sctl-tree/include/sctl/tree.hpp

#ifndef _SCTL_TREE_
#define _SCTL_TREE_

#include SCTL_INCLUDE(common.hpp)
#include SCTL_INCLUDE(morton.hpp)
#include SCTL_INCLUDE(comm.hpp)

#include <fstream>
#include <algorithm>

namespace SCTL_NAMESPACE {

struct VTUData {
  typedef float VTKReal;

  // Point data
  Vector<VTKReal> coord;  // always 3D
  Vector<VTKReal> value;

  // Cell data
  Vector<int32_t> connect;
  Vector<int32_t> offset;
  Vector<uint8_t> types;

  void WriteVTK(const std::string& fname, Comm comm = Comm::Self()) const {
    typedef typename VTUData::VTKReal VTKReal;

    Integer rank = comm.Rank();
    Integer np = comm.Size();

    Long value_dof = 0;
    {  // Write vtu file.
      std::ofstream vtufile;
      {  // Open file for writing.
        std::stringstream vtufname;
        vtufname << fname << std::setfill('0') << std::setw(6) << rank << ".vtu";
        vtufile.open(vtufname.str().c_str());
        if (vtufile.fail()) return;
      }
      {  // Write to file.
        Long pt_cnt = coord.Dim() / 3;
        Long cell_cnt = types.Dim();
        value_dof = (pt_cnt ? value.Dim() / pt_cnt : 0);

        Vector<int32_t> mpi_rank;
        {  // Set  mpi_rank
          Integer new_myrank = rank;
          mpi_rank.ReInit(pt_cnt);
          for (Long i = 0; i < mpi_rank.Dim(); i++) mpi_rank[i] = new_myrank;
        }

        bool isLittleEndian;
        {  // Set isLittleEndian
          uint16_t number = 0x1;
          uint8_t *numPtr = (uint8_t *)&number;
          isLittleEndian = (numPtr[0] == 1);
        }

        Long data_size = 0;
        vtufile << "<?xml version=\"1.0\"?>\n";
        vtufile << "<VTKFile type=\"UnstructuredGrid\" version=\"0.1\" byte_order=\"" << (isLittleEndian ? "LittleEndian" : "BigEndian") << "\">\n";
        // ===========================================================================
        vtufile << "  <UnstructuredGrid>\n";
        vtufile << "    <Piece NumberOfPoints=\"" << pt_cnt << "\" NumberOfCells=\"" << cell_cnt << "\">\n";
        //---------------------------------------------------------------------------
        vtufile << "      <Points>\n";
        vtufile << "        <DataArray type=\"Float" << sizeof(VTKReal) * 8 << "\" NumberOfComponents=\"3\" Name=\"Position\" format=\"appended\" offset=\"" << data_size << "\" />\n";
        data_size += sizeof(uint32_t) + coord.Dim() * sizeof(VTKReal);
        vtufile << "      </Points>\n";
        //---------------------------------------------------------------------------
        vtufile << "      <PointData>\n";
        if (value_dof) {  // value
          vtufile << "        <DataArray type=\"Float" << sizeof(VTKReal) * 8 << "\" NumberOfComponents=\"" << value_dof << "\" Name=\"value\" format=\"appended\" offset=\"" << data_size << "\" />\n";
          data_size += sizeof(uint32_t) + value.Dim() * sizeof(VTKReal);
        }
        {  // mpi_rank
          vtufile << "        <DataArray type=\"Int32\" NumberOfComponents=\"1\" Name=\"mpi_rank\" format=\"appended\" offset=\"" << data_size << "\" />\n";
          data_size += sizeof(uint32_t) + pt_cnt * sizeof(int32_t);
        }
        vtufile << "      </PointData>\n";
        //---------------------------------------------------------------------------
        //---------------------------------------------------------------------------
        vtufile << "      <Cells>\n";
        vtufile << "        <DataArray type=\"Int32\" Name=\"connectivity\" format=\"appended\" offset=\"" << data_size << "\" />\n";
        data_size += sizeof(uint32_t) + connect.Dim() * sizeof(int32_t);
        vtufile << "        <DataArray type=\"Int32\" Name=\"offsets\" format=\"appended\" offset=\"" << data_size << "\" />\n";
        data_size += sizeof(uint32_t) + offset.Dim() * sizeof(int32_t);
        vtufile << "        <DataArray type=\"UInt8\" Name=\"types\" format=\"appended\" offset=\"" << data_size << "\" />\n";
        data_size += sizeof(uint32_t) + types.Dim() * sizeof(uint8_t);
        vtufile << "      </Cells>\n";
        //---------------------------------------------------------------------------
        vtufile << "    </Piece>\n";
        vtufile << "  </UnstructuredGrid>\n";
        // ===========================================================================
        vtufile << "  <AppendedData encoding=\"raw\">\n";
        vtufile << "    _";

