--- branches/development/src/parallel/ForceMatrixDecomposition.cpp 2012/05/24 14:17:42 1721 +++ branches/development/src/parallel/ForceMatrixDecomposition.cpp 2012/06/22 20:01:37 1761 @@ -95,7 +95,7 @@ namespace OpenMD { storageLayout_ = sman_->getStorageLayout(); ff_ = info_->getForceField(); nLocal_ = snap_->getNumberOfAtoms(); - + nGroups_ = info_->getNLocalCutoffGroups(); // gather the information for atomtype IDs (atids): idents = info_->getIdentArray(); @@ -109,7 +109,13 @@ namespace OpenMD { PairList* oneTwo = info_->getOneTwoInteractions(); PairList* oneThree = info_->getOneThreeInteractions(); PairList* oneFour = info_->getOneFourInteractions(); - + + if (needVelocities_) + snap_->cgData.setStorageLayout(DataStorage::dslPosition | + DataStorage::dslVelocity); + else + snap_->cgData.setStorageLayout(DataStorage::dslPosition); + #ifdef IS_MPI MPI::Intracomm row = rowComm.getComm(); @@ -145,8 +151,13 @@ namespace OpenMD { cgRowData.resize(nGroupsInRow_); cgRowData.setStorageLayout(DataStorage::dslPosition); cgColData.resize(nGroupsInCol_); - cgColData.setStorageLayout(DataStorage::dslPosition); - + if (needVelocities_) + // we only need column velocities if we need them. + cgColData.setStorageLayout(DataStorage::dslPosition | + DataStorage::dslVelocity); + else + cgColData.setStorageLayout(DataStorage::dslPosition); + identsRow.resize(nAtomsInRow_); identsCol.resize(nAtomsInCol_); @@ -164,6 +175,9 @@ namespace OpenMD { pot_row.resize(nAtomsInRow_); pot_col.resize(nAtomsInCol_); + + expot_row.resize(nAtomsInRow_); + expot_col.resize(nAtomsInCol_); AtomRowToGlobal.resize(nAtomsInRow_); AtomColToGlobal.resize(nAtomsInCol_); @@ -450,7 +464,6 @@ namespace OpenMD { } } - groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) { int i, j; #ifdef IS_MPI @@ -474,6 +487,8 @@ namespace OpenMD { void ForceMatrixDecomposition::zeroWorkArrays() { pairwisePot = 0.0; embeddingPot = 0.0; + excludedPot = 0.0; + excludedSelfPot = 0.0; #ifdef IS_MPI if (storageLayout_ & DataStorage::dslForce) { @@ -492,6 +507,12 @@ namespace OpenMD { fill(pot_col.begin(), pot_col.end(), Vector (0.0)); + fill(expot_row.begin(), expot_row.end(), + Vector (0.0)); + + fill(expot_col.begin(), expot_col.end(), + Vector (0.0)); + if (storageLayout_ & DataStorage::dslParticlePot) { fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(), 0.0); @@ -600,6 +621,17 @@ namespace OpenMD { cgPlanVectorColumn->gather(snap_->cgData.position, cgColData.position); + + + if (needVelocities_) { + // gather up the atomic velocities + AtomPlanVectorColumn->gather(snap_->atomData.velocity, + atomColData.velocity); + + cgPlanVectorColumn->gather(snap_->cgData.velocity, + cgColData.velocity); + } + // if needed, gather the atomic rotation matrices if (storageLayout_ & DataStorage::dslAmat) { @@ -759,22 +791,82 @@ namespace OpenMD { vector pot_temp(nLocal_, Vector (0.0)); + vector expot_temp(nLocal_, + Vector (0.0)); // scatter/gather pot_row into the members of my column AtomPlanPotRow->scatter(pot_row, pot_temp); + AtomPlanPotRow->scatter(expot_row, expot_temp); - for (int ii = 0; ii < pot_temp.size(); ii++ ) + for (int ii = 0; ii < pot_temp.size(); ii++ ) pairwisePot += pot_temp[ii]; - + + for (int ii = 0; ii < expot_temp.size(); ii++ ) + excludedPot += expot_temp[ii]; + + if (storageLayout_ & DataStorage::dslParticlePot) { + // This is the pairwise contribution to the particle pot. The + // embedding contribution is