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 1721 by gezelter, Thu May 24 14:17:42 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 508 | Line 525 | namespace OpenMD {
525             atomColData.skippedCharge.end(), 0.0);
526      }
527  
528 +    if (storageLayout_ & DataStorage::dslFlucQForce) {      
529 +      fill(atomRowData.flucQFrc.begin(),
530 +           atomRowData.flucQFrc.end(), 0.0);
531 +      fill(atomColData.flucQFrc.begin(),
532 +           atomColData.flucQFrc.end(), 0.0);
533 +    }
534 +
535 +    if (storageLayout_ & DataStorage::dslElectricField) {    
536 +      fill(atomRowData.electricField.begin(),
537 +           atomRowData.electricField.end(), V3Zero);
538 +      fill(atomColData.electricField.begin(),
539 +           atomColData.electricField.end(), V3Zero);
540 +    }
541 +
542 +    if (storageLayout_ & DataStorage::dslFlucQForce) {    
543 +      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(),
544 +           0.0);
545 +      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(),
546 +           0.0);
547 +    }
548 +
549   #endif
550      // even in parallel, we need to zero out the local arrays:
551  
# Line 520 | Line 558 | namespace OpenMD {
558        fill(snap_->atomData.density.begin(),
559             snap_->atomData.density.end(), 0.0);
560      }
561 +
562      if (storageLayout_ & DataStorage::dslFunctional) {
563        fill(snap_->atomData.functional.begin(),
564             snap_->atomData.functional.end(), 0.0);
565      }
566 +
567      if (storageLayout_ & DataStorage::dslFunctionalDerivative) {      
568        fill(snap_->atomData.functionalDerivative.begin(),
569             snap_->atomData.functionalDerivative.end(), 0.0);
570      }
571 +
572      if (storageLayout_ & DataStorage::dslSkippedCharge) {      
573        fill(snap_->atomData.skippedCharge.begin(),
574             snap_->atomData.skippedCharge.end(), 0.0);
575      }
576 <    
576 >
577 >    if (storageLayout_ & DataStorage::dslElectricField) {      
578 >      fill(snap_->atomData.electricField.begin(),
579 >           snap_->atomData.electricField.end(), V3Zero);
580 >    }
581    }
582  
583  
# Line 572 | Line 617 | namespace OpenMD {
617                                     atomColData.electroFrame);
618      }
619  
620 +    // if needed, gather the atomic fluctuating charge values
621 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {
622 +      AtomPlanRealRow->gather(snap_->atomData.flucQPos,
623 +                              atomRowData.flucQPos);
624 +      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
625 +                                 atomColData.flucQPos);
626 +    }
627 +
628   #endif      
629    }
630    
# Line 594 | Line 647 | namespace OpenMD {
647        for (int i = 0; i < n; i++)
648          snap_->atomData.density[i] += rho_tmp[i];
649      }
650 +
651 +    if (storageLayout_ & DataStorage::dslElectricField) {
652 +      
653 +      AtomPlanVectorRow->scatter(atomRowData.electricField,
654 +                                 snap_->atomData.electricField);
655 +      
656 +      int n = snap_->atomData.electricField.size();
657 +      vector<Vector3d> field_tmp(n, V3Zero);
658 +      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
659 +      for (int i = 0; i < n; i++)
660 +        snap_->atomData.electricField[i] += field_tmp[i];
661 +    }
662   #endif
663    }
664  
# Line 668 | Line 733 | namespace OpenMD {
733        }
734        
735        AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
736 <      for (int i = 0; i < ns; i++)
736 >      for (int i = 0; i < ns; i++)
737          snap_->atomData.skippedCharge[i] += skch_tmp[i];
738 +            
739      }
740      
741 +    if (storageLayout_ & DataStorage::dslFlucQForce) {
742 +
743 +      int nq = snap_->atomData.flucQFrc.size();
744 +      vector<RealType> fqfrc_tmp(nq, 0.0);
745 +
746 +      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
747 +      for (int i = 0; i < nq; i++) {
748 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
749 +        fqfrc_tmp[i] = 0.0;
750 +      }
751 +      
752 +      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
753 +      for (int i = 0; i < nq; i++)
754 +        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
755 +            
756 +    }
757 +
758      nLocal_ = snap_->getNumberOfAtoms();
759  
760      vector<potVec> pot_temp(nLocal_,
# Line 699 | Line 782 | namespace OpenMD {
782        pairwisePot[ii] = ploc2;
783      }
784  
785 +    for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) {
786 +      RealType ploc1 = embeddingPot[ii];
787 +      RealType ploc2 = 0.0;
788 +      MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM);
789 +      embeddingPot[ii] = ploc2;
790 +    }
791 +
792   #endif
793  
794    }
# Line 811 | Line 901 | namespace OpenMD {
901     */
902    bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) {
903      int unique_id_1, unique_id_2;
904 <    
904 >        
905   #ifdef IS_MPI
906      // in MPI, we have to look up the unique IDs for each atom
907      unique_id_1 = AtomRowToGlobal[atom1];
908      unique_id_2 = AtomColToGlobal[atom2];
909 + #else
910 +    unique_id_1 = AtomLocalToGlobal[atom1];
911 +    unique_id_2 = AtomLocalToGlobal[atom2];
912 + #endif  
913  
820 –    // this situation should only arise in MPI simulations
914      if (unique_id_1 == unique_id_2) return true;
915 <    
915 >
916 > #ifdef IS_MPI
917      // this prevents us from doing the pair on multiple processors
918      if (unique_id_1 < unique_id_2) {
919        if ((unique_id_1 + unique_id_2) % 2 == 0) return true;
920      } else {
921 <      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
921 >      if ((unique_id_1 + unique_id_2) % 2 == 1) return true;
922      }
923   #endif
924 +    
925      return false;
926    }
927  
# Line 840 | Line 935 | namespace OpenMD {
935     * field) must still be handled for these pairs.
