portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 07:17:06 +01:00
Code Releases Activity

Imported changes from feature/gparity_HMC branch:

Browse Source 
Added a bounds-check function for the RHMC with arbitrary power
	Added a pseudofermion action for the rational ratio with an arbitrary power and a mixed-precision variant of the same. The existing one-flavor rational ratio class now uses the general class under the hood
	To support testing of the two-flavor even-odd ratio pseudofermion, separated the functionality of generating the random field and performing the heatbath step, and added a method to obtain the pseudofermion field
	Added a new HMC runner start type: CheckpointStartReseed, which reseeds the RNG from scratch, allowing for the creation of new evolution streams from an existing checkpoint. Added log output of seeds used when the RNG is seeded.
	EOFA changes:
		To support mixed-precision inversion, generalized the class to maintain a separate solver for the L and R operators in the heatbath (separate solvers are already implemented for the other stages)
		To support mixed-precision, the action of setting the operator shift coefficients is now maintained in a virtual function. A derived class for mixed-precision solvers ensures the coefficients are applied to both the double and single-prec operators
		The ||^2 of the random source is now stored by the heatbath and compared to the initial action when it is computed. These should be equal but may differ if the rational bounds are not chosen correctly, hence serving as a useful and free test
		Fixed calculation of M_eofa (previously incomplete and #if'd out)
		Added functionality to compute M_eofa^-1 to complement the calculation of M_eofa (both are equally expensive!)
		To support testing, separated the functionality of generating the random field and performing the heatbath step, and added a method to obtain the pseudofermion field
	Added a test program which computes the G-parity force using the 1 and 2 flavor implementations and compares the result. Test supports DWF, EOFA and DSDR actions, chosen by a command line option.
	The Mobius EOFA force test now also checks the rational approximation used for the heatbath
	Added a test program for the mixed precision EOFA compared to the double-prec implementation,
	G-parity HMC test now applied GPBC in the y direction and not the t direction (GPBC in t are no longer supported) and checkpoints after every configuration
	Added a test program which computes the two-flavor G-parity action (via RHMC) with both the 1 and 2 flavor implementations and checks they agree
	Added a test program to check the implementation of M_eofa^{-1}
This commit is contained in:
Christopher Kelly
2022-06-22 10:27:48 -04:00
parent 1ad54d049d
commit fd933420c6
17 changed files with 1920 additions and 305 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
Grid/qcd/action/ActionParams.h
View File

@ -37,6 +37,7 @@ NAMESPACE_BEGIN(Grid);
// These can move into a params header and be given MacroMagic serialisation
struct GparityWilsonImplParams {
Coordinate twists;
//mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
GparityWilsonImplParams() : twists(Nd, 0) {};
};
@ -66,7 +67,8 @@ struct StaggeredImplParams {
RealD, mdtolerance,
int, degree,
int, precision,
int, BoundsCheckFreq);
int, BoundsCheckFreq,
RealD, BoundsCheckTol);
// MaxIter and tolerance, vectors??
@ -78,7 +80,8 @@ struct StaggeredImplParams {
int _degree = 10,
int _precision = 64,
int _BoundsCheckFreq=20,
RealD mdtol = 1.0e-6)
RealD mdtol = 1.0e-6,
double _BoundsCheckTol=1e-6)
: lo(_lo),
hi(_hi),
MaxIter(_maxit),
@ -86,9 +89,52 @@ struct StaggeredImplParams {
mdtolerance(mdtol),
degree(_degree),
precision(_precision),
BoundsCheckFreq(_BoundsCheckFreq){};
BoundsCheckFreq(_BoundsCheckFreq),
BoundsCheckTol(_BoundsCheckTol){};
};
/*Action parameters for the generalized rational action
The approximation is for (M^dag M)^{1/inv_pow}
where inv_pow is the denominator of the fractional power.
Default inv_pow=2 for square root, making this equivalent to
the OneFlavourRational action
*/
struct RationalActionParams : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams,
int, inv_pow,
RealD, lo, //low eigenvalue bound of rational approx
RealD, hi, //high eigenvalue bound of rational approx
int, MaxIter, //maximum iterations in msCG
RealD, action_tolerance, //msCG tolerance in action evaluation
int, action_degree, //rational approx tolerance in action evaluation
RealD, md_tolerance, //msCG tolerance in MD integration
int, md_degree, //rational approx tolerance in MD integration
int, precision, //precision of floating point arithmetic
int, BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
// constructor
RationalActionParams(int _inv_pow = 2,
RealD _lo = 0.0,
RealD _hi = 1.0,
int _maxit = 1000,
RealD _action_tolerance = 1.0e-8,
int _action_degree = 10,
RealD _md_tolerance = 1.0e-8,
int _md_degree = 10,
int _precision = 64,
int _BoundsCheckFreq=20)
: inv_pow(_inv_pow),
lo(_lo),
hi(_hi),
MaxIter(_maxit),
action_tolerance(_action_tolerance),
action_degree(_action_degree),
md_tolerance(_md_tolerance),
md_degree(_md_degree),
precision(_precision),
BoundsCheckFreq(_BoundsCheckFreq){};
};
NAMESPACE_END(Grid);
#endif

60
Grid/qcd/action/pseudofermion/Bounds.h
View File

@ -65,13 +65,65 @@ NAMESPACE_BEGIN(Grid);
X=X-Y;
RealD Nd = norm2(X);
std::cout << "************************* "<<std::endl;
std::cout << " noise = "<<Nx<<std::endl;
std::cout << " (MdagM^-1/2)^2 noise = "<<Nz<<std::endl;
std::cout << " MdagM (MdagM^-1/2)^2 noise = "<<Ny<<std::endl;
std::cout << " noise - MdagM (MdagM^-1/2)^2 noise = "<<Nd<<std::endl;
std::cout << " | noise |^2 = "<<Nx<<std::endl;
std::cout << " | (MdagM^-1/2)^2 noise |^2 = "<<Nz<<std::endl;
std::cout << " | MdagM (MdagM^-1/2)^2 noise |^2 = "<<Ny<<std::endl;
std::cout << " | noise - MdagM (MdagM^-1/2)^2 noise |^2 = "<<Nd<<std::endl;
std::cout << " | noise - MdagM (MdagM^-1/2)^2 noise|/|noise| = " << std::sqrt(Nd/Nx) << std::endl;
std::cout << "************************* "<<std::endl;
assert( (std::sqrt(Nd/Nx)<tol) && " InverseSqrtBoundsCheck ");
}
/* For a HermOp = M^dag M, check the approximation of HermOp^{-1/inv_pow}
by computing |X - HermOp * [ Hermop^{-1/inv_pow} ]^{inv_pow} X| < tol
for noise X (aka GaussNoise).
ApproxNegPow should be the rational approximation for X^{-1/inv_pow}
*/
template<class Field> void InversePowerBoundsCheck(int inv_pow,
int MaxIter,double tol,
LinearOperatorBase<Field> &HermOp,
Field &GaussNoise,
MultiShiftFunction &ApproxNegPow)
{
GridBase *FermionGrid = GaussNoise.Grid();
Field X(FermionGrid);
Field Y(FermionGrid);
Field Z(FermionGrid);
Field tmp1(FermionGrid), tmp2(FermionGrid);
X=GaussNoise;
RealD Nx = norm2(X);
ConjugateGradientMultiShift<Field> msCG(MaxIter,ApproxNegPow);
tmp1 = X;
Field* in = &tmp1;
Field* out = &tmp2;
for(int i=0;i<inv_pow;i++){ //apply [ Hermop^{-1/inv_pow} ]^{inv_pow} X = HermOp^{-1} X
msCG(HermOp, *in, *out); //backwards conventions!
if(i!=inv_pow-1) std::swap(in, out);
}
Z = *out;
RealD Nz = norm2(Z);
HermOp.HermOp(Z,Y);
RealD Ny = norm2(Y);
X=X-Y;
RealD Nd = norm2(X);
std::cout << "************************* "<<std::endl;
std::cout << " | noise |^2 = "<<Nx<<std::endl;
std::cout << " | (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2 = "<<Nz<<std::endl;
std::cout << " | MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2 = "<<Ny<<std::endl;
std::cout << " | noise - MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2 = "<<Nd<<std::endl;
std::cout << " | noise - MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |/| noise | = "<<std::sqrt(Nd/Nx)<<std::endl;
std::cout << "************************* "<<std::endl;
assert( (std::sqrt(Nd/Nx)<tol) && " InversePowerBoundsCheck ");
}
NAMESPACE_END(Grid);

