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 1601 by gezelter, Thu Aug 4 20:04:35 2011 UTC vs.
Revision 1713 by gezelter, Sat May 19 14:21:02 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 218 | Line 234 | namespace OpenMD {
234        }      
235      }
236  
237 < #endif
222 <
223 <    // allocate memory for the parallel objects
224 <    atypesLocal.resize(nLocal_);
225 <
226 <    for (int i = 0; i < nLocal_; i++)
227 <      atypesLocal[i] = ff_->getAtomType(idents[i]);
228 <
229 <    groupList_.clear();
230 <    groupList_.resize(nGroups_);
231 <    for (int i = 0; i < nGroups_; i++) {
232 <      int gid = cgLocalToGlobal[i];
233 <      for (int j = 0; j < nLocal_; j++) {
234 <        int aid = AtomLocalToGlobal[j];
235 <        if (globalGroupMembership[aid] == gid) {
236 <          groupList_[i].push_back(j);
237 <        }
238 <      }      
239 <    }
240 <
237 > #else
238      excludesForAtom.clear();
239      excludesForAtom.resize(nLocal_);
240      toposForAtom.clear();
# Line 270 | Line 267 | namespace OpenMD {
267          }
268        }      
269      }
270 <    
270 > #endif
271 >
272 >    // allocate memory for the parallel objects
273 >    atypesLocal.resize(nLocal_);
274 >
275 >    for (int i = 0; i < nLocal_; i++)
276 >      atypesLocal[i] = ff_->getAtomType(idents[i]);
277 >
278 >    groupList_.clear();
279 >    groupList_.resize(nGroups_);
280 >    for (int i = 0; i < nGroups_; i++) {
281 >      int gid = cgLocalToGlobal[i];
282 >      for (int j = 0; j < nLocal_; j++) {
283 >        int aid = AtomLocalToGlobal[j];
284 >        if (globalGroupMembership[aid] == gid) {
285 >          groupList_[i].push_back(j);
286 >        }
287 >      }      
288 >    }
289 >
290 >
291      createGtypeCutoffMap();
292  
293    }
# Line 506 | Line 523 | namespace OpenMD {
523             atomRowData.skippedCharge.end(), 0.0);
524        fill(atomColData.skippedCharge.begin(),
525             atomColData.skippedCharge.end(), 0.0);
526 +    }
527 +
528 +    if (storageLayout_ & DataStorage::dslElectricField) {    
529 +      fill(atomRowData.electricField.begin(),
530 +           atomRowData.electricField.end(), V3Zero);
531 +      fill(atomColData.electricField.begin(),
532 +           atomColData.electricField.end(), V3Zero);
533 +    }
534 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
535 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
536 +           0.0);
537 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
538 +           0.0);
539      }
540  
541   #endif
# Line 520 | Line 550 | namespace OpenMD {
550        fill(snap_->atomData.density.begin(),
551             snap_->atomData.density.end(), 0.0);
552      }
553 +
554      if (storageLayout_ & DataStorage::dslFunctional) {
555        fill(snap_->atomData.functional.begin(),
556             snap_->atomData.functional.end(), 0.0);
557      }
558 +
559      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
560        fill(snap_->atomData.functionalDerivative.begin(),
561             snap_->atomData.functionalDerivative.end(), 0.0);
562      }
563 +
564      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
565        fill(snap_->atomData.skippedCharge.begin(),
566             snap_->atomData.skippedCharge.end(), 0.0);
567      }
568 <    
568 >
569 >    if (storageLayout_ & DataStorage::dslElectricField) {      
570 >      fill(snap_->atomData.electricField.begin(),
571 >           snap_->atomData.electricField.end(), V3Zero);
572 >    }
573    }
574  
575  
# Line 572 | Line 609 | namespace OpenMD {
609                                     atomColData.electroFrame);
610      }
611  
612 +    // if needed, gather the atomic fluctuating charge values
613 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
614 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
615 +                              atomRowData.flucQPos);
616 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
617 +                                 atomColData.flucQPos);
618 +    }
619 +
620   #endif      
621    }
622    
# Line 593 | Line 638 | namespace OpenMD {
638        AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
639        for (int i = 0; i < n; i++)
640          snap_->atomData.density[i] += rho_tmp[i];
641 +    }
642 +
643 +    if (storageLayout_ & DataStorage::dslElectricField) {
644 +      
645 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
646 +                                 snap_->atomData.electricField);
647 +      
648 +      int n = snap_->atomData.electricField.size();
