mdwf_solver.cc

Go to the documentation of this file.

// -*- C++ -*-
/*! \file
 *  \brief DWF/SSE double-prec solver
 */
 
#include "actions/ferm/qprop/mdwf_solver.h"
 
extern "C" {
#include <qop-mdwf3.h>
};
 
#include "eoprec_constdet_wilstype_fermact_w.h"
#include "io/aniso_io.h"
#include "actions/ferm/invert/syssolver_cg_params.h"
 
 
using namespace QDP;
namespace Chroma 
{ 
  //! AVP's DWF Solver interface
  /*!
   * \ingroup qprop
   *
   * @{
   */
 
  /* Utility functions -- stick these in local anonymous namespace */
 
  namespace { 
 
    double gaugeReader(int mu,  
                       const int latt_coord[4],
                       int row,
                       int col,
                       int reim,
                       void *env)
    {
      START_CODE();
        /* Translate arg */
      multi1d<LatticeColorMatrix>& u = *(multi1d<LatticeColorMatrix>*)env;
      
      // Get node and index
      multi1d<int> coord(Nd);
      coord = latt_coord;
      int node = Layout::nodeNumber(coord);
      int linear = Layout::linearSiteIndex(coord);
      
      if (node != Layout::nodeNumber()) {
        
        QDPIO::cerr << __func__ << ": wrong coordinates for this node" << std::endl;
        QDP_abort(1);
      }
      
      // Get the value
      // NOTE: it would be nice to use the "peek" functions, but they will
      // broadcast to all nodes the value since they are platform independent.
      // We don't want that, so we poke into the on-node data
      double val = (reim == 0) ? 
        toDouble(u[mu].elem(linear).elem().elem(row,col).real()) : 
        toDouble(u[mu].elem(linear).elem().elem(row,col).imag());
 
 
      END_CODE();      
      return val;
 
    }
    
    
    // Fermion Reader function - user supplied
    double fermionReader(const int latt_coord[5],
                         int color,
                         int spin,
                         int reim,
                         void *env)
    {
      START_CODE();
 
      /* Translate arg */
      multi1d<LatticeFermion>& psi = *(multi1d<LatticeFermion>*)env;
      
      // Get node and index
      int s = latt_coord[Nd];
      multi1d<int> coord(Nd);
      coord = latt_coord;
      int node = Layout::nodeNumber(coord);
      int linear = Layout::linearSiteIndex(coord);
      
      if (node != Layout::nodeNumber()) {
        
        QDPIO::cerr << __func__ << ": wrong coordinates for this node" << std::endl;
        QDP_abort(1);
      }
      
      // Get the value
      // NOTE: it would be nice to use the "peek" functions, but they will
      // broadcast to all nodes the value since they are platform independent.
      // We don't want that, so we poke into the on-node data
      double val = (reim == 0) ? 
        double(psi[s].elem(linear).elem(spin).elem(color).real()) : 
        double(psi[s].elem(linear).elem(spin).elem(color).imag());
      
      // if (spin >= Ns/2)
      //        val *= -1;
 
      END_CODE();
      return val;
    }
    
      
      
    
    // Fermion Writer function - user supplied
    void fermionWriter(const int latt_coord[5],
                       int color, 
                       int spin,
                       int reim,
                       double val,
                       void *env )
      
    {
      START_CODE();
 
      /* Translate arg */
      multi1d<LatticeFermion>& psi = *(multi1d<LatticeFermion>*)env;
      
      // Get node and index
      int s = latt_coord[Nd];
      multi1d<int> coord(Nd);
      coord = latt_coord;
      int node = Layout::nodeNumber(coord);
      int linear = Layout::linearSiteIndex(coord);
      
      if (node != Layout::nodeNumber()) {
        QDPIO::cerr << __func__ << ": wrong coordinates for this node" << std::endl;
        QDP_abort(1);
      }
      
      // Rescale
      //      if (spin >= Ns/2)
      //        val *= -1;
      
      // val *= -2.0;
      
