Merge pull request #409 from giltirn/feature/dirichlet-gparity-stage

Import round 5

Grid/GridQCDcore.h

@@ -36,6 +36,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/qcd/QCD.h>
 #include <Grid/qcd/spin/Spin.h>
+#include <Grid/qcd/gparity/Gparity.h>
 #include <Grid/qcd/utils/Utils.h>
 #include <Grid/qcd/representations/Representations.h>
 NAMESPACE_CHECK(GridQCDCore);

Grid/GridStd.h

@@ -16,6 +16,7 @@
 #include <functional>
 #include <stdio.h>
 #include <stdlib.h>
+#include <strings.h>
 #include <stdio.h>
 #include <signal.h>
 #include <ctime>

Grid/algorithms/Algorithms.h

@@ -54,6 +54,7 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
 #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>

Grid/algorithms/iterative/ConjugateGradient.h

@@ -120,6 +120,9 @@ public:
     SolverTimer.Start();
     int k;
     for (k = 1; k <= MaxIterations; k++) {
+
+      GridStopWatch IterationTimer;
+      IterationTimer.Start();
       c = cp;
 
       MatrixTimer.Start();
@@ -152,8 +155,14 @@ public:
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
-      std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+      IterationTimer.Stop();
+      if ( (k % 500) == 0 ) {
+	std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
                 << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
+      } else { 
+	std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+		  << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
+      }
 
       // Stopping condition
       if (cp <= rsq) {
@@ -170,13 +179,13 @@ public:
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogIterative << "Time breakdown "<<std::endl;
-	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
+	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -49,6 +49,7 @@ NAMESPACE_BEGIN(Grid);
     Integer TotalInnerIterations; //Number of inner CG iterations
     Integer TotalOuterIterations; //Number of restarts
     Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
+    RealD TrueResidual;
 
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
@@ -68,6 +69,7 @@ NAMESPACE_BEGIN(Grid);
     }
   
   void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::cout << GridLogMessage << "MixedPrecisionConjugateGradient: Starting mixed precision CG with outer tolerance " << Tolerance << " and inner tolerance " << InnerTolerance << std::endl;
     TotalInnerIterations = 0;
 	
     GridStopWatch TotalTimer;
@@ -97,6 +99,7 @@ NAMESPACE_BEGIN(Grid);
     FieldF sol_f(SinglePrecGrid);
     sol_f.Checkerboard() = cb;
     
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting initial inner CG with tolerance " << inner_tol << std::endl;
     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
     CG_f.ErrorOnNoConverge = false;
 
@@ -130,6 +133,7 @@ NAMESPACE_BEGIN(Grid);
 	(*guesser)(src_f, sol_f);
 
       //Inner CG
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " << outer_iter << " starting inner CG with tolerance " << inner_tol << std::endl;
       CG_f.Tolerance = inner_tol;
       InnerCGtimer.Start();
       CG_f(Linop_f, src_f, sol_f);
@@ -150,6 +154,7 @@ NAMESPACE_BEGIN(Grid);
     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
     CG_d(Linop_d, src_d_in, sol_d);
     TotalFinalStepIterations = CG_d.IterationsToComplete;
+    TrueResidual = CG_d.TrueResidual;
 
     TotalTimer.Stop();
     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -44,7 +44,7 @@ public:
 
   using OperatorFunction<Field>::operator();
 
-  RealD   Tolerance;
+  //  RealD   Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -52,7 +52,7 @@ public:
   MultiShiftFunction shifts;
   std::vector<RealD> TrueResidualShift;
 
-  ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
+  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
     MaxIterations(maxit),
     shifts(_shifts)
   { 
@@ -182,6 +182,9 @@ public:
     for(int s=0;s<nshift;s++) {
       axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
     }
+
+    std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
+    
   
   ///////////////////////////////////////
   // Timers
@@ -321,8 +324,8 @@ public:
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -0,0 +1,409 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@bnl.gov>
+
+    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 GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+    
+//For safety the a final regular CG is applied to clean up if necessary
+
+//Linop to add shift to input linop, used in cleanup CG
+namespace ConjugateGradientMultiShiftMixedPrecSupport{
+template<typename Field>
+class ShiftedLinop: public LinearOperatorBase<Field>{
+public:
+  LinearOperatorBase<Field> &linop_base;
+  RealD shift;
+
+  ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
+
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
+  
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+
+  void HermOp(const Field &in, Field &out){
+    linop_base.HermOp(in, out);
+    axpy(out, shift, in, out);
+  }    
+
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+};
+};
+
+
+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 ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq
+				       ) : 
+    MaxIterations(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GridBase *DoublePrecGrid = src_d.Grid();
+
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    assert(psi_d.size()==nshift);
+    assert(mass.size()==nshift);
+    assert(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF r_f(SinglePrecGrid);
+    FieldF p_f(SinglePrecGrid);
+    FieldF tmp_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+    FieldF src_f(SinglePrecGrid);
+    precisionChange(src_f, src_d);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      assert( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      ps_d[s] = src_d;
+    }
+    // r and p for primary
+    r_f=src_f; //residual maintained in single
+    p_f=src_f;
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+  
+    //MdagM+m[0]
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    axpy(mmp_f,mass[0],p_f,mmp_f);
+    RealD rn = norm2(p_f);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_f,b,mmp_f,r_f);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterations;k++){    
+      a = c /cp;
+
+      //Update double precision search direction by residual
+      PrecChangeTimer.Start();
+      precisionChange(r_d, r_f);
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_d[s],a,ps_d[s],r_d);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+
+      cp=c;
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f); 
+      d=real(innerProduct(p_f,mmp_f));    
+      MatrixTimer.Stop();  
+
+      AXPYTimer.Start();
+      axpy(mmp_f,mass[0],p_f,mmp_f);
+      AXPYTimer.Stop();
+      RealD rn = norm2(p_f);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+    
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update double precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
+	}
+      }
+
+      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
+      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
+      AXPYTimer.Stop();
+
+      c = c_f;
+
+      if(k % ReliableUpdateFreq == 0){
+	//Replace r with true residual
+	MatrixTimer.Start();  
+	Linop_d.HermOp(psi_d[0],mmp_d); 
+	MatrixTimer.Stop();  
+
+	AXPYTimer.Start();
+	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
+
+	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
+	AXPYTimer.Stop();
+
+	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
+	
+	PrecChangeTimer.Start();
+	precisionChange(r_f, r_d);
+	PrecChangeTimer.Stop();
+	c = c_d;
+      }
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsq[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged ){
+
+	SolverTimer.Stop();
+	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
+	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
+      
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+
+   
+    }
+    // ugly hack
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    //  assert(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -44,6 +44,7 @@ public:
 				  int, MinRes);    // Must restart
 };
 
+//This class is the input parameter class for some testing programs
 struct LocalCoherenceLanczosParams : Serializable {
 public:
   GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
@@ -145,16 +146,24 @@ public:
   LinearOperatorBase<FineField> &_Linop;
   RealD                             _coarse_relax_tol;
   std::vector<FineField>        &_subspace;
+
+  int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
+                                //As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
+                                //To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
+                                //out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
+                                //NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
   
   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
 					   OperatorFunction<FineField>   &smoother,
 					   LinearOperatorBase<FineField> &Linop,
 					   std::vector<FineField>        &subspace,
-					   RealD coarse_relax_tol=5.0e3) 
+					   RealD coarse_relax_tol=5.0e3,
+					   int largestEvalIdxForReport=-1) 
     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
-      _coarse_relax_tol(coarse_relax_tol)  
+      _coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
   {    };
 
+  //evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
   int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
     CoarseField v(B);
@@ -177,12 +186,26 @@ public:
 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
 	     <<std::endl;
 
+    if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
+      std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
+      RealD tmp_eval;
+      ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
+    }
+    
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
     return conv;
   }
-  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+
+  //This function is called at the end of the coarse grid Lanczos. It promotes the coarse eigenvector 'B' to the fine grid,
+  //applies a smoother to the result then computes the computes the *fine grid* eigenvalue (output as 'eval').
+
+  //evalMaxApprox should be the approximation of the largest eval of the fine Hermop. However when this function is called by IRL it actually passes the largest eval of the *Chebyshev* operator (as this is the max approx used for the TestConvergence above)
+  //As the largest eval of the Chebyshev is typically several orders of magnitude larger this makes the convergence test pass even when it should not.
+  //We therefore ignore evalMaxApprox here and use a value of 1.0 (note this value is already used by TestCoarse)
+  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)  
   {
+    evalMaxApprox = 1.0; //cf above
     GridBase *FineGrid = _subspace[0].Grid();    
     int checkerboard   = _subspace[0].Checkerboard();
     FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
@@ -201,13 +224,13 @@ public:
     eval   = vnum/vden;
     fv -= eval*fB;
     RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
-
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
+    
     std::cout.precision(13);
     std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
 	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv << " target " << eresid*eresid
 	     <<std::endl;
-    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
     if( (vv<eresid*eresid) ) return 1;
     return 0;
   }
@@ -285,6 +308,10 @@ public:
     evals_coarse.resize(0);
   };
 
+  //The block inner product is the inner product on the fine grid locally summed over the blocks
+  //to give a Lattice<Scalar> on the coarse grid. This function orthnormalizes the fine-grid subspace
+  //vectors under the block inner product. This step must be performed after computing the fine grid
+  //eigenvectors and before computing the coarse grid eigenvectors.    
   void Orthogonalise(void ) {
     CoarseScalar InnerProd(_CoarseGrid);
     std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
@@ -328,6 +355,8 @@ public:
     }
   }
 
+  //While this method serves to check the coarse eigenvectors, it also recomputes the eigenvalues from the smoothed reconstructed eigenvectors
+  //hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
   {
     assert(evals_fine.size() == nbasis);
@@ -376,25 +405,31 @@ public:
     evals_fine.resize(nbasis);
     subspace.resize(nbasis,_FineGrid);
   }
+
+
+  //cheby_op: Parameters of the fine grid Chebyshev polynomial used for the Lanczos acceleration
+  //cheby_smooth: Parameters of a separate Chebyshev polynomial used after the Lanczos has completed to smooth out high frequency noise in the reconstructed fine grid eigenvectors prior to computing the eigenvalue
+  //relax: Reconstructed eigenvectors (post smoothing) are naturally not as precise as true eigenvectors. This factor acts as a multiplier on the stopping condition when determining whether the results satisfy the user provided stopping condition
   void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
 		  int Nstop, int Nk, int Nm,RealD resid, 
 		  RealD MaxIt, RealD betastp, int MinRes)
   {
-    Chebyshev<FineField>                          Cheby(cheby_op);
-    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace);
-    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
+    Chebyshev<FineField>                          Cheby(cheby_op); //Chebyshev of fine operator on fine grid
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace); //Fine operator on coarse grid with intermediate fine grid conversion
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); //Chebyshev of fine operator on coarse grid with intermediate fine grid conversion
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
 
-    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1); 
 
     evals_coarse.resize(Nm);
     evec_coarse.resize(Nm,_CoarseGrid);
 
     CoarseField src(_CoarseGrid);     src=1.0; 
 
+    //Note the "tester" here is also responsible for generating the fine grid eigenvalues which are output into the "evals_coarse" array
     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
     int Nconv=0;
     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
@@ -405,6 +440,14 @@ public:
       std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl;
     }
   }
+
+  //Get the fine eigenvector 'i' by reconstruction
+  void getFineEvecEval(FineField &evec, RealD &eval, const int i) const{
+    blockPromote(evec_coarse[i],evec,subspace);  
+    eval = evals_coarse[i];
+  }
+    
+    
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/PowerMethod.h

@@ -29,6 +29,8 @@ template<class Field> class PowerMethod
       RealD vnum = real(innerProduct(src_n,tmp)); // HermOp. 
       RealD vden = norm2(src_n); 
       RealD na = vnum/vden; 
+
+      std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
       
       if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) { 
  	evalMaxApprox = na; 

Grid/communicator/Communicator_base.h

@@ -53,10 +53,11 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
+  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processor_coor;  // linear processor coordinate
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -97,8 +98,9 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                      ProcessorCount(void)    ;
+  int                ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -142,16 +144,16 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,
+			       int xmit_to_rank,int do_xmit,
 			       void *recv,
-			       int recv_from_rank,
+			       int recv_from_rank,int do_recv,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,
+				    int xmit_to_rank,int do_xmit,
 				    void *recv,
-				    int recv_from_rank,
+				    int recv_from_rank,int do_recv,
 				    int bytes,int dir);
   
   

Grid/communicator/Communicator_mpi3.cc

@@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
   // Remap using the shared memory optimising routine
   // The remap creates a comm which must be freed
   ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,12 +124,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
-
+  Coordinate shm_processors       (_ndimension,1);
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
   for(int d=0;d<parent_ndimension;d++){
     parent_processor_coor[pad+d]=parent._processor_coor[d];
     parent_processors    [pad+d]=parent._processors[d];
+    shm_processors       [pad+d]=parent._shm_processors[d];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -154,6 +155,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     ccoor[d] = parent_processor_coor[d] % processors[d];
     scoor[d] = parent_processor_coor[d] / processors[d];
     ssize[d] = parent_processors[d]     / processors[d];
+    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@@ -335,22 +337,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest,
+						     int dest, int dox,
 						     void *recv,
-						     int from,
+						     int from, int dor,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,
+							 int dest,int dox,
 							 void *recv,
-							 int from,
+							 int from,int dor,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -370,30 +372,33 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+from*32;
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-    assert(ierr==0);
-    list.push_back(rrq);
-    off_node_bytes+=bytes;
+  if ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      list.push_back(rrq);
+      off_node_bytes+=bytes;
+    }
   }
-
-  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+_processor*32;
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-    assert(ierr==0);
-    list.push_back(xrq);
-    off_node_bytes+=bytes;
-  } else {
-    // TODO : make a OMP loop on CPU, call threaded bcopy
-    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-    assert(shm!=NULL);
-    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
-    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+  
+  if (dox) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+_processor*32;
+      ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      list.push_back(xrq);
+      off_node_bytes+=bytes;
+    } else {
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+    }
   }
-
+  
   if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);
+    list.resize(0);
   }
 
   return off_node_bytes;

Grid/communicator/Communicator_none.cc

@@ -45,12 +45,14 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
+  _shm_processors = Coordinate(processors.size(),1);
   srank=0;
   SetCommunicator(communicator_world);
 }
 
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
+  _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
   _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
@@ -111,18 +113,18 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,
+						     int xmit_to_rank,int dox,
 						     void *recv,
-						     int recv_from_rank,
+						     int recv_from_rank,int dor,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,
+							 int xmit_to_rank,int dox,
 							 void *recv,
-							 int recv_from_rank,
+							 int recv_from_rank,int dor,
 							 int bytes, int dir)
 {
   return 2.0*bytes;

Grid/communicator/SharedMemory.h

@@ -93,9 +93,10 @@ public:
   // Create an optimal reordered communicator that makes MPI_Cart_create get it right
   //////////////////////////////////////////////////////////////////////////////////////
   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
   static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
   ///////////////////////////////////////////////////
   // Provide shared memory facilities off comm world

Grid/communicator/SharedMemoryMPI.cc

@@ -152,7 +152,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   //////////////////////////////////////////////////////////////////////////////
   // Look and see if it looks like an HPE 8600 based on hostname conventions
@@ -165,8 +165,8 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
 
-  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
 static inline int divides(int a,int b)
 {
@@ -221,7 +221,7 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
     dim=(dim+1) %ndimension;
   }
 }
-void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -294,7 +294,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-
+  SHM = ShmDims;
+  
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -341,7 +342,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
 }
-void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -353,6 +354,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
+  SHM=ShmDims;
+
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryNone.cc

@@ -48,9 +48,10 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   _ShmSetup=1;
 }
 
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   optimal_comm = WorldComm;
+  SHM = Coordinate(processors.size(),1);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////

Grid/lattice/Lattice.h

@@ -46,3 +46,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_unary.h>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
+#include <Grid/lattice/Lattice_crc.h>

Grid/lattice/Lattice_crc.h

@@ -0,0 +1,55 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/Lattice_crc.h
+
+    Copyright (C) 2021
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
+{
+  auto ff = localNorm2(f);
+  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
+  typedef typename vobj::tensor_reduced normtype;
+  typedef typename normtype::scalar_object scalar;
+  std::vector<scalar> sff;
+  sliceSum(ff,sff,mu);
+  for(int t=0;t<sff.size();t++){
+    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
+  }
+}
+
+template<class vobj> uint32_t crc(Lattice<vobj> & buf)
+{
+  autoView( buf_v , buf, CpuRead);
+  return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
+}
+
+#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
+
+NAMESPACE_END(Grid);
+
+

Grid/lattice/Lattice_peekpoke.h

@@ -125,6 +125,12 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+template<class vobj>
+typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
+  typename vobj::scalar_object s;
+  peekSite(s,l,site);
+  return s;
+}        
 template<class vobj,class sobj>
 void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         

Grid/lattice/Lattice_reduction.h

@@ -232,6 +232,7 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   const uint64_t sites = grid->oSites();
   
   // Might make all code paths go this way.
+#if 0
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
@@ -241,15 +242,31 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
     autoView( right_v,right, AcceleratorRead);
 
     // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, 1,{
-	auto x_l = left_v[ss];
-	auto y_l = right_v[ss];
-	inner_tmp_v[ss]=innerProductD(x_l,y_l);
+    accelerator_for( ss, sites, nsimd,{
+	auto x_l = left_v(ss);
+	auto y_l = right_v(ss);
+	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
     });
   }
+#else
+  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+    
+  {
+    autoView( left_v , left, AcceleratorRead);
+    autoView( right_v,right, AcceleratorRead);
 
+    // GPU - SIMT lane compliance...
+    accelerator_for( ss, sites, nsimd,{
+	auto x_l = left_v(ss);
+	auto y_l = right_v(ss);
+	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
+    });
+  }
+#endif
   // This is in single precision and fails some tests
-  auto anrm = sum(inner_tmp_v,sites);  
+  auto anrm = sumD(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
@@ -283,7 +300,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   conformable(x,y);
 
   typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_typeD vector_type;
+  //  typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
   GridBase *grid = x.Grid();
@@ -295,17 +312,29 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-
+#if 0
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
 
-  accelerator_for( ss, sites, 1,{
-      auto tmp = a*x_v[ss]+b*y_v[ss];
-      inner_tmp_v[ss]=innerProductD(tmp,tmp);
-      z_v[ss]=tmp;
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
   });
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
+#else
+  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
+  });
+  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
+#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }

Grid/lattice/Lattice_rng.h

@@ -424,9 +424,32 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
+#if 1
+    thread_for( lidx, _grid->lSites(), {
 
+	int gidx;
+	int o_idx;
+	int i_idx;
+	int rank;
+	Coordinate pcoor;
+	Coordinate lcoor;
+	Coordinate gcoor;
+	_grid->LocalIndexToLocalCoor(lidx,lcoor);
+	pcoor=_grid->ThisProcessorCoor();
+	_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
+	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
+
+	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
+	assert(rank == _grid->ThisRank() );
+	
+	int l_idx=generator_idx(o_idx,i_idx);
+	_generators[l_idx] = master_engine;
+	Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
+    });
+#else
     // Everybody loops over global volume.
     thread_for( gidx, _grid->_gsites, {
+
 	// Where is it?
 	int rank;
 	int o_idx;
@@ -443,6 +466,7 @@ public:
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
     });
+#endif
 #else 
     ////////////////////////////////////////////////////////////////
     // Machine and thread decomposition dependent seeding is efficient

Grid/lattice/Lattice_transfer.h

@@ -855,7 +855,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
 
 
 template<class vobj>
-void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
+void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 {
   typedef typename vobj::scalar_object sobj;
 

Grid/log/Log.cc

@@ -65,29 +65,34 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
+GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(0);
+  GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
+  GridLogMemory.Active(0); // at least the messages should be always on
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
+  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
+    if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -182,6 +182,8 @@ extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogHMC;
+extern GridLogger GridLogMemory;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/NerscIO.h

@@ -42,9 +42,11 @@ using namespace Grid;
 ////////////////////////////////////////////////////////////////////////////////
 class NerscIO : public BinaryIO { 
 public:
-
   typedef Lattice<vLorentzColourMatrixD> GaugeField;
 
+  // Enable/disable exiting if the plaquette in the header does not match the value computed (default true)
+  static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; }
+
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -203,7 +205,7 @@ public:
       std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
       exit(0);
     }
-    assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
+    if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
     assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
     assert(nersc_csum == header.checksum );
       

Grid/perfmon/PerfCount.h

@@ -72,17 +72,9 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
-#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
-#define __SSC_STOP  __SSC_MARK(0x110)
-#define __SSC_START __SSC_MARK(0x111)
-
 
 #else
 
-#define __SSC_MARK(mark) 
-#define __SSC_STOP  
-#define __SSC_START 
-
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

@@ -39,9 +39,9 @@ NAMESPACE_BEGIN(Grid)
 // C++11 time facilities better?
 inline double usecond(void) {
   struct timeval tv;
-#ifdef TIMERS_ON
+  tv.tv_sec = 0;
+  tv.tv_usec = 0;
   gettimeofday(&tv,NULL);
-#endif
   return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
 }
 

Grid/qcd/QCD.h

@@ -63,6 +63,7 @@ static constexpr int Ngp=2; // gparity index range
 #define ColourIndex  (2)
 #define SpinIndex    (1)
 #define LorentzIndex (0)
+#define GparityFlavourIndex (0)
 
 // Also should make these a named enum type
 static constexpr int DaggerNo=0;
@@ -87,6 +88,8 @@ template<typename T> struct isCoarsened {
 template <typename T> using IfCoarsened    = Invoke<std::enable_if< isCoarsened<T>::value,int> > ;
 template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ;
 
+const int GparityFlavourTensorIndex = 3; //TensorLevel counts from the bottom!
+
 // ChrisK very keen to add extra space for Gparity doubling.
 //
 // Also add domain wall index, in a way where Wilson operator 
@@ -110,8 +113,10 @@ template<typename vtype> using iHalfSpinColourVector      = iScalar<iVector<iVec
     template<typename vtype> using iSpinColourSpinColourMatrix  = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
 
 
+template<typename vtype> using iGparityFlavourVector                = iVector<iScalar<iScalar<vtype> >, Ngp>;
 template<typename vtype> using iGparitySpinColourVector       = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
 template<typename vtype> using iGparityHalfSpinColourVector   = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
+template<typename vtype> using iGparityFlavourMatrix = iMatrix<iScalar<iScalar<vtype> >, Ngp>;
 
 // Spin matrix
 typedef iSpinMatrix<Complex  >          SpinMatrix;
@@ -176,6 +181,16 @@ typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
 typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
 
+//G-parity flavour matrix
+typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
+typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
+
+typedef iGparityFlavourMatrix<vComplex> vGparityFlavourMatrix;
+typedef iGparityFlavourMatrix<vComplexF> vGparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<vComplexD> vGparityFlavourMatrixD;
+
+
 // Spin vector
 typedef iSpinVector<Complex >           SpinVector;
 typedef iSpinVector<ComplexF>           SpinVectorF;
@@ -220,6 +235,16 @@ typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
 typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
 typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
 typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
+
+//G-parity flavour vector
+typedef iGparityFlavourVector<Complex >         GparityFlavourVector;
+typedef iGparityFlavourVector<ComplexF>         GparityFlavourVectorF;
+typedef iGparityFlavourVector<ComplexD>         GparityFlavourVectorD;
+
+typedef iGparityFlavourVector<vComplex >         vGparityFlavourVector;
+typedef iGparityFlavourVector<vComplexF>         vGparityFlavourVectorF;
+typedef iGparityFlavourVector<vComplexD>         vGparityFlavourVectorD;
+
     
 // singlets
 typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type.

