portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Merge remote-tracking branch 'upstream/develop' into feature/kl2QED

Browse Source
This commit is contained in:
Vera Guelpers 2019-05-23 12:41:54 +01:00
commit e675c6a48c
14 changed files with 466 additions and 70 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
Grid/algorithms/LinearOperator.h
View File

@ -212,9 +212,8 @@ namespace Grid {
}; };
template<class Matrix,class Field> template<class Matrix,class Field>
class SchurDiagMooeeOperator : public SchurOperatorBase<Field> { class SchurDiagMooeeOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
public: public:
Matrix &_Mat;
SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){}; SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) { virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in._grid); Field tmp(in._grid);

7
Grid/algorithms/iterative/ConjugateGradient.h
View File

@ -89,6 +89,8 @@ class ConjugateGradient : public OperatorFunction<Field> {
// Check if guess is really REALLY good :) // Check if guess is really REALLY good :)
if (cp <= rsq) { if (cp <= rsq) {
std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
IterationsToComplete = 0;
return; return;
} }
@ -104,7 +106,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
SolverTimer.Start(); SolverTimer.Start();
int k; int k;
for (k = 1; k <= MaxIterations*1000; k++) { for (k = 1; k <= MaxIterations; k++) {
c = cp; c = cp;
MatrixTimer.Start(); MatrixTimer.Start();
@ -165,8 +167,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
return; return;
} }
} }
std::cout << GridLogMessage << "ConjugateGradient did NOT converge" std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
<< std::endl;
if (ErrorOnNoConverge) assert(0); if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k; IterationsToComplete = k;

14
Grid/algorithms/iterative/ConjugateGradientMixedPrec.h
View File

@ -30,8 +30,11 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
namespace Grid { namespace Grid {
//Mixed precision restarted defect correction CG //Mixed precision restarted defect correction CG
template<class FieldD,class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> template<class FieldD,class FieldF,
typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> { class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
public: public:
RealD Tolerance; RealD Tolerance;
@ -50,7 +53,12 @@ namespace Grid {
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
LinearFunction<FieldF> *guesser; LinearFunction<FieldF> *guesser;
MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) : MixedPrecisionConjugateGradient(RealD tol,
Integer maxinnerit,
Integer maxouterit,
GridBase* _sp_grid,
LinearOperatorBase<FieldF> &_Linop_f,
LinearOperatorBase<FieldD> &_Linop_d) :
Linop_f(_Linop_f), Linop_d(_Linop_d), Linop_f(_Linop_f), Linop_d(_Linop_d),
Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid), Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
OuterLoopNormMult(100.), guesser(NULL){ }; OuterLoopNormMult(100.), guesser(NULL){ };
@ -149,6 +157,8 @@ namespace Grid {
} }
}; };
} }
#endif #endif

6
Grid/algorithms/iterative/Deflation.h
View File

@ -35,7 +35,11 @@ class ZeroGuesser: public LinearFunction<Field> {
public: public:
virtual void operator()(const Field &src, Field &guess) { guess = zero; }; virtual void operator()(const Field &src, Field &guess) { guess = zero; };
}; };
template<class Field>
class DoNothingGuesser: public LinearFunction<Field> {
public:
virtual void operator()(const Field &src, Field &guess) { };
};
template<class Field> template<class Field>
class SourceGuesser: public LinearFunction<Field> { class SourceGuesser: public LinearFunction<Field> {
public: public:

2
Grid/algorithms/iterative/SchurRedBlack.h
View File

@ -261,7 +261,7 @@ namespace Grid {
} }
///////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////
// Override in derived. Not virtual as template methods // Override in derived.
///////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////
virtual void RedBlackSource (Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o) =0; virtual void RedBlackSource (Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o) =0;
virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol) =0; virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol) =0;

2
Grid/communicator/SharedMemory.h
View File

@ -103,6 +103,8 @@ class GlobalSharedMemory {
////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////
static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorHypercube(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorSharedMemory(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
/////////////////////////////////////////////////// ///////////////////////////////////////////////////
// Provide shared memory facilities off comm world // Provide shared memory facilities off comm world
/////////////////////////////////////////////////// ///////////////////////////////////////////////////

24
Grid/communicator/SharedMemoryMPI.cc
View File

@ -132,7 +132,22 @@ int Log2Size(int TwoToPower,int MAXLOG2)
} }
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm) void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
{ {
#ifdef HYPERCUBE //////////////////////////////////////////////////////////////////////////////
// Look and see if it looks like an HPE 8600 based on hostname conventions
//////////////////////////////////////////////////////////////////////////////
const int namelen = _POSIX_HOST_NAME_MAX;
char name[namelen];
int R;
int I;
int N;
gethostname(name,namelen);
int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
if(nscan==3) OptimalCommunicatorHypercube(processors,optimal_comm);
else OptimalCommunicatorSharedMemory(processors,optimal_comm);
}
void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
{
//////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////
// Assert power of two shm_size. // Assert power of two shm_size.
//////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////
@ -253,7 +268,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
///////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////////
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm); int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
assert(ierr==0); assert(ierr==0);
#else }
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
{
//////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////
// Assert power of two shm_size. // Assert power of two shm_size.
//////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////
@ -306,7 +323,6 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
///////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////////
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm); int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
assert(ierr==0); assert(ierr==0);
#endif
} }
//////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////
// SHMGET // SHMGET
@ -337,7 +353,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
int errsv = errno; int errsv = errno;
printf("Errno %d\n",errsv); printf("Errno %d\n",errsv);
printf("key %d\n",key); printf("key %d\n",key);
printf("size %lld\n",size); printf("size %ld\n",size);
printf("flags %d\n",flags); printf("flags %d\n",flags);
perror("shmget"); perror("shmget");
exit(1); exit(1);

