navier_stokes_preconditioners.cc

Go to the documentation of this file.

// LIC// ====================================================================
// LIC// This file forms part of oomph-lib, the object-oriented,
// LIC// multi-physics finite-element library, available
// LIC// at http://www.oomph-lib.org.
// LIC//
// LIC// Copyright (C) 2006-2022 Matthias Heil and Andrew Hazel
// LIC//
// LIC// This library is free software; you can redistribute it and/or
// LIC// modify it under the terms of the GNU Lesser General Public
// LIC// License as published by the Free Software Foundation; either
// LIC// version 2.1 of the License, or (at your option) any later version.
// LIC//
// LIC// This library is distributed in the hope that it will be useful,
// LIC// but WITHOUT ANY WARRANTY; without even the implied warranty of
// LIC// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
// LIC// Lesser General Public License for more details.
// LIC//
// LIC// You should have received a copy of the GNU Lesser General Public
// LIC// License along with this library; if not, write to the Free Software
// LIC// Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
// LIC// 02110-1301  USA.
// LIC//
// LIC// The authors may be contacted at oomph-lib@maths.man.ac.uk.
// LIC//
// LIC//====================================================================
#include "navier_stokes_preconditioners.h"
 
namespace oomph
{
  /// ////////////////////////////////////////////////////////////////////
  /// ////////////////////////////////////////////////////////////////////
  /// ////////////////////////////////////////////////////////////////////
 
 
  //======start_of_namespace============================================
  /// Namespace for exact solution for pressure advection diffusion
  /// problem
  //====================================================================
  namespace PressureAdvectionDiffusionValidation
  {
    /// Flag for solution
    unsigned Flag = 0;
 
    /// Peclet number -- overwrite with actual Reynolds number
    double Peclet = 0.0;
 
    /// Wind
    void wind_function(const Vector<double>& x, Vector<double>& wind)
    {
      if (Flag == 0)
      {
        wind[0] = sin(6.0 * x[1]);
        wind[1] = cos(6.0 * x[0]);
      }
      else
      {
        wind[0] = 1.0;
        wind[1] = 0.0;
      }
    }
 
    /// Exact solution as a Vector
    void get_exact_u(const Vector<double>& x, Vector<double>& u)
    {
      u.resize(3);
      wind_function(x, u);
      if (Flag == 0)
      {
        u[2] = x[0] * x[0] * pow(1.0 - x[0], 2.0) * x[1] * x[1] *
               pow(1.0 - x[1], 2.0);
      }
      else
      {
        u[2] = 0.1E1 - Peclet * x[0] * (0.1E1 - 0.5 * x[0]);
      }
    }
 
    /// Exact solution as a scalar
    void get_exact_u(const Vector<double>& x, double& u)
    {
      if (Flag == 0)
      {
        u = x[0] * x[0] * pow(1.0 - x[0], 2.0) * x[1] * x[1] *
            pow(1.0 - x[1], 2.0);
      }
      else
      {
        u = 0.1E1 - Peclet * x[0] * (0.1E1 - 0.5 * x[0]);
      }
    }
 
    /// Source function required to make the solution above an exact solution
    double source_function(const Vector<double>& x_vect)
    {
      double x[2];
      x[0] = x_vect[0];
      x[1] = x_vect[1];
 
 
      double source = 0.0;
 
      if (Flag == 0)
      {
        source =
          Peclet *
            (sin(0.6E1 * x[1]) * (2.0 * x[0] * pow(1.0 - x[0], 2.0) * x[1] *
                                    x[1] * pow(1.0 - x[1], 2.0) -
                                  2.0 * x[0] * x[0] * (1.0 - x[0]) * x[1] *
                                    x[1] * pow(1.0 - x[1], 2.0)) +
             cos(0.6E1 * x[0]) * (2.0 * x[0] * x[0] * pow(1.0 - x[0], 2.0) *
                                    x[1] * pow(1.0 - x[1], 2.0) -
                                  2.0 * x[0] * x[0] * pow(1.0 - x[0], 2.0) *
                                    x[1] * x[1] * (1.0 - x[1]))) -
          2.0 * pow(1.0 - x[0], 2.0) * x[1] * x[1] * pow(1.0 - x[1], 2.0) +
          8.0 * x[0] * (1.0 - x[0]) * x[1] * x[1] * pow(1.0 - x[1], 2.0) -
          2.0 * x[0] * x[0] * x[1] * x[1] * pow(1.0 - x[1], 2.0) -
          2.0 * x[0] * x[0] * pow(1.0 - x[0], 2.0) * pow(1.0 - x[1], 2.0) +
          8.0 * x[0] * x[0] * pow(1.0 - x[0], 2.0) * x[1] * (1.0 - x[1]) -
          2.0 * x[0] * x[0] * pow(1.0 - x[0], 2.0) * x[1] * x[1];
      }
      else
      {
        source = Peclet * (-0.1E1 * Peclet * (0.1E1 - 0.5 * x[0]) +
                           0.5 * Peclet * x[0]) -
                 0.1E1 * Peclet;
      }
 
      return source;
    }
 
 
  } // namespace PressureAdvectionDiffusionValidation
 
 
  /// /////////////////////////////////////////////////////////////
  /// /////////////////////////////////////////////////////////////
  /// /////////////////////////////////////////////////////////////
 
 
  //===========================================================================
  /// Setup the least-squares commutator Navier Stokes preconditioner. This
  /// extracts blocks corresponding to the velocity and pressure unknowns,
  /// creates the matrices actually needed in the application of the
  /// preconditioner and deletes what can be deleted... Note that
  /// this preconditioner needs a CRDoubleMatrix.
  //============================================================================
  void NavierStokesSchurComplementPreconditioner::setup()
  {
    // For debugging...
    bool doc_block_matrices = false;
 
    // For output timing results - to be removed soon. Ray
    bool raytime_flag = false;
 
    //%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
    // NOTE: In the interest of minimising memory usage, several containers
    //       are recycled, therefore their content/meaning changes
    //       throughout this function. The code is carefully annotated
    //       but you'll have to read it line by line!
    //%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
    double t_clean_up_memory_start = TimingHelpers::timer();
    // make sure any old data is deleted
    clean_up_memory();
    double t_clean_up_memory_end = TimingHelpers::timer();
    double clean_up_memory_time =
      t_clean_up_memory_end - t_clean_up_memory_start;
    if (raytime_flag)
    {
      oomph_info << "LSC: clean_up_memory_time: " << clean_up_memory_time
                 << std::endl;
    }
 
 
#ifdef PARANOID
    // paranoid check that the navier stokes mesh pt has been set
    if (Navier_stokes_mesh_pt == 0)
    {
      std::ostringstream error_message;
      error_message << "The navier stokes elements mesh pointer must be set.\n"
                    << "Use method set_navier_stokes_mesh(...)";
      throw OomphLibError(
        error_message.str(), OOMPH_CURRENT_FUNCTION, OOMPH_EXCEPTION_LOCATION);
    }
#endif
 
 
    // Set the mesh
    this->set_nmesh(1);
    this->set_mesh(0,
                   Navier_stokes_mesh_pt,
                   Allow_multiple_element_type_in_navier_stokes_mesh);
 
    // Get blocks
    // ----------
 
    // In comes the current Jacobian. Recast it to a CR double matrix;
    // shout if that can't be done.
    CRDoubleMatrix* cr_matrix_pt = dynamic_cast<CRDoubleMatrix*>(matrix_pt());
 