Grid/qcd/action/ActionBase.h

@@ -40,6 +40,29 @@ class Action
 
 public:
   bool is_smeared = false;
+  RealD deriv_norm_sum;
+  RealD deriv_max_sum;
+  int   deriv_num;
+  RealD deriv_us;
+  RealD S_us;
+  RealD refresh_us;
+  void  reset_timer(void)        {
+    deriv_us = S_us = refresh_us = 0.0;
+    deriv_num=0;
+    deriv_norm_sum = deriv_max_sum=0.0;
+  }
+  void  deriv_log(RealD nrm, RealD max) { deriv_max_sum+=max; deriv_norm_sum+=nrm; deriv_num++;}
+  RealD deriv_max_average(void)         { return deriv_max_sum/deriv_num; };
+  RealD deriv_norm_average(void)        { return deriv_norm_sum/deriv_num; };
+  RealD deriv_timer(void)        { return deriv_us; };
+  RealD S_timer(void)            { return deriv_us; };
+  RealD refresh_timer(void)      { return deriv_us; };
+  void deriv_timer_start(void)   { deriv_us-=usecond(); }
+  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
+  void refresh_timer_start(void) { refresh_us-=usecond(); }
+  void refresh_timer_stop(void)  { refresh_us+=usecond(); }
+  void S_timer_start(void)       { S_us-=usecond(); }
+  void S_timer_stop(void)        { S_us+=usecond(); }
   // Heatbath?
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action

Grid/qcd/action/ActionCore.h

@@ -37,6 +37,10 @@ NAMESPACE_CHECK(ActionSet);
 #include <Grid/qcd/action/ActionParams.h>
 NAMESPACE_CHECK(ActionParams);
 
+#include <Grid/qcd/action/filters/MomentumFilter.h>
+#include <Grid/qcd/action/filters/DirichletFilter.h>
+#include <Grid/qcd/action/filters/DDHMCFilter.h>
+
 ////////////////////////////////////////////
 // Gauge Actions
 ////////////////////////////////////////////

Grid/qcd/action/ActionParams.h

@@ -37,24 +37,33 @@ NAMESPACE_BEGIN(Grid);
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
   Coordinate twists;
-  GparityWilsonImplParams() : twists(Nd, 0) {};
+                     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  GparityWilsonImplParams() : twists(Nd, 0), dirichlet(Nd, 0) {};
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
+    dirichlet.resize(Nd,0);
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
+    dirichlet.resize(Nd,0);
   }
 };
 
 struct StaggeredImplParams {
-  StaggeredImplParams()  {};
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  StaggeredImplParams()
+  {
+    dirichlet.resize(Nd,0);
+  };
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -63,9 +72,11 @@ struct StaggeredImplParams {
 				    RealD, hi, 
 				    int,   MaxIter, 
 				    RealD, tolerance, 
+				    RealD, mdtolerance, 
 				    int,   degree, 
 				    int,   precision,
-				    int,   BoundsCheckFreq);
+				    int,   BoundsCheckFreq,
+				    RealD, BoundsCheckTol);
     
   // MaxIter and tolerance, vectors??
     
@@ -76,16 +87,62 @@ struct StaggeredImplParams {
 				RealD tol      = 1.0e-8, 
                            	int _degree    = 10,
 				int _precision = 64,
-				int _BoundsCheckFreq=20)
+				int _BoundsCheckFreq=20,
+				RealD mdtol    = 1.0e-6,
+				double _BoundsCheckTol=1e-6)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
+        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

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -71,6 +71,7 @@ public:
   RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
   RealD MassPlus(void) { return mass_plus; };
   RealD MassMinus(void) { return mass_minus; };
+
   void  SetMass(RealD _mass) { 
     mass_plus=mass_minus=_mass; 
     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs

Grid/qcd/action/fermion/FermionOperator.h

@@ -49,6 +49,8 @@ public:
 
   virtual FermionField &tmp(void) = 0;
 
+  virtual void DirichletBlock(const Coordinate & _Block) { assert(0); };
+  
   GridBase * Grid(void)   { return FermionGrid(); };   // this is all the linalg routines need to know
   GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -30,6 +30,18 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+/*
+  Policy implementation for G-parity boundary conditions
+
+  Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular
+  field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first
+  with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field
+  to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to
+  either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor.
+
+  Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
+  mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+ */
 template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal>
 class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > {
 public:
@@ -113,7 +125,7 @@ public:
     || ((distance== 1)&&(icoor[direction]==1))
     || ((distance==-1)&&(icoor[direction]==0));
 
-    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction
 
     //Apply the links
     int f_upper = permute_lane ? 1 : 0;
@@ -139,10 +151,10 @@ public:
     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code
     assert((sl == 1) || (sl == 2));
 
-    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
-
+    //If this site is an global boundary site, perform the G-parity flavor twist
+    if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) {
       if ( sl == 2 ) {
-       
+	//Only do the twist for lanes on the edge of the physical node
 	ExtractBuffer<sobj> vals(Nsimd);
 
 	extract(chi,vals);
@@ -197,6 +209,19 @@ public:
     reg = memory;
   }
 
+
+  //Poke 'poke_f0' onto flavor 0 and 'poke_f1' onto flavor 1 in direction mu of the doubled gauge field Uds
+  inline void pokeGparityDoubledGaugeField(DoubledGaugeField &Uds, const GaugeLinkField &poke_f0, const GaugeLinkField &poke_f1, const int mu){
+    autoView(poke_f0_v, poke_f0, CpuRead);
+    autoView(poke_f1_v, poke_f1, CpuRead);
+    autoView(Uds_v, Uds, CpuWrite);
+    thread_foreach(ss,poke_f0_v,{
+	Uds_v[ss](0)(mu) = poke_f0_v[ss]();
+	Uds_v[ss](1)(mu) = poke_f1_v[ss]();
+      });
+  }
+    
+
   inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
   {
     conformable(Uds.Grid(),GaugeGrid);
@@ -207,14 +232,19 @@ public:
     GaugeLinkField Uconj(GaugeGrid);
    
     Lattice<iScalar<vInteger> > coor(GaugeGrid);
-        
-    for(int mu=0;mu<Nd;mu++){
-          
-      LatticeCoordinate(coor,mu);
+
+    //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
+    //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs        
+    for(int mu=0;mu<Nd-1;mu++){
+
+      if( Params.twists[mu] ){
+	LatticeCoordinate(coor,mu);
+      }
           
       U     = PeekIndex<LorentzIndex>(Umu,mu);
       Uconj = conjugate(U);
      
+      // Implement the isospin rotation sign on the boundary between f=1 and f=0
       // This phase could come from a simple bc 1,1,-1,1 ..
       int neglink = GaugeGrid->GlobalDimensions()[mu]-1;
       if ( Params.twists[mu] ) { 
@@ -229,7 +259,7 @@ public:
 	thread_foreach(ss,U_v,{
 	    Uds_v[ss](0)(mu) = U_v[ss]();
 	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
-	  });
+	});
       }
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
@@ -260,6 +290,38 @@ public:
         });
       }
     }
+
+    { //periodic / antiperiodic temporal BCs
+      int mu = Nd-1;
+      int L   = GaugeGrid->GlobalDimensions()[mu];
+      int Lmu = L - 1;
+
+      LatticeCoordinate(coor, mu);
+
+      U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links
+      
+      GaugeLinkField *Upoke = &U;
+
+      if(Params.twists[mu]){ //antiperiodic
+	Utmp =  where(coor == Lmu, -U, U);
+	Upoke = &Utmp;
+      }
+    
+      Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links      
+      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu);
+
+      //Get the barrel-shifted field
+      Utmp = adj(Cshift(U, mu, -1)); //is a forward shift!
+      Upoke = &Utmp;
+
+      if(Params.twists[mu]){
+	U = where(coor == 0, -Utmp, Utmp);  //boundary phase
+	Upoke = &U;
+      }
+      
+      Uconj = conjugate(*Upoke);
+      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4);
+    }
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
@@ -298,28 +360,48 @@ public:
   inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
     assert(0);
   }
-  
+ 
   inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
-
-    int Ls = Btilde.Grid()->_fdimensions[0];
-        
-    GaugeLinkField tmp(mat.Grid());
-    tmp = Zero();
+    int Ls=Btilde.Grid()->_fdimensions[0];
+    
     {
-      autoView( tmp_v , tmp, CpuWrite);
-      autoView( Atilde_v , Atilde, CpuRead);
-      autoView( Btilde_v , Btilde, CpuRead);
-      thread_for(ss,tmp.Grid()->oSites(),{
-	  for (int s = 0; s < Ls; s++) {
-	    int sF = s + Ls * ss;
-	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
-	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
-	  }
-	});
+      GridBase *GaugeGrid = mat.Grid();
+      Lattice<iScalar<vInteger> > coor(GaugeGrid);
+
+      if( Params.twists[mu] ){
+	LatticeCoordinate(coor,mu);
+      }
+
+      autoView( mat_v , mat, AcceleratorWrite);
+      autoView( Btilde_v , Btilde, AcceleratorRead);
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{	  
+  	  int sU=sss;
+  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
+  	  ColorMatrixType sum;
+  	  zeroit(sum);
+  	  for(int s=0;s<Ls;s++){
+  	    int sF = s+Ls*sU;
+  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
+	      //Flavor 0
+  	      auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector
+  	      auto aa = coalescedRead(Atilde_v[sF](0)(spn) );
+  	      sum = sum + outerProduct(bb,aa);
+
+  	      //Flavor 1
+  	      bb = coalescedRead(Btilde_v[sF](1)(spn) );
+  	      aa = coalescedRead(Atilde_v[sF](1)(spn) );
+  	      sum = sum + conjugate(outerProduct(bb,aa));
+  	    }
+  	  }	    
+  	  coalescedWrite(mat_v[sU](mu)(), sum);
+  	});
     }
-    PokeIndex<LorentzIndex>(mat, tmp, mu);
-    return;
   }
+
+
+  
+
   
 };
 

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -297,7 +297,7 @@ public:
   void ZeroCountersi(void)  {  }
   void Reporti(int calls)  {  }
 
-  std::vector<int> surface_list;
+  //  Vector<int> surface_list;
 
   WilsonStencil(GridBase *grid,
 		int npoints,
@@ -307,10 +307,11 @@ public:
     : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
   { 
     ZeroCountersi();
-    surface_list.resize(0);
+    //    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
 
+  /*
   void BuildSurfaceList(int Ls,int vol4){
 
     // find same node for SHM
@@ -331,7 +332,8 @@ public:
       }
     }
   }
-
+  */
+  
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
   {

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,6 +75,10 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  int Dirichlet;
+  Coordinate Block; 
+
+  /********** Deprecate timers **********/
   void Report(void);
   void ZeroCounters(void);
   double DhopCalls;
@@ -173,7 +177,10 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-    
+
+  virtual void DirichletBlock(const Coordinate & block)
+  {
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h

@@ -66,18 +66,17 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   M5Dcalls++;
   M5Dtime-=usecond();
 
-  uint64_t nloop = grid->oSites()/Ls;
+  uint64_t nloop = grid->oSites();
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t ss= sss*Ls;
+    uint64_t s = sss%Ls;
+    uint64_t ss= sss-s;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1, tmp2;
-    for(int s=0;s<Ls;s++){
-      uint64_t idx_u = ss+((s+1)%Ls);
-      uint64_t idx_l = ss+((s+Ls-1)%Ls);
-      spProj5m(tmp1,psi(idx_u));
-      spProj5p(tmp2,psi(idx_l));
-      coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
-    }
+    uint64_t idx_u = ss+((s+1)%Ls);
+    uint64_t idx_l = ss+((s+Ls-1)%Ls);
+    spProj5m(tmp1,psi(idx_u));
+    spProj5p(tmp2,psi(idx_l));
+    coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
   });
   M5Dtime+=usecond();
 }
@@ -108,18 +107,17 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   M5Dcalls++;
   M5Dtime-=usecond();
 
-  uint64_t nloop = grid->oSites()/Ls;
+  uint64_t nloop = grid->oSites();
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t ss=sss*Ls;
+    uint64_t s = sss%Ls;
+    uint64_t ss= sss-s;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1,tmp2;
-    for(int s=0;s<Ls;s++){
-      uint64_t idx_u = ss+((s+1)%Ls);
-      uint64_t idx_l = ss+((s+Ls-1)%Ls);
-      spProj5p(tmp1,psi(idx_u));
-      spProj5m(tmp2,psi(idx_l));
-      coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
-    }
+    uint64_t idx_u = ss+((s+1)%Ls);
+    uint64_t idx_l = ss+((s+Ls-1)%Ls);
+    spProj5p(tmp1,psi(idx_u));
+    spProj5m(tmp2,psi(idx_l));
+    coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
   });
   M5Dtime+=usecond();
 }

Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h

@@ -60,7 +60,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   UmuOdd (_FourDimRedBlackGrid),
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid),
-  _tmp(&FiveDimRedBlackGrid)
+  _tmp(&FiveDimRedBlackGrid),
+  Dirichlet(0)
 {
   // some assertions
   assert(FiveDimGrid._ndimension==5);
@@ -91,6 +92,19 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
     assert(FourDimRedBlackGrid._simd_layout[d]  ==FourDimGrid._simd_layout[d]);
   }
 
+  if ( p.dirichlet.size() == Nd+1) {
+    Coordinate block = p.dirichlet;
+    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
+      Dirichlet = 1;
+      Block = block;
+    }
+  } else {
+    Coordinate block(Nd+1,0);
+    Block = block;
+  }
+
+  ZeroCounters();
+
   if (Impl::LsVectorised) { 
 
     int nsimd = Simd::Nsimd();
@@ -218,6 +232,14 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
+  if ( Dirichlet ) {
+    std::cout << GridLogMessage << " Dirichlet BCs 5d " <<Block<<std::endl;
+    Coordinate GaugeBlock(Nd);
+    for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
+    std::cout << GridLogMessage << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
+    DirichletFilter<GaugeField> Filter(GaugeBlock);
+    Filter.applyFilter(HUmu);
+  }
   Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
   pickCheckerboard(Even,UmuEven,Umu);
   pickCheckerboard(Odd ,UmuOdd,Umu);

Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h

@@ -440,6 +440,17 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
+#define KERNEL_CALL_EXT(A)						\
+  const uint64_t    NN = Nsite*Ls;					\
+  const uint64_t    sz = st.surface_list.size();			\
+  auto ptr = &st.surface_list[0];					\
+  accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
+      int sF = ptr[ss];							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
+    });									\
+  accelerator_barrier();
+
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
     int sU = ss;							\

Grid/qcd/action/filters/DDHMCFilter.h

@@ -0,0 +1,102 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/DirichletFilter.h
+
+Copyright (C) 2015
+
+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 */
+//--------------------------------------------------------------------
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+////////////////////////////////////////////////////
+// DDHMC filter with sub-block size B[mu]
+////////////////////////////////////////////////////
+
+template<typename GaugeField>
+struct DDHMCFilter: public MomentumFilterBase<GaugeField>
+{
+  Coordinate Block;
+  int Width;
+  
+  DDHMCFilter(const Coordinate &_Block,int _Width=2): Block(_Block) { Width=_Width; }
+
+  void applyFilter(GaugeField &U) const override
+  {
+    GridBase *grid = U.Grid();
+    Coordinate Global=grid->GlobalDimensions();
+    GaugeField zzz(grid); zzz = Zero();
+    LatticeInteger coor(grid); 
+    
+    auto zzz_mu = PeekIndex<LorentzIndex>(zzz,0);
+    ////////////////////////////////////////////////////
+    // Zero BDY layers
+    ////////////////////////////////////////////////////
+    std::cout<<GridLogMessage<<" DDHMC Force Filter Block "<<Block<<" width " <<Width<<std::endl;
+    for(int mu=0;mu<Nd;mu++) {
+
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+	LatticeCoordinate(coor,mu);
+
+	////////////////////////////////
+	// OmegaBar - zero all links contained in slice B-1,0 and
+	// mu links connecting to Omega
+	////////////////////////////////
+	if ( Width==1) { 
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-2),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+	if ( Width==2) { 
+	  U    = where(mod(coor,B1)==Integer(B1-2),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-3),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+	if ( Width==3) { 
+	  U    = where(mod(coor,B1)==Integer(B1-3),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-2),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(2)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-4),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+      }
+
+    }
+   
+  }
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/filters/DirichletFilter.h

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/DirichletFilter.h
+
+Copyright (C) 2015
+
+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 */
+//--------------------------------------------------------------------
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<typename MomentaField>
+struct DirichletFilter: public MomentumFilterBase<MomentaField>
+{
+  typedef typename MomentaField::vector_type vector_type; //SIMD-vectorized complex type
+  typedef typename MomentaField::scalar_type scalar_type; //scalar complex type
+
+  typedef iScalar<iScalar<iScalar<vector_type> > >            ScalarType; //complex phase for each site
+  
+  Coordinate Block;
+  
+  DirichletFilter(const Coordinate &_Block): Block(_Block){}
+
+  void applyFilter(MomentaField &P) const override
+  {
+    GridBase *grid = P.Grid();
+    typedef decltype(PeekIndex<LorentzIndex>(P, 0)) LatCM;
+    ////////////////////////////////////////////////////
+    // Zero strictly links crossing between domains
+    ////////////////////////////////////////////////////
+    LatticeInteger coor(grid); 
+    LatCM zz(grid); zz = Zero();
+    for(int mu=0;mu<Nd;mu++) {
+      if ( (Block[mu]) && (Block[mu] <= grid->GlobalDimensions()[mu] ) ) {
+	// If costly could provide Grid earlier and precompute masks
+	std::cout << GridLogMessage << " Dirichlet in mu="<<mu<<std::endl;
+	LatticeCoordinate(coor,mu);
+	auto P_mu = PeekIndex<LorentzIndex>(P, mu);
+	P_mu = where(mod(coor,Block[mu])==Integer(Block[mu]-1),zz,P_mu);
+	PokeIndex<LorentzIndex>(P, P_mu, mu);
+      }
+    }
+  }
+};
+
+
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/filters/MomentumFilter.h

@@ -37,7 +37,7 @@ NAMESPACE_BEGIN(Grid);
 
 template<typename MomentaField>
 struct MomentumFilterBase{
-  virtual void applyFilter(MomentaField &P) const;
+  virtual void applyFilter(MomentaField &P) const = 0;
 };
 
 //Do nothing

Grid/qcd/action/gauge/GaugeImplementations.h

@@ -69,6 +69,11 @@ public:
     return PeriodicBC::ShiftStaple(Link,mu);
   }
 
+  //Same as Cshift for periodic BCs
+  static inline GaugeLinkField CshiftLink(const GaugeLinkField &Link, int mu, int shift){
+    return PeriodicBC::CshiftLink(Link,mu,shift);
+  }
+
   static inline bool isPeriodicGaugeField(void) { return true; }
 };
 
@@ -110,6 +115,11 @@ public:
       return PeriodicBC::CovShiftBackward(Link, mu, field);
   }
 
+  //If mu is a conjugate BC direction
+  //Out(x) = U^dag_\mu(x-mu)  | x_\mu != 0
+  //       = U^T_\mu(L-1)  | x_\mu == 0
+  //else
+  //Out(x) = U^dag_\mu(x-mu mod L)
   static inline GaugeLinkField
   CovShiftIdentityBackward(const GaugeLinkField &Link, int mu)
   {
@@ -129,6 +139,13 @@ public:
       return PeriodicBC::CovShiftIdentityForward(Link,mu);
   }
 
+
+  //If mu is a conjugate BC direction
+  //Out(x) = S_\mu(x+mu)  | x_\mu != L-1
+  //       = S*_\mu(x+mu)  | x_\mu == L-1
+  //else
+  //Out(x) = S_\mu(x+mu mod L)
+  //Note: While this is used for Staples it is also applicable for shifting gauge links or gauge transformation matrices
   static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu)
   {
     assert(_conjDirs.size() == Nd);
@@ -138,6 +155,27 @@ public:
       return PeriodicBC::ShiftStaple(Link,mu);
   }
 
+  //Boundary-aware C-shift of gauge links / gauge transformation matrices
+  //For conjugate BC direction
+  //shift = 1
+  //Out(x) = U_\mu(x+\hat\mu)  | x_\mu != L-1
+  //       = U*_\mu(0)  | x_\mu == L-1
+  //shift = -1
+  //Out(x) = U_\mu(x-mu)  | x_\mu != 0
+  //       = U*_\mu(L-1)  | x_\mu == 0
+  //else
+  //shift = 1
+  //Out(x) = U_\mu(x+\hat\mu mod L)
+  //shift = -1
+  //Out(x) = U_\mu(x-\hat\mu mod L)
+  static inline GaugeLinkField CshiftLink(const GaugeLinkField &Link, int mu, int shift){
+    assert(_conjDirs.size() == Nd);
+    if(_conjDirs[mu]) 
+      return ConjugateBC::CshiftLink(Link,mu,shift);
+    else     
+      return PeriodicBC::CshiftLink(Link,mu,shift);
+  }
+
   static inline void       setDirections(std::vector<int> &conjDirs) { _conjDirs=conjDirs; }
   static inline std::vector<int> getDirections(void) { return _conjDirs; }
   static inline bool isPeriodicGaugeField(void) { return false; }

Grid/qcd/action/pseudofermion/Bounds.h

@@ -13,6 +13,31 @@ NAMESPACE_BEGIN(Grid);
       std::cout << GridLogMessage << "Pseudofermion action lamda_max "<<lambda_max<<"( bound "<<hi<<")"<<std::endl;
       assert( (lambda_max < hi) && " High Bounds Check on operator failed" );
     }
+
+     template<class Field> void ChebyBoundsCheck(LinearOperatorBase<Field> &HermOp,
+						 Field &GaussNoise,
+						 RealD lo,RealD hi) 
+    {
+      int orderfilter = 1000;
+      Chebyshev<Field> Cheb(lo,hi,orderfilter);
+
+      GridBase *FermionGrid = GaussNoise.Grid();
+
+      Field X(FermionGrid);
+      Field Z(FermionGrid);
+
+      X=GaussNoise;
+      RealD Nx = norm2(X);
+      Cheb(HermOp,X,Z);
+      RealD Nz = norm2(Z);
+
+      std::cout << "************************* "<<std::endl;
+      std::cout << " noise                    = "<<Nx<<std::endl;
+      std::cout << " Cheb x noise             = "<<Nz<<std::endl;
+      std::cout << " Ratio                    = "<<Nz/Nx<<std::endl;
+      std::cout << "************************* "<<std::endl;
+      assert( ((Nz/Nx)<1.0) && " ChebyBoundsCheck ");
+    }
       
     template<class Field> void InverseSqrtBoundsCheck(int MaxIter,double tol,
 						       LinearOperatorBase<Field> &HermOp,
@@ -40,13 +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);
 

