| 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 |  |  | 
| 170 |  | AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal); | 
| 171 |  | AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal); | 
| 172 |  |  | 
| 157 | – | cerr << "Atoms in Local:\n"; | 
| 158 | – | for (int i = 0; i < AtomLocalToGlobal.size(); i++) { | 
| 159 | – | cerr << "i =\t" << i << "\t localAt =\t" << AtomLocalToGlobal[i] << "\n"; | 
| 160 | – | } | 
| 161 | – | cerr << "Atoms in Row:\n"; | 
| 162 | – | for (int i = 0; i < AtomRowToGlobal.size(); i++) { | 
| 163 | – | cerr << "i =\t" << i << "\t rowAt =\t" << AtomRowToGlobal[i] << "\n"; | 
| 164 | – | } | 
| 165 | – | cerr << "Atoms in Col:\n"; | 
| 166 | – | for (int i = 0; i < AtomColToGlobal.size(); i++) { | 
| 167 | – | cerr << "i =\t" << i << "\t colAt =\t" << AtomColToGlobal[i] << "\n"; | 
| 168 | – | } | 
| 169 | – |  | 
| 173 |  | cgRowToGlobal.resize(nGroupsInRow_); | 
| 174 |  | cgColToGlobal.resize(nGroupsInCol_); | 
| 175 |  | cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal); | 
| 176 |  | cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal); | 
| 174 | – |  | 
| 175 | – | cerr << "Gruops in Local:\n"; | 
| 176 | – | for (int i = 0; i < cgLocalToGlobal.size(); i++) { | 
| 177 | – | cerr << "i =\t" << i << "\t localCG =\t" << cgLocalToGlobal[i] << "\n"; | 
| 178 | – | } | 
| 179 | – | cerr << "Groups in Row:\n"; | 
| 180 | – | for (int i = 0; i < cgRowToGlobal.size(); i++) { | 
| 181 | – | cerr << "i =\t" << i << "\t rowCG =\t" << cgRowToGlobal[i] << "\n"; | 
| 182 | – | } | 
| 183 | – | cerr << "Groups in Col:\n"; | 
| 184 | – | for (int i = 0; i < cgColToGlobal.size(); i++) { | 
| 185 | – | cerr << "i =\t" << i << "\t colCG =\t" << cgColToGlobal[i] << "\n"; | 
| 186 | – | } | 
| 177 |  |  | 
| 188 | – |  | 
| 178 |  | massFactorsRow.resize(nAtomsInRow_); | 
| 179 |  | massFactorsCol.resize(nAtomsInCol_); | 
| 180 |  | AtomPlanRealRow->gather(massFactors, massFactorsRow); | 
| 234 |  | } | 
| 235 |  | } | 
| 236 |  |  | 
| 237 | < | #endif | 
| 249 | < |  | 
| 250 | < | // allocate memory for the parallel objects | 
| 251 | < | atypesLocal.resize(nLocal_); | 
| 252 | < |  | 
| 253 | < | for (int i = 0; i < nLocal_; i++) | 
| 254 | < | atypesLocal[i] = ff_->getAtomType(idents[i]); | 
| 255 | < |  | 
| 256 | < | groupList_.clear(); | 
| 257 | < | groupList_.resize(nGroups_); | 
| 258 | < | for (int i = 0; i < nGroups_; i++) { | 
| 259 | < | int gid = cgLocalToGlobal[i]; | 
| 260 | < | for (int j = 0; j < nLocal_; j++) { | 
| 261 | < | int aid = AtomLocalToGlobal[j]; | 
| 262 | < | if (globalGroupMembership[aid] == gid) { | 
| 263 | < | groupList_[i].push_back(j); | 
| 264 | < | } | 
| 265 | < | } | 
| 266 | < | } | 
| 267 | < |  | 
| 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 |  | } | 
| 566 |  |  | 
| 567 |  | // gather up the cutoff group positions | 
| 568 |  |  | 
| 579 | – | cerr  << "before gather\n"; | 
| 580 | – | for (int i = 0; i < snap_->cgData.position.size(); i++) { | 
| 581 | – | cerr << "cgpos = " << snap_->cgData.position[i] << "\n"; | 
| 582 | – | } | 