 
#ifdef PARANOID
    if (cr_matrix_pt == 0)
    {
      std::ostringstream error_message;
      error_message
        << "NavierStokesSchurComplementPreconditioner only works with "
        << "CRDoubleMatrix matrices" << std::endl;
      throw OomphLibError(
        error_message.str(), OOMPH_CURRENT_FUNCTION, OOMPH_EXCEPTION_LOCATION);
    }
#endif
 
 
    if (doc_block_matrices)
    {
      std::stringstream junk;
      junk << "j_matrix" << comm_pt()->my_rank() << ".dat";
      oomph_info << "About to output " << junk.str() << std::endl;
      cr_matrix_pt->sparse_indexed_output_with_offset(junk.str());
      oomph_info << "Done output of " << junk.str() << std::endl;
    }
 
 
    // Set up block look up schemes (done automatically in the
    // BlockPreconditioner base class, based on the information
    // provided in the block-preconditionable elements in the problem)
 
    // this preconditioner has two types of block:
    // type 0: velocity - corresponding to DOFs 0 to n-2
    // type 1: pressure - corresponding to DOF n-1
    double t_block_setup_start = TimingHelpers::timer();
    unsigned ndof_types = 0;
 
    if (this->is_subsidiary_block_preconditioner())
    {
      ndof_types = this->ndof_types();
    }
    else
    {
      // This is the upper-most master block preconditioner, the Navier-Stokes
      // mesh is in position 0
      ndof_types = this->ndof_types_in_mesh(0);
    }
 
    Vector<unsigned> dof_to_block_map(ndof_types);
    dof_to_block_map[ndof_types - 1] = 1;
 
    this->block_setup(dof_to_block_map);
 
    double t_block_setup_finish = TimingHelpers::timer();
    double block_setup_time = t_block_setup_finish - t_block_setup_start;
    if (Doc_time)
    {
      oomph_info << "Time for block_setup(...) [sec]: " << block_setup_time
                 << "\n";
    }
 
    if (raytime_flag)
    {
      oomph_info << "LSC: block_setup: " << block_setup_time << std::endl;
    }
 
 
    // determine whether the F preconditioner is a block preconditioner (and
    // therefore a subsidiary preconditioner)
    BlockPreconditioner<CRDoubleMatrix>* F_block_preconditioner_pt =
      dynamic_cast<BlockPreconditioner<CRDoubleMatrix>*>(F_preconditioner_pt);
    F_preconditioner_is_block_preconditioner = true;
    if (F_block_preconditioner_pt == 0)
    {
      F_preconditioner_is_block_preconditioner = false;
    }
 
    // Get B (the divergence block)
    double t_get_B_start = TimingHelpers::timer();
    CRDoubleMatrix* b_pt = new CRDoubleMatrix;
    this->get_block(1, 0, *b_pt);
 
    double t_get_B_finish = TimingHelpers::timer();
    double get_B_time = t_get_B_finish - t_get_B_start;
    if (Doc_time)
    {
      oomph_info << "Time to get B [sec]: " << get_B_time << "\n";
    }
 
    if (raytime_flag)
    {
      oomph_info << "LSC: get block B get_B_time: " << get_B_time << std::endl;
    }
 
    if (doc_block_matrices)
    {
      std::stringstream junk;
      junk << "b_matrix" << comm_pt()->my_rank() << ".dat";
      b_pt->sparse_indexed_output_with_offset(junk.str());
      oomph_info << "Done output of " << junk.str() << std::endl;
    }
 
 
    // get the inverse velocity and pressure mass matrices
    CRDoubleMatrix* inv_v_mass_pt = 0;
    CRDoubleMatrix* inv_p_mass_pt = 0;
 
    double ivmm_assembly_start_t = TimingHelpers::timer();
    if (Use_LSC)
    {
      // We only need the velocity mass matrix
      assemble_inv_press_and_veloc_mass_matrix_diagonal(
        inv_p_mass_pt, inv_v_mass_pt, false);
    }
    else
    {
      // We need both mass matrices
      assemble_inv_press_and_veloc_mass_matrix_diagonal(
        inv_p_mass_pt, inv_v_mass_pt, true);
    }
 
    double ivmm_assembly_finish_t = TimingHelpers::timer();
 
    double ivmm_assembly_time = ivmm_assembly_finish_t - ivmm_assembly_start_t;
    if (Doc_time)
    {
      oomph_info << "Time to assemble inverse diagonal velocity and pressure"
                 << "mass matrices) [sec]: " << ivmm_assembly_time << "\n";
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: ivmm_assembly_time: " << ivmm_assembly_time
                 << std::endl;
    }
 
 
    if (doc_block_matrices)
    {
      std::stringstream junk;
      junk << "inv_v_mass_matrix" << comm_pt()->my_rank() << ".dat";
      inv_v_mass_pt->sparse_indexed_output_with_offset(junk.str());
      oomph_info << "Done output of " << junk.str() << std::endl;
    }
 
 
    // Get gradient matrix Bt
    CRDoubleMatrix* bt_pt = new CRDoubleMatrix;
    double t_get_Bt_start = TimingHelpers::timer();
    this->get_block(0, 1, *bt_pt);
    double t_get_Bt_finish = TimingHelpers::timer();
 
    double t_get_Bt_time = t_get_Bt_finish - t_get_Bt_start;
    if (Doc_time)
    {
      oomph_info << "Time to get Bt [sec]: " << t_get_Bt_time << std::endl;
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: get block Bt: " << t_get_Bt_time << std::endl;
    }
 
    if (doc_block_matrices)
    {
      std::stringstream junk;
      junk << "bt_matrix" << comm_pt()->my_rank() << ".dat";
      bt_pt->sparse_indexed_output_with_offset(junk.str());
      oomph_info << "Done output of " << junk.str() << std::endl;
    }
 
 
    // Build pressure Poisson matrix
    CRDoubleMatrix* p_matrix_pt = new CRDoubleMatrix;
 
    // Multiply inverse velocity mass matrix by gradient matrix B^T
    double t_QBt_matrix_start = TimingHelpers::timer();
    CRDoubleMatrix* qbt_pt = new CRDoubleMatrix;
    inv_v_mass_pt->multiply(*bt_pt, *qbt_pt);
    delete bt_pt;
    bt_pt = 0;
 
    // Store product in bt_pt
    bt_pt = qbt_pt;
    double t_QBt_matrix_finish = TimingHelpers::timer();
 
    double t_QBt_time = t_QBt_matrix_finish - t_QBt_matrix_start;
    if (Doc_time)
    {
      oomph_info << "Time to generate QBt [sec]: " << t_QBt_time << std::endl;
    }
    delete inv_v_mass_pt;
    inv_v_mass_pt = 0;
    if (raytime_flag)
    {
      oomph_info << "LSC: t_QBt_time (matrix multiplicaton): " << t_QBt_time
                 << std::endl;
    }
 
    // Multiply B from left by divergence matrix B and store result in
    // pressure Poisson matrix.
    double t_p_matrix_start = TimingHelpers::timer();
    b_pt->multiply(*bt_pt, *p_matrix_pt);
    double t_p_matrix_finish = TimingHelpers::timer();
 
    double t_p_time = t_p_matrix_finish - t_p_matrix_start;
    if (Doc_time)
    {
      oomph_info << "Time to generate P [sec]: " << t_p_time << std::endl;
    }
    // Kill divergence matrix because we don't need it any more
    delete b_pt;
    b_pt = 0;
 
    if (raytime_flag)
    {
      oomph_info << "LSC: t_p_time (matrix multiplication): " << t_p_time
                 << std::endl;
    }
 
