afopr_Wilson-tmpl.h

Go to the documentation of this file.

 
#include "lib/Parameters/commonParameters.h"
#include "lib/IO/bridgeIO.h"
using Bridge::vout;
 
template<typename AFIELD>
const std::string AFopr_Wilson<AFIELD>::class_name
                                           = "AFopr_Wilson<AFIELD>";
 
//====================================================================
namespace{
  inline void set_kernel_thread(int& ith, int& nth, int& ith_kernel)
  {
    ith = ThreadManager::get_thread_id();
    nth = ThreadManager::get_num_threads();
    ith_kernel = 0;
    if(nth > 1) ith_kernel = 1;
  }
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::init(const Parameters& params)
{
  ThreadManager::assert_single_thread(class_name);
 
  std::string vlevel;
  if (!params.fetch_string("verbose_level", vlevel)) {
    m_vl = vout.set_verbose_level(vlevel);
  } else {
    m_vl = CommonParameters::Vlevel();
  }
 
  vout.general(m_vl, "%s: construction\n", class_name.c_str());
  vout.increase_indent();
 
  m_Nc = CommonParameters::Nc();
  m_Nd = CommonParameters::Nd();
  m_Nst  = CommonParameters::Nvol();
  m_Ndim = CommonParameters::Ndim();
  int Nvc = m_Nc * 2;
  int Ndf = 2 * m_Nc * m_Nc;
  int Nx  = CommonParameters::Nx();
  int Ny  = CommonParameters::Ny();
  int Nz  = CommonParameters::Nz();
  int Nt  = CommonParameters::Nt();
 
  m_Nsize[0] = Nx;
  m_Nsize[1] = Ny;
  m_Nsize[2] = Nz;
  m_Nsize[3] = Nt;
 
  // switches for coomunication
  int req_comm = 0;  //  0: communication only if necessary
                     //  1: communication is enforced any time
  if (!params.fetch_int("require_communication", req_comm)) {
    vout.general(m_vl, "req_comm = %d (input)\n", req_comm);
  } else {
    vout.general(m_vl, "req_comm = %d (default)\n", req_comm);
  }
 
  do_comm_any = 0;
  for(int mu = 0; mu < m_Ndim; ++mu){
    do_comm[mu] = 1;
    if(req_comm == 0 && Communicator::npe(mu) == 1) do_comm[mu] = 0;
    do_comm_any += do_comm[mu];
    vout.general(m_vl, "do_comm[%d] = %d\n", mu, do_comm[mu]);
  }
 
  m_Nbdsize.resize(m_Ndim);
  int Nd2 = m_Nd/2;
  m_Nbdsize[0] = Nvc * Nd2 * ceil_nwp(Ny * Nz * Nt);
  m_Nbdsize[1] = Nvc * Nd2 * ceil_nwp(Nx * Nz * Nt);
  m_Nbdsize[2] = Nvc * Nd2 * ceil_nwp(Nx * Ny * Nt);
  m_Nbdsize[3] = Nvc * Nd2 * ceil_nwp(Nx * Ny * Nz);
 
  setup_channels();
 
  // gauge configuration.
  m_U.reset(Ndf, m_Nst, m_Ndim);
 
  // working vectors.
  int NinF = 2 * m_Nc * m_Nd;
  m_v2.reset(NinF, m_Nst, 1);
 
  set_parameters(params);
 
  vout.decrease_indent();
  vout.general(m_vl, "%s: construction finished.\n",
               class_name.c_str());
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::setup_channels()
{
  ThreadManager::assert_single_thread(class_name);
 
  chsend_up.resize(m_Ndim);
  chrecv_up.resize(m_Ndim);
  chsend_dn.resize(m_Ndim);
  chrecv_dn.resize(m_Ndim);
 
  for(int mu = 0; mu < m_Ndim; ++mu){
 
    int Nvsize = m_Nbdsize[mu] * sizeof(real_t);
 
