }
686  
# Line 683 | Line 755 | namespace OpenMD {
755        }
756        
757        AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
758 <      for (int i = 0; i < ns; i++)
758 >      for (int i = 0; i < ns; i++)
759          snap_->atomData.skippedCharge[i] += skch_tmp[i];
760 +            
761      }
762      
763 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
764 +
765 +      int nq = snap_->atomData.flucQFrc.size();
766 +      vector<RealType> fqfrc_tmp(nq, 0.0);
767 +
768 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
769 +      for (int i = 0; i < nq; i++) {
770 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
771 +        fqfrc_tmp[i] = 0.0;
772 +      }
773 +      
774 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
775 +      for (int i = 0; i < nq; i++)
776 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
777 +            
778 +    }
779 +
780      nLocal_ = snap_->getNumberOfAtoms();
781  
782      vector<potVec> pot_temp(nLocal_,
# Line 698 | Line 788 | namespace OpenMD {
788  
789      for (int ii = 0;  ii < pot_temp.size(); ii++ )
790        pairwisePot += pot_temp[ii];
791 <    
791 >        
792 >    if (storageLayout_ & DataStorage::dslParticlePot) {
793 >      // This is the pairwise contribution to the particle pot.  The
794 >      // embedding contribution is added in each of the low level
795 >      // non-bonded routines.  In single processor, this is done in
796 >      // unpackInteractionData, not in collectData.
797 >      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
798 >        for (int i = 0; i < nLocal_; i++) {
799 >          // factor of two is because the total potential terms are divided
800 >          // by 2 in parallel due to row/ column scatter      
801 >          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
802 >        }
803 >      }
804 >    }
805 >
806      fill(pot_temp.begin(), pot_temp.end(),
807           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
808        
# Line 706 | Line 810 | namespace OpenMD {
810      
811      for (int ii = 0;  ii < pot_temp.size(); ii++ )
812        pairwisePot += pot_temp[ii];    
813 +
814 +    if (storageLayout_ & DataStorage::dslParticlePot) {
815 +      // This is the pairwise contribution to the particle pot.  The
816 +      // embedding contribution is added in each of the low level
817 +      // non-bonded routines.  In single processor, this is done in
818 +      // unpackInteractionData, not in collectData.
819 +      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
820 +        for (int i = 0; i < nLocal_; i++) {
821 +          // factor of two is because the total potential terms are divided
822 +          // by 2 in parallel due to row/ column scatter      
823 +          snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii);
824 +        }
825 +      }
826 +    }
827      
828 +    if (storageLayout_ & DataStorage::dslParticlePot) {
829 +      int npp = snap_->atomData.particlePot.size();
830 +      vector<RealType> ppot_temp(npp, 0.0);
831 +
832 +      // This is the direct or embedding contribution to the particle
833 +      // pot.
834 +      
835 +      AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp);
836 +      for (int i = 0; i < npp; i++) {
837 +        snap_->atomData.particlePot[i] += ppot_temp[i];
838 +      }
839 +
840 +      fill(ppot_temp.begin(), ppot_temp.end(), 0.0);
841 +      
842 +      AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp);
843 +      for (int i = 0; i < npp; i++) {
844 +        snap_->atomData.particlePot[i] += ppot_temp[i];
845 +      }
846 +    }
847 +
848      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
849        RealType ploc1 = pairwisePot[ii];
850        RealType ploc2 = 0.0;
# Line 714 | Line 852 | namespace OpenMD {
852        pairwisePot[ii] = ploc2;
853      }
854  
855 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
856 +      RealType ploc1 = embeddingPot[ii];
857 +      RealType ploc2 = 0.0;
858 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
859 +      embeddingPot[ii] = ploc2;
860 +    }
861 +    
862 +    // Here be dragons.
863 +    MPI::Intracomm col = colComm.getComm();
864 +
865 +    col.Allreduce(MPI::IN_PLACE,
866 +                  &snap_->frameData.conductiveHeatFlux[0], 3,
867 +                  MPI::REALTYPE, MPI::SUM);
868 +
869 +
870   #endif
871  
872    }
# Line 756 | Line 909 | namespace OpenMD {
909      
910      snap_->wrapVector(d);
911      return d;    
912 +  }
913 +
914 +  Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
915 + #ifdef IS_MPI
916 +    return cgColData.velocity[cg2];
917 + #else
918 +    return snap_->cgData.velocity[cg2];
919 + #endif
920    }
921  
922 +  Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
923 + #ifdef IS_MPI
924 +    return atomColData.velocity[atom2];
925 + #else
926 +    return snap_->atomData.velocity[atom2];
927 + #endif
928 +  }
929  
930 +
931    Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
932  
933      Vector3d d;
# Line 826 | Line 995 | namespace OpenMD {
995     */
996    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
997      int unique_id_1, unique_id_2;
998 <    
998 >        
999   #ifdef IS_MPI
1000      // in MPI, we have to look up the unique IDs for each atom
1001      unique_id_1 = AtomRowToGlobal[atom1];
1002      unique_id_2 = AtomColToGlobal[atom2];
1003 + #else
1004 +    unique_id_1 = AtomLocalToGlobal[atom1];
1005 +    unique_id_2 = AtomLocalToGlobal[atom2];
1006 + #endif  
1007  
835 –    // this situation should only arise in MPI simulations
1008      if (unique_id_1 == unique_id_2) return true;
1009 <    
1009 >
1010 > #ifdef IS_MPI
1011      // this prevents us from doing the pair on multiple processors
1012      if (unique_id_1 < unique_id_2) {
1013        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
1014      } else {
1015 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1015 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1016      }
1017   #endif
1018 +    
1019      return false;
1020    }
1021  
# Line 855 | Line 1029 | namespace OpenMD {
1029     * field) must still be handled for these pairs.
