--- branches/development/src/parallel/ForceMatrixDecomposition.cpp 2011/07/12 15:25:07 1591 +++ branches/development/src/parallel/ForceMatrixDecomposition.cpp 2011/08/30 15:45:35 1616 @@ -47,11 +47,48 @@ namespace OpenMD { using namespace std; namespace OpenMD { + ForceMatrixDecomposition::ForceMatrixDecomposition(SimInfo* info, InteractionManager* iMan) : ForceDecomposition(info, iMan) { + + // In a parallel computation, row and colum scans must visit all + // surrounding cells (not just the 14 upper triangular blocks that + // are used when the processor can see all pairs) +#ifdef IS_MPI + cellOffsets_.clear(); + cellOffsets_.push_back( Vector3i(-1,-1,-1) ); + cellOffsets_.push_back( Vector3i( 0,-1,-1) ); + cellOffsets_.push_back( Vector3i( 1,-1,-1) ); + cellOffsets_.push_back( Vector3i(-1, 0,-1) ); + cellOffsets_.push_back( Vector3i( 0, 0,-1) ); + cellOffsets_.push_back( Vector3i( 1, 0,-1) ); + cellOffsets_.push_back( Vector3i(-1, 1,-1) ); + cellOffsets_.push_back( Vector3i( 0, 1,-1) ); + cellOffsets_.push_back( Vector3i( 1, 1,-1) ); + cellOffsets_.push_back( Vector3i(-1,-1, 0) ); + cellOffsets_.push_back( Vector3i( 0,-1, 0) ); + cellOffsets_.push_back( Vector3i( 1,-1, 0) ); + cellOffsets_.push_back( Vector3i(-1, 0, 0) ); + cellOffsets_.push_back( Vector3i( 0, 0, 0) ); + cellOffsets_.push_back( Vector3i( 1, 0, 0) ); + cellOffsets_.push_back( Vector3i(-1, 1, 0) ); + cellOffsets_.push_back( Vector3i( 0, 1, 0) ); + cellOffsets_.push_back( Vector3i( 1, 1, 0) ); + cellOffsets_.push_back( Vector3i(-1,-1, 1) ); + cellOffsets_.push_back( Vector3i( 0,-1, 1) ); + cellOffsets_.push_back( Vector3i( 1,-1, 1) ); + cellOffsets_.push_back( Vector3i(-1, 0, 1) ); + cellOffsets_.push_back( Vector3i( 0, 0, 1) ); + cellOffsets_.push_back( Vector3i( 1, 0, 1) ); + cellOffsets_.push_back( Vector3i(-1, 1, 1) ); + cellOffsets_.push_back( Vector3i( 0, 1, 1) ); + cellOffsets_.push_back( Vector3i( 1, 1, 1) ); +#endif + } + + /** * distributeInitialData is essentially a copy of the older fortran * SimulationSetup */ - void ForceMatrixDecomposition::distributeInitialData() { snap_ = sman_->getCurrentSnapshot(); storageLayout_ = sman_->getStorageLayout(); @@ -74,28 +111,31 @@ namespace OpenMD { #ifdef IS_MPI - AtomCommIntRow = new Communicator(nLocal_); - AtomCommRealRow = new Communicator(nLocal_); - AtomCommVectorRow = new Communicator(nLocal_); - AtomCommMatrixRow = new Communicator(nLocal_); - AtomCommPotRow = new Communicator(nLocal_); + MPI::Intracomm row = rowComm.getComm(); + MPI::Intracomm col = colComm.getComm(); - AtomCommIntColumn = new Communicator(nLocal_); - AtomCommRealColumn = new Communicator(nLocal_); - AtomCommVectorColumn = new Communicator(nLocal_); - AtomCommMatrixColumn = new Communicator(nLocal_); - AtomCommPotColumn = new Communicator(nLocal_); + AtomPlanIntRow = new Plan(row, nLocal_); + AtomPlanRealRow = new Plan(row, nLocal_); + AtomPlanVectorRow = new Plan(row, nLocal_); + AtomPlanMatrixRow = new Plan(row, nLocal_); + AtomPlanPotRow = new Plan(row, nLocal_); - cgCommIntRow = new Communicator(nGroups_); - cgCommVectorRow = new Communicator(nGroups_); - cgCommIntColumn = new Communicator(nGroups_); - cgCommVectorColumn = new