| 583 | – |  | 
| 569 |  | cgPlanVectorRow->gather(snap_->cgData.position, | 
| 570 |  | cgRowData.position); | 
| 586 | – |  | 
| 587 | – | cerr  << "after gather\n"; | 
| 588 | – | for (int i = 0; i < cgRowData.position.size(); i++) { | 
| 589 | – | cerr << "cgRpos = " << cgRowData.position[i] << "\n"; | 
| 590 | – | } | 
| 571 |  |  | 
| 572 |  | cgPlanVectorColumn->gather(snap_->cgData.position, | 
| 573 |  | cgColData.position); | 
| 594 | – | for (int i = 0; i < cgColData.position.size(); i++) { | 
| 595 | – | cerr << "cgCpos = " << cgColData.position[i] << "\n"; | 
| 596 | – | } | 
| 574 |  |  | 
| 575 |  |  | 
| 576 |  | // if needed, gather the atomic rotation matrices | 
| 685 |  | } | 
| 686 |  |  | 
| 687 |  | AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp); | 
| 688 | < | for (int i = 0; i < ns; i++) | 
| 688 | > | for (int i = 0; i < ns; i++) | 
| 689 |  | snap_->atomData.skippedCharge[i] += skch_tmp[i]; | 
| 690 | + |  | 
| 691 |  | } | 
| 692 |  |  | 
| 693 |  | nLocal_ = snap_->getNumberOfAtoms(); | 
| 709 |  |  | 
| 710 |  | for (int ii = 0;  ii < pot_temp.size(); ii++ ) | 
| 711 |  | pairwisePot += pot_temp[ii]; | 
| 712 | + |  | 
| 713 | + | for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { | 
| 714 | + | RealType ploc1 = pairwisePot[ii]; | 
| 715 | + | RealType ploc2 = 0.0; | 
| 716 | + | MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); | 
| 717 | + | pairwisePot[ii] = ploc2; | 
| 718 | + | } | 
| 719 | + |  | 
| 720 | + | for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { | 
| 721 | + | RealType ploc1 = embeddingPot[ii]; | 
| 722 | + | RealType ploc2 = 0.0; | 
| 723 | + | MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); | 
| 724 | + | embeddingPot[ii] = ploc2; | 
| 725 | + | } | 
| 726 | + |  | 
| 727 |  | #endif | 
| 728 |  |  | 
| 736 | – | cerr << "pairwisePot = " <<  pairwisePot << "\n"; | 
| 729 |  | } | 
| 730 |  |  | 
| 731 |  | int ForceMatrixDecomposition::getNAtomsInRow() { | 
| 760 |  |  | 
| 761 |  | #ifdef IS_MPI | 
| 762 |  | d = cgColData.position[cg2] - cgRowData.position[cg1]; | 
| 771 | – | cerr << "cg1 = " << cg1 << "\tcg1p = " << cgRowData.position[cg1] << "\n"; | 
| 772 | – | cerr << "cg2 = " << cg2 << "\tcg2p = " << cgColData.position[cg2] << "\n"; | 
| 763 |  | #else | 
| 764 |  | d = snap_->cgData.position[cg2] - snap_->cgData.position[cg1]; | 
| 775 | – | cerr << "cg1 = " << cg1 << "\tcg1p = " << snap_->cgData.position[cg1] << "\n"; | 
| 776 | – | cerr << "cg2 = " << cg2 << "\tcg2p = " << snap_->cgData.position[cg2] << "\n"; | 
| 765 |  | #endif | 
| 766 |  |  | 
| 767 |  | snap_->wrapVector(d); | 
| 836 |  | */ | 
| 837 |  | bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) { | 
| 838 |  | int unique_id_1, unique_id_2; | 
| 839 | < |  | 
| 852 | < |  | 
| 853 | < | cerr << "sap with atom1, atom2 =\t" << atom1 << "\t" << atom2 << "\n"; | 
| 839 | > |  | 
| 840 |  | #ifdef IS_MPI | 
| 841 |  | // in MPI, we have to look up the unique IDs for each atom | 
| 842 |  | unique_id_1 = AtomRowToGlobal[atom1]; | 
| 843 |  | unique_id_2 = AtomColToGlobal[atom2]; | 