83
Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h
View File

@ -58,13 +58,29 @@ namespace QCD{
bool use_heatbath_forecasting; bool use_heatbath_forecasting;
AbstractEOFAFermion<Impl>& Lop; // the basic LH operator AbstractEOFAFermion<Impl>& Lop; // the basic LH operator
AbstractEOFAFermion<Impl>& Rop; // the basic RH operator AbstractEOFAFermion<Impl>& Rop; // the basic RH operator
SchurRedBlackDiagMooeeSolve<FermionField> Solver; SchurRedBlackDiagMooeeSolve<FermionField> SolverHB;
SchurRedBlackDiagMooeeSolve<FermionField> SolverL;
SchurRedBlackDiagMooeeSolve<FermionField> SolverR;
SchurRedBlackDiagMooeeSolve<FermionField> DerivativeSolverL;
SchurRedBlackDiagMooeeSolve<FermionField> DerivativeSolverR;
FermionField Phi; // the pseudofermion field for this trajectory FermionField Phi; // the pseudofermion field for this trajectory
public: public:
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop, AbstractEOFAFermion<Impl>& _Rop, ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
OperatorFunction<FermionField>& S, Params& p, bool use_fc=false) : Lop(_Lop), Rop(_Rop), AbstractEOFAFermion<Impl>& _Rop,
Solver(S, false, true), Phi(_Lop.FermionGrid()), param(p), use_heatbath_forecasting(use_fc) OperatorFunction<FermionField>& HeatbathCG,
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),
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)
{ {
AlgRemez remez(param.lo, param.hi, param.precision); AlgRemez remez(param.lo, param.hi, param.precision);
@ -98,6 +114,9 @@ namespace QCD{
// We generate a Gaussian noise vector \eta, and then compute // We generate a Gaussian noise vector \eta, and then compute
// \Phi = M_{\rm EOFA}^{-1/2} * \eta // \Phi = M_{\rm EOFA}^{-1/2} * \eta
// using a rational approximation to the inverse square root // using a rational approximation to the inverse square root
//
// 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, GridParallelRNG& pRNG) virtual void refresh(const GaugeField& U, GridParallelRNG& pRNG)
{ {
Lop.ImportGauge(U); Lop.ImportGauge(U);
@ -118,7 +137,6 @@ namespace QCD{
RealD scale = std::sqrt(0.5); RealD scale = std::sqrt(0.5);
gaussian(pRNG,eta); gaussian(pRNG,eta);
eta = eta * scale; eta = eta * scale;
printf("Heatbath source vector: <\\eta|\\eta> = %1.15e\n", norm2(eta));
// \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta // \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta
RealD N(PowerNegHalf.norm); RealD N(PowerNegHalf.norm);
@ -139,11 +157,11 @@ namespace QCD{
if(use_heatbath_forecasting){ // Forecast CG guess using solutions from previous poles if(use_heatbath_forecasting){ // Forecast CG guess using solutions from previous poles
Lop.Mdag(CG_src, Forecast_src); Lop.Mdag(CG_src, Forecast_src);
CG_soln = Forecast(Lop, Forecast_src, prev_solns); CG_soln = Forecast(Lop, Forecast_src, prev_solns);
Solver(Lop, CG_src, CG_soln); SolverHB(Lop, CG_src, CG_soln);
prev_solns.push_back(CG_soln); prev_solns.push_back(CG_soln);
} else { } else {
CG_soln = zero; // Just use zero as the initial guess CG_soln = zero; // Just use zero as the initial guess
Solver(Lop, CG_src, CG_soln); SolverHB(Lop, CG_src, CG_soln);
} }
Lop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back 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]; tmp[1] = tmp[1] + ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Lop.k ) * tmp[0];
@ -166,11 +184,11 @@ namespace QCD{
if(use_heatbath_forecasting){ if(use_heatbath_forecasting){
Rop.Mdag(CG_src, Forecast_src); Rop.Mdag(CG_src, Forecast_src);
CG_soln = Forecast(Rop, Forecast_src, prev_solns); CG_soln = Forecast(Rop, Forecast_src, prev_solns);
Solver(Rop, CG_src, CG_soln); SolverHB(Rop, CG_src, CG_soln);
prev_solns.push_back(CG_soln); prev_solns.push_back(CG_soln);
} else { } else {
CG_soln = zero; CG_soln = zero;
Solver(Rop, CG_src, CG_soln); SolverHB(Rop, CG_src, CG_soln);
} }
Rop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back 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]; tmp[1] = tmp[1] - ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Rop.k ) * tmp[0];
@ -182,8 +200,47 @@ namespace QCD{
// Reset shift coefficients for energy and force evals // Reset shift coefficients for energy and force evals
Lop.RefreshShiftCoefficients(0.0); Lop.RefreshShiftCoefficients(0.0);
Rop.RefreshShiftCoefficients(-1.0); Rop.RefreshShiftCoefficients(-1.0);
// Bounds check
RealD EtaDagEta = norm2(eta);
// RealD PhiDagMPhi= norm2(eta);
}; };
void Meofa(const GaugeField& U,const FermionField &phi, FermionField & Mphi)
{
#if 0
Lop.ImportGauge(U);
Rop.ImportGauge(U);
FermionField spProj_Phi(Lop.FermionGrid());
FermionField mPhi(Lop.FermionGrid());
std::vector<FermionField> tmp(2, Lop.FermionGrid());
mPhi = phi;
// 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);
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();
// 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);
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
}
// EOFA action: see Eqn. (10) of arXiv:1706.05843 // EOFA action: see Eqn. (10) of arXiv:1706.05843
virtual RealD S(const GaugeField& U) virtual RealD S(const GaugeField& U)
{ {
@ -201,7 +258,7 @@ namespace QCD{
Lop.Omega(spProj_Phi, tmp[0], -1, 0); Lop.Omega(spProj_Phi, tmp[0], -1, 0);
G5R5(tmp[1], tmp[0]); G5R5(tmp[1], tmp[0]);
tmp[0] = zero; tmp[0] = zero;
Solver(Lop, tmp[1], tmp[0]); SolverL(Lop, tmp[1], tmp[0]);
Lop.Dtilde(tmp[0], tmp[1]); // We actually solved Cayley preconditioned system: transform back Lop.Dtilde(tmp[0], tmp[1]); // We actually solved Cayley preconditioned system: transform back
Lop.Omega(tmp[1], tmp[0], -1, 1); Lop.Omega(tmp[1], tmp[0], -1, 1);
action -= Lop.k * innerProduct(spProj_Phi, tmp[0]).real(); action -= Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
@ -212,7 +269,7 @@ namespace QCD{
Rop.Omega(spProj_Phi, tmp[0], 1, 0); Rop.Omega(spProj_Phi, tmp[0], 1, 0);
G5R5(tmp[1], tmp[0]); G5R5(tmp[1], tmp[0]);
tmp[0] = zero; tmp[0] = zero;
Solver(Rop, tmp[1], tmp[0]); SolverR(Rop, tmp[1], tmp[0]);
Rop.Dtilde(tmp[0], tmp[1]); Rop.Dtilde(tmp[0], tmp[1]);
Rop.Omega(tmp[1], tmp[0], 1, 1); Rop.Omega(tmp[1], tmp[0], 1, 1);
action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real(); action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real();
@ -245,7 +302,7 @@ namespace QCD{
Lop.Omega(spProj_Phi, Omega_spProj_Phi, -1, 0); Lop.Omega(spProj_Phi, Omega_spProj_Phi, -1, 0);
G5R5(CG_src, Omega_spProj_Phi); G5R5(CG_src, Omega_spProj_Phi);
spProj_Phi = zero; spProj_Phi = zero;
Solver(Lop, CG_src, spProj_Phi); DerivativeSolverL(Lop, CG_src, spProj_Phi);
Lop.Dtilde(spProj_Phi, Chi); Lop.Dtilde(spProj_Phi, Chi);
G5R5(g5_R5_Chi, Chi); G5R5(g5_R5_Chi, Chi);
Lop.MDeriv(force, g5_R5_Chi, Chi, DaggerNo); Lop.MDeriv(force, g5_R5_Chi, Chi, DaggerNo);
@ -257,7 +314,7 @@ namespace QCD{
Rop.Omega(spProj_Phi, Omega_spProj_Phi, 1, 0); Rop.Omega(spProj_Phi, Omega_spProj_Phi, 1, 0);
G5R5(CG_src, Omega_spProj_Phi); G5R5(CG_src, Omega_spProj_Phi);
spProj_Phi = zero; spProj_Phi = zero;
Solver(Rop, CG_src, spProj_Phi); DerivativeSolverR(Rop, CG_src, spProj_Phi);
Rop.Dtilde(spProj_Phi, Chi); Rop.Dtilde(spProj_Phi, Chi);
G5R5(g5_R5_Chi, Chi); G5R5(g5_R5_Chi, Chi);
Lop.MDeriv(force, g5_R5_Chi, Chi, DaggerNo); Lop.MDeriv(force, g5_R5_Chi, Chi, DaggerNo);