    chsend_dn[mu].send_init(Nvsize, mu, -1);
    chsend_up[mu].send_init(Nvsize, mu,  1);
#ifdef USE_MPI
    chrecv_up[mu].recv_init(Nvsize, mu,  1);
    chrecv_dn[mu].recv_init(Nvsize, mu, -1);
#else
    void* buf_up = (void*)chsend_dn[mu].ptr();
    chrecv_up[mu].recv_init(Nvsize, mu,  1, buf_up);
    void* buf_dn = (void*)chsend_up[mu].ptr();
    chrecv_dn[mu].recv_init(Nvsize, mu, -1, buf_dn);
#endif
 
    if(do_comm[mu] == 1){
      chset_send.append(chsend_up[mu]);
      chset_send.append(chsend_dn[mu]);
      chset_recv.append(chrecv_up[mu]);
      chset_recv.append(chrecv_dn[mu]);
    }
 
    // openacc device memory allocation
#ifdef USE_MPI
    real_t* buf_dn1 = (real_t*)chsend_dn[mu].ptr();
    real_t* buf_dn2 = (real_t*)chrecv_dn[mu].ptr();
    real_t* buf_up1 = (real_t*)chsend_up[mu].ptr();
    real_t* buf_up2 = (real_t*)chrecv_up[mu].ptr();
    BridgeACC::afield_init(buf_dn1, m_Nbdsize[mu]);
    BridgeACC::afield_init(buf_dn2, m_Nbdsize[mu]);
    BridgeACC::afield_init(buf_up1, m_Nbdsize[mu]);
    BridgeACC::afield_init(buf_up2, m_Nbdsize[mu]);
#else
    BridgeACC::afield_init((real_t*)buf_up, m_Nbdsize[mu]);
    BridgeACC::afield_init((real_t*)buf_dn, m_Nbdsize[mu]);
#endif
 
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::tidyup()
{
  ThreadManager::assert_single_thread(class_name);
 
  // openacc device memory clean up
  for(int mu = 0; mu < m_Ndim; ++mu){
 
#ifdef USE_MPI
    real_t* buf_dn1 = (real_t*)chsend_dn[mu].ptr();
    real_t* buf_dn2 = (real_t*)chrecv_dn[mu].ptr();
    real_t* buf_up1 = (real_t*)chrecv_up[mu].ptr();
    real_t* buf_up2 = (real_t*)chsend_up[mu].ptr();
    BridgeACC::afield_tidyup(buf_dn1, m_Nbdsize[mu]);
    BridgeACC::afield_tidyup(buf_dn2, m_Nbdsize[mu]);
    BridgeACC::afield_tidyup(buf_up1, m_Nbdsize[mu]);
    BridgeACC::afield_tidyup(buf_up2, m_Nbdsize[mu]);
#else
    real_t* buf_up = (real_t*)chsend_up[mu].ptr();
    real_t* buf_dn = (real_t*)chsend_dn[mu].ptr();
    BridgeACC::afield_tidyup(buf_up, m_Nbdsize[mu]);
    BridgeACC::afield_tidyup(buf_dn, m_Nbdsize[mu]);
#endif
 
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::set_parameters(const Parameters& params)
{
#pragma omp barrier
 
  std::string vlevel;
  if (!params.fetch_string("verbose_level", vlevel)) {
    m_vl = vout.set_verbose_level(vlevel);
  }
 
  //- fetch and check input parameters
  double kappa;
  std::vector<int> bc;
 
  int err = 0;
  err += params.fetch_double("hopping_parameter", kappa);
  err += params.fetch_int_vector("boundary_condition", bc);
  if (err) {
    vout.crucial(m_vl, "Error at %s: input parameter not found.\n",
                                                class_name.c_str());
    exit(EXIT_FAILURE);
 }
 