936     */
937    bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
938 <    int unique_id_2;
939 < #ifdef IS_MPI
940 <    // 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
938 >
939 >    // excludesForAtom was constructed to use row/column indices in the MPI
940 >    // version, and to use local IDs in the non-MPI version:
941      
942      for (vector<int>::iterator i = excludesForAtom[atom1].begin();
943           i != excludesForAtom[atom1].end(); ++i) {
944 <      if ( (*i) == unique_id_2 ) return true;
944 >      if ( (*i) == atom2 ) return true;
945      }
946  
947      return false;
# Line 925 | Line 1015 | namespace OpenMD {
1015        idat.skippedCharge2 = &(atomColData.skippedCharge[atom2]);
1016      }
1017  
1018 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1019 +      idat.flucQ1 = &(atomRowData.flucQPos[atom1]);
1020 +      idat.flucQ2 = &(atomColData.flucQPos[atom2]);
1021 +    }
1022 +
1023   #else
1024 +    
1025  
1026 +    // cerr << "atoms = " << atom1 << " " << atom2 << "\n";
1027 +    // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n";
1028 +    // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n";
1029 +
1030      idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]);
1031      //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]),
1032      //                         ff_->getAtomType(idents[atom2]) );
# Line 970 | Line 1070 | namespace OpenMD {
1070        idat.skippedCharge1 = &(snap_->atomData.skippedCharge[atom1]);
1071        idat.skippedCharge2 = &(snap_->atomData.skippedCharge[atom2]);
1072      }
1073 +
1074 +    if (storageLayout_ & DataStorage::dslFlucQPosition) {              
1075 +      idat.flucQ1 = &(snap_->atomData.flucQPos[atom1]);
1076 +      idat.flucQ2 = &(snap_->atomData.flucQPos[atom2]);
1077 +    }
1078 +
1079   #endif
1080    }
1081  
1082    
1083    void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) {    
1084   #ifdef IS_MPI
1085 <    pot_row[atom1] += 0.5 *  *(idat.pot);
1086 <    pot_col[atom2] += 0.5 *  *(idat.pot);
1085 >    pot_row[atom1] += RealType(0.5) *  *(idat.pot);
1086 >    pot_col[atom2] += RealType(0.5) *  *(idat.pot);
1087  
1088      atomRowData.force[atom1] += *(idat.f1);
1089      atomColData.force[atom2] -= *(idat.f1);
1090 +
1091 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1092 +      atomRowData.flucQFrc[atom1] += *(idat.dVdFQ1);
1093 +      atomColData.flucQFrc[atom2] += *(idat.dVdFQ2);
1094 +    }
1095 +
1096 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1097 +      atomRowData.electricField[atom1] += *(idat.eField1);
1098 +      atomColData.electricField[atom2] += *(idat.eField2);
1099 +    }
1100 +
1101 +    // should particle pot be done here also?