 
    // Build the matvec operator for QBt
    double t_QBt_MV_start = TimingHelpers::timer();
    QBt_mat_vec_pt = new MatrixVectorProduct;
    this->setup_matrix_vector_product(QBt_mat_vec_pt, bt_pt, 1);
    double t_QBt_MV_finish = TimingHelpers::timer();
 
    double t_p_time2 = t_QBt_MV_finish - t_QBt_MV_start;
    if (Doc_time)
    {
      oomph_info << "Time to build QBt matrix vector operator [sec]: "
                 << t_p_time2 << std::endl;
    }
 
    // Kill gradient matrix B^T (it's been overwritten anyway and
    // needs to be recomputed afresh below)
    delete bt_pt;
    bt_pt = 0;
 
    if (raytime_flag)
    {
      oomph_info << "LSC: QBt (setup MV product): " << t_p_time2 << std::endl;
    }
 
    // Do we need the Fp stuff?
    if (!Use_LSC)
    {
      // Get pressure advection diffusion matrix Fp and store in
      // a "big" matrix (same size as the problem's Jacobian)
      double t_get_Fp_start = TimingHelpers::timer();
      CRDoubleMatrix full_fp_matrix;
      get_pressure_advection_diffusion_matrix(full_fp_matrix);
 
      // Now extract the pressure pressure block
      CRDoubleMatrix* fp_matrix_pt = new CRDoubleMatrix;
      this->get_block_other_matrix(1, 1, &full_fp_matrix, *fp_matrix_pt);
      double t_get_Fp_finish = TimingHelpers::timer();
      if (Doc_time)
      {
        double t_get_Fp_time = t_get_Fp_finish - t_get_Fp_start;
        oomph_info << "Time to get Fp [sec]: " << t_get_Fp_time << std::endl;
      }
 
      // Build vector product of pressure advection diffusion matrix with
      // inverse pressure mass matrix
      CRDoubleMatrix* fp_qp_inv_pt = new CRDoubleMatrix;
      fp_matrix_pt->multiply(*inv_p_mass_pt, *fp_qp_inv_pt);
 
      // Build the matvec operator for E = F_p Q_p^{-1}
      double t_Fp_Qp_inv_MV_start = TimingHelpers::timer();
      E_mat_vec_pt = new MatrixVectorProduct;
      this->setup_matrix_vector_product(E_mat_vec_pt, fp_qp_inv_pt, 1);
      double t_Fp_Qp_inv_MV_finish = TimingHelpers::timer();
      if (Doc_time)
      {
        double t_p_time = t_Fp_Qp_inv_MV_finish - t_Fp_Qp_inv_MV_start;
        oomph_info << "Time to build Fp Qp^{-1} matrix vector operator [sec]: "
                   << t_p_time << std::endl;
      }
      // Kill pressure advection diffusion and inverse pressure mass matrices
      delete inv_p_mass_pt;
      inv_p_mass_pt = 0;
      delete fp_qp_inv_pt;
      fp_qp_inv_pt = 0;
    }
 
 
    // Get momentum block F
    CRDoubleMatrix* f_pt = new CRDoubleMatrix;
    double t_get_F_start = TimingHelpers::timer();
    this->get_block(0, 0, *f_pt);
    double t_get_F_finish = TimingHelpers::timer();
 
    double t_get_F_time = t_get_F_finish - t_get_F_start;
    if (Doc_time)
    {
      oomph_info << "Time to get F [sec]: " << t_get_F_time << std::endl;
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: get_block t_get_F_time: " << t_get_F_time
                 << std::endl;
    }
 
    // form the matrix vector product helper
    double t_F_MV_start = TimingHelpers::timer();
    F_mat_vec_pt = new MatrixVectorProduct;
    this->setup_matrix_vector_product(F_mat_vec_pt, f_pt, 0);
    double t_F_MV_finish = TimingHelpers::timer();
 
    double t_F_MV_time = t_F_MV_finish - t_F_MV_start;
    if (Doc_time)
    {
      oomph_info << "Time to build F Matrix Vector Operator [sec]: "
                 << t_F_MV_time << std::endl;
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: MV product setup t_F_MV_time: " << t_F_MV_time
                 << std::endl;
    }
 
 
    // if F is a block preconditioner then we can delete the F matrix
    if (F_preconditioner_is_block_preconditioner)
    {
      delete f_pt;
      f_pt = 0;
    }
 
    // Rebuild Bt (remember that we temporarily overwrote
    // it by its product with the inverse velocity mass matrix)
    t_get_Bt_start = TimingHelpers::timer();
    bt_pt = new CRDoubleMatrix;
    this->get_block(0, 1, *bt_pt);
    t_get_Bt_finish = TimingHelpers::timer();
    double t_get_Bt_time2 = t_get_Bt_finish - t_get_Bt_start;
    if (Doc_time)
    {
      oomph_info << "Time to get Bt [sec]: " << t_get_Bt_time2 << std::endl;
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: get_block t_get_Bt_time2: " << t_get_Bt_time2
                 << std::endl;
    }
 
 
    // form the matrix vector operator for Bt
    double t_Bt_MV_start = TimingHelpers::timer();
    Bt_mat_vec_pt = new MatrixVectorProduct;
    this->setup_matrix_vector_product(Bt_mat_vec_pt, bt_pt, 1);
 
    //  if(Doc_time)
    //   {
    //    oomph_info << "Time to build Bt Matrix Vector Operator [sec]: "
    //               << t_Bt_MV_time << std::endl;
    //   }
 
    delete bt_pt;
    bt_pt = 0;
 
    double t_Bt_MV_finish = TimingHelpers::timer();
 
    double t_Bt_MV_time = t_Bt_MV_finish - t_Bt_MV_start;
    if (raytime_flag)
    {
      oomph_info << "LSC: MV product setup t_Bt_MV_time: " << t_Bt_MV_time
                 << std::endl;
    }
 
    // if the P preconditioner has not been setup
    if (P_preconditioner_pt == 0)
    {
      P_preconditioner_pt = new SuperLUPreconditioner;
      Using_default_p_preconditioner = true;
    }
 
    // Setup the preconditioner for the Pressure matrix
    double t_p_prec_start = TimingHelpers::timer();
 
    if (doc_block_matrices)
    {
      std::stringstream junk;
      junk << "p_matrix" << comm_pt()->my_rank() << ".dat";
      p_matrix_pt->sparse_indexed_output_with_offset(junk.str());
      oomph_info << "Done output of " << junk.str() << std::endl;
    }
 
    P_preconditioner_pt->setup(p_matrix_pt);
    delete p_matrix_pt;
    p_matrix_pt = 0;
    double t_p_prec_finish = TimingHelpers::timer();
 
    double t_p_prec_time = t_p_prec_finish - t_p_prec_start;
    if (Doc_time)
    {
      oomph_info << "P sub-preconditioner setup time [sec]: " << t_p_prec_time
                 << "\n";
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: p_prec setup time: " << t_p_prec_time << std::endl;
    }
 