  //- setting gamma matrix representation
  std::string repr;
  err = params.fetch_string("gamma_matrix_type", repr);
  if(err){
    vout.general(m_vl, "  gamma_matrix_type is not given - set to Dirac\n");
    m_repr = DIRAC;
  }else if(repr == "Dirac"){
    m_repr = DIRAC;
  }else if(repr == "Chiral"){
    m_repr = CHIRAL;
  }else{
    vout.crucial(m_vl, "Error in %s: irrelevant gamma_matrix_type: %s\n",
                class_name.c_str(), repr.c_str());
    exit(EXIT_FAILURE);
  }
 
  set_parameters(real_t(kappa), bc);
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::set_parameters(const real_t CKs,
                                          const std::vector<int> bc)
{
  assert(bc.size() == m_Ndim);
 
#pragma omp barrier
 
  int ith = ThreadManager::get_thread_id();
 
  if (ith == 0) {
    m_CKs = CKs;
    m_boundary.resize(m_Ndim);
    for (int mu = 0; mu < m_Ndim; ++mu) {
      m_boundary[mu] = bc[mu];
    }
 
    for (int mu = 0; mu < m_Ndim; ++mu) {
      m_bc[mu]  = 1;
      if(do_comm[mu] > 0){ // do communication
        if(Communicator::ipe(mu) == 0) m_bc[mu]  = m_boundary[mu];
        m_bc2[mu] = 0;
      }else{  // no communication
        m_bc[mu]  = 0;    // for boundary part (dummy)
        m_bc2[mu] = m_boundary[mu];  // for bulk part
      }
    }
  }
#pragma omp barrier
 
  //- print input parameters
  vout.general(m_vl, "%s: input parameters\n", class_name.c_str());
  if(m_repr == DIRAC){
    vout.general(m_vl, "  gamma-matrix type = Dirac\n");
  }else{
    vout.general(m_vl, "  gamma-matrix type = Chiral\n");
  }
  vout.general(m_vl, "  kappa = %8.4f\n", m_CKs);
  for (int mu = 0; mu < m_Ndim; ++mu) {
    vout.general(m_vl, "  boundary[%d] = %2d\n", mu, m_boundary[mu]);
  }
 
#pragma omp barrier
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::set_config(Field *u)
{
  int nth = ThreadManager::get_num_threads();
 
  vout.paranoiac(m_vl, "%s: set_config is called: num_threads = %d\n",
                 class_name.c_str(), nth);
 
  if (nth > 1) {
    set_config_impl(u);
  } else {
    set_config_omp(u);
  }
 
  vout.paranoiac(m_vl, "%s: set_config finished\n", class_name.c_str());
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::set_config_omp(Field *u)
{
  vout.paranoiac(m_vl, "  set_config_omp is called.\n");
 
#pragma omp parallel
 {
  set_config_impl(u);
 }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::set_config_impl(Field* u)
{
#pragma omp barrier
 
  int ith = ThreadManager::get_thread_id();
 
  if (ith == 0) m_conf = u;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  convert_gauge(index_lex, m_U, *u);
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::get_parameters(Parameters& params) const
{
  params.set_double("hopping_parameter", double(m_CKs));
  params.set_int_vector("boundary_condition", m_boundary);
 
  std::string repr;
  if(m_repr == DIRAC) repr  = "Dirac";
  if(m_repr == CHIRAL) repr = "Chiral";
  params.set_string("gamma_matrix_type", repr);
 
  params.set_string("verbose_level", vout.get_verbose_level(m_vl));
}
 
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::set_mode(std::string mode)
{
#pragma omp barrier
 
  int ith = ThreadManager::get_thread_id();
  if (ith == 0) m_mode = mode;
 
