C++ cudaColorSpinorField类代码示例

void DiracStaggered::MdagM(cudaColorSpinorField &out, const cudaColorSpinorField &in) const
{

  if (!initDslash){
    initDslashConstants(*fatGauge, in.Stride());
    initStaggeredConstants(*fatGauge, *longGauge);
  }
  
  bool reset = newTmp(&tmp1, in);
  
  cudaColorSpinorField* mytmp = dynamic_cast<cudaColorSpinorField*>(&(tmp1->Even()));
  cudaColorSpinorField* ineven = dynamic_cast<cudaColorSpinorField*>(&(in.Even()));
  cudaColorSpinorField* inodd = dynamic_cast<cudaColorSpinorField*>(&(in.Odd()));
  cudaColorSpinorField* outeven = dynamic_cast<cudaColorSpinorField*>(&(out.Even()));
  cudaColorSpinorField* outodd = dynamic_cast<cudaColorSpinorField*>(&(out.Odd()));
  
  //even
  Dslash(*mytmp, *ineven, QUDA_ODD_PARITY);  
  DslashXpay(*outeven, *mytmp, QUDA_EVEN_PARITY, *ineven, 4*mass*mass);
  
  //odd
  Dslash(*mytmp, *inodd, QUDA_EVEN_PARITY);  
  DslashXpay(*outodd, *mytmp, QUDA_ODD_PARITY, *inodd, 4*mass*mass);    

  deleteTmp(&tmp1, reset);
}

void FaceBuffer::exchangeFacesStart(cudaColorSpinorField &in, int parity,
				    int dagger, int dir, cudaStream_t *stream_p)
{
  if(!commDimPartitioned(dir)){
    return ;
  }

  in.allocateGhostBuffer();   // allocate the ghost buffer if not yet allocated
  
  stream = stream_p;
  
  int back_nbr[4] = {X_BACK_NBR, Y_BACK_NBR, Z_BACK_NBR,T_BACK_NBR};
  int fwd_nbr[4] = {X_FWD_NBR, Y_FWD_NBR, Z_FWD_NBR,T_FWD_NBR};
  int uptags[4] = {XUP, YUP, ZUP, TUP};
  int downtags[4] = {XDOWN, YDOWN, ZDOWN, TDOWN};
  
  // Prepost all receives
  recv_request1[dir] = comm_recv_with_tag(pageable_back_nbr_spinor[dir], nbytes[dir], back_nbr[dir], uptags[dir]);
  recv_request2[dir] = comm_recv_with_tag(pageable_fwd_nbr_spinor[dir], nbytes[dir], fwd_nbr[dir], downtags[dir]);
  
  // gather for backwards send
  in.packGhost(back_nbr_spinor_sendbuf[dir], dir, QUDA_BACKWARDS, 
	       (QudaParity)parity, dagger, &stream[2*dir + sendBackStrmIdx]); CUERR;  
  
  // gather for forwards send
  in.packGhost(fwd_nbr_spinor_sendbuf[dir], dir, QUDA_FORWARDS, 
	       (QudaParity)parity, dagger, &stream[2*dir + sendFwdStrmIdx]); CUERR;
}

void Dirac::checkFullSpinor(const cudaColorSpinorField &out, const cudaColorSpinorField &in) const
{
   if (in.SiteSubset() != QUDA_FULL_SITE_SUBSET || out.SiteSubset() != QUDA_FULL_SITE_SUBSET) {
    errorQuda("ColorSpinorFields are not full fields: in = %d, out = %d", 
	      in.SiteSubset(), out.SiteSubset());
  } 
}

void DiracWilsonPC::prepare(cudaColorSpinorField* &src, cudaColorSpinorField* &sol,
			    cudaColorSpinorField &x, cudaColorSpinorField &b, 
			    const QudaSolutionType solType) const
{
  // we desire solution to preconditioned system
  if (solType == QUDA_MATPC_SOLUTION || solType == QUDA_MATPCDAG_MATPC_SOLUTION) {
    src = &b;
    sol = &x;
  } else {
    // we desire solution to full system
    if (matpcType == QUDA_MATPC_EVEN_EVEN) {
      // src = b_e + k D_eo b_o
      DslashXpay(x.Odd(), b.Odd(), QUDA_EVEN_PARITY, b.Even(), kappa);
      src = &(x.Odd());
      sol = &(x.Even());
    } else if (matpcType == QUDA_MATPC_ODD_ODD) {
      // src = b_o + k D_oe b_e
      DslashXpay(x.Even(), b.Even(), QUDA_ODD_PARITY, b.Odd(), kappa);
      src = &(x.Even());
      sol = &(x.Odd());
    } else {
      errorQuda("MatPCType %d not valid for DiracWilsonPC", matpcType);
    }
    // here we use final solution to store parity solution and parity source
    // b is now up for grabs if we want
  }

}

void FaceBuffer::exchangeFacesPack(cudaColorSpinorField &in, int parity,
				   int dagger, int dir, cudaStream_t *stream_p)
{
  int dim = dir/2;
  if(!commDimPartitioned(dim)) return;

  in.allocateGhostBuffer();   // allocate the ghost buffer if not yet allocated
  stream = stream_p;

  if (dir%2==0) { // sending backwards
#ifdef QMP_COMMS
    // Prepost receive
    QMP_start(mh_from_fwd[dim]);
#endif
    // gather for backwards send
    in.packGhost(dim, QUDA_BACKWARDS, (QudaParity)parity, dagger, &stream[2*dim+sendBackStrmIdx]);  
  } else { // sending forwards