229
Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h
View File

@ -44,6 +44,10 @@ NAMESPACE_BEGIN(Grid);
// Exact one flavour implementation of DWF determinant ratio //
///////////////////////////////////////////////////////////////
//Note: using mixed prec CG for the heatbath solver in this action class will not work
// because the L, R operators must have their shift coefficients updated throughout the heatbath step
// You will find that the heatbath solver simply won't converge.
// To use mixed precision here use the ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction variant below
template<class Impl>
class ExactOneFlavourRatioPseudoFermionAction : public Action<typename Impl::GaugeField>
{
@ -57,37 +61,60 @@ NAMESPACE_BEGIN(Grid);
bool use_heatbath_forecasting;
AbstractEOFAFermion<Impl>& Lop; // the basic LH operator
AbstractEOFAFermion<Impl>& Rop; // the basic RH operator
SchurRedBlackDiagMooeeSolve<FermionField> SolverHB;
SchurRedBlackDiagMooeeSolve<FermionField> SolverHBL;
SchurRedBlackDiagMooeeSolve<FermionField> SolverHBR;
SchurRedBlackDiagMooeeSolve<FermionField> SolverL;
SchurRedBlackDiagMooeeSolve<FermionField> SolverR;
SchurRedBlackDiagMooeeSolve<FermionField> DerivativeSolverL;
SchurRedBlackDiagMooeeSolve<FermionField> DerivativeSolverR;
FermionField Phi; // the pseudofermion field for this trajectory
RealD norm2_eta; //|eta|^2 where eta is the random gaussian field used to generate the pseudofermion field
bool initial_action; //true for the first call to S after refresh, for which the identity S = |eta|^2 holds provided the rational approx is good
public:
//Used in the heatbath, refresh the shift coefficients of the L (LorR=0) or R (LorR=1) operator
virtual void heatbathRefreshShiftCoefficients(int LorR, RealD to){
AbstractEOFAFermion<Impl>&op = LorR == 0 ? Lop : Rop;
op.RefreshShiftCoefficients(to);
}
//Use the same solver for L,R in all cases
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
AbstractEOFAFermion<Impl>& _Rop,
OperatorFunction<FermionField>& CG,
Params& p,
bool use_fc=false)
: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,CG,CG,CG,CG,CG,p,use_fc) {};
: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,CG,CG,CG,CG,CG,CG,p,use_fc) {};
//Use the same solver for L,R in the heatbath but different solvers elsewhere
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
AbstractEOFAFermion<Impl>& _Rop,
OperatorFunction<FermionField>& HeatbathCG,
OperatorFunction<FermionField>& HeatbathCG,
OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR,
OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR,
Params& p,
bool use_fc=false)
: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,HeatbathCG,HeatbathCG, ActionCGL, ActionCGR, DerivCGL,DerivCGR,p,use_fc) {};
//Use different solvers for L,R in all cases
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
AbstractEOFAFermion<Impl>& _Rop,
OperatorFunction<FermionField>& HeatbathCGL, OperatorFunction<FermionField>& HeatbathCGR,
OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR,
OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR,
Params& p,
bool use_fc=false) :
Lop(_Lop),
Rop(_Rop),
SolverHB(HeatbathCG,false,true),
SolverHBL(HeatbathCGL,false,true), SolverHBR(HeatbathCGR,false,true),
SolverL(ActionCGL, false, true), SolverR(ActionCGR, false, true),
DerivativeSolverL(DerivCGL, false, true), DerivativeSolverR(DerivCGR, false, true),
Phi(_Lop.FermionGrid()),
param(p),
use_heatbath_forecasting(use_fc)
use_heatbath_forecasting(use_fc),
initial_action(false)
{
AlgRemez remez(param.lo, param.hi, param.precision);
@ -97,6 +124,8 @@ NAMESPACE_BEGIN(Grid);
PowerNegHalf.Init(remez, param.tolerance, true);
};
const FermionField &getPhi() const{ return Phi; }
virtual std::string action_name() { return "ExactOneFlavourRatioPseudoFermionAction"; }
virtual std::string LogParameters() {
@ -117,6 +146,19 @@ NAMESPACE_BEGIN(Grid);
else{ for(int s=0; s<Ls; ++s){ axpby_ssp_pminus(out, 0.0, in, 1.0, in, s, s); } }
}
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
// P(eta_o) = e^{- eta_o^dag eta_o}
//
// e^{x^2/2 sig^2} => sig^2 = 0.5.
//
RealD scale = std::sqrt(0.5);
FermionField eta (Lop.FermionGrid());
gaussian(pRNG,eta); eta = eta * scale;
refresh(U,eta);
}
// EOFA heatbath: see Eqn. (29) of arXiv:1706.05843
// We generate a Gaussian noise vector \eta, and then compute
// \Phi = M_{\rm EOFA}^{-1/2} * \eta
@ -124,12 +166,10 @@ NAMESPACE_BEGIN(Grid);
//
// As a check of rational require \Phi^dag M_{EOFA} \Phi == eta^dag M^-1/2^dag M M^-1/2 eta = eta^dag eta
//
virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
{
void refresh(const GaugeField &U, const FermionField &eta) {
Lop.ImportGauge(U);
Rop.ImportGauge(U);
FermionField eta (Lop.FermionGrid());
FermionField CG_src (Lop.FermionGrid());
FermionField CG_soln (Lop.FermionGrid());
FermionField Forecast_src(Lop.FermionGrid());
@ -140,11 +180,6 @@ NAMESPACE_BEGIN(Grid);
if(use_heatbath_forecasting){ prev_solns.reserve(param.degree); }
ChronoForecast<AbstractEOFAFermion<Impl>, FermionField> Forecast;
// Seed with Gaussian noise vector (var = 0.5)
RealD scale = std::sqrt(0.5);
gaussian(pRNG,eta);
eta = eta * scale;
// \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta
RealD N(PowerNegHalf.norm);
for(int k=0; k<param.degree; ++k){ N += PowerNegHalf.residues[k] / ( 1.0 + PowerNegHalf.poles[k] ); }
@ -160,15 +195,15 @@ NAMESPACE_BEGIN(Grid);
tmp[1] = Zero();
for(int k=0; k<param.degree; ++k){
gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
Lop.RefreshShiftCoefficients(-gamma_l);
heatbathRefreshShiftCoefficients(0, -gamma_l);
if(use_heatbath_forecasting){ // Forecast CG guess using solutions from previous poles
Lop.Mdag(CG_src, Forecast_src);
CG_soln = Forecast(Lop, Forecast_src, prev_solns);
SolverHB(Lop, CG_src, CG_soln);
SolverHBL(Lop, CG_src, CG_soln);
prev_solns.push_back(CG_soln);
} else {
CG_soln = Zero(); // Just use zero as the initial guess
SolverHB(Lop, CG_src, CG_soln);
SolverHBL(Lop, CG_src, CG_soln);
}
Lop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
tmp[1] = tmp[1] + ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Lop.k ) * tmp[0];
@ -187,15 +222,15 @@ NAMESPACE_BEGIN(Grid);
if(use_heatbath_forecasting){ prev_solns.clear(); } // empirically, LH solns don't help for RH solves
for(int k=0; k<param.degree; ++k){
gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
Rop.RefreshShiftCoefficients(-gamma_l*PowerNegHalf.poles[k]);
heatbathRefreshShiftCoefficients(1, -gamma_l*PowerNegHalf.poles[k]);
if(use_heatbath_forecasting){
Rop.Mdag(CG_src, Forecast_src);
CG_soln = Forecast(Rop, Forecast_src, prev_solns);
SolverHB(Rop, CG_src, CG_soln);
SolverHBR(Rop, CG_src, CG_soln);
prev_solns.push_back(CG_soln);
} else {
CG_soln = Zero();
SolverHB(Rop, CG_src, CG_soln);
SolverHBR(Rop, CG_src, CG_soln);
}
Rop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
tmp[1] = tmp[1] - ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Rop.k ) * tmp[0];
@ -205,49 +240,117 @@ NAMESPACE_BEGIN(Grid);
Phi = Phi + tmp[1];
// Reset shift coefficients for energy and force evals