1030     */
1031    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
1032 <    int unique_id_2;
1033 < #ifdef IS_MPI
1034 <    // in MPI, we have to look up the unique IDs for the row atom.
861 <    unique_id_2 = AtomColToGlobal[atom2];
862 < #else
863 <    // in the normal loop, the atom numbers are unique
864 <    unique_id_2 = atom2;
865 < #endif
1032 >
1033 >    // excludesForAtom was constructed to use row/column indices in the MPI
1034 >    // version, and to use local IDs in the non-MPI version:
1035      
1036      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
1037           i != excludesForAtom[atom1].end(); ++i) {
1038 <      if ( (*i) == unique_id_2 ) return true;
1038 >      if ( (*i) == atom2 ) return true;
1039      }
1040  
1041      return false;
# Line 940 | Line 1109 | namespace OpenMD {
1109        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1110      }
1111  
1112 < #else
1112 >    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1113 >      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1114 >      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1115 >    }
1116  
1117 + #else
1118 +    
1119      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
946 –    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
947 –    //                         ff_->getAtomType(idents[atom2]) );
1120  
1121      if (storageLayout_ & DataStorage::dslAmat) {
1122        idat.A1 = &(snap_->atomData.aMat[atom1]);
# Line 985 | Line 1157 | namespace OpenMD {
1157        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1158        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1159      }
1160 +
1161 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1162 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1163 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1164 +    }
1165 +
1166   #endif
1167    }
1168  
1169    
1170    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1171   #ifdef IS_MPI
1172 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1173 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1172 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1173 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1174  
1175      atomRowData.force[atom1] += *(idat.f1);
1176      atomColData.force[atom2] -= *(idat.f1);
1177 +
1178 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1179 +      atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1180 +      atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1181 +    }
1182 +
1183 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1184 +      atomRowData.electricField[atom1] += *(idat.eField1);
1185 +      atomColData.electricField[atom2] += *(idat.eField2);
1186 +    }
1187 +
1188   #else
1189      pairwisePot += *(idat.pot);
1190  
1191      snap_->atomData.force[atom1] += *(idat.f1);
1192      snap_->atomData.force[atom2] -= *(idat.f1);
1193 +
1194 +    if (idat.doParticlePot) {
1195 +      // This is the pairwise contribution to the particle pot.  The
1196 +      // embedding contribution is added in each of the low level
1197 +      // non-bonded routines.  In parallel, this calculation is done
1198 +      // in collectData, not in unpackInteractionData.
1199 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1200 +      snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1201 +    }
1202 +    
1203 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1204 +      snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1205 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1206 +    }
1207 +
1208 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1209 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1210 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1211 +    }
1212 +
1213   #endif
1214      
1215    }
# Line 1185 | Line 1394 | namespace OpenMD {
1394                  }
1395                }
1396   #else
1188 –              
1397                for (vector<int>::iterator j1 = cellList_[m1].begin();
1398                     j1 != cellList_[m1].end(); ++j1) {
1399                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1400                       j2 != cellList_[m2].end(); ++j2) {
1401 <                  
1401 >    
1402                    // Always do this if we're in different cells or if
1403 <                  // we're in the same cell and the global index of the
1404 <                  // j2 cutoff group is less than the j1 cutoff group
1405 <                  
1406 <                  if (m2 != m1 || (*j2) < (*j1)) {
1403 >                  // we're in the same cell and the global index of
1404 >                  // the j2 cutoff group is greater than or equal to
1405 >                  // the j1 cutoff group.  Note that Rappaport's code
1406 >                  // has a "less than" conditional here, but that
1407 >                  // deals with atom-by-atom computation.  OpenMD
1408 >                  // allows atoms within a single cutoff group to
1409 >                  // interact with each other.
1410 >
1411 >
1412 >
1413 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1414 >
1415                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1416                      snap_->wrapVector(dr);
1417                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1214 | Line 1430 | namespace OpenMD {
1430        // branch to do all cutoff group pairs
1431   #ifdef IS_MPI
1432        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1433 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1433 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1434            dr = cgColData.position[j2] - cgRowData.position[j1];
1435            snap_->wrapVector(dr);
1436            cuts = getGroupCutoffs( j1, j2 );
# Line 1222 | Line 1438 | namespace OpenMD {
1438              neighborList.push_back(make_pair(j1, j2));
1439            }
1440          }
1441 <      }
1441 >      }      
1442   #else
1443 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1444 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1443 >      // include all groups here.
1444 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1445 >        // include self group interactions j2 == j1
1446 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1447            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1448            snap_->wrapVector(dr);
1449            cuts = getGroupCutoffs( j1, j2 );
1450            if (dr.lengthSquare() < cuts.third) {
1451              neighborList.push_back(make_pair(j1, j2));
1452            }
1453 <        }
1454 <      }        
1453 >        }    
1454 >      }
1455   #endif
1456      }
1457        

Diff Legend

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