      // Set the value
      // NOTE: it would be nice to use the "peek" functions, but they will
      // broadcast to all nodes the value since they are platform independent.
      // We don't want that, so we poke into the on-node data
      if (reim == 0)
        psi[s].elem(linear).elem(spin).elem(color).real() = val;
      else
        psi[s].elem(linear).elem(spin).elem(color).imag() = val;
 
      END_CODE();
      return;
    }
    
    // Env is for the interface spec. I ignore it completely
    void sublattice_func(int lo[],
                         int hi[],
                         const int node[],
                         void *env) 
    {
      START_CODE();
      // Given my node coordinates in node[]
      // produce the lo/hi pair
      
      // This is the size on the local subgrid. 
      // For QDP++ they are all the same.
      const multi1d<int>& local_subgrid=Layout::subgridLattSize();
      for(int i=0; i <Nd; i++) { 
        // The lowest coordinate is just the node coordinate
        // times the local subgrid size in that direction
        lo[i]=node[i]*local_subgrid[i];
        
        // The high is the start of the next 'corner'
        // I would say my hi is hi[i]-1 but am following the 
        // document conventions
        hi[i]=(node[i]+1)*local_subgrid[i];
      }
      
      END_CODE();
      return;
    }
  } // End of anonymous namespace
 
  //! Solver the linear system
  /*!
   * \param psi      quark propagator ( Modify )
   * \param chi      source ( Read )
   * \return number of CG iterations
   */
  SystemSolverResults_t MDWFQpropT::operator() (multi1d<LatticeFermion>& psi, 
                                                const multi1d<LatticeFermion>& chi) const
  {
    START_CODE();
 
    SystemSolverResults_t res;
    res.n_count = 0;
    int out_iters_single=0;
    int out_iters_double=0;
    
    /* Stuff for flopcounting */
    double time_sec;
    long long  flops;
    long long  sent; /* Messages? Bytes? Faces? */
    long long received; /* Messages? Bytes? Receives? */
    
    
    /* Single Precision Branch. */
    {
      double out_eps_single;
      
      // OK Try to solve for the single precision 
      // Cast to single precision gauge field
      QOP_F3_MDWF_Gauge *sprec_gauge = NULL;
      if ( QOP_F3_MDWF_import_gauge(&sprec_gauge,
                                    state,
                                    gaugeReader,
                                    (void *)&u) != 0 ) { 
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Fermion fields */
      QOP_F3_MDWF_Fermion *sprec_rhs;
      QOP_F3_MDWF_Fermion *sprec_x0;
      QOP_F3_MDWF_Fermion *sprec_soln;
      
      if( QOP_F3_MDWF_import_fermion(&sprec_rhs, 
                                     state,
                                     fermionReader,
                                     (void *)&chi) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      if( QOP_F3_MDWF_import_fermion(&sprec_x0, 
                                     state,
                                     fermionReader,
                                     (void *)&psi) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      if( QOP_F3_MDWF_allocate_fermion(&sprec_soln, 
                                       state) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Do the solve */
      
      double target_epsilon = toDouble(invParam.RsdCG*invParam.RsdCG);
      int max_iteration = invParam.MaxCG;
      int status; 
      
      QDPIO::cout << "MDWFQpropT: Beginning Single Precision Solve" << std::endl;
      if( ( status=QOP_F3_MDWF_DDW_CG(sprec_soln,
                                      &out_iters_single,
                                      &out_eps_single,
                                      params, 
                                      sprec_x0,
                                      sprec_gauge,
                                      sprec_rhs,
                                      max_iteration,
                                      target_epsilon, 
                                      QOP_MDWF_LOG_NONE)) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Get Perf counters before solve */
      if( QOP_MDWF_performance(&time_sec,
                               &flops,
                               &sent,
                               &received,
                               state) != 0 ) { 
        QDPIO::cerr << "MDWF_Error: "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Report status */
      QDPIO::cout << "MDWFQpropT Single Prec : status=" << status 
                  << " iterations=" << out_iters_single
                  << " resulting epsilon=" << sqrt(out_eps_single) << std::endl;
      
      /* Report Flops */
      FlopCounter flopcount_single;
      flopcount_single.reset();
      flopcount_single.addFlops(flops);
      flopcount_single.report("MDWFQpropT_Single_Prec:", time_sec);
      