 
    // Set up solver for solution of system with momentum matrix
    // ----------------------------------------------------------
 
    // if the F preconditioner has not been setup
    if (F_preconditioner_pt == 0)
    {
      F_preconditioner_pt = new SuperLUPreconditioner;
      Using_default_f_preconditioner = true;
    }
 
    // if F is a block preconditioner
    double t_f_prec_start = TimingHelpers::timer();
    if (F_preconditioner_is_block_preconditioner)
    {
      unsigned nvelocity_dof_types =
        Navier_stokes_mesh_pt->finite_element_pt(0)->dim();
 
      Vector<unsigned> dof_map(nvelocity_dof_types);
      for (unsigned i = 0; i < nvelocity_dof_types; i++)
      {
        dof_map[i] = i;
      }
 
      F_block_preconditioner_pt->turn_into_subsidiary_block_preconditioner(
        this, dof_map);
 
      F_block_preconditioner_pt->setup(matrix_pt());
    }
    // otherwise F is not a block preconditioner
    else
    {
      F_preconditioner_pt->setup(f_pt);
      delete f_pt;
      f_pt = 0;
    }
    double t_f_prec_finish = TimingHelpers::timer();
    double t_f_prec_time = t_f_prec_finish - t_f_prec_start;
    if (Doc_time)
    {
      oomph_info << "F sub-preconditioner setup time [sec]: " << t_f_prec_time
                 << "\n";
    }
    if (raytime_flag)
    {
      oomph_info << "LSC: f_prec setup time: " << t_f_prec_time << std::endl;
    }
 
    // Remember that the preconditioner has been setup so
    // the stored information can be wiped when we
    // come here next...
    Preconditioner_has_been_setup = true;
  }
 
 
  //=======================================================================
  /// Apply preconditioner to r.
  //=======================================================================
  void NavierStokesSchurComplementPreconditioner::preconditioner_solve(
    const DoubleVector& r, DoubleVector& z)
  {
#ifdef PARANOID
    if (Preconditioner_has_been_setup == false)
    {
      std::ostringstream error_message;
      error_message << "setup must be called before using preconditioner_solve";
      throw OomphLibError(
        error_message.str(), OOMPH_CURRENT_FUNCTION, OOMPH_EXCEPTION_LOCATION);
    }
    if (z.built())
    {
      if (z.nrow() != r.nrow())
      {
        std::ostringstream error_message;
        error_message << "The vectors z and r must have the same number of "
                      << "of global rows";
        throw OomphLibError(error_message.str(),
                            OOMPH_CURRENT_FUNCTION,
                            OOMPH_EXCEPTION_LOCATION);
      }
    }
#endif
 
    // if z is not setup then give it the same distribution
    if (!z.distribution_pt()->built())
    {
      z.build(r.distribution_pt(), 0.0);
    }
 
    // Step 1 - apply approximate Schur inverse to pressure unknowns (block 1)
    // -----------------------------------------------------------------------
 
    // Working vectors
    DoubleVector temp_vec;
    DoubleVector another_temp_vec;
    DoubleVector yet_another_temp_vec;
 
    // Copy pressure values from residual vector to temp_vec:
    // Loop over all entries in the global vector (this one
    // includes velocity and pressure dofs in some random fashion)
    this->get_block_vector(1, r, temp_vec);
 
    // NOTE: The vector temp_vec now contains the vector r_p.
 
    // LSC version
    if (Use_LSC)
    {
      // Solve first pressure Poisson system
#ifdef PARANOID
      // check a solver has been set
      if (P_preconditioner_pt == 0)
      {
        std::ostringstream error_message;
        error_message << "P_preconditioner_pt has not been set.";
        throw OomphLibError(error_message.str(),
                            OOMPH_CURRENT_FUNCTION,
                            OOMPH_EXCEPTION_LOCATION);
      }
#endif
 
      // use some Preconditioner's preconditioner_solve function
      P_preconditioner_pt->preconditioner_solve(temp_vec, another_temp_vec);
 
      // NOTE: The vector another_temp_vec now contains the vector P^{-1} r_p
 
      // Multiply another_temp_vec by matrix E and stick the result into
      // temp_vec
      temp_vec.clear();
      QBt_mat_vec_pt->multiply(another_temp_vec, temp_vec);
      another_temp_vec.clear();
      F_mat_vec_pt->multiply(temp_vec, another_temp_vec);
      temp_vec.clear();
      QBt_mat_vec_pt->multiply_transpose(another_temp_vec, temp_vec);
 
 
      // NOTE: The vector temp_vec now contains E P^{-1} r_p
 
      // Solve second pressure Poisson system using preconditioner_solve
      another_temp_vec.clear();
      P_preconditioner_pt->preconditioner_solve(temp_vec, another_temp_vec);
 
      // NOTE: The vector another_temp_vec now contains z_p = P^{-1} E P^{-1}
      // r_p
      //       as required (apart from the sign which we'll fix in the
      //       next step.
    }
    // Fp version
    else
    {
      // Multiply temp_vec by matrix E and stick the result into
      // yet_another_temp_vec
      E_mat_vec_pt->multiply(temp_vec, yet_another_temp_vec);
 
      // NOTE: The vector yet_another_temp_vec now contains Fp Qp^{-1} r_p
 
      // Solve pressure Poisson system
#ifdef PARANOID
      // check a solver has been set
      if (P_preconditioner_pt == 0)
      {
        std::ostringstream error_message;
        error_message << "P_preconditioner_pt has not been set.";
        throw OomphLibError(error_message.str(),
                            OOMPH_CURRENT_FUNCTION,
                            OOMPH_EXCEPTION_LOCATION);
      }
#endif
 
      // Solve second pressure Poisson system using preconditioner_solve
      another_temp_vec.clear();
      P_preconditioner_pt->preconditioner_solve(yet_another_temp_vec,
                                                another_temp_vec);
 
      // NOTE: The vector another_temp_vec now contains
      //       z_p = P^{-1} Fp Qp^{-1} r_p
      //       as required (apart from the sign which we'll fix in the
      //       next step.
    }
 
    // Now copy another_temp_vec (i.e. z_p) back into the global vector z.
    // Loop over all entries in the global results vector z:
    temp_vec.build(another_temp_vec.distribution_pt(), 0.0);
    temp_vec -= another_temp_vec;
    return_block_vector(1, temp_vec, z);
 
 
    // Step 2 - apply preconditioner to velocity unknowns (block 0)
    // ------------------------------------------------------------
 
    // Recall that another_temp_vec (computed above) contains the
    // negative of the solution of the Schur complement systen, -z_p.
    // Multiply by G (stored in Block_matrix_pt(0,1) and store
    // result in temp_vec (vector resizes itself).
    temp_vec.clear();
    Bt_mat_vec_pt->multiply(another_temp_vec, temp_vec);
 
    // NOTE: temp_vec now contains -G z_p
 
    // The vector another_temp_vec is no longer needed -- re-use it to store
    // velocity quantities:
    another_temp_vec.clear();
 
    // Loop over all enries in the global vector and find the
    // entries associated with the velocities:
    get_block_vector(0, r, another_temp_vec);
    another_temp_vec += temp_vec;
 
