ForceMatrixDecomposition.cpp

/*
 * Copyright (c) 2004-present, The University of Notre Dame. All rights
 * reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice,
 *    this list of conditions and the following disclaimer.
 *
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * 3. Neither the name of the copyright holder nor the names of its
 *    contributors may be used to endorse or promote products derived from
 *    this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 *
 * SUPPORT OPEN SCIENCE!  If you use OpenMD or its source code in your
 * research, please cite the following paper when you publish your work:
 *
 * [1] Drisko et al., J. Open Source Softw. 9, 7004 (2024).
 *
 * Good starting points for code and simulation methodology are:
 *
 * [2] Meineke, et al., J. Comp. Chem. 26, 252-271 (2005).
 * [3] Fennell & Gezelter, J. Chem. Phys. 124, 234104 (2006).
 * [4] Sun, Lin & Gezelter, J. Chem. Phys. 128, 234107 (2008).
 * [5] Vardeman, Stocker & Gezelter, J. Chem. Theory Comput. 7, 834 (2011).
 * [6] Kuang & Gezelter, Mol. Phys., 110, 691-701 (2012).
 * [7] Lamichhane, Gezelter & Newman, J. Chem. Phys. 141, 134109 (2014).
 * [8] Bhattarai, Newman & Gezelter, Phys. Rev. B 99, 094106 (2019).
 * [9] Drisko & Gezelter, J. Chem. Theory Comput. 20, 4986-4997 (2024).
 */
#include "parallel/ForceMatrixDecomposition.hpp"
 
#include "brains/PairList.hpp"
#include "brains/SnapshotManager.hpp"
#include "math/SquareMatrix3.hpp"
#include "nonbonded/NonBondedInteraction.hpp"
 
using namespace std;
namespace OpenMD {
 
  ForceMatrixDecomposition::ForceMatrixDecomposition(SimInfo* info,
                                                     InteractionManager* iMan) :
      ForceDecomposition(info, iMan) {
    // Row and colum scans must visit all surrounding cells
    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));
  }
 
  ForceMatrixDecomposition::~ForceMatrixDecomposition() {
#ifdef IS_MPI
    delete AtomPlanIntRow;
    delete AtomPlanRealRow;
    delete AtomPlanVectorRow;
    delete AtomPlanMatrixRow;
    delete AtomPlanPotRow;
    delete AtomPlanIntColumn;
    delete AtomPlanRealColumn;
    delete AtomPlanVectorColumn;
    delete AtomPlanMatrixColumn;
    delete AtomPlanPotColumn;
    delete cgPlanIntRow;
    delete cgPlanVectorRow;
    delete cgPlanIntColumn;
    delete cgPlanVectorColumn;
#endif
  }
 
  /**
   * distributeInitialData is essentially a copy of the older fortran
   * SimulationSetup
   */
  void ForceMatrixDecomposition::distributeInitialData() {
    snap_              = sman_->getCurrentSnapshot();
    atomStorageLayout_ = sman_->getAtomStorageLayout();
    ff_                = info_->getForceField();
    nLocal_            = snap_->getNumberOfAtoms();
 
    nGroups_ = info_->getNLocalCutoffGroups();
    // gather the information for atomtype IDs (atids):
    idents                            = info_->getIdentArray();
    regions                           = info_->getRegions();
    AtomLocalToGlobal                 = info_->getGlobalAtomIndices();
    cgLocalToGlobal                   = info_->getGlobalGroupIndices();
    vector<int> globalGroupMembership = info_->getGlobalGroupMembership();
 
    massFactors = info_->getMassFactors();
 
    PairList* excludes = info_->getExcludedInteractions();
    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_Comm row = rowComm.getComm();
    MPI_Comm col = colComm.getComm();
 
    AtomPlanIntRow    = new Plan<int>(row, nLocal_);
    AtomPlanRealRow   = new Plan<RealType>(row, nLocal_);
    AtomPlanVectorRow = new Plan<Vector3d>(row, nLocal_);
    AtomPlanMatrixRow = new Plan<Mat3x3d>(row, nLocal_);
    AtomPlanPotRow    = new Plan<potVec>(row, nLocal_);
 
    AtomPlanIntColumn    = new Plan<int>(col, nLocal_);
    AtomPlanRealColumn   = new Plan<RealType>(col, nLocal_);
    AtomPlanVectorColumn = new Plan<Vector3d>(col, nLocal_);
    AtomPlanMatrixColumn = new Plan<Mat3x3d>(col, nLocal_);
    AtomPlanPotColumn    = new Plan<potVec>(col, nLocal_);
 
    cgPlanIntRow       = new Plan<int>(row, nGroups_);
    cgPlanVectorRow    = new Plan<Vector3d>(row, nGroups_);
    cgPlanIntColumn    = new Plan<int>(col, nGroups_);
    cgPlanVectorColumn = new Plan<Vector3d>(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(atomStorageLayout_);
    atomColData.resize(nAtomsInCol_);
    atomColData.setStorageLayout(atomStorageLayout_);
    cgRowData.resize(nGroupsInRow_);
    cgRowData.setStorageLayout(DataStorage::dslPosition);
    cgColData.resize(nGroupsInCol_);
    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_);
 
    AtomPlanIntRow->gather(idents, identsRow);
    AtomPlanIntColumn->gather(idents, identsCol);
 
    regionsRow.resize(nAtomsInRow_);
    regionsCol.resize(nAtomsInCol_);
 
    AtomPlanIntRow->gather(regions, regionsRow);
    AtomPlanIntColumn->gather(regions, regionsCol);
 
    // allocate memory for the parallel objects
    atypesRow.resize(nAtomsInRow_);
    atypesCol.resize(nAtomsInCol_);
 
    for (int i = 0; i < nAtomsInRow_; i++)
      atypesRow[i] = ff_->getAtomType(identsRow[i]);
    for (int i = 0; i < nAtomsInCol_; i++)
      atypesCol[i] = ff_->getAtomType(identsCol[i]);
 
    pot_row.resize(nAtomsInRow_);
    pot_col.resize(nAtomsInCol_);
 
    expot_row.resize(nAtomsInRow_);
    expot_col.resize(nAtomsInCol_);
 
    selepot_row.resize(nAtomsInRow_);
    selepot_col.resize(nAtomsInCol_);
 
