void DiracCloverPC::reconstruct(cudaColorSpinorField &x, const cudaColorSpinorField &b,
const QudaSolutionType solType) const
{
if (solType == QUDA_MATPC_SOLUTION || solType == QUDA_MATPCDAG_MATPC_SOLUTION) {
return;
}
checkFullSpinor(x, b);
bool reset = newTmp(&tmp1, b.Even());
// create full solution
if (matpcType == QUDA_MATPC_EVEN_EVEN ||
matpcType == QUDA_MATPC_EVEN_EVEN_ASYMMETRIC) {
// x_o = A_oo^-1 (b_o + k D_oe x_e)
DiracWilson::DslashXpay(*tmp1, x.Even(), QUDA_ODD_PARITY, b.Odd(), kappa);
CloverInv(x.Odd(), *tmp1, QUDA_ODD_PARITY);
} else if (matpcType == QUDA_MATPC_ODD_ODD ||
matpcType == QUDA_MATPC_ODD_ODD_ASYMMETRIC) {
// x_e = A_ee^-1 (b_e + k D_eo x_o)
DiracWilson::DslashXpay(*tmp1, x.Odd(), QUDA_EVEN_PARITY, b.Even(), kappa);
CloverInv(x.Even(), *tmp1, QUDA_EVEN_PARITY);
} else {
errorQuda("MatPCType %d not valid for DiracCloverPC", matpcType);
}
deleteTmp(&tmp1, reset);
}
// Apply the even-odd preconditioned clover-improved Dirac operator
void DiracCloverPC::M(cudaColorSpinorField &out, const cudaColorSpinorField &in) const
{
double kappa2 = -kappa*kappa;
// FIXME: For asymmetric, a "DslashCxpay" kernel would improve performance.
bool reset = newTmp(&tmp1, in);
if (matpcType == QUDA_MATPC_EVEN_EVEN_ASYMMETRIC) {
bool reset = newTmp(&tmp2, in);
// DiracCloverPC::Dslash applies A^{-1}Dslash
Dslash(*tmp1, in, QUDA_ODD_PARITY);
Clover(*tmp2, in, QUDA_EVEN_PARITY);
// DiracWilson::Dslash applies only Dslash
DiracWilson::DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, *tmp2, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD_ASYMMETRIC) {
// FIXME: It would be nice if I could do something like: cudaColorSpinorField tmp3( in.param() );
// to save copying the data from 'in'
bool reset = newTmp(&tmp2, in);
// DiracCloverPC::Dslash applies A^{-1}Dslash
Dslash(*tmp1, in, QUDA_EVEN_PARITY);
Clover(*tmp2, in, QUDA_ODD_PARITY);
// DiracWilson::Dslash applies only Dslash
DiracWilson::DslashXpay(out, *tmp1, QUDA_ODD_PARITY, *tmp2, kappa2);
} else if (!dagger) { // symmetric preconditioning
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
Dslash(*tmp1, in, QUDA_ODD_PARITY);
DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
Dslash(*tmp1, in, QUDA_EVEN_PARITY);
DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else {
errorQuda("Invalid matpcType");
}
} else { // symmetric preconditioning, dagger
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
CloverInv(out, in, QUDA_EVEN_PARITY);
Dslash(*tmp1, out, QUDA_ODD_PARITY);
DiracWilson::DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
CloverInv(out, in, QUDA_ODD_PARITY);
Dslash(*tmp1, out, QUDA_EVEN_PARITY);
DiracWilson::DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else {
errorQuda("MatPCType %d not valid for DiracCloverPC", matpcType);
}
}
deleteTmp(&tmp1, reset);
}
// Apply the even-odd preconditioned clover-improved Dirac operator
void DiracCloverPC::M(cudaColorSpinorField &out, const cudaColorSpinorField &in) const
{
double kappa2 = -kappa*kappa;
// FIXME: For asymmetric, a "DslashCxpay" kernel would improve performance.
