minvcg_accumulate_array.cc

Go to the documentation of this file.

 
/*! \file
 *  \brief Multishift Conjugate-Gradient algorithm for a Linear Operator
 */
 
#include "linearop.h"
#include "actions/ferm/invert/minvcg_accumulate_array.h"
 
using namespace QDP::Hints;
 
namespace Chroma 
{
 
 
  //! Multishift Conjugate-Gradient (CG1) algorithm for a  Linear Operator
  /*! \ingroup invert
   *
   * This subroutine uses the Conjugate Gradient (CG) algorithm to find
   * the solution of the set of linear equations
   * Method used is described in  Jegerlehner, hep-lat/9708029
 
   * We are searching in a subspace orthogonal to the eigenvectors EigVec
   * of A. The source chi is assumed to already be orthogonal!
 
   *        Chi  =  A . Psi
 
   * Algorithm:
 
   *  Psi[0] :=  0;                                   Zeroed
   *  r[0]   :=  Chi;                           Initial residual
   *  p[1]   :=  Chi ;                        Initial direction
   *  b[0]   := |r[0]|**2 / <p[0],Ap[0]> ;
   *  z[0]   := 1 / (1 - (shift - shift(0))*b) 
   *  bs[0]  := b[0] * z[0]  
   *  r[1] += b[k] A . p[0] ;                 New residual
   *  Psi[1] = - b[k] p[k] ;                  Starting solution std::vector
   *  IF |r[0]| <= RsdCG |Chi| THEN RETURN;        Converged?
   *  FOR k FROM 1 TO MaxCG DO                         CG iterations
   *      a[k] := |r[k]|**2 / |r[k-1]|**2 ;
   *      p[k] := r[k] + a[k] p[k-1];                  New direction
   *      b[k+1] := |r[k]|**2 / <p[k],Ap[k]> ;
   *      r[k+1] += b[k+1] A . p[k] ;                  New residual
   *      Psi[k+1] -= b[k+1] p[k] ;                    New solution std::vector
   *      IF |[k+1]| <= RsdCG |Chi| THEN RETURN;    Converged?
 
   * Arguments:
 
   *  A         Hermitian linear operator      (Read)
   *  Chi               Source                         (Read)
   *  Psi               array of solutions             (Write)
   *  shifts        shifts of form  A + mass       (Read)
   *  RsdCG       residual accuracy              (Read/Write)
   *  n_count     Number of CG iteration               (Write)
 
   * Local Variables:
 
   *  p                Direction std::vector
   *  r                Residual std::vector
   *  cp               | r[k] |**2
   *  c                | r[k-1] |**2
   *  k                CG iteration counter
   *  a                a[k]
   *  b                b[k+1]
   *  d                < p[k], A.p[k] >
   *  Ap               Temporary for  M.p
 
   *  MaxCG       Maximum number of CG iterations allowed
 
   * Subroutines:
   *  A        Apply matrix hermitian A to std::vector 
   */
 
  template<typename T, typename C>
  void MInvCGAccum_a(const C& A, 
                     const multi1d<T>& chi, 
                     multi1d<T>& psi,
                     const Real& norm,
                     const multi1d<Real>& residues,
                     const multi1d<Real>& poles, 
                     const Real& RsdCG, 
                     const int MaxCG,
                     int& n_count)
  {
    START_CODE();
 
    // Setup
    const Subset& sub = A.subset();
    int n_shift = poles.size();
    int N = A.size();
 
    FlopCounter flopcount;
    StopWatch swatch;
 
 
    if (n_shift == 0) {
      QDPIO::cerr << "MinvCG: You must supply at least 1 shift. shift.size() = " << n_shift << std::endl;
      QDP_abort(1);
    }
 
    /* Now find the smallest mass */
    int isz = 0;
    for(int findit=1; findit < n_shift; ++findit) {
      if ( toBool( poles[findit] < poles[isz])  ) { 
        isz = findit;
      }
    }
 
    /* Keep the solutions here */
    multi1d< multi1d<T> > X(n_shift);
 
    // Now arrange the memory
    // The outer size of the multi1d is n_shift
    X.resize(n_shift);
 