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion.h

@@ -0,0 +1,163 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundaryBoson.h
+
+    Copyright (C) 2021
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class ImplD,class ImplF>
+class DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion : public Action<typename ImplD::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(ImplD);
+
+private:
+  SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
+  RealD InnerStoppingCondition;
+  RealD ActionStoppingCondition;
+  RealD DerivativeStoppingCondition;
+  FermionField Phi; // the pseudo fermion field for this trajectory
+public:
+  DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_NumOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
+    : NumOp(_NumOp), 
+      DerivativeStoppingCondition(_DerivativeTol),
+      ActionStoppingCondition(_ActionTol),
+      InnerStoppingCondition(_InnerTol),
+      Phi(_NumOp.FermionGrid()) {};
+
+  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion";}
+
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    return sstream.str();
+  }  
+  
+  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
+  {
+    // P(phi) = e^{- phi^dag P^dag P phi}
+    //
+    // NumOp == P
+    //
+    // Take phi = P^{-1} eta  ; eta = P Phi
+    //
+    // 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) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    NumOp.tolinner=InnerStoppingCondition;
+    NumOp.tol=ActionStoppingCondition;
+    NumOp.ImportGauge(U);
+
+    FermionField eta(NumOp.FermionGrid());
+
+    gaussian(pRNG,eta);    eta=eta*scale;
+    
+    NumOp.ProjectBoundaryBar(eta);
+    //DumpSliceNorm("eta",eta);
+    NumOp.RInv(eta,Phi);
+
+    //DumpSliceNorm("Phi",Phi);
+
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag Pdag P phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    NumOp.tolinner=InnerStoppingCondition;
+    NumOp.tol=ActionStoppingCondition;
+    NumOp.ImportGauge(U);
+
+    FermionField Y(NumOp.FermionGrid());
+
+    NumOp.R(Phi,Y);
+
+    RealD action = norm2(Y);
+
+    return action;
+  };
+
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
+  {
+    NumOp.tolinner=InnerStoppingCondition;
+    NumOp.tol=DerivativeStoppingCondition;
+    NumOp.ImportGauge(U);
+
+    GridBase *fgrid = NumOp.FermionGrid();
+    GridBase *ugrid = NumOp.GaugeGrid();
+
+    FermionField  X(fgrid);
+    FermionField  Y(fgrid);
+    FermionField  tmp(fgrid);
+
+    GaugeField   force(ugrid);	
+
+    FermionField DobiDdbPhi(fgrid);      // Vector A in my notes
+    FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
+    FermionField DoidP_Phi(fgrid);    // Vector E in my notes
+    FermionField DobidDddDoidP_Phi(fgrid);    // Vector F in my notes
+    
+    FermionField P_Phi(fgrid);
+    
+    // P term
+    NumOp.dBoundaryBar(Phi,tmp);
+    NumOp.dOmegaBarInv(tmp,DobiDdbPhi);        // Vector A
+    NumOp.dBoundary(DobiDdbPhi,tmp);
+    NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi);      // Vector B
+    P_Phi  = Phi - DoiDdDobiDdbPhi;
+    NumOp.ProjectBoundaryBar(P_Phi);
+    
+    // P^dag P term
+    NumOp.dOmegaDagInv(P_Phi,DoidP_Phi); // Vector E
+    NumOp.dBoundaryDag(DoidP_Phi,tmp);
+    NumOp.dOmegaBarDagInv(tmp,DobidDddDoidP_Phi);   // Vector F
+    NumOp.dBoundaryBarDag(DobidDddDoidP_Phi,tmp);
+
+    X = DobiDdbPhi;
+    Y = DobidDddDoidP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    X = DoiDdDobiDdbPhi;
+    Y = DoidP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    dSdU *= -1.0;
+
+  };
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourPseudoFermion.h

@@ -0,0 +1,158 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.h
+
+    Copyright (C) 2021
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class ImplD,class ImplF>
+class DomainDecomposedBoundaryTwoFlavourPseudoFermion : public Action<typename ImplD::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(ImplD);
+
+private:
+  SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
+  RealD ActionStoppingCondition;
+  RealD DerivativeStoppingCondition;
+  RealD InnerStoppingCondition;
+
+  FermionField Phi; // the pseudo fermion field for this trajectory
+
+  RealD refresh_action;
+public:
+  DomainDecomposedBoundaryTwoFlavourPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_DenOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol = 1.0e-6 )
+    : DenOp(_DenOp),
+      DerivativeStoppingCondition(_DerivativeTol),
+      ActionStoppingCondition(_ActionTol),
+      InnerStoppingCondition(_InnerTol),
+      Phi(_DenOp.FermionGrid()) {};
+      
+  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourPseudoFermion";}
+
+ 
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    return sstream.str();
+  }  
+  
+  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
+  {
+    // P(phi) = e^{- phi^dag Rdag^-1 R^-1 phi}
+    //
+    // DenOp == R
+    //
+    // Take phi = R eta  ; eta = R^-1 Phi
+    //
+    // 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) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol     =ActionStoppingCondition;
+    DenOp.ImportGauge(U);
+
+    FermionField eta(DenOp.FermionGrid());
+
+    gaussian(pRNG,eta);    eta=eta*scale;
+    
+    DenOp.ProjectBoundaryBar(eta);
+    DenOp.R(eta,Phi);
+    //DumpSliceNorm("Phi",Phi);
+    refresh_action = norm2(eta);
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag Rdag^-1 R^-1 phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol=ActionStoppingCondition;
+    DenOp.ImportGauge(U);
+
+    FermionField X(DenOp.FermionGrid());
+
+    DenOp.RInv(Phi,X);
+
+    RealD action = norm2(X);
+
+    return action;
+  };
+
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
+  {
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol=DerivativeStoppingCondition;
+    DenOp.ImportGauge(U);
+
+    GridBase *fgrid = DenOp.FermionGrid();
+    GridBase *ugrid = DenOp.GaugeGrid();
+
+    FermionField  X(fgrid);
+    FermionField  Y(fgrid);
+    FermionField  tmp(fgrid);
+
+    GaugeField   force(ugrid);	
+
+    FermionField DiDdb_Phi(fgrid);      // Vector C in my notes
+    FermionField DidRinv_Phi(fgrid);    // Vector D in my notes
+    FermionField Rinv_Phi(fgrid);
+
+//   FermionField RinvDagRinv_Phi(fgrid);
+//   FermionField DdbdDidRinv_Phi(fgrid);
+
+    // R^-1 term
+    DenOp.dBoundaryBar(Phi,tmp);
+    DenOp.Dinverse(tmp,DiDdb_Phi);            // Vector C
+    Rinv_Phi = Phi - DiDdb_Phi;
+    DenOp.ProjectBoundaryBar(Rinv_Phi); 
+ 
+    // R^-dagger R^-1 term
+    DenOp.DinverseDag(Rinv_Phi,DidRinv_Phi); // Vector D
+/*
+    DenOp.dBoundaryBarDag(DidRinv_Phi,DdbdDidRinv_Phi);
+    RinvDagRinv_Phi = Rinv_Phi - DdbdDidRinv_Phi;
+    DenOp.ProjectBoundaryBar(RinvDagRinv_Phi);
+*/
+    X = DiDdb_Phi;
+    Y = DidRinv_Phi;
+    DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=force;
+    DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+    DumpSliceNorm("force",dSdU);
+    dSdU *= -1.0;
+  };
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion.h

@@ -0,0 +1,237 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.h
+
+    Copyright (C) 2021
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class ImplD,class ImplF>
+class DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion : public Action<typename ImplD::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(ImplD);
+
+private:
+  SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
+  SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
+
+  RealD InnerStoppingCondition;
+  RealD ActionStoppingCondition;
+  RealD DerivativeStoppingCondition;
+  
+  FermionField Phi; // the pseudo fermion field for this trajectory
+
+public:
+  DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_NumOp, 
+						       SchurFactoredFermionOperator<ImplD,ImplF>  &_DenOp,
+						       RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
+    : NumOp(_NumOp), DenOp(_DenOp),
+      Phi(_NumOp.PeriodicFermOpD.FermionGrid()),
+      InnerStoppingCondition(_InnerTol),
+      DerivativeStoppingCondition(_DerivativeTol),
+      ActionStoppingCondition(_ActionTol)
+  {};
+      
+  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion";}
+ 
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    return sstream.str();
+  }  
+  
+  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
+  {
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField eta(NumOp.PeriodicFermOpD.FermionGrid());
+    FermionField tmp(NumOp.PeriodicFermOpD.FermionGrid());
+
+    // P(phi) = e^{- phi^dag P^dag Rdag^-1 R^-1 P phi}
+    //
+    // NumOp == P
+    // DenOp == R
+    //
+    // Take phi = P^{-1} R eta  ; eta = R^-1 P Phi
+    //
+    // 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) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    gaussian(pRNG,eta);    eta=eta*scale;
+    
+    NumOp.ProjectBoundaryBar(eta);
+    NumOp.tolinner=InnerStoppingCondition;
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol = ActionStoppingCondition;
+    NumOp.tol = ActionStoppingCondition;
+    DenOp.R(eta,tmp);
+    NumOp.RInv(tmp,Phi);
+    DumpSliceNorm("Phi",Phi);
+
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag Pdag Rdag^-1 R^-1 P phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField X(NumOp.PeriodicFermOpD.FermionGrid());
+    FermionField Y(NumOp.PeriodicFermOpD.FermionGrid());
+
+    NumOp.tolinner=InnerStoppingCondition;
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol = ActionStoppingCondition;
+    NumOp.tol = ActionStoppingCondition;
+    NumOp.R(Phi,Y);
+    DenOp.RInv(Y,X);
+
+    RealD action = norm2(X);
+    //    std::cout << " DD boundary action is " <<action<<std::endl;
+
+    return action;
+  };
+
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
+  {
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    GridBase *fgrid = NumOp.PeriodicFermOpD.FermionGrid();
+    GridBase *ugrid = NumOp.PeriodicFermOpD.GaugeGrid();
+
+    FermionField  X(fgrid);
+    FermionField  Y(fgrid);
+    FermionField  tmp(fgrid);
+
+    GaugeField   force(ugrid);	
+
+    FermionField DobiDdbPhi(fgrid);      // Vector A in my notes
+    FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
+    FermionField DiDdbP_Phi(fgrid);      // Vector C in my notes
+    FermionField DidRinvP_Phi(fgrid);    // Vector D in my notes
+    FermionField DdbdDidRinvP_Phi(fgrid);
+    FermionField DoidRinvDagRinvP_Phi(fgrid);    // Vector E in my notes
+    FermionField DobidDddDoidRinvDagRinvP_Phi(fgrid);    // Vector F in my notes
+    
+    FermionField P_Phi(fgrid);
+    FermionField RinvP_Phi(fgrid);
+    FermionField RinvDagRinvP_Phi(fgrid);
+    FermionField PdagRinvDagRinvP_Phi(fgrid);
+
+    //    RealD action = S(U);
+    NumOp.tolinner=InnerStoppingCondition;
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol = DerivativeStoppingCondition;
+    NumOp.tol = DerivativeStoppingCondition;
+    
+    // P term
+    NumOp.dBoundaryBar(Phi,tmp);
+    NumOp.dOmegaBarInv(tmp,DobiDdbPhi);        // Vector A
+    NumOp.dBoundary(DobiDdbPhi,tmp);
+    NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi);      // Vector B
+    P_Phi  = Phi - DoiDdDobiDdbPhi;
+    NumOp.ProjectBoundaryBar(P_Phi);
+
+    // R^-1 P term
+    DenOp.dBoundaryBar(P_Phi,tmp);
+    DenOp.Dinverse(tmp,DiDdbP_Phi);            // Vector C
+    RinvP_Phi = P_Phi - DiDdbP_Phi;
+    DenOp.ProjectBoundaryBar(RinvP_Phi); // Correct to here
+
+ 
+    // R^-dagger R^-1 P term
+    DenOp.DinverseDag(RinvP_Phi,DidRinvP_Phi); // Vector D
+    DenOp.dBoundaryBarDag(DidRinvP_Phi,DdbdDidRinvP_Phi);
+    RinvDagRinvP_Phi = RinvP_Phi - DdbdDidRinvP_Phi;
+    DenOp.ProjectBoundaryBar(RinvDagRinvP_Phi);
+
+    
+    // P^dag R^-dagger R^-1 P term
+    NumOp.dOmegaDagInv(RinvDagRinvP_Phi,DoidRinvDagRinvP_Phi); // Vector E
+    NumOp.dBoundaryDag(DoidRinvDagRinvP_Phi,tmp);
+    NumOp.dOmegaBarDagInv(tmp,DobidDddDoidRinvDagRinvP_Phi);   // Vector F
+    NumOp.dBoundaryBarDag(DobidDddDoidRinvDagRinvP_Phi,tmp);
+    PdagRinvDagRinvP_Phi = RinvDagRinvP_Phi- tmp;
+    NumOp.ProjectBoundaryBar(PdagRinvDagRinvP_Phi);
+
+    /*
+    std::cout << "S eval  "<< action << std::endl;
+    std::cout << "S - IP1 "<< innerProduct(Phi,PdagRinvDagRinvP_Phi) << std::endl;
+    std::cout << "S - IP2 "<< norm2(RinvP_Phi) << std::endl;
+
+    NumOp.R(Phi,tmp);
+    tmp = tmp - P_Phi;
+    std::cout << "diff1 "<<norm2(tmp) <<std::endl;
+    
+    
+    DenOp.RInv(P_Phi,tmp);
+    tmp = tmp - RinvP_Phi;
+    std::cout << "diff2 "<<norm2(tmp) <<std::endl;
+
+    DenOp.RDagInv(RinvP_Phi,tmp);
+    tmp  = tmp - RinvDagRinvP_Phi;
+    std::cout << "diff3 "<<norm2(tmp) <<std::endl;
+
+    DenOp.RDag(RinvDagRinvP_Phi,tmp);
+    tmp  = tmp - PdagRinvDagRinvP_Phi;
+    std::cout << "diff4 "<<norm2(tmp) <<std::endl;
+    */
+    
+    dSdU=Zero();
+
+    X = DobiDdbPhi;
+    Y = DobidDddDoidRinvDagRinvP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    X = DoiDdDobiDdbPhi;
+    Y = DoidRinvDagRinvP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    X = DiDdbP_Phi;
+    Y = DidRinvP_Phi;
+    DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    dSdU *= -1.0;
+
+  };
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h

@@ -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

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h

@@ -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

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h

@@ -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

Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h

@@ -40,249 +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
+      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()),
-      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(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}
-
-      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);
-
-	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 = PhiOdd;
-	  HighBoundCheck(MdagM,gauss,param.hi);
-	  InverseSqrtBoundsCheck(param.MaxIter,param.tolerance*100,MdagM,gauss,PowerNegHalf);
-	}
-
-	//  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);

Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h

@@ -49,10 +49,12 @@ NAMESPACE_BEGIN(Grid);
       Params param;
 
       MultiShiftFunction PowerHalf   ;
-      MultiShiftFunction PowerNegHalf;
       MultiShiftFunction PowerQuarter;
+      MultiShiftFunction PowerNegHalf;
       MultiShiftFunction PowerNegQuarter;
 
+      MultiShiftFunction MDPowerQuarter;
+      MultiShiftFunction MDPowerNegHalf;
     private:
      
       FermionOperator<Impl> & NumOp;// the basic operator
@@ -73,11 +75,13 @@ NAMESPACE_BEGIN(Grid);
 	remez.generateApprox(param.degree,1,2);
 	PowerHalf.Init(remez,param.tolerance,false);
 	PowerNegHalf.Init(remez,param.tolerance,true);
+	MDPowerNegHalf.Init(remez,param.mdtolerance,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);
+   	MDPowerQuarter.Init(remez,param.mdtolerance,false);
 	PowerNegQuarter.Init(remez,param.tolerance,true);
       };
 
@@ -204,8 +208,8 @@ NAMESPACE_BEGIN(Grid);
 
       virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
 
-	const int n_f  = PowerNegHalf.poles.size();
-	const int n_pv = PowerQuarter.poles.size();
+	const int n_f  = MDPowerNegHalf.poles.size();
+	const int n_pv = MDPowerQuarter.poles.size();
 
 	std::vector<FermionField> MpvPhi_k     (n_pv,NumOp.FermionGrid());
 	std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionGrid());
@@ -224,8 +228,8 @@ NAMESPACE_BEGIN(Grid);
 	MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagM(DenOp);
 	MdagMLinearOperator<FermionOperator<Impl> ,FermionField> VdagV(NumOp);
 
-	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
-	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
+	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,MDPowerQuarter);
+	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,MDPowerNegHalf);
 
 	msCG_V(VdagV,Phi,MpvPhi_k,MpvPhi);
 	msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
@@ -244,7 +248,7 @@ NAMESPACE_BEGIN(Grid);
 
 	//(1)
 	for(int k=0;k<n_f;k++){
-	  ak = PowerNegHalf.residues[k];
+	  ak = MDPowerNegHalf.residues[k];
 	  DenOp.M(MfMpvPhi_k[k],Y);
 	  DenOp.MDeriv(tmp , MfMpvPhi_k[k], Y,DaggerYes );  dSdU=dSdU+ak*tmp;
 	  DenOp.MDeriv(tmp , Y, MfMpvPhi_k[k], DaggerNo );  dSdU=dSdU+ak*tmp;
@@ -254,7 +258,7 @@ NAMESPACE_BEGIN(Grid);
 	//(3)
 	for(int k=0;k<n_pv;k++){
 
-          ak = PowerQuarter.residues[k];
+          ak = MDPowerQuarter.residues[k];
 	  
 	  NumOp.M(MpvPhi_k[k],Y);
 	  NumOp.MDeriv(tmp,MpvMfMpvPhi_k[k],Y,DaggerYes); dSdU=dSdU+ak*tmp;

Grid/qcd/action/pseudofermion/PseudoFermion.h

@@ -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>
 

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -75,24 +75,22 @@ NAMESPACE_BEGIN(Grid);
           conformable(_NumOp.GaugeRedBlackGrid(), _DenOp.GaugeRedBlackGrid());
         };
 
-      virtual std::string action_name(){return "TwoFlavourEvenOddRatioPseudoFermionAction";}
+      virtual std::string action_name(){
+	std::stringstream sstream;
+	sstream<<"TwoFlavourEvenOddRatioPseudoFermionAction det("<<DenOp.Mass()<<") / det("<<NumOp.Mass()<<")";
+	return sstream.str();
+      }
 
       virtual std::string LogParameters(){
 	std::stringstream sstream;
-	sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
+	sstream<< GridLogMessage << "["<<action_name()<<"] -- No further parameters "<<std::endl;
 	return sstream.str();
       } 
 
       
-      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.
@@ -100,12 +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);
 
@@ -124,10 +132,6 @@ NAMESPACE_BEGIN(Grid);
         // Even det factors
         DenOp.MooeeDag(etaEven,tmp);
         NumOp.MooeeInvDag(tmp,PhiEven);
-
-        PhiOdd =PhiOdd*scale;
-        PhiEven=PhiEven*scale;
-        
       };
 
       //////////////////////////////////////////////////////

Grid/qcd/action/pseudofermion/TwoFlavourRatioEO4DPseudoFermion.h

@@ -0,0 +1,203 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/TwoFlavourRatio.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class Impl>
+class TwoFlavourRatioEO4DPseudoFermionAction : public Action<typename Impl::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+private:
+  typedef FermionOperator<Impl> FermOp;
+  FermionOperator<Impl> & NumOp;// the basic operator
+  FermionOperator<Impl> & DenOp;// the basic operator
+
+  OperatorFunction<FermionField> &DerivativeSolver;
+  OperatorFunction<FermionField> &DerivativeDagSolver;
+  OperatorFunction<FermionField> &ActionSolver;
+  OperatorFunction<FermionField> &HeatbathSolver;
+
+  FermionField phi4; // the pseudo fermion field for this trajectory
+
+public:
+  TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
+					 FermionOperator<Impl>  &_DenOp, 
+					 OperatorFunction<FermionField> & DS,
+					 OperatorFunction<FermionField> & AS ) : 
+    TwoFlavourRatioEO4DPseudoFermionAction(_NumOp,_DenOp, DS,DS,AS,AS) {};
+  TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
+					 FermionOperator<Impl>  &_DenOp, 
+					 OperatorFunction<FermionField> & DS,
+					 OperatorFunction<FermionField> & DDS,
+					 OperatorFunction<FermionField> & AS,
+					 OperatorFunction<FermionField> & HS
+				       ) : NumOp(_NumOp),
+					   DenOp(_DenOp),
+					   DerivativeSolver(DS),
+					   DerivativeDagSolver(DDS),
+					   ActionSolver(AS),
+					   HeatbathSolver(HS),
+					   phi4(_NumOp.GaugeGrid())
+  {};
+      
+  virtual std::string action_name(){return "TwoFlavourRatioEO4DPseudoFermionAction";}
+
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
+    return sstream.str();
+  }  
+      
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
+
+    // P(phi) = e^{- phi^dag (V^dag M^-dag)_11  (M^-1 V)_11 phi}
+    //
+    // NumOp == V
+    // DenOp == M
+    //
+    // Take phi = (V^{-1} M)_11 eta  ; eta = (M^{-1} V)_11 Phi
+    //
+    // 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) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    FermionField eta4(NumOp.GaugeGrid());
+    FermionField eta5(NumOp.FermionGrid());
+    FermionField tmp(NumOp.FermionGrid());
+    FermionField phi5(NumOp.FermionGrid());
+
+    gaussian(pRNG,eta4);
+    NumOp.ImportFourDimPseudoFermion(eta4,eta5);
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(HeatbathSolver);
+
+    DenOp.M(eta5,tmp);               // M eta
+    PrecSolve(NumOp,tmp,phi5);  // phi = V^-1 M eta
+    phi5=phi5*scale;
+    std::cout << GridLogMessage << "4d pf refresh "<< norm2(phi5)<<"\n";
+    // Project to 4d
+    NumOp.ExportFourDimPseudoFermion(phi5,phi4);
+      
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag (V^dag M^-dag)_11  (M^-1 V)_11 phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField Y4(NumOp.GaugeGrid());
+    FermionField X(NumOp.FermionGrid());
+    FermionField Y(NumOp.FermionGrid());
+    FermionField phi5(NumOp.FermionGrid());
+	
+    MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagMOp(DenOp);
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(ActionSolver);
+
+    NumOp.ImportFourDimPseudoFermion(phi4,phi5);
+    NumOp.M(phi5,X);              // X= V phi
+    PrecSolve(DenOp,X,Y);    // Y= (MdagM)^-1 Mdag Vdag phi = M^-1 V phi
+    NumOp.ExportFourDimPseudoFermion(Y,Y4);
+
+    RealD action = norm2(Y4);
+
+    return action;
+  };
+
+  //////////////////////////////////////////////////////
+  // dS/du = 2 Re phi^dag (V^dag M^-dag)_11  (M^-1 d V)_11  phi
+  //       - 2 Re phi^dag (dV^dag M^-dag)_11  (M^-1 dM M^-1 V)_11  phi
+  //////////////////////////////////////////////////////
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
+
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField  X(NumOp.FermionGrid());
+    FermionField  Y(NumOp.FermionGrid());
+    FermionField       phi(NumOp.FermionGrid());
+    FermionField      Vphi(NumOp.FermionGrid());
+    FermionField  MinvVphi(NumOp.FermionGrid());
+    FermionField      tmp4(NumOp.GaugeGrid());
+    FermionField  MdagInvMinvVphi(NumOp.FermionGrid());
+
+    GaugeField   force(NumOp.GaugeGrid());	
+
+    //Y=V phi
+    //X = (Mdag V phi
+    //Y = (Mdag M)^-1 Mdag V phi = M^-1 V Phi
+    NumOp.ImportFourDimPseudoFermion(phi4,phi);
+    NumOp.M(phi,Vphi);               //  V phi
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(DerivativeSolver);
+    PrecSolve(DenOp,Vphi,MinvVphi);// M^-1 V phi
+    std::cout << GridLogMessage << "4d deriv solve "<< norm2(MinvVphi)<<"\n";
+
+    // Projects onto the physical space and back
+    NumOp.ExportFourDimPseudoFermion(MinvVphi,tmp4);
+    NumOp.ImportFourDimPseudoFermion(tmp4,Y);
+
+    SchurRedBlackDiagMooeeDagSolve<FermionField> PrecDagSolve(DerivativeDagSolver);
+    // X = proj M^-dag V phi
+    // Need an adjoint solve
+    PrecDagSolve(DenOp,Y,MdagInvMinvVphi);
+    std::cout << GridLogMessage << "4d deriv solve dag "<< norm2(MdagInvMinvVphi)<<"\n";
+    
+    // phi^dag (Vdag Mdag^-1) (M^-1 dV)  phi
+    NumOp.MDeriv(force ,MdagInvMinvVphi , phi, DaggerNo );  dSdU=force;
+  
+    // phi^dag (dVdag Mdag^-1) (M^-1 V)  phi
+    NumOp.MDeriv(force , phi, MdagInvMinvVphi ,DaggerYes  );  dSdU=dSdU+force;
+
+    //    - 2 Re phi^dag (dV^dag M^-dag)_11  (M^-1 dM M^-1 V)_11  phi
+    DenOp.MDeriv(force,MdagInvMinvVphi,MinvVphi,DaggerNo);   dSdU=dSdU-force;
+    DenOp.MDeriv(force,MinvVphi,MdagInvMinvVphi,DaggerYes);  dSdU=dSdU-force;
+
+    dSdU *= -1.0; 
+    //dSdU = - Ta(dSdU);
+    
+  };
+};
+
+NAMESPACE_END(Grid);
+
+

