| 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" | 
| 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 |  |  | 
| 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(); | 
| 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 |  | } | 
| 537 |  | fill(snap_->atomData.density.begin(), | 
| 538 |  | snap_->atomData.density.end(), 0.0); | 
| 539 |  | } | 
| 540 | + |  | 
| 541 |  | if (storageLayout_ & DataStorage::dslFunctional) { | 
| 542 |  | fill(snap_->atomData.functional.begin(), | 
| 543 |  | snap_->atomData.functional.end(), 0.0); | 
| 544 |  | } | 
| 545 | + |  | 
| 546 |  | if (storageLayout_ & DataStorage::dslFunctionalDerivative) { | 
| 547 |  | fill(snap_->atomData.functionalDerivative.begin(), | 
| 548 |  | snap_->atomData.functionalDerivative.end(), 0.0); | 
| 549 |  | } | 
| 550 | + |  | 
| 551 |  | if (storageLayout_ & DataStorage::dslSkippedCharge) { | 
| 552 |  | fill(snap_->atomData.skippedCharge.begin(), | 
| 553 |  | snap_->atomData.skippedCharge.end(), 0.0); | 
| 554 |  | } | 
| 535 | – |  | 
| 555 |  | } | 
| 556 |  |  | 
| 557 |  |  | 
| 687 |  | } | 
| 688 |  |  | 
| 689 |  | AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp); | 
| 690 | < | for (int i = 0; i < ns; i++) | 
| 690 | > | for (int i = 0; i < ns; i++) | 
| 691 |  | snap_->atomData.skippedCharge[i] += skch_tmp[i]; | 
| 692 | + |  | 
| 693 |  | } | 
| 694 |  |  | 
| 695 |  | nLocal_ = snap_->getNumberOfAtoms(); | 
| 719 |  | pairwisePot[ii] = ploc2; | 
| 720 |  | } | 
| 721 |  |  | 
| 722 | + | for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { | 
| 723 | + | RealType ploc1 = embeddingPot[ii]; | 
| 724 | + | RealType ploc2 = 0.0; | 
| 725 | + | MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); | 
| 726 | + | embeddingPot[ii] = ploc2; | 
| 727 | + | } | 
| 728 | + |  | 
| 729 |  | #endif | 
| 730 |  |  | 
| 731 |  | } | 
| 838 |  | */ | 
| 839 |  | bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) { | 
| 840 |  | int unique_id_1, unique_id_2; | 
| 841 | < |  | 
| 841 | > |  | 
| 842 |  | #ifdef IS_MPI | 
| 843 |  | // in MPI, we have to look up the unique IDs for each atom | 
| 844 |  | unique_id_1 = AtomRowToGlobal[atom1]; | 
| 845 |  | unique_id_2 = AtomColToGlobal[atom2]; | 
| 846 | + | #else | 
| 847 | + | unique_id_1 = AtomLocalToGlobal[atom1]; | 
| 848 | + | unique_id_2 = AtomLocalToGlobal[atom2]; | 
| 849 | + | #endif | 
| 850 |  |  | 
| 820 | – | // this situation should only arise in MPI simulations | 
| 851 |  | if (unique_id_1 == unique_id_2) return true; | 
| 852 | < |  | 
| 852 | > |  | 
| 853 | > | #ifdef IS_MPI | 
| 854 |  | // this prevents us from doing the pair on multiple processors | 
| 855 |  | if (unique_id_1 < unique_id_2) { | 
| 856 |  | if ((unique_id_1 + unique_id_2) % 2 == 0) return true; | 
| 857 |  | } else { | 
| 858 | < | if ((unique_id_1 + unique_id_2) % 2 == 1) return true; | 
| 858 | > | if ((unique_id_1 + unique_id_2) % 2 == 1) return true; | 
| 859 |  | } | 
| 860 |  | #endif | 
| 861 | + |  | 
| 862 |  | return false; | 
| 863 |  | } | 
| 864 |  |  | 