    // For this algorithm, all the X have to be 0 to start
    // Only is that way the initial residuum r = chi
    for(int i= 0; i < n_shift; ++i) { 
      X[i].resize(N);
    }
 
    psi.resize(N);
    for(int i=0; i < N; i++) {
      psi[i] = zero;
    }
 
 
    multi1d<T> r(N);
    multi1d< multi1d<T> > p(n_shift);
 
    for(int s = 0; s < n_shift; ++s) {
      p[s].resize(N); 
    }
 
    multi1d<T> Ap(N);
    multi1d<T> chi_internal(N);
 
    // Now arrange the memory:
    // These guys get most hits so locate them first
    moveToFastMemoryHint(Ap);
    moveToFastMemoryHint(p[isz]); 
    moveToFastMemoryHint(r);
    moveToFastMemoryHint(X[isz]);
    moveToFastMemoryHint(chi_internal);
 
    { 
      multi1d<T> blocker1(N); moveToFastMemoryHint(blocker1);
      multi1d<T> blocker2(N); moveToFastMemoryHint(blocker2);
  
      // Now the rest of the p-s
      for(int i=0; i < n_shift; i++) { 
        if( i!=isz ) { 
          moveToFastMemoryHint(p[i]);
          moveToFastMemoryHint(X[i]);
        }
      }
      
      // Blockers get freed here
    }
 
    moveToFastMemoryHint(psi);
 
    // initialise X-s
    for(int i=0; i < n_shift; i++) { 
      for(int n=0; n < N; ++n) {
        X[i][n][sub] = zero;
      }
    }
 
    // initialise chi internal
    for(int i=0; i < N; i++) { 
      chi_internal[i][sub] = chi[i];
    }
 
    // -------- All memory setup and copies done. Timer starts here
    QDPIO::cout << "MinvCG starting" << std::endl;
    flopcount.reset();
    swatch.reset();
    swatch.start();
 
    // If chi has zero norm then the result is zero
    Double chi_norm_sq = norm2(chi_internal,sub);
    flopcount.addSiteFlops(4*Nc*Ns*N,sub);
 
    Double chi_norm = sqrt(chi_norm_sq);
 
    if( toBool( chi_norm < fuzz )) { 
      n_count = 0;
      swatch.stop();
      QDPIO::cout << "MinvCG: Finished. Iters taken = " << n_count << std::endl;
      flopcount.report("MinvCGArray", swatch.getTimeInSeconds());
 
 
      // Accumulate psi
      for(int n=0; n < N; n++) { 
        psi[n][sub]=norm*chi_internal[n];
      }
      for(int s=0; s < n_shift; s++) {
        for(int n=0; n < N; n++) { 
          psi[n][sub] += residues[s]*X[s][n];
        }
      }
 
      // Revert X-s
      for(int i=0; i < n_shift; i++) { 
        revertFromFastMemoryHint(X[i], false);
      }
 
      revertFromFastMemoryHint(psi, true);
      
 
      END_CODE();
      return;
    }
 
    multi1d<Double> rsd_sq(n_shift);
    multi1d<Double> rsdcg_sq(n_shift);
 
    Double cp = chi_norm_sq;
    for(int s = 0; s < n_shift; ++s)  {
      rsdcg_sq[s] = RsdCG * RsdCG;  // RsdCG^2
      rsd_sq[s] = Real(cp) * rsdcg_sq[s]; // || chi ||^2 RsdCG^2
    }
 
  
    // r[0] := p[0] := Chi 
    for(int n=0; n < N; ++n) {
      r[n][sub] = chi_internal[n];
 
      for(int s=0; s < n_shift; s++) {
        p[s][n][sub] = chi_internal[n];
      }
    }
 
 
 
    //  b[0] := - | r[0] |**2 / < p[0], Ap[0] > ;/
    //  First compute  d  =  < p, A.p > 
    //  Ap = A . p  */
    A(Ap, p[isz], PLUS);
    flopcount.addFlops(A.nFlops());
 
    for(int n=0; n < N; ++n) {
      Ap[n][sub] += poles[isz] * p[isz][n];
    }
    flopcount.addSiteFlops(4*Nc*Ns*N, sub);
 
    /*  d =  < p, A.p >  */
    Double d = innerProductReal(p[isz], Ap, sub);
    flopcount.addSiteFlops(4*Nc*Ns*N ,sub);
 
    Double b = -cp/d;
 