Communicator(nGroups_); + AtomPlanIntColumn = new Plan(col, nLocal_); + AtomPlanRealColumn = new Plan(col, nLocal_); + AtomPlanVectorColumn = new Plan(col, nLocal_); + AtomPlanMatrixColumn = new Plan(col, nLocal_); + AtomPlanPotColumn = new Plan(col, nLocal_); - nAtomsInRow_ = AtomCommIntRow->getSize(); - nAtomsInCol_ = AtomCommIntColumn->getSize(); - nGroupsInRow_ = cgCommIntRow->getSize(); - nGroupsInCol_ = cgCommIntColumn->getSize(); + cgPlanIntRow = new Plan(row, nGroups_); + cgPlanVectorRow = new Plan(row, nGroups_); + cgPlanIntColumn = new Plan(col, nGroups_); + cgPlanVectorColumn = new Plan(col, nGroups_); + nAtomsInRow_ = AtomPlanIntRow->getSize(); + nAtomsInCol_ = AtomPlanIntColumn->getSize(); + nGroupsInRow_ = cgPlanIntRow->getSize(); + nGroupsInCol_ = cgPlanIntColumn->getSize(); + // Modify the data storage objects with the correct layouts and sizes: atomRowData.resize(nAtomsInRow_); atomRowData.setStorageLayout(storageLayout_); @@ -109,8 +149,8 @@ namespace OpenMD { identsRow.resize(nAtomsInRow_); identsCol.resize(nAtomsInCol_); - AtomCommIntRow->gather(idents, identsRow); - AtomCommIntColumn->gather(idents, identsCol); + AtomPlanIntRow->gather(idents, identsRow); + AtomPlanIntColumn->gather(idents, identsCol); // allocate memory for the parallel objects atypesRow.resize(nAtomsInRow_); @@ -126,18 +166,18 @@ namespace OpenMD { AtomRowToGlobal.resize(nAtomsInRow_); AtomColToGlobal.resize(nAtomsInCol_); - AtomCommIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal); - AtomCommIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal); - + AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal); + AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal); + cgRowToGlobal.resize(nGroupsInRow_); cgColToGlobal.resize(nGroupsInCol_); - cgCommIntRow->gather(cgLocalToGlobal, cgRowToGlobal); - cgCommIntColumn->gather(cgLocalToGlobal, cgColToGlobal); + cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal); + cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal); massFactorsRow.resize(nAtomsInRow_); massFactorsCol.resize(nAtomsInCol_); - AtomCommRealRow->gather(massFactors, massFactorsRow); - AtomCommRealColumn->gather(massFactors, massFactorsCol); + AtomPlanRealRow->gather(massFactors, massFactorsRow); + AtomPlanRealColumn->gather(massFactors, massFactorsCol); groupListRow_.clear(); groupListRow_.resize(nGroupsInRow_); @@ -193,26 +233,7 @@ namespace OpenMD { } } -#endif - - // allocate memory for the parallel objects - atypesLocal.resize(nLocal_); - - for (int i = 0; i < nLocal_; i++) - atypesLocal[i] = ff_->getAtomType(idents[i]); - - groupList_.clear(); - groupList_.resize(nGroups_); - for (int i = 0; i < nGroups_; i++) { - int gid = cgLocalToGlobal[i]; - for (int j = 0; j < nLocal_; j++) { - int aid = AtomLocalToGlobal[j]; - if (globalGroupMembership[aid] == gid) { - groupList_[i].push_back(j); - } - } - } - +#else excludesForAtom.clear(); excludesForAtom.resize(nLocal_); toposForAtom.clear(); @@ -245,7 +266,27 @@ namespace OpenMD { } } } - +#endif + + // allocate memory for the parallel objects + atypesLocal.resize(nLocal_); + + for (int i = 0; i < nLocal_; i++) + atypesLocal[i] = ff_->getAtomType(idents[i]); + + groupList_.clear(); + groupList_.resize(nGroups_); + for (int i = 0; i < nGroups_; i++) { + int gid = cgLocalToGlobal[i]; + for (int j = 0; j < nLocal_; j++) { + int aid = AtomLocalToGlobal[j]; + if (globalGroupMembership[aid] == gid) { + groupList_[i].push_back(j); + } + } + } + + createGtypeCutoffMap(); } @@ -253,9 +294,11 @@ namespace OpenMD { void ForceMatrixDecomposition::createGtypeCutoffMap() { RealType tol = 1e-6; + largestRcut_ = 0.0; RealType rc; int atid; set atypes = info_->getSimulatedAtomTypes(); + map atypeCutoff; for (set::iterator at = atypes.begin(); @@ -263,10 +306,10 @@ namespace OpenMD { atid = (*at)->getIdent(); if (userChoseCutoff_) atypeCutoff[atid] = userCutoff_; - else + else atypeCutoff[atid] = interactionMan_->getSuggestedCutoffRadius(*at); } - + vector gTypeCutoffs; // first we do a single loop over the cutoff groups to find the // largest cutoff for any atypes present in this group. @@ -326,19 +369,16 @@ namespace OpenMD { vector groupCutoff(nGroups_, 0.0); groupToGtype.resize(nGroups_); for (int cg1 = 0; cg1 < nGroups_; cg1++) { - groupCutoff[cg1] = 0.0; vector atomList = getAtomsInGroupRow(cg1); - for (vector::iterator ia = atomList.begin(); ia != atomList.end(); ++ia) { int atom1 = (*ia); atid = idents[atom1]; - if (atypeCutoff[atid] > groupCutoff[cg1]) { + if (atypeCutoff[atid] > groupCutoff[cg1]) groupCutoff[cg1] = atypeCutoff[atid]; - } } - + bool gTypeFound = false; for (int gt = 0; gt < gTypeCutoffs.size(); gt++) { if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) { @@ -346,7 +386,7 @@ namespace OpenMD { gTypeFound = true; } } - if (!gTypeFound) { + if (!gTypeFound) { gTypeCutoffs.push_back( groupCutoff[cg1] ); groupToGtype[cg1] = gTypeCutoffs.size() - 1; } @@ -390,13 +430,9 @@ namespace OpenMD { pair key = make_pair(i,j); gTypeCutoffMap[key].first = thisRcut; - if (thisRcut > largestRcut_) largestRcut_ = thisRcut; - gTypeCutoffMap[key].second = thisRcut*thisRcut; - gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2); - // sanity check if (userChoseCutoff_) { @@ -522,30 +558,33 @@ namespace OpenMD { #ifdef IS_MPI // gather up the atomic positions - AtomCommVectorRow->gather(snap_->atomData.position, + AtomPlanVectorRow->gather(snap_->atomData.position, atomRowData.position); - AtomCommVectorColumn->gather(snap_->atomData.position, + AtomPlanVectorColumn->gather(snap_->atomData.position, atomColData.position); // gather up the cutoff group positions - cgCommVectorRow->gather(snap_->cgData.position, + + cgPlanVectorRow->gather(snap_->cgData.position, cgRowData.position); - cgCommVectorColumn->gather(snap_->cgData.position, + + cgPlanVectorColumn->gather(snap_->cgData.position, cgColData.position); + // if needed, gather the atomic rotation matrices if (storageLayout_ & DataStorage::dslAmat) { - AtomCommMatrixRow->gather(snap_->atomData.aMat, + AtomPlanMatrixRow->gather(snap_->atomData.aMat, atomRowData.aMat); - AtomCommMatrixColumn->gather(snap_->atomData.aMat, + AtomPlanMatrixColumn->gather(snap_->atomData.aMat, atomColData.aMat); } // if needed, gather the atomic eletrostatic frames if (storageLayout_ & DataStorage::dslElectroFrame) { - AtomCommMatrixRow->gather(snap_->atomData.electroFrame, + AtomPlanMatrixRow->gather(snap_->atomData.electroFrame, atomRowData.electroFrame); - AtomCommMatrixColumn->gather(snap_->atomData.electroFrame, + AtomPlanMatrixColumn->gather(snap_->atomData.electroFrame, atomColData.electroFrame); } @@ -562,12 +601,12 @@ namespace OpenMD { if (storageLayout_ & DataStorage::dslDensity) { - AtomCommRealRow->scatter(atomRowData.density, + AtomPlanRealRow->scatter(atomRowData.density, snap_->atomData.density); int n = snap_->atomData.density.size(); vector rho_tmp(n, 0.0); - AtomCommRealColumn->scatter(atomColData.density, rho_tmp); + AtomPlanRealColumn->scatter(atomColData.density, rho_tmp); for (int i = 0; i < n; i++) snap_->atomData.density[i] += rho_tmp[i]; } @@ -583,16 +622,16 @@ namespace OpenMD { storageLayout_ = sman_->getStorageLayout(); #ifdef IS_MPI if (storageLayout_ & DataStorage::dslFunctional) { - AtomCommRealRow->gather(snap_->atomData.functional, + AtomPlanRealRow->gather(snap_->atomData.functional, atomRowData.functional); - AtomCommRealColumn->gather(snap_->atomData.functional, + AtomPlanRealColumn->gather(snap_->atomData.functional, atomColData.functional); } if (storageLayout_ & DataStorage::dslFunctionalDerivative) { - AtomCommRealRow->gather(snap_->atomData.functionalDerivative, + AtomPlanRealRow->gather(snap_->atomData.functionalDerivative, atomRowData.functionalDerivative); - AtomCommRealColumn->gather(snap_->atomData.functionalDerivative, + AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative, atomColData.functionalDerivative); } #endif @@ -606,28 +645,29 @@ namespace OpenMD { int n = snap_->atomData.force.size(); vector frc_tmp(n, V3Zero); - AtomCommVectorRow->scatter(atomRowData.force, frc_tmp); + AtomPlanVectorRow->scatter(atomRowData.force, frc_tmp); for (int i = 0; i < n; i++) { snap_->atomData.force[i] += frc_tmp[i]; frc_tmp[i] = 0.0; } - AtomCommVectorColumn->scatter(atomColData.force, frc_tmp); - for (int i = 0; i < n; i++) + AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp); + for (int i = 0; i < n; i++) { snap_->atomData.force[i] += frc_tmp[i]; + } if (storageLayout_ & DataStorage::dslTorque) { int nt = snap_->atomData.torque.size(); vector trq_tmp(nt, V3Zero); - AtomCommVectorRow->scatter(atomRowData.torque, trq_tmp); + AtomPlanVectorRow->scatter(atomRowData.torque, trq_tmp); for (int i = 0; i < nt; i++) { snap_->atomData.torque[i] += trq_tmp[i]; trq_tmp[i] = 0.0; } - AtomCommVectorColumn->scatter(atomColData.torque, trq_tmp); + AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp); for (int i = 0; i < nt; i++) snap_->atomData.torque[i] += trq_tmp[i]; } @@ -637,15 +677,16 @@ namespace OpenMD { int ns = snap_->atomData.skippedCharge.size(); vector skch_tmp(ns, 0.0); - AtomCommRealRow->scatter(atomRowData.skippedCharge, skch_tmp); + AtomPlanRealRow->scatter(atomRowData.skippedCharge, skch_tmp); for (int i = 0; i < ns; i++) { snap_->atomData.skippedCharge[i] += skch_tmp[i]; skch_tmp[i] = 0.0; } - AtomCommRealColumn->scatter(atomColData.skippedCharge, skch_tmp); - for (int i = 0; i < ns; i++) + AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp); + for (int i = 0; i < ns; i++) snap_->atomData.skippedCharge[i] += skch_tmp[i]; + } nLocal_ = snap_->getNumberOfAtoms(); @@ -655,7 +696,7 @@ namespace OpenMD { // scatter/gather pot_row into the members of my column - AtomCommPotRow->scatter(pot_row, pot_temp); + AtomPlanPotRow->scatter(pot_row, pot_temp); for (int ii = 0; ii < pot_temp.size(); ii++ ) pairwisePot += pot_temp[ii]; @@ -663,10 +704,25 @@ namespace OpenMD { fill(pot_temp.begin(), pot_temp.end(), Vector (0.0)); - AtomCommPotColumn->scatter(pot_col, pot_temp); + AtomPlanPotColumn->scatter(pot_col, pot_temp); for (int ii = 0; ii < pot_temp.size(); ii++ ) pairwisePot += pot_temp[ii]; + + for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { + RealType ploc1 = pairwisePot[ii]; + RealType ploc2 = 0.0; + MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); + pairwisePot[ii] = ploc2; + } + + for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { + RealType ploc1 = embeddingPot[ii]; + RealType ploc2 = 0.0; + MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); + embeddingPot[ii] = ploc2; + } + #endif } @@ -779,22 +835,27 @@ namespace OpenMD { */ bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) { int unique_id_1, unique_id_2; - + #ifdef IS_MPI // in MPI, we have to look up the unique IDs for each atom unique_id_1 = AtomRowToGlobal[atom1]; unique_id_2 = AtomColToGlobal[atom2]; +#else + unique_id_1 = AtomLocalToGlobal[atom1]; + unique_id_2 = AtomLocalToGlobal[atom2]; +#endif - // this situation should only arise in MPI simulations if (unique_id_1 == unique_id_2) return true; - + +#ifdef IS_MPI // this prevents us from doing the pair on multiple processors if (unique_id_1 < unique_id_2) { if ((unique_id_1 + unique_id_2) % 2 == 0) return true; } else { - if ((unique_id_1 + unique_id_2) % 2 == 1) return true; + if ((unique_id_1 + unique_id_2) % 2 == 1) return true; } #endif + return false; } @@ -808,19 +869,13 @@ namespace OpenMD { * field) must still be handled for these pairs. */ bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) { - int unique_id_2; - -#ifdef IS_MPI - // in MPI, we have to look up the unique IDs for the row atom. - unique_id_2 = AtomColToGlobal[atom2]; -#else - // in the normal loop, the atom numbers are unique - unique_id_2 = atom2; -#endif + + // excludesForAtom was constructed to use row/column indices in the MPI + // version, and to use local IDs in the non-MPI version: for (vector::iterator i = excludesForAtom[atom1].begin(); i != excludesForAtom[atom1].end(); ++i) { - if ( (*i) == unique_id_2 ) return true; + if ( (*i) == atom2 ) return true; } return false; @@ -1036,7 +1091,6 @@ namespace OpenMD { // add this cutoff group to the list of groups in this cell; cellListRow_[cellIndex].push_back(i); } - for (int i = 0; i < nGroupsInCol_; i++) { rs = cgColData.position[i]; @@ -1061,6 +1115,7 @@ namespace OpenMD { // add this cutoff group to the list of groups in this cell; cellListCol_[cellIndex].push_back(i); } + #else for (int i = 0; i < nGroups_; i++) { rs = snap_->cgData.position[i]; @@ -1081,11 +1136,12 @@ namespace OpenMD { whichCell.z() = nCells_.z() * scaled.z(); // find single index of this cell: - cellIndex = Vlinear(whichCell, nCells_); + cellIndex = Vlinear(whichCell, nCells_); // add this cutoff group to the list of groups in