649 +      vector<Vector3d> field_tmp(n, V3Zero);
650 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
651 +      for (int i = 0; i < n; i++)
652 +        snap_->atomData.electricField[i] += field_tmp[i];
653      }
654   #endif
655    }
# Line 668 | Line 725 | namespace OpenMD {
725        }
726        
727        AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
728 <      for (int i = 0; i < ns; i++)
728 >      for (int i = 0; i < ns; i++)
729          snap_->atomData.skippedCharge[i] += skch_tmp[i];
730 +            
731      }
732      
733 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
734 +
735 +      int nq = snap_->atomData.flucQFrc.size();
736 +      vector<RealType> fqfrc_tmp(nq, 0.0);
737 +
738 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
739 +      for (int i = 0; i < nq; i++) {
740 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
741 +        fqfrc_tmp[i] = 0.0;
742 +      }
743 +      
744 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
745 +      for (int i = 0; i < nq; i++)
746 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
747 +            
748 +    }
749 +
750      nLocal_ = snap_->getNumberOfAtoms();
751  
752      vector<potVec> pot_temp(nLocal_,
# Line 699 | Line 774 | namespace OpenMD {
774        pairwisePot[ii] = ploc2;
775      }
776  
777 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
778 +      RealType ploc1 = embeddingPot[ii];
779 +      RealType ploc2 = 0.0;
780 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
781 +      embeddingPot[ii] = ploc2;
782 +    }
783 +
784   #endif
785  
786    }
# Line 811 | Line 893 | namespace OpenMD {
893     */
894    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
895      int unique_id_1, unique_id_2;
896 <    
896 >        
897   #ifdef IS_MPI
898      // in MPI, we have to look up the unique IDs for each atom
899      unique_id_1 = AtomRowToGlobal[atom1];
900      unique_id_2 = AtomColToGlobal[atom2];
901 <
902 <    // this situation should only arise in MPI simulations
901 > #else
902 >    unique_id_1 = AtomLocalToGlobal[atom1];
903 >    unique_id_2 = AtomLocalToGlobal[atom2];
904 > #endif  
905 >
906      if (unique_id_1 == unique_id_2) return true;
907 <    
907 >
908 > #ifdef IS_MPI
909      // this prevents us from doing the pair on multiple processors
910      if (unique_id_1 < unique_id_2) {
911        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
912      } else {
913 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
913 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
914      }
915   #endif
916 +    
917      return false;
918    }
919  
# Line 840 | Line 927 | namespace OpenMD {
927     * field) must still be handled for these pairs.
928     */
929    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
930 <    int unique_id_2;
931 < #ifdef IS_MPI
932 <    // in MPI, we have to look up the unique IDs for the row atom.
846 <    unique_id_2 = AtomColToGlobal[atom2];
847 < #else
848 <    // in the normal loop, the atom numbers are unique
849 <    unique_id_2 = atom2;
850 < #endif
930 >
931 >    // excludesForAtom was constructed to use row/column indices in the MPI
932 >    // version, and to use local IDs in the non-MPI version:
933      
934      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
935           i != excludesForAtom[atom1].end(); ++i) {
936 <      if ( (*i) == unique_id_2 ) return true;
936 >      if ( (*i) == atom2 ) return true;
937      }
938  
939      return false;
# Line 926 | Line 1008 | namespace OpenMD {
1008      }
1009  
1010   #else
1011 +    
1012  
1013 +    // cerr << "atoms = " << atom1 << " " << atom2 << "\n";
1014 +    // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n";
1015 +    // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n";
1016 +
1017      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1018      //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
1019      //                         ff_->getAtomType(idents[atom2]) );
# Line 976 | Line 1063 | namespace OpenMD {
1063    
1064    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1065   #ifdef IS_MPI
1066 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1067 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1066 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1067 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1068  
1069      atomRowData.force[atom1] += *(idat.f1);
1070      atomColData.force[atom2] -= *(idat.f1);
1071 +
1072 +    // should particle pot be done here also?