#pragma omp barrier
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult(AFIELD &v, const AFIELD &w)
{
  if (m_mode == "D") {
    mult_D(v, w);
    // mult_D_alt(v, w);
  } else if (m_mode == "DdagD") {
    mult_DdagD(v, w);
  } else if (m_mode == "Ddag") {
    mult_Ddag(v, w);
  } else if (m_mode == "H") {
    mult_H(v, w);
  } else {
    vout.crucial(m_vl, "%s: undefined mode = %s\n",
                 class_name.c_str(), m_mode.c_str());
    exit(EXIT_FAILURE);
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_dag(AFIELD &v, const AFIELD &w)
{
  if (m_mode == "D") {
    mult_Ddag(v, w);
  } else if (m_mode == "DdagD") {
    mult_DdagD(v, w);
  } else if (m_mode == "Ddag") {
    mult_D(v, w);
  } else if (m_mode == "H") {
    mult_H(v, w);
  } else {
    vout.crucial(m_vl, "%s: undefined mode = %s\n",
                 class_name.c_str(), m_mode.c_str());
    exit(EXIT_FAILURE);
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult(AFIELD &v, const AFIELD &w,
                                const std::string mode)
{
  if(mode == "D"){
    mult_D(v, w);
  }else if(mode == "Ddag"){
    mult_Ddag(v, w);
  }else if(mode == "DdagD"){
    mult_DdagD(v, w);
  }else if(mode == "H"){
    mult_H(v, w);
  }else{
    vout.crucial(m_vl, "%s: illegal mode is given to mult with mode\n",
                 class_name.c_str());
    exit(EXIT_FAILURE);
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_dag(AFIELD &v, const AFIELD &w,
                                    const std::string mode)
{
  if(mode == "D"){
    mult_Ddag(v, w);
  }else if(mode == "Ddag"){
    mult_D(v, w);
  }else if(mode == "DdagD"){
    mult_DdagD(v, w);
  }else if(mode == "H"){
    mult_H(v, w);
  }else{
    vout.crucial(m_vl, "%s: illegal mode is given to mult with mode\n",
                 class_name.c_str());
    exit(EXIT_FAILURE);
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_gm5(AFIELD &v, const AFIELD &w)
{
#pragma omp barrier
 
  int ith = ThreadManager::get_thread_id();
 
  if (ith == 0){
    real_t* vp = v.ptr(0);
    real_t* wp = const_cast<AFIELD*>(&w)->ptr(0);
 
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_gm5_dirac(vp, wp, m_Nsize, NC);
    }else{
      BridgeACC::mult_wilson_gm5_chiral(vp, wp, m_Nsize, NC);
    }
  }
 
#pragma omp barrier
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_D(AFIELD &v, const AFIELD &w)
{
  mult_DorH(v, w, 0);  // D
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_DdagD(AFIELD &v, const AFIELD &w)
{
  mult_DorH(m_v2, w, 1);  // H
  mult_DorH(v, m_v2, 1);  // H
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_Ddag(AFIELD &v, const AFIELD &w)
{
  mult_gm5(m_v2, w);
  mult_DorH(v, m_v2, 1);  // H
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_H(AFIELD &v, const AFIELD &w)
{
  mult_DorH(v, w, 1);  // H
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_H_alt(AFIELD &v, const AFIELD &w)
{
  mult_D_alt(m_v2, w);
  mult_gm5(v, m_v2);
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_DorH(AFIELD &v, const AFIELD &w,
                                     const int flag)
{
#pragma omp barrier
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  real_t *v2 = v.ptr(0);
  real_t *v1 = const_cast<AFIELD*>(&w)->ptr(0);
  real_t *u  = m_U.ptr(0);
 
  if (do_comm_any > 0 && ith == ith_kernel){
 
    real_t *buf1xp = (real_t*)chsend_dn[0].ptr();
    real_t *buf1xm = (real_t*)chsend_up[0].ptr();
 
    real_t *buf1yp = (real_t*)chsend_dn[1].ptr();
    real_t *buf1ym = (real_t*)chsend_up[1].ptr();
 
    real_t *buf1zp = (real_t*)chsend_dn[2].ptr();
    real_t *buf1zm = (real_t*)chsend_up[2].ptr();
 
    real_t *buf1tp = (real_t*)chsend_dn[3].ptr();
    real_t *buf1tm = (real_t*)chsend_up[3].ptr();
 