10
Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h
View File

@ -46,6 +46,7 @@ namespace Grid{
OperatorFunction<FermionField> &DerivativeSolver; OperatorFunction<FermionField> &DerivativeSolver;
OperatorFunction<FermionField> &ActionSolver; OperatorFunction<FermionField> &ActionSolver;
OperatorFunction<FermionField> &HeatbathSolver;
FermionField PhiOdd; // the pseudo fermion field for this trajectory FermionField PhiOdd; // the pseudo fermion field for this trajectory
FermionField PhiEven; // the pseudo fermion field for this trajectory FermionField PhiEven; // the pseudo fermion field for this trajectory
@ -55,10 +56,17 @@ namespace Grid{
FermionOperator<Impl> &_DenOp, FermionOperator<Impl> &_DenOp,
OperatorFunction<FermionField> & DS, OperatorFunction<FermionField> & DS,
OperatorFunction<FermionField> & AS ) : OperatorFunction<FermionField> & AS ) :
TwoFlavourEvenOddRatioPseudoFermionAction(_NumOp,_DenOp, DS,AS,AS) {};
TwoFlavourEvenOddRatioPseudoFermionAction(FermionOperator<Impl> &_NumOp,
FermionOperator<Impl> &_DenOp,
OperatorFunction<FermionField> & DS,
OperatorFunction<FermionField> & AS, OperatorFunction<FermionField> & HS) :
NumOp(_NumOp), NumOp(_NumOp),
DenOp(_DenOp), DenOp(_DenOp),
DerivativeSolver(DS), DerivativeSolver(DS),
ActionSolver(AS), ActionSolver(AS),
HeatbathSolver(HS),
PhiEven(_NumOp.FermionRedBlackGrid()), PhiEven(_NumOp.FermionRedBlackGrid()),
PhiOdd(_NumOp.FermionRedBlackGrid()) PhiOdd(_NumOp.FermionRedBlackGrid())
{ {
@ -111,7 +119,7 @@ namespace Grid{
// Odd det factors // Odd det factors
Mpc.MpcDag(etaOdd,PhiOdd); Mpc.MpcDag(etaOdd,PhiOdd);
tmp=zero; tmp=zero;
ActionSolver(Vpc,PhiOdd,tmp); HeatbathSolver(Vpc,PhiOdd,tmp);
Vpc.Mpc(tmp,PhiOdd); Vpc.Mpc(tmp,PhiOdd);
// Even det factors // Even det factors