1073   #else
1074      pairwisePot += *(idat.pot);
1075  
1076      snap_->atomData.force[atom1] += *(idat.f1);
1077      snap_->atomData.force[atom2] -= *(idat.f1);
1078 +
1079 +    if (idat.doParticlePot) {
1080 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1081 +      snap_->atomData.particlePot[atom2] -= *(idat.vpair) * *(idat.sw);
1082 +    }
1083 +      
1084   #endif
1085      
1086    }
# Line 1091 | Line 1186 | namespace OpenMD {
1186          // add this cutoff group to the list of groups in this cell;
1187          cellListCol_[cellIndex].push_back(i);
1188        }
1189 +    
1190   #else
1191        for (int i = 0; i < nGroups_; i++) {
1192          rs = snap_->cgData.position[i];
# Line 1116 | Line 1212 | namespace OpenMD {
1212          // add this cutoff group to the list of groups in this cell;
1213          cellList_[cellIndex].push_back(i);
1214        }
1215 +
1216   #endif
1217  
1218        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1128 | Line 1225 | namespace OpenMD {
1225                   os != cellOffsets_.end(); ++os) {
1226                
1227                Vector3i m2v = m1v + (*os);
1228 <              
1228 >            
1229 >
1230                if (m2v.x() >= nCells_.x()) {
1231                  m2v.x() = 0;          
1232                } else if (m2v.x() < 0) {
# Line 1146 | Line 1244 | namespace OpenMD {
1244                } else if (m2v.z() < 0) {
1245                  m2v.z() = nCells_.z() - 1;
1246                }
1247 <              
1247 >
1248                int m2 = Vlinear (m2v, nCells_);
1249                
1250   #ifdef IS_MPI
# Line 1155 | Line 1253 | namespace OpenMD {
1253                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1254                       j2 != cellListCol_[m2].end(); ++j2) {
1255                    
1256 <                  // In parallel, we need to visit *all* pairs of row &
1257 <                  // column indicies and will truncate later on.
1256 >                  // In parallel, we need to visit *all* pairs of row
1257 >                  // & column indicies and will divide labor in the
1258 >                  // force evaluation later.
1259                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1260                    snap_->wrapVector(dr);
1261                    cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1166 | Line 1265 | namespace OpenMD {
1265                  }
1266                }
1267   #else
1169 –              
1268                for (vector<int>::iterator j1 = cellList_[m1].begin();
1269                     j1 != cellList_[m1].end(); ++j1) {
1270                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1271                       j2 != cellList_[m2].end(); ++j2) {
1272 <                  
1272 >    
1273                    // Always do this if we're in different cells or if
1274 <                  // we're in the same cell and the global index of the
1275 <                  // j2 cutoff group is less than the j1 cutoff group
1276 <                  
1277 <                  if (m2 != m1 || (*j2) < (*j1)) {
1274 >                  // we're in the same cell and the global index of
1275 >                  // the j2 cutoff group is greater than or equal to
1276 >                  // the j1 cutoff group.  Note that Rappaport's code
1277 >                  // has a "less than" conditional here, but that
1278 >                  // deals with atom-by-atom computation.  OpenMD
1279 >                  // allows atoms within a single cutoff group to
1280 >                  // interact with each other.
1281 >
1282 >
1283 >
1284 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1285 >
1286                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1287                      snap_->wrapVector(dr);
1288                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1195 | Line 1301 | namespace OpenMD {
1301        // branch to do all cutoff group pairs
1302   #ifdef IS_MPI
1303        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1304 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1304 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1305            dr = cgColData.position[j2] - cgRowData.position[j1];
1306            snap_->wrapVector(dr);
1307            cuts = getGroupCutoffs( j1, j2 );
# Line 1203 | Line 1309 | namespace OpenMD {
1309              neighborList.push_back(make_pair(j1, j2));
1310            }
1311          }
1312 <      }
1312 >      }      
1313   #else
1314 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1315 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1314 >      // include all groups here.
1315 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1316 >        // include self group interactions j2 == j1
1317 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1318            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1319            snap_->wrapVector(dr);
1320            cuts = getGroupCutoffs( j1, j2 );
1321            if (dr.lengthSquare() < cuts.third) {
1322              neighborList.push_back(make_pair(j1, j2));
1323            }
1324 <        }
1325 <      }        
1324 >        }    
1325 >      }
1326   #endif
1327      }
1328        

Diff Legend

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