Lop.RefreshShiftCoefficients(0.0);
Rop.RefreshShiftCoefficients(-1.0);
heatbathRefreshShiftCoefficients(0, 0.0);
heatbathRefreshShiftCoefficients(1, -1.0);
//Mark that the next call to S is the first after refresh
initial_action = true;
// Bounds check
RealD EtaDagEta = norm2(eta);
norm2_eta = EtaDagEta;
// RealD PhiDagMPhi= norm2(eta);
};
void Meofa(const GaugeField& U,const FermionField &phi, FermionField & Mphi)
void Meofa(const GaugeField& U,const FermionField &in, FermionField & out)
{
#if 0
Lop.ImportGauge(U);
Rop.ImportGauge(U);
FermionField spProj_Phi(Lop.FermionGrid());
FermionField mPhi(Lop.FermionGrid());
FermionField spProj_in(Lop.FermionGrid());
std::vector<FermionField> tmp(2, Lop.FermionGrid());
mPhi = phi;
out = in;
// LH term: S = S - k <\Phi| P_{-} \Omega_{-}^{\dagger} H(mf)^{-1} \Omega_{-} P_{-} |\Phi>
spProj(Phi, spProj_Phi, -1, Lop.Ls);
Lop.Omega(spProj_Phi, tmp[0], -1, 0);
spProj(in, spProj_in, -1, Lop.Ls);
Lop.Omega(spProj_in, tmp[0], -1, 0);
G5R5(tmp[1], tmp[0]);
tmp[0] = Zero();
SolverL(Lop, tmp[1], tmp[0]);
Lop.Dtilde(tmp[0], tmp[1]); // We actually solved Cayley preconditioned system: transform back
Lop.Omega(tmp[1], tmp[0], -1, 1);
mPhi = mPhi - Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
spProj(tmp[0], tmp[1], -1, Lop.Ls);
out = out - Lop.k * tmp[1];
// RH term: S = S + k <\Phi| P_{+} \Omega_{+}^{\dagger} ( H(mb)
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{-} P_{-} |\Phi>
spProj(Phi, spProj_Phi, 1, Rop.Ls);
Rop.Omega(spProj_Phi, tmp[0], 1, 0);
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} |\Phi>
spProj(in, spProj_in, 1, Rop.Ls);
Rop.Omega(spProj_in, tmp[0], 1, 0);
G5R5(tmp[1], tmp[0]);
tmp[0] = Zero();
SolverR(Rop, tmp[1], tmp[0]);
Rop.Dtilde(tmp[0], tmp[1]);
Rop.Omega(tmp[1], tmp[0], 1, 1);
action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real();
#endif
spProj(tmp[0], tmp[1], 1, Rop.Ls);
out = out + Rop.k * tmp[1];
}
//Due to the structure of EOFA, it is no more expensive to compute the inverse of Meofa
//To ensure correctness we can simply reuse the heatbath code but use the rational approx
//f(x) = 1/x which corresponds to alpha_0=0, alpha_1=1, beta_1=0 => gamma_1=1
void MeofaInv(const GaugeField &U, const FermionField &in, FermionField &out) {
Lop.ImportGauge(U);
Rop.ImportGauge(U);
FermionField CG_src (Lop.FermionGrid());
FermionField CG_soln (Lop.FermionGrid());
std::vector<FermionField> tmp(2, Lop.FermionGrid());
// \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta
// = 1 * \eta
out = in;
// LH terms:
// \Phi = \Phi + k \sum_{k=1}^{N_{p}} P_{-} \Omega_{-}^{\dagger} ( H(mf)
// - \gamma_{l} \Delta_{-}(mf,mb) P_{-} )^{-1} \Omega_{-} P_{-} \eta
spProj(in, tmp[0], -1, Lop.Ls);
Lop.Omega(tmp[0], tmp[1], -1, 0);
G5R5(CG_src, tmp[1]);
{
heatbathRefreshShiftCoefficients(0, -1.); //-gamma_1 = -1.
CG_soln = Zero(); // Just use zero as the initial guess
SolverHBL(Lop, CG_src, CG_soln);
Lop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
tmp[1] = Lop.k * tmp[0];
}
Lop.Omega(tmp[1], tmp[0], -1, 1);
spProj(tmp[0], tmp[1], -1, Lop.Ls);
out = out + tmp[1];
// RH terms:
// \Phi = \Phi - k \sum_{k=1}^{N_{p}} P_{+} \Omega_{+}^{\dagger} ( H(mb)
// - \beta_l\gamma_{l} \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} \eta
spProj(in, tmp[0], 1, Rop.Ls);
Rop.Omega(tmp[0], tmp[1], 1, 0);
G5R5(CG_src, tmp[1]);
{
heatbathRefreshShiftCoefficients(1, 0.); //-gamma_1 * beta_1 = 0
CG_soln = Zero();
SolverHBR(Rop, CG_src, CG_soln);
Rop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
tmp[1] = - Rop.k * tmp[0];
}
Rop.Omega(tmp[1], tmp[0], 1, 1);
spProj(tmp[0], tmp[1], 1, Rop.Ls);
out = out + tmp[1];
// Reset shift coefficients for energy and force evals
heatbathRefreshShiftCoefficients(0, 0.0);
heatbathRefreshShiftCoefficients(1, -1.0);
};
// EOFA action: see Eqn. (10) of arXiv:1706.05843
virtual RealD S(const GaugeField& U)
{
@ -271,7 +374,7 @@ NAMESPACE_BEGIN(Grid);
action -= Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
// RH term: S = S + k <\Phi| P_{+} \Omega_{+}^{\dagger} ( H(mb)
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{-} P_{-} |\Phi>
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} |\Phi>
spProj(Phi, spProj_Phi, 1, Rop.Ls);
Rop.Omega(spProj_Phi, tmp[0], 1, 0);
G5R5(tmp[1], tmp[0]);
@ -281,6 +384,26 @@ NAMESPACE_BEGIN(Grid);
Rop.Omega(tmp[1], tmp[0], 1, 1);
action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real();
if(initial_action){
//For the first call to S after refresh, S = |eta|^2. We can use this to ensure the rational approx is good
RealD diff = action - norm2_eta;
//S_init = eta^dag M^{-1/2} M M^{-1/2} eta
//S_init - eta^dag eta = eta^dag ( M^{-1/2} M M^{-1/2} - 1 ) eta
//If approximate solution
//S_init - eta^dag eta = eta^dag ( [M^{-1/2}+\delta M^{-1/2}] M [M^{-1/2}+\delta M^{-1/2}] - 1 ) eta
// \approx eta^dag ( \delta M^{-1/2} M^{1/2} + M^{1/2}\delta M^{-1/2} ) eta
// We divide out |eta|^2 to remove source scaling but the tolerance on this check should still be somewhat higher than the actual approx tolerance
RealD test = fabs(diff)/norm2_eta; //test the quality of the rational approx
std::cout << GridLogMessage << action_name() << " initial action " << action << " expect " << norm2_eta << "; diff " << diff << std::endl;
std::cout << GridLogMessage << action_name() << "[ eta^dag ( M^{-1/2} M M^{-1/2} - 1 ) eta ]/|eta^2| = " << test << " expect 0 (tol " << param.BoundsCheckTol << ")" << std::endl;
assert( ( test < param.BoundsCheckTol ) && " Initial action check failed" );
initial_action = false;
}
return action;
};
@ -329,6 +452,40 @@ NAMESPACE_BEGIN(Grid);
};
};
template<class ImplD, class ImplF>
class ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction : public ExactOneFlavourRatioPseudoFermionAction<ImplD>{
public:
INHERIT_IMPL_TYPES(ImplD);
typedef OneFlavourRationalParams Params;
private:
AbstractEOFAFermion<ImplF>& LopF; // the basic LH operator
AbstractEOFAFermion<ImplF>& RopF; // the basic RH operator
public:
virtual std::string action_name() { return "ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction"; }
//Used in the heatbath, refresh the shift coefficients of the L (LorR=0) or R (LorR=1) operator
virtual void heatbathRefreshShiftCoefficients(int LorR, RealD to){
AbstractEOFAFermion<ImplF> &op = LorR == 0 ? LopF : RopF;
op.RefreshShiftCoefficients(to);
this->ExactOneFlavourRatioPseudoFermionAction<ImplD>::heatbathRefreshShiftCoefficients(LorR,to);
}
ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction(AbstractEOFAFermion<ImplF>& _LopF,
AbstractEOFAFermion<ImplF>& _RopF,
AbstractEOFAFermion<ImplD>& _LopD,
AbstractEOFAFermion<ImplD>& _RopD,
OperatorFunction<FermionField>& HeatbathCGL, OperatorFunction<FermionField>& HeatbathCGR,
OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR,
OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR,
Params& p,
bool use_fc=false) :
LopF(_LopF), RopF(_RopF), ExactOneFlavourRatioPseudoFermionAction<ImplD>(_LopD, _RopD, HeatbathCGL, HeatbathCGR, ActionCGL, ActionCGR, DerivCGL, DerivCGR, p, use_fc){}
};
NAMESPACE_END(Grid);
#endif