23
Grid/qcd/hmc/integrators/Integrator.h
View File

@ -54,7 +54,7 @@ public:
template <class ReaderClass, typename std::enable_if<isReader<ReaderClass>::value, int >::type = 0 > template <class ReaderClass, typename std::enable_if<isReader<ReaderClass>::value, int >::type = 0 >
IntegratorParameters(ReaderClass & Reader){ IntegratorParameters(ReaderClass & Reader){
std::cout << "Reading integrator\n"; std::cout << GridLogMessage << "Reading integrator\n";
read(Reader, "Integrator", *this); read(Reader, "Integrator", *this);
} }
@ -132,7 +132,7 @@ class Integrator {
double end_full = usecond(); double end_full = usecond();
double time_full = (end_full - start_full) / 1e3; double time_full = (end_full - start_full) / 1e3;
double time_force = (end_force - start_force) / 1e3; double time_force = (end_force - start_force) / 1e3;
std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)" << std::endl; std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)" << std::endl;
} }
// Force from the other representations // Force from the other representations
@ -237,8 +237,7 @@ class Integrator {
for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) { for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
// get gauge field from the SmearingPolicy and // get gauge field from the SmearingPolicy and
// based on the boolean is_smeared in actionID // based on the boolean is_smeared in actionID
Field& Us = Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
as[level].actions.at(actionID)->refresh(Us, pRNG); as[level].actions.at(actionID)->refresh(Us, pRNG);
} }
@ -251,13 +250,11 @@ class Integrator {
// over the representations // over the representations
struct _S { struct _S {
template <class FieldType, class Repr> template <class FieldType, class Repr>
void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, int level, RealD& H) {
int level, RealD& H) {
for (int a = 0; a < repr_set.size(); ++a) { for (int a = 0; a < repr_set.size(); ++a) {
RealD Hterm = repr_set.at(a)->S(Rep.U); RealD Hterm = repr_set.at(a)->S(Rep.U);
std::cout << GridLogMessage << "S Level " << level << " term " << a std::cout << GridLogMessage << "S Level " << level << " term " << a << " H Hirep = " << Hterm << std::endl;
<< " H Hirep = " << Hterm << std::endl;
H += Hterm; H += Hterm;
} }
@ -267,9 +264,10 @@ class Integrator {
// Calculate action // Calculate action
RealD S(Field& U) { // here also U not used RealD S(Field& U) { // here also U not used
std::cout << GridLogIntegrator << "Integrator action\n";
RealD H = - FieldImplementation::FieldSquareNorm(P)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom RealD H = - FieldImplementation::FieldSquareNorm(P)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
std::cout << " Momentum hamiltonian "<< -H<<std::endl;
RealD Hterm; RealD Hterm;
// Actions // Actions
@ -278,9 +276,9 @@ class Integrator {
// get gauge field from the SmearingPolicy and // get gauge field from the SmearingPolicy and
// based on the boolean is_smeared in actionID // based on the boolean is_smeared in actionID
Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared); Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
Hterm = as[level].actions.at(actionID)->S(Us); Hterm = as[level].actions.at(actionID)->S(Us);
std::cout << GridLogMessage << "S Level " << level << " term " std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
<< actionID << " H = " << Hterm << std::endl;
H += Hterm; H += Hterm;
} }
as[level].apply(S_hireps, Representations, level, H); as[level].apply(S_hireps, Representations, level, H);
@ -305,8 +303,7 @@ class Integrator {
// Check the clocks all match on all levels // Check the clocks all match on all levels
for (int level = 0; level < as.size(); ++level) { for (int level = 0; level < as.size(); ++level) {
assert(fabs(t_U - t_P[level]) < 1.0e-6); // must be the same assert(fabs(t_U - t_P[level]) < 1.0e-6); // must be the same
std::cout << GridLogIntegrator << " times[" << level std::cout << GridLogIntegrator << " times[" << level << "]= " << t_P[level] << " " << t_U << std::endl;
<< "]= " << t_P[level] << " " << t_U << std::endl;
} }
// and that we indeed got to the end of the trajectory // and that we indeed got to the end of the trajectory

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

@ -231,8 +231,7 @@ class ForceGradient : public Integrator<FieldImplementation, SmearingPolicy,
Field Pfg(U._grid); Field Pfg(U._grid);
Ufg = U; Ufg = U;
Pfg = zero; Pfg = zero;
std::cout << GridLogIntegrator << "FG update " << fg_dt << " " << ep std::cout << GridLogIntegrator << "FG update " << fg_dt << " " << ep << std::endl;
<< std::endl;
// prepare_fg; no prediction/result cache for now // prepare_fg; no prediction/result cache for now
// could relax CG stopping conditions for the // could relax CG stopping conditions for the
// derivatives in the small step since the force gets multiplied by // derivatives in the small step since the force gets multiplied by
@ -271,8 +270,7 @@ class ForceGradient : public Integrator<FieldImplementation, SmearingPolicy,
this->step(U, level + 1, first_step, 0); this->step(U, level + 1, first_step, 0);
} }
this->FG_update_P(U, level, 2 * Chi / ((1.0 - 2.0 * lambda) * eps), this->FG_update_P(U, level, 2 * Chi / ((1.0 - 2.0 * lambda) * eps), (1.0 - 2.0 * lambda) * eps);
(1.0 - 2.0 * lambda) * eps);
if (level == fl) { // lowest level if (level == fl) { // lowest level
this->update_U(U, 0.5 * eps); this->update_U(U, 0.5 * eps);