    AtomRowToGlobal.resize(nAtomsInRow_);
    AtomColToGlobal.resize(nAtomsInCol_);
    AtomPlanIntRow->gather(AtomLocalToGlobal, AtomRowToGlobal);
    AtomPlanIntColumn->gather(AtomLocalToGlobal, AtomColToGlobal);
 
    cgRowToGlobal.resize(nGroupsInRow_);
    cgColToGlobal.resize(nGroupsInCol_);
    cgPlanIntRow->gather(cgLocalToGlobal, cgRowToGlobal);
    cgPlanIntColumn->gather(cgLocalToGlobal, cgColToGlobal);
 
    massFactorsRow.resize(nAtomsInRow_);
    massFactorsCol.resize(nAtomsInCol_);
    AtomPlanRealRow->gather(massFactors, massFactorsRow);
    AtomPlanRealColumn->gather(massFactors, massFactorsCol);
 
    groupListRow_.clear();
    groupListRow_.resize(nGroupsInRow_);
    for (int i = 0; i < nGroupsInRow_; i++) {
      int gid = cgRowToGlobal[i];
      for (int j = 0; j < nAtomsInRow_; j++) {
        int aid = AtomRowToGlobal[j];
        if (globalGroupMembership[aid] == gid) groupListRow_[i].push_back(j);
      }
    }
 
    groupListCol_.clear();
    groupListCol_.resize(nGroupsInCol_);
    for (int i = 0; i < nGroupsInCol_; i++) {
      int gid = cgColToGlobal[i];
      for (int j = 0; j < nAtomsInCol_; j++) {
        int aid = AtomColToGlobal[j];
        if (globalGroupMembership[aid] == gid) groupListCol_[i].push_back(j);
      }
    }
 
    excludesForAtom.clear();
    excludesForAtom.resize(nAtomsInRow_);
    toposForAtom.clear();
    toposForAtom.resize(nAtomsInRow_);
    topoDist.clear();
    topoDist.resize(nAtomsInRow_);
    for (int i = 0; i < nAtomsInRow_; i++) {
      int iglob = AtomRowToGlobal[i];
 
      for (int j = 0; j < nAtomsInCol_; j++) {
        int jglob = AtomColToGlobal[j];
 
        if (excludes->hasPair(iglob, jglob)) excludesForAtom[i].push_back(j);
 
        if (oneTwo->hasPair(iglob, jglob)) {
          toposForAtom[i].push_back(j);
          topoDist[i].push_back(1);
        } else {
          if (oneThree->hasPair(iglob, jglob)) {
            toposForAtom[i].push_back(j);
            topoDist[i].push_back(2);
          } else {
            if (oneFour->hasPair(iglob, jglob)) {
              toposForAtom[i].push_back(j);
              topoDist[i].push_back(3);
            }
          }
        }
      }
    }
 
#else
    excludesForAtom.clear();
    excludesForAtom.resize(nLocal_);
    toposForAtom.clear();
    toposForAtom.resize(nLocal_);
    topoDist.clear();
    topoDist.resize(nLocal_);
 
    for (int i = 0; i < nLocal_; i++) {
      int iglob = AtomLocalToGlobal[i];
 
      for (int j = 0; j < nLocal_; j++) {
        int jglob = AtomLocalToGlobal[j];
 
        if (excludes->hasPair(iglob, jglob)) excludesForAtom[i].push_back(j);
 
        if (oneTwo->hasPair(iglob, jglob)) {
          toposForAtom[i].push_back(j);
          topoDist[i].push_back(1);
        } else {
          if (oneThree->hasPair(iglob, jglob)) {
            toposForAtom[i].push_back(j);
            topoDist[i].push_back(2);
          } else {
            if (oneFour->hasPair(iglob, jglob)) {
              toposForAtom[i].push_back(j);
              topoDist[i].push_back(3);
            }
          }
        }
      }
    }
#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); }
      }
    }
  }
 
  int ForceMatrixDecomposition::getTopologicalDistance(int atom1, int atom2) {
    for (unsigned int j = 0; j < toposForAtom[atom1].size(); j++) {
      if (toposForAtom[atom1][j] == atom2) return topoDist[atom1][j];
    }
    return 0;
  }
 
  void ForceMatrixDecomposition::zeroWorkArrays() {
    pairwisePot     = 0.0;
    selfPot         = 0.0;
    excludedPot     = 0.0;
    excludedSelfPot = 0.0;
    selectedPot     = 0.0;
    selectedSelfPot = 0.0;
 
#ifdef IS_MPI
    if (atomStorageLayout_ & DataStorage::dslForce) {
      fill(atomRowData.force.begin(), atomRowData.force.end(), V3Zero);
      fill(atomColData.force.begin(), atomColData.force.end(), V3Zero);
    }
 
    if (atomStorageLayout_ & DataStorage::dslTorque) {
      fill(atomRowData.torque.begin(), atomRowData.torque.end(), V3Zero);
      fill(atomColData.torque.begin(), atomColData.torque.end(), V3Zero);
    }
 
    fill(pot_row.begin(), pot_row.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    fill(pot_col.begin(), pot_col.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    fill(expot_row.begin(), expot_row.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    fill(expot_col.begin(), expot_col.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    fill(selepot_row.begin(), selepot_row.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    fill(selepot_col.begin(), selepot_col.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    if (atomStorageLayout_ & DataStorage::dslParticlePot) {
      fill(atomRowData.particlePot.begin(), atomRowData.particlePot.end(), 0.0);
      fill(atomColData.particlePot.begin(), atomColData.particlePot.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslDensity) {
      fill(atomRowData.density.begin(), atomRowData.density.end(), 0.0);
      fill(atomColData.density.begin(), atomColData.density.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslFunctional) {
      fill(atomRowData.functional.begin(), atomRowData.functional.end(), 0.0);
      fill(atomColData.functional.begin(), atomColData.functional.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
      fill(atomRowData.functionalDerivative.begin(),
           atomRowData.functionalDerivative.end(), 0.0);
      fill(atomColData.functionalDerivative.begin(),
           atomColData.functionalDerivative.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
      fill(atomRowData.skippedCharge.begin(), atomRowData.skippedCharge.end(),
           0.0);
      fill(atomColData.skippedCharge.begin(), atomColData.skippedCharge.end(),
           0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslFlucQForce) {
      fill(atomRowData.flucQFrc.begin(), atomRowData.flucQFrc.end(), 0.0);
      fill(atomColData.flucQFrc.begin(), atomColData.flucQFrc.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslElectricField) {
      fill(atomRowData.electricField.begin(), atomRowData.electricField.end(),
           V3Zero);
      fill(atomColData.electricField.begin(), atomColData.electricField.end(),
           V3Zero);
    }
 
    if (atomStorageLayout_ & DataStorage::dslSitePotential) {
      fill(atomRowData.sitePotential.begin(), atomRowData.sitePotential.end(),
           0.0);
      fill(atomColData.sitePotential.begin(), atomColData.sitePotential.end(),
           0.0);
    }
 