    /* Compute the shifted bs and z */
    multi1d<Double> bs(n_shift);
    multi2d<Double> z(2, n_shift);
    int iz;
 
    z[0][isz] = Double(1);
    z[1][isz] = Double(1);
    bs[isz] = b;
    iz = 1;
 
    for(int s = 0; s < n_shift; ++s)
    {
      if( s != isz ) {
        z[1-iz][s] = Double(1);
        z[iz][s] = Double(1) / (Double(1) - (Double(poles[s])-Double(poles[isz]))*b);
        bs[s] = b * z[iz][s];
      }
    }
 
    //  r[1] += b[0] A . p[0]; 
    for(int n=0; n < N; ++n) {
      r[n][sub] += Real(b)* Ap[n];
    }
    flopcount.addSiteFlops(4*Nc*Ns*N, sub);
 
    //  X[1] -= b[0] p[0] = - b[0] chi;
    // X[0] are all 0 so I can write this as a -= bs*chi
    for(int s = 0; s < n_shift; ++s) {
      for(int n=0; n < N; ++n) {
        X[s][n][sub] -= Real(bs[s])*chi_internal[n];
      }
    }
    flopcount.addSiteFlops(4*Nc*Ns*N*n_shift, sub);
 
 
    //  c = |r[1]|^2   
    Double c = norm2(r,sub);                    
    flopcount.addSiteFlops(4*Nc*Ns*N,sub);
 
 
    // Check convergence of first solution
    multi1d<bool> convsP(n_shift);
    for(int s = 0; s < n_shift; ++s) {
      convsP[s] = false;
    }
 
    bool convP = toBool( c < rsd_sq[isz] );
 
 
    //  FOR k FROM 1 TO MaxCG DO
    //  IF |X[k+1] - X[k]| <= RsdCG |X[k+1]| THEN RETURN; 
    Double z0, z1;
    Double ztmp;
    Double cs;
    Double a;
    Double as;
    Double  bp;
    int k;
 
 
    for(k = 1; k <= MaxCG && !convP ; ++k)
    {
      //  a[k+1] := |r[k]|**2 / |r[k-1]|**2 ; 
      a = c/cp;
 
      //  p[k+1] := r[k+1] + a[k+1] p[k]; 
      //  Compute the shifted as */
      //  ps[k+1] := zs[k+1] r[k+1] + a[k+1] ps[k];
      for(int s = 0; s < n_shift; ++s) {
 
        // Always update p[isz] even if isz is converged
        // since the other p-s depend on it.
        if (s == isz) {
 
          for(int n=0; n < N; ++n) {
            p[s][n][sub] = r[n] + Real(a)*p[s][n];
            //      p[s][n][sub] *= Real(a);            
            //    p[s][n][sub] += r[n];                 
          }
          flopcount.addSiteFlops(4*Nc*Ns*N,sub);
 
        }
        else {
          // Don't update other p-s if converged.
          if( ! convsP[s] ) { 
            as = a * z[iz][s]*bs[s] / (z[1-iz][s]*b);
          
            for(int n=0; n < N; ++n) {
              p[s][n][sub] = Real(as)*p[s][n] + Real(z[iz][s])*r[n];
              //p[s][n][sub] *= Real(as);                
              //p[s][n][sub] += Real(z[iz][s])*r[n];     
            }
            flopcount.addSiteFlops(6*Nc*Ns*N, sub);
          }
        }
 
      }
 
      //  cp  =  | r[k] |**2 
      cp = c;
 
      //  b[k] := | r[k] |**2 / < p[k], Ap[k] > ;
      //  First compute  d  =  < p, A.p >  
      //  Ap = A . p 
      A(Ap, p[isz], PLUS);
      flopcount.addFlops(A.nFlops());
 
      for(int n=0; n < N; ++n) {
        Ap[n][sub] += poles[isz] *p[isz][n];
      }
      flopcount.addSiteFlops(4*Nc*Ns*N, sub);
 
      /*  d =  < p, A.p >  */
      d = innerProductReal(p[isz], Ap, sub);
      flopcount.addSiteFlops(4*Nc*Ns*N, sub);
    
      bp = b;
      b = -cp/d;
 