277
HMC/Mobius2p1fEOFA.cc
View File

@ -30,6 +30,137 @@ directory
/* END LEGAL */ /* END LEGAL */
#include <Grid/Grid.h> #include <Grid/Grid.h>
#define MIXED_PRECISION
namespace Grid{
namespace QCD{
/*
* Need a plan for gauge field update for mixed precision in HMC (2x speed up)
* -- Store the single prec action operator.
* -- Clone the gauge field from the operator function argument.
* -- Build the mixed precision operator dynamically from the passed operator and single prec clone.
*/
template<class FermionOperatorD, class FermionOperatorF, class SchurOperatorD, class SchurOperatorF>
class MixedPrecisionConjugateGradientOperatorFunction : public OperatorFunction<typename FermionOperatorD::FermionField> {
public:
typedef typename FermionOperatorD::FermionField FieldD;
typedef typename FermionOperatorF::FermionField FieldF;
RealD Tolerance;
RealD InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
Integer MaxInnerIterations;
Integer MaxOuterIterations;
GridBase* SinglePrecGrid4; //Grid for single-precision fields
GridBase* SinglePrecGrid5; //Grid for single-precision fields
RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
FermionOperatorF &FermOpF;
FermionOperatorD &FermOpD;;
SchurOperatorF &LinOpF;
SchurOperatorD &LinOpD;
Integer TotalInnerIterations; //Number of inner CG iterations
Integer TotalOuterIterations; //Number of restarts
Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
MixedPrecisionConjugateGradientOperatorFunction(RealD tol,
Integer maxinnerit,
Integer maxouterit,
GridBase* _sp_grid4,
GridBase* _sp_grid5,
FermionOperatorF &_FermOpF,
FermionOperatorD &_FermOpD,
SchurOperatorF &_LinOpF,
SchurOperatorD &_LinOpD):
LinOpF(_LinOpF),
LinOpD(_LinOpD),
FermOpF(_FermOpF),
FermOpD(_FermOpD),
Tolerance(tol),
InnerTolerance(tol),
MaxInnerIterations(maxinnerit),
MaxOuterIterations(maxouterit),
SinglePrecGrid4(_sp_grid4),
SinglePrecGrid5(_sp_grid5),
OuterLoopNormMult(100.)
{
/* Debugging instances of objects; references are stored
std::cout << GridLogMessage << " Mixed precision CG wrapper LinOpF " <<std::hex<< &LinOpF<<std::dec <<std::endl;
std::cout << GridLogMessage << " Mixed precision CG wrapper LinOpD " <<std::hex<< &LinOpD<<std::dec <<std::endl;
std::cout << GridLogMessage << " Mixed precision CG wrapper FermOpF " <<std::hex<< &FermOpF<<std::dec <<std::endl;
std::cout << GridLogMessage << " Mixed precision CG wrapper FermOpD " <<std::hex<< &FermOpD<<std::dec <<std::endl;
*/
};
void operator()(LinearOperatorBase<FieldD> &LinOpU, const FieldD &src, FieldD &psi) {
std::cout << GridLogMessage << " Mixed precision CG wrapper operator() "<<std::endl;
SchurOperatorD * SchurOpU = static_cast<SchurOperatorD *>(&LinOpU);
// std::cout << GridLogMessage << " Mixed precision CG wrapper operator() FermOpU " <<std::hex<< &(SchurOpU->_Mat)<<std::dec <<std::endl;
// std::cout << GridLogMessage << " Mixed precision CG wrapper operator() FermOpD " <<std::hex<< &(LinOpD._Mat) <<std::dec <<std::endl;
// Assumption made in code to extract gauge field
// We could avoid storing LinopD reference alltogether ?
assert(&(SchurOpU->_Mat)==&(LinOpD._Mat));
////////////////////////////////////////////////////////////////////////////////////
// Must snarf a single precision copy of the gauge field in Linop_d argument
////////////////////////////////////////////////////////////////////////////////////
typedef typename FermionOperatorF::GaugeField GaugeFieldF;
typedef typename FermionOperatorF::GaugeLinkField GaugeLinkFieldF;
typedef typename FermionOperatorD::GaugeField GaugeFieldD;
typedef typename FermionOperatorD::GaugeLinkField GaugeLinkFieldD;
GridBase * GridPtrF = SinglePrecGrid4;
GridBase * GridPtrD = FermOpD.Umu._grid;
GaugeFieldF U_f (GridPtrF);
GaugeLinkFieldF Umu_f(GridPtrF);
// std::cout << " Dim gauge field "<<GridPtrF->Nd()<<std::endl; // 4d
// std::cout << " Dim gauge field "<<GridPtrD->Nd()<<std::endl; // 4d
////////////////////////////////////////////////////////////////////////////////////
// Moving this to a Clone method of fermion operator would allow to duplicate the
// physics parameters and decrease gauge field copies
////////////////////////////////////////////////////////////////////////////////////
GaugeLinkFieldD Umu_d(GridPtrD);
for(int mu=0;mu<Nd*2;mu++){
Umu_d = PeekIndex<LorentzIndex>(FermOpD.Umu, mu);
precisionChange(Umu_f,Umu_d);
PokeIndex<LorentzIndex>(FermOpF.Umu, Umu_f, mu);
}
pickCheckerboard(Even,FermOpF.UmuEven,FermOpF.Umu);
pickCheckerboard(Odd ,FermOpF.UmuOdd ,FermOpF.Umu);
////////////////////////////////////////////////////////////////////////////////////
// Could test to make sure that LinOpF and LinOpD agree to single prec?
////////////////////////////////////////////////////////////////////////////////////
/*
GridBase *Fgrid = psi._grid;
FieldD tmp2(Fgrid);
FieldD tmp1(Fgrid);
LinOpU.Op(src,tmp1);
LinOpD.Op(src,tmp2);
std::cout << " Double gauge field "<< norm2(FermOpD.Umu)<<std::endl;
std::cout << " Single gauge field "<< norm2(FermOpF.Umu)<<std::endl;
std::cout << " Test of operators "<<norm2(tmp1)<<std::endl;
std::cout << " Test of operators "<<norm2(tmp2)<<std::endl;
tmp1=tmp1-tmp2;
std::cout << " Test of operators diff "<<norm2(tmp1)<<std::endl;
*/
////////////////////////////////////////////////////////////////////////////////////
// Make a mixed precision conjugate gradient
////////////////////////////////////////////////////////////////////////////////////
MixedPrecisionConjugateGradient<FieldD,FieldF> MPCG(Tolerance,MaxInnerIterations,MaxOuterIterations,SinglePrecGrid5,LinOpF,LinOpD);
