Exemplo n.º 1
0
  void DiracStaggered::Dslash(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
			      const QudaParity parity) const
  {
    checkParitySpinor(in, out);

    initSpinorConstants(in, profile);
    setFace(face); // FIXME: temporary hack maintain C linkage for dslashCuda
    staggeredDslashCuda(&out, fatGauge, longGauge, &in, parity, dagger, 0, 0, commDim, profile);
  
    flops += 1146ll*in.Volume();
  }
Exemplo n.º 2
0
void DiracWilson::Dslash(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
			 const QudaParity parity) const
{
  initSpinorConstants(in);
  checkParitySpinor(in, out);
  checkSpinorAlias(in, out);

  setFace(face); // FIXME: temporary hack maintain C linkage for dslashCuda

  wilsonDslashCuda(&out, gauge, &in, parity, dagger, 0, 0.0, commDim);

  flops += 1320ll*in.Volume();
}
Exemplo n.º 3
0
  // Public method to apply the clover term only
  void DiracClover::Clover(cudaColorSpinorField &out, const cudaColorSpinorField &in, const QudaParity parity) const
  {
    initSpinorConstants(in);
    checkParitySpinor(in, out);

    // regular clover term
    FullClover cs;
    cs.even = clover.even; cs.odd = clover.odd; cs.evenNorm = clover.evenNorm; cs.oddNorm = clover.oddNorm;
    cs.precision = clover.precision; cs.bytes = clover.bytes, cs.norm_bytes = clover.norm_bytes;
    cloverCuda(&out, gauge, cs, &in, parity);

    flops += 504*in.Volume();
  }
Exemplo n.º 4
0
  void DiracWilson::DslashXpay(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
			       const QudaParity parity, const cudaColorSpinorField &x,
			       const double &k) const
  {
    initSpinorConstants(in, profile);
    checkParitySpinor(in, out);
    checkSpinorAlias(in, out);

    setFace(face1,face2); // FIXME: temporary hack maintain C linkage for dslashCuda

    wilsonDslashCuda(&out, gauge, &in, parity, dagger, &x, k, commDim, profile);

    flops += 1368ll*in.Volume();
  }
Exemplo n.º 5
0
  // Public method
  void DiracCloverPC::CloverInv(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
				const QudaParity parity) const
  {
    initSpinorConstants(in);
    checkParitySpinor(in, out);

    // needs to be cloverinv
    FullClover cs;
    cs.even = clover.evenInv; cs.odd = clover.oddInv; cs.evenNorm = clover.evenInvNorm; cs.oddNorm = clover.oddInvNorm;
    cs.precision = clover.precision; cs.bytes = clover.bytes, cs.norm_bytes = clover.norm_bytes;
    cloverCuda(&out, gauge, cs, &in, parity);

    flops += 504*in.Volume();
  }
Exemplo n.º 6
0
  // apply hopping term, then clover: (A_ee^-1 D_eo) or (A_oo^-1 D_oe),
  // and likewise for dagger: (A_ee^-1 D^dagger_eo) or (A_oo^-1 D^dagger_oe)
  // NOTE - this isn't Dslash dagger since order should be reversed!
  void DiracCloverPC::Dslash(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
			     const QudaParity parity) const
  {
    initSpinorConstants(in);
    checkParitySpinor(in, out);
    checkSpinorAlias(in, out);

    setFace(face); // FIXME: temporary hack maintain C linkage for dslashCuda

    FullClover cs;
    cs.even = clover.evenInv; cs.odd = clover.oddInv; cs.evenNorm = clover.evenInvNorm; cs.oddNorm = clover.oddInvNorm;
    cs.precision = clover.precision; cs.bytes = clover.bytes, cs.norm_bytes = clover.norm_bytes;
    cloverDslashCuda(&out, gauge, cs, &in, parity, dagger, 0, 0.0, commDim);

    flops += (1320+504)*in.Volume();
  }
Exemplo n.º 7
0
//!NEW : added setFace(),   domainWallDslashCuda() got an extra argument  
void DiracDomainWall::Dslash(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
			     const QudaParity parity) const
{
  if ( in.Ndim() != 5 || out.Ndim() != 5) errorQuda("Wrong number of dimensions\n");
  checkParitySpinor(in, out);
  checkSpinorAlias(in, out);
 
  initSpinorConstants(in);
  setFace(face); // FIXME: temporary hack maintain C linkage for dslashCuda  
  domainWallDslashCuda(&out, gauge, &in, parity, dagger, 0, mass, 0, commDim);   

  long long Ls = in.X(4);
  long long bulk = (Ls-2)*(in.Volume()/Ls);
  long long wall = 2*in.Volume()/Ls;
  flops += 1320LL*(long long)in.Volume() + 96LL*bulk + 120LL*wall;
}