fopr_Wilson_eo_impl.h

Go to the documentation of this file.

#ifndef FOPR_WILSON_EO_IMPL_INCLUDED
#define FOPR_WILSON_EO_IMPL_INCLUDED

#include "fopr_Wilson_eo.h"

#include "channel.h"

#include "bridgeIO.h"
using Bridge::vout;


class Fopr_Wilson_eo::Fopr_Wilson_eo_impl
{
 public:
  static const std::string class_name;

 private:
  int m_Nc, m_Nd, m_Nvc, m_Ndf;
  int m_Nvol, m_Nvol2, m_Ndim;
  int m_Nx, m_Ny, m_Nz, m_Nt, m_Nx2;

  double              m_kappa;      
  std::vector<int>    m_boundary;   
  std::vector<double> m_boundary2;  

  Bridge::VerboseLevel m_vl;

  std::string m_mode;    
  std::string m_repr;    

  Index_eo m_index;
  Field_G  *m_Ueo;
  Field_G  *m_U;    

  std::vector<int> m_Leo;
  Field            m_v1, m_v2, m_w1, m_w2; 

  double *vcp1_xp, *vcp2_xp, *vcp1_xm, *vcp2_xm;
  double *vcp1_yp, *vcp2_yp, *vcp1_ym, *vcp2_ym;
  double *vcp1_zp, *vcp2_zp, *vcp1_zm, *vcp2_zm;
  double *vcp1_tp, *vcp2_tp, *vcp1_tm, *vcp2_tm;

  // use async data transfer
  std::vector<int>       m_npe;
  std::vector<Channel *> m_fw_send;
  std::vector<Channel *> m_fw_recv;
  std::vector<Channel *> m_bw_send;
  std::vector<Channel *> m_bw_recv;

  void (Fopr_Wilson_eo::Fopr_Wilson_eo_impl::*m_gm5)(Field&, const Field&);
  void (Fopr_Wilson_eo::Fopr_Wilson_eo_impl::*m_gm5_self)(Field&);
  void (Fopr_Wilson_eo::Fopr_Wilson_eo_impl::*m_mult_tp)(Field&, const Field&, const int ieo);
  void (Fopr_Wilson_eo::Fopr_Wilson_eo_impl::*m_mult_tm)(Field&, const Field&, const int ieo);

 public:

  Fopr_Wilson_eo_impl(std::string repr)
  { init(repr);  }

  ~Fopr_Wilson_eo_impl()
  { tidyup();  }

  void set_parameters(const double kappa, const std::vector<int> bc);

  void set_config(Field *U);

  void prePropD(Field&, Field&, const Field&);
  void postPropD(Field&, const Field&, const Field&);
  void prePropDag(Field&, Field&, const Field&);
  void postPropDag(Field&, const Field&, const Field&);

  void D(Field& v, const Field& f);
  void Ddag(Field& v, const Field& f);
  void DdagD(Field& v, const Field& f);
  void DDdag(Field& v, const Field& f);
  void H(Field& v, const Field& f);

  // ieo=0: even <-- odd
  // ieo=1: odd  <-- even

  void Meo(Field&, const Field&, const int ieo);
  void Mdageo(Field&, const Field&, const int ieo);
  void MeoMoe(Field&, const Field&);
  void Meo_gm5(Field&, const Field&, const int ieo);

  void mult_gm5(Field&, const Field&);
  void mult_gm5(Field&);
  void gm5_dirac(Field&, const Field&);
  void gm5_chiral(Field&, const Field&);
  void gm5_self_dirac(Field&);
  void gm5_self_chiral(Field&);

  void gm5p(const int mu, Field&, const Field& v);

  int field_nvol() { return m_Nvol2; }
  int field_nin()  { return m_Nvc * m_Nd; }
  int field_nex()  { return 1; }

  double flop_count();

  void mult_p(int mu, Field_F&, const Field_F&, const int ieo);
  void mult_m(int mu, Field_F&, const Field_F&, const int ieo);

 private:
  void init(const std::string);
  void tidyup();

  void mult_xp(Field&, const Field&, const int ieo);
  void mult_xm(Field&, const Field&, const int ieo);
  void mult_yp(Field&, const Field&, const int ieo);
  void mult_ym(Field&, const Field&, const int ieo);
  void mult_zp(Field&, const Field&, const int ieo);
  void mult_zm(Field&, const Field&, const int ieo);
  void mult_tp_dirac(Field&, const Field&, const int ieo);
  void mult_tm_dirac(Field&, const Field&, const int ieo);
  void mult_tp_chiral(Field&, const Field&, const int ieo);
  void mult_tm_chiral(Field&, const Field&, const int ieo);

  void clear_impl(Field&);
  void scal_impl(Field&, double);

  // member data for threading
  int m_Mz, m_Mt;
  int m_Nthread, m_Ntask;
  int m_Ntask_z, m_Ntask_t;

  struct mult_arg
  {
    int isite;
    int isite_cpx, isite_cpy, isite_cpz, isite_cpt;
    int kz0, kz1, kt0, kt1;
  };
  std::vector<mult_arg> m_arg;

  // member functions for threading
  void setup_thread();
  void mult_xp1_thread(int, double *, const double *, int);
  void mult_xp2_thread(int, double *, const double *, int);
  void mult_xpb_thread(int, double *, const double *, int);
  void mult_xm1_thread(int, double *, const double *, int);
  void mult_xm2_thread(int, double *, const double *, int);
  void mult_xmb_thread(int, double *, const double *, int);
  void mult_yp1_thread(int, double *, const double *, int);
  void mult_yp2_thread(int, double *, const double *, int);
  void mult_ypb_thread(int, double *, const double *, int);
  void mult_ym1_thread(int, double *, const double *, int);
  void mult_ym2_thread(int, double *, const double *, int);
  void mult_ymb_thread(int, double *, const double *, int);
  void mult_zp1_thread(int, double *, const double *, int);
  void mult_zp2_thread(int, double *, const double *, int);
  void mult_zpb_thread(int, double *, const double *, int);
  void mult_zm1_thread(int, double *, const double *, int);
  void mult_zm2_thread(int, double *, const double *, int);
  void mult_zmb_thread(int, double *, const double *, int);
  void mult_tp1_dirac_thread(int, double *, const double *, int);
  void mult_tp2_dirac_thread(int, double *, const double *, int);
  void mult_tpb_dirac_thread(int, double *, const double *, int);
  void mult_tm1_dirac_thread(int, double *, const double *, int);
  void mult_tm2_dirac_thread(int, double *, const double *, int);
  void mult_tmb_dirac_thread(int, double *, const double *, int);
  void mult_tp1_chiral_thread(int, double *, const double *, int);
  void mult_tp2_chiral_thread(int, double *, const double *, int);
  void mult_tpb_chiral_thread(int, double *, const double *, int);
  void mult_tm1_chiral_thread(int, double *, const double *, int);
  void mult_tm2_chiral_thread(int, double *, const double *, int);
  void mult_tmb_chiral_thread(int, double *, const double *, int);
  void scal_thread(int, double *, double);
  void clear_thread(int, double *);
  void gm5_dirac_thread(int, double *, const double *);
  void gm5_chiral_thread(int, double *, const double *);
  void gm5_dirac_thread(int, double *);
  void gm5_chiral_thread(int, double *);
};
#endif