    // NOTE:  The vector another_temp_vec now contains r_u - G z_p
 
    // Solve momentum system
#ifdef PARANOID
    // check a solver has been set
    if (F_preconditioner_pt == 0)
    {
      std::ostringstream error_message;
      error_message << "F_preconditioner_pt has not been set.";
      throw OomphLibError(
        error_message.str(), OOMPH_CURRENT_FUNCTION, OOMPH_EXCEPTION_LOCATION);
    }
#endif
 
    // use some Preconditioner's preconditioner solve
    // and return
    if (F_preconditioner_is_block_preconditioner)
    {
      return_block_vector(0, another_temp_vec, z);
      F_preconditioner_pt->preconditioner_solve(z, z);
    }
    else
    {
      F_preconditioner_pt->preconditioner_solve(another_temp_vec, temp_vec);
      return_block_vector(0, temp_vec, z);
    }
  }
 
 
  //========================================================================
  /// Helper function to assemble the inverse diagonals of the pressure and
  /// velocity mass matrix from the elemental contributions defined in
  /// NavierStokesElementWithDiagonalMassMatrices::
  /// get_pressure_and_velocity_mass_matrix_diagonal(...)
  /// If do_both=true, both are computed, otherwise only the velocity
  /// mass matrix (the LSC version of the preconditioner only needs
  /// that one)
  //========================================================================
  void NavierStokesSchurComplementPreconditioner::
    assemble_inv_press_and_veloc_mass_matrix_diagonal(
      CRDoubleMatrix*& inv_p_mass_pt,
      CRDoubleMatrix*& inv_v_mass_pt,
      const bool& do_both)
  {
    // determine the velocity rows required by this processor
    unsigned v_first_row = this->block_distribution_pt(0)->first_row();
    unsigned v_nrow_local = this->block_distribution_pt(0)->nrow_local();
    unsigned v_nrow = this->block_distribution_pt(0)->nrow();
 
    // create storage for the diagonals
    double* v_values = new double[v_nrow_local];
    for (unsigned i = 0; i < v_nrow_local; i++)
    {
      v_values[i] = 0.0;
    }
 
    // Equivalent information for pressure mass matrix (only needed for
    // Fp version)
    unsigned p_first_row = 0;
    unsigned p_nrow_local = 0;
    unsigned p_nrow = 0;
    double* p_values = 0;
    if (!Use_LSC)
    {
      // determine the pressure rows required by this processor
      p_first_row = this->block_distribution_pt(1)->first_row();
      p_nrow_local = this->block_distribution_pt(1)->nrow_local();
      p_nrow = this->block_distribution_pt(1)->nrow();
 
      // create storage for the diagonals
      p_values = new double[p_nrow_local];
      for (unsigned i = 0; i < p_nrow_local; i++)
      {
        p_values[i] = 0.0;
      }
    }
 
    // if the problem is distributed
    bool distributed = false;
#ifdef OOMPH_HAS_MPI
    if (problem_pt()->distributed() ||
        this->master_distribution_pt()->distributed())
    {
      distributed = true;
    }
#endif
 
    // next we get the diagonal velocity mass matrix data
    if (distributed)
    {
#ifdef OOMPH_HAS_MPI
 
      // the number of processors
      unsigned nproc = comm_pt()->nproc();
 
      // and my rank
      unsigned my_rank = comm_pt()->my_rank();
 
      // determine the rows for which we have lookup rows
 
      // if the problem is NOT distributed then we only classify global
      // equations on this processor to avoid duplication (as every processor
      // holds every element)
      unsigned first_lookup_row = 0;
      unsigned last_lookup_row = 0;
      first_lookup_row = this->master_distribution_pt()->first_row();
      last_lookup_row =
        first_lookup_row + this->master_distribution_pt()->nrow_local() - 1;
 
      // find number of local elements
      unsigned n_el = Navier_stokes_mesh_pt->nelement();
 
      // get the master distribution pt
      const LinearAlgebraDistribution* master_distribution_pt =
        this->master_distribution_pt();
 
      // Do the two blocks (0: veloc; 1: press)
      unsigned max_block = 0;
      if (!Use_LSC) max_block = 1;
      for (unsigned block_index = 0; block_index <= max_block; block_index++)
      {
        // Local working variables: Default to velocity
        unsigned v_or_p_first_row = v_first_row;
        double* v_or_p_values = v_values;
        // Switch to pressure
        if (block_index == 1)
        {
          v_or_p_first_row = p_first_row;
          v_or_p_values = p_values;
        }
 
 
        // the diagonal mass matrix contributions that have been
        // classified and should be sent to another processor
        Vector<double>* classified_contributions_send =
          new Vector<double>[nproc];
 
        // the corresponding block indices
        Vector<unsigned>* classified_indices_send = new Vector<unsigned>[nproc];
 
        // the matrix contributions that cannot be classified by this processor
        // and therefore must be sent to another for classification
        Vector<double>* unclassified_contributions_send =
          new Vector<double>[nproc];
 
        // the corresponding global indices that require classification
        Vector<unsigned>* unclassified_indices_send =
          new Vector<unsigned>[nproc];
 
        // get the velocity or pressure distribution pt
        const LinearAlgebraDistribution* velocity_or_press_dist_pt =
          this->block_distribution_pt(block_index);
 
        // get the contribution for each element
        for (unsigned e = 0; e < n_el; e++)
        {
          // Get element
          GeneralisedElement* el_pt = Navier_stokes_mesh_pt->element_pt(e);
 
          // check that the element is not halo
          if (!el_pt->is_halo())
          {
            // find number of degrees of freedom in the element
            // (this is slightly too big because it includes the
            // pressure dofs but this doesn't matter)
            unsigned el_dof = el_pt->ndof();
 
            // Allocate local storage for the element's contribution to the
            // mass matrix diagonal
            Vector<double> el_vmm_diagonal(el_dof, 0.0);
            Vector<double> el_pmm_diagonal(el_dof, 0.0);
 
            unsigned which_one = 2;
            if (block_index == 1) which_one = 1;
 
            NavierStokesElementWithDiagonalMassMatrices* cast_el_pt = 0;
            cast_el_pt =
              dynamic_cast<NavierStokesElementWithDiagonalMassMatrices*>(el_pt);
            if (cast_el_pt != 0)
            {
              cast_el_pt->get_pressure_and_velocity_mass_matrix_diagonal(
                el_pmm_diagonal, el_vmm_diagonal, which_one);
            }
 
            // get the contribution for each dof
            for (unsigned i = 0; i < el_dof; i++)
            {
              // Get the equation number
              unsigned eqn_number = el_pt->eqn_number(i);
 
              // if I have lookup information on this processor
              if ((eqn_number >= first_lookup_row) &&
                  (eqn_number <= last_lookup_row))
              {
                // Only use the dofs that we're dealing with here
                if (this->block_number(eqn_number) == int(block_index))
                {
                  // get the index in the block
                  unsigned index = this->index_in_block(eqn_number);
 