372
Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h Normal file
View File

@ -0,0 +1,372 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h
Copyright (C) 2015
Author: Christopher Kelly <ckelly@bnl.gov>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_H
#define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_H
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////
// Generic rational approximation for ratios of operators
/////////////////////////////////////////////////////////
/* S_f = -log( det( [M^dag M]/[V^dag V] )^{1/inv_pow} )
= chi^dag ( [M^dag M]/[V^dag V] )^{-1/inv_pow} chi\
= chi^dag ( [V^dag V]^{-1/2} [M^dag M] [V^dag V]^{-1/2} )^{-1/inv_pow} chi\
= chi^dag [V^dag V]^{1/(2*inv_pow)} [M^dag M]^{-1/inv_pow} [V^dag V]^{1/(2*inv_pow)} chi\
S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
BIG WARNING:
Here V^dag V is referred to in this code as the "numerator" operator and M^dag M is the *denominator* operator.
this refers to their position in the pseudofermion action, which is the *inverse* of what appears in the determinant
Thus for DWF the numerator operator is the Pauli-Villars operator
Here P/Q \sim R_{1/(2*inv_pow)} ~ (V^dagV)^{1/(2*inv_pow)}
Here N/D \sim R_{-1/inv_pow} ~ (M^dagM)^{-1/inv_pow}
*/
template<class Impl>
class GeneralEvenOddRatioRationalPseudoFermionAction : public Action<typename Impl::GaugeField> {
public:
INHERIT_IMPL_TYPES(Impl);
typedef RationalActionParams Params;
Params param;
//For action evaluation
MultiShiftFunction ApproxPowerAction ; //rational approx for X^{1/inv_pow}
MultiShiftFunction ApproxNegPowerAction; //rational approx for X^{-1/inv_pow}
MultiShiftFunction ApproxHalfPowerAction; //rational approx for X^{1/(2*inv_pow)}
MultiShiftFunction ApproxNegHalfPowerAction; //rational approx for X^{-1/(2*inv_pow)}
//For the MD integration
MultiShiftFunction ApproxPowerMD ; //rational approx for X^{1/inv_pow}
MultiShiftFunction ApproxNegPowerMD; //rational approx for X^{-1/inv_pow}
MultiShiftFunction ApproxHalfPowerMD; //rational approx for X^{1/(2*inv_pow)}
MultiShiftFunction ApproxNegHalfPowerMD; //rational approx for X^{-1/(2*inv_pow)}
private:
FermionOperator<Impl> & NumOp;// the basic operator
FermionOperator<Impl> & DenOp;// the basic operator
FermionField PhiEven; // the pseudo fermion field for this trajectory
FermionField PhiOdd; // the pseudo fermion field for this trajectory
//Generate the approximation to x^{1/inv_pow} (->approx) and x^{-1/inv_pow} (-> approx_inv) by an approx_degree degree rational approximation
//CG_tolerance is used to issue a warning if the approximation error is larger than the tolerance of the CG and is otherwise just stored in the MultiShiftFunction for use by the multi-shift
static void generateApprox(MultiShiftFunction &approx, MultiShiftFunction &approx_inv, int inv_pow, int approx_degree, double CG_tolerance, AlgRemez &remez){
std::cout<<GridLogMessage << "Generating degree "<< approx_degree<<" approximation for x^(1/" << inv_pow << ")"<<std::endl;
double error = remez.generateApprox(approx_degree,1,inv_pow);
if(error > CG_tolerance)
std::cout<<GridLogMessage << "WARNING: Remez approximation has a larger error " << error << " than the CG tolerance " << CG_tolerance << "! Try increasing the number of poles" << std::endl;
approx.Init(remez, CG_tolerance,false);
approx_inv.Init(remez, CG_tolerance,true);
}
protected:
static constexpr bool Numerator = true;
static constexpr bool Denominator = false;
//Allow derived classes to override the multishift CG
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionField &in, FermionField &out){
SchurDifferentiableOperator<Impl> schurOp(numerator ? NumOp : DenOp);
ConjugateGradientMultiShift<FermionField> msCG(MaxIter, approx);
msCG(schurOp,in, out);
}
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionField &in, std::vector<FermionField> &out_elems, FermionField &out){
SchurDifferentiableOperator<Impl> schurOp(numerator ? NumOp : DenOp);
ConjugateGradientMultiShift<FermionField> msCG(MaxIter, approx);
msCG(schurOp,in, out_elems, out);
}
//Allow derived classes to override the gauge import
virtual void ImportGauge(const GaugeField &U){
NumOp.ImportGauge(U);
DenOp.ImportGauge(U);
}
public:
GeneralEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl> &_NumOp,
FermionOperator<Impl> &_DenOp,
const Params & p
) :
NumOp(_NumOp),
DenOp(_DenOp),
PhiOdd (_NumOp.FermionRedBlackGrid()),
PhiEven(_NumOp.FermionRedBlackGrid()),
param(p)
{
std::cout<<GridLogMessage << action_name() << " initialize: starting" << std::endl;
AlgRemez remez(param.lo,param.hi,param.precision);
//Generate approximations for action eval
generateApprox(ApproxPowerAction, ApproxNegPowerAction, param.inv_pow, param.action_degree, param.action_tolerance, remez);
generateApprox(ApproxHalfPowerAction, ApproxNegHalfPowerAction, 2*param.inv_pow, param.action_degree, param.action_tolerance, remez);
//Generate approximations for MD
if(param.md_degree != param.action_degree){ //note the CG tolerance is unrelated to the stopping condition of the Remez algorithm
generateApprox(ApproxPowerMD, ApproxNegPowerMD, param.inv_pow, param.md_degree, param.md_tolerance, remez);
generateApprox(ApproxHalfPowerMD, ApproxNegHalfPowerMD, 2*param.inv_pow, param.md_degree, param.md_tolerance, remez);
}else{
std::cout<<GridLogMessage << "Using same rational approximations for MD as for action evaluation" << std::endl;
ApproxPowerMD = ApproxPowerAction;
ApproxNegPowerMD = ApproxNegPowerAction;
for(int i=0;i<ApproxPowerMD.tolerances.size();i++)
ApproxNegPowerMD.tolerances[i] = ApproxPowerMD.tolerances[i] = param.md_tolerance; //used for multishift
ApproxHalfPowerMD = ApproxHalfPowerAction;
ApproxNegHalfPowerMD = ApproxNegHalfPowerAction;
for(int i=0;i<ApproxPowerMD.tolerances.size();i++)
ApproxNegHalfPowerMD.tolerances[i] = ApproxHalfPowerMD.tolerances[i] = param.md_tolerance;
}
std::cout<<GridLogMessage << action_name() << " initialize: complete" << std::endl;
};
virtual std::string action_name(){return "GeneralEvenOddRatioRationalPseudoFermionAction";}
virtual std::string LogParameters(){
std::stringstream sstream;
sstream << GridLogMessage << "["<<action_name()<<"] Power : 1/" << param.inv_pow << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Low :" << param.lo << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] High :" << param.hi << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Max iterations :" << param.MaxIter << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Tolerance (Action) :" << param.action_tolerance << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Degree (Action) :" << param.action_degree << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Tolerance (MD) :" << param.md_tolerance << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Degree (MD) :" << param.md_degree << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Precision :" << param.precision << std::endl;
return sstream.str();
}
//Access the fermion field
const FermionField &getPhiOdd() const{ return PhiOdd; }
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
FermionField eta(NumOp.FermionGrid());
// P(eta) \propto e^{- eta^dag eta}
//