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_1_dirac(
                           buf1xp, buf1xm, buf1yp, buf1ym, 
                           buf1zp, buf1zm, buf1tp, buf1tm,
                           u, v1, m_Nsize, m_bc, do_comm);
    }else{
       BridgeACC::mult_wilson_1_chiral(
                           buf1xp, buf1xm, buf1yp, buf1ym, 
                           buf1zp, buf1zm, buf1tp, buf1tm,
                           u, v1, m_Nsize, m_bc, do_comm);
    }
 
  }
#pragma omp barrier
 
  if(do_comm_any > 0 && ith == 0){
    chset_recv.start();
    chset_send.start();
  }
 
  // bulk part
  if (ith == ith_kernel){
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_D_dirac(v2, u, v1,
                                     m_CKs, m_Nsize, m_bc2, flag);
    }else{
      BridgeACC::mult_wilson_D_chiral(v2, u, v1,
                                      m_CKs, m_Nsize, m_bc2, flag);
    }
  }
 
  if(do_comm_any > 0 && ith == 0){
    chset_send.wait();
    chset_recv.wait();
  }
#pragma omp barrier
 
  if(do_comm_any > 0 && ith == ith_kernel){
 
    real_t *buf2xp = (real_t*)chrecv_up[0].ptr();
    real_t *buf2xm = (real_t*)chrecv_dn[0].ptr();
    real_t *buf2yp = (real_t*)chrecv_up[1].ptr();
    real_t *buf2ym = (real_t*)chrecv_dn[1].ptr();
    real_t *buf2zp = (real_t*)chrecv_up[2].ptr();
    real_t *buf2zm = (real_t*)chrecv_dn[2].ptr();
    real_t *buf2tp = (real_t*)chrecv_up[3].ptr();
    real_t *buf2tm = (real_t*)chrecv_dn[3].ptr();
 
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_2_dirac(v2, u, 
                          buf2xp, buf2xm, buf2yp, buf2ym, 
                          buf2zp, buf2zm, buf2tp, buf2tm,
                          m_CKs, m_Nsize, m_bc, do_comm, flag);
    }else{
      BridgeACC::mult_wilson_2_chiral(v2, u,
                          buf2xp, buf2xm, buf2yp, buf2ym,
                          buf2zp, buf2zm, buf2tp, buf2tm,
                          m_CKs, m_Nsize, m_bc, do_comm, flag);
    }
 
  }
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_D_alt(AFIELD &v, const AFIELD &w)
{
#pragma omp barrier
 
  real_t *vp = v.ptr(0);
  real_t *wp = const_cast<AFIELD*>(&w)->ptr(0);
 
  v.set(real_t(0.0));
  mult_xp(vp, wp);
  mult_xm(vp, wp);
  mult_yp(vp, wp);
  mult_ym(vp, wp);
  mult_zp(vp, wp);
  mult_zm(vp, wp);
  mult_tp(vp, wp);
  mult_tm(vp, wp);
  aypx(-m_CKs, v, w);
 
#pragma omp barrier
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_up(int mu, AFIELD &v, const AFIELD &w)
{
  real_t *vp = v.ptr(0);
  real_t *wp = const_cast<AFIELD*>(&w)->ptr(0);
 
  if(mu == 0){
    mult_xp(vp, wp);
  }else if(mu == 1){
    mult_yp(vp, wp);
  }else if(mu == 2){
    mult_zp(vp, wp);
  }else if(mu == 3){
    mult_tp(vp, wp);
  }else{
    vout.crucial(m_vl, "%s: mult_up for %d direction is undefined.",
                 class_name.c_str(), mu);
    exit(EXIT_FAILURE);
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_dn(int mu, AFIELD &v, const AFIELD &w)
{
  real_t *vp = v.ptr(0);
  real_t *wp = const_cast<AFIELD*>(&w)->ptr(0);
 