added in each of the low level + // non-bonded routines. In single processor, this is done in + // unpackInteractionData, not in collectData. + for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { + for (int i = 0; i < nLocal_; i++) { + // factor of two is because the total potential terms are divided + // by 2 in parallel due to row/ column scatter + snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii); + } + } + } + fill(pot_temp.begin(), pot_temp.end(), Vector (0.0)); + fill(expot_temp.begin(), expot_temp.end(), + Vector (0.0)); AtomPlanPotColumn->scatter(pot_col, pot_temp); + AtomPlanPotColumn->scatter(expot_col, expot_temp); for (int ii = 0; ii < pot_temp.size(); ii++ ) pairwisePot += pot_temp[ii]; + + for (int ii = 0; ii < expot_temp.size(); ii++ ) + excludedPot += expot_temp[ii]; + + if (storageLayout_ & DataStorage::dslParticlePot) { + // This is the pairwise contribution to the particle pot. The + // embedding contribution is added in each of the low level + // non-bonded routines. In single processor, this is done in + // unpackInteractionData, not in collectData. + for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { + for (int i = 0; i < nLocal_; i++) { + // factor of two is because the total potential terms are divided + // by 2 in parallel due to row/ column scatter + snap_->atomData.particlePot[i] += 2.0 * pot_temp[i](ii); + } + } + } + if (storageLayout_ & DataStorage::dslParticlePot) { + int npp = snap_->atomData.particlePot.size(); + vector ppot_temp(npp, 0.0); + + // This is the direct or embedding contribution to the particle + // pot. + + AtomPlanRealRow->scatter(atomRowData.particlePot, ppot_temp); + for (int i = 0; i < npp; i++) { + snap_->atomData.particlePot[i] += ppot_temp[i]; + } + + fill(ppot_temp.begin(), ppot_temp.end(), 0.0); + + AtomPlanRealColumn->scatter(atomColData.particlePot, ppot_temp); + for (int i = 0; i < npp; i++) { + snap_->atomData.particlePot[i] += ppot_temp[i]; + } + } + for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { RealType ploc1 = pairwisePot[ii]; RealType ploc2 = 0.0; @@ -783,16 +875,51 @@ namespace OpenMD { } for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { - RealType ploc1 = embeddingPot[ii]; + RealType ploc1 = excludedPot[ii]; RealType ploc2 = 0.0; MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); - embeddingPot[ii] = ploc2; + excludedPot[ii] = ploc2; } + // Here be dragons. + MPI::Intracomm col = colComm.getComm(); + + col.Allreduce(MPI::IN_PLACE, + &snap_->frameData.conductiveHeatFlux[0], 3, + MPI::REALTYPE, MPI::SUM); + + #endif } + /** + * Collects information obtained during the post-pair (and embedding + * functional) loops onto local data structures. + */ + void ForceMatrixDecomposition::collectSelfData() { + snap_ = sman_->getCurrentSnapshot(); + storageLayout_ = sman_->getStorageLayout(); + +#ifdef IS_MPI + 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; + } + for (int ii = 0; ii < N_INTERACTION_FAMILIES; ii++) { + RealType ploc1 = excludedSelfPot[ii]; + RealType ploc2 = 0.0; + MPI::COMM_WORLD.Allreduce(&ploc1, &ploc2, 1, MPI::REALTYPE, MPI::SUM); + excludedSelfPot[ii] = ploc2; + } +#endif + + } + + + int ForceMatrixDecomposition::getNAtomsInRow() { #ifdef IS_MPI return nAtomsInRow_; @@ -831,9 +958,25 @@ namespace OpenMD { snap_->wrapVector(d); return d; + } + + Vector3d ForceMatrixDecomposition::getGroupVelocityColumn(int cg2){ +#ifdef IS_MPI + return