Grid/qcd/gparity/Gparity.h

@@ -0,0 +1,6 @@
+#ifndef GRID_GPARITY_H_
+#define GRID_GPARITY_H_
+
+#include<Grid/qcd/gparity/GparityFlavour.h>
+
+#endif

Grid/qcd/gparity/GparityFlavour.cc

@@ -0,0 +1,34 @@
+#include <Grid/Grid.h>
+
+NAMESPACE_BEGIN(Grid);
+
+const std::array<const GparityFlavour, 3> GparityFlavour::sigma_mu = {{
+    GparityFlavour(GparityFlavour::Algebra::SigmaX),
+    GparityFlavour(GparityFlavour::Algebra::SigmaY),
+    GparityFlavour(GparityFlavour::Algebra::SigmaZ)
+    }};
+
+const std::array<const GparityFlavour, 6> GparityFlavour::sigma_all = {{
+  GparityFlavour(GparityFlavour::Algebra::Identity),
+  GparityFlavour(GparityFlavour::Algebra::SigmaX),
+  GparityFlavour(GparityFlavour::Algebra::SigmaY),
+  GparityFlavour(GparityFlavour::Algebra::SigmaZ),
+  GparityFlavour(GparityFlavour::Algebra::ProjPlus),
+  GparityFlavour(GparityFlavour::Algebra::ProjMinus)
+}};
+
+const std::array<const char *, GparityFlavour::nSigma> GparityFlavour::name = {{
+    "SigmaX",
+    "MinusSigmaX",
+    "SigmaY",
+    "MinusSigmaY",
+    "SigmaZ",
+    "MinusSigmaZ",
+    "Identity",
+    "MinusIdentity",
+    "ProjPlus",
+    "MinusProjPlus",
+    "ProjMinus",
+    "MinusProjMinus"}};
+
+NAMESPACE_END(Grid);

Grid/qcd/gparity/GparityFlavour.h

@@ -0,0 +1,475 @@
+#ifndef GRID_QCD_GPARITY_FLAVOUR_H
+#define GRID_QCD_GPARITY_FLAVOUR_H
+
+//Support for flavour-matrix operations acting on the G-parity flavour index
+
+#include <array>
+
+NAMESPACE_BEGIN(Grid);
+
+class GparityFlavour {
+  public:
+    GRID_SERIALIZABLE_ENUM(Algebra, undef,
+                           SigmaX, 0,
+			   MinusSigmaX, 1,
+                           SigmaY, 2,
+			   MinusSigmaY, 3,
+                           SigmaZ, 4,
+			   MinusSigmaZ, 5,
+			   Identity, 6,
+			   MinusIdentity, 7,
+			   ProjPlus, 8,
+			   MinusProjPlus, 9,
+			   ProjMinus, 10,
+			   MinusProjMinus, 11
+			   );
+    static constexpr unsigned int nSigma = 12;
+    static const std::array<const char *, nSigma>                name;
+    static const std::array<const GparityFlavour, 3>             sigma_mu;
+    static const std::array<const GparityFlavour, 6>            sigma_all;
+    Algebra                                                      g;
+  public:
+  accelerator GparityFlavour(Algebra initg): g(initg) {}  
+};
+
+
+
+// 0 1  x   vector
+// 1 0
+template<class vtype>
+accelerator_inline void multFlavourSigmaX(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = rhs(1);
+  ret(1) = rhs(0);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourSigmaX(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = rhs(1,0);
+  ret(0,1) = rhs(1,1);
+  ret(1,0) = rhs(0,0);
+  ret(1,1) = rhs(0,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourSigmaX(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = rhs(0,1);
+  ret(0,1) = rhs(0,0);
+  ret(1,0) = rhs(1,1);
+  ret(1,1) = rhs(1,0);
+};
+
+
+template<class vtype>
+accelerator_inline void multFlavourMinusSigmaX(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = -rhs(1);
+  ret(1) = -rhs(0);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourMinusSigmaX(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -rhs(1,0);
+  ret(0,1) = -rhs(1,1);
+  ret(1,0) = -rhs(0,0);
+  ret(1,1) = -rhs(0,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourMinusSigmaX(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -rhs(0,1);
+  ret(0,1) = -rhs(0,0);
+  ret(1,0) = -rhs(1,1);
+  ret(1,1) = -rhs(1,0);
+};
+
+
+
+
+
+// 0 -i  x   vector
+// i 0
+template<class vtype>
+accelerator_inline void multFlavourSigmaY(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = timesMinusI(rhs(1));
+  ret(1) = timesI(rhs(0));
+};
+template<class vtype>
+accelerator_inline void lmultFlavourSigmaY(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = timesMinusI(rhs(1,0));
+  ret(0,1) = timesMinusI(rhs(1,1));
+  ret(1,0) = timesI(rhs(0,0));
+  ret(1,1) = timesI(rhs(0,1));
+};
+template<class vtype>
+accelerator_inline void rmultFlavourSigmaY(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = timesI(rhs(0,1));
+  ret(0,1) = timesMinusI(rhs(0,0));
+  ret(1,0) = timesI(rhs(1,1));
+  ret(1,1) = timesMinusI(rhs(1,0));
+};
+
+template<class vtype>
+accelerator_inline void multFlavourMinusSigmaY(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = timesI(rhs(1));
+  ret(1) = timesMinusI(rhs(0));
+};
+template<class vtype>
+accelerator_inline void lmultFlavourMinusSigmaY(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = timesI(rhs(1,0));
+  ret(0,1) = timesI(rhs(1,1));
+  ret(1,0) = timesMinusI(rhs(0,0));
+  ret(1,1) = timesMinusI(rhs(0,1));
+};
+template<class vtype>
+accelerator_inline void rmultFlavourMinusSigmaY(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = timesMinusI(rhs(0,1));
+  ret(0,1) = timesI(rhs(0,0));
+  ret(1,0) = timesMinusI(rhs(1,1));
+  ret(1,1) = timesI(rhs(1,0));
+};
+
+
+
+
+
+// 1 0  x   vector
+// 0 -1
+template<class vtype>
+accelerator_inline void multFlavourSigmaZ(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = rhs(0);
+  ret(1) = -rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourSigmaZ(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = rhs(0,0);
+  ret(0,1) = rhs(0,1);
+  ret(1,0) = -rhs(1,0);
+  ret(1,1) = -rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourSigmaZ(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = rhs(0,0);
+  ret(0,1) = -rhs(0,1);
+  ret(1,0) = rhs(1,0);
+  ret(1,1) = -rhs(1,1);
+};
+
+
+template<class vtype>
+accelerator_inline void multFlavourMinusSigmaZ(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = -rhs(0);
+  ret(1) = rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourMinusSigmaZ(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -rhs(0,0);
+  ret(0,1) = -rhs(0,1);
+  ret(1,0) = rhs(1,0);
+  ret(1,1) = rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourMinusSigmaZ(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -rhs(0,0);
+  ret(0,1) = rhs(0,1);
+  ret(1,0) = -rhs(1,0);
+  ret(1,1) = rhs(1,1);
+};
+
+
+
+
+
+
+template<class vtype>
+accelerator_inline void multFlavourIdentity(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = rhs(0);
+  ret(1) = rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourIdentity(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = rhs(0,0);
+  ret(0,1) = rhs(0,1);
+  ret(1,0) = rhs(1,0);
+  ret(1,1) = rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourIdentity(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = rhs(0,0);
+  ret(0,1) = rhs(0,1);
+  ret(1,0) = rhs(1,0);
+  ret(1,1) = rhs(1,1);
+};
+
+template<class vtype>
+accelerator_inline void multFlavourMinusIdentity(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = -rhs(0);
+  ret(1) = -rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourMinusIdentity(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -rhs(0,0);
+  ret(0,1) = -rhs(0,1);
+  ret(1,0) = -rhs(1,0);
+  ret(1,1) = -rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourMinusIdentity(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -rhs(0,0);
+  ret(0,1) = -rhs(0,1);
+  ret(1,0) = -rhs(1,0);
+  ret(1,1) = -rhs(1,1);
+};
+
+
+
+
+
+//G-parity flavour projection 1/2(1+\sigma_2)
+//1 -i
+//i  1
+template<class vtype>
+accelerator_inline void multFlavourProjPlus(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = 0.5*rhs(0) + 0.5*timesMinusI(rhs(1));
+  ret(1) = 0.5*timesI(rhs(0)) + 0.5*rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourProjPlus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = 0.5*rhs(0,0) + 0.5*timesMinusI(rhs(1,0));
+  ret(0,1) = 0.5*rhs(0,1) + 0.5*timesMinusI(rhs(1,1));
+  ret(1,0) = 0.5*timesI(rhs(0,0)) + 0.5*rhs(1,0);
+  ret(1,1) = 0.5*timesI(rhs(0,1)) + 0.5*rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourProjPlus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = 0.5*rhs(0,0) + 0.5*timesI(rhs(0,1));
+  ret(0,1) = 0.5*timesMinusI(rhs(0,0)) + 0.5*rhs(0,1);
+  ret(1,0) = 0.5*rhs(1,0) + 0.5*timesI(rhs(1,1));
+  ret(1,1) = 0.5*timesMinusI(rhs(1,0)) + 0.5*rhs(1,1);
+};
+
+
+template<class vtype>
+accelerator_inline void multFlavourMinusProjPlus(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = -0.5*rhs(0) + 0.5*timesI(rhs(1));
+  ret(1) = 0.5*timesMinusI(rhs(0)) - 0.5*rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourMinusProjPlus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -0.5*rhs(0,0) + 0.5*timesI(rhs(1,0));
+  ret(0,1) = -0.5*rhs(0,1) + 0.5*timesI(rhs(1,1));
+  ret(1,0) = 0.5*timesMinusI(rhs(0,0)) - 0.5*rhs(1,0);
+  ret(1,1) = 0.5*timesMinusI(rhs(0,1)) - 0.5*rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourMinusProjPlus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -0.5*rhs(0,0) + 0.5*timesMinusI(rhs(0,1));
+  ret(0,1) = 0.5*timesI(rhs(0,0)) - 0.5*rhs(0,1);
+  ret(1,0) = -0.5*rhs(1,0) + 0.5*timesMinusI(rhs(1,1));
+  ret(1,1) = 0.5*timesI(rhs(1,0)) - 0.5*rhs(1,1);
+};
+
+
+
+
+
+//G-parity flavour projection 1/2(1-\sigma_2)
+//1 i
+//-i  1
+template<class vtype>
+accelerator_inline void multFlavourProjMinus(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = 0.5*rhs(0) + 0.5*timesI(rhs(1));
+  ret(1) = 0.5*timesMinusI(rhs(0)) + 0.5*rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourProjMinus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = 0.5*rhs(0,0) + 0.5*timesI(rhs(1,0));
+  ret(0,1) = 0.5*rhs(0,1) + 0.5*timesI(rhs(1,1));
+  ret(1,0) = 0.5*timesMinusI(rhs(0,0)) + 0.5*rhs(1,0);
+  ret(1,1) = 0.5*timesMinusI(rhs(0,1)) + 0.5*rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourProjMinus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = 0.5*rhs(0,0) + 0.5*timesMinusI(rhs(0,1));
+  ret(0,1) = 0.5*timesI(rhs(0,0)) + 0.5*rhs(0,1);
+  ret(1,0) = 0.5*rhs(1,0) + 0.5*timesMinusI(rhs(1,1));
+  ret(1,1) = 0.5*timesI(rhs(1,0)) + 0.5*rhs(1,1);
+};
+
+
+template<class vtype>
+accelerator_inline void multFlavourMinusProjMinus(iVector<vtype, Ngp> &ret, const iVector<vtype, Ngp> &rhs)
+{
+  ret(0) = -0.5*rhs(0) + 0.5*timesMinusI(rhs(1));
+  ret(1) = 0.5*timesI(rhs(0)) - 0.5*rhs(1);
+};
+template<class vtype>
+accelerator_inline void lmultFlavourMinusProjMinus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -0.5*rhs(0,0) + 0.5*timesMinusI(rhs(1,0));
+  ret(0,1) = -0.5*rhs(0,1) + 0.5*timesMinusI(rhs(1,1));
+  ret(1,0) = 0.5*timesI(rhs(0,0)) - 0.5*rhs(1,0);
+  ret(1,1) = 0.5*timesI(rhs(0,1)) - 0.5*rhs(1,1);
+};
+template<class vtype>
+accelerator_inline void rmultFlavourMinusProjMinus(iMatrix<vtype, Ngp> &ret, const iMatrix<vtype, Ngp> &rhs)
+{
+  ret(0,0) = -0.5*rhs(0,0) + 0.5*timesI(rhs(0,1));
+  ret(0,1) = 0.5*timesMinusI(rhs(0,0)) - 0.5*rhs(0,1);
+  ret(1,0) = -0.5*rhs(1,0) + 0.5*timesI(rhs(1,1));
+  ret(1,1) = 0.5*timesMinusI(rhs(1,0)) - 0.5*rhs(1,1);
+};
+
+
+
+
+
+
+
+
+
+
+template<class vtype> 
+accelerator_inline auto operator*(const GparityFlavour &G, const iVector<vtype, Ngp> &arg)
+->typename std::enable_if<matchGridTensorIndex<iVector<vtype, Ngp>, GparityFlavourTensorIndex>::value, iVector<vtype, Ngp>>::type
+{
+  iVector<vtype, Ngp> ret;
+
+  switch (G.g) 
+  {
+  case GparityFlavour::Algebra::SigmaX:
+    multFlavourSigmaX(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaX:
+    multFlavourMinusSigmaX(ret, arg); break;
+  case GparityFlavour::Algebra::SigmaY:
+    multFlavourSigmaY(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaY:
+    multFlavourMinusSigmaY(ret, arg); break;
+  case GparityFlavour::Algebra::SigmaZ:
+    multFlavourSigmaZ(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaZ:
+    multFlavourMinusSigmaZ(ret, arg); break;
+  case GparityFlavour::Algebra::Identity:
+    multFlavourIdentity(ret, arg); break;
+  case GparityFlavour::Algebra::MinusIdentity:
+    multFlavourMinusIdentity(ret, arg); break;
+  case GparityFlavour::Algebra::ProjPlus:
+    multFlavourProjPlus(ret, arg); break;
+  case GparityFlavour::Algebra::MinusProjPlus:
+    multFlavourMinusProjPlus(ret, arg); break;
+  case GparityFlavour::Algebra::ProjMinus:
+    multFlavourProjMinus(ret, arg); break;
+  case GparityFlavour::Algebra::MinusProjMinus:
+    multFlavourMinusProjMinus(ret, arg); break;
+  default: assert(0);
+  }
+ 
+  return ret;
+}
+
+template<class vtype> 
+accelerator_inline auto operator*(const GparityFlavour &G, const iMatrix<vtype, Ngp> &arg)
+->typename std::enable_if<matchGridTensorIndex<iMatrix<vtype, Ngp>, GparityFlavourTensorIndex>::value, iMatrix<vtype, Ngp>>::type
+{
+  iMatrix<vtype, Ngp> ret;
+
+  switch (G.g) 
+  {
+  case GparityFlavour::Algebra::SigmaX:
+    lmultFlavourSigmaX(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaX:
+    lmultFlavourMinusSigmaX(ret, arg); break;
+  case GparityFlavour::Algebra::SigmaY:
+    lmultFlavourSigmaY(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaY:
+    lmultFlavourMinusSigmaY(ret, arg); break;
+  case GparityFlavour::Algebra::SigmaZ:
+    lmultFlavourSigmaZ(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaZ:
+    lmultFlavourMinusSigmaZ(ret, arg); break;
+  case GparityFlavour::Algebra::Identity:
+    lmultFlavourIdentity(ret, arg); break;
+  case GparityFlavour::Algebra::MinusIdentity:
+    lmultFlavourMinusIdentity(ret, arg); break;
+  case GparityFlavour::Algebra::ProjPlus:
+    lmultFlavourProjPlus(ret, arg); break;
+  case GparityFlavour::Algebra::MinusProjPlus:
+    lmultFlavourMinusProjPlus(ret, arg); break;
+  case GparityFlavour::Algebra::ProjMinus:
+    lmultFlavourProjMinus(ret, arg); break;
+  case GparityFlavour::Algebra::MinusProjMinus:
+    lmultFlavourMinusProjMinus(ret, arg); break;  
+  default: assert(0);
+  }
+  
+  return ret;
+}
+
+template<class vtype> 
+accelerator_inline auto operator*(const iMatrix<vtype, Ngp> &arg, const GparityFlavour &G)
+->typename std::enable_if<matchGridTensorIndex<iMatrix<vtype, Ngp>, GparityFlavourTensorIndex>::value, iMatrix<vtype, Ngp>>::type
+{
+  iMatrix<vtype, Ngp> ret;
+
+  switch (G.g) 
+  {
+  case GparityFlavour::Algebra::SigmaX:
+    rmultFlavourSigmaX(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaX:
+    rmultFlavourMinusSigmaX(ret, arg); break;
+  case GparityFlavour::Algebra::SigmaY:
+    rmultFlavourSigmaY(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaY:
+    rmultFlavourMinusSigmaY(ret, arg); break;
+  case GparityFlavour::Algebra::SigmaZ:
+    rmultFlavourSigmaZ(ret, arg); break;
+  case GparityFlavour::Algebra::MinusSigmaZ:
+    rmultFlavourMinusSigmaZ(ret, arg); break;
+  case GparityFlavour::Algebra::Identity:
+    rmultFlavourIdentity(ret, arg); break;
+  case GparityFlavour::Algebra::MinusIdentity:
+    rmultFlavourMinusIdentity(ret, arg); break;
+  case GparityFlavour::Algebra::ProjPlus:
+    rmultFlavourProjPlus(ret, arg); break;
+  case GparityFlavour::Algebra::MinusProjPlus:
+    rmultFlavourMinusProjPlus(ret, arg); break;
+  case GparityFlavour::Algebra::ProjMinus:
+    rmultFlavourProjMinus(ret, arg); break;
+  case GparityFlavour::Algebra::MinusProjMinus:
+    rmultFlavourMinusProjMinus(ret, arg); break;
+  default: assert(0);
+  }
+
+  return ret;
+}
+
+NAMESPACE_END(Grid);
+
+#endif // include guard

Grid/qcd/hmc/GenericHMCrunner.h

@@ -129,18 +129,10 @@ public:
     Runner(S);
   }
 
-  //////////////////////////////////////////////////////////////////
-
-private:
-  template <class SmearingPolicy>
-  void Runner(SmearingPolicy &Smearing) {
-    auto UGrid = Resources.GetCartesian();
-    Resources.AddRNGs();
-    Field U(UGrid);
-
-    // Can move this outside?
-    typedef IntegratorType<SmearingPolicy> TheIntegrator;
-    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
+  //Use the checkpointer to initialize the RNGs and the gauge field, writing the resulting gauge field into U.
+  //This is called automatically by Run but may be useful elsewhere, e.g. for integrator tuning experiments
+  void initializeGaugeFieldAndRNGs(Field &U){
+    if(!Resources.haveRNGs()) Resources.AddRNGs();
 
     if (Parameters.StartingType == "HotStart") {
       // Hot start
@@ -159,14 +151,43 @@ private:
       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);
     }
+  }
+
+
+
+  //////////////////////////////////////////////////////////////////
+
+private:
+  template <class SmearingPolicy>
+  void Runner(SmearingPolicy &Smearing) {
+    auto UGrid = Resources.GetCartesian();
+    Field U(UGrid);
+
+    initializeGaugeFieldAndRNGs(U);
+
+    typedef IntegratorType<SmearingPolicy> TheIntegrator;
+    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
+
+    // Sets the momentum filter
+    MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
 
     Smearing.set_Field(U);
 