        int32_t block_size;
        {  // coord
          block_size = coord.Dim() * sizeof(VTKReal);
          vtufile.write((char *)&block_size, sizeof(int32_t));
          if (coord.Dim()) vtufile.write((char *)&coord[0], coord.Dim() * sizeof(VTKReal));
        }
        if (value_dof) {  // value
          block_size = value.Dim() * sizeof(VTKReal);
          vtufile.write((char *)&block_size, sizeof(int32_t));
          if (value.Dim()) vtufile.write((char *)&value[0], value.Dim() * sizeof(VTKReal));
        }
        {  // mpi_rank
          block_size = mpi_rank.Dim() * sizeof(int32_t);
          vtufile.write((char *)&block_size, sizeof(int32_t));
          if (mpi_rank.Dim()) vtufile.write((char *)&mpi_rank[0], mpi_rank.Dim() * sizeof(int32_t));
        }
        {  // block_size
          block_size = connect.Dim() * sizeof(int32_t);
          vtufile.write((char *)&block_size, sizeof(int32_t));
          if (connect.Dim()) vtufile.write((char *)&connect[0], connect.Dim() * sizeof(int32_t));
        }
        {  // offset
          block_size = offset.Dim() * sizeof(int32_t);
          vtufile.write((char *)&block_size, sizeof(int32_t));
          if (offset.Dim()) vtufile.write((char *)&offset[0], offset.Dim() * sizeof(int32_t));
        }
        {  // types
          block_size = types.Dim() * sizeof(uint8_t);
          vtufile.write((char *)&block_size, sizeof(int32_t));
          if (types.Dim()) vtufile.write((char *)&types[0], types.Dim() * sizeof(uint8_t));
        }

        vtufile << "\n";
        vtufile << "  </AppendedData>\n";
        // ===========================================================================
        vtufile << "</VTKFile>\n";
      }
      vtufile.close();  // close file
    }
    if (!rank) {  // Write pvtu file
      std::ofstream pvtufile;
      {  // Open file for writing
        std::stringstream pvtufname;
        pvtufname << fname << ".pvtu";
        pvtufile.open(pvtufname.str().c_str());
        if (pvtufile.fail()) return;
      }
      {  // Write to file.
        pvtufile << "<?xml version=\"1.0\"?>\n";
        pvtufile << "<VTKFile type=\"PUnstructuredGrid\">\n";
        pvtufile << "  <PUnstructuredGrid GhostLevel=\"0\">\n";
        pvtufile << "      <PPoints>\n";
        pvtufile << "        <PDataArray type=\"Float" << sizeof(VTKReal) * 8 << "\" NumberOfComponents=\"3\" Name=\"Position\"/>\n";
        pvtufile << "      </PPoints>\n";
        pvtufile << "      <PPointData>\n";
        if (value_dof) {  // value
          pvtufile << "        <PDataArray type=\"Float" << sizeof(VTKReal) * 8 << "\" NumberOfComponents=\"" << value_dof << "\" Name=\"value\"/>\n";
        }
        {  // mpi_rank
          pvtufile << "        <PDataArray type=\"Int32\" NumberOfComponents=\"1\" Name=\"mpi_rank\"/>\n";
        }
        pvtufile << "      </PPointData>\n";
        {
          // Extract filename from path.
          std::stringstream vtupath;
          vtupath << '/' << fname;
          std::string pathname = vtupath.str();
          std::string fname_ = pathname.substr(pathname.find_last_of("/\\") + 1);
          // char *fname_ = (char*)strrchr(vtupath.str().c_str(), '/') + 1;
          // std::string fname_ =
          // boost::filesystem::path(fname).filename().string().
          for (Integer i = 0; i < np; i++) pvtufile << "      <Piece Source=\"" << fname_ << std::setfill('0') << std::setw(6) << i << ".vtu\"/>\n";
        }
        pvtufile << "  </PUnstructuredGrid>\n";
        pvtufile << "</VTKFile>\n";
      }
      pvtufile.close();  // close file
    }
  };
};

template <class Real, Integer DIM> class Tree {
  public:

    struct NodeAttr {
      unsigned char Leaf : 1, Ghost : 1;
    };

