portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-18 15:57:05 +01:00
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
parent ae5ad986e2 918e673078
commit e675c6a48c
14 changed files with 466 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
Grid/algorithms/LinearOperator.h
View File

@ -178,7 +178,7 @@ namespace Grid {
//////////////////////////////////////////////////////////
template<class Field>
class SchurOperatorBase : public LinearOperatorBase<Field> {
class SchurOperatorBase : public LinearOperatorBase<Field> {
public:
virtual RealD Mpc (const Field &in, Field &out) =0;
virtual RealD MpcDag (const Field &in, Field &out) =0;
@ -211,10 +211,9 @@ namespace Grid {
}
};
template<class Matrix,class Field>
class SchurDiagMooeeOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
class SchurDiagMooeeOperator : public SchurOperatorBase<Field> {
public:
Matrix &_Mat;
SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
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 :)
if (cp <= rsq) {
std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
IterationsToComplete = 0;
return;
}
@ -104,7 +106,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
SolverTimer.Start();
int k;
for (k = 1; k <= MaxIterations*1000; k++) {
for (k = 1; k <= MaxIterations; k++) {
c = cp;
MatrixTimer.Start();
@ -165,8 +167,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
return;
}
}
std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
<< std::endl;
std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k;

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

@ -30,8 +30,11 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
namespace Grid {
//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> {
public:
RealD Tolerance;
@ -50,7 +53,12 @@ namespace Grid {
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
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),
Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
OuterLoopNormMult(100.), guesser(NULL){ };
@ -149,6 +157,8 @@ namespace Grid {
}
};
}
#endif

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

@ -35,7 +35,11 @@ class ZeroGuesser: public LinearFunction<Field> {
public:
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>
class SourceGuesser: public LinearFunction<Field> {
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 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 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
///////////////////////////////////////////////////

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)
{
#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.
////////////////////////////////////////////////////////////////
@ -253,7 +268,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
/////////////////////////////////////////////////////////////////
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
assert(ierr==0);
#else
}
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
{
////////////////////////////////////////////////////////////////
// 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);
assert(ierr==0);
#endif
}
////////////////////////////////////////////////////////////////////////////////////////////
// SHMGET
@ -337,7 +353,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
int errsv = errno;
printf("Errno %d\n",errsv);
printf("key %d\n",key);
printf("size %lld\n",size);
printf("size %ld\n",size);
printf("flags %d\n",flags);
perror("shmget");
exit(1);

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

@ -58,13 +58,29 @@ namespace QCD{
bool use_heatbath_forecasting;
AbstractEOFAFermion<Impl>& Lop; // the basic LH 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
public:
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop, AbstractEOFAFermion<Impl>& _Rop,
OperatorFunction<FermionField>& S, Params& p, bool use_fc=false) : Lop(_Lop), Rop(_Rop),
Solver(S, false, true), Phi(_Lop.FermionGrid()), param(p), use_heatbath_forecasting(use_fc)
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
AbstractEOFAFermion<Impl>& _Rop,
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);
@ -98,6 +114,9 @@ namespace QCD{
// We generate a Gaussian noise vector \eta, and then compute
// \Phi = M_{\rm EOFA}^{-1/2} * \eta
// 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)
{
Lop.ImportGauge(U);
@ -118,7 +137,6 @@ namespace QCD{
RealD scale = std::sqrt(0.5);
gaussian(pRNG,eta);
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
RealD N(PowerNegHalf.norm);
@ -139,11 +157,11 @@ namespace QCD{
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);
Solver(Lop, CG_src, CG_soln);
SolverHB(Lop, CG_src, CG_soln);
prev_solns.push_back(CG_soln);
} else {
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
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){
Rop.Mdag(CG_src, Forecast_src);
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);
} else {
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
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
Lop.RefreshShiftCoefficients(0.0);
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
virtual RealD S(const GaugeField& U)
{
@ -201,7 +258,7 @@ namespace QCD{
Lop.Omega(spProj_Phi, tmp[0], -1, 0);
G5R5(tmp[1], tmp[0]);
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.Omega(tmp[1], tmp[0], -1, 1);
action -= Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
@ -212,7 +269,7 @@ namespace QCD{
Rop.Omega(spProj_Phi, tmp[0], 1, 0);
G5R5(tmp[1], tmp[0]);
tmp[0] = zero;
Solver(Rop, tmp[1], tmp[0]);
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();
@ -245,7 +302,7 @@ namespace QCD{
Lop.Omega(spProj_Phi, Omega_spProj_Phi, -1, 0);
G5R5(CG_src, Omega_spProj_Phi);
spProj_Phi = zero;
Solver(Lop, CG_src, spProj_Phi);
DerivativeSolverL(Lop, CG_src, spProj_Phi);
Lop.Dtilde(spProj_Phi, Chi);
G5R5(g5_R5_Chi, Chi);
Lop.MDeriv(force, g5_R5_Chi, Chi, DaggerNo);
@ -257,7 +314,7 @@ namespace QCD{
Rop.Omega(spProj_Phi, Omega_spProj_Phi, 1, 0);
G5R5(CG_src, Omega_spProj_Phi);
spProj_Phi = zero;
Solver(Rop, CG_src, spProj_Phi);
DerivativeSolverR(Rop, CG_src, spProj_Phi);
Rop.Dtilde(spProj_Phi, Chi);
G5R5(g5_R5_Chi, Chi);
Lop.MDeriv(force, g5_R5_Chi, Chi, DaggerNo);