  if(mu == 0){
    mult_xm(vp, wp);
  }else if(mu == 1){
    mult_ym(vp, wp);
  }else if(mu == 2){
    mult_zm(vp, wp);
  }else if(mu == 3){
    mult_tm(vp, wp);
  }else{
    vout.crucial(m_vl, "%s: mult_dn for %d direction is undefined.",
                 class_name.c_str(), mu);
    exit(EXIT_FAILURE);
  }
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_xp(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 0;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_dn[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_up[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::mult_wilson_xp1(buf1, v1, m_Nsize, m_bc, NC);
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_up[idir].start();
    chsend_dn[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    BridgeACC::mult_wilson_xpb(v2, u, v1, m_Nsize, m_bc2, NC);
  }
 
  if(do_comm[idir] > 0 && ith == 0){
    chsend_dn[idir].wait();
    chrecv_up[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
      BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
      BridgeACC::mult_wilson_xp2(v2, u, buf2, m_Nsize, m_bc, NC);
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_xm(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 0;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_up[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_dn[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::mult_wilson_xm1(buf1, u, v1, m_Nsize, m_bc, NC);
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_dn[idir].start();
    chsend_up[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    BridgeACC::mult_wilson_xmb(v2, u, v1, m_Nsize, m_bc2, NC);
  }
 
  if(do_comm[idir] > 0 && ith == 0){
    chsend_up[idir].wait();
    chrecv_dn[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    BridgeACC::mult_wilson_xm2(v2, buf2, m_Nsize, m_bc, NC);
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_yp(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 1;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_dn[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_up[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::mult_wilson_yp1(buf1, v1, m_Nsize, m_bc, NC);
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_up[idir].start();
    chsend_dn[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    BridgeACC::mult_wilson_ypb(v2, u, v1, m_Nsize, m_bc2, NC);
  }
 
 if(do_comm[idir] > 0 && ith == 0){
    chsend_dn[idir].wait();
    chrecv_up[idir].wait();
  }
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    BridgeACC::mult_wilson_yp2(v2, u, buf2, m_Nsize, m_bc, NC);
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_ym(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 1;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_up[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_dn[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::mult_wilson_ym1(buf1, u, v1, m_Nsize, m_bc, NC);
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_dn[idir].start();
    chsend_up[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    BridgeACC::mult_wilson_ymb(v2, u, v1, m_Nsize, m_bc2, NC);
  }
  
  if(do_comm[idir] > 0 && ith == 0){
    chsend_up[idir].wait();
    chrecv_dn[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    BridgeACC::mult_wilson_ym2(v2, buf2, m_Nsize, m_bc, NC);
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_zp(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 2;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_dn[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_up[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::mult_wilson_zp1(buf1, v1, m_Nsize, m_bc, NC);
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_up[idir].start();
    chsend_dn[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    BridgeACC::mult_wilson_zpb(v2, u, v1, m_Nsize, m_bc2, NC);
  }
 
  if(do_comm[idir] > 0 && ith == 0){
    chsend_dn[idir].wait();
    chrecv_up[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    BridgeACC::mult_wilson_zp2(v2, u, buf2, m_Nsize, m_bc, NC);
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_zm(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 2;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_up[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_dn[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::mult_wilson_zm1(buf1, u, v1, m_Nsize, m_bc, NC);
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_dn[idir].start();
    chsend_up[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    BridgeACC::mult_wilson_zmb(v2, u, v1, m_Nsize, m_bc2, NC);
  }
 
  if(do_comm[idir] > 0 && ith == 0){
    chsend_up[idir].wait();
    chrecv_dn[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    BridgeACC::mult_wilson_zm2(v2, buf2, m_Nsize, m_bc, NC);
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_tp(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 3;
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_dn[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_up[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_tp1_dirac(buf1, v1, m_Nsize, m_bc, NC);
    }else{
      BridgeACC::mult_wilson_tp1_chiral(buf1, v1, m_Nsize, m_bc, NC);
    }
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_up[idir].start();
    chsend_dn[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_tpb_dirac(v2, u, v1, m_Nsize, m_bc2, NC);
    }else{
      BridgeACC::mult_wilson_tpb_chiral(v2, u, v1, m_Nsize, m_bc2, NC);
    }
  }
 