cgColData.velocity[cg2]; +#else + return snap_->cgData.velocity[cg2]; +#endif } + Vector3d ForceMatrixDecomposition::getAtomVelocityColumn(int atom2){ +#ifdef IS_MPI + return atomColData.velocity[atom2]; +#else + return snap_->atomData.velocity[atom2]; +#endif + } + Vector3d ForceMatrixDecomposition::getAtomToGroupVectorRow(int atom1, int cg1){ Vector3d d; @@ -899,16 +1042,20 @@ namespace OpenMD { * We need to exclude some overcounted interactions that result from * the parallel decomposition. */ - bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2) { - int unique_id_1, unique_id_2; + bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1, int cg2) { + int unique_id_1, unique_id_2, group1, group2; #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]; + group1 = cgRowToGlobal[cg1]; + group2 = cgColToGlobal[cg2]; #else unique_id_1 = AtomLocalToGlobal[atom1]; unique_id_2 = AtomLocalToGlobal[atom2]; + group1 = cgLocalToGlobal[cg1]; + group2 = cgLocalToGlobal[cg2]; #endif if (unique_id_1 == unique_id_2) return true; @@ -919,6 +1066,12 @@ namespace OpenMD { if ((unique_id_1 + unique_id_2) % 2 == 0) return true; } else { if ((unique_id_1 + unique_id_2) % 2 == 1) return true; + } +#endif + +#ifndef IS_MPI + if (group1 == group2) { + if (unique_id_1 < unique_id_2) return true; } #endif @@ -1022,14 +1175,7 @@ namespace OpenMD { #else - - // cerr << "atoms = " << atom1 << " " << atom2 << "\n"; - // cerr << "pos1 = " << snap_->atomData.position[atom1] << "\n"; - // cerr << "pos2 = " << snap_->atomData.position[atom2] << "\n"; - idat.atypes = make_pair( atypesLocal[atom1], atypesLocal[atom2]); - //idat.atypes = make_pair( ff_->getAtomType(idents[atom1]), - // ff_->getAtomType(idents[atom2]) ); if (storageLayout_ & DataStorage::dslAmat) { idat.A1 = &(snap_->atomData.aMat[atom1]); @@ -1084,13 +1230,15 @@ namespace OpenMD { #ifdef IS_MPI pot_row[atom1] += RealType(0.5) * *(idat.pot); pot_col[atom2] += RealType(0.5) * *(idat.pot); + expot_row[atom1] += RealType(0.5) * *(idat.excludedPot); + expot_col[atom2] += RealType(0.5) * *(idat.excludedPot); atomRowData.force[atom1] += *(idat.f1); atomColData.force[atom2] -= *(idat.f1); if (storageLayout_ & DataStorage::dslFlucQForce) { - atomRowData.flucQFrc[atom1] += *(idat.dVdFQ1); - atomColData.flucQFrc[atom2] += *(idat.dVdFQ2); + atomRowData.flucQFrc[atom1] -= *(idat.dVdFQ1); + atomColData.flucQFrc[atom2] -= *(idat.dVdFQ2); } if (storageLayout_ & DataStorage::dslElectricField) { @@ -1098,20 +1246,24 @@ namespace OpenMD { atomColData.electricField[atom2] += *(idat.eField2); } - // should particle pot be done here also? #else pairwisePot += *(idat.pot); + excludedPot += *(idat.excludedPot); snap_->atomData.force[atom1] += *(idat.f1); snap_->atomData.force[atom2] -= *(idat.f1); if (idat.doParticlePot) { + // This is the pairwise contribution to the particle pot. The + // embedding contribution is added in each of the low level + // non-bonded routines. In parallel, this calculation is done + // in collectData, not in unpackInteractionData. snap_->atomData.particlePot[atom1] += *(idat.vpair) * *(idat.sw); - snap_->atomData.particlePot[atom2] -= *(idat.vpair) * *(idat.sw); + snap_->atomData.particlePot[atom2] += *(idat.vpair) * *(idat.sw); } if (storageLayout_ & DataStorage::dslFlucQForce) { - snap_->atomData.flucQFrc[atom1] += *(idat.dVdFQ1); + snap_->atomData.flucQFrc[atom1] -= *(idat.dVdFQ1); snap_->atomData.flucQFrc[atom2] -= *(idat.dVdFQ2); }