    //struct PtrTree {
    //  Long p2n;
    //  Long parent;
    //  Long child[1 << DIM];
    //  Long nbr[pvfmm::pow<DIM>(3)];
    //};

    static constexpr Integer Dim() {
      return DIM;
    }

    Tree(const Comm& comm_ = Comm::Self()) {
      comm = comm_;
      Integer rank = comm.Rank();
      Integer np = comm.Size();

      Vector<Real> coord;
      { // Set coord
        Long N0 = 1;
        while (sctl::pow<DIM,Long>(N0) < np) N0++;
        Long N = sctl::pow<DIM,Long>(N0);
        Long start = N * (rank+0) / np;
        Long end   = N * (rank+1) / np;
        coord.ReInit((end-start)*DIM);
        for (Long i = start; i < end; i++) {
          Long  idx = i;
          for (Integer k = 0; k < DIM; k++) {
            coord[(i-start)*DIM+k] = (idx % N0) / (Real)N0;
            idx /= N0;
          }
        }
      }
      this->UpdateRefinement(coord);
    }

    ~Tree() {
      #ifdef SCTL_MEMDEBUG
      for (auto& pair : node_data) {
        SCTL_ASSERT(node_cnt.find(pair.first) != node_cnt.end());
      }
      #endif
    }

    const Vector<Morton<DIM>>& GetPartitionMID() const {
      return mins;
    }
    const Vector<Morton<DIM>>& GetNodeMID() const {
      return node_mid;
    }
    const Vector<NodeAttr>& GetNodeAttr() const {
      return node_attr;
    }
    const Comm& GetComm() const {
      return comm;
    }

    void UpdateRefinement(const Vector<Real>& coord, Long M = 1) {
      Integer np = comm.Size();
      Integer rank = comm.Rank();

      Morton<DIM> pt_mid0;
      Vector<Morton<DIM>> pt_mid;
      { // Construct sorted pt_mid
        Long Npt = coord.Dim() / DIM;
        pt_mid.ReInit(Npt);
        for (Long i = 0; i < Npt; i++) {
          pt_mid[i] = Morton<DIM>(coord.begin() + i*DIM);
        }
        Vector<Morton<DIM>> sorted_mid;
        comm.HyperQuickSort(pt_mid, sorted_mid);
        pt_mid.Swap(sorted_mid);
        SCTL_ASSERT(pt_mid.Dim());
        pt_mid0 = pt_mid[0];
      }
      { // Update M = global_min(pt_mid.Dim(), M)
        Long M0, M1, Npt = pt_mid.Dim();
        comm.Allreduce(Ptr2ConstItr<Long>(&M,1), Ptr2Itr<Long>(&M0,1), 1, Comm::CommOp::MIN);
        comm.Allreduce(Ptr2ConstItr<Long>(&Npt,1), Ptr2Itr<Long>(&M1,1), 1, Comm::CommOp::MIN);
        M = std::min(M0,M1);
        SCTL_ASSERT(M > 0);
      }
      { // pt_mid <-- [M points from rank-1; pt_mid; M points from rank+1]
        Long send_size0 = (rank+1<np ? M : 0);
        Long send_size1 = (rank  > 0 ? M : 0);
        Long recv_size0 = (rank  > 0 ? M : 0);
        Long recv_size1 = (rank+1<np ? M : 0);
        Vector<Morton<DIM>> pt_mid_(recv_size0 + pt_mid.Dim() + recv_size1);
        memcopy(pt_mid_.begin()+recv_size0, pt_mid.begin(), pt_mid.Dim());