#endif
    // even in parallel, we need to zero out the local arrays:
 
    if (atomStorageLayout_ & DataStorage::dslParticlePot) {
      fill(snap_->atomData.particlePot.begin(),
           snap_->atomData.particlePot.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslDensity) {
      fill(snap_->atomData.density.begin(), snap_->atomData.density.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslFunctional) {
      fill(snap_->atomData.functional.begin(), snap_->atomData.functional.end(),
           0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
      fill(snap_->atomData.functionalDerivative.begin(),
           snap_->atomData.functionalDerivative.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
      fill(snap_->atomData.skippedCharge.begin(),
           snap_->atomData.skippedCharge.end(), 0.0);
    }
 
    if (atomStorageLayout_ & DataStorage::dslElectricField) {
      fill(snap_->atomData.electricField.begin(),
           snap_->atomData.electricField.end(), V3Zero);
    }
    if (atomStorageLayout_ & DataStorage::dslSitePotential) {
      fill(snap_->atomData.sitePotential.begin(),
           snap_->atomData.sitePotential.end(), 0.0);
    }
  }
 
  void ForceMatrixDecomposition::distributeData() {
#ifdef IS_MPI
 
    snap_              = sman_->getCurrentSnapshot();
    atomStorageLayout_ = sman_->getAtomStorageLayout();
 
    bool needsCG = true;
    if (info_->getNCutoffGroups() != info_->getNAtoms()) needsCG = false;
 
    // gather up the atomic positions
    AtomPlanVectorRow->gather(snap_->atomData.position, atomRowData.position);
    AtomPlanVectorColumn->gather(snap_->atomData.position,
                                 atomColData.position);
 
    // gather up the cutoff group positions
 
    if (needsCG) {
      cgPlanVectorRow->gather(snap_->cgData.position, cgRowData.position);
 
      cgPlanVectorColumn->gather(snap_->cgData.position, cgColData.position);
    }
 
    if (needVelocities_) {
      // gather up the atomic velocities
      AtomPlanVectorColumn->gather(snap_->atomData.velocity,
                                   atomColData.velocity);
 
      if (needsCG) {
        cgPlanVectorColumn->gather(snap_->cgData.velocity, cgColData.velocity);
      }
    }
 
    // if needed, gather the atomic rotation matrices
    if (atomStorageLayout_ & DataStorage::dslAmat) {
      AtomPlanMatrixRow->gather(snap_->atomData.aMat, atomRowData.aMat);
      AtomPlanMatrixColumn->gather(snap_->atomData.aMat, atomColData.aMat);
    }
 
    // if needed, gather the atomic eletrostatic information
    if (atomStorageLayout_ & DataStorage::dslDipole) {
      AtomPlanVectorRow->gather(snap_->atomData.dipole, atomRowData.dipole);
      AtomPlanVectorColumn->gather(snap_->atomData.dipole, atomColData.dipole);
    }
 
    if (atomStorageLayout_ & DataStorage::dslQuadrupole) {
      AtomPlanMatrixRow->gather(snap_->atomData.quadrupole,
                                atomRowData.quadrupole);
      AtomPlanMatrixColumn->gather(snap_->atomData.quadrupole,
                                   atomColData.quadrupole);
    }
 
    // if needed, gather the atomic fluctuating charge values
    if (atomStorageLayout_ & DataStorage::dslFlucQPosition) {
      AtomPlanRealRow->gather(snap_->atomData.flucQPos, atomRowData.flucQPos);
      AtomPlanRealColumn->gather(snap_->atomData.flucQPos,
                                 atomColData.flucQPos);
    }
 
#endif
  }
 
  /* collects information obtained during the pre-pair loop onto local
   * data structures.
   */
  void ForceMatrixDecomposition::collectIntermediateData() {
#ifdef IS_MPI
 
    snap_              = sman_->getCurrentSnapshot();
    atomStorageLayout_ = sman_->getAtomStorageLayout();
 
    if (atomStorageLayout_ & DataStorage::dslDensity) {
      AtomPlanRealRow->scatter(atomRowData.density, snap_->atomData.density);
 
      int n = snap_->atomData.density.size();
      vector<RealType> rho_tmp(n, 0.0);
      AtomPlanRealColumn->scatter(atomColData.density, rho_tmp);
      for (int i = 0; i < n; i++)
        snap_->atomData.density[i] += rho_tmp[i];
    }
 
    // this isn't necessary if we don't have polarizable atoms, but
    // we'll leave it here for now.
    if (atomStorageLayout_ & DataStorage::dslElectricField) {
      AtomPlanVectorRow->scatter(atomRowData.electricField,
                                 snap_->atomData.electricField);
 
      int n = snap_->atomData.electricField.size();
      vector<Vector3d> field_tmp(n, V3Zero);
      AtomPlanVectorColumn->scatter(atomColData.electricField, field_tmp);
      for (int i = 0; i < n; i++)
        snap_->atomData.electricField[i] += field_tmp[i];
    }
#endif
  }
 
  /*
   * redistributes information obtained during the pre-pair loop out to
   * row and column-indexed data structures
   */
  void ForceMatrixDecomposition::distributeIntermediateData() {
#ifdef IS_MPI
    snap_              = sman_->getCurrentSnapshot();
    atomStorageLayout_ = sman_->getAtomStorageLayout();
 
    if (atomStorageLayout_ & DataStorage::dslFunctional) {
      AtomPlanRealRow->gather(snap_->atomData.functional,
                              atomRowData.functional);
      AtomPlanRealColumn->gather(snap_->atomData.functional,
                                 atomColData.functional);
    }
 
    if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
      AtomPlanRealRow->gather(snap_->atomData.functionalDerivative,
                              atomRowData.functionalDerivative);
      AtomPlanRealColumn->gather(snap_->atomData.functionalDerivative,
                                 atomColData.functionalDerivative);
    }
#endif
  }
 
  void ForceMatrixDecomposition::collectData() {
#ifdef IS_MPI
    snap_              = sman_->getCurrentSnapshot();
    atomStorageLayout_ = sman_->getAtomStorageLayout();
 
    int n = snap_->atomData.force.size();
    vector<Vector3d> frc_tmp(n, V3Zero);
 