// The gaussian function draws from P(x) \propto e^{- x^2 / 2 } [i.e. sigma=1]
// Thus eta = x/sqrt{2} = x * sqrt(1/2)
RealD scale = std::sqrt(0.5);
gaussian(pRNG,eta); eta=eta*scale;
refresh(U,eta);
}
//Allow for manual specification of random field for testing
void refresh(const GaugeField &U, const FermionField &eta) {
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
//
// P(phi) = e^{- phi^dag (VdagV)^1/(2*inv_pow) (MdagM)^-1/inv_pow (VdagV)^1/(2*inv_pow) phi}
// = e^{- phi^dag (VdagV)^1/(2*inv_pow) (MdagM)^-1/(2*inv_pow) (MdagM)^-1/(2*inv_pow) (VdagV)^1/(2*inv_pow) phi}
//
// Phi = (VdagV)^-1/(2*inv_pow) Mdag^{1/(2*inv_pow)} eta
std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
FermionField etaOdd (NumOp.FermionRedBlackGrid());
FermionField etaEven(NumOp.FermionRedBlackGrid());
FermionField tmp(NumOp.FermionRedBlackGrid());
pickCheckerboard(Even,etaEven,eta);
pickCheckerboard(Odd,etaOdd,eta);
ImportGauge(U);
// MdagM^1/(2*inv_pow) eta
std::cout<<GridLogMessage << action_name() << " refresh: doing (M^dag M)^{1/" << 2*param.inv_pow << "} eta" << std::endl;
multiShiftInverse(Denominator, ApproxHalfPowerAction, param.MaxIter, etaOdd, tmp);
// VdagV^-1/(2*inv_pow) MdagM^1/(2*inv_pow) eta
std::cout<<GridLogMessage << action_name() << " refresh: doing (V^dag V)^{-1/" << 2*param.inv_pow << "} ( (M^dag M)^{1/" << 2*param.inv_pow << "} eta)" << std::endl;
multiShiftInverse(Numerator, ApproxNegHalfPowerAction, param.MaxIter, tmp, PhiOdd);
assert(NumOp.ConstEE() == 1);
assert(DenOp.ConstEE() == 1);
PhiEven = Zero();
std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
};
//////////////////////////////////////////////////////
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
//////////////////////////////////////////////////////
virtual RealD S(const GaugeField &U) {
std::cout<<GridLogMessage << action_name() << " compute action: starting" << std::endl;
ImportGauge(U);
FermionField X(NumOp.FermionRedBlackGrid());
FermionField Y(NumOp.FermionRedBlackGrid());
// VdagV^1/(2*inv_pow) Phi
std::cout<<GridLogMessage << action_name() << " compute action: doing (V^dag V)^{1/" << 2*param.inv_pow << "} Phi" << std::endl;
multiShiftInverse(Numerator, ApproxHalfPowerAction, param.MaxIter, PhiOdd,X);
// MdagM^-1/(2*inv_pow) VdagV^1/(2*inv_pow) Phi
std::cout<<GridLogMessage << action_name() << " compute action: doing (M^dag M)^{-1/" << 2*param.inv_pow << "} ( (V^dag V)^{1/" << 2*param.inv_pow << "} Phi)" << std::endl;
multiShiftInverse(Denominator, ApproxNegHalfPowerAction, param.MaxIter, X,Y);
// Randomly apply rational bounds checks.
int rcheck = rand();
auto grid = NumOp.FermionGrid();
auto r=rand();
grid->Broadcast(0,r);
if ( param.BoundsCheckFreq != 0 && (r % param.BoundsCheckFreq)==0 ) {
std::cout<<GridLogMessage << action_name() << " compute action: doing bounds check" << std::endl;
FermionField gauss(NumOp.FermionRedBlackGrid());
gauss = PhiOdd;
SchurDifferentiableOperator<Impl> MdagM(DenOp);
std::cout<<GridLogMessage << action_name() << " compute action: checking high bounds" << std::endl;
HighBoundCheck(MdagM,gauss,param.hi);
std::cout<<GridLogMessage << action_name() << " compute action: full approximation" << std::endl;
InversePowerBoundsCheck(param.inv_pow,param.MaxIter,param.action_tolerance*100,MdagM,gauss,ApproxNegPowerAction);
std::cout<<GridLogMessage << action_name() << " compute action: bounds check complete" << std::endl;
}
// Phidag VdagV^1/(2*inv_pow) MdagM^-1/(2*inv_pow) MdagM^-1/(2*inv_pow) VdagV^1/(2*inv_pow) Phi
RealD action = norm2(Y);
std::cout<<GridLogMessage << action_name() << " compute action: complete" << std::endl;
return action;
};
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
//
// Here, M is some 5D operator and V is the Pauli-Villars field
// N and D makeup the rat. poly of the M term and P and & makeup the rat.poly of the denom term
//
// Need
// dS_f/dU = chi^dag d[P/Q] N/D P/Q chi
// + chi^dag P/Q d[N/D] P/Q chi
// + chi^dag P/Q N/D d[P/Q] chi
//
// P/Q is expressed as partial fraction expansion:
//
// a0 + \sum_k ak/(V^dagV + bk)
//
// d[P/Q] is then
//
// \sum_k -ak [V^dagV+bk]^{-1} [ dV^dag V + V^dag dV ] [V^dag V + bk]^{-1}
//
// and similar for N/D.
//
// Need
// MpvPhi_k = [Vdag V + bk]^{-1} chi
// MpvPhi = {a0 + \sum_k ak [Vdag V + bk]^{-1} }chi
//
// MfMpvPhi_k = [MdagM+bk]^{-1} MpvPhi
// MfMpvPhi = {a0 + \sum_k ak [Mdag M + bk]^{-1} } MpvPhi
//
// MpvMfMpvPhi_k = [Vdag V + bk]^{-1} MfMpvchi
//
virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
std::cout<<GridLogMessage << action_name() << " deriv: starting" << std::endl;
const int n_f = ApproxNegPowerMD.poles.size();
const int n_pv = ApproxHalfPowerMD.poles.size();
std::vector<FermionField> MpvPhi_k (n_pv,NumOp.FermionRedBlackGrid());
std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionRedBlackGrid());
std::vector<FermionField> MfMpvPhi_k (n_f ,NumOp.FermionRedBlackGrid());
FermionField MpvPhi(NumOp.FermionRedBlackGrid());
FermionField MfMpvPhi(NumOp.FermionRedBlackGrid());
FermionField MpvMfMpvPhi(NumOp.FermionRedBlackGrid());
FermionField Y(NumOp.FermionRedBlackGrid());
GaugeField tmp(NumOp.GaugeGrid());
ImportGauge(U);
std::cout<<GridLogMessage << action_name() << " deriv: doing (V^dag V)^{1/" << 2*param.inv_pow << "} Phi" << std::endl;
multiShiftInverse(Numerator, ApproxHalfPowerMD, param.MaxIter, PhiOdd,MpvPhi_k,MpvPhi);
std::cout<<GridLogMessage << action_name() << " deriv: doing (M^dag M)^{-1/" << param.inv_pow << "} ( (V^dag V)^{1/" << 2*param.inv_pow << "} Phi)" << std::endl;
multiShiftInverse(Denominator, ApproxNegPowerMD, param.MaxIter, MpvPhi,MfMpvPhi_k,MfMpvPhi);
std::cout<<GridLogMessage << action_name() << " deriv: doing (V^dag V)^{1/" << 2*param.inv_pow << "} ( (M^dag M)^{-1/" << param.inv_pow << "} (V^dag V)^{1/" << 2*param.inv_pow << "} Phi)" << std::endl;
multiShiftInverse(Numerator, ApproxHalfPowerMD, param.MaxIter, MfMpvPhi,MpvMfMpvPhi_k,MpvMfMpvPhi);
SchurDifferentiableOperator<Impl> MdagM(DenOp);
SchurDifferentiableOperator<Impl> VdagV(NumOp);
RealD ak;
dSdU = Zero();
// With these building blocks
//
// dS/dU =
// \sum_k -ak MfMpvPhi_k^dag [ dM^dag M + M^dag dM ] MfMpvPhi_k (1)
// + \sum_k -ak MpvMfMpvPhi_k^\dag [ dV^dag V + V^dag dV ] MpvPhi_k (2)
// -ak MpvPhi_k^dag [ dV^dag V + V^dag dV ] MpvMfMpvPhi_k (3)
//(1)
std::cout<<GridLogMessage << action_name() << " deriv: doing dS/dU part (1)" << std::endl;
for(int k=0;k<n_f;k++){
ak = ApproxNegPowerMD.residues[k];
MdagM.Mpc(MfMpvPhi_k[k],Y);
MdagM.MpcDagDeriv(tmp , MfMpvPhi_k[k], Y ); dSdU=dSdU+ak*tmp;
MdagM.MpcDeriv(tmp , Y, MfMpvPhi_k[k] ); dSdU=dSdU+ak*tmp;
}
//(2)
//(3)
std::cout<<GridLogMessage << action_name() << " deriv: doing dS/dU part (2)+(3)" << std::endl;
for(int k=0;k<n_pv;k++){
ak = ApproxHalfPowerMD.residues[k];
VdagV.Mpc(MpvPhi_k[k],Y);
VdagV.MpcDagDeriv(tmp,MpvMfMpvPhi_k[k],Y); dSdU=dSdU+ak*tmp;
VdagV.MpcDeriv (tmp,Y,MpvMfMpvPhi_k[k]); dSdU=dSdU+ak*tmp;
VdagV.Mpc(MpvMfMpvPhi_k[k],Y); // V as we take Ydag
VdagV.MpcDeriv (tmp,Y, MpvPhi_k[k]); dSdU=dSdU+ak*tmp;
VdagV.MpcDagDeriv(tmp,MpvPhi_k[k], Y); dSdU=dSdU+ak*tmp;
}
//dSdU = Ta(dSdU);
std::cout<<GridLogMessage << action_name() << " deriv: complete" << std::endl;
};
};
NAMESPACE_END(Grid);
#endif