                  // determine which processor requires the block index
                  for (unsigned p = 0; p < nproc; p++)
                  {
                    if ((index >= velocity_or_press_dist_pt->first_row(p)) &&
                        (index < (velocity_or_press_dist_pt->first_row(p) +
                                  velocity_or_press_dist_pt->nrow_local(p))))
                    {
                      // if it is required by this processor then add the
                      // contribution
                      if (p == my_rank)
                      {
                        if (block_index == 0)
                        {
                          v_or_p_values[index - v_or_p_first_row] +=
                            el_vmm_diagonal[i];
                        }
                        else if (block_index == 1)
                        {
                          v_or_p_values[index - v_or_p_first_row] +=
                            el_pmm_diagonal[i];
                        }
                      }
                      // otherwise store it for communication
                      else
                      {
                        if (block_index == 0)
                        {
                          classified_contributions_send[p].push_back(
                            el_vmm_diagonal[i]);
                          classified_indices_send[p].push_back(index);
                        }
                        else if (block_index == 1)
                        {
                          classified_contributions_send[p].push_back(
                            el_pmm_diagonal[i]);
                          classified_indices_send[p].push_back(index);
                        }
                      }
                    }
                  }
                }
              }
              // if we do not have the lookup information on this processor
              // then we send the mass matrix contribution to a processor
              // which we know to have the lookup information
              // the assumption: the processor for which the master block
              // preconditioner distribution will definitely hold the lookup
              // data for eqn_number (although others may)
              else if (problem_pt()->distributed())
              {
                // determine which processor requires the block index
                unsigned p = 0;
                while (
                  !(eqn_number >= master_distribution_pt->first_row(p) &&
                    (eqn_number < (master_distribution_pt->first_row(p) +
                                   master_distribution_pt->nrow_local(p)))))
                {
                  p++;
                }
 
                // store the data
                if (block_index == 0)
                {
                  unclassified_contributions_send[p].push_back(
                    el_vmm_diagonal[i]);
                  unclassified_indices_send[p].push_back(eqn_number);
                }
                else if (block_index == 1)
                {
                  unclassified_contributions_send[p].push_back(
                    el_pmm_diagonal[i]);
                  unclassified_indices_send[p].push_back(eqn_number);
                }
              }
            }
          }
        }
 
        // next the unclassified contributions are communicated to
        // processors that can classify them
 
        // first determine how many unclassified rows are to be sent to
        // each processor
        unsigned* n_unclassified_send = new unsigned[nproc];
        for (unsigned p = 0; p < nproc; p++)
        {
          if (p == my_rank)
          {
            n_unclassified_send[p] = 0;
          }
          else
          {
            n_unclassified_send[p] = unclassified_contributions_send[p].size();
          }
        }
 
        // then all-to-all com number of unclassified to be sent / recv
        unsigned* n_unclassified_recv = new unsigned[nproc];
        MPI_Alltoall(n_unclassified_send,
                     1,
                     MPI_UNSIGNED,
                     n_unclassified_recv,
                     1,
                     MPI_UNSIGNED,
                     comm_pt()->mpi_comm());
 
        // the base displacement for the sends
        MPI_Aint base_displacement;
        MPI_Get_address(v_or_p_values, &base_displacement);
 
        // allocate storage for the data to be received
        // and post the sends and recvs
        Vector<double*> unclassified_contributions_recv(nproc);
        Vector<unsigned*> unclassified_indices_recv(nproc);
        Vector<MPI_Request> unclassified_recv_requests;
        Vector<MPI_Request> unclassified_send_requests;
        Vector<unsigned> unclassified_recv_proc;
        for (unsigned p = 0; p < nproc; p++)
        {
          if (p != my_rank)
          {
            // recv
            if (n_unclassified_recv[p] > 0)
            {
              unclassified_contributions_recv[p] =
                new double[n_unclassified_recv[p]];
              unclassified_indices_recv[p] =
                new unsigned[n_unclassified_recv[p]];
 
              // data for the struct data type
              MPI_Datatype recv_types[2];
              MPI_Aint recv_displacements[2];
              int recv_sz[2];
 
              // contributions
              MPI_Type_contiguous(
                n_unclassified_recv[p], MPI_DOUBLE, &recv_types[0]);
              MPI_Type_commit(&recv_types[0]);
              MPI_Get_address(unclassified_contributions_recv[p],
                              &recv_displacements[0]);
              recv_displacements[0] -= base_displacement;
              recv_sz[0] = 1;
 
              // indices
              MPI_Type_contiguous(
                n_unclassified_recv[p], MPI_UNSIGNED, &recv_types[1]);
              MPI_Type_commit(&recv_types[1]);
              MPI_Get_address(unclassified_indices_recv[p],
                              &recv_displacements[1]);
              recv_displacements[1] -= base_displacement;
              recv_sz[1] = 1;
 
              // build the final recv type
              MPI_Datatype final_recv_type;
              MPI_Type_create_struct(
                2, recv_sz, recv_displacements, recv_types, &final_recv_type);
              MPI_Type_commit(&final_recv_type);
 
              // and recv
              MPI_Request req;
              MPI_Irecv(v_or_p_values,
                        1,
                        final_recv_type,
                        p,
                        0,
                        comm_pt()->mpi_comm(),
                        &req);
              unclassified_recv_requests.push_back(req);
              unclassified_recv_proc.push_back(p);
              MPI_Type_free(&recv_types[0]);
              MPI_Type_free(&recv_types[1]);
              MPI_Type_free(&final_recv_type);
            }
 
            // send
            if (n_unclassified_send[p] > 0)
            {
              // data for the struct data type
              MPI_Datatype send_types[2];
              MPI_Aint send_displacements[2];
              int send_sz[2];
 
              // contributions
              MPI_Type_contiguous(
                n_unclassified_send[p], MPI_DOUBLE, &send_types[0]);
              MPI_Type_commit(&send_types[0]);
              MPI_Get_address(&unclassified_contributions_send[p][0],
                              &send_displacements[0]);
              send_displacements[0] -= base_displacement;
              send_sz[0] = 1;
 
              // indices
              MPI_Type_contiguous(
                n_unclassified_send[p], MPI_UNSIGNED, &send_types[1]);
              MPI_Type_commit(&send_types[1]);
              MPI_Get_address(&unclassified_indices_send[p][0],
                              &send_displacements[1]);
              send_displacements[1] -= base_displacement;
              send_sz[1] = 1;
 
              // build the final send type
              MPI_Datatype final_send_type;
              MPI_Type_create_struct(
                2, send_sz, send_displacements, send_types, &final_send_type);
              MPI_Type_commit(&final_send_type);
 
              // and send
              MPI_Request req;
              MPI_Isend(v_or_p_values,
                        1,
                        final_send_type,
                        p,
                        0,
                        comm_pt()->mpi_comm(),
                        &req);
              unclassified_send_requests.push_back(req);
              MPI_Type_free(&send_types[0]);
              MPI_Type_free(&send_types[1]);
              MPI_Type_free(&final_send_type);
            }
          }
        }
 
        // next classify the data as it is received
        unsigned n_unclassified_recv_req = unclassified_recv_requests.size();
        while (n_unclassified_recv_req > 0)
        {
          // get the processor number and remove the completed request
          // for the vector of requests
          int req_num;
          MPI_Waitany(n_unclassified_recv_req,
                      &unclassified_recv_requests[0],
                      &req_num,
                      MPI_STATUS_IGNORE);
          unsigned p = unclassified_recv_proc[req_num];
          unclassified_recv_requests.erase(unclassified_recv_requests.begin() +
                                           req_num);
          unclassified_recv_proc.erase(unclassified_recv_proc.begin() +
                                       req_num);
          n_unclassified_recv_req--;
 
          // next classify the dofs
          // and store them for sending to other processors if required
          unsigned n_recv = n_unclassified_recv[p];
          for (unsigned i = 0; i < n_recv; i++)
          {
            unsigned eqn_number = unclassified_indices_recv[p][i];
            // Only deal with our block unknowns
            if (this->block_number(eqn_number) == int(block_index))
            {
              // get the index in the block
              unsigned index = this->index_in_block(eqn_number);
 