bool reset = newTmp(&tmp1, in);
if (matpcType == QUDA_MATPC_EVEN_EVEN_ASYMMETRIC) {
Dslash(*tmp1, in, QUDA_ODD_PARITY);
Clover(out, in, QUDA_EVEN_PARITY);
#ifdef MULTI_GPU // not safe to alias because of partial updates
cudaColorSpinorField tmp3(in);
#else // safe since out is not read after writing
cudaColorSpinorField &tmp3 = out;
#endif
DiracWilson::DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, tmp3, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD_ASYMMETRIC) {
Dslash(*tmp1, in, QUDA_EVEN_PARITY);
Clover(out, in, QUDA_ODD_PARITY);
#ifdef MULTI_GPU // not safe to alias because of partial updates
cudaColorSpinorField tmp3(in);
#else // safe since out is not read after writing
cudaColorSpinorField &tmp3 = out;
#endif
DiracWilson::DslashXpay(out, *tmp1, QUDA_ODD_PARITY, tmp3, kappa2);
} else if (!dagger) { // symmetric preconditioning
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
Dslash(*tmp1, in, QUDA_ODD_PARITY);
DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
Dslash(*tmp1, in, QUDA_EVEN_PARITY);
DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else {
errorQuda("Invalid matpcType");
}
} else { // symmetric preconditioning, dagger
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
CloverInv(out, in, QUDA_EVEN_PARITY);
Dslash(*tmp1, out, QUDA_ODD_PARITY);
DiracWilson::DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
CloverInv(out, in, QUDA_ODD_PARITY);
Dslash(*tmp1, out, QUDA_EVEN_PARITY);
DiracWilson::DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else {
errorQuda("MatPCType %d not valid for DiracCloverPC", matpcType);
}
}
deleteTmp(&tmp1, reset);
}
// Apply the even-odd preconditioned clover-improved Dirac operator
void DiracCloverPC::M(cudaColorSpinorField &out, const cudaColorSpinorField &in) const
{
double kappa2 = -kappa*kappa;
bool reset1 = newTmp(&tmp1, in);
if (matpcType == QUDA_MATPC_EVEN_EVEN_ASYMMETRIC) {
// DiracCloverPC::Dslash applies A^{-1}Dslash
Dslash(*tmp1, in, QUDA_ODD_PARITY);
// DiracClover::DslashXpay applies (A - kappa^2 D)
DiracClover::DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD_ASYMMETRIC) {
// DiracCloverPC::Dslash applies A^{-1}Dslash
Dslash(*tmp1, in, QUDA_EVEN_PARITY);
// DiracClover::DslashXpay applies (A - kappa^2 D)
DiracClover::DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else if (!dagger) { // symmetric preconditioning
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
Dslash(*tmp1, in, QUDA_ODD_PARITY);
DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
Dslash(*tmp1, in, QUDA_EVEN_PARITY);
DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else {
errorQuda("Invalid matpcType");
}
} else { // symmetric preconditioning, dagger
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
CloverInv(out, in, QUDA_EVEN_PARITY);
Dslash(*tmp1, out, QUDA_ODD_PARITY);
DiracWilson::DslashXpay(out, *tmp1, QUDA_EVEN_PARITY, in, kappa2);
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
CloverInv(out, in, QUDA_ODD_PARITY);
Dslash(*tmp1, out, QUDA_EVEN_PARITY);
DiracWilson::DslashXpay(out, *tmp1, QUDA_ODD_PARITY, in, kappa2);
} else {
errorQuda("MatPCType %d not valid for DiracCloverPC", matpcType);
}
}
deleteTmp(&tmp1, reset1);
}
void DiracCloverPC::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;
return;
}
bool reset = newTmp(&tmp1, b.Even());
// we desire solution to full system
if (matpcType == QUDA_MATPC_EVEN_EVEN) {
// src = A_ee^-1 (b_e + k D_eo A_oo^-1 b_o)
src = &(x.Odd());
CloverInv(*src, b.Odd(), QUDA_ODD_PARITY);
DiracWilson::DslashXpay(*tmp1, *src, QUDA_EVEN_PARITY, b.Even(), kappa);
CloverInv(*src, *tmp1, QUDA_EVEN_PARITY);
sol = &(x.Even());
} else if (matpcType == QUDA_MATPC_ODD_ODD) {
// src = A_oo^-1 (b_o + k D_oe A_ee^-1 b_e)
src = &(x.Even());
CloverInv(*src, b.Even(), QUDA_EVEN_PARITY);
DiracWilson::DslashXpay(*tmp1, *src, QUDA_ODD_PARITY, b.Odd(), kappa);
CloverInv(*src, *tmp1, QUDA_ODD_PARITY);
sol = &(x.Odd());
} else if (matpcType == QUDA_MATPC_EVEN_EVEN_ASYMMETRIC) {
// src = b_e + k D_eo A_oo^-1 b_o
src = &(x.Odd());
CloverInv(*tmp1, b.Odd(), QUDA_ODD_PARITY); // safe even when *tmp1 = b.odd
DiracWilson::DslashXpay(*src, *tmp1, QUDA_EVEN_PARITY, b.Even(), kappa);
sol = &(x.Even());
} else if (matpcType == QUDA_MATPC_ODD_ODD_ASYMMETRIC) {
// src = b_o + k D_oe A_ee^-1 b_e
src = &(x.Even());
CloverInv(*tmp1, b.Even(), QUDA_EVEN_PARITY); // safe even when *tmp1 = b.even
DiracWilson::DslashXpay(*src, *tmp1, QUDA_ODD_PARITY, b.Odd(), kappa);
sol = &(x.Odd());
} else {
errorQuda("MatPCType %d not valid for DiracCloverPC", matpcType);
}
// here we use final solution to store parity solution and parity source
// b is now up for grabs if we want
deleteTmp(&tmp1, reset);
}