| 844 | + | #else | 
| 845 | + | unique_id_1 = AtomLocalToGlobal[atom1]; | 
| 846 | + | unique_id_2 = AtomLocalToGlobal[atom2]; | 
| 847 | + | #endif | 
| 848 |  |  | 
| 859 | – | cerr << "sap with uid1, uid2 =\t" << unique_id_1 << "\t" << unique_id_2 << "\n"; | 
| 860 | – | // this situation should only arise in MPI simulations | 
| 849 |  | if (unique_id_1 == unique_id_2) return true; | 
| 850 | < |  | 
| 850 | > |  | 
| 851 | > | #ifdef IS_MPI | 
| 852 |  | // this prevents us from doing the pair on multiple processors | 
| 853 |  | if (unique_id_1 < unique_id_2) { | 
| 854 |  | if ((unique_id_1 + unique_id_2) % 2 == 0) return true; | 
| 855 |  | } else { | 
| 856 | < | if ((unique_id_1 + unique_id_2) % 2 == 1) return true; | 
| 856 | > | if ((unique_id_1 + unique_id_2) % 2 == 1) return true; | 
| 857 |  | } | 
| 858 |  | #endif | 
| 859 | + |  | 
| 860 |  | return false; | 
| 861 |  | } | 
| 862 |  |  | 
| 870 |  | * field) must still be handled for these pairs. | 
| 871 |  | */ | 
| 872 |  | bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) { | 
| 873 | < | int unique_id_2; | 
| 874 | < | #ifdef IS_MPI | 
| 875 | < | // in MPI, we have to look up the unique IDs for the row atom. | 
| 886 | < | unique_id_2 = AtomColToGlobal[atom2]; | 
| 887 | < | #else | 
| 888 | < | // in the normal loop, the atom numbers are unique | 
| 889 | < | unique_id_2 = atom2; | 
| 890 | < | #endif | 
| 873 | > |  | 
| 874 | > | // excludesForAtom was constructed to use row/column indices in the MPI | 
| 875 | > | // version, and to use local IDs in the non-MPI version: | 
| 876 |  |  | 
| 877 |  | for (vector<int>::iterator i = excludesForAtom[atom1].begin(); | 
| 878 |  | i != excludesForAtom[atom1].end(); ++i) { | 
| 879 | < | if ( (*i) == unique_id_2 ) return true; | 
| 879 | > | if ( (*i) == atom2 ) return true; | 
| 880 |  | } | 
| 881 |  |  | 
| 882 |  | return false; | 
| 1116 |  | // add this cutoff group to the list of groups in this cell; | 
| 1117 |  | cellListCol_[cellIndex].push_back(i); | 
| 1118 |  | } | 
| 1119 | + |  | 
| 1120 |  | #else | 
| 1121 |  | for (int i = 0; i < nGroups_; i++) { | 
| 1122 |  | rs = snap_->cgData.position[i]; | 
| 1142 |  | // add this cutoff group to the list of groups in this cell; | 
| 1143 |  | cellList_[cellIndex].push_back(i); | 
| 1144 |  | } | 
| 1145 | + |  | 
| 1146 |  | #endif | 
| 1147 |  |  | 
| 1148 |  | for (int m1z = 0; m1z < nCells_.z(); m1z++) { | 
| 1155 |  | os != cellOffsets_.end(); ++os) { | 
| 1156 |  |  | 
| 1157 |  | Vector3i m2v = m1v + (*os); | 
| 1158 | < |  | 
| 1158 | > |  | 
| 1159 | > |  | 
| 1160 |  | if (m2v.x() >= nCells_.x()) { | 
| 1161 |  | m2v.x() = 0; | 
| 1162 |  | } else if (m2v.x() < 0) { | 
| 1174 |  | } else if (m2v.z() < 0) { | 
| 1175 |  | m2v.z() = nCells_.z() - 1; | 
| 1176 |  | } | 
| 1177 | < |  | 
| 1177 | > |  | 
| 1178 |  | int m2 = Vlinear (m2v, nCells_); | 
| 1179 |  |  | 
| 1180 |  | #ifdef IS_MPI | 
| 1183 |  | for (vector<int>::iterator j2 = cellListCol_[m2].begin(); | 
| 1184 |  | j2 != cellListCol_[m2].end(); ++j2) { | 
| 1185 |  |  | 
| 1186 | < | // In parallel, we need to visit *all* pairs of row & | 