    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;
    }
 
    AtomPlanVectorColumn->scatter(atomColData.force, frc_tmp);
    for (int i = 0; i < n; i++) {
      snap_->atomData.force[i] += frc_tmp[i];
    }
 
    if (atomStorageLayout_ & DataStorage::dslTorque) {
      int nt = snap_->atomData.torque.size();
      vector<Vector3d> trq_tmp(nt, V3Zero);
 
      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;
      }
 
      AtomPlanVectorColumn->scatter(atomColData.torque, trq_tmp);
      for (int i = 0; i < nt; i++)
        snap_->atomData.torque[i] += trq_tmp[i];
    }
 
    if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
      int ns = snap_->atomData.skippedCharge.size();
      vector<RealType> skch_tmp(ns, 0.0);
 
      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;
      }
 
      AtomPlanRealColumn->scatter(atomColData.skippedCharge, skch_tmp);
      for (int i = 0; i < ns; i++)
        snap_->atomData.skippedCharge[i] += skch_tmp[i];
    }
 
    if (atomStorageLayout_ & DataStorage::dslFlucQForce) {
      int nq = snap_->atomData.flucQFrc.size();
      vector<RealType> fqfrc_tmp(nq, 0.0);
 
      AtomPlanRealRow->scatter(atomRowData.flucQFrc, fqfrc_tmp);
      for (int i = 0; i < nq; i++) {
        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
        fqfrc_tmp[i] = 0.0;
      }
 
      AtomPlanRealColumn->scatter(atomColData.flucQFrc, fqfrc_tmp);
      for (int i = 0; i < nq; i++)
        snap_->atomData.flucQFrc[i] += fqfrc_tmp[i];
    }
 
    if (atomStorageLayout_ & DataStorage::dslElectricField) {
      int nef = snap_->atomData.electricField.size();
      vector<Vector3d> efield_tmp(nef, V3Zero);
 
      AtomPlanVectorRow->scatter(atomRowData.electricField, efield_tmp);
      for (int i = 0; i < nef; i++) {
        snap_->atomData.electricField[i] += efield_tmp[i];
        efield_tmp[i] = 0.0;
      }
 
      AtomPlanVectorColumn->scatter(atomColData.electricField, efield_tmp);
      for (int i = 0; i < nef; i++)
        snap_->atomData.electricField[i] += efield_tmp[i];
    }
 
    if (atomStorageLayout_ & DataStorage::dslSitePotential) {
      int nsp = snap_->atomData.sitePotential.size();
      vector<RealType> sp_tmp(nsp, 0.0);
 
      AtomPlanRealRow->scatter(atomRowData.sitePotential, sp_tmp);
      for (int i = 0; i < nsp; i++) {
        snap_->atomData.sitePotential[i] += sp_tmp[i];
        sp_tmp[i] = 0.0;
      }
 
      AtomPlanRealColumn->scatter(atomColData.sitePotential, sp_tmp);
      for (int i = 0; i < nsp; i++)
        snap_->atomData.sitePotential[i] += sp_tmp[i];
    }
 
    nLocal_ = snap_->getNumberOfAtoms();
 
    vector<potVec> pot_temp(nLocal_,
                            Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
    vector<potVec> expot_temp(nLocal_,
                              Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
    vector<potVec> selepot_temp(nLocal_,
                                Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    // scatter/gather pot_row into the members of my column
 
    AtomPlanPotRow->scatter(pot_row, pot_temp);
    AtomPlanPotRow->scatter(expot_row, expot_temp);
    AtomPlanPotRow->scatter(selepot_row, selepot_temp);
 
    for (std::size_t ii = 0; ii < pot_temp.size(); ii++)
      pairwisePot += pot_temp[ii];
 
    for (std::size_t ii = 0; ii < expot_temp.size(); ii++)
      excludedPot += expot_temp[ii];
 
    for (std::size_t ii = 0; ii < selepot_temp.size(); ii++)
      selectedPot += selepot_temp[ii];
 
    if (atomStorageLayout_ & 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<RealType, N_INTERACTION_FAMILIES>(0.0));
    fill(expot_temp.begin(), expot_temp.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
    fill(selepot_temp.begin(), selepot_temp.end(),
         Vector<RealType, N_INTERACTION_FAMILIES>(0.0));
 
    AtomPlanPotColumn->scatter(pot_col, pot_temp);
    AtomPlanPotColumn->scatter(expot_col, expot_temp);
    AtomPlanPotColumn->scatter(selepot_col, selepot_temp);
 
    for (std::size_t ii = 0; ii < pot_temp.size(); ii++)
      pairwisePot += pot_temp[ii];
 
    for (std::size_t ii = 0; ii < expot_temp.size(); ii++)
      excludedPot += expot_temp[ii];
 
    for (std::size_t ii = 0; ii < selepot_temp.size(); ii++)
      selectedPot += selepot_temp[ii];
 
    if (atomStorageLayout_ & 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 (atomStorageLayout_ & DataStorage::dslParticlePot) {
      int npp = snap_->atomData.particlePot.size();
      vector<RealType> 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];
      }
    }
 
    MPI_Allreduce(MPI_IN_PLACE, &pairwisePot[0], N_INTERACTION_FAMILIES,
                  MPI_REALTYPE, MPI_SUM, MPI_COMM_WORLD);
 
    MPI_Allreduce(MPI_IN_PLACE, &excludedPot[0], N_INTERACTION_FAMILIES,
                  MPI_REALTYPE, MPI_SUM, MPI_COMM_WORLD);
 
    MPI_Allreduce(MPI_IN_PLACE, &selectedPot[0], N_INTERACTION_FAMILIES,
                  MPI_REALTYPE, MPI_SUM, MPI_COMM_WORLD);
 
    // Here be dragons.
    MPI_Comm col = colComm.getComm();
 
    MPI_Allreduce(MPI_IN_PLACE, &snap_->frameData.conductiveHeatFlux[0], 3,
                  MPI_REALTYPE, MPI_SUM, col);
#endif
  }
 