std::cout << GridLogMessage << "Calling mixed precision Conjugate Gradient" <<std::endl;
MPCG(src,psi);
}
};
}};
int main(int argc, char **argv) { int main(int argc, char **argv) {
using namespace Grid; using namespace Grid;
using namespace Grid::QCD; using namespace Grid::QCD;
@ -42,7 +173,11 @@ int main(int argc, char **argv) {
// Typedefs to simplify notation // Typedefs to simplify notation
typedef WilsonImplR FermionImplPolicy; typedef WilsonImplR FermionImplPolicy;
typedef MobiusFermionR FermionAction; typedef MobiusFermionR FermionAction;
typedef MobiusFermionF FermionActionF;
typedef MobiusEOFAFermionR FermionEOFAAction;
typedef MobiusEOFAFermionF FermionEOFAActionF;
typedef typename FermionAction::FermionField FermionField; typedef typename FermionAction::FermionField FermionField;
typedef typename FermionActionF::FermionField FermionFieldF;
typedef Grid::XmlReader Serialiser; typedef Grid::XmlReader Serialiser;
@ -54,12 +189,12 @@ int main(int argc, char **argv) {
MD.name = std::string("Force Gradient"); MD.name = std::string("Force Gradient");
// typedef GenericHMCRunner<MinimumNorm2> HMCWrapper; // typedef GenericHMCRunner<MinimumNorm2> HMCWrapper;
// MD.name = std::string("MinimumNorm2"); // MD.name = std::string("MinimumNorm2");
MD.MDsteps = 8; MD.MDsteps = 6;
MD.trajL = 1.0; MD.trajL = 1.0;
HMCparameters HMCparams; HMCparameters HMCparams;
HMCparams.StartTrajectory = 70; HMCparams.StartTrajectory = 590;
HMCparams.Trajectories = 200; HMCparams.Trajectories = 1000;
HMCparams.NoMetropolisUntil= 0; HMCparams.NoMetropolisUntil= 0;
// "[HotStart, ColdStart, TepidStart, CheckpointStart]\n"; // "[HotStart, ColdStart, TepidStart, CheckpointStart]\n";
// HMCparams.StartingType =std::string("ColdStart"); // HMCparams.StartingType =std::string("ColdStart");
@ -97,42 +232,53 @@ int main(int argc, char **argv) {
RealD b = 1.0; RealD b = 1.0;
RealD c = 0.0; RealD c = 0.0;
std::vector<Real> hasenbusch({ 0.1, 0.3 }); std::vector<Real> hasenbusch({ 0.1, 0.3, 0.6 });
auto GridPtr = TheHMC.Resources.GetCartesian(); auto GridPtr = TheHMC.Resources.GetCartesian();
auto GridRBPtr = TheHMC.Resources.GetRBCartesian(); auto GridRBPtr = TheHMC.Resources.GetRBCartesian();
auto FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr); auto FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr);
auto FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr); auto FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr);
std::vector<int> latt = GridDefaultLatt();
std::vector<int> mpi = GridDefaultMpi();
std::vector<int> simdF = GridDefaultSimd(Nd,vComplexF::Nsimd());
std::vector<int> simdD = GridDefaultSimd(Nd,vComplexD::Nsimd());
auto GridPtrF = SpaceTimeGrid::makeFourDimGrid(latt,simdF,mpi);
auto GridRBPtrF = SpaceTimeGrid::makeFourDimRedBlackGrid(GridPtrF);
auto FGridF = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtrF);
auto FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtrF);
IwasakiGaugeActionR GaugeAction(beta); IwasakiGaugeActionR GaugeAction(beta);
// temporarily need a gauge field // temporarily need a gauge field
LatticeGaugeField U(GridPtr); LatticeGaugeField U(GridPtr);
LatticeGaugeFieldF UF(GridPtrF);
// These lines are unecessary if BC are all periodic // These lines are unecessary if BC are all periodic
std::vector<Complex> boundary = {1,1,1,-1}; std::vector<Complex> boundary = {1,1,1,-1};
FermionAction::ImplParams Params(boundary); FermionAction::ImplParams Params(boundary);
FermionActionF::ImplParams ParamsF(boundary);
double ActionStoppingCondition = 1e-10; double ActionStoppingCondition = 1e-10;
double DerivativeStoppingCondition = 1e-7; double DerivativeStoppingCondition = 1e-6;
double MaxCGIterations = 30000; double MaxCGIterations = 30000;
ConjugateGradient<FermionField> ActionCG(ActionStoppingCondition,MaxCGIterations);
ConjugateGradient<FermionField> DerivativeCG(DerivativeStoppingCondition,MaxCGIterations);
//////////////////////////////////// ////////////////////////////////////
// Collect actions // Collect actions
//////////////////////////////////// ////////////////////////////////////
ActionLevel<HMCWrapper::Field> Level1(1); ActionLevel<HMCWrapper::Field> Level1(1);
ActionLevel<HMCWrapper::Field> Level2(4); ActionLevel<HMCWrapper::Field> Level2(8);
//////////////////////////////////// ////////////////////////////////////
// Strange action // Strange action
//////////////////////////////////// ////////////////////////////////////
typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
typedef SchurDiagMooeeOperator<FermionAction ,FermionField > LinearOperatorD;
typedef SchurDiagMooeeOperator<FermionEOFAActionF,FermionFieldF> LinearOperatorEOFAF;
typedef SchurDiagMooeeOperator<FermionEOFAAction ,FermionField > LinearOperatorEOFAD;
// FermionAction StrangeOp (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params); typedef MixedPrecisionConjugateGradientOperatorFunction<MobiusFermionD,MobiusFermionF,LinearOperatorD,LinearOperatorF> MxPCG;
// FermionAction StrangePauliVillarsOp(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass, M5,b,c, Params); typedef MixedPrecisionConjugateGradientOperatorFunction<MobiusEOFAFermionD,MobiusEOFAFermionF,LinearOperatorEOFAD,LinearOperatorEOFAF> MxPCG_EOFA;
// OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermion(StrangePauliVillarsOp,StrangeOp,OFRp);
// Level1.push_back(&StrangePseudoFermion);
// DJM: setup for EOFA ratio (Mobius) // DJM: setup for EOFA ratio (Mobius)
OneFlavourRationalParams OFRp; OneFlavourRationalParams OFRp;
@ -143,9 +289,67 @@ int main(int argc, char **argv) {