| 1187 | < | // column indicies and will truncate later on. | 
| 1186 | > | // In parallel, we need to visit *all* pairs of row | 
| 1187 | > | // & column indicies and will divide labor in the | 
| 1188 | > | // force evaluation later. | 
| 1189 |  | dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)]; | 
| 1190 |  | snap_->wrapVector(dr); | 
| 1191 |  | cuts = getGroupCutoffs( (*j1), (*j2) ); | 
| 1195 |  | } | 
| 1196 |  | } | 
| 1197 |  | #else | 
| 1209 | – |  | 
| 1198 |  | for (vector<int>::iterator j1 = cellList_[m1].begin(); | 
| 1199 |  | j1 != cellList_[m1].end(); ++j1) { | 
| 1200 |  | for (vector<int>::iterator j2 = cellList_[m2].begin(); | 
| 1201 |  | j2 != cellList_[m2].end(); ++j2) { | 
| 1202 | < |  | 
| 1202 | > |  | 
| 1203 |  | // Always do this if we're in different cells or if | 
| 1204 | < | // we're in the same cell and the global index of the | 
| 1205 | < | // j2 cutoff group is less than the j1 cutoff group | 
| 1206 | < |  | 
| 1207 | < | if (m2 != m1 || (*j2) < (*j1)) { | 
| 1204 | > | // we're in the same cell and the global index of | 
| 1205 | > | // the j2 cutoff group is greater than or equal to | 
| 1206 | > | // the j1 cutoff group.  Note that Rappaport's code | 
| 1207 | > | // has a "less than" conditional here, but that | 
| 1208 | > | // deals with atom-by-atom computation.  OpenMD | 
| 1209 | > | // allows atoms within a single cutoff group to | 
| 1210 | > | // interact with each other. | 
| 1211 | > |  | 
| 1212 | > |  | 
| 1213 | > |  | 
| 1214 | > | if (m2 != m1 || (*j2) >= (*j1) ) { | 
| 1215 | > |  | 
| 1216 |  | dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)]; | 
| 1217 |  | snap_->wrapVector(dr); | 
| 1218 |  | cuts = getGroupCutoffs( (*j1), (*j2) ); | 
| 1231 |  | // branch to do all cutoff group pairs | 
| 1232 |  | #ifdef IS_MPI | 
| 1233 |  | for (int j1 = 0; j1 < nGroupsInRow_; j1++) { | 
| 1234 | < | for (int j2 = 0; j2 < nGroupsInCol_; j2++) { | 
| 1234 | > | for (int j2 = 0; j2 < nGroupsInCol_; j2++) { | 
| 1235 |  | dr = cgColData.position[j2] - cgRowData.position[j1]; | 
| 1236 |  | snap_->wrapVector(dr); | 
| 1237 |  | cuts = getGroupCutoffs( j1, j2 ); | 
| 1239 |  | neighborList.push_back(make_pair(j1, j2)); | 
| 1240 |  | } | 
| 1241 |  | } | 
| 1242 | < | } | 
| 1242 | > | } | 
| 1243 |  | #else | 
| 1244 | < | for (int j1 = 0; j1 < nGroups_ - 1; j1++) { | 
| 1245 | < | for (int j2 = j1 + 1; j2 < nGroups_; j2++) { | 
| 1244 | > | // include all groups here. | 
| 1245 | > | for (int j1 = 0; j1 < nGroups_; j1++) { | 
| 1246 | > | // include self group interactions j2 == j1 | 
| 1247 | > | for (int j2 = j1; j2 < nGroups_; j2++) { | 
| 1248 |  | dr = snap_->cgData.position[j2] - snap_->cgData.position[j1]; | 
| 1249 |  | snap_->wrapVector(dr); | 
| 1250 |  | cuts = getGroupCutoffs( j1, j2 ); | 
| 1251 |  | if (dr.lengthSquare() < cuts.third) { | 
| 1252 |  | neighborList.push_back(make_pair(j1, j2)); | 
| 1253 |  | } | 
| 1254 | < | } | 
| 1255 | < | } | 
| 1254 | > | } | 
| 1255 | > | } | 
| 1256 |  | #endif | 
| 1257 |  | } | 
| 1258 |  |  |