  /**
   * Collects information obtained during the post-pair (and embedding
   * functional) loops onto local data structures.
   */
  void ForceMatrixDecomposition::collectSelfData() {
#ifdef IS_MPI
    MPI_Allreduce(MPI_IN_PLACE, &selfPot[0], N_INTERACTION_FAMILIES,
                  MPI_REALTYPE, MPI_SUM, MPI_COMM_WORLD);
 
    MPI_Allreduce(MPI_IN_PLACE, &excludedSelfPot[0], N_INTERACTION_FAMILIES,
                  MPI_REALTYPE, MPI_SUM, MPI_COMM_WORLD);
 
    MPI_Allreduce(MPI_IN_PLACE, &selectedSelfPot[0], N_INTERACTION_FAMILIES,
                  MPI_REALTYPE, MPI_SUM, MPI_COMM_WORLD);
#endif
  }
 
  int& ForceMatrixDecomposition::getNAtomsInRow() {
#ifdef IS_MPI
    return nAtomsInRow_;
#else
    return nLocal_;
#endif
  }
 
  /**
   * returns the list of atoms belonging to this group.
   */
  vector<int>& ForceMatrixDecomposition::getAtomsInGroupRow(int cg1) {
#ifdef IS_MPI
    return groupListRow_[cg1];
#else
    return groupList_[cg1];
#endif
  }
 
  vector<int>& ForceMatrixDecomposition::getAtomsInGroupColumn(int cg2) {
#ifdef IS_MPI
    return groupListCol_[cg2];
#else
    return groupList_[cg2];
#endif
  }
 
  Vector3d ForceMatrixDecomposition::getIntergroupVector(int cg1, int cg2) {
    Vector3d d;
#ifdef IS_MPI
    d = cgColData.position[cg2] - cgRowData.position[cg1];
#else
    d = snap_->cgData.position[cg2] - snap_->cgData.position[cg1];
#endif
 
    if (usePeriodicBoundaryConditions_) { 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;
 
#ifdef IS_MPI
    d = cgRowData.position[cg1] - atomRowData.position[atom1];
#else
    d = snap_->cgData.position[cg1] - snap_->atomData.position[atom1];
#endif
    if (usePeriodicBoundaryConditions_) { snap_->wrapVector(d); }
    return d;
  }
 
  Vector3d ForceMatrixDecomposition::getAtomToGroupVectorColumn(int atom2,
                                                                int cg2) {
    Vector3d d;
 
#ifdef IS_MPI
    d = cgColData.position[cg2] - atomColData.position[atom2];
#else
    d = snap_->cgData.position[cg2] - snap_->atomData.position[atom2];
#endif
    if (usePeriodicBoundaryConditions_) { snap_->wrapVector(d); }
    return d;
  }
 
  RealType& ForceMatrixDecomposition::getMassFactorRow(int atom1) {
#ifdef IS_MPI
    return massFactorsRow[atom1];
#else
    return massFactors[atom1];
#endif
  }
 
  RealType& ForceMatrixDecomposition::getMassFactorColumn(int atom2) {
#ifdef IS_MPI
    return massFactorsCol[atom2];
#else
    return massFactors[atom2];
#endif
  }
 
  Vector3d ForceMatrixDecomposition::getInteratomicVector(int atom1,
                                                          int atom2) {
    Vector3d d;
 
#ifdef IS_MPI
    d = atomColData.position[atom2] - atomRowData.position[atom1];
#else
    d = snap_->atomData.position[atom2] - snap_->atomData.position[atom1];
#endif
    if (usePeriodicBoundaryConditions_) { snap_->wrapVector(d); }
    return d;
  }
 
  vector<int>& ForceMatrixDecomposition::getExcludesForAtom(int atom1) {
    return excludesForAtom[atom1];
  }
 
  /**
   * We need to exclude some overcounted interactions that result from
   * the parallel decomposition.
   */
#ifdef IS_MPI
  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int, int) {
    // in MPI, we have to look up the unique IDs for each atom
    int unique_id_1 = AtomRowToGlobal[atom1];
    int unique_id_2 = AtomColToGlobal[atom2];
 
    if (unique_id_1 == unique_id_2) return true;
 
    // 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;
    }
 
    return false;
  }
 
#else
  bool ForceMatrixDecomposition::skipAtomPair(int atom1, int atom2, int cg1,
                                              int cg2) {
    int unique_id_1 = AtomLocalToGlobal[atom1];
    int unique_id_2 = AtomLocalToGlobal[atom2];
    int group1      = cgLocalToGlobal[cg1];
    int group2      = cgLocalToGlobal[cg2];
 
    if (unique_id_1 == unique_id_2) return true;
 
    if (group1 == group2) {
      if (unique_id_1 < unique_id_2) return true;
    }
 
    return false;
  }
#endif
 
  /**
   * We need to handle the interactions for atoms who are involved in
   * the same rigid body as well as some short range interactions
   * (bonds, bends, torsions) differently from other interactions.
   * We'll still visit the pairwise routines, but with a flag that
   * tells those routines to exclude the pair from direct long range
   * interactions.  Some indirect interactions (notably reaction
   * field) must still be handled for these pairs.
   */
  bool ForceMatrixDecomposition::excludeAtomPair(int atom1, int atom2) {
    // excludesForAtom was constructed to use row/column indices in the MPI
    // version, and to use local IDs in the non-MPI version:
 
    for (vector<int>::iterator i = excludesForAtom[atom1].begin();
         i != excludesForAtom[atom1].end(); ++i) {
      if ((*i) == atom2) return true;
    }
 
    return false;
  }
 
  void ForceMatrixDecomposition::addForceToAtomRow(int atom1, Vector3d fg) {
#ifdef IS_MPI
    atomRowData.force[atom1] += fg;
#else
    snap_->atomData.force[atom1] += fg;
#endif
  }
 
  void ForceMatrixDecomposition::addForceToAtomColumn(int atom2, Vector3d fg) {
#ifdef IS_MPI
    atomColData.force[atom2] += fg;
#else
    snap_->atomData.force[atom2] += fg;
#endif
  }
 