| 872 |  | * field) must still be handled for these pairs. | 
| 873 |  | */ | 
| 874 |  | bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) { | 
| 875 | < | int unique_id_2; | 
| 876 | < | #ifdef IS_MPI | 
| 877 | < | // 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 | 
| 875 | > |  | 
| 876 | > | // excludesForAtom was constructed to use row/column indices in the MPI | 
| 877 | > | // version, and to use local IDs in the non-MPI version: | 
| 878 |  |  | 
| 879 |  | for (vector<int>::iterator i = excludesForAtom[atom1].begin(); | 
| 880 |  | i != excludesForAtom[atom1].end(); ++i) { | 
| 881 | < | if ( (*i) == unique_id_2 ) return true; | 
| 881 | > | if ( (*i) == atom2 ) return true; | 
| 882 |  | } | 
| 883 |  |  | 
| 884 |  | return false; | 
| 953 |  | } | 
| 954 |  |  | 
| 955 |  | #else | 
| 956 | + |  | 
| 957 |  |  | 
| 958 | + | // cerr << "atoms = " << atom1 << " " << atom2 << "\n"; | 
| 959 | + | // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n"; | 
| 960 | + | // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n"; | 
| 961 | + |  | 
| 962 |  | idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]); | 
| 963 |  | //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]), | 
| 964 |  | //                         ff_->getAtomType(idents[atom2]) ); | 
| 1008 |  |  | 
| 1009 |  | void ForceMatrixDecomposition::unpackInteractionData(InteractionData &idat, int atom1, int atom2) { | 
| 1010 |  | #ifdef IS_MPI | 
| 1011 | < | pot_row[atom1] += 0.5 *  *(idat.pot); | 
| 1012 | < | pot_col[atom2] += 0.5 *  *(idat.pot); | 
| 1011 | > | pot_row[atom1] += RealType(0.5) *  *(idat.pot); | 
| 1012 | > | pot_col[atom2] += RealType(0.5) *  *(idat.pot); | 
| 1013 |  |  | 
| 1014 |  | atomRowData.force[atom1] += *(idat.f1); | 
| 1015 |  | atomColData.force[atom2] -= *(idat.f1); | 
| 1123 |  | // add this cutoff group to the list of groups in this cell; | 
| 1124 |  | cellListCol_[cellIndex].push_back(i); | 
| 1125 |  | } | 
| 1126 | + |  | 
| 1127 |  | #else | 
| 1128 |  | for (int i = 0; i < nGroups_; i++) { | 
| 1129 |  | rs = snap_->cgData.position[i]; | 
| 1149 |  | // add this cutoff group to the list of groups in this cell; | 
| 1150 |  | cellList_[cellIndex].push_back(i); | 
| 1151 |  | } | 
| 1152 | + |  | 
| 1153 |  | #endif | 
| 1154 |  |  | 
| 1155 |  | for (int m1z = 0; m1z < nCells_.z(); m1z++) { | 
| 1162 |  | os != cellOffsets_.end(); ++os) { | 
| 1163 |  |  | 
| 1164 |  | Vector3i m2v = m1v + (*os); | 
| 1165 | < |  | 
| 1165 | > |  | 
| 1166 | > |  | 
| 1167 |  | if (m2v.x() >= nCells_.x()) { | 
| 1168 |  | m2v.x() = 0; | 
| 1169 |  | } else if (m2v.x() < 0) { | 
| 1181 |  | } else if (m2v.z() < 0) { | 
| 1182 |  | m2v.z() = nCells_.z() - 1; | 
| 1183 |  | } | 
| 1184 | < |  | 
| 1184 | > |  | 
| 1185 |  | int m2 = Vlinear (m2v, nCells_); | 
| 1186 |  |  | 
| 1187 |  | #ifdef IS_MPI | 
| 1190 |  | for (vector<int>::iterator j2 = cellListCol_[m2].begin(); | 
| 1191 |  | j2 != cellListCol_[m2].end(); ++j2) { | 
| 1192 |  |  | 
| 1193 | < | // In parallel, we need to visit *all* pairs of row & | 