        void* recv_req0 = comm.Irecv(pt_mid_.begin(), recv_size0, (rank+np-1)%np, 0);
        void* recv_req1 = comm.Irecv(pt_mid_.begin() + recv_size0 + pt_mid.Dim(), recv_size1, (rank+1)%np, 1);
        void* send_req0 = comm.Isend(pt_mid .begin() + pt_mid.Dim() - send_size0, send_size0, (rank+1)%np, 0);
        void* send_req1 = comm.Isend(pt_mid .begin(), send_size1, (rank+np-1)%np, 1);
        comm.Wait(recv_req0);
        comm.Wait(recv_req1);
        comm.Wait(send_req0);
        comm.Wait(send_req1);
        pt_mid.Swap(pt_mid_);
      }
      { // Build linear MortonID tree from pt_mid
        node_mid.ReInit(0);
        Long idx = 0;
        Morton<DIM> m0;
        Morton<DIM> mend = Morton<DIM>().Next();
        while (m0 < mend) {
          Integer d = m0.Depth();
          Morton<DIM> m1 = (idx + M < pt_mid.Dim() ? pt_mid[idx+M] : Morton<DIM>().Next());
          while (d < Morton<DIM>::MAX_DEPTH && m0.Ancestor(d) == m1.Ancestor(d)) d++;
          m0 = m0.Ancestor(d);
          node_mid.PushBack(m0);
          m0 = m0.Next();
          idx = std::lower_bound(pt_mid.begin(), pt_mid.end(), m0) - pt_mid.begin();
        }
      }
      { // Set mins
        mins.ReInit(np);
        Long min_idx = std::lower_bound(node_mid.begin(), node_mid.end(), pt_mid0) - node_mid.begin() - 1;
        if (min_idx < 0) min_idx = 0;
        comm.Allgather(node_mid.begin() + min_idx, 1, mins.begin(), 1);
      }
      { // Set node_mid, node_attr
        Morton<DIM> m0 = (rank      ? mins[rank]   : Morton<DIM>()       );
        Morton<DIM> m1 = (rank+1<np ? mins[rank+1] : Morton<DIM>().Next());
        Long Nnodes = node_mid.Dim();
        node_attr.ReInit(Nnodes);
        for (Long i = 0; i < Nnodes; i++) {
          node_attr[i].Leaf = 1;
          node_attr[i].Ghost = (node_mid[i] < m0 || node_mid[i] >= m1);
        }
      }

      { // Add non-leaf nodes and place-holder for ghost nodes
        // TODO
      }
      { // Update node_data, node_cnt
        // TODO
      }
    }

    void AddData(const std::string& name, const Vector<Real>& data, const Vector<Long>& cnt) {
      Long dof = 0;
      { // Check dof
        StaticArray<Long,2> Nl, Ng;
        Nl[0] = data.Dim();
        Nl[1] = omp_par::reduce(cnt.begin(), cnt.Dim());
        comm.Allreduce((ConstIterator<Long>)Nl, (Iterator<Long>)Ng, 2, Comm::CommOp::SUM);
        if (Ng[1]) dof = Ng[0] / Ng[1];
        SCTL_ASSERT(Nl[0] == Nl[1] * dof);
        SCTL_ASSERT(Ng[0] == Ng[1] * dof);
      }
      if (dof) SCTL_ASSERT(cnt.Dim() == node_mid.Dim());

      SCTL_ASSERT(node_data.find(name) == node_data.end());
      node_data[name] = data;
      node_cnt [name] =  cnt;
    }

    void GetData(Iterator<Vector<Real>>& data, ConstIterator<Vector<Long>>& cnt, const std::string& name) {
      auto data_ = node_data.find(name);
      const auto cnt_ = node_cnt.find(name);
      SCTL_ASSERT(data_ != node_data.end());
      SCTL_ASSERT( cnt_ != node_cnt .end());
      data = Ptr2Itr<Vector<Real>>(&data_->second,1);
      cnt  = Ptr2ConstItr<Vector<Long>>(& cnt_->second,1);
    }
    void GetData(ConstIterator<Vector<Real>>& data, ConstIterator<Vector<Long>>& cnt, const std::string& name) const {
      const auto data_ = node_data.find(name);
      const auto cnt_ = node_cnt.find(name);
      SCTL_ASSERT(data_ != node_data.end());
      SCTL_ASSERT( cnt_ != node_cnt .end());
      data = Ptr2ConstItr<Vector<Real>>(&data_->second,1);
      cnt  = Ptr2ConstItr<Vector<Long>>(& cnt_->second,1);
    }