12
Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h
View File

@ -46,6 +46,7 @@ namespace Grid{
OperatorFunction<FermionField> &DerivativeSolver;
OperatorFunction<FermionField> &ActionSolver;
OperatorFunction<FermionField> &HeatbathSolver;
FermionField PhiOdd; // the pseudo fermion field for this trajectory
FermionField PhiEven; // the pseudo fermion field for this trajectory
@ -54,11 +55,18 @@ namespace Grid{
TwoFlavourEvenOddRatioPseudoFermionAction(FermionOperator<Impl> &_NumOp,
FermionOperator<Impl> &_DenOp,
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),
DenOp(_DenOp),
DerivativeSolver(DS),
ActionSolver(AS),
HeatbathSolver(HS),
PhiEven(_NumOp.FermionRedBlackGrid()),
PhiOdd(_NumOp.FermionRedBlackGrid())
{
@ -111,7 +119,7 @@ namespace Grid{
// Odd det factors
Mpc.MpcDag(etaOdd,PhiOdd);
tmp=zero;
ActionSolver(Vpc,PhiOdd,tmp);
HeatbathSolver(Vpc,PhiOdd,tmp);
Vpc.Mpc(tmp,PhiOdd);
// 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 >
IntegratorParameters(ReaderClass & Reader){
std::cout << "Reading integrator\n";
std::cout << GridLogMessage << "Reading integrator\n";
read(Reader, "Integrator", *this);
}
@ -132,7 +132,7 @@ class Integrator {
double end_full = usecond();
double time_full = (end_full - start_full) / 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
@ -237,8 +237,7 @@ class Integrator {
for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
// get gauge field from the SmearingPolicy and
// 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);
as[level].actions.at(actionID)->refresh(Us, pRNG);
}
@ -251,13 +250,11 @@ class Integrator {
// over the representations
struct _S {
template <class FieldType, class Repr>
void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep,
int level, RealD& H) {
void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, int level, RealD& H) {
for (int a = 0; a < repr_set.size(); ++a) {
RealD Hterm = repr_set.at(a)->S(Rep.U);
std::cout << GridLogMessage << "S Level " << level << " term " << a
<< " H Hirep = " << Hterm << std::endl;
std::cout << GridLogMessage << "S Level " << level << " term " << a << " H Hirep = " << Hterm << std::endl;
H += Hterm;
}
@ -267,9 +264,10 @@ class Integrator {
// Calculate action
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
std::cout << " Momentum hamiltonian "<< -H<<std::endl;
RealD Hterm;
// Actions
@ -278,9 +276,9 @@ class Integrator {
// get gauge field from the SmearingPolicy and
// based on the boolean is_smeared in actionID
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);
std::cout << GridLogMessage << "S Level " << level << " term "
<< actionID << " H = " << Hterm << std::endl;
std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
H += Hterm;
}
as[level].apply(S_hireps, Representations, level, H);
@ -305,8 +303,7 @@ class Integrator {
// Check the clocks all match on all levels
for (int level = 0; level < as.size(); ++level) {
assert(fabs(t_U - t_P[level]) < 1.0e-6); // must be the same
std::cout << GridLogIntegrator << " times[" << level
<< "]= " << t_P[level] << " " << t_U << std::endl;
std::cout << GridLogIntegrator << " times[" << level << "]= " << t_P[level] << " " << t_U << std::endl;
}
// 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);
Ufg = U;
Pfg = zero;
std::cout << GridLogIntegrator << "FG update " << fg_dt << " " << ep
<< std::endl;
std::cout << GridLogIntegrator << "FG update " << fg_dt << " " << ep << std::endl;
// prepare_fg; no prediction/result cache for now
// could relax CG stopping conditions for the
// 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->FG_update_P(U, level, 2 * Chi / ((1.0 - 2.0 * lambda) * eps),
(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);
if (level == fl) { // lowest level
this->update_U(U, 0.5 * eps);