93
Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h Normal file
View File

@ -0,0 +1,93 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h
Copyright (C) 2015
Author: Christopher Kelly <ckelly@bnl.gov>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
#define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Generic rational approximation for ratios of operators utilizing the mixed precision multishift algorithm
// cf. GeneralEvenOddRational.h for details
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class ImplD, class ImplF>
class GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<ImplD> {
private:
typedef typename ImplD::FermionField FermionFieldD;
typedef typename ImplF::FermionField FermionFieldF;
FermionOperator<ImplD> & NumOpD;
FermionOperator<ImplD> & DenOpD;
FermionOperator<ImplF> & NumOpF;
FermionOperator<ImplF> & DenOpF;
Integer ReliableUpdateFreq;
protected:
//Allow derived classes to override the multishift CG
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
msCG(schurOpD, in, out);
}
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
msCG(schurOpD, in, out_elems, out);
}
//Allow derived classes to override the gauge import
virtual void ImportGauge(const typename ImplD::GaugeField &Ud){
typename ImplF::GaugeField Uf(NumOpF.GaugeGrid());
precisionChange(Uf, Ud);
NumOpD.ImportGauge(Ud);
DenOpD.ImportGauge(Ud);
NumOpF.ImportGauge(Uf);
DenOpF.ImportGauge(Uf);
}
public:
GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction(FermionOperator<ImplD> &_NumOpD, FermionOperator<ImplD> &_DenOpD,
FermionOperator<ImplF> &_NumOpF, FermionOperator<ImplF> &_DenOpF,
const RationalActionParams & p, Integer _ReliableUpdateFreq
) : GeneralEvenOddRatioRationalPseudoFermionAction<ImplD>(_NumOpD, _DenOpD, p),
ReliableUpdateFreq(_ReliableUpdateFreq), NumOpD(_NumOpD), DenOpD(_DenOpD), NumOpF(_NumOpF), DenOpF(_DenOpF){}
virtual std::string action_name(){return "GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction";}
};
NAMESPACE_END(Grid);
#endif