    void DeleteData(const std::string& name) {
      SCTL_ASSERT(node_data.find(name) != node_data.end());
      SCTL_ASSERT(node_cnt .find(name) != node_cnt .end());
      node_data.erase(name);
      node_cnt .erase(name);
    }

    void WriteTreeVTK(std::string fname, bool show_ghost = false) const {
      typedef typename VTUData::VTKReal VTKReal;
      VTUData vtu_data;
      if (DIM <= 3) {  // Set vtu data
        static const Integer Ncorner = (1u << DIM);

        Vector<VTKReal> &coord = vtu_data.coord;
        //Vector<VTKReal> &value = vtu_data.value;

        Vector<int32_t> &connect = vtu_data.connect;
        Vector<int32_t> &offset = vtu_data.offset;
        Vector<uint8_t> &types = vtu_data.types;

        StaticArray<VTKReal, DIM> c;
        Long point_cnt = coord.Dim() / 3;
        Long connect_cnt = connect.Dim();
        for (Long nid = 0; nid < node_mid.Dim(); nid++) {
          const Morton<DIM> &mid = node_mid[nid];
          const NodeAttr &attr = node_attr[nid];
          if (!show_ghost && attr.Ghost) continue;
          if (!attr.Leaf) continue;

          mid.Coord((Iterator<VTKReal>)c);
          VTKReal s = sctl::pow<VTKReal>(0.5, mid.Depth());
          for (Integer j = 0; j < Ncorner; j++) {
            for (Integer i = 0; i < DIM; i++) coord.PushBack(c[i] + (j & (1u << i) ? 1 : 0) * s);
            for (Integer i = DIM; i < 3; i++) coord.PushBack(0);
            connect.PushBack(point_cnt);
            connect_cnt++;
            point_cnt++;
          }
          offset.PushBack(connect_cnt);
          if (DIM == 2)
            types.PushBack(8);
          else if (DIM == 3)
            types.PushBack(11);
          else
            types.PushBack(4);
        }
      }
      vtu_data.WriteVTK(fname, comm);
    }

  protected:

    void GetData(Iterator<Vector<Real>>& data, Iterator<Vector<Long>>& cnt, const std::string& name) {
      auto data_ = node_data.find(name);
      const auto cnt_ = node_cnt.find(name);
      SCTL_ASSERT(data_ != node_data.end());
      SCTL_ASSERT( cnt_ != node_cnt .end());
      data = Ptr2Itr<Vector<Real>>(&data_->second,1);
      cnt  = Ptr2Itr<Vector<Long>>(& cnt_->second,1);
    }

    static void scan(Vector<Long>& dsp, const Vector<Long>& cnt) {
      dsp.ReInit(cnt.Dim());
      dsp[0] = 0;
      omp_par::scan(cnt.begin(), dsp.begin(), cnt.Dim());
    }

    //template <typename A, typename B> struct SortPair {
    //  int operator<(const SortPair<A, B> &p1) const { return key < p1.key; }
    //  A key;
    //  B data;
    //};

  private:

    Vector<Morton<DIM>> mins;
    Vector<Morton<DIM>> node_mid;
    Vector<NodeAttr> node_attr;

    std::map<std::string, Vector<Real>> node_data;
    std::map<std::string, Vector<Long>> node_cnt;

    Comm comm;
};

template <class Real, Integer DIM, class BaseTree = Tree<Real,DIM>> class PtTree : public BaseTree {
  public:

    PtTree(const Comm& comm = Comm::Self()) : BaseTree(comm) {}

    ~PtTree() {
      #ifdef SCTL_MEMDEBUG
      for (auto& pair : data_pt_name) {
        ConstIterator<Vector<Real>> data;
        ConstIterator<Vector<Long>> cnt;
        this->GetData(data, cnt, pair.second);
        SCTL_ASSERT(scatter_idx.find(pair.second) != scatter_idx.end());
      }
      #endif
    }

    void UpdateRefinement(const Vector<Real>& coord, Long M = 1) {
      const auto& comm = this->GetComm();
      Long start_node_idx, end_node_idx;
      { // Set start_node_idx, end_node_idx
        const auto& mins = this->GetPartitionMID();
        const auto& node_mid = this->GetNodeMID();
        Integer np = comm.Size();
        Integer rank = comm.Rank();
        start_node_idx = std::lower_bound(node_mid.begin(), node_mid.end(), mins[rank]) - node_mid.begin();
        end_node_idx = std::lower_bound(node_mid.begin(), node_mid.end(), (rank+1==np ? Morton<DIM>().Next() : mins[rank+1])) - node_mid.begin();
      }

      BaseTree::UpdateRefinement(coord, M);

      const auto& mins = this->GetPartitionMID();
      const auto& node_mid = this->GetNodeMID();
      for (const auto& pair : pt_mid) {
        const auto& pt_name = pair.first;
        auto& pt_mid_ = pt_mid[pt_name];
        auto& scatter_idx_ = scatter_idx[pt_name];
        comm.PartitionS(pt_mid_, mins[comm.Rank()]);
        comm.PartitionN(scatter_idx_, pt_mid_.Dim());

        Vector<Long> pt_cnt(node_mid.Dim());
        for (Long i = 0; i < node_mid.Dim(); i++) { // Set pt_cnt
          Long start = std::lower_bound(pt_mid_.begin(), pt_mid_.end(), node_mid[i]) - pt_mid_.begin();
          Long end = std::lower_bound(pt_mid_.begin(), pt_mid_.end(), (i+1==node_mid.Dim() ? Morton<DIM>().Next() : node_mid[i+1])) - pt_mid_.begin();
          if (i == 0) SCTL_ASSERT(start == 0);
          if (i+1 == node_mid.Dim()) SCTL_ASSERT(end == pt_mid_.Dim());
          pt_cnt[i] = end - start;
        }

        for (const auto& pair : data_pt_name) {
          if (pair.second == pt_name) {
            const auto& data_name = pair.first;

            Iterator<Vector<Real>> data;
            Iterator<Vector<Long>> cnt;
            this->GetData(data, cnt, data_name);

            { // Update data
              Long dof = 0;
              { // Set dof
                StaticArray<Long,2> Nl = {0, 0}, Ng;
                Nl[0] = data->Dim();
                for (Long i = 0; i < cnt->Dim(); i++) Nl[1] += cnt[0][i];
                comm.Allreduce((ConstIterator<Long>)Nl, (Iterator<Long>)Ng, 2, Comm::CommOp::SUM);
                dof = Ng[0] / std::max<Long>(Ng[1],1);
              }
              Long offset = 0, count = 0;
              SCTL_ASSERT(0 <= start_node_idx);
              SCTL_ASSERT(start_node_idx <= end_node_idx);
              SCTL_ASSERT(end_node_idx <= cnt->Dim());
              for (Long i = 0; i < start_node_idx; i++) offset += cnt[0][i];
              for (Long i = start_node_idx; i < end_node_idx; i++) count += cnt[0][i];
              offset *= dof;
              count *= dof;

              Vector<Real> data_(count, data->begin() + offset);
              comm.PartitionN(data_, pt_mid_.Dim());
              data->Swap(data_);
            }
            cnt[0] = pt_cnt;
          }
        }
      }
    }

    void AddParticles(const std::string& name, const Vector<Real>& coord) {
      const auto& mins = this->GetPartitionMID();
      const auto& node_mid = this->GetNodeMID();
      const auto& comm = this->GetComm();

      SCTL_ASSERT(scatter_idx.find(name) == scatter_idx.end());
      Vector<Long>& scatter_idx_ = scatter_idx[name];

      Long N = coord.Dim() / DIM;
      SCTL_ASSERT(coord.Dim() == N * DIM);
      Nlocal[name] = N;

      Vector<Morton<DIM>>& pt_mid_ = pt_mid[name];
      if (pt_mid_.Dim() != N) pt_mid_.ReInit(N);
      for (Long i = 0; i < N; i++) {
        pt_mid_[i] = Morton<DIM>(coord.begin() + i*DIM);
      }
      comm.SortScatterIndex(pt_mid_, scatter_idx_, &mins[comm.Rank()]);
      comm.ScatterForward(pt_mid_, scatter_idx_);
      AddParticleData(name, name, coord);