Grid/qcd/hmc/HMC.h

@@ -34,6 +34,7 @@ directory
 			    * @brief Classes for Hybrid Monte Carlo update
 			    *
 			    * @author Guido Cossu
+			    * @author Peter Boyle
 			    */
 			   //--------------------------------------------------------------------
 #pragma once
@@ -115,22 +116,17 @@ private:
 
     random(sRNG, rn_test);
 
-    std::cout << GridLogMessage
-              << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "exp(-dH) = " << prob
-              << "  Random = " << rn_test << "\n";
-    std::cout << GridLogMessage
-              << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+    std::cout << GridLogHMC << "exp(-dH) = " << prob << "  Random = " << rn_test << "\n";
+    std::cout << GridLogHMC << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
 
     if ((prob > 1.0) || (rn_test <= prob)) {  // accepted
-      std::cout << GridLogMessage << "Metropolis_test -- ACCEPTED\n";
-      std::cout << GridLogMessage
-                << "--------------------------------------------------\n";
+      std::cout << GridLogHMC << "Metropolis_test -- ACCEPTED\n";
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
-      std::cout << GridLogMessage << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogMessage
-                << "--------------------------------------------------\n";
+      std::cout << GridLogHMC << "Metropolis_test -- REJECTED\n";
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
       return false;
     }
   }
@@ -139,19 +135,68 @@ private:
   // Evolution
   /////////////////////////////////////////////////////////
   RealD evolve_hmc_step(Field &U) {
-    TheIntegrator.refresh(U, sRNG, pRNG);  // set U and initialize P and phi's
 
-    RealD H0 = TheIntegrator.S(U);  // initial state action
+    GridBase *Grid = U.Grid();
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // Mainly for DDHMC perform a random translation of U modulo volume
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Random shifting gauge field by [";
+    for(int d=0;d<Grid->Nd();d++) {
+
+      int L = Grid->GlobalDimensions()[d];
+
+      RealD rn_uniform;  random(sRNG, rn_uniform);
+
+      int shift = (int) (rn_uniform*L);
+
+      std::cout << shift;
+      if(d<Grid->Nd()-1) std::cout <<",";
+      else               std::cout <<"]\n";
+      
+      U = Cshift(U,d,shift);
+    }
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    TheIntegrator.reset_timer();
+    
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // set U and initialize P and phi's
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Refresh momenta and pseudofermions";
+    TheIntegrator.refresh(U, sRNG, pRNG);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // initial state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Compute initial action";
+    RealD H0 = TheIntegrator.S(U);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
     std::streamsize current_precision = std::cout.precision();
     std::cout.precision(15);
-    std::cout << GridLogMessage << "Total H before trajectory = " << H0 << "\n";
+    std::cout << GridLogHMC << "Total H before trajectory = " << H0 << "\n";
     std::cout.precision(current_precision);
 
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << " Molecular Dynamics evolution ";
     TheIntegrator.integrate(U);
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
-    RealD H1 = TheIntegrator.S(U);  // updated state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // updated state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Compute final action";
+    RealD H1 = TheIntegrator.S(U);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
+
+    
     ///////////////////////////////////////////////////////////
     if(0){
       std::cout << "------------------------- Reversibility test" << std::endl;
@@ -163,17 +208,16 @@ private:
     }
     ///////////////////////////////////////////////////////////
 
-
     std::cout.precision(15);
-    std::cout << GridLogMessage << "Total H after trajectory  = " << H1
-	      << "  dH = " << H1 - H0 << "\n";
+
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+    std::cout << GridLogHMC << "Total H after trajectory  = " << H1 << "  dH = " << H1 - H0 << "\n";
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+
     std::cout.precision(current_precision);
     
     return (H1 - H0);
   }
-  
-
-  
 
 public:
   /////////////////////////////////////////
@@ -195,10 +239,13 @@ public:
 
     // Actual updates (evolve a copy Ucopy then copy back eventually)
     unsigned int FinalTrajectory = Params.Trajectories + Params.NoMetropolisUntil + Params.StartTrajectory;
+
     for (int traj = Params.StartTrajectory; traj < FinalTrajectory; ++traj) {
-      std::cout << GridLogMessage << "-- # Trajectory = " << traj << "\n";
+
+      std::cout << GridLogHMC << "-- # Trajectory = " << traj << "\n";
+
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
-      	std::cout << GridLogMessage << "-- Thermalization" << std::endl;
+      	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
       }
       
       double t0=usecond();
@@ -207,20 +254,19 @@ public:
       DeltaH = evolve_hmc_step(Ucopy);
       // Metropolis-Hastings test
       bool accept = true;
-      if (traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
+      if (Params.MetropolisTest && traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
         accept = metropolis_test(DeltaH);
       } else {
-      	std::cout << GridLogMessage << "Skipping Metropolis test" << std::endl;
+      	std::cout << GridLogHMC << "Skipping Metropolis test" << std::endl;
       }
 
       if (accept)
         Ucur = Ucopy; 
       
-     
-      
       double t1=usecond();
-      std::cout << GridLogMessage << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
+      std::cout << GridLogHMC << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
 
+      TheIntegrator.print_timer();
 
       for (int obs = 0; obs < Observables.size(); obs++) {
       	std::cout << GridLogDebug << "Observables # " << obs << std::endl;
@@ -228,7 +274,7 @@ public:
       	std::cout << GridLogDebug << "Observables pointer " << Observables[obs] << std::endl;
         Observables[obs]->TrajectoryComplete(traj + 1, Ucur, sRNG, pRNG);
       }
-      std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
+      std::cout << GridLogHMC << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
     }
   }
 

Grid/qcd/hmc/HMCModules.h

@@ -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);
   }
 };

Grid/qcd/hmc/HMCResourceManager.h

@@ -72,6 +72,8 @@ class HMCResourceManager {
   typedef HMCModuleBase< BaseHmcCheckpointer<ImplementationPolicy> > CheckpointerBaseModule;
   typedef HMCModuleBase< HmcObservable<typename ImplementationPolicy::Field> > ObservableBaseModule;
   typedef ActionModuleBase< Action<typename ImplementationPolicy::Field>, GridModule > ActionBaseModule;
+  typedef typename ImplementationPolicy::Field MomentaField;
+  typedef typename ImplementationPolicy::Field Field;  
 
   // Named storage for grid pairs (std + red-black)
   std::unordered_map<std::string, GridModule> Grids;
@@ -80,6 +82,9 @@ class HMCResourceManager {
   // SmearingModule<ImplementationPolicy> Smearing;
   std::unique_ptr<CheckpointerBaseModule> CP;
 
+  // Momentum filter
+  std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> > Filter;
+  
   // A vector of HmcObservable modules
   std::vector<std::unique_ptr<ObservableBaseModule> > ObservablesList;
 
@@ -90,6 +95,7 @@ class HMCResourceManager {
 
   bool have_RNG;
   bool have_CheckPointer;
+  bool have_Filter;
 
   // NOTE: operator << is not overloaded for std::vector<string> 
   // so this function is necessary
@@ -101,7 +107,7 @@ class HMCResourceManager {
 
 
 public:
-  HMCResourceManager() : have_RNG(false), have_CheckPointer(false) {}
+  HMCResourceManager() : have_RNG(false), have_CheckPointer(false), have_Filter(false) {}
 
   template <class ReaderClass, class vector_type = vComplex >
   void initialize(ReaderClass &Read){
@@ -129,6 +135,7 @@ public:
     RNGModuleParameters RNGpar(Read);
     SetRNGSeeds(RNGpar);
 
+  
     // Observables
     auto &ObsFactory = HMC_ObservablesModuleFactory<observable_string, typename ImplementationPolicy::Field, ReaderClass>::getInstance(); 
     Read.push(observable_string);// here must check if existing...
@@ -208,6 +215,16 @@ public:
     AddGrid(s, Mod);
   }
 
+  void SetMomentumFilter( MomentumFilterBase<typename ImplementationPolicy::Field> * MomFilter) {
+    assert(have_Filter==false);
+    Filter = std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> >(MomFilter);
+    have_Filter = true;
+  }
+  MomentumFilterBase<typename ImplementationPolicy::Field> *GetMomentumFilter(void) {
+    if ( !have_Filter)
+      SetMomentumFilter(new MomentumFilterNone<typename ImplementationPolicy::Field>());
+    return Filter.get();
+  }
 
   GridCartesian* GetCartesian(std::string s = "") {
     if (s.empty()) s = Grids.begin()->first;
@@ -226,6 +243,9 @@ public:
   //////////////////////////////////////////////////////
   // Random number generators
   //////////////////////////////////////////////////////
+  
+  //Return true if the RNG objects have been instantiated
+  bool haveRNGs() const{ return have_RNG; }
 
   void AddRNGs(std::string s = "") {
     // Couple the RNGs to the GridModule tagged by s

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,7 +33,6 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
-#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -67,6 +66,7 @@ public:
 template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
+
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 
@@ -119,36 +119,58 @@ protected:
     }
   } update_P_hireps{};
 
+ 
   void update_P(MomentaField& Mom, Field& U, int level, double ep) {
     // input U actually not used in the fundamental case
     // Fundamental updates, include smearing
 
     for (int a = 0; a < as[level].actions.size(); ++a) {
+
       double start_full = usecond();
       Field force(U.Grid());
       conformable(U.Grid(), Mom.Grid());
 
       Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
+      as[level].actions.at(a)->deriv_timer_start();
       as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
+      as[level].actions.at(a)->deriv_timer_stop();
 
       std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
+      auto name = as[level].actions.at(a)->action_name();
       if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
+
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
-      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
-      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << std::endl;
+
+      MomFilter->applyFilter(force);
+      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<< std::endl;
+      DumpSliceNorm("force ",force,Nd-1);
+      
+      Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
+      Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
+
+      Real force_max   = std::sqrt(maxLocalNorm2(force));
+      Real impulse_max = force_max * ep * HMC_MOMENTUM_DENOMINATOR;    
+
+      as[level].actions.at(a)->deriv_log(force_abs,force_max);
+      
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force average: " << force_abs <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force max    : " << force_max <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt average  : " << impulse_abs <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt max      : " << impulse_max <<" "<<name<<std::endl;
+
       Mom -= force * ep* HMC_MOMENTUM_DENOMINATOR;; 
       double end_full = usecond();
       double time_full  = (end_full - start_full) / 1e3;
       double time_force = (end_force - start_force) / 1e3;
       std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)"  << std::endl;
+
     }
 
     // Force from the other representations
     as[level].apply(update_P_hireps, Representations, Mom, U, ep);
 
-    MomFilter->applyFilter(Mom);
   }
 
   void update_U(Field& U, double ep) 
@@ -162,8 +184,12 @@ protected:
   
   void update_U(MomentaField& Mom, Field& U, double ep) 
   {
+    MomentaField MomFiltered(Mom.Grid());
+    MomFiltered = Mom;
+    MomFilter->applyFilter(MomFiltered);
+
     // exponential of Mom*U in the gauge fields case
-    FieldImplementation::update_field(Mom, U, ep);
+    FieldImplementation::update_field(MomFiltered, U, ep);
 
     // Update the smeared fields, can be implemented as observer
     Smearer.set_Field(U);
@@ -206,6 +232,66 @@ public:
   const MomentaField & getMomentum() const{ return P; }
   
 
+  void reset_timer(void)
+  {
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+        as[level].actions.at(actionID)->reset_timer();
+      }
+    }
+  }
+  void print_timer(void)
+  {
+    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::" << std::endl;
+    std::cout << GridLogMessage << " Refresh cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->refresh_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Action cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->S_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Force cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->deriv_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Force average size "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] : "
+		  <<" force max " << as[level].actions.at(actionID)->deriv_max_average()
+		  <<" norm "      << as[level].actions.at(actionID)->deriv_norm_average()
+		  <<" calls "     << as[level].actions.at(actionID)->deriv_num
+		  << std::endl;
+      }
+    }
+    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
+  }
+  
   void print_parameters()
   {
     std::cout << GridLogMessage << "[Integrator] Name : "<< integrator_name() << std::endl;
@@ -224,7 +310,6 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
-
   }
 
   void reverse_momenta()
@@ -249,15 +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
@@ -267,15 +356,19 @@ public:
       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
+	auto name = as[level].actions.at(actionID)->action_name();
+        std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
+
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
+	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
+	as[level].actions.at(actionID)->refresh_timer_stop();
       }
 
       // Refresh the higher representation actions
       as[level].apply(refresh_hireps, Representations, sRNG, pRNG);
     }
 
-    MomFilter->applyFilter(P);
   }
 
   // to be used by the actionlevel class to iterate
@@ -310,7 +403,9 @@ public:
         // 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;
+	        as[level].actions.at(actionID)->S_timer_start();
         Hterm = as[level].actions.at(actionID)->S(Us);
+   	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
       }

Grid/qcd/observables/topological_charge.h

@@ -99,7 +99,7 @@ public:
 	// using wilson flow by default here
 	WilsonFlow<PeriodicGimplR> WF(Pars.Smearing.steps, Pars.Smearing.step_size, Pars.Smearing.meas_interval);
 	WF.smear_adaptive(Usmear, U, Pars.Smearing.maxTau);
-	Real T0   = WF.energyDensityPlaquette(Usmear);
+	Real T0   = WF.energyDensityPlaquette(Pars.Smearing.maxTau, Usmear);
 	std::cout << GridLogMessage << std::setprecision(std::numeric_limits<Real>::digits10 + 1)
 		  << "T0                : [ " << traj << " ] "<< T0 << std::endl;
       }

Grid/qcd/smearing/WilsonFlow.h

@@ -7,6 +7,7 @@ Source file: ./lib/qcd/modules/plaquette.h
 Copyright (C) 2017
 
 Author: Guido Cossu <guido.cossu@ed.ac.uk>
+Author: Christopher Kelly <ckelly@bnl.gov>
 
 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
@@ -33,28 +34,44 @@ NAMESPACE_BEGIN(Grid);
 
 template <class Gimpl>
 class WilsonFlow: public Smear<Gimpl>{
+public:
+  //Store generic measurements to take during smearing process using std::function
+  typedef std::function<void(int, RealD, const typename Gimpl::GaugeField &)> FunctionType;  //int: step,  RealD: flow time,  GaugeField : the gauge field
+  
+private:
   unsigned int Nstep;
-  unsigned int measure_interval;
-  mutable RealD epsilon, taus;
-
+  RealD epsilon; //for regular smearing this is the time step, for adaptive it is the initial time step
+ 
+  std::vector< std::pair<int, FunctionType> > functions; //The int maps to the measurement frequency
 
   mutable WilsonGaugeAction<Gimpl> SG;
 
-  void evolve_step(typename Gimpl::GaugeField&) const;
-  void evolve_step_adaptive(typename Gimpl::GaugeField&, RealD);
-  RealD tau(unsigned int t)const {return epsilon*(t+1.0); }
+  //Evolve the gauge field by 1 step and update tau
+  void evolve_step(typename Gimpl::GaugeField &U, RealD &tau) const;
+  //Evolve the gauge field by 1 step and update tau and the current time step eps
+  void evolve_step_adaptive(typename Gimpl::GaugeField&U, RealD &tau, RealD &eps, RealD maxTau) const;
 
 public:
   INHERIT_GIMPL_TYPES(Gimpl)
 
+  void resetActions(){ functions.clear(); }
+
+  void addMeasurement(int meas_interval, FunctionType meas){ functions.push_back({meas_interval, meas}); }
+
+  //Set the class to perform the default measurements: 
+  //the plaquette energy density every step
+  //the plaquette topological charge every 'topq_meas_interval' steps
+  //and output to stdout
+  void setDefaultMeasurements(int topq_meas_interval = 1);
+
   explicit WilsonFlow(unsigned int Nstep, RealD epsilon, unsigned int interval = 1):
   Nstep(Nstep),
     epsilon(epsilon),
-    measure_interval(interval),
     SG(WilsonGaugeAction<Gimpl>(3.0)) {
     // WilsonGaugeAction with beta 3.0
     assert(epsilon > 0.0);
     LogMessage();
+    setDefaultMeasurements(interval);
   }
 
   void LogMessage() {
@@ -73,9 +90,29 @@ public:
     // undefined for WilsonFlow
   }
 
-  void smear_adaptive(GaugeField&, const GaugeField&, RealD maxTau);
-  RealD energyDensityPlaquette(unsigned int step, const GaugeField& U) const;
-  RealD energyDensityPlaquette(const GaugeField& U) const;
+  void smear_adaptive(GaugeField&, const GaugeField&, RealD maxTau) const;
+
+  //Compute t^2 <E(t)> for time t from the plaquette
+  static RealD energyDensityPlaquette(const RealD t, const GaugeField& U);
+
+  //Compute t^2 <E(t)> for time t from the 1x1 cloverleaf form
+  //t is the Wilson flow time
+  static RealD energyDensityCloverleaf(const RealD t, const GaugeField& U);
+  
+  //Evolve the gauge field by Nstep steps of epsilon and return the energy density computed every interval steps
+  //The smeared field is output as V
+  std::vector<RealD> flowMeasureEnergyDensityPlaquette(GaugeField &V, const GaugeField& U, int measure_interval = 1);
+
+  //Version that does not return the smeared field
+  std::vector<RealD> flowMeasureEnergyDensityPlaquette(const GaugeField& U, int measure_interval = 1);
+
+
+  //Evolve the gauge field by Nstep steps of epsilon and return the Cloverleaf energy density computed every interval steps
+  //The smeared field is output as V
+  std::vector<RealD> flowMeasureEnergyDensityCloverleaf(GaugeField &V, const GaugeField& U, int measure_interval = 1);
+
+  //Version that does not return the smeared field
+  std::vector<RealD> flowMeasureEnergyDensityCloverleaf(const GaugeField& U, int measure_interval = 1);
 };
 
 
@@ -83,7 +120,7 @@ public:
 // Implementations
 ////////////////////////////////////////////////////////////////////////////////
 template <class Gimpl>
-void WilsonFlow<Gimpl>::evolve_step(typename Gimpl::GaugeField &U) const{
+void WilsonFlow<Gimpl>::evolve_step(typename Gimpl::GaugeField &U, RealD &tau) const{
   GaugeField Z(U.Grid());
   GaugeField tmp(U.Grid());
   SG.deriv(U, Z);
@@ -99,12 +136,13 @@ void WilsonFlow<Gimpl>::evolve_step(typename Gimpl::GaugeField &U) const{
   SG.deriv(U, tmp); Z += tmp;                 // 4/3*(17/36*Z0 -8/9*Z1) +Z2
   Z *= 3.0/4.0;                               // Z = 17/36*Z0 -8/9*Z1 +3/4*Z2
   Gimpl::update_field(Z, U, -2.0*epsilon);    // V(t+e) = exp(ep*Z)*W2
+  tau += epsilon;
 }
 
 template <class Gimpl>
-void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, RealD maxTau) {
-  if (maxTau - taus < epsilon){
-    epsilon = maxTau-taus;
+void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, RealD &tau, RealD &eps, RealD maxTau) const{
+  if (maxTau - tau < eps){
+    eps = maxTau-tau;
   }
   //std::cout << GridLogMessage << "Integration epsilon : " << epsilon << std::endl;
   GaugeField Z(U.Grid());
@@ -114,95 +152,151 @@ void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, Real
   SG.deriv(U, Z);
   Zprime = -Z;
   Z *= 0.25;                                  // Z0 = 1/4 * F(U)
-  Gimpl::update_field(Z, U, -2.0*epsilon);    // U = W1 = exp(ep*Z0)*W0
+  Gimpl::update_field(Z, U, -2.0*eps);    // U = W1 = exp(ep*Z0)*W0
 
   Z *= -17.0/8.0;
   SG.deriv(U, tmp); Z += tmp;                 // -17/32*Z0 +Z1
   Zprime += 2.0*tmp;
   Z *= 8.0/9.0;                               // Z = -17/36*Z0 +8/9*Z1
-  Gimpl::update_field(Z, U, -2.0*epsilon);    // U_= W2 = exp(ep*Z)*W1
+  Gimpl::update_field(Z, U, -2.0*eps);    // U_= W2 = exp(ep*Z)*W1
     
 
   Z *= -4.0/3.0;
   SG.deriv(U, tmp); Z += tmp;                 // 4/3*(17/36*Z0 -8/9*Z1) +Z2
   Z *= 3.0/4.0;                               // Z = 17/36*Z0 -8/9*Z1 +3/4*Z2
-  Gimpl::update_field(Z, U, -2.0*epsilon);    // V(t+e) = exp(ep*Z)*W2
+  Gimpl::update_field(Z, U, -2.0*eps);    // V(t+e) = exp(ep*Z)*W2
 
   // Ramos 
-  Gimpl::update_field(Zprime, Uprime, -2.0*epsilon); // V'(t+e) = exp(ep*Z')*W0
+  Gimpl::update_field(Zprime, Uprime, -2.0*eps); // V'(t+e) = exp(ep*Z')*W0
   // Compute distance as norm^2 of the difference
   GaugeField diffU = U - Uprime;
   RealD diff = norm2(diffU);
   // adjust integration step
     
-  taus += epsilon;
+  tau += eps;
   //std::cout << GridLogMessage << "Adjusting integration step with distance: " << diff << std::endl;
     
-  epsilon = epsilon*0.95*std::pow(1e-4/diff,1./3.);
+  eps = eps*0.95*std::pow(1e-4/diff,1./3.);
   //std::cout << GridLogMessage << "New epsilon : " << epsilon << std::endl;
 
 }
 
+
 template <class Gimpl>
-RealD WilsonFlow<Gimpl>::energyDensityPlaquette(unsigned int step, const GaugeField& U) const {
-  RealD td = tau(step);
-  return 2.0 * td * td * SG.S(U)/U.Grid()->gSites();
+RealD WilsonFlow<Gimpl>::energyDensityPlaquette(const RealD t, const GaugeField& U){
+  static WilsonGaugeAction<Gimpl> SG(3.0);
+  return 2.0 * t * t * SG.S(U)/U.Grid()->gSites();
+}
+
+//Compute t^2 <E(t)> for time from the 1x1 cloverleaf form
+template <class Gimpl>
+RealD WilsonFlow<Gimpl>::energyDensityCloverleaf(const RealD t, const GaugeField& U){
+  typedef typename Gimpl::GaugeLinkField GaugeMat;
+  typedef typename Gimpl::GaugeField GaugeLorentz;
+
+  assert(Nd == 4);
+  //E = 1/2 tr( F_munu F_munu )
+  //However as  F_numu = -F_munu, only need to sum the trace of the squares of the following 6 field strengths:
+  //F_01 F_02 F_03   F_12 F_13  F_23
+  GaugeMat F(U.Grid());
+  LatticeComplexD R(U.Grid());
+  R = Zero();
+  
+  for(int mu=0;mu<3;mu++){
+    for(int nu=mu+1;nu<4;nu++){
+      WilsonLoops<Gimpl>::FieldStrength(F, U, mu, nu);
+      R = R + trace(F*F);
+    }
+  }
+  ComplexD out = sum(R);
+  out = t*t*out / RealD(U.Grid()->gSites());
+  return -real(out); //minus sign necessary for +ve energy
+}
+
+
+template <class Gimpl>
+std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityPlaquette(GaugeField &V, const GaugeField& U, int measure_interval){
+  std::vector<RealD> out;
+  resetActions();
+  addMeasurement(measure_interval, [&out](int step, RealD t, const typename Gimpl::GaugeField &U){ 
+      std::cout << GridLogMessage << "[WilsonFlow] Computing plaquette energy density for step " << step << std::endl;
+      out.push_back( energyDensityPlaquette(t,U) );
+    });      
+  smear(V,U);
+  return out;
 }
 
 template <class Gimpl>
-RealD WilsonFlow<Gimpl>::energyDensityPlaquette(const GaugeField& U) const {
-  return 2.0 * taus * taus * SG.S(U)/U.Grid()->gSites();
+std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityPlaquette(const GaugeField& U, int measure_interval){
+  GaugeField V(U);
+  return flowMeasureEnergyDensityPlaquette(V,U, measure_interval);
 }
 