  // filling interaction blocks with pointers
  void ForceMatrixDecomposition::fillInteractionData(InteractionData& idat,
                                                     int atom1, int atom2,
                                                     bool newAtom1) {
    idat.excluded = excludeAtomPair(atom1, atom2);
 
    if (newAtom1) {
#ifdef IS_MPI
      idat.atid1 = identsRow[atom1];
      idat.atid2 = identsCol[atom2];
 
      if (regionsRow[atom1] >= 0 && regionsCol[atom2] >= 0) {
        idat.sameRegion = (regionsRow[atom1] == regionsCol[atom2]);
      } else {
        idat.sameRegion = false;
      }
 
      if (atomStorageLayout_ & DataStorage::dslAmat) {
        idat.A1 = atomRowData.aMat[atom1];
        idat.A2 = atomColData.aMat[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslTorque) {
        idat.t1 = atomRowData.torque[atom1];
        idat.t2 = atomColData.torque[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDipole) {
        idat.D_1 = atomRowData.dipole[atom1];
        idat.D_2 = atomColData.dipole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslQuadrupole) {
        idat.Q_1 = atomRowData.quadrupole[atom1];
        idat.Q_2 = atomColData.quadrupole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDensity) {
        idat.rho1 = atomRowData.density[atom1];
        idat.rho2 = atomColData.density[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctional) {
        idat.frho1 = atomRowData.functional[atom1];
        idat.frho2 = atomColData.functional[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
        idat.dfrho1 = atomRowData.functionalDerivative[atom1];
        idat.dfrho2 = atomColData.functionalDerivative[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslParticlePot) {
        idat.particlePot1 = atomRowData.particlePot[atom1];
        idat.particlePot2 = atomColData.particlePot[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
        idat.skippedCharge1 = atomRowData.skippedCharge[atom1];
        idat.skippedCharge2 = atomColData.skippedCharge[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFlucQPosition) {
        idat.flucQ1 = atomRowData.flucQPos[atom1];
        idat.flucQ2 = atomColData.flucQPos[atom2];
      }
 
#else
 
      idat.atid1 = idents[atom1];
      idat.atid2 = idents[atom2];
 
      if (regions[atom1] >= 0 && regions[atom2] >= 0) {
        idat.sameRegion = (regions[atom1] == regions[atom2]);
      } else {
        idat.sameRegion = false;
      }
 
      if (atomStorageLayout_ & DataStorage::dslAmat) {
        idat.A1 = snap_->atomData.aMat[atom1];
        idat.A2 = snap_->atomData.aMat[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslTorque) {
        idat.t1 = snap_->atomData.torque[atom1];
        idat.t2 = snap_->atomData.torque[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDipole) {
        idat.D_1 = snap_->atomData.dipole[atom1];
        idat.D_2 = snap_->atomData.dipole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslQuadrupole) {
        idat.Q_1 = snap_->atomData.quadrupole[atom1];
        idat.Q_2 = snap_->atomData.quadrupole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDensity) {
        idat.rho1 = snap_->atomData.density[atom1];
        idat.rho2 = snap_->atomData.density[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctional) {
        idat.frho1 = snap_->atomData.functional[atom1];
        idat.frho2 = snap_->atomData.functional[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
        idat.dfrho1 = snap_->atomData.functionalDerivative[atom1];
        idat.dfrho2 = snap_->atomData.functionalDerivative[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslParticlePot) {
        idat.particlePot1 = snap_->atomData.particlePot[atom1];
        idat.particlePot2 = snap_->atomData.particlePot[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
        idat.skippedCharge1 = snap_->atomData.skippedCharge[atom1];
        idat.skippedCharge2 = snap_->atomData.skippedCharge[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFlucQPosition) {
        idat.flucQ1 = snap_->atomData.flucQPos[atom1];
        idat.flucQ2 = snap_->atomData.flucQPos[atom2];
      }
#endif
 
    } else {
      // atom1 is not new, so don't bother updating properties of that atom:
#ifdef IS_MPI
      idat.atid2 = identsCol[atom2];
 
      if (regionsRow[atom1] >= 0 && regionsCol[atom2] >= 0) {
        idat.sameRegion = (regionsRow[atom1] == regionsCol[atom2]);
      } else {
        idat.sameRegion = false;
      }
 
      if (atomStorageLayout_ & DataStorage::dslAmat) {
        idat.A2 = atomColData.aMat[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslTorque) {
        idat.t2 = atomColData.torque[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDipole) {
        idat.D_2 = atomColData.dipole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslQuadrupole) {
        idat.Q_2 = atomColData.quadrupole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDensity) {
        idat.rho2 = atomColData.density[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctional) {
        idat.frho2 = atomColData.functional[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
        idat.dfrho2 = atomColData.functionalDerivative[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslParticlePot) {
        idat.particlePot2 = atomColData.particlePot[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
        idat.skippedCharge2 = atomColData.skippedCharge[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFlucQPosition) {
        idat.flucQ2 = atomColData.flucQPos[atom2];
      }
 
#else
      idat.atid2 = idents[atom2];
 
      if (regions[atom1] >= 0 && regions[atom2] >= 0) {
        idat.sameRegion = (regions[atom1] == regions[atom2]);
      } else {
        idat.sameRegion = false;
      }
 
      if (atomStorageLayout_ & DataStorage::dslAmat) {
        idat.A2 = snap_->atomData.aMat[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslTorque) {
        idat.t2 = snap_->atomData.torque[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDipole) {
        idat.D_2 = snap_->atomData.dipole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslQuadrupole) {
        idat.Q_2 = snap_->atomData.quadrupole[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslDensity) {
        idat.rho2 = snap_->atomData.density[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctional) {
        idat.frho2 = snap_->atomData.functional[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFunctionalDerivative) {
        idat.dfrho2 = snap_->atomData.functionalDerivative[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslParticlePot) {
        idat.particlePot2 = snap_->atomData.particlePot[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
        idat.skippedCharge2 = snap_->atomData.skippedCharge[atom2];
      }
 
      if (atomStorageLayout_ & DataStorage::dslFlucQPosition) {
        idat.flucQ2 = snap_->atomData.flucQPos[atom2];
      }
 