#ifdef QMP_COMMS
    // Prepost receive
    QMP_start(mh_from_back[dim]);
#endif
    // gather for forwards send
    in.packGhost(dim, QUDA_FORWARDS, (QudaParity)parity, dagger, &stream[2*dim+sendFwdStrmIdx]);
  }
}

void FaceBuffer::pack(cudaColorSpinorField &in, int parity, int dagger, int dim, cudaStream_t *stream_p)
{
  if(!commDimPartitioned(dim)) return;

  in.allocateGhostBuffer();   // allocate the ghost buffer if not yet allocated
  stream = stream_p;

  in.packGhost(dim, (QudaParity)parity, dagger, &stream[Nstream-1]);  
}

  // Full staggered operator
  void DiracStaggered::M(cudaColorSpinorField &out, const cudaColorSpinorField &in) const
  {
    bool reset = newTmp(&tmp1, in.Even());

    DslashXpay(out.Even(), in.Odd(), QUDA_EVEN_PARITY, *tmp1, 2*mass);  
    DslashXpay(out.Odd(), in.Even(), QUDA_ODD_PARITY, *tmp1, 2*mass);
  
    deleteTmp(&tmp1, reset);
  }

  void DiracClover::checkParitySpinor(const cudaColorSpinorField &out, const cudaColorSpinorField &in) const
  {
    Dirac::checkParitySpinor(out, in);

    if (out.Volume() != clover.VolumeCB()) {
      errorQuda("Parity spinor volume %d doesn't match clover checkboard volume %d",
		out.Volume(), clover.VolumeCB());
    }
  }

void FaceBuffer::scatter(cudaColorSpinorField &out, int dagger, int dir) {
  int dim = dir / 2;
  if(!commDimPartitioned(dim)) return;

  if (dir%2 == 0) {
    out.unpackGhost(fwd_nbr_spinor[dim], dim, QUDA_FORWARDS,  dagger, &stream[2*dim + recFwdStrmIdx]); CUERR;
  } else {
    out.unpackGhost(back_nbr_spinor[dim], dim, QUDA_BACKWARDS,  dagger, &stream[2*dim + recBackStrmIdx]); CUERR;
  }
}

void FaceBuffer::gather(cudaColorSpinorField &in, int dagger, int dir)
{
  int dim = dir/2;
  if(!commDimPartitioned(dim)) return;

  if (dir%2==0){ // backwards send
    in.sendGhost(back_nbr_spinor_sendbuf[dim], dim, QUDA_BACKWARDS, dagger, &stream[2*dim + sendBackStrmIdx]);
  } else { // forwards send
    in.sendGhost(fwd_nbr_spinor_sendbuf[dim], dim, QUDA_FORWARDS, dagger, &stream[2*dim + sendFwdStrmIdx]); 
  }
}

void FaceBuffer::scatter(cudaColorSpinorField &out, int dagger, int dir)
{
  int dim = dir/2;
  if(!commDimPartitioned(dim)) return;

  if (dir%2==0) {// receive from forwards
    out.unpackGhost(from_fwd_face[dim], dim, QUDA_FORWARDS, dagger, &stream[2*dim+recFwdStrmIdx]); // 0, 2, 4, 6
  } else { // receive from backwards
    out.unpackGhost(from_back_face[dim], dim, QUDA_BACKWARDS, dagger, &stream[2*dim+recBackStrmIdx]); // 1, 3, 5, 7
  }
}

void FaceBuffer::exchangeFacesStart(cudaColorSpinorField &in, int dagger, int dir)
{
  int dim = dir/2;
  if(!commDimPartitioned(dim)) return;

  if (dir%2==0) {
    // backwards copy to host
    in.sendGhost(my_back_face[dim], dim, QUDA_BACKWARDS, dagger, &stream[2*dim+sendBackStrmIdx]);  
  } else {
    // forwards copy to host
    in.sendGhost(my_fwd_face[dim], dim, QUDA_FORWARDS, dagger, &stream[2*dim+sendFwdStrmIdx]);
  }
}

void DiracStaggered::Dslash(cudaColorSpinorField &out, const cudaColorSpinorField &in, 
			 const QudaParity parity) const
{
  if (!initDslash) {
    initDslashConstants(*fatGauge, in.Stride());
    initStaggeredConstants(*fatGauge, *longGauge);
  }
  checkParitySpinor(in, out);

  setFace(face); // FIXME: temporary hack maintain C linkage for dslashCuda
  staggeredDslashCuda(&out, *fatGauge, *longGauge, &in, parity, dagger, 0, 0, commDim);
  
  flops += 1146*in.Volume();
}

// Public method to apply the clover term only
void DiracClover::Clover(cudaColorSpinorField &out, const cudaColorSpinorField &in, const QudaParity parity) const
{
  if (!initDslash) initDslashConstants(gauge, in.Stride());
  if (!initClover) initCloverConstants(clover.Stride());
  checkParitySpinor(in, out, clover);

  // 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();
}

void FaceBuffer::pack(cudaColorSpinorField &in, int parity,
		      int dagger, int dir, cudaStream_t *stream_p)
{
  int dim = dir/2;
  if(!commDimPartitioned(dim)) return;

  in.allocateGhostBuffer();   // allocate the ghost buffer if not yet allocated  
  stream = stream_p;

  if (dir%2==0){ // backwards send
    in.packGhost(dim, QUDA_BACKWARDS, (QudaParity)parity, dagger, &stream[2*dim + sendBackStrmIdx]);
  } else { // forwards send
    in.packGhost(dim, QUDA_FORWARDS, (QudaParity)parity, dagger, &stream[2*dim + sendFwdStrmIdx]);
  }
}