| 1194 | < | // column indicies and will truncate later on. | 
| 1193 | > | // In parallel, we need to visit *all* pairs of row | 
| 1194 | > | // & column indicies and will divide labor in the | 
| 1195 | > | // force evaluation later. | 
| 1196 |  | dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)]; | 
| 1197 |  | snap_->wrapVector(dr); | 
| 1198 |  | cuts = getGroupCutoffs( (*j1), (*j2) ); | 
| 1202 |  | } | 
| 1203 |  | } | 
| 1204 |  | #else | 
| 1169 | – |  | 
| 1205 |  | for (vector<int>::iterator j1 = cellList_[m1].begin(); | 
| 1206 |  | j1 != cellList_[m1].end(); ++j1) { | 
| 1207 |  | for (vector<int>::iterator j2 = cellList_[m2].begin(); | 
| 1208 |  | j2 != cellList_[m2].end(); ++j2) { | 
| 1209 | < |  | 
| 1209 | > |  | 
| 1210 |  | // Always do this if we're in different cells or if | 
| 1211 | < | // we're in the same cell and the global index of the | 
| 1212 | < | // j2 cutoff group is less than the j1 cutoff group | 
| 1213 | < |  | 
| 1214 | < | if (m2 != m1 || (*j2) < (*j1)) { | 
| 1211 | > | // we're in the same cell and the global index of | 
| 1212 | > | // the j2 cutoff group is greater than or equal to | 
| 1213 | > | // the j1 cutoff group.  Note that Rappaport's code | 
| 1214 | > | // has a "less than" conditional here, but that | 
| 1215 | > | // deals with atom-by-atom computation.  OpenMD | 
| 1216 | > | // allows atoms within a single cutoff group to | 
| 1217 | > | // interact with each other. | 
| 1218 | > |  | 
| 1219 | > |  | 
| 1220 | > |  | 
| 1221 | > | if (m2 != m1 || (*j2) >= (*j1) ) { | 
| 1222 | > |  | 
| 1223 |  | dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)]; | 
| 1224 |  | snap_->wrapVector(dr); | 
| 1225 |  | cuts = getGroupCutoffs( (*j1), (*j2) ); | 
| 1238 |  | // branch to do all cutoff group pairs | 
| 1239 |  | #ifdef IS_MPI | 
| 1240 |  | for (int j1 = 0; j1 < nGroupsInRow_; j1++) { | 
| 1241 | < | for (int j2 = 0; j2 < nGroupsInCol_; j2++) { | 
| 1241 | > | for (int j2 = 0; j2 < nGroupsInCol_; j2++) { | 
| 1242 |  | dr = cgColData.position[j2] - cgRowData.position[j1]; | 
| 1243 |  | snap_->wrapVector(dr); | 
| 1244 |  | cuts = getGroupCutoffs( j1, j2 ); | 
| 1246 |  | neighborList.push_back(make_pair(j1, j2)); | 
| 1247 |  | } | 
| 1248 |  | } | 
| 1249 | < | } | 
| 1249 | > | } | 
| 1250 |  | #else | 
| 1251 | < | for (int j1 = 0; j1 < nGroups_ - 1; j1++) { | 
| 1252 | < | for (int j2 = j1 + 1; j2 < nGroups_; j2++) { | 
| 1251 | > | // include all groups here. | 
| 1252 | > | for (int j1 = 0; j1 < nGroups_; j1++) { | 
| 1253 | > | // include self group interactions j2 == j1 | 
| 1254 | > | for (int j2 = j1; j2 < nGroups_; j2++) { | 
| 1255 |  | dr = snap_->cgData.position[j2] - snap_->cgData.position[j1]; | 
| 1256 |  | snap_->wrapVector(dr); | 
| 1257 |  | cuts = getGroupCutoffs( j1, j2 ); | 
| 1258 |  | if (dr.lengthSquare() < cuts.third) { | 
| 1259 |  | neighborList.push_back(make_pair(j1, j2)); | 
| 1260 |  | } | 
| 1261 | < | } | 
| 1262 | < | } | 
| 1261 | > | } | 
| 1262 | > | } | 
| 1263 |  | #endif | 
| 1264 |  | } | 
| 1265 |  |  |