1102   #else
1103      pairwisePot += *(idat.pot);
1104  
1105      snap_->atomData.force[atom1] += *(idat.f1);
1106      snap_->atomData.force[atom2] -= *(idat.f1);
1107 +
1108 +    if (idat.doParticlePot) {
1109 +      snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw);
1110 +      snap_->atomData.particlePot[atom2] -= *(idat.vpair) * *(idat.sw);
1111 +    }
1112 +    
1113 +    if (storageLayout_ & DataStorage::dslFlucQForce) {              
1114 +      snap_->atomData.flucQFrc[atom1] += *(idat.dVdFQ1);
1115 +      snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2);
1116 +    }
1117 +
1118 +    if (storageLayout_ & DataStorage::dslElectricField) {              
1119 +      snap_->atomData.electricField[atom1] += *(idat.eField1);
1120 +      snap_->atomData.electricField[atom2] += *(idat.eField2);
1121 +    }
1122 +
1123   #endif
1124      
1125    }
# Line 1091 | Line 1225 | namespace OpenMD {
1225          // add this cutoff group to the list of groups in this cell;
1226          cellListCol_[cellIndex].push_back(i);
1227        }
1228 +    
1229   #else
1230        for (int i = 0; i < nGroups_; i++) {
1231          rs = snap_->cgData.position[i];
# Line 1116 | Line 1251 | namespace OpenMD {
1251          // add this cutoff group to the list of groups in this cell;
1252          cellList_[cellIndex].push_back(i);
1253        }
1254 +
1255   #endif
1256  
1257        for (int m1z = 0; m1z < nCells_.z(); m1z++) {
# Line 1128 | Line 1264 | namespace OpenMD {
1264                   os != cellOffsets_.end(); ++os) {
1265                
1266                Vector3i m2v = m1v + (*os);
1267 <              
1267 >            
1268 >
1269                if (m2v.x() >= nCells_.x()) {
1270                  m2v.x() = 0;          
1271                } else if (m2v.x() < 0) {
# Line 1146 | Line 1283 | namespace OpenMD {
1283                } else if (m2v.z() < 0) {
1284                  m2v.z() = nCells_.z() - 1;
1285                }
1286 <              
1286 >
1287                int m2 = Vlinear (m2v, nCells_);
1288                
1289   #ifdef IS_MPI
# Line 1155 | Line 1292 | namespace OpenMD {
1292                  for (vector<int>::iterator j2 = cellListCol_[m2].begin();
1293                       j2 != cellListCol_[m2].end(); ++j2) {
1294                    
1295 <                  // In parallel, we need to visit *all* pairs of row &
1296 <                  // column indicies and will truncate later on.
1295 >                  // In parallel, we need to visit *all* pairs of row
1296 >                  // & column indicies and will divide labor in the
1297 >                  // force evaluation later.
1298                    dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
1299                    snap_->wrapVector(dr);
1300                    cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1166 | Line 1304 | namespace OpenMD {
1304                  }
1305                }
1306   #else
1169 –              
1307                for (vector<int>::iterator j1 = cellList_[m1].begin();
1308                     j1 != cellList_[m1].end(); ++j1) {
1309                  for (vector<int>::iterator j2 = cellList_[m2].begin();
1310                       j2 != cellList_[m2].end(); ++j2) {
1311 <                  
1311 >    
1312                    // Always do this if we're in different cells or if
1313 <                  // we're in the same cell and the global index of the
1314 <                  // j2 cutoff group is less than the j1 cutoff group
1315 <                  
1316 <                  if (m2 != m1 || (*j2) < (*j1)) {
1313 >                  // we're in the same cell and the global index of
1314 >                  // the j2 cutoff group is greater than or equal to
1315 >                  // the j1 cutoff group.  Note that Rappaport's code
1316 >                  // has a "less than" conditional here, but that
1317 >                  // deals with atom-by-atom computation.  OpenMD
1318 >                  // allows atoms within a single cutoff group to
1319 >                  // interact with each other.
1320 >
1321 >
1322 >
1323 >                  if (m2 != m1 || (*j2) >= (*j1) ) {
1324 >
1325                      dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
1326                      snap_->wrapVector(dr);
1327                      cuts = getGroupCutoffs( (*j1), (*j2) );
# Line 1195 | Line 1340 | namespace OpenMD {
1340        // branch to do all cutoff group pairs
1341   #ifdef IS_MPI
1342        for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
1343 <        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {      
1343 >        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {    
1344            dr = cgColData.position[j2] - cgRowData.position[j1];
1345            snap_->wrapVector(dr);
1346            cuts = getGroupCutoffs( j1, j2 );
# Line 1203 | Line 1348 | namespace OpenMD {
1348              neighborList.push_back(make_pair(j1, j2));
1349            }
1350          }
1351 <      }
1351 >      }      
1352   #else
1353 <      for (int j1 = 0; j1 < nGroups_ - 1; j1++) {
1354 <        for (int j2 = j1 + 1; j2 < nGroups_; j2++) {
1353 >      // include all groups here.
1354 >      for (int j1 = 0; j1 < nGroups_; j1++) {
1355 >        // include self group interactions j2 == j1
1356 >        for (int j2 = j1; j2 < nGroups_; j2++) {
1357            dr = snap_->cgData.position[j2] - snap_->cgData.position[j1];
1358            snap_->wrapVector(dr);
1359            cuts = getGroupCutoffs( j1, j2 );
1360            if (dr.lengthSquare() < cuts.third) {
1361              neighborList.push_back(make_pair(j1, j2));
1362            }
1363 <        }
1364 <      }        
1363 >        }    
1364 >      }
1365   #endif
1366      }
1367        

Diff Legend

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