      /* Export the solution */
      if( QOP_F3_MDWF_export_fermion(fermionWriter, 
                                     &psi,
                                     sprec_soln) != 0 ) { 
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Now I can free the fermions */
      QOP_F3_MDWF_free_fermion(&sprec_soln);
      QOP_F3_MDWF_free_fermion(&sprec_x0);
      QOP_F3_MDWF_free_fermion(&sprec_rhs);
      QOP_F3_MDWF_free_gauge(&sprec_gauge);
      
      res.n_count = out_iters_single;
      
    }
 
    /* DoublePrecision Branch. */
    {
      double out_eps_double;
      
      // OK Try to solve for the single precision 
      // Cast to single precision gauge field
      QOP_D3_MDWF_Gauge *dprec_gauge = NULL;
      if ( QOP_D3_MDWF_import_gauge(&dprec_gauge,
                                    state,
                                    gaugeReader,
                                    (void *)&u) != 0 ) { 
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Fermion fields */
      QOP_D3_MDWF_Fermion *dprec_rhs;  // Right hand side
      QOP_D3_MDWF_Fermion *dprec_x0;   // Guess
      QOP_D3_MDWF_Fermion *dprec_soln; // result
      
      if( QOP_D3_MDWF_import_fermion(&dprec_rhs, 
                                     state,
                                     fermionReader,
                                     (void *)&chi) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      if( QOP_D3_MDWF_import_fermion(&dprec_x0, 
                                     state,
                                     fermionReader,
                                     (void *)&psi) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      if( QOP_D3_MDWF_allocate_fermion(&dprec_soln, 
                                       state) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Do the solve */
      double target_epsilon = toDouble(invParam.RsdCGRestart*invParam.RsdCGRestart);
      int max_iteration = invParam.MaxCGRestart;
      int status; 
      
      
      QDPIO::cout << "MDWFQpropT: Beginning Double Precision Solve" << std::endl;
      if( ( status=QOP_D3_MDWF_DDW_CG(dprec_soln,
                                      &out_iters_double,
                                      &out_eps_double,
                                      params, 
                                      dprec_x0,
                                      dprec_gauge,
                                      dprec_rhs,
                                      max_iteration,
                                      target_epsilon, 
                                      QOP_MDWF_LOG_NONE)) != 0 ) {
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Get Perf counters after solve */
      if( QOP_MDWF_performance(&time_sec,
                               &flops,
                               &sent,
                               &received,
                               state) != 0 ) { 
        QDPIO::cerr << "MDWF_Error: "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      /* Reoirt Status */
      QDPIO::cout << "MDWFQpropT Double Prec: status=" << status 
                  << " iterations=" << out_iters_double
                  << " resulting epsilon=" << sqrt(out_eps_double) << std::endl;
      
      /* Report Flops */
      FlopCounter flopcount_double;
      flopcount_double.reset();
      flopcount_double.addFlops(flops);
      flopcount_double.report("MDWFQpropT_Double_Prec:", time_sec);
      
      /* Export the solution */
      if( QOP_D3_MDWF_export_fermion(fermionWriter, 
                                     &psi,
                                     dprec_soln) != 0 ) { 
        QDPIO::cerr << "MDWF Error:  "<< QOP_MDWF_error(state) << std::endl;
        QDP_abort(1);
      }
      
      /* Now I can free the fermions */
      QOP_D3_MDWF_free_fermion(&dprec_soln);
      QOP_D3_MDWF_free_fermion(&dprec_x0);
      QOP_D3_MDWF_free_fermion(&dprec_rhs);
      QOP_D3_MDWF_free_gauge(&dprec_gauge);
      
      /* Add the double prec iteration count onto the global count */
      res.n_count += out_iters_double;
    }
    
    // Compute actual residual
    {
      multi1d<LatticeFermion>  r(N5);
      A->unprecLinOp(r, psi, PLUS);
      r -= chi;
      res.resid = sqrt(norm2(r));
    }
    QDPIO::cout << "MDWF Final: single_iters=" << out_iters_single << " double_iters=" << out_iters_double << " total_iters=" << res.n_count << std::endl;
    QDPIO::cout << "MDWF Final: final absolute unprec residuum="<<res.resid<<std::endl;
    