262
Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h
View File

@ -40,257 +40,31 @@ NAMESPACE_BEGIN(Grid);
// Here N/D \sim R_{-1/2} ~ (M^dagM)^{-1/2}
template<class Impl>
class OneFlavourEvenOddRatioRationalPseudoFermionAction : public Action<typename Impl::GaugeField> {
class OneFlavourEvenOddRatioRationalPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<Impl> {
public:
INHERIT_IMPL_TYPES(Impl);
typedef OneFlavourRationalParams Params;
Params param;
MultiShiftFunction PowerHalf ;
MultiShiftFunction PowerNegHalf;
MultiShiftFunction PowerQuarter;
MultiShiftFunction PowerNegQuarter;
private:
FermionOperator<Impl> & NumOp;// the basic operator
FermionOperator<Impl> & DenOp;// the basic operator
FermionField PhiEven; // the pseudo fermion field for this trajectory
FermionField PhiOdd; // the pseudo fermion field for this trajectory
FermionField Noise; // spare noise field for bounds check
static RationalActionParams transcribe(const Params &in){
RationalActionParams out;
out.inv_pow = 2;
out.lo = in.lo;
out.hi = in.hi;
out.MaxIter = in.MaxIter;
out.action_tolerance = out.md_tolerance = in.tolerance;
out.action_degree = out.md_degree = in.degree;
out.precision = in.precision;
out.BoundsCheckFreq = in.BoundsCheckFreq;
return out;
}
public:
OneFlavourEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl> &_NumOp,
FermionOperator<Impl> &_DenOp,
Params & p
) :
NumOp(_NumOp),
DenOp(_DenOp),
PhiOdd (_NumOp.FermionRedBlackGrid()),
PhiEven(_NumOp.FermionRedBlackGrid()),
Noise(_NumOp.FermionRedBlackGrid()),
param(p)
{
AlgRemez remez(param.lo,param.hi,param.precision);
FermionOperator<Impl> &_DenOp,
const Params & p
) :
GeneralEvenOddRatioRationalPseudoFermionAction<Impl>(_NumOp, _DenOp, transcribe(p)){}
// MdagM^(+- 1/2)
std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/2)"<<std::endl;
remez.generateApprox(param.degree,1,2);
PowerHalf.Init(remez,param.tolerance,false);
PowerNegHalf.Init(remez,param.tolerance,true);
// MdagM^(+- 1/4)
std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/4)"<<std::endl;
remez.generateApprox(param.degree,1,4);
PowerQuarter.Init(remez,param.tolerance,false);
PowerNegQuarter.Init(remez,param.tolerance,true);
};
virtual std::string action_name(){
std::stringstream sstream;
sstream<< "OneFlavourEvenOddRatioRationalPseudoFermionAction det("<< DenOp.Mass() << ") / det("<<NumOp.Mass()<<")";
return sstream.str();
}
virtual std::string LogParameters(){
std::stringstream sstream;
sstream << GridLogMessage << "["<<action_name()<<"] Low :" << param.lo << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] High :" << param.hi << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Max iterations :" << param.MaxIter << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Tolerance :" << param.tolerance << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Degree :" << param.degree << std::endl;
sstream << GridLogMessage << "["<<action_name()<<"] Precision :" << param.precision << std::endl;
return sstream.str();
}
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
//
// P(phi) = e^{- phi^dag (VdagV)^1/4 (MdagM)^-1/2 (VdagV)^1/4 phi}
// = e^{- phi^dag (VdagV)^1/4 (MdagM)^-1/4 (MdagM)^-1/4 (VdagV)^1/4 phi}
//
// Phi = (VdagV)^-1/4 Mdag^{1/4} eta
//
// P(eta) = e^{- eta^dag eta}
//
// e^{x^2/2 sig^2} => sig^2 = 0.5.
//
// So eta should be of width sig = 1/sqrt(2).
RealD scale = std::sqrt(0.5);
FermionField eta(NumOp.FermionGrid());
FermionField etaOdd (NumOp.FermionRedBlackGrid());
FermionField etaEven(NumOp.FermionRedBlackGrid());
FermionField tmp(NumOp.FermionRedBlackGrid());
gaussian(pRNG,eta); eta=eta*scale;
pickCheckerboard(Even,etaEven,eta);
pickCheckerboard(Odd,etaOdd,eta);
Noise = etaOdd;
NumOp.ImportGauge(U);
DenOp.ImportGauge(U);
// MdagM^1/4 eta
SchurDifferentiableOperator<Impl> MdagM(DenOp);
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerQuarter);
msCG_M(MdagM,etaOdd,tmp);
// VdagV^-1/4 MdagM^1/4 eta
SchurDifferentiableOperator<Impl> VdagV(NumOp);
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerNegQuarter);
msCG_V(VdagV,tmp,PhiOdd);
assert(NumOp.ConstEE() == 1);
assert(DenOp.ConstEE() == 1);
PhiEven = Zero();
};
//////////////////////////////////////////////////////
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
//////////////////////////////////////////////////////
virtual RealD S(const GaugeField &U) {
NumOp.ImportGauge(U);
DenOp.ImportGauge(U);
FermionField X(NumOp.FermionRedBlackGrid());
FermionField Y(NumOp.FermionRedBlackGrid());
// VdagV^1/4 Phi
SchurDifferentiableOperator<Impl> VdagV(NumOp);
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
msCG_V(VdagV,PhiOdd,X);
// MdagM^-1/4 VdagV^1/4 Phi
SchurDifferentiableOperator<Impl> MdagM(DenOp);
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegQuarter);
msCG_M(MdagM,X,Y);
// Randomly apply rational bounds checks.
auto grid = NumOp.FermionGrid();
auto r=rand();
grid->Broadcast(0,r);
if ( (r%param.BoundsCheckFreq)==0 ) {
FermionField gauss(NumOp.FermionRedBlackGrid());
gauss = Noise;
HighBoundCheck(MdagM,gauss,param.hi);
InverseSqrtBoundsCheck(param.MaxIter,param.tolerance*100,MdagM,gauss,PowerNegHalf);
ChebyBoundsCheck(MdagM,Noise,param.lo,param.hi);
}
// Phidag VdagV^1/4 MdagM^-1/4 MdagM^-1/4 VdagV^1/4 Phi
RealD action = norm2(Y);
return action;
};
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
//
// Here, M is some 5D operator and V is the Pauli-Villars field
// N and D makeup the rat. poly of the M term and P and & makeup the rat.poly of the denom term
//
// Need
// dS_f/dU = chi^dag d[P/Q] N/D P/Q chi
// + chi^dag P/Q d[N/D] P/Q chi
// + chi^dag P/Q N/D d[P/Q] chi
//
// P/Q is expressed as partial fraction expansion:
//
// a0 + \sum_k ak/(V^dagV + bk)
//
// d[P/Q] is then
//
// \sum_k -ak [V^dagV+bk]^{-1} [ dV^dag V + V^dag dV ] [V^dag V + bk]^{-1}
//
// and similar for N/D.
//
// Need
// MpvPhi_k = [Vdag V + bk]^{-1} chi
// MpvPhi = {a0 + \sum_k ak [Vdag V + bk]^{-1} }chi
//
// MfMpvPhi_k = [MdagM+bk]^{-1} MpvPhi
// MfMpvPhi = {a0 + \sum_k ak [Mdag M + bk]^{-1} } MpvPhi
//
// MpvMfMpvPhi_k = [Vdag V + bk]^{-1} MfMpvchi
//
virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
const int n_f = PowerNegHalf.poles.size();
const int n_pv = PowerQuarter.poles.size();
std::vector<FermionField> MpvPhi_k (n_pv,NumOp.FermionRedBlackGrid());
std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionRedBlackGrid());
std::vector<FermionField> MfMpvPhi_k (n_f ,NumOp.FermionRedBlackGrid());
FermionField MpvPhi(NumOp.FermionRedBlackGrid());
FermionField MfMpvPhi(NumOp.FermionRedBlackGrid());
FermionField MpvMfMpvPhi(NumOp.FermionRedBlackGrid());
FermionField Y(NumOp.FermionRedBlackGrid());
GaugeField tmp(NumOp.GaugeGrid());
NumOp.ImportGauge(U);
DenOp.ImportGauge(U);
SchurDifferentiableOperator<Impl> VdagV(NumOp);
SchurDifferentiableOperator<Impl> MdagM(DenOp);
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
msCG_V(VdagV,PhiOdd,MpvPhi_k,MpvPhi);
msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
msCG_V(VdagV,MfMpvPhi,MpvMfMpvPhi_k,MpvMfMpvPhi);
RealD ak;
dSdU = Zero();
// With these building blocks
//
// dS/dU =
// \sum_k -ak MfMpvPhi_k^dag [ dM^dag M + M^dag dM ] MfMpvPhi_k (1)
// + \sum_k -ak MpvMfMpvPhi_k^\dag [ dV^dag V + V^dag dV ] MpvPhi_k (2)
// -ak MpvPhi_k^dag [ dV^dag V + V^dag dV ] MpvMfMpvPhi_k (3)
//(1)
for(int k=0;k<n_f;k++){
ak = PowerNegHalf.residues[k];
MdagM.Mpc(MfMpvPhi_k[k],Y);
MdagM.MpcDagDeriv(tmp , MfMpvPhi_k[k], Y ); dSdU=dSdU+ak*tmp;
MdagM.MpcDeriv(tmp , Y, MfMpvPhi_k[k] ); dSdU=dSdU+ak*tmp;
}
//(2)
//(3)
for(int k=0;k<n_pv;k++){
ak = PowerQuarter.residues[k];
VdagV.Mpc(MpvPhi_k[k],Y);
VdagV.MpcDagDeriv(tmp,MpvMfMpvPhi_k[k],Y); dSdU=dSdU+ak*tmp;
VdagV.MpcDeriv (tmp,Y,MpvMfMpvPhi_k[k]); dSdU=dSdU+ak*tmp;
VdagV.Mpc(MpvMfMpvPhi_k[k],Y); // V as we take Ydag
VdagV.MpcDeriv (tmp,Y, MpvPhi_k[k]); dSdU=dSdU+ak*tmp;
VdagV.MpcDagDeriv(tmp,MpvPhi_k[k], Y); dSdU=dSdU+ak*tmp;
}
//dSdU = Ta(dSdU);
};
virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}
};
NAMESPACE_END(Grid);