  if(do_comm[idir] > 0 && ith == 0){
    chsend_dn[idir].wait();
    chrecv_up[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_tp2_dirac(v2, u, buf2, m_Nsize, m_bc, NC);
    }else{
      BridgeACC::mult_wilson_tp2_chiral(v2, u, buf2, m_Nsize, m_bc, NC);
    }
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
void AFopr_Wilson<AFIELD>::mult_tm(real_t *v2, real_t *v1)
{
#pragma omp barrier
 
  int idir = 3;
 
  int ith, nth, ith_kernel;
  set_kernel_thread(ith, nth, ith_kernel);
 
  AIndex_lex<real_t,AFIELD::IMPL> index_lex;
 
  real_t *buf1 = (real_t*)chsend_up[idir].ptr();
  real_t *buf2 = (real_t*)chrecv_dn[idir].ptr();
  real_t *u = m_U.ptr(index_lex.idx_G(0, 0, idir));
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_tm1_dirac(buf1, u, v1, m_Nsize, m_bc, NC);
    }else{
      BridgeACC::mult_wilson_tm1_chiral(buf1, u, v1, m_Nsize, m_bc, NC);
    }
    BridgeACC::copy_from_device(buf1, m_Nbdsize[idir]);
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == 0){
    chrecv_dn[idir].start();
    chsend_up[idir].start();
  }
 
  // bulk part
  if(ith == ith_kernel){
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_tmb_dirac(v2, u, v1, m_Nsize, m_bc2, NC);
    }else{
      BridgeACC::mult_wilson_tmb_chiral(v2, u, v1, m_Nsize, m_bc2, NC);
    }
  }
 
  if(do_comm[idir] > 0 && ith == 0){
    chsend_up[idir].wait();
    chrecv_dn[idir].wait();
  }
 
#pragma omp barrier
 
  if(do_comm[idir] > 0 && ith == ith_kernel){
    BridgeACC::copy_to_device(buf2, m_Nbdsize[idir]);
    if(m_repr == DIRAC){
      BridgeACC::mult_wilson_tm2_dirac(v2, buf2, m_Nsize, m_bc, NC);
    }else{
      BridgeACC::mult_wilson_tm2_chiral(v2, buf2, m_Nsize, m_bc, NC);
    }
  }
 
#pragma omp barrier
 
}
 
//====================================================================
template<typename AFIELD>
double AFopr_Wilson<AFIELD>::flop_count()
{
  return flop_count(m_mode);
}
 
//====================================================================
template<typename AFIELD>
double AFopr_Wilson<AFIELD>::flop_count(const std::string mode)
{
  // The following flop counting adopts the original nmuber of the
  // arithmetic operations, while the kernel code may employ the SU(3)
  // third row reconstruction. This aims at comparison to the codes
  // in other branches.  [14 Feb 2025 H.M.]
 
  int    Lvol = CommonParameters::Lvol();
  double flop_site, flop;
 
  if (m_repr == DIRAC) {
    flop_site = static_cast<double>(
      m_Nc * m_Nd * (4 + 6 * (4 * m_Nc + 2) + 2 * (4 * m_Nc + 1)));
  } else if (m_repr == CHIRAL) {
    flop_site = static_cast<double>(
      m_Nc * m_Nd * (4 + 8 * (4 * m_Nc + 2)));
  } else {
    vout.crucial(m_vl, "%s: gamma matrix repr is undefined.\n",
                 class_name.c_str());
    exit(EXIT_FAILURE);
  }
 
  flop = flop_site * static_cast<double>(Lvol);
  if ((mode == "DdagD") || (mode == "DDdag")) flop *= 2.0;
 
  return flop;
}
 
//============================================================END=====