/Users/mdipierro/fermiqcd/development/Libraries/fermiqcd_cg_inverter.h

Go to the documentation of this file.

class CG2 {
 public:  
  template <class fieldT, class fieldG>
  static inversion_stats inverter(fieldT &psi_out, 
                                  fieldT &psi_in, 
                                  fieldG &U, 
                                  coefficients &coeff, 
                                  mdp_real absolute_precision=mdp_precision,
                                  mdp_real relative_precision=0,
                                  int max_steps=2000,
                                  bool qdaggerq=false) {
    mpi.begin_function("ConjugateGradientInverter");
    int             step=0;
    double          residue, rresidue=-1, old_rresidue;
    double          time=mpi.time();
    inversion_stats stats;

    fieldT r(psi_in);
    fieldT rnew(psi_in);
    fieldT p(psi_in);
    fieldT pnew(psi_in);
    fieldT t(psi_in);
    fieldT s(psi_in);
    fieldT solnew(psi_in);
    fieldT b(psi_in);
        
    double alpha, beta, rrtmp, rrtmp2, psdot;

    //mpi << "\tstep\tresidue\t\ttime (sec)\n";


    int mulQ=1, guesszero=1,debug=0;

    psi_out=0; 
    r=psi_in;
    p=r;    

    //mpi << "\t<target>\n" 
        //    << "\t\t<max_steps>" << max_steps << "</max_steps>\n"
        //    << "\t\t<absolute_precision>" << absolute_precision << "</absolute_precision>\n"
        //    << "\t\t<relative_precision>" << relative_precision << "</relative_precision>\n"
        //    << "\t</target>\n";
         
    do {
      p.update();
      mul_Q(t,p,U,coeff); t/=2.0*coeff["kappa"]; t.update();
      coeff["sign"]=-1; 
      mul_Q(s,t,U,coeff);
      coeff["sign"]=1;
      s/=2.0*coeff["kappa"];
      rrtmp=real_scalar_product(r,r);           
      psdot=real_scalar_product(s,p);          
      alpha=rrtmp/psdot;
      rnew=r;
      mdp_add_scaled_field(rnew, -alpha, s);    
      solnew=psi_out;
      mdp_add_scaled_field(solnew, alpha, p);
      rrtmp2=real_scalar_product(rnew,rnew);      
      beta=rrtmp2/rrtmp;          
      pnew=rnew;
      mdp_add_scaled_field(pnew, beta, p);
      old_rresidue=rresidue;
      residue=sqrt(rrtmp2/(psi_in.global_size()));        
      rresidue=relative_residue(rnew,psi_out);         
      p=pnew;
      r=rnew;
      psi_out=solnew;     
      step++;
      
      //mpi << "\t\t<step>" << step << "</step>\n"
      //    << "\t\t<residue>" << residue << "</residue>\n"
      //    << "\t\t<relative_residue>" << rresidue << "</relative_residue>\n"
      //    << "\t\t<time>" << mpi.time()-time << "</time>\n\n";

      // debug: cout << residue << endl;

      //if((step>10) && (rresidue==old_rresidue))
      //    error("not converging"); 
      step++;
      
    } while (residue>absolute_precision && 
             rresidue>relative_precision && 
             step<max_steps);
    
    if(!qdaggerq) {
      solnew.update();
      mul_Q(psi_out,solnew,U,coeff);
    }

    psi_out.update();
    
    stats.target_absolute_precision=absolute_precision;
    stats.target_relative_precision=relative_precision;
    stats.max_steps=max_steps;
    stats.absolute_precision=residue;
    stats.relative_precision=rresidue;
    stats.residue=residue;
    stats.steps=step;
    stats.mul_Q_steps=2*step+1;
    stats.time=mpi.time()-time;

    mpi << "\t<stats>\n" 
        << "\t\t<max_steps>" << step << "</max_steps>\n"
        << "\t\t<absolute_precision>" << residue << "</absolute_precision>\n"
        << "\t\t<relative_precision>" << rresidue << "</relative_precision>\n"
        << "\t\t<time>" << stats.time << "</time>\n"
        << "\t</stats>\n";
    
    mpi.end_function("ConjugateGradientInverter");
    return stats;
  }
};

---