+template <class Gimpl>
+std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityCloverleaf(GaugeField &V, const GaugeField& U, int measure_interval){
+  std::vector<RealD> out;
+  resetActions();
+  addMeasurement(measure_interval, [&out](int step, RealD t, const typename Gimpl::GaugeField &U){ 
+      std::cout << GridLogMessage << "[WilsonFlow] Computing Cloverleaf energy density for step " << step << std::endl;
+      out.push_back( energyDensityCloverleaf(t,U) );
+    });      
+  smear(V,U);
+  return out;
+}
+
+template <class Gimpl>
+std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityCloverleaf(const GaugeField& U, int measure_interval){
+  GaugeField V(U);
+  return flowMeasureEnergyDensityCloverleaf(V,U, measure_interval);
+}
+
+
 
 //#define WF_TIMING 
-
-
-
 template <class Gimpl>
-void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const {
+void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const{
   out = in;
-  for (unsigned int step = 1; step <= Nstep; step++) {
+  RealD taus = 0.;
+  for (unsigned int step = 1; step <= Nstep; step++) { //step indicates the number of smearing steps applied at the time of measurement
     auto start = std::chrono::high_resolution_clock::now();
-    evolve_step(out);
+    evolve_step(out, taus);
     auto end = std::chrono::high_resolution_clock::now();
     std::chrono::duration<double> diff = end - start;
 #ifdef WF_TIMING
     std::cout << "Time to evolve " << diff.count() << " s\n";
 #endif
-    std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "
-		  << step << "  " << tau(step) << "  " 
-	      << energyDensityPlaquette(step,out) << std::endl;
-    if( step % measure_interval == 0){
-      std::cout << GridLogMessage << "[WilsonFlow] Top. charge           : "
-		<< step << "  " 
-		<< WilsonLoops<PeriodicGimplR>::TopologicalCharge(out) << std::endl;
-    }
+    //Perform measurements
+    for(auto const &meas : functions)
+      if( step % meas.first == 0 ) meas.second(step,taus,out);
   }
 }
 
 template <class Gimpl>
-void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, RealD maxTau){
+void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, RealD maxTau) const{
   out = in;
-  taus = epsilon;
+  RealD taus = 0.;
+  RealD eps = epsilon;
   unsigned int step = 0;
   do{
     step++;
     //std::cout << GridLogMessage << "Evolution time :"<< taus << std::endl;
-    evolve_step_adaptive(out, maxTau);
-    std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "
-		  << step << "  " << taus << "  "
-	      << energyDensityPlaquette(out) << std::endl;
-    if( step % measure_interval == 0){
-      std::cout << GridLogMessage << "[WilsonFlow] Top. charge           : "
-		<< step << "  " 
-		<< WilsonLoops<PeriodicGimplR>::TopologicalCharge(out) << std::endl;
-    }
+    evolve_step_adaptive(out, taus, eps, maxTau);
+    //Perform measurements
+    for(auto const &meas : functions)
+      if( step % meas.first == 0 ) meas.second(step,taus,out);
   } while (taus < maxTau);
-
-
-
 }
 
+template <class Gimpl>
+void WilsonFlow<Gimpl>::setDefaultMeasurements(int topq_meas_interval){
+  addMeasurement(1, [](int step, RealD t, const typename Gimpl::GaugeField &U){
+      std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "  << step << "  " << t << "  " << energyDensityPlaquette(t,U) << std::endl;
+    });
+  addMeasurement(topq_meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
+      std::cout << GridLogMessage << "[WilsonFlow] Top. charge           : "  << step << "  " << WilsonLoops<Gimpl>::TopologicalCharge(U) << std::endl;
+    });
+}
+
+
 NAMESPACE_END(Grid);
 

Grid/qcd/utils/CovariantCshift.h

@@ -88,6 +88,12 @@ namespace PeriodicBC {
     return CovShiftBackward(Link,mu,arg);
   }
 
+  //Boundary-aware C-shift of gauge links / gauge transformation matrices
+  template<class gauge> Lattice<gauge>
+  CshiftLink(const Lattice<gauge> &Link, int mu, int shift)
+  {
+    return Cshift(Link, mu, shift);
+  }
 
 }
 
@@ -158,6 +164,9 @@ namespace ConjugateBC {
     //    std::cout<<"Gparity::CovCshiftBackward mu="<<mu<<std::endl;
     return Cshift(tmp,mu,-1);// moves towards positive mu
   }
+
+  //Out(x) = U^dag_\mu(x-mu)  | x_\mu != 0
+  //       = U^T_\mu(L-1)  | x_\mu == 0
   template<class gauge> Lattice<gauge>
   CovShiftIdentityBackward(const Lattice<gauge> &Link, int mu) {
     GridBase *grid = Link.Grid();
@@ -176,6 +185,9 @@ namespace ConjugateBC {
     return Link;
   }
 
+  //Out(x) = S_\mu(x+\hat\mu)  | x_\mu != L-1
+  //       = S*_\mu(0)  | x_\mu == L-1
+  //Note: While this is used for Staples it is also applicable for shifting gauge links or gauge transformation matrices
   template<class gauge> Lattice<gauge>
   ShiftStaple(const Lattice<gauge> &Link, int mu)
   {
@@ -208,6 +220,35 @@ namespace ConjugateBC {
     return CovShiftBackward(Link,mu,arg);
   }
 
+  //Boundary-aware C-shift of gauge links / gauge transformation matrices
+  //shift = 1
+  //Out(x) = U_\mu(x+\hat\mu)  | x_\mu != L-1
+  //       = U*_\mu(0)  | x_\mu == L-1
+  //shift = -1
+  //Out(x) = U_\mu(x-mu)  | x_\mu != 0
+  //       = U*_\mu(L-1)  | x_\mu == 0
+  template<class gauge> Lattice<gauge>
+  CshiftLink(const Lattice<gauge> &Link, int mu, int shift)
+  {
+    GridBase *grid = Link.Grid();
+    int Lmu = grid->GlobalDimensions()[mu] - 1;
+
+    Lattice<iScalar<vInteger>> coor(grid);
+    LatticeCoordinate(coor, mu);
+
+    Lattice<gauge> tmp(grid);
+    if(shift == 1){
+      tmp = Cshift(Link, mu, 1);
+      tmp = where(coor == Lmu, conjugate(tmp), tmp);
+      return tmp;
+    }else if(shift == -1){
+      tmp = Link;
+      tmp = where(coor == Lmu, conjugate(tmp), tmp);
+      return Cshift(tmp, mu, -1);
+    }else assert(0 && "Invalid shift value");
+    return tmp; //shuts up the compiler fussing about the return type
+  }
+
 }
 
 

Grid/qcd/utils/GaugeFix.h

@@ -40,27 +40,45 @@ public:
   typedef typename Gimpl::GaugeLinkField GaugeMat;
   typedef typename Gimpl::GaugeField GaugeLorentz;
 
-  static void GaugeLinkToLieAlgebraField(const std::vector<GaugeMat> &U,std::vector<GaugeMat> &A) {
-    for(int mu=0;mu<Nd;mu++){
-      Complex cmi(0.0,-1.0);
-      A[mu] = Ta(U[mu]) * cmi;
-    }
+  //A_\mu(x) = -i Ta(U_\mu(x) )   where Ta(U) = 1/2( U - U^dag ) - 1/2N tr(U - U^dag)  is the traceless antihermitian part. This is an O(A^3) approximation to the logarithm of U
+  static void GaugeLinkToLieAlgebraField(const GaugeMat &U, GaugeMat &A) {
+    Complex cmi(0.0,-1.0);
+    A = Ta(U) * cmi;
   }
-  static void DmuAmu(const std::vector<GaugeMat> &A,GaugeMat &dmuAmu,int orthog) {
+  
+  //The derivative of the Lie algebra field
+  static void DmuAmu(const std::vector<GaugeMat> &U, GaugeMat &dmuAmu,int orthog) {
+    GridBase* grid = U[0].Grid();
+    GaugeMat Ax(grid);
+    GaugeMat Axm1(grid);
+    GaugeMat Utmp(grid);
+
     dmuAmu=Zero();
     for(int mu=0;mu<Nd;mu++){
       if ( mu != orthog ) {
-	dmuAmu = dmuAmu + A[mu] - Cshift(A[mu],mu,-1);
+	//Rather than define functionality to work out how the BCs apply to A_\mu we simply use the BC-aware Cshift to the gauge links and compute A_\mu(x) and A_\mu(x-1) separately
+	//Ax = A_\mu(x)
+	GaugeLinkToLieAlgebraField(U[mu], Ax);
+	
+	//Axm1 = A_\mu(x_\mu-1)
+	Utmp = Gimpl::CshiftLink(U[mu], mu, -1);
+	GaugeLinkToLieAlgebraField(Utmp, Axm1);
+	
+	//Derivative
+	dmuAmu = dmuAmu + Ax - Axm1;
       }
     }
   }  
 
+  //Fix the gauge field Umu
+  //0 < alpha < 1 is related to the step size, cf https://arxiv.org/pdf/1405.5812.pdf
   static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
     SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
   }
   static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
+  //Fix the gauge field Umu and also return the gauge transformation from the original gauge field, xform
 
     GridBase *grid = Umu.Grid();
 
@@ -123,28 +141,25 @@ public:
       }
     }
     std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
-    if (err_on_no_converge) assert(0);
+    if (err_on_no_converge)
+      assert(0 && "Gauge fixing did not converge within the specified number of iterations");
   };
-  static Real SteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform,Real & alpha, GaugeMat & dmuAmu,int orthog) {
+  static Real SteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform, Real alpha, GaugeMat & dmuAmu,int orthog) {
     GridBase *grid = U[0].Grid();
 
-    std::vector<GaugeMat> A(Nd,grid);
     GaugeMat g(grid);
-
-    GaugeLinkToLieAlgebraField(U,A);
-    ExpiAlphaDmuAmu(A,g,alpha,dmuAmu,orthog);
-
+    ExpiAlphaDmuAmu(U,g,alpha,dmuAmu,orthog);
 
     Real vol = grid->gSites();
     Real trG = TensorRemove(sum(trace(g))).real()/vol/Nc;
 
     xform = g*xform ;
-    SU<Nc>::GaugeTransform(U,g);
+    SU<Nc>::GaugeTransform<Gimpl>(U,g);
 
     return trG;
   }
 
-  static Real FourierAccelSteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform,Real & alpha, GaugeMat & dmuAmu,int orthog) {
+  static Real FourierAccelSteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform, Real alpha, GaugeMat & dmuAmu,int orthog) {
 
     GridBase *grid = U[0].Grid();
 
@@ -159,11 +174,7 @@ public:
 
     GaugeMat g(grid);
     GaugeMat dmuAmu_p(grid);
-    std::vector<GaugeMat> A(Nd,grid);
-
-    GaugeLinkToLieAlgebraField(U,A);
-
-    DmuAmu(A,dmuAmu,orthog);
+    DmuAmu(U,dmuAmu,orthog);
 
     std::vector<int> mask(Nd,1);
     for(int mu=0;mu<Nd;mu++) if (mu==orthog) mask[mu]=0;
@@ -207,16 +218,16 @@ public:
     Real trG = TensorRemove(sum(trace(g))).real()/vol/Nc;
 
     xform = g*xform ;
-    SU<Nc>::GaugeTransform(U,g);
+    SU<Nc>::GaugeTransform<Gimpl>(U,g);
 
     return trG;
   }
 
-  static void ExpiAlphaDmuAmu(const std::vector<GaugeMat> &A,GaugeMat &g,Real & alpha, GaugeMat &dmuAmu,int orthog) {
+  static void ExpiAlphaDmuAmu(const std::vector<GaugeMat> &U,GaugeMat &g, Real alpha, GaugeMat &dmuAmu,int orthog) {
     GridBase *grid = g.Grid();
     Complex cialpha(0.0,-alpha);
     GaugeMat ciadmam(grid);
-    DmuAmu(A,dmuAmu,orthog);
+    DmuAmu(U,dmuAmu,orthog);
     ciadmam = dmuAmu*cialpha;
     SU<Nc>::taExp(ciadmam,g);
   }  

Grid/qcd/utils/SUn.h

@@ -694,32 +694,32 @@ public:
  * Adjoint rep gauge xform
  */
 
-  template<typename GaugeField,typename GaugeMat>
-  static void GaugeTransform( GaugeField &Umu, GaugeMat &g){
+  template<typename Gimpl>
+  static void GaugeTransform(typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
     GridBase *grid = Umu.Grid();
     conformable(grid,g.Grid());
 
-    GaugeMat U(grid);
-    GaugeMat ag(grid); ag = adj(g);
+    typename Gimpl::GaugeLinkField U(grid);
+    typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
 
     for(int mu=0;mu<Nd;mu++){
       U= PeekIndex<LorentzIndex>(Umu,mu);
-      U = g*U*Cshift(ag, mu, 1);
+      U = g*U*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
       PokeIndex<LorentzIndex>(Umu,U,mu);
     }
   }
-  template<typename GaugeMat>
-  static void GaugeTransform( std::vector<GaugeMat> &U, GaugeMat &g){
+  template<typename Gimpl>
+  static void GaugeTransform( std::vector<typename Gimpl::GaugeLinkField> &U, typename Gimpl::GaugeLinkField &g){
     GridBase *grid = g.Grid();
-    GaugeMat ag(grid); ag = adj(g);
+    typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
     for(int mu=0;mu<Nd;mu++){
-      U[mu] = g*U[mu]*Cshift(ag, mu, 1);
+      U[mu] = g*U[mu]*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
     }
   }
-  template<typename GaugeField,typename GaugeMat>
-  static void RandomGaugeTransform(GridParallelRNG &pRNG, GaugeField &Umu, GaugeMat &g){
+  template<typename Gimpl>
+  static void RandomGaugeTransform(GridParallelRNG &pRNG, typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
     LieRandomize(pRNG,g,1.0);
-    GaugeTransform(Umu,g);
+    GaugeTransform<Gimpl>(Umu,g);
   }
 
   // Projects the algebra components a lattice matrix (of dimension ncol*ncol -1 )

Grid/qcd/utils/WilsonLoops.h

@@ -125,6 +125,57 @@ public:
     return sumplaq / vol / faces / Nc; // Nd , Nc dependent... FIXME
   }
 
+  //////////////////////////////////////////////////
+  // sum over all spatial planes of plaquette
+  //////////////////////////////////////////////////
+  static void siteSpatialPlaquette(ComplexField &Plaq,
+                            const std::vector<GaugeMat> &U) {
+    ComplexField sitePlaq(U[0].Grid());
+    Plaq = Zero();
+    for (int mu = 1; mu < Nd-1; mu++) {
+      for (int nu = 0; nu < mu; nu++) {
+        traceDirPlaquette(sitePlaq, U, mu, nu);
+        Plaq = Plaq + sitePlaq;
+      }
+    }
+  }
+
+  ////////////////////////////////////
+  // sum over all x,y,z and over all spatial planes of plaquette
+  //////////////////////////////////////////////////
+  static std::vector<RealD> timesliceSumSpatialPlaquette(const GaugeLorentz &Umu) {
+    std::vector<GaugeMat> U(Nd, Umu.Grid());
+    // inefficient here
+    for (int mu = 0; mu < Nd; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    ComplexField Plaq(Umu.Grid());
+
+    siteSpatialPlaquette(Plaq, U);
+    typedef typename ComplexField::scalar_object sobj;
+    std::vector<sobj> Tq;
+    sliceSum(Plaq, Tq, Nd-1);
+
+    std::vector<Real> out(Tq.size());
+    for(int t=0;t<Tq.size();t++) out[t] = TensorRemove(Tq[t]).real();
+    return out;
+  }
+  
+  //////////////////////////////////////////////////
+  // average over all x,y,z and over all spatial planes of plaquette
+  //////////////////////////////////////////////////
+  static std::vector<RealD> timesliceAvgSpatialPlaquette(const GaugeLorentz &Umu) {
+    std::vector<RealD> sumplaq = timesliceSumSpatialPlaquette(Umu);
+    int Lt = Umu.Grid()->FullDimensions()[Nd-1];
+    assert(sumplaq.size() == Lt);
+    double vol = Umu.Grid()->gSites() / Lt;
+    double faces = (1.0 * (Nd - 1)* (Nd - 2)) / 2.0;
+    for(int t=0;t<Lt;t++)
+      sumplaq[t] = sumplaq[t] / vol / faces / Nc; // Nd , Nc dependent... FIXME
+    return sumplaq;
+  }
+
   //////////////////////////////////////////////////
   // average over all x,y,z the temporal loop
   //////////////////////////////////////////////////
@@ -362,11 +413,11 @@ public:
     GaugeMat u = PeekIndex<LorentzIndex>(Umu, mu);  // some redundant copies
     GaugeMat vu = v*u;
       //FS = 0.25*Ta(u*v + Cshift(vu, mu, -1));
-      FS = (u*v + Cshift(vu, mu, -1));
+      FS = (u*v + Gimpl::CshiftLink(vu, mu, -1));
       FS = 0.125*(FS - adj(FS));
   }
 
-  static Real TopologicalCharge(GaugeLorentz &U){
+  static Real TopologicalCharge(const GaugeLorentz &U){
     // 4d topological charge
     assert(Nd==4);
     // Bx = -iF(y,z), By = -iF(z,y), Bz = -iF(x,y)
@@ -389,6 +440,203 @@ public:
   }
 