this cell; cellList_[cellIndex].push_back(i); } + #endif for (int m1z = 0; m1z < nCells_.z(); m1z++) { @@ -1098,7 +1154,8 @@ namespace OpenMD { os != cellOffsets_.end(); ++os) { Vector3i m2v = m1v + (*os); - + + if (m2v.x() >= nCells_.x()) { m2v.x() = 0; } else if (m2v.x() < 0) { @@ -1116,7 +1173,7 @@ namespace OpenMD { } else if (m2v.z() < 0) { m2v.z() = nCells_.z() - 1; } - + int m2 = Vlinear (m2v, nCells_); #ifdef IS_MPI @@ -1125,32 +1182,36 @@ namespace OpenMD { for (vector::iterator j2 = cellListCol_[m2].begin(); j2 != cellListCol_[m2].end(); ++j2) { - // Always do this if we're in different cells or if - // we're in the same cell and the global index of the - // j2 cutoff group is less than the j1 cutoff group - - if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) { - dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)]; - snap_->wrapVector(dr); - cuts = getGroupCutoffs( (*j1), (*j2) ); - if (dr.lengthSquare() < cuts.third) { - neighborList.push_back(make_pair((*j1), (*j2))); - } - } + // In parallel, we need to visit *all* pairs of row + // & column indicies and will divide labor in the + // force evaluation later. + dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)]; + snap_->wrapVector(dr); + cuts = getGroupCutoffs( (*j1), (*j2) ); + if (dr.lengthSquare() < cuts.third) { + neighborList.push_back(make_pair((*j1), (*j2))); + } } } #else - for (vector::iterator j1 = cellList_[m1].begin(); j1 != cellList_[m1].end(); ++j1) { for (vector::iterator j2 = cellList_[m2].begin(); j2 != cellList_[m2].end(); ++j2) { - + // Always do this if we're in different cells or if - // we're in the same cell and the global index of the - // j2 cutoff group is less than the j1 cutoff group - - if (m2 != m1 || (*j2) < (*j1)) { + // we're in the same cell and the global index of + // the j2 cutoff group is greater than or equal to + // the j1 cutoff group. Note that Rappaport's code + // has a "less than" conditional here, but that + // deals with atom-by-atom computation. OpenMD + // allows atoms within a single cutoff group to + // interact with each other. + + + + if (m2 != m1 || (*j2) >= (*j1) ) { + dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)]; snap_->wrapVector(dr); cuts = getGroupCutoffs( (*j1), (*j2) ); @@ -1169,7 +1230,7 @@ namespace OpenMD { // branch to do all cutoff group pairs #ifdef IS_MPI for (int j1 = 0; j1 < nGroupsInRow_; j1++) { - for (int j2 = 0; j2 < nGroupsInCol_; j2++) { + for (int j2 = 0; j2 < nGroupsInCol_; j2++) { dr = cgColData.position[j2] - cgRowData.position[j1]; snap_->wrapVector(dr); cuts = getGroupCutoffs( j1, j2 ); @@ -1177,18 +1238,20 @@ namespace OpenMD { neighborList.push_back(make_pair(j1, j2)); } } - } + } #else - for (int j1 = 0; j1 < nGroups_ - 1; j1++) { - for (int j2 = j1 + 1; j2 < nGroups_; j2++) { + // include all groups here. + for (int j1 = 0; j1 < nGroups_; j1++) { + // include self group interactions j2 == j1 + for (int j2 = j1; j2 < nGroups_; j2++) { dr = snap_->cgData.position[j2] - snap_->cgData.position[j1]; snap_->wrapVector(dr); cuts = getGroupCutoffs( j1, j2 ); if (dr.lengthSquare() < cuts.third) { neighborList.push_back(make_pair(j1, j2)); } - } - } + } + } #endif }