              // determine which processor requires the block index
              for (unsigned pp = 0; pp < nproc; pp++)
              {
                if ((index >= velocity_or_press_dist_pt->first_row(pp)) &&
                    (index < (velocity_or_press_dist_pt->first_row(pp) +
                              velocity_or_press_dist_pt->nrow_local(pp))))
                {
                  // if it is required by this processor then add the
                  // contribution
                  if (pp == my_rank)
                  {
                    v_or_p_values[index - v_or_p_first_row] +=
                      unclassified_contributions_recv[p][i];
                  }
                  // otherwise store it for communication
                  else
                  {
                    double v = unclassified_contributions_recv[p][i];
                    classified_contributions_send[pp].push_back(v);
                    classified_indices_send[pp].push_back(index);
                  }
                }
              }
            }
          }
 
          // clean up
          delete[] unclassified_contributions_recv[p];
          delete[] unclassified_indices_recv[p];
        }
        delete[] n_unclassified_recv;
 
        // now all indices have been classified
 
        // next the classified contributions are communicated to
        // processors that require them
 
        // first determine how many classified rows are to be sent to
        // each processor
        unsigned* n_classified_send = new unsigned[nproc];
        for (unsigned p = 0; p < nproc; p++)
        {
          if (p == my_rank)
          {
            n_classified_send[p] = 0;
          }
          else
          {
            n_classified_send[p] = classified_contributions_send[p].size();
          }
        }
 
        // then all-to-all number of classified to be sent / recv
        unsigned* n_classified_recv = new unsigned[nproc];
        MPI_Alltoall(n_classified_send,
                     1,
                     MPI_UNSIGNED,
                     n_classified_recv,
                     1,
                     MPI_UNSIGNED,
                     comm_pt()->mpi_comm());
 
        // allocate storage for the data to be received
        // and post the sends and recvs
        Vector<double*> classified_contributions_recv(nproc);
        Vector<unsigned*> classified_indices_recv(nproc);
        Vector<MPI_Request> classified_recv_requests;
        Vector<MPI_Request> classified_send_requests;
        Vector<unsigned> classified_recv_proc;
        for (unsigned p = 0; p < nproc; p++)
        {
          if (p != my_rank)
          {
            // recv
            if (n_classified_recv[p] > 0)
            {
              classified_contributions_recv[p] =
                new double[n_classified_recv[p]];
              classified_indices_recv[p] = new unsigned[n_classified_recv[p]];
 
              // data for the struct data type
              MPI_Datatype recv_types[2];
              MPI_Aint recv_displacements[2];
              int recv_sz[2];
 
              // contributions
              MPI_Type_contiguous(
                n_classified_recv[p], MPI_DOUBLE, &recv_types[0]);
              MPI_Type_commit(&recv_types[0]);
              MPI_Get_address(classified_contributions_recv[p],
                              &recv_displacements[0]);
              recv_displacements[0] -= base_displacement;
              recv_sz[0] = 1;
 
              // indices
              MPI_Type_contiguous(
                n_classified_recv[p], MPI_UNSIGNED, &recv_types[1]);
              MPI_Type_commit(&recv_types[1]);
              MPI_Get_address(classified_indices_recv[p],
                              &recv_displacements[1]);
              recv_displacements[1] -= base_displacement;
              recv_sz[1] = 1;
 
              // build the final recv type
              MPI_Datatype final_recv_type;
              MPI_Type_create_struct(
                2, recv_sz, recv_displacements, recv_types, &final_recv_type);
              MPI_Type_commit(&final_recv_type);
 
              // and recv
              MPI_Request req;
              MPI_Irecv(v_or_p_values,
                        1,
                        final_recv_type,
                        p,
                        0,
                        comm_pt()->mpi_comm(),
                        &req);
              classified_recv_requests.push_back(req);
              classified_recv_proc.push_back(p);
              MPI_Type_free(&recv_types[0]);
              MPI_Type_free(&recv_types[1]);
              MPI_Type_free(&final_recv_type);
            }
 
            // send
            if (n_classified_send[p] > 0)
            {
              // data for the struct data type
              MPI_Datatype send_types[2];
              MPI_Aint send_displacements[2];
              int send_sz[2];
 
              // contributions
              MPI_Type_contiguous(
                n_classified_send[p], MPI_DOUBLE, &send_types[0]);
              MPI_Type_commit(&send_types[0]);
              MPI_Get_address(&classified_contributions_send[p][0],
                              &send_displacements[0]);
              send_displacements[0] -= base_displacement;
              send_sz[0] = 1;
 
              // indices
              MPI_Type_contiguous(
                n_classified_send[p], MPI_UNSIGNED, &send_types[1]);
              MPI_Type_commit(&send_types[1]);
              MPI_Get_address(&classified_indices_send[p][0],
                              &send_displacements[1]);
              send_displacements[1] -= base_displacement;
              send_sz[1] = 1;
 
              // build the final send type
              MPI_Datatype final_send_type;
              MPI_Type_create_struct(
                2, send_sz, send_displacements, send_types, &final_send_type);
              MPI_Type_commit(&final_send_type);
 
              // and send
              MPI_Request req;
              MPI_Isend(v_or_p_values,
                        1,
                        final_send_type,
                        p,
                        0,
                        comm_pt()->mpi_comm(),
                        &req);
              classified_send_requests.push_back(req);
              MPI_Type_free(&send_types[0]);
              MPI_Type_free(&send_types[1]);
              MPI_Type_free(&final_send_type);
            }
          }
        }
 
        // next classify the data as it is received
        unsigned n_classified_recv_req = classified_recv_requests.size();
        while (n_classified_recv_req > 0)
        {
          // get the processor number and remove the completed request
          // for the vector of requests
          int req_num;
          MPI_Waitany(n_classified_recv_req,
                      &classified_recv_requests[0],
                      &req_num,
                      MPI_STATUS_IGNORE);
          unsigned p = classified_recv_proc[req_num];
          classified_recv_requests.erase(classified_recv_requests.begin() +
                                         req_num);
          classified_recv_proc.erase(classified_recv_proc.begin() + req_num);
          n_classified_recv_req--;
 
          // next classify the dofs
          // and store them for sending to other processors if required
          unsigned n_recv = n_classified_recv[p];
          for (unsigned i = 0; i < n_recv; i++)
          {
            v_or_p_values[classified_indices_recv[p][i] - v_or_p_first_row] +=
              classified_contributions_recv[p][i];
          }
 
          // clean up
          delete[] classified_contributions_recv[p];
          delete[] classified_indices_recv[p];
        }
 
        // wait for the unclassified sends to complete
        unsigned n_unclassified_send_req = unclassified_send_requests.size();
        if (n_unclassified_send_req > 0)
        {
          MPI_Waitall(n_unclassified_send_req,
                      &unclassified_send_requests[0],
                      MPI_STATUS_IGNORE);
        }
        delete[] unclassified_contributions_send;
        delete[] unclassified_indices_send;
        delete[] n_unclassified_send;
 