#endif
    }
  }
 
  void ForceMatrixDecomposition::unpackInteractionData(InteractionData& idat,
                                                       int atom1, int atom2) {
#ifdef IS_MPI
    pot_row[atom1] += 0.5 * idat.pot;
    pot_col[atom2] += 0.5 * idat.pot;
    expot_row[atom1] += 0.5 * idat.excludedPot;
    expot_col[atom2] += 0.5 * idat.excludedPot;
    selepot_row[atom1] += 0.5 * idat.selePot;
    selepot_col[atom2] += 0.5 * idat.selePot;
 
    atomRowData.force[atom1] += idat.f1;
    atomColData.force[atom2] -= idat.f1;
 
    if (atomStorageLayout_ & DataStorage::dslFlucQForce) {
      atomRowData.flucQFrc[atom1] -= idat.dVdFQ1;
      atomColData.flucQFrc[atom2] -= idat.dVdFQ2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslElectricField) {
      atomRowData.electricField[atom1] += idat.eField1;
      atomColData.electricField[atom2] += idat.eField2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslSitePotential) {
      atomRowData.sitePotential[atom1] += idat.sPot1;
      atomColData.sitePotential[atom2] += idat.sPot2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslTorque) {
      atomRowData.torque[atom1] = idat.t1;
      atomColData.torque[atom2] = idat.t2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
      atomRowData.skippedCharge[atom1] = idat.skippedCharge1;
      atomColData.skippedCharge[atom2] = idat.skippedCharge2;
    }
 
#else
    pairwisePot += idat.pot;
    excludedPot += idat.excludedPot;
    selectedPot += idat.selePot;
 
    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
      // self and 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;
    }
 
    if (atomStorageLayout_ & DataStorage::dslFlucQForce) {
      snap_->atomData.flucQFrc[atom1] -= idat.dVdFQ1;
      snap_->atomData.flucQFrc[atom2] -= idat.dVdFQ2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslElectricField) {
      snap_->atomData.electricField[atom1] += idat.eField1;
      snap_->atomData.electricField[atom2] += idat.eField2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslSitePotential) {
      snap_->atomData.sitePotential[atom1] += idat.sPot1;
      snap_->atomData.sitePotential[atom2] += idat.sPot2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslTorque) {
      snap_->atomData.torque[atom1] = idat.t1;
      snap_->atomData.torque[atom2] = idat.t2;
    }
 
    if (atomStorageLayout_ & DataStorage::dslSkippedCharge) {
      snap_->atomData.skippedCharge[atom1] = idat.skippedCharge1;
      snap_->atomData.skippedCharge[atom2] = idat.skippedCharge2;
    }
 
#endif
  }
  void ForceMatrixDecomposition::unpackPrePairData(InteractionData& idat,
                                                   int atom1, int atom2) {
#ifdef IS_MPI
 
    if (atomStorageLayout_ & DataStorage::dslDensity) {
      atomRowData.density[atom1] = idat.rho1;
      atomColData.density[atom2] = idat.rho2;
    }
 
#else
 
    if (atomStorageLayout_ & DataStorage::dslDensity) {
      snap_->atomData.density[atom1] = idat.rho1;
      snap_->atomData.density[atom2] = idat.rho2;
    }
 
#endif
  }
 
  /*
   * buildNeighborList
   *
   * Constructs the Verlet neighbor list for a force-matrix
   * decomposition.  In this case, each processor is responsible for
   * row-site interactions with column-sites.
   *
   * neighborList is returned as a packed array of neighboring
   * column-ordered CutoffGroups.  The starting position in
   * neighborList for each row-ordered CutoffGroup is given by the
   * returned vector point.
   */
  void ForceMatrixDecomposition::buildNeighborList(
      vector<int>& neighborList, vector<int>& point,
      vector<Vector3d>& savedPositions) {
    neighborList.clear();
    point.clear();
    int len = 0;
 
    bool doAllPairs = false;
 
    Snapshot* snap_ = sman_->getCurrentSnapshot();
    Mat3x3d box;
    Mat3x3d invBox;
 
    Vector3d rs, scaled, dr;
    Vector3i whichCell;
    int cellIndex;
 
#ifdef IS_MPI
    cellListRow_.clear();
    cellListCol_.clear();
    point.resize(nGroupsInRow_ + 1);
#else
    cellList_.clear();
    point.resize(nGroups_ + 1);
#endif
 
    if (!usePeriodicBoundaryConditions_) {
      box    = snap_->getBoundingBox();
      invBox = snap_->getInvBoundingBox();
    } else {
      box    = snap_->getHmat();
      invBox = snap_->getInvHmat();
    }
 
    Vector3d A = box.getColumn(0);
    Vector3d B = box.getColumn(1);
    Vector3d C = box.getColumn(2);
 
    // Required for triclinic cells
    Vector3d AxB = cross(A, B);
    Vector3d BxC = cross(B, C);
    Vector3d CxA = cross(C, A);
 
    // unit vectors perpendicular to the faces of the triclinic cell:
    AxB.normalize();
    BxC.normalize();
    CxA.normalize();
 
    // A set of perpendicular lengths in triclinic cells:
    RealType Wa = abs(dot(A, BxC));
    RealType Wb = abs(dot(B, CxA));
    RealType Wc = abs(dot(C, AxB));
 
    nCells_.x() = int(Wa / rList_);
    nCells_.y() = int(Wb / rList_);
    nCells_.z() = int(Wc / rList_);
 
    // handle small boxes where the cell offsets can end up repeating cells
    if (nCells_.x() < 3) doAllPairs = true;
    if (nCells_.y() < 3) doAllPairs = true;
    if (nCells_.z() < 3) doAllPairs = true;
 
    int nCtot = nCells_.x() * nCells_.y() * nCells_.z();
 
#ifdef IS_MPI
    cellListRow_.resize(nCtot);
    cellListCol_.resize(nCtot);
#else
    cellList_.resize(nCtot);
#endif
 
    if (!doAllPairs) {
#ifdef IS_MPI
 
      for (int i = 0; i < nGroupsInRow_; i++) {
        rs = cgRowData.position[i];
 
        // scaled positions relative to the box vectors
        scaled = invBox * rs;
 
        // wrap the vector back into the unit box by subtracting integer box
        // numbers
        for (int j = 0; j < 3; j++) {
          scaled[j] -= roundMe(scaled[j]);
          scaled[j] += 0.5;
          // Handle the special case when an object is exactly on the
          // boundary (a scaled coordinate of 1.0 is the same as
          // scaled coordinate of 0.0)
          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
        }
 
        // find xyz-indices of cell that cutoffGroup is in.
        whichCell.x() = int(nCells_.x() * scaled.x());
        whichCell.y() = int(nCells_.y() * scaled.y());
        whichCell.z() = int(nCells_.z() * scaled.z());
 
        // find single index of this cell:
        cellIndex = Vlinear(whichCell, nCells_);
 
        // 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];
 
        // scaled positions relative to the box vectors
        scaled = invBox * rs;
 
