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 1616 by gezelter, Tue Aug 30 15:45:35 2011 UTC vs.
Revision 1760 by gezelter, Thu Jun 21 19:26:46 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 164 | 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);
# Line 448 | Line 463 | namespace OpenMD {
463        }
464      }
465    }
451 –
466  
467    groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
468      int i, j;  
# Line 473 | Line 487 | namespace OpenMD {
487    void ForceMatrixDecomposition::zeroWorkArrays() {
488      pairwisePot = 0.0;
489      embeddingPot = 0.0;
490 +    excludedPot = 0.0;
491  
492   #ifdef IS_MPI
493      if (storageLayout_ & DataStorage::dslForce) {
# Line 489 | Line 504 | namespace OpenMD {
504           Vector<RealType, N_INTERACTION_FAMILIES> (0.0));
505  
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) {    
# Line 522 | Line 543 | namespace OpenMD {
543             atomRowData.skippedCharge.end(), 0.0);
544        fill(atomColData.skippedCharge.begin(),
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 536 | 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 571 | Line 620 | namespace OpenMD {
620      cgPlanVectorColumn->gather(snap_->cgData.position,
621                                 cgColData.position);
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      
635      // if needed, gather the atomic rotation matrices
636      if (storageLayout_ & DataStorage::dslAmat) {
# Line 588 | 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 610 | 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 689 | Line 769 | namespace OpenMD {
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];    
831 +
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;
# Line 717 | Line 874 | namespace OpenMD {
874      }
875  
876      for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
877 <      RealType ploc1 = embeddingPot[ii];
877 >      RealType ploc1 = excludedPot[ii];
878        RealType ploc2 = 0.0;
879        MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
880 <      embeddingPot[ii] = ploc2;
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 + #endif
911 +    
912 +  }
913 +
914 +
915 +
916    int ForceMatrixDecomposition::getNAtomsInRow() {  
917   #ifdef IS_MPI
918      return nAtomsInRow_;
# Line 767 | Line 953 | namespace OpenMD {
953      return d;    
954    }
955  
956 +  Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){
957 + #ifdef IS_MPI
958 +    return cgColData.velocity[cg2];
959 + #else
960 +    return snap_->cgData.velocity[cg2];
961 + #endif
962 +  }
963  
964 +  Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){
965 + #ifdef IS_MPI
966 +    return atomColData.velocity[atom2];
967 + #else
968 +    return snap_->atomData.velocity[atom2];
969 + #endif
970 +  }
971 +
972 +
973    Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){
974  
975      Vector3d d;
# Line 833 | Line 1035 | namespace OpenMD {
1035     * We need to exclude some overcounted interactions that result from
1036     * the parallel decomposition.
1037     */
1038 <  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
1039 <    int unique_id_1, unique_id_2;
1038 >  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) {
1039 >    int unique_id_1, unique_id_2, group1, group2;
1040          
1041   #ifdef IS_MPI
1042      // in MPI, we have to look up the unique IDs for each atom
1043      unique_id_1 = AtomRowToGlobal[atom1];
1044      unique_id_2 = AtomColToGlobal[atom2];
1045 +    group1 = cgRowToGlobal[cg1];
1046 +    group2 = cgColToGlobal[cg2];
1047   #else
1048      unique_id_1 = AtomLocalToGlobal[atom1];
1049      unique_id_2 = AtomLocalToGlobal[atom2];
1050 +    group1 = cgLocalToGlobal[cg1];
1051 +    group2 = cgLocalToGlobal[cg2];
1052   #endif  
1053  
1054      if (unique_id_1 == unique_id_2) return true;
# Line 854 | Line 1060 | namespace OpenMD {
1060      } else {
1061        if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
1062      }
1063 + #endif    
1064 +
1065 + #ifndef IS_MPI
1066 +    if (group1 == group2) {
1067 +      if (unique_id_1 < unique_id_2) return true;
1068 +    }
1069   #endif
1070      
1071      return false;
# Line 949 | Line 1161 | namespace OpenMD {
1161        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1162      }
1163  
1164 < #else
1164 >    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1165 >      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1166 >      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1167 >    }
1168  
1169 + #else
1170 +    
1171      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
955 –    //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
956 –    //                         ff_->getAtomType(idents[atom2]) );
1172  
1173      if (storageLayout_ & DataStorage::dslAmat) {
1174        idat.A1 = &(snap_->atomData.aMat[atom1]);
# Line 994 | Line 1209 | namespace OpenMD {
1209        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1210        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1211      }
1212 +
1213 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1214 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1215 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1216 +    }
1217 +
1218   #endif
1219    }
1220  
1221    
1222    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1223   #ifdef IS_MPI
1224 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1225 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1224 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1225 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1226 >    expot_row[atom1] += RealType(0.5) *  *(idat.excludedPot);
1227 >    expot_col[atom2] += RealType(0.5) *  *(idat.excludedPot);
1228  
1229      atomRowData.force[atom1] += *(idat.f1);
1230      atomColData.force[atom2] -= *(idat.f1);
1231 +
1232 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1233 +      atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1234 +      atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1235 +    }
1236 +
1237 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1238 +      atomRowData.electricField[atom1] += *(idat.eField1);
1239 +      atomColData.electricField[atom2] += *(idat.eField2);
1240 +    }
1241 +
1242   #else
1243      pairwisePot += *(idat.pot);
1244 +    excludedPot += *(idat.excludedPot);
1245  
1246      snap_->atomData.force[atom1] += *(idat.f1);
1247      snap_->atomData.force[atom2] -= *(idat.f1);
1248 +
1249 +    if (idat.doParticlePot) {
1250 +      // This is the pairwise contribution to the particle pot.  The
1251 +      // embedding contribution is added in each of the low level
1252 +      // non-bonded routines.  In parallel, this calculation is done
1253 +      // in collectData, not in unpackInteractionData.
1254 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1255 +      snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw);
1256 +    }
1257 +    
1258 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1259 +      snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1);
1260 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1261 +    }
1262 +
1263 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1264 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1265 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1266 +    }
1267 +
1268   #endif
1269      
1270    }

Diff Legend

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