      { // Set node_cnt
        Iterator<Vector<Real>> data_;
        Iterator<Vector<Long>> cnt_;
        this->GetData(data_,cnt_,name);
        cnt_[0].ReInit(node_mid.Dim());
        for (Long i = 0; i < node_mid.Dim(); i++) {
          Long start = std::lower_bound(pt_mid_.begin(), pt_mid_.end(), node_mid[i]) - pt_mid_.begin();
          Long end = std::lower_bound(pt_mid_.begin(), pt_mid_.end(), (i+1==node_mid.Dim() ? Morton<DIM>().Next() : node_mid[i+1])) - pt_mid_.begin();
          if (i == 0) SCTL_ASSERT(start == 0);
          if (i+1 == node_mid.Dim()) SCTL_ASSERT(end == pt_mid_.Dim());
          cnt_[0][i] = end - start;
        }
      }
    }

    void AddParticleData(const std::string& data_name, const std::string& particle_name, const Vector<Real>& data) {
      SCTL_ASSERT(scatter_idx.find(particle_name) != scatter_idx.end());
      SCTL_ASSERT(data_pt_name.find(data_name) == data_pt_name.end());
      data_pt_name[data_name] = particle_name;

      Iterator<Vector<Real>> data_;
      Iterator<Vector<Long>> cnt_;
      this->AddData(data_name, Vector<Real>(), Vector<Long>());
      this->GetData(data_,cnt_,data_name);
      { // Set data_[0]
        data_[0] = data;
        this->GetComm().ScatterForward(data_[0], scatter_idx[particle_name]);
      }
      if (data_name != particle_name) { // Set cnt_[0]
        Iterator<Vector<Real>> pt_coord;
        Iterator<Vector<Long>> pt_cnt;
        this->GetData(pt_coord, pt_cnt, particle_name);
        cnt_[0] = pt_cnt[0];
      }
    }

    void GetParticleData(Vector<Real>& data, const std::string& data_name) const {
      SCTL_ASSERT(data_pt_name.find(data_name) != data_pt_name.end());
      const std::string& particle_name = data_pt_name.find(data_name)->second;
      SCTL_ASSERT(scatter_idx.find(particle_name) != scatter_idx.end());
      const auto& scatter_idx_ = scatter_idx.find(particle_name)->second;
      const Long Nlocal_ = Nlocal.find(particle_name)->second;

      const auto& mins = this->GetPartitionMID();
      const auto& node_mid = this->GetNodeMID();
      const auto& comm = this->GetComm();

      Long dof = 0;
      Vector<Long> dsp;
      ConstIterator<Vector<Long>> cnt_;
      ConstIterator<Vector<Real>> data_;
      this->GetData(data_, cnt_, data_name);
      SCTL_ASSERT(cnt_->Dim() == node_mid.Dim());
      BaseTree::scan(dsp, cnt_[0]);
      { // Set dof
        Long Nn = node_mid.Dim();
        StaticArray<Long,2> Ng, Nl = {data_->Dim(), dsp[Nn-1]+cnt_[0][Nn-1]};
        comm.Allreduce((ConstIterator<Long>)Nl, (Iterator<Long>)Ng, 2, Comm::CommOp::SUM);
        if (Ng[1]) dof = Ng[0] / Ng[1];
      }
      { // Set data
        Integer np = comm.Size();
        Integer rank = comm.Rank();
        Long N0 = std::lower_bound(node_mid.begin(), node_mid.end(), mins[rank]) - node_mid.begin();
        Long N1 = std::lower_bound(node_mid.begin(), node_mid.end(), (rank+1==np ? Morton<DIM>().Next() : mins[rank+1])) - node_mid.begin();
        Long start = dsp[N0] * dof;
        Long end = (N1<dsp.Dim() ? dsp[N1] : dsp[N1-1]+cnt_[0][N1-1]) * dof;
        data.ReInit(end-start, (Iterator<Real>)data_->begin()+start);
        comm.ScatterReverse(data, scatter_idx_, Nlocal_ * dof);
      }
    }