        // wait for the classified sends to complete
        unsigned n_classified_send_req = classified_send_requests.size();
        if (n_classified_send_req > 0)
        {
          MPI_Waitall(n_classified_send_req,
                      &classified_send_requests[0],
                      MPI_STATUS_IGNORE);
        }
        delete[] classified_indices_send;
        delete[] classified_contributions_send;
        delete[] n_classified_recv;
        delete[] n_classified_send;
 
        // Copy the values back where they belong
        if (block_index == 0)
        {
          v_values = v_or_p_values;
        }
        else if (block_index == 1)
        {
          p_values = v_or_p_values;
        }
      }
 
#endif
    }
    // or if the problem is not distributed
    else
    {
      // find number of elements
      unsigned n_el = Navier_stokes_mesh_pt->nelement();
 
      // Fp needs pressure and velocity mass matrices
      unsigned which_one = 0;
      if (Use_LSC) which_one = 2;
 
      // get the contribution for each element
      for (unsigned e = 0; e < n_el; e++)
      {
        // Get element
        GeneralisedElement* el_pt = Navier_stokes_mesh_pt->element_pt(e);
 
        // find number of degrees of freedom in the element
        // (this is slightly too big because it includes the
        // pressure dofs but this doesn't matter)
        unsigned el_dof = el_pt->ndof();
 
        // allocate local storage for the element's contribution to the
        // pressure and velocity mass matrix diagonal
        Vector<double> el_vmm_diagonal(el_dof, 0.0);
        Vector<double> el_pmm_diagonal(el_dof, 0.0);
 
        NavierStokesElementWithDiagonalMassMatrices* cast_el_pt = 0;
        cast_el_pt =
          dynamic_cast<NavierStokesElementWithDiagonalMassMatrices*>(el_pt);
        if (cast_el_pt != 0)
        {
          cast_el_pt->get_pressure_and_velocity_mass_matrix_diagonal(
            el_pmm_diagonal, el_vmm_diagonal, which_one);
        }
        else
        {
#ifdef PARANOID
          std::ostringstream error_message;
          error_message
            << "Navier-Stokes mesh contains element that is not of type \n"
            << "NavierStokesElementWithDiagonalMassMatrices. \n"
            << "The element is in fact of type " << typeid(*el_pt).name()
            << "\nWe'll assume that it does not make a used contribution \n"
            << "to the inverse diagonal mass matrix used in the "
               "preconditioner\n"
            << "and (to avoid divisions by zero) fill in dummy unit entries.\n"
            << "[This case currently arises with flux control problems\n"
            << "where for simplicity the NetFluxControlElement has been added "
               "\n"
            << "to the Navier Stokes mesh -- this should probably be changed "
               "at\n"
            << "some point -- if you get this warning in any other context\n"
            << "you should check your code very carefully]\n";
          OomphLibWarning(error_message.str(),
                          "NavierStokesSchurComplementPreconditioner::assemble_"
                          "inv_press_and_veloc_mass_matrix_diagonal()",
                          OOMPH_EXCEPTION_LOCATION);
#endif
 
          // Fill in dummy entries to stop division by zero below
          for (unsigned j = 0; j < el_dof; j++)
          {
            el_vmm_diagonal[j] = 1.0;
            el_pmm_diagonal[j] = 1.0;
          }
        }
 
        // Get the contribution for each dof
        for (unsigned i = 0; i < el_dof; i++)
        {
          // Get the equation number
          unsigned eqn_number = el_pt->eqn_number(i);
 
          // Get the velocity dofs
          if (this->block_number(eqn_number) == 0)
          {
            // get the index in the block
            unsigned index = this->index_in_block(eqn_number);
 
            // if it is required on this processor
            if ((index >= v_first_row) &&
                (index < (v_first_row + v_nrow_local)))
            {
              v_values[index - v_first_row] += el_vmm_diagonal[i];
            }
          }
          // Get the pressure dofs
          else if (this->block_number(eqn_number) == 1)
          {
            if (!Use_LSC)
            {
              // get the index in the block
              unsigned index = this->index_in_block(eqn_number);
 
              // if it is required on this processor
              if ((index >= p_first_row) &&
                  (index < (p_first_row + p_nrow_local)))
              {
                p_values[index - p_first_row] += el_pmm_diagonal[i];
              }
            }
          }
        }
      }
    }
 
    // Create column index and row start for velocity mass matrix
    int* v_column_index = new int[v_nrow_local];
    int* v_row_start = new int[v_nrow_local + 1];
    for (unsigned i = 0; i < v_nrow_local; i++)
    {
#ifdef PARANOID
      if (v_values[i] == 0.0)
      {
        std::ostringstream error_message;
        error_message << "Zero entry in diagonal of velocity mass matrix\n"
                      << "Index: " << i << std::endl;
        throw OomphLibError(error_message.str(),
                            OOMPH_CURRENT_FUNCTION,
                            OOMPH_EXCEPTION_LOCATION);
      }
#endif
      v_values[i] = 1.0 / v_values[i];
      v_column_index[i] = v_first_row + i;
      v_row_start[i] = i;
    }
    v_row_start[v_nrow_local] = v_nrow_local;
 
    // Build the velocity mass matrix
    inv_v_mass_pt = new CRDoubleMatrix(this->block_distribution_pt(0));
    inv_v_mass_pt->build_without_copy(
      v_nrow, v_nrow_local, v_values, v_column_index, v_row_start);
 
    // Create pressure mass matrix
    if (!Use_LSC)
    {
      // Create column index and row start for pressure mass matrix
      int* p_column_index = new int[p_nrow_local];
      int* p_row_start = new int[p_nrow_local + 1];
      for (unsigned i = 0; i < p_nrow_local; i++)
      {
#ifdef PARANOID
        if (p_values[i] == 0.0)
        {
          std::ostringstream error_message;
          error_message << "Zero entry in diagonal of pressure mass matrix\n"
                        << "Index: " << i << std::endl;
          throw OomphLibError(error_message.str(),
                              OOMPH_CURRENT_FUNCTION,
                              OOMPH_EXCEPTION_LOCATION);
        }
#endif
        p_values[i] = 1.0 / p_values[i];
 
        p_column_index[i] = p_first_row + i;
        p_row_start[i] = i;
      }
      p_row_start[p_nrow_local] = p_nrow_local;
 
      // Build the pressure mass matrix
      inv_p_mass_pt = new CRDoubleMatrix(this->block_distribution_pt(1));
      inv_p_mass_pt->build_without_copy(
        p_nrow, p_nrow_local, p_values, p_column_index, p_row_start);
    }
  }
 
  //=========================================================================
  /// Helper function to delete preconditioner data.
  //=========================================================================
  void NavierStokesSchurComplementPreconditioner::clean_up_memory()
  {
    if (Preconditioner_has_been_setup)
    {
      // delete matvecs
      delete Bt_mat_vec_pt;
      Bt_mat_vec_pt = 0;
 
      delete F_mat_vec_pt;
      F_mat_vec_pt = 0;
 
      delete QBt_mat_vec_pt;
      QBt_mat_vec_pt = 0;
 
      delete E_mat_vec_pt;
      E_mat_vec_pt = 0;
 
      // delete stuff from velocity solve
      if (Using_default_f_preconditioner)
      {
        delete F_preconditioner_pt;
        F_preconditioner_pt = 0;
      }
 
      // delete stuff from Schur complement approx
      if (Using_default_p_preconditioner)
      {
        delete P_preconditioner_pt;
        P_preconditioner_pt = 0;
      }
    }
  }
 
 
} // namespace oomph