      // Compute the shifted bs and z 
      bs[isz] = b;
      iz = 1 - iz;
      for(int s = 0; s < n_shift; s++) {
      
        if (s != isz && !convsP[s] ) {
          z0 = z[1-iz][s];
          z1 = z[iz][s];
          z[iz][s] = z0*z1*bp;
          z[iz][s] /= b*a*(z1-z0) + z1*bp*(Double(1) - (poles[s] - poles[isz])*b);
          bs[s] = b*z[iz][s]/z0;
        }
      }
 
      //  r[k+1] += b[k] A . p[k] ; 
      for(int n=0; n < N; ++n) {
        r[n][sub] += Real(b)*Ap[n];
      }
      flopcount.addSiteFlops(4*Nc*Ns*N, sub);
 
 
      // Check convergence
      //  c  =  | r[k] |**2 
      c = norm2(r,sub);                 
      flopcount.addSiteFlops(4*Nc*Ns*N, sub);
 
      //    IF |psi[k+1] - psi[k]| <= RsdCG |psi[k+1]| THEN RETURN;
      // or IF |r[k+1]| <= RsdCG |chi| THEN RETURN;
      convP = true;
      for(int s = 0; s < n_shift; s++) {
        if (! convsP[s] ) {
 
 
          // Convergence methods 
          // Check norm of shifted residuals 
          Double css = c * z[iz][s]* z[iz][s];
          convsP[s] = toBool(  css < rsd_sq[s] );
 
        }
        convP &= convsP[s];
      }
 
      // Check accumulated result -- modify X-s in the same loop
      multi1d<T> Delta(N);
      for(int n=0; n < N; n++) { 
        Delta[n][sub] = zero;
        psi[n][sub] = norm*chi_internal[n];
      }
      for(int s=0 ; s < n_shift; s++) { 
        if ( !convsP[s] ) { 
          multi1d<T> tmp(N);
          for(int n=0; n < N; n++) { 
            tmp[n][sub] = Real(bs[s])*p[s][n];
            X[s][n][sub] -= tmp[n];
            Delta[n][sub] += residues[s]*tmp[n];
          }
        }
        // Always use the all the new X-s to update psi.
        for(int n=0; n < N; n++) { 
          psi[n][sub] += residues[s]*X[s][n];
        }
      }
 
      Double delta_norm = norm2(Delta, sub);
      Double psi_norm = norm2(psi,sub);
      convP |= toBool( delta_norm < rsdcg_sq[0]*psi_norm);
      n_count = k;
    }
 
    swatch.stop();
    QDPIO::cout << "MinvCGAccumArray finished: " << n_count << " iterations " << std::endl;
    flopcount.report("MinvCGAccumArray", swatch.getTimeInSeconds());
 
    for(int i=0; i < n_shift; i++) { 
      revertFromFastMemoryHint(psi[i],true);
    }
 
    if (n_count == MaxCG) {
      QDPIO::cout << "too many CG iterationns: " << n_count << std::endl;
      QDP_abort(1);
 
    }
    END_CODE();
    return;
  }
 
 
  /*! \ingroup invert */
  template<>
  void MInvCGAccum(const LinearOperatorArray<LatticeFermion>& M,
                   const multi1d<LatticeFermion>& chi, 
                   multi1d<LatticeFermion>& psi, 
                   const Real& norm,
                   const multi1d<Real>& residues,
                   const multi1d<Real>& poles,
                   const Real& RsdCG, 
                   const int MaxCG,
                   int &n_count)
  {
    MInvCGAccum_a(M, chi, psi, norm, residues, poles, RsdCG, MaxCG, n_count);
  }
 
 
  /*! \ingroup invert */
  template<>
  void MInvCGAccum(const DiffLinearOperatorArray<LatticeFermion, 
                                            multi1d<LatticeColorMatrix>, 
                                            multi1d<LatticeColorMatrix> >& M,
                   const multi1d<LatticeFermion>& chi, 
                   multi1d<LatticeFermion>& psi, 
                   const Real& norm,
                   const multi1d<Real>& residues,
                   const multi1d<Real>& poles,
                   const Real& RsdCG, 
                   const int MaxCG,
                   int &n_count)
  {
    MInvCGAccum_a(M, chi, psi, norm, residues, poles, RsdCG, MaxCG, n_count);
  }
 
}  // end namespace Chroma