 
+  //Clover-leaf Wilson loop combination for arbitrary mu-extent M and nu extent N,  mu >= nu
+  //cf  https://arxiv.org/pdf/hep-lat/9701012.pdf Eq 7  for 1x2 Wilson loop    
+  //Clockwise ordering
+  static void CloverleafMxN(GaugeMat &FS, const GaugeMat &Umu, const GaugeMat &Unu, int mu, int nu, int M, int N){  
+#define Fmu(A) Gimpl::CovShiftForward(Umu, mu, A)
+#define Bmu(A) Gimpl::CovShiftBackward(Umu, mu, A)
+#define Fnu(A) Gimpl::CovShiftForward(Unu, nu, A)
+#define Bnu(A) Gimpl::CovShiftBackward(Unu, nu, A)
+#define FmuI Gimpl::CovShiftIdentityForward(Umu, mu)
+#define BmuI Gimpl::CovShiftIdentityBackward(Umu, mu)
+#define FnuI Gimpl::CovShiftIdentityForward(Unu, nu)
+#define BnuI Gimpl::CovShiftIdentityBackward(Unu, nu)
+
+    //Upper right loop
+    GaugeMat tmp = BmuI;
+    for(int i=1;i<M;i++)
+      tmp = Bmu(tmp);
+    for(int j=0;j<N;j++)
+      tmp = Bnu(tmp);
+    for(int i=0;i<M;i++)
+      tmp = Fmu(tmp);
+    for(int j=0;j<N;j++)
+      tmp = Fnu(tmp);
+      
+    FS = tmp;
+
+    //Upper left loop
+    tmp = BnuI;
+    for(int j=1;j<N;j++)
+      tmp = Bnu(tmp);
+    for(int i=0;i<M;i++)
+      tmp = Fmu(tmp);
+    for(int j=0;j<N;j++)
+      tmp = Fnu(tmp);
+    for(int i=0;i<M;i++)
+      tmp = Bmu(tmp);
+      
+    FS = FS + tmp;
+
+    //Lower right loop
+    tmp = FnuI;
+    for(int j=1;j<N;j++)
+      tmp = Fnu(tmp);
+    for(int i=0;i<M;i++)
+      tmp = Bmu(tmp);
+    for(int j=0;j<N;j++)
+      tmp = Bnu(tmp);
+    for(int i=0;i<M;i++)
+      tmp = Fmu(tmp);
+      
+    FS = FS + tmp;
+
+    //Lower left loop
+    tmp = FmuI;
+    for(int i=1;i<M;i++)
+      tmp = Fmu(tmp);
+    for(int j=0;j<N;j++)
+      tmp = Fnu(tmp);
+    for(int i=0;i<M;i++)
+      tmp = Bmu(tmp);
+    for(int j=0;j<N;j++)
+      tmp = Bnu(tmp);
+
+    FS = FS + tmp;
+
+#undef Fmu
+#undef Bmu
+#undef Fnu
+#undef Bnu
+#undef FmuI
+#undef BmuI
+#undef FnuI
+#undef BnuI
+  }
+
+  //Field strength from MxN Wilson loop
+  //Note F_numu = - F_munu
+  static void FieldStrengthMxN(GaugeMat &FS, const GaugeLorentz &U, int mu, int nu, int M, int N){  
+    GaugeMat Umu = PeekIndex<LorentzIndex>(U, mu);
+    GaugeMat Unu = PeekIndex<LorentzIndex>(U, nu);
+    if(M == N){
+      GaugeMat F(Umu.Grid());
+      CloverleafMxN(F, Umu, Unu, mu, nu, M, N);
+      FS = 0.125 * ( F - adj(F) );
+    }else{
+      //Average over both orientations
+      GaugeMat horizontal(Umu.Grid()), vertical(Umu.Grid());
+      CloverleafMxN(horizontal, Umu, Unu, mu, nu, M, N);
+      CloverleafMxN(vertical, Umu, Unu, mu, nu, N, M);
+      FS = 0.0625 * ( horizontal - adj(horizontal) + vertical - adj(vertical) );
+    }
+  }
+
+  //Topological charge contribution from MxN Wilson loops
+  //cf  https://arxiv.org/pdf/hep-lat/9701012.pdf  Eq 6
+  //output is the charge by timeslice: sum over timeslices to obtain the total
+  static std::vector<Real> TimesliceTopologicalChargeMxN(const GaugeLorentz &U, int M, int N){
+    assert(Nd == 4);
+    std::vector<std::vector<GaugeMat*> > F(Nd,std::vector<GaugeMat*>(Nd,nullptr));
+    //Note F_numu = - F_munu
+    //hence we only need to loop over mu,nu,rho,sigma that aren't related by permuting mu,nu  or rho,sigma
+    //Use nu > mu
+    for(int mu=0;mu<Nd-1;mu++){
+      for(int nu=mu+1; nu<Nd; nu++){
+	F[mu][nu] = new GaugeMat(U.Grid());
+	FieldStrengthMxN(*F[mu][nu], U, mu, nu, M, N);
+      }
+    }
+    Real coeff = -1./(32 * M_PI*M_PI * M*M * N*N); //overall sign to match CPS and Grid conventions, possibly related to time direction = 3 vs 0
+
+    static const int combs[3][4] = { {0,1,2,3}, {0,2,1,3}, {0,3,1,2} };
+    static const int signs[3] = { 1, -1, 1 }; //epsilon_{mu nu rho sigma}
+
+    ComplexField fsum(U.Grid());
+    fsum = Zero();
+    for(int c=0;c<3;c++){
+      int mu = combs[c][0], nu = combs[c][1], rho = combs[c][2], sigma = combs[c][3];
+      int eps = signs[c];
+      fsum = fsum + (8. * coeff * eps) * trace( (*F[mu][nu]) * (*F[rho][sigma]) ); 
+    }
+
+    for(int mu=0;mu<Nd-1;mu++)
+      for(int nu=mu+1; nu<Nd; nu++)
+	delete F[mu][nu];
+    
+    typedef typename ComplexField::scalar_object sobj;
+    std::vector<sobj> Tq;
+    sliceSum(fsum, Tq, Nd-1);
+
+    std::vector<Real> out(Tq.size());
+    for(int t=0;t<Tq.size();t++) out[t] = TensorRemove(Tq[t]).real();
+    return out;
+  }
+  static Real TopologicalChargeMxN(const GaugeLorentz &U, int M, int N){
+    std::vector<Real> Tq = TimesliceTopologicalChargeMxN(U,M,N);
+    Real out(0);
+    for(int t=0;t<Tq.size();t++) out += Tq[t];
+    return out;
+  }
+
+  //Generate the contributions to the 5Li topological charge from Wilson loops of the following sizes
+  //Use coefficients from hep-lat/9701012
+  //1x1 : c1=(19.-55.*c5)/9.
+  //2x2 : c2=(1-64.*c5)/9.
+  //1x2 : c3=(-64.+640.*c5)/45.
+  //1x3 : c4=1./5.-2.*c5
+  //3x3 : c5=1./20.
+  //Output array outer index contains the loops in the above order
+  //Inner index is the time coordinate
+  static std::vector<std::vector<Real> > TimesliceTopologicalCharge5LiContributions(const GaugeLorentz &U){
+    static const int exts[5][2] = { {1,1}, {2,2}, {1,2}, {1,3}, {3,3} };       
+    std::vector<std::vector<Real> > out(5);
+    for(int i=0;i<5;i++){	
+      out[i] = TimesliceTopologicalChargeMxN(U,exts[i][0],exts[i][1]);
+    }
+    return out;
+  }   
+
+  static std::vector<Real> TopologicalCharge5LiContributions(const GaugeLorentz &U){   
+    static const int exts[5][2] = { {1,1}, {2,2}, {1,2}, {1,3}, {3,3} };
+    std::vector<Real> out(5);
+    std::cout << GridLogMessage << "Computing topological charge" << std::endl;
+    for(int i=0;i<5;i++){
+      out[i] = TopologicalChargeMxN(U,exts[i][0],exts[i][1]);
+      std::cout << GridLogMessage << exts[i][0] << "x" << exts[i][1] << " Wilson loop contribution " << out[i] << std::endl;
+    }
+    return out;
+  }
+
+  //Compute the 5Li topological charge
+  static std::vector<Real> TimesliceTopologicalCharge5Li(const GaugeLorentz &U){
+    std::vector<std::vector<Real> > loops = TimesliceTopologicalCharge5LiContributions(U);
+
+    double c5=1./20.;
+    double c4=1./5.-2.*c5;
+    double c3=(-64.+640.*c5)/45.;
+    double c2=(1-64.*c5)/9.;
+    double c1=(19.-55.*c5)/9.;
+
+    int Lt = loops[0].size();
+    std::vector<Real> out(Lt,0.);
+    for(int t=0;t<Lt;t++)
+      out[t] += c1*loops[0][t] + c2*loops[1][t] + c3*loops[2][t] + c4*loops[3][t] + c5*loops[4][t];
+    return out;
+  }
+
+  static Real TopologicalCharge5Li(const GaugeLorentz &U){
+    std::vector<Real> Qt = TimesliceTopologicalCharge5Li(U);
+    Real Q = 0.;
+    for(int t=0;t<Qt.size();t++) Q += Qt[t];
+    std::cout << GridLogMessage << "5Li Topological charge: " << Q << std::endl;
+    return Q;
+  }
+
+
+
+
   //////////////////////////////////////////////////////
   // Similar to above for rectangle is required
   //////////////////////////////////////////////////////

Grid/stencil/SimpleCompressor.h

@@ -52,6 +52,11 @@ public:
     return arg;
   }
 };
+class SimpleStencilParams{
+public:
+  Coordinate dirichlet;
+  SimpleStencilParams() {};
+};
 
 NAMESPACE_END(Grid);
 

Grid/stencil/Stencil.h

@@ -133,6 +133,8 @@ class CartesianStencilAccelerator {
   int           _osites;
   StencilVector _directions;
   StencilVector _distances;
+  StencilVector _comms_send;
+  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -226,6 +228,8 @@ public:
     void * recv_buf;
     Integer to_rank;
     Integer from_rank;
+    Integer do_send;
+    Integer do_recv;
     Integer bytes;
   };
   struct Merge {
@@ -240,7 +244,20 @@ public:
     cobj * mpi_p;
     Integer buffer_size;
   };
-
+  struct CopyReceiveBuffer {
+    void * from_p;
+    void * to_p;
+    Integer bytes;
+  };
+  struct CachedTransfer {
+    Integer direction;
+    Integer OrthogPlane;
+    Integer DestProc;
+    Integer bytes;
+    Integer lane;
+    Integer cb;
+    void *recv_buf;
+  };
 
 protected:
   GridBase *                        _grid;
@@ -271,7 +288,8 @@ public:
   std::vector<Merge> MergersSHM;
   std::vector<Decompress> Decompressions;
   std::vector<Decompress> DecompressionsSHM;
-
+  std::vector<CopyReceiveBuffer> CopyReceiveBuffers ;
+  std::vector<CachedTransfer> CachedTransfers;
   ///////////////////////////////////////////////////////////
   // Unified Comms buffers for all directions
   ///////////////////////////////////////////////////////////
@@ -284,29 +302,6 @@ public:
   int u_comm_offset;
   int _unified_buffer_size;
 
-  /////////////////////////////////////////
-  // Timing info; ugly; possibly temporary
-  /////////////////////////////////////////
-  double commtime;
-  double mpi3synctime;
-  double mpi3synctime_g;
-  double shmmergetime;
-  double gathertime;
-  double gathermtime;
-  double halogtime;
-  double mergetime;
-  double decompresstime;
-  double comms_bytes;
-  double shm_bytes;
-  double splicetime;
-  double nosplicetime;
-  double calls;
-  std::vector<double> comm_bytes_thr;
-  std::vector<double> shm_bytes_thr;
-  std::vector<double> comm_time_thr;
-  std::vector<double> comm_enter_thr;
-  std::vector<double> comm_leave_thr;
-
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
@@ -333,11 +328,12 @@ public:
   //////////////////////////////////////////
   // Comms packet queue for asynch thread
   // Use OpenMP Tasks for cleaner ???
+  // must be called *inside* parallel region
   //////////////////////////////////////////
+  /*
   void CommunicateThreaded()
   {
 #ifdef GRID_OMP
-    // must be called in parallel region
     int mythread = omp_get_thread_num();
     int nthreads = CartesianCommunicator::nCommThreads;
 #else
@@ -346,65 +342,29 @@ public:
 #endif
     if (nthreads == -1) nthreads = 1;
     if (mythread < nthreads) {
-      comm_enter_thr[mythread] = usecond();
       for (int i = mythread; i < Packets.size(); i += nthreads) {
 	uint64_t bytes = _grid->StencilSendToRecvFrom(Packets[i].send_buf,
 						      Packets[i].to_rank,
 						      Packets[i].recv_buf,
 						      Packets[i].from_rank,
 						      Packets[i].bytes,i);
-	comm_bytes_thr[mythread] += bytes;
-	shm_bytes_thr[mythread] += 2*Packets[i].bytes-bytes; // Send + Recv.
-
       }
-      comm_leave_thr[mythread]= usecond();
-      comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
     }
   }
-
-  void CollateThreads(void)
-  {
-    int nthreads = CartesianCommunicator::nCommThreads;
-    double first=0.0;
-    double last =0.0;
-
-    for(int t=0;t<nthreads;t++) {
-
-      double t0 = comm_enter_thr[t];
-      double t1 = comm_leave_thr[t];
-      comms_bytes+=comm_bytes_thr[t];
-      shm_bytes  +=shm_bytes_thr[t];
-
-      comm_enter_thr[t] = 0.0;
-      comm_leave_thr[t] = 0.0;
-      comm_time_thr[t]   = 0.0;
-      comm_bytes_thr[t]=0;
-      shm_bytes_thr[t]=0;
-
-      if ( first == 0.0 ) first = t0;                   // first is t0
-      if ( (t0 > 0.0) && ( t0 < first ) ) first = t0;   // min time seen
-
-      if ( t1 > last ) last = t1;                       // max time seen
-
-    }
-    commtime+= last-first;
-  }
+  */
   ////////////////////////////////////////////////////////////////////////
   // Non blocking send and receive. Necessarily parallel.
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
     reqs.resize(Packets.size());
-    commtime-=usecond();
     for(int i=0;i<Packets.size();i++){
-      uint64_t bytes=_grid->StencilSendToRecvFromBegin(reqs[i],
-						     Packets[i].send_buf,
-						     Packets[i].to_rank,
-						     Packets[i].recv_buf,
-						     Packets[i].from_rank,
-						     Packets[i].bytes,i);
-      comms_bytes+=bytes;
-      shm_bytes  +=2*Packets[i].bytes-bytes;
+      _grid->StencilSendToRecvFromBegin(reqs[i],
+					Packets[i].send_buf,
+					Packets[i].to_rank,Packets[i].do_send,
+					Packets[i].recv_buf,
+					Packets[i].from_rank,Packets[i].do_recv,
+					Packets[i].bytes,i);
     }
   }
 
@@ -413,7 +373,6 @@ public:
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
-    commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
@@ -421,28 +380,27 @@ public:
   void Communicate(void)
   {
     if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
-      thread_region {
-	// must be called in parallel region
-	int mythread  = thread_num();
-	int maxthreads= thread_max();
-	int nthreads = CartesianCommunicator::nCommThreads;
-	assert(nthreads <= maxthreads);
-	if (nthreads == -1) nthreads = 1;
-	if (mythread < nthreads) {
-	  for (int i = mythread; i < Packets.size(); i += nthreads) {
-	    double start = usecond();
-	    uint64_t bytes= _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-							 Packets[i].to_rank,
-							 Packets[i].recv_buf,
-							 Packets[i].from_rank,
-							 Packets[i].bytes,i);
-	    comm_bytes_thr[mythread] += bytes;
-	    shm_bytes_thr[mythread]  += Packets[i].bytes - bytes;
-	    comm_time_thr[mythread]  += usecond() - start;
-	  }
-	}
-      }
-    } else { // Concurrent and non-threaded asynch calls to MPI
+      /////////////////////////////////////////////////////////
+      // several way threaded on different communicators.
+      // Cannot combine with Dirichlet operators
+      // This scheme is needed on Intel Omnipath for best performance
+      // Deprecate once there are very few omnipath clusters
+      /////////////////////////////////////////////////////////
+      int nthreads = CartesianCommunicator::nCommThreads;
+      int old = GridThread::GetThreads();
+      GridThread::SetThreads(nthreads);
+      thread_for(i,Packets.size(),{
+	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
+				       Packets[i].to_rank,Packets[i].do_send,
+				       Packets[i].recv_buf,
+				       Packets[i].from_rank,Packets[i].do_recv,
+				       Packets[i].bytes,i);
+      });
+      GridThread::SetThreads(old);
+    } else { 
+      /////////////////////////////////////////////////////////
+      // Concurrent and non-threaded asynch calls to MPI
+      /////////////////////////////////////////////////////////
       std::vector<std::vector<CommsRequest_t> > reqs;
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
@@ -484,31 +442,23 @@ public:
       sshift[1] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Odd);
       if ( sshift[0] == sshift[1] ) {
 	if (splice_dim) {
-	  splicetime-=usecond();
-	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
+	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
-	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
+	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp;
-	  nosplicetime+=usecond();
 	}
       } else {
 	if(splice_dim){
-	  splicetime-=usecond();
 	  // if checkerboard is unfavourable take two passes
 	  // both with block stride loop iteration
-	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx);
-	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx);
+	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx,point);
+	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
-	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
-	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
+	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx,point);
+	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  nosplicetime+=usecond();
 	}
       }
     }
@@ -518,13 +468,10 @@ public:
   template<class compressor>
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
-    mpi3synctime_g-=usecond();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime_g+=usecond();
 
     // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
-    halogtime-=usecond();
 
     u_comm_offset=0;
 
@@ -538,7 +485,6 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
-    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -551,14 +497,70 @@ public:
     Mergers.resize(0);
     MergersSHM.resize(0);
     Packets.resize(0);
-    calls++;
+    CopyReceiveBuffers.resize(0);
+    CachedTransfers.resize(0);
   }
-  void AddPacket(void *xmit,void * rcv, Integer to,Integer from,Integer bytes){
+  void AddCopy(void *from,void * to, Integer bytes)
+  {
+    CopyReceiveBuffer obj;
+    obj.from_p = from;
+    obj.to_p = to;
+    obj.bytes= bytes;
+    CopyReceiveBuffers.push_back(obj);
+  }
+  void CommsCopy()
+  {
+    //    These are device resident MPI buffers.
+    for(int i=0;i<CopyReceiveBuffers.size();i++){
+      cobj *from=(cobj *)CopyReceiveBuffers[i].from_p;
+      cobj *to  =(cobj *)CopyReceiveBuffers[i].to_p;
+      Integer words = CopyReceiveBuffers[i].bytes/sizeof(cobj);
+
+      accelerator_forNB(j, words, cobj::Nsimd(), {
+	  coalescedWrite(to[j] ,coalescedRead(from [j]));
+      });
+    }
+  }
+  
+  Integer CheckForDuplicate(Integer direction, Integer OrthogPlane, Integer DestProc, void *recv_buf,Integer lane,Integer bytes,Integer cb)
+  {
+    CachedTransfer obj;
+    obj.direction   = direction;
+    obj.OrthogPlane = OrthogPlane;
+    obj.DestProc    = DestProc;
+    obj.recv_buf    = recv_buf;
+    obj.lane        = lane;
+    obj.bytes       = bytes;
+    obj.cb          = cb;
+
+    for(int i=0;i<CachedTransfers.size();i++){
+      if (   (CachedTransfers[i].direction  ==direction)
+	   &&(CachedTransfers[i].OrthogPlane==OrthogPlane)
+	   &&(CachedTransfers[i].DestProc   ==DestProc)
+	   &&(CachedTransfers[i].bytes      ==bytes)
+	   &&(CachedTransfers[i].lane       ==lane)
+	   &&(CachedTransfers[i].cb         ==cb)
+	     ){
+
+	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
+	return 1;
+      }
+    }
+
+    CachedTransfers.push_back(obj);
+    return 0;
+  }
+  void AddPacket(void *xmit,void * rcv,
+		 Integer to, Integer do_send,
+		 Integer from, Integer do_recv,
+		 Integer bytes){
     Packet p;
     p.send_buf = xmit;
     p.recv_buf = rcv;
     p.to_rank  = to;
     p.from_rank= from;
+    p.do_send  = do_send;
+    p.do_recv  = do_recv;
     p.bytes    = bytes;
     Packets.push_back(p);
   }
@@ -578,22 +580,17 @@ public:
     mv.push_back(m);
   }
   template<class decompressor>  void CommsMerge(decompressor decompress)    {
+    CommsCopy();
     CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    mpi3synctime-=usecond();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime+=usecond();
-    shmmergetime-=usecond();
     CommsMerge(decompress,MergersSHM,DecompressionsSHM);
-    shmmergetime+=usecond();
   }
 
   template<class decompressor>
-  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
-
-    
-    mergetime-=usecond();
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd)
+  {
     for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
@@ -603,9 +600,7 @@ public:
 	  decompress.Exchange(mp,vp0,vp1,type,o);
       });
     }
-    mergetime+=usecond();
 
-    decompresstime-=usecond();
     for(int i=0;i<dd.size();i++){
       auto kp = dd[i].kernel_p;
       auto mp = dd[i].mpi_p;
@@ -613,7 +608,6 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
-    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -646,22 +640,60 @@ public:
 	}
       }
       if(local == 0) {
-	surface_list.push_back(site);
+	for(int s=0;s<Ls;s++){
+	  surface_list.push_back(site*Ls+s);
+	}
+      }
+    }
+  }
+  /// Introduce a block structure and switch off comms on boundaries
+  void DirichletBlock(const Coordinate &dirichlet_block)
+  {
+    for(int ii=0;ii<this->_npoints;ii++){
+      int dimension    = this->_directions[ii];
+      int displacement = this->_distances[ii];
+      int gd = _grid->_gdimensions[dimension];
+      int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      int pc = _grid->_processor_coor[dimension];
+      int ld = fd/pd;
+      ///////////////////////////////////////////
+      // Figure out dirichlet send and receive
+      // on this leg of stencil.
+      ///////////////////////////////////////////
+      int comm_dim        = _grid->_processors[dimension] >1 ;
+      int block = dirichlet_block[dimension];
+      this->_comms_send[ii] = comm_dim;
+      this->_comms_recv[ii] = comm_dim;
+      if ( block && comm_dim ) {
+	assert(abs(displacement) < ld );
+	// Quiesce communication across block boundaries
+	if( displacement > 0 ) {
+	  // High side, low side
+	  // | <--B--->|
+	  // |    |    |
+	  //           noR
+	  // noS
+	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_recv[ii] = 0;
+	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_send[ii] = 0;
+	} else {
+	  // High side, low side
+	  // | <--B--->|
+	  // |    |    |
+	  //           noS
+	  // noR
+	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_send[ii] = 0;
+	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_recv[ii] = 0;
+	}
       }
     }
   }
-
   CartesianStencil(GridBase *grid,
 		   int npoints,
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
 		   Parameters p)
-    : shm_bytes_thr(npoints),
-      comm_bytes_thr(npoints),
-      comm_enter_thr(npoints),
-      comm_leave_thr(npoints),
-      comm_time_thr(npoints)
   {
     face_table_computed=0;
     _grid    = grid;
@@ -675,8 +707,12 @@ public:
     this->_simd_layout = _grid->_simd_layout; // copy simd_layout to give access to Accelerator Kernels
     this->_directions = StencilVector(directions);
     this->_distances  = StencilVector(distances);
+    this->_comms_send.resize(npoints); 
+    this->_comms_recv.resize(npoints); 
     this->same_node.resize(npoints);
 
+    if ( p.dirichlet.size() ) DirichletBlock(p.dirichlet); // comms send/recv set up
+
     _unified_buffer_size=0;
     surface_list.resize(0);
 
@@ -693,15 +729,16 @@ public:
       int displacement = distances[i];
       int shift = displacement;
 
+      int gd = _grid->_gdimensions[dimension];
       int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      //      int ld = gd/pd;
       int rd = _grid->_rdimensions[dimension];
+      int pc = _grid->_processor_coor[dimension];
       this->_permute_type[point]=_grid->PermuteType(dimension);
 
       this->_checkerboard = checkerboard;
 
-      //////////////////////////
-      // the permute type
-      //////////////////////////
       int simd_layout     = _grid->_simd_layout[dimension];
       int comm_dim        = _grid->_processors[dimension] >1 ;
       int splice_dim      = _grid->_simd_layout[dimension]>1 && (comm_dim);
@@ -710,7 +747,6 @@ public:
       assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
 
       int sshift[2];
-
       //////////////////////////
       // Underlying approach. For each local site build
       // up a table containing the npoint "neighbours" and whether they
@@ -811,6 +847,7 @@ public:
     GridBase *grid=_grid;
     const int Nsimd = grid->Nsimd();
 
+    int comms_recv      = this->_comms_recv[point];
     int fd              = _grid->_fdimensions[dimension];
     int ld              = _grid->_ldimensions[dimension];
     int rd              = _grid->_rdimensions[dimension];
@@ -836,12 +873,14 @@ public:
     for(int x=0;x<rd;x++){
 
       int permute_type=grid->PermuteType(dimension);
+      int permute_slice;
 
       int sx        =  (x+sshift)%rd;
 
       int offnode = 0;
       if ( simd_layout > 1 ) {
 
+	permute_slice=1;
 	for(int i=0;i<Nsimd;i++){
 
 	  int inner_bit = (Nsimd>>(permute_type+1));
@@ -858,6 +897,7 @@ public:
       } else {
 	int comm_proc = ((x+sshift)/rd)%pd;
 	offnode = (comm_proc!= 0);
+	permute_slice=0;
       }
 
       int wraparound=0;
@@ -867,25 +907,31 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-      if (!offnode) {
 
-	int permute_slice=0;
+      // Wrap locally dirichlet support case OR node local
+      if ( offnode==0 ) {
+
+	permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-
+	
       } else {
 
+	if ( comms_recv ) {
+
+	  ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
+
+	} else { 
+
+	  CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
+
+	}
+
+      }
+      
+      if ( offnode ) {
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
-
-	//	int rank           = grid->_processor;
-	//	int recv_from_rank;
-	//	int xmit_to_rank;
-
-	int unified_buffer_offset = _unified_buffer_size;
 	_unified_buffer_size    += words;
-
-	ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
-
       }
     }
   }
@@ -984,11 +1030,14 @@ public:
   }
 
   template<class compressor>
-  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
+  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx, int point)
   {
     typedef typename cobj::vector_type vector_type;
     typedef typename cobj::scalar_type scalar_type;
 
+    int comms_send   = this->_comms_send[point] ;
+    int comms_recv   = this->_comms_recv[point] ;
+
     assert(rhs.Grid()==_grid);
     //	  conformable(_grid,rhs.Grid());
 
@@ -1011,78 +1060,93 @@ public:
 
       int sx        = (x+sshift)%rd;
       int comm_proc = ((x+sshift)/rd)%pd;
-
+      
       if (comm_proc) {
-
+	
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
 
 	int bytes =  words * compress.CommDatumSize();
 
 	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
-	if ( !face_table_computed ) {
-	  face_table.resize(face_idx+1);
-	  std::vector<std::pair<int,int> >  face_table_host ;
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
-	  face_table[face_idx].resize(face_table_host.size());
-	  acceleratorCopyToDevice(&face_table_host[0],
-				  &face_table[face_idx][0],
-				  face_table[face_idx].size()*sizeof(face_table_host[0]));
-	}
+	int comm_off = u_comm_offset;
 