2
Grid/qcd/action/pseudofermion/PseudoFermion.h
View File

@ -40,6 +40,8 @@ directory
#include <Grid/qcd/action/pseudofermion/OneFlavourRational.h>
#include <Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h>
#include <Grid/qcd/action/pseudofermion/OneFlavourEvenOddRational.h>
#include <Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h>
#include <Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h>
#include <Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h>
#include <Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h>

28
Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h
View File

@ -88,15 +88,9 @@ NAMESPACE_BEGIN(Grid);
}
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
const FermionField &getPhiOdd() const{ return PhiOdd; }
// P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
//
// NumOp == V
// DenOp == M
//
// Take phi_o = Vpcdag^{-1} Mpcdag eta_o ; eta_o = Mpcdag^{-1} Vpcdag Phi
//
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
// P(eta_o) = e^{- eta_o^dag eta_o}
//
// e^{x^2/2 sig^2} => sig^2 = 0.5.
@ -104,12 +98,22 @@ NAMESPACE_BEGIN(Grid);
RealD scale = std::sqrt(0.5);
FermionField eta (NumOp.FermionGrid());
gaussian(pRNG,eta); eta = eta * scale;
refresh(U,eta);
}
void refresh(const GaugeField &U, const FermionField &eta) {
// P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
//
// NumOp == V
// DenOp == M
//
FermionField etaOdd (NumOp.FermionRedBlackGrid());
FermionField etaEven(NumOp.FermionRedBlackGrid());
FermionField tmp (NumOp.FermionRedBlackGrid());
gaussian(pRNG,eta);
pickCheckerboard(Even,etaEven,eta);
pickCheckerboard(Odd,etaOdd,eta);
@ -128,10 +132,6 @@ NAMESPACE_BEGIN(Grid);
// Even det factors
DenOp.MooeeDag(etaEven,tmp);
NumOp.MooeeInvDag(tmp,PhiEven);
PhiOdd =PhiOdd*scale;
PhiEven=PhiEven*scale;
};
//////////////////////////////////////////////////////

12
Grid/qcd/hmc/GenericHMCrunner.h
View File

@ -151,12 +151,22 @@ public:
Resources.GetCheckPointer()->CheckpointRestore(Parameters.StartTrajectory, U,
Resources.GetSerialRNG(),
Resources.GetParallelRNG());
} else if (Parameters.StartingType == "CheckpointStartReseed") {
// Same as CheckpointRestart but reseed the RNGs using the fixed integer seeding used for ColdStart and HotStart
// Useful for creating new evolution streams from an existing stream
// WARNING: Unfortunately because the checkpointer doesn't presently allow us to separately restore the RNG and gauge fields we have to load
// an existing RNG checkpoint first; make sure one is available and named correctly
Resources.GetCheckPointer()->CheckpointRestore(Parameters.StartTrajectory, U,
Resources.GetSerialRNG(),
Resources.GetParallelRNG());
Resources.SeedFixedIntegers();
} else {
// others
std::cout << GridLogError << "Unrecognized StartingType\n";
std::cout
<< GridLogError
<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart]\n";
<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart, CheckpointStartReseed]\n";
exit(1);
}
}

2
Grid/qcd/hmc/HMCModules.h
View File

@ -80,7 +80,9 @@ public:
std::cout << GridLogError << "Seeds not initialized" << std::endl;
exit(1);
}
std::cout << GridLogMessage << "Reseeding serial RNG with seed vector " << SerialSeeds << std::endl;
sRNG_.SeedFixedIntegers(SerialSeeds);
std::cout << GridLogMessage << "Reseeding parallel RNG with seed vector " << ParallelSeeds << std::endl;
pRNG_->SeedFixedIntegers(ParallelSeeds);
}
};

6
Grid/qcd/hmc/integrators/Integrator.h
View File

@ -334,15 +334,19 @@ public:
void refresh(Field& U, GridSerialRNG & sRNG, GridParallelRNG& pRNG)
{
assert(P.Grid() == U.Grid());
std::cout << GridLogIntegrator << "Integrator refresh\n";
std::cout << GridLogIntegrator << "Integrator refresh" << std::endl;
std::cout << GridLogIntegrator << "Generating momentum" << std::endl;
FieldImplementation::generate_momenta(P, sRNG, pRNG);
// Update the smeared fields, can be implemented as observer
// necessary to keep the fields updated even after a reject
// of the Metropolis
std::cout << GridLogIntegrator << "Updating smeared fields" << std::endl;
Smearer.set_Field(U);
// Set the (eventual) representations gauge fields
std::cout << GridLogIntegrator << "Updating representations" << std::endl;
Representations.update(U);
// The Smearer is attached to a pointer of the gauge field