    void DeleteParticleData(const std::string& data_name) {
      SCTL_ASSERT(data_pt_name.find(data_name) != data_pt_name.end());
      auto particle_name = data_pt_name[data_name];
      if (data_name == particle_name) {
        std::vector<std::string> data_name_lst;
        for (auto& pair : data_pt_name) {
          if (pair.second == particle_name) {
            data_name_lst.push_back(pair.first);
          }
        }
        for (auto x : data_name_lst) {
          if (x != particle_name) {
            DeleteParticleData(x);
          }
        }
        Nlocal.erase(particle_name);
      }
      this->DeleteData(data_name);
      data_pt_name.erase(data_name);
    }

    void WriteParticleVTK(std::string fname, std::string data_name, bool show_ghost = false) const {
      typedef typename VTUData::VTKReal VTKReal;
      const auto& node_mid = this->GetNodeMID();
      const auto& node_attr = this->GetNodeAttr();

      VTUData vtu_data;
      if (DIM <= 3) {  // Set vtu data
        SCTL_ASSERT(data_pt_name.find(data_name) != data_pt_name.end());
        std::string particle_name = data_pt_name.find(data_name)->second;

        ConstIterator<Vector<Real>> pt_coord = NullIterator<Vector<Real>>();
        ConstIterator<Vector<Real>> pt_value = NullIterator<Vector<Real>>();
        ConstIterator<Vector<Long>> pt_cnt = NullIterator<Vector<Long>>();
        Vector<Long> pt_dsp;
        Long value_dof = 0;
        { // Set pt_coord, pt_cnt, pt_dsp
          this->GetData(pt_coord, pt_cnt, particle_name);
          Tree<Real,DIM>::scan(pt_dsp, pt_cnt[0]);
        }
        if (particle_name != data_name) { // Set pt_value, value_dof
          ConstIterator<Vector<Long>> pt_cnt = NullIterator<Vector<Long>>();
          this->GetData(pt_value, pt_cnt, data_name);
          Long Npt = omp_par::reduce(pt_cnt->begin(), pt_cnt->Dim());
          value_dof = pt_value->Dim() / std::max<Long>(Npt,1);
        }

        Vector<VTKReal> &coord = vtu_data.coord;
        Vector<VTKReal> &value = vtu_data.value;

        Vector<int32_t> &connect = vtu_data.connect;
        Vector<int32_t> &offset = vtu_data.offset;
        Vector<uint8_t> &types = vtu_data.types;

        Long point_cnt = coord.Dim() / DIM;
        Long connect_cnt = connect.Dim();
        value.ReInit(point_cnt * value_dof);
        value.SetZero();

        SCTL_ASSERT(node_mid.Dim() == node_attr.Dim());
        SCTL_ASSERT(node_mid.Dim() == pt_cnt->Dim());
        for (Long i = 0; i < node_mid.Dim(); i++) {
          if (!show_ghost && node_attr[i].Ghost) continue;
          if (!node_attr[i].Leaf) continue;

          for (Long j = 0; j < pt_cnt[0][i]; j++) {
            ConstIterator<Real> pt_coord_ = pt_coord->begin() + (pt_dsp[i] + j) * DIM;
            ConstIterator<Real> pt_value_ = (value_dof ? pt_value->begin() + (pt_dsp[i] + j) * value_dof : NullIterator<Real>());

            for (Integer k = 0; k < DIM; k++) coord.PushBack((VTKReal)pt_coord_[k]);
            for (Integer k = DIM; k < 3; k++) coord.PushBack(0);
            for (Integer k = 0; k < value_dof; k++) value.PushBack((VTKReal)pt_value_[k]);
            connect.PushBack(point_cnt);
            connect_cnt++;
            point_cnt++;

            offset.PushBack(connect_cnt);
            types.PushBack(1);
          }
        }
      }
      vtu_data.WriteVTK(fname, this->GetComm());
    }

  private:

    std::map<std::string, Long> Nlocal;
    std::map<std::string, Vector<Morton<DIM>>> pt_mid;
    std::map<std::string, Vector<Long>> scatter_idx;
    std::map<std::string, std::string> data_pt_name;
};

}

#endif //_SCTL_TREE_