-	//      	int rank           = _grid->_processor;
 	int recv_from_rank;
 	int xmit_to_rank;
+	cobj *recv_buf;
+	cobj *send_buf;
 	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
 	assert (xmit_to_rank   != _grid->ThisRank());
 	assert (recv_from_rank != _grid->ThisRank());
 
-	cobj *recv_buf;
-	if ( compress.DecompressionStep() ) {
-	  recv_buf=u_simd_recv_buf[0];
-	} else {
-	  recv_buf=this->u_recv_buf_p;
+	if( comms_send ) {
+
+	  if ( !face_table_computed ) {
+	    face_table.resize(face_idx+1);
+	    std::vector<std::pair<int,int> >  face_table_host ;
+	    Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
+	    face_table[face_idx].resize(face_table_host.size());
+	    acceleratorCopyToDevice(&face_table_host[0],
+				    &face_table[face_idx][0],
+				    face_table[face_idx].size()*sizeof(face_table_host[0]));
+	  }
+
+
+	  if ( compress.DecompressionStep() ) {
+	    recv_buf=u_simd_recv_buf[0];
+	  } else {
+	    recv_buf=this->u_recv_buf_p;
+	  }
+
+	  send_buf = this->u_send_buf_p; // Gather locally, must send
+	
+	  ////////////////////////////////////////////////////////
+	  // Gather locally
+	  ////////////////////////////////////////////////////////
+	  assert(send_buf!=NULL);
+
+	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,comm_off,so);
 	}
 
-	cobj *send_buf;
-	send_buf = this->u_send_buf_p; // Gather locally, must send
+	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,bytes,cbmask);
+	if ( (!duplicate) ) { // Force comms for now
 
-	////////////////////////////////////////////////////////
-	// Gather locally
-	////////////////////////////////////////////////////////
-	gathertime-=usecond();
-	assert(send_buf!=NULL);
-	Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so); face_idx++;
-	gathertime+=usecond();
-
-	///////////////////////////////////////////////////////////
-	// Build a list of things to do after we synchronise GPUs
-	// Start comms now???
-	///////////////////////////////////////////////////////////
-	AddPacket((void *)&send_buf[u_comm_offset],
-		  (void *)&recv_buf[u_comm_offset],
-		  xmit_to_rank,
-		  recv_from_rank,
-		  bytes);
-
-	if ( compress.DecompressionStep() ) {
-	  AddDecompress(&this->u_recv_buf_p[u_comm_offset],
-			&recv_buf[u_comm_offset],
+	  ///////////////////////////////////////////////////////////
+	  // Build a list of things to do after we synchronise GPUs
+	  // Start comms now???
+	  ///////////////////////////////////////////////////////////
+	  AddPacket((void *)&send_buf[comm_off],
+		    (void *)&recv_buf[comm_off],
+		    xmit_to_rank, comms_send,
+		    recv_from_rank, comms_recv,
+		    bytes);
+	}
+	
+	if ( compress.DecompressionStep()  && comms_recv ) {
+	  AddDecompress(&this->u_recv_buf_p[comm_off],
+			&recv_buf[comm_off],
 			words,Decompressions);
 	}
+	
 	u_comm_offset+=words;
+	face_idx++;
+
       }
     }
     return 0;
   }
 
   template<class compressor>
-  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
+  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx,int point)
   {
     const int Nsimd = _grid->Nsimd();
 
     const int maxl =2;// max layout in a direction
+
+    int comms_send   = this->_comms_send[point] ;
+    int comms_recv   = this->_comms_recv[point] ;
+
     int fd = _grid->_fdimensions[dimension];
     int rd = _grid->_rdimensions[dimension];
     int ld = _grid->_ldimensions[dimension];
@@ -1097,7 +1161,6 @@ public:
 
 
     int permute_type=_grid->PermuteType(dimension);
-    //    std::cout << "SimdNew permute type "<<permute_type<<std::endl;
 
     ///////////////////////////////////////////////
     // Simd direction uses an extract/merge pair
@@ -1131,8 +1194,9 @@ public:
 
       if ( any_offnode ) {
 
+	int comm_off = u_comm_offset;
 	for(int i=0;i<maxl;i++){
-	  spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
+	  spointers[i] = (cobj *) &u_simd_send_buf[i][comm_off];
 	}
 
 	int sx   = (x+sshift)%rd;
@@ -1141,18 +1205,17 @@ public:
 	  face_table.resize(face_idx+1);
 	  std::vector<std::pair<int,int> >  face_table_host ;
 				
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
+	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
 	  face_table[face_idx].resize(face_table_host.size());
 	  acceleratorCopyToDevice(&face_table_host[0],
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
 	}
-	gathermtime-=usecond();
 
-	Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
+	if ( comms_send || comms_recv )
+	  Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
 	face_idx++;
 
-	gathermtime+=usecond();
 	//spointers[0] -- low
 	//spointers[1] -- high
 
@@ -1169,8 +1232,8 @@ public:
 	  int nbr_plane = nbr_ic;
 	  assert (sx == nbr_ox);
 
-	  auto rp = &u_simd_recv_buf[i        ][u_comm_offset];
-	  auto sp = &u_simd_send_buf[nbr_plane][u_comm_offset];
+	  auto rp = &u_simd_recv_buf[i        ][comm_off];
+	  auto sp = &u_simd_send_buf[nbr_plane][comm_off];
 
 	  if(nbr_proc){
 
@@ -1181,8 +1244,13 @@ public:
 
 	    rpointers[i] = rp;
 
-	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
-
+	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,bytes,cbmask);
+	    if ( !duplicate  ) { 
+	      AddPacket((void *)sp,(void *)rp,
+			xmit_to_rank,comms_send,
+			recv_from_rank,comms_recv,
+			bytes);
+	    }
 
 	  } else {
 
@@ -1191,9 +1259,12 @@ public:
 	  }
 	}
 
-	AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
+	if ( comms_recv ) {
+	  AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
+	}
 
 	u_comm_offset     +=buffer_size;
+
       }
     }
     return 0;

Grid/tensors/Tensor_exp.h

@@ -55,7 +55,7 @@ template<class vtype, int N> accelerator_inline iVector<vtype, N> Exponentiate(c
 
 
 // Specialisation: Cayley-Hamilton exponential for SU(3)
-#ifndef GRID_CUDA
+#ifndef GRID_ACCELERATED
 template<class vtype, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0>::type * =nullptr> 
 accelerator_inline iMatrix<vtype,3> Exponentiate(const iMatrix<vtype,3> &arg, RealD alpha  , Integer Nexp = DEFAULT_MAT_EXP )
 {

Grid/tensors/Tensor_extract_merge.h

@@ -208,5 +208,46 @@ void merge(vobj &vec,const ExtractPointerArray<sobj> &extracted, int offset)
 }
 
 
+
+//////////////////////////////////////////////////////////////////////////////////
+//Copy a single lane of a SIMD tensor type from one object to another
+//Output object must be of the same tensor type but may be of a different precision (i.e. it can have a different root data type)
+///////////////////////////////////////////////////////////////////////////////////
+template<class vobjOut, class vobjIn>
+accelerator_inline 
+void copyLane(vobjOut & __restrict__ vecOut, int lane_out, const vobjIn & __restrict__ vecIn, int lane_in)
+{
+  static_assert( std::is_same<typename vobjOut::DoublePrecision, typename vobjIn::DoublePrecision>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+
+  typedef typename vobjOut::vector_type ovector_type;  
+  typedef typename vobjIn::vector_type ivector_type;  
+  constexpr int owords=sizeof(vobjOut)/sizeof(ovector_type);
+  constexpr int iwords=sizeof(vobjIn)/sizeof(ivector_type);
+  static_assert( owords == iwords, "copyLane: Expected number of vector words in input and output objects to be equal" );
+
+  typedef typename vobjOut::scalar_type oscalar_type;  
+  typedef typename vobjIn::scalar_type iscalar_type;  
+  typedef typename ExtractTypeMap<oscalar_type>::extract_type oextract_type;
+  typedef typename ExtractTypeMap<iscalar_type>::extract_type iextract_type;
+
+  typedef oextract_type * opointer;
+  typedef iextract_type * ipointer;
+
+  constexpr int oNsimd=ovector_type::Nsimd();
+  constexpr int iNsimd=ivector_type::Nsimd();
+
+  iscalar_type itmp;
+  oscalar_type otmp;
+
+  opointer __restrict__  op = (opointer)&vecOut;
+  ipointer __restrict__  ip = (ipointer)&vecIn;
+  for(int w=0;w<owords;w++){
+    memcpy( (char*)&itmp, (char*)(ip + lane_in + iNsimd*w), sizeof(iscalar_type) );
+    otmp = itmp; //potential precision change
+    memcpy( (char*)(op + lane_out + oNsimd*w), (char*)&otmp, sizeof(oscalar_type) );
+  }
+}
+
+
 NAMESPACE_END(Grid);
 

Grid/threads/Accelerator.cc

@@ -6,9 +6,17 @@ uint32_t accelerator_threads=2;
 uint32_t acceleratorThreads(void)       {return accelerator_threads;};
 void     acceleratorThreads(uint32_t t) {accelerator_threads = t;};
 
+#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
+#define ENV_LOCAL_RANK_SLURM   "SLURM_LOCALID"
+#define ENV_RANK_SLURM         "SLURM_PROCID"
+#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
+
 #ifdef GRID_CUDA
 cudaDeviceProp *gpu_props;
 cudaStream_t copyStream;
+cudaStream_t cpuStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -17,12 +25,6 @@ void acceleratorInit(void)
 
   char * localRankStr = NULL;
   int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
-#define ENV_LOCAL_RANK_SLURM   "SLURM_LOCALID"
-#define ENV_RANK_SLURM         "SLURM_PROCID"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_SLURM  )) != NULL) { world_rank = atoi(localRankStr);}
@@ -97,6 +99,7 @@ void acceleratorInit(void)
 
   cudaSetDevice(device);
   cudaStreamCreate(&copyStream);
+  cudaStreamCreate(&cpuStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -111,6 +114,7 @@ void acceleratorInit(void)
 #ifdef GRID_HIP
 hipDeviceProp_t *gpu_props;
 hipStream_t copyStream;
+hipStream_t cpuStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -119,10 +123,6 @@ void acceleratorInit(void)
 
   char * localRankStr = NULL;
   int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
-#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
   // We extract the local rank initialization using an environment variable
   if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL)
   {
@@ -134,8 +134,10 @@ void acceleratorInit(void)
   }
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
+  if ((localRankStr = getenv(ENV_RANK_SLURM  )) != NULL) { world_rank = atoi(localRankStr);}
 
-  printf("world_rank %d has %d devices\n",world_rank,nDevices);
+  if ( world_rank == 0 ) 
+    printf("world_rank %d has %d devices\n",world_rank,nDevices);
   size_t totalDeviceMem=0;
   for (int i = 0; i < nDevices; i++) {
 
@@ -181,6 +183,7 @@ void acceleratorInit(void)
 #endif
   hipSetDevice(device);
   hipStreamCreate(&copyStream);
+  hipStreamCreate(&cpuStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -208,10 +211,7 @@ void acceleratorInit(void)
   
   char * localRankStr = NULL;
   int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
-#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
+
   // We extract the local rank initialization using an environment variable
   if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL)
   {

Grid/threads/Accelerator.h

@@ -107,6 +107,7 @@ void     acceleratorInit(void);
 
 extern int acceleratorAbortOnGpuError;
 extern cudaStream_t copyStream;
+extern cudaStream_t cpuStream;
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
@@ -134,7 +135,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_threads(nsimd,acceleratorThreads(),1);			\
     dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
-    LambdaApply<<<cu_blocks,cu_threads>>>(num1,num2,nsimd,lambda);	\
+    LambdaApply<<<cu_blocks,cu_threads,0,cpuStream>>>(num1,num2,nsimd,lambda);	\
   }
 
 #define accelerator_for6dNB(iter1, num1,				\
@@ -153,7 +154,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_blocks (num1,num2,num3);					\
     dim3 cu_threads(num4,num5,num6);					\
-    Lambda6Apply<<<cu_blocks,cu_threads>>>(num1,num2,num3,num4,num5,num6,lambda); \
+    Lambda6Apply<<<cu_blocks,cu_threads,0,cpuStream>>>(num1,num2,num3,num4,num5,num6,lambda); \
   }
 
 template<typename lambda>  __global__
@@ -189,7 +190,7 @@ void Lambda6Apply(uint64_t num1, uint64_t num2, uint64_t num3,
 
 #define accelerator_barrier(dummy)					\
   {									\
-    cudaDeviceSynchronize();						\
+    cudaStreamSynchronize(cpuStream);					\
     cudaError err = cudaGetLastError();					\
     if ( cudaSuccess != err ) {						\
       printf("accelerator_barrier(): Cuda error %s \n",			\
@@ -339,6 +340,7 @@ NAMESPACE_BEGIN(Grid);
 #define accelerator_inline __host__ __device__ inline
 
 extern hipStream_t copyStream;
+extern hipStream_t cpuStream;
 /*These routines define mapping from thread grid to loop & vector lane indexing */
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
@@ -360,12 +362,12 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
     dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
     if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
       hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads,		\
-            0,0,						\
-            num1,num2,nsimd, lambda);				\
+			 0,cpuStream,					\
+			 num1,num2,nsimd, lambda);			\
     } else { \
       hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
-            0,0,						\
-            num1,num2,nsimd, lambda);				\
+			 0,cpuStream,					\
+			 num1,num2,nsimd, lambda);			\
     } \
   }
 
@@ -398,7 +400,7 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 
 #define accelerator_barrier(dummy)				\
   {								\
-    hipDeviceSynchronize();					\
+    hipStreamSynchronize(cpuStream);					\
     auto err = hipGetLastError();				\
     if ( err != hipSuccess ) {					\
       printf("After hipDeviceSynchronize() : HIP error %s \n", hipGetErrorString( err )); \
@@ -441,7 +443,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
-  hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToDevice,copyStream);
+  hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);
 }
 inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
 
@@ -461,6 +463,8 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
   accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, { __VA_ARGS__ } ); \
   accelerator_barrier(dummy);
 
+#define GRID_ACCELERATED
+
 #endif
 
 //////////////////////////////////////////////

HMC/Mobius2p1fIDSDRGparityEOFA_40ID.cc

@@ -0,0 +1,918 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./HMC/Mobius2p1fIDSDRGparityEOFA.cc
+
+Copyright (C) 2015-2016
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <pabobyle@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 */
+#include <Grid/Grid.h>
+
+using namespace Grid;
+
+//Production binary for the 40ID G-parity ensemble
+
+struct RatQuoParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(RatQuoParameters,
+				  double, bnd_lo,
+				  double, bnd_hi,
+				  Integer, action_degree,
+				  double, action_tolerance,
+				  Integer, md_degree,
+				  double, md_tolerance,
+				  Integer, reliable_update_freq,
+				  Integer, bnd_check_freq);
+  RatQuoParameters() { 
+    bnd_lo = 1e-2;
+    bnd_hi = 30;
+    action_degree = 10;
+    action_tolerance = 1e-10;
+    md_degree = 10;
+    md_tolerance = 1e-8;
+    bnd_check_freq = 20;
+    reliable_update_freq = 50;
+  }
+
+  void Export(RationalActionParams &into) const{
+    into.lo = bnd_lo;
+    into.hi = bnd_hi;
+    into.action_degree = action_degree;
+    into.action_tolerance = action_tolerance;
+    into.md_degree = md_degree;
+    into.md_tolerance = md_tolerance;
+    into.BoundsCheckFreq = bnd_check_freq;
+  }
+};
+
+struct EOFAparameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(EOFAparameters,
+				  OneFlavourRationalParams, rat_params,
+				  double, action_tolerance,
+				  double, action_mixcg_inner_tolerance,
+				  double, md_tolerance,
+				  double, md_mixcg_inner_tolerance);
+
+  EOFAparameters() { 
+    action_mixcg_inner_tolerance = 1e-8;
+    action_tolerance = 1e-10;
+    md_tolerance = 1e-8;
+    md_mixcg_inner_tolerance = 1e-8;
+
+    rat_params.lo = 1.0;
+    rat_params.hi = 25.0;
+    rat_params.MaxIter  = 50000;
+    rat_params.tolerance= 1.0e-9;
+    rat_params.degree   = 14;
+    rat_params.precision= 50;
+  }
+};
+
+struct EvolParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(EvolParameters,
+                                  Integer, StartTrajectory,
+                                  Integer, Trajectories,
+				  Integer, SaveInterval,
+				  Integer, Steps,
+				  RealD, TrajectoryLength,
+                                  bool, MetropolisTest,
+				  std::string, StartingType,
+				  std::vector<Integer>, GparityDirs,
+				  std::vector<EOFAparameters>, eofa_l,
+				  RatQuoParameters, rat_quo_s,
+				  RatQuoParameters, rat_quo_DSDR);
+
+  EvolParameters() {
+    //For initial thermalization; afterwards user should switch Metropolis on and use StartingType=CheckpointStart
+    MetropolisTest    = false;
+    StartTrajectory   = 0;
+    Trajectories      = 50;
+    SaveInterval = 5;
+    StartingType      = "ColdStart";
+    GparityDirs.resize(3, 1); //1 for G-parity, 0 for periodic
+    Steps = 5;
+    TrajectoryLength = 1.0;
+  }
+};
+
+bool fileExists(const std::string &fn){
+  std::ifstream f(fn);
+  return f.good();
+}
+
+
+
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+				  double, alpha,
+				  double, beta,
+				  double, mu,
+				  int, ord,
+				  int, n_stop,
+				  int, n_want,
+				  int, n_use,
+				  double, tolerance);
+
+  LanczosParameters() {
+    alpha = 35;
+    beta = 5;
+    mu = 0;
+    ord = 100;
+    n_stop = 10;
+    n_want = 10;
+    n_use = 15;
+    tolerance = 1e-6;
+  }
+};
+
+
+
+template<typename FermionActionD, typename FermionFieldD>
+void computeEigenvalues(std::string param_file,
+			GridCartesian* Grid, GridRedBlackCartesian* rbGrid, const LatticeGaugeFieldD &latt,  //expect lattice to have been initialized to something
+			FermionActionD &action, GridParallelRNG &rng){
+  
+  LanczosParameters params;
+  if(fileExists(param_file)){
+    std::cout << GridLogMessage << " Reading " << param_file << std::endl;
+    Grid::XmlReader rd(param_file);
+    read(rd, "LanczosParameters", params);
+  }else if(!GlobalSharedMemory::WorldRank){
+    std::cout << GridLogMessage << " File " << param_file << " does not exist" << std::endl;
+    std::cout << GridLogMessage << " Writing xml template to " << param_file << ".templ" << std::endl;
+    Grid::XmlWriter wr(param_file + ".templ");
+    write(wr, "LanczosParameters", params);
+  }
+
+  FermionFieldD gauss_o(rbGrid);
+  FermionFieldD gauss(Grid);
+  gaussian(rng, gauss);
+  pickCheckerboard(Odd, gauss_o, gauss);
+
+  action.ImportGauge(latt);
+
+  SchurDiagMooeeOperator<FermionActionD, FermionFieldD> hermop(action);
+  PlainHermOp<FermionFieldD> hermop_wrap(hermop);
+  //ChebyshevLanczos<FermionFieldD> Cheb(params.alpha, params.beta, params.mu, params.ord);
+  assert(params.mu == 0.0);
+
+  Chebyshev<FermionFieldD> Cheb(params.beta*params.beta, params.alpha*params.alpha, params.ord+1);
+  FunctionHermOp<FermionFieldD> Cheb_wrap(Cheb, hermop);
+
+  std::cout << "IRL: alpha=" << params.alpha << " beta=" << params.beta << " mu=" << params.mu << " ord=" << params.ord << std::endl;
+  ImplicitlyRestartedLanczos<FermionFieldD> IRL(Cheb_wrap, hermop_wrap, params.n_stop, params.n_want, params.n_use, params.tolerance, 50000);
+
+  std::vector<RealD> eval(params.n_use);
+  std::vector<FermionFieldD> evec(params.n_use, rbGrid);
+  int Nconv;
+  IRL.calc(eval, evec, gauss_o, Nconv);
+
+  std::cout << "Eigenvalues:" << std::endl;
+  for(int i=0;i<params.n_want;i++){
+    std::cout << i << " " << eval[i] << std::endl;
+  }
+}
+
+
+//Check the quality of the RHMC approx
+//action_or_md toggles checking the action (0), MD (1) or both (2) setups
+template<typename FermionActionD, typename FermionFieldD, typename RHMCtype>
+void checkRHMC(GridCartesian* Grid, GridRedBlackCartesian* rbGrid, const LatticeGaugeFieldD &latt,  //expect lattice to have been initialized to something
+	       FermionActionD &numOp, FermionActionD &denOp, RHMCtype &rhmc, GridParallelRNG &rng,
+	       int inv_pow, const std::string &quark_descr, int action_or_md){
+  assert(action_or_md == 0 || action_or_md == 1 || action_or_md == 2);
+  
+  FermionFieldD gauss_o(rbGrid);
+  FermionFieldD gauss(Grid);
+  gaussian(rng, gauss);
+  pickCheckerboard(Odd, gauss_o, gauss);
+
+  numOp.ImportGauge(latt);
+  denOp.ImportGauge(latt);
+
+  typedef typename FermionActionD::Impl_t FermionImplPolicyD;
+  SchurDifferentiableOperator<FermionImplPolicyD> MdagM(numOp);
+  SchurDifferentiableOperator<FermionImplPolicyD> VdagV(denOp);
+
+  PowerMethod<FermionFieldD> power_method;
+  RealD lambda_max;
+
+  std::cout << "Starting: Get RHMC high bound approx for " << quark_descr << " numerator" << std::endl;
+
+  lambda_max = power_method(MdagM,gauss_o);
+  std::cout << GridLogMessage << "Got lambda_max "<<lambda_max<<std::endl;
+
+  std::cout << "Starting: Get RHMC high bound approx for " << quark_descr << " denominator" << std::endl;
+  lambda_max = power_method(VdagV,gauss_o);
+  std::cout << GridLogMessage << "Got lambda_max "<<lambda_max<<std::endl;
+
+  if(action_or_md == 0 || action_or_md == 2){
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 50000, 1e16, MdagM,gauss_o, rhmc.ApproxNegPowerAction); //use large tolerance to prevent exit on fail; we are trying to tune here!
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 50000, 1e16, MdagM,gauss_o, rhmc.ApproxNegHalfPowerAction);
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 50000, 1e16, VdagV,gauss_o, rhmc.ApproxNegPowerAction);
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 50000, 1e16, VdagV,gauss_o, rhmc.ApproxNegHalfPowerAction);
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+  }
+
+  std::cout << "-------------------------------------------------------------------------------" << std::endl;
+
+  if(action_or_md == 1 || action_or_md == 2){
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 50000, 1e16, MdagM,gauss_o, rhmc.ApproxNegPowerMD); 
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 50