    END_CODE();
    return res;
  }
  
  
  /* INIT Function */
  void MDWFQpropT::init(Handle< FermState<T,P,Q> > fermstate, const GroupXML_t& inv) 
  {
    START_CODE();
 
    if( Nd != 4 ) {
      QDPIO::cout << "This will only work for Nd=4" << std::endl;
      QDP_abort(1);
    }
 
    if( Nc != 3 ) {
      QDPIO::cout << "This will only work for Nc=3" << std::endl;
      QDP_abort(1);
    }
    
    // Read the XML for the CG params
    try {
      std::istringstream  is(inv.xml);
      XMLReader  paramtop(is);
      read(paramtop, inv.path, invParam);
    }
    catch (const std::string& e) {
      QDPIO::cerr << "CGDWFQpropT: only support a CG inverter" << std::endl;
      QDP_abort(1);
    }
    
    
    // I need to call Andrews init function.
    // For this I need a function that can tell me the 
    multi1d<int> lattice(5);
    multi1d<int> network(4);
    multi1d<int> node_coords(4);
    int master_p;
    
    // Lattice is the 5D lattice size
    for(int mu=0; mu < Nd; mu++) { 
      lattice[mu] = Layout::lattSize()[mu];
    }
    lattice[Nd]=N5;
    
    for(int mu=0; mu < Nd; mu++) { 
      network[mu] = Layout::logicalSize()[mu];
      node_coords[mu] = Layout::nodeCoord()[mu];
    }
    
    // Master_p has to be zero on master node and nonzero
    // elsewhere. This is odd.
    if( Layout::primaryNode() ) { 
      master_p = 0;
    }
    else { 
      master_p = 1;
    }
 
    // Announce a version just to be nice
    QDPIO::cout << "MDWFQpropT: Initializing MDWF Library Version " << QOP_MDWF_version() << std::endl;
 
    // OK. Let's call Andrew's routine
    if(  QOP_MDWF_init(&state, lattice.slice(), network.slice(),
                       node_coords.slice(), master_p, sublattice_func, 
                       NULL) != 0 ) { 
      // Nonzero return value => error
      QDPIO::cerr << "MDWF Error: " << QOP_MDWF_error(state) << std::endl;
      QDP_abort(1);
    }
    
    // Set up the masses etc...
    u.resize(Nd);
    u = fermstate->getLinks();
    Real ff = Real(1);
    
    if (anisoParam.anisoP) {
      ff = where(anisoParam.anisoP, anisoParam.nu / anisoParam.xi_0, Real(1));
      for(int mu=0; mu < u.size(); ++mu) {
        if (mu != anisoParam.t_dir)
          u[mu] *= ff;
      }
    }
    
    // Set the Shamir parameters for now
    {
      double a5 = (double)1;
      
      // Convention change. Now in the internal code it is 5 - OverMass
      // To allow using anisotropy, I subtract off the 5 and 'add it back on'
      // suitable for the anisotropic case
      double M5  = (double)(-5) + toDouble((double)1 + a5*((double)1 + (double)(Nd-1)*ff - OverMass));
      
      double m_f = toDouble(Mass);
      
      if(  QOP_MDWF_set_Shamir(&params, state, a5, M5 ,m_f) != 0){
        QDPIO::cerr << "MDWF Error: " << QOP_MDWF_error(state)<< std::endl;
        QDP_abort(1);
      }
      
    }
    
    END_CODE();
    return;
  }
  
  // Finalize - destructor call
  void MDWFQpropT::fini(void)  
  {
    START_CODE();
    QDPIO::cout << "MDWFQpropT: Finalizing MDWF Library Version " << QOP_MDWF_version() << std::endl;
 
    if (params != NULL) { 
      QOP_MDWF_free_parameters(&params);
    }
    
    if (state != NULL) { 
      QOP_MDWF_fini(&state);
    }
 
    END_CODE();
    return;
  }
    
 
  /*! @} */   // end of group qprop
}