        // wrap the vector back into the unit box by subtracting integer box
        // numbers
        for (int j = 0; j < 3; j++) {
          scaled[j] -= roundMe(scaled[j]);
          scaled[j] += 0.5;
          // Handle the special case when an object is exactly on the
          // boundary (a scaled coordinate of 1.0 is the same as
          // scaled coordinate of 0.0)
          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
        }
 
        // find xyz-indices of cell that cutoffGroup is in.
        whichCell.x() = int(nCells_.x() * scaled.x());
        whichCell.y() = int(nCells_.y() * scaled.y());
        whichCell.z() = int(nCells_.z() * scaled.z());
 
        // find single index of this cell:
        cellIndex = Vlinear(whichCell, nCells_);
 
        // 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];
 
        // scaled positions relative to the box vectors
        scaled = invBox * rs;
 
        // wrap the vector back into the unit box by subtracting integer box
        // numbers
        for (int j = 0; j < 3; j++) {
          scaled[j] -= roundMe(scaled[j]);
          scaled[j] += 0.5;
          // Handle the special case when an object is exactly on the
          // boundary (a scaled coordinate of 1.0 is the same as
          // scaled coordinate of 0.0)
          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
        }
 
        // find xyz-indices of cell that cutoffGroup is in.
        whichCell.x() = int(nCells_.x() * scaled.x());
        whichCell.y() = int(nCells_.y() * scaled.y());
        whichCell.z() = int(nCells_.z() * scaled.z());
 
        // find single index of this cell:
        cellIndex = Vlinear(whichCell, nCells_);
 
        // add this cutoff group to the list of groups in this cell;
        cellList_[cellIndex].push_back(i);
      }
 
#endif
 
#ifdef IS_MPI
      for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
        rs = cgRowData.position[j1];
#else
 
      for (int j1 = 0; j1 < nGroups_; j1++) {
        rs = snap_->cgData.position[j1];
#endif
        point[j1] = len;
 
        // scaled positions relative to the box vectors
        scaled = invBox * rs;
 
        // wrap the vector back into the unit box by subtracting integer box
        // numbers
        for (int j = 0; j < 3; j++) {
          scaled[j] -= roundMe(scaled[j]);
          scaled[j] += 0.5;
          // Handle the special case when an object is exactly on the
          // boundary (a scaled coordinate of 1.0 is the same as
          // scaled coordinate of 0.0)
          if (scaled[j] >= 1.0) scaled[j] -= 1.0;
        }
 
        // find xyz-indices of cell that cutoffGroup is in.
        whichCell.x() = int(nCells_.x() * scaled.x());
        whichCell.y() = int(nCells_.y() * scaled.y());
        whichCell.z() = int(nCells_.z() * scaled.z());
 
        for (vector<Vector3i>::iterator os = cellOffsets_.begin();
             os != cellOffsets_.end(); ++os) {
          Vector3i m2v = whichCell + (*os);
 
          if (m2v.x() >= nCells_.x()) {
            m2v.x() = 0;
          } else if (m2v.x() < 0) {
            m2v.x() = nCells_.x() - 1;
          }
 
          if (m2v.y() >= nCells_.y()) {
            m2v.y() = 0;
          } else if (m2v.y() < 0) {
            m2v.y() = nCells_.y() - 1;
          }
 
          if (m2v.z() >= nCells_.z()) {
            m2v.z() = 0;
          } else if (m2v.z() < 0) {
            m2v.z() = nCells_.z() - 1;
          }
          int m2 = Vlinear(m2v, nCells_);
#ifdef IS_MPI
          for (vector<int>::iterator j2 = cellListCol_[m2].begin();
               j2 != cellListCol_[m2].end(); ++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)] - rs;
            if (usePeriodicBoundaryConditions_) { snap_->wrapVector(dr); }
            if (dr.lengthSquare() < rListSq_) {
              neighborList.push_back((*j2));
              ++len;
            }
          }
#else
          for (vector<int>::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 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 ((*j2) >= j1) {
              dr = snap_->cgData.position[(*j2)] - rs;
              if (usePeriodicBoundaryConditions_) { snap_->wrapVector(dr); }
              if (dr.lengthSquare() < rListSq_) {
                neighborList.push_back((*j2));
                ++len;
              }
            }
          }
#endif
        }
      }
    } else {
      // branch to do all cutoff group pairs
#ifdef IS_MPI
      for (int j1 = 0; j1 < nGroupsInRow_; j1++) {
        point[j1] = len;
        rs        = cgRowData.position[j1];
        for (int j2 = 0; j2 < nGroupsInCol_; j2++) {
          dr = cgColData.position[j2] - rs;
          if (usePeriodicBoundaryConditions_) { snap_->wrapVector(dr); }
          if (dr.lengthSquare() < rListSq_) {
            neighborList.push_back(j2);
            ++len;
          }
        }
      }
#else
      // include all groups here.
      for (int j1 = 0; j1 < nGroups_; j1++) {
        point[j1] = len;
        rs        = snap_->cgData.position[j1];
        // include self group interactions j2 == j1
        for (int j2 = j1; j2 < nGroups_; j2++) {
          dr = snap_->cgData.position[j2] - rs;
          if (usePeriodicBoundaryConditions_) { snap_->wrapVector(dr); }
          if (dr.lengthSquare() < rListSq_) {
            neighborList.push_back(j2);
            ++len;
          }
        }
      }
#endif
    }
 
#ifdef IS_MPI
    point[nGroupsInRow_] = len;
#else
    point[nGroups_] = len;
#endif
 
    // save the local cutoff group positions for the check that is
    // done on each loop:
    savedPositions.clear();
    savedPositions.reserve(nGroups_);
    for (int i = 0; i < nGroups_; i++)
      savedPositions.push_back(snap_->cgData.position[i]);
  }
 
  int ForceMatrixDecomposition::getGlobalIDRow(int atom1) {
#ifdef IS_MPI
    return AtomRowToGlobal[atom1];
#else
    return atom1;
#endif
  }
 
  int ForceMatrixDecomposition::getGlobalIDCol(int atom2) {
#ifdef IS_MPI
    return AtomColToGlobal[atom2];
#else
    return atom2;
#endif
  }
 
  int ForceMatrixDecomposition::getGlobalID(int atom1) {
#ifdef IS_MPI
    return AtomLocalToGlobal[atom1];
#else
    return atom1;
#endif
  }
}  // namespace OpenMD