OFRp.degree = 14; OFRp.degree = 14;
OFRp.precision= 50; OFRp.precision= 50;
MobiusEOFAFermionR Strange_Op_L (U , *FGrid , *FrbGrid , *GridPtr , *GridRBPtr , strange_mass, strange_mass, pv_mass, 0.0, -1, M5, b, c); MobiusEOFAFermionR Strange_Op_L (U , *FGrid , *FrbGrid , *GridPtr , *GridRBPtr , strange_mass, strange_mass, pv_mass, 0.0, -1, M5, b, c);
MobiusEOFAFermionF Strange_Op_LF(UF, *FGridF, *FrbGridF, *GridPtrF, *GridRBPtrF, strange_mass, strange_mass, pv_mass, 0.0, -1, M5, b, c);
MobiusEOFAFermionR Strange_Op_R (U , *FGrid , *FrbGrid , *GridPtr , *GridRBPtr , pv_mass, strange_mass, pv_mass, -1.0, 1, M5, b, c); MobiusEOFAFermionR Strange_Op_R (U , *FGrid , *FrbGrid , *GridPtr , *GridRBPtr , pv_mass, strange_mass, pv_mass, -1.0, 1, M5, b, c);
ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> EOFA(Strange_Op_L, Strange_Op_R, ActionCG, OFRp, true); MobiusEOFAFermionF Strange_Op_RF(UF, *FGridF, *FrbGridF, *GridPtrF, *GridRBPtrF, pv_mass, strange_mass, pv_mass, -1.0, 1, M5, b, c);
ConjugateGradient<FermionField> ActionCG(ActionStoppingCondition,MaxCGIterations);
ConjugateGradient<FermionField> DerivativeCG(DerivativeStoppingCondition,MaxCGIterations);
#ifdef MIXED_PRECISION
const int MX_inner = 1000;
// Mixed precision EOFA
LinearOperatorEOFAD Strange_LinOp_L (Strange_Op_L);
LinearOperatorEOFAD Strange_LinOp_R (Strange_Op_R);
LinearOperatorEOFAF Strange_LinOp_LF(Strange_Op_LF);
LinearOperatorEOFAF Strange_LinOp_RF(Strange_Op_RF);
MxPCG_EOFA ActionCGL(ActionStoppingCondition,
MX_inner,
MaxCGIterations,
GridPtrF,
FrbGridF,
Strange_Op_LF,Strange_Op_L,
Strange_LinOp_LF,Strange_LinOp_L);
MxPCG_EOFA DerivativeCGL(DerivativeStoppingCondition,
MX_inner,
MaxCGIterations,
GridPtrF,
FrbGridF,
Strange_Op_LF,Strange_Op_L,
Strange_LinOp_LF,Strange_LinOp_L);
MxPCG_EOFA ActionCGR(ActionStoppingCondition,
MX_inner,
MaxCGIterations,
GridPtrF,
FrbGridF,
Strange_Op_RF,Strange_Op_R,
Strange_LinOp_RF,Strange_LinOp_R);
MxPCG_EOFA DerivativeCGR(DerivativeStoppingCondition,
MX_inner,
MaxCGIterations,
GridPtrF,
FrbGridF,
Strange_Op_RF,Strange_Op_R,
Strange_LinOp_RF,Strange_LinOp_R);
ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy>
EOFA(Strange_Op_L, Strange_Op_R,
ActionCG,
ActionCGL, ActionCGR,
DerivativeCGL, DerivativeCGR,
OFRp, true);
#else
ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy>
EOFA(Strange_Op_L, Strange_Op_R,
ActionCG, ActionCG,
DerivativeCG, DerivativeCG,
OFRp, true);
#endif
Level1.push_back(&EOFA); Level1.push_back(&EOFA);
//////////////////////////////////// ////////////////////////////////////
@ -162,15 +366,62 @@ int main(int argc, char **argv) {
} }
light_num.push_back(pv_mass); light_num.push_back(pv_mass);
//////////////////////////////////////////////////////////////
// Forced to replicate the MxPCG and DenominatorsF etc.. because
// there is no convenient way to "Clone" physics params from double op
// into single op for any operator pair.
// Same issue prevents using MxPCG in the Heatbath step
//////////////////////////////////////////////////////////////
std::vector<FermionAction *> Numerators; std::vector<FermionAction *> Numerators;
std::vector<FermionAction *> Denominators; std::vector<FermionAction *> Denominators;
std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients; std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
std::vector<MxPCG *> ActionMPCG;
std::vector<MxPCG *> MPCG;
std::vector<FermionActionF *> DenominatorsF;
std::vector<LinearOperatorD *> LinOpD;
std::vector<LinearOperatorF *> LinOpF;
for(int h=0;h<n_hasenbusch+1;h++){ for(int h=0;h<n_hasenbusch+1;h++){
std::cout << GridLogMessage << " 2f quotient Action "<< light_num[h] << " / " << light_den[h]<< std::endl; std::cout << GridLogMessage << " 2f quotient Action "<< light_num[h] << " / " << light_den[h]<< std::endl;
Numerators.push_back (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, Params)); Numerators.push_back (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, Params));
Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, Params)); Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, Params));
#ifdef MIXED_PRECISION
////////////////////////////////////////////////////////////////////////////
// Mixed precision CG for 2f force
////////////////////////////////////////////////////////////////////////////
DenominatorsF.push_back(new FermionActionF(UF,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_den[h],M5,b,c, ParamsF));
LinOpD.push_back(new LinearOperatorD(*Denominators[h]));
LinOpF.push_back(new LinearOperatorF(*DenominatorsF[h]));
MPCG.push_back(new MxPCG(DerivativeStoppingCondition,
MX_inner,
MaxCGIterations,
GridPtrF,
FrbGridF,
*DenominatorsF[h],*Denominators[h],
*LinOpF[h], *LinOpD[h]) );
ActionMPCG.push_back(new MxPCG(ActionStoppingCondition,
MX_inner,
MaxCGIterations,
GridPtrF,
FrbGridF,
*DenominatorsF[h],*Denominators[h],
*LinOpF[h], *LinOpD[h]) );
// Heatbath not mixed yet. As inverts numerators not so important as raised mass.
Quotients.push_back (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],*MPCG[h],*ActionMPCG[h],ActionCG));
#else
////////////////////////////////////////////////////////////////////////////
// Standard CG for 2f force
////////////////////////////////////////////////////////////////////////////
Quotients.push_back (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],DerivativeCG,ActionCG)); Quotients.push_back (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],DerivativeCG,ActionCG));
#endif
} }
for(int h=0;h<n_hasenbusch+1;h++){ for(int h=0;h<n_hasenbusch+1;h++){

56
HMC/README
View File

@ -1,14 +1,54 @@
* Sign off 2+1f HMC with Hasenbush and strange RHMC
- Wilson plaquette cross checked against CPS and literature GwilsonFnone
- Timesteps matched
- Use 16^3x32
******************************************************************** ********************************************************************
* From previous CPS runs: TODO:
******************************************************************** ********************************************************************
i) Got mixed precision in 2f and EOFA force and action solves.
But need mixed precision in the heatbath solve. Best for Fermop to have a "clone" method, to
reduce the number of solver and action objects. Needed ideally for the EOFA heatbath.
15% perhaps
Combine with 2x trajectory length?
ii) Rational on EOFA HB -- relax order
-- Test the approx as per David email
Resume / roll.sh
----------------------------------------------------------------
- 16^3 Currently 10 traj per hour
- EOFA use a different derivative solver from action solver
- EOFA fix Davids hack to the SchurRedBlack guessing
*** Reduce precision/tolerance in EOFA with second CG param. (10% speed up)
*** Force gradient - reduced precision solve for the gradient (4/3x speedup)
*** Need a plan for gauge field update for mixed precision in HMC (2x speed up)
-- Store the single prec action operator.
-- Clone the gauge field from the operator function argument.
-- Build the mixed precision operator dynamically from the passed operator and single prec clone.
*** Mixed precision CG into EOFA portion
*** Further reduce precision in forces to 10^-6 ?
*** Overall: a 3x or so is still possible => 500s -> 160s and 20 traj per hour on 16^3.
- Use mixed precision CG in HMC
- SchurRedBlack.h: stop use of operator function; use LinearOperator or similar instead.
- Or make an OperatorFunction for mixed precision as a wrapper
********************************************************************
* Signed off 2+1f HMC with Hasenbush and strange RHMC 16^3 x 32 DWF Ls=16 Plaquette 0.5883 ish
* Signed off 2+1f HMC with Hasenbush and strange EOFA 16^3 x 32 DWF Ls=16 Plaquette 0.5883 ish
* Wilson plaquette cross checked against CPS and literature GwilsonFnone
********************************************************************
********************************************************************
* RHMC: Timesteps & eigenranges matched from previous CPS 16^3 x 32 runs:
********************************************************************
****
Strange (m=0.04) has eigenspan Strange (m=0.04) has eigenspan
**** ****
16^3 done as 1+1+1 with separate PV's. 16^3 done as 1+1+1 with separate PV's.

13
tests/IO/Test_serialisation.cc
View File

@ -370,5 +370,18 @@ int main(int argc,char **argv)
tensorConvTest(rng, SpinMatrix); tensorConvTest(rng, SpinMatrix);
tensorConvTest(rng, SpinVector); tensorConvTest(rng, SpinVector);
{
HMCparameters HMCparams;
HMCparams.StartingType =std::string("CheckpointStart");
HMCparams.StartTrajectory =7;
HMCparams.Trajectories =1000;
HMCparams.NoMetropolisUntil=0;
HMCparams.MD.name =std::string("Force Gradient");
HMCparams.MD.MDsteps = 10;
HMCparams.MD.trajL = 1.0;
XmlWriter HMCwr("HMCparameters.xml");
write(HMCwr,"HMCparameters",HMCparams);
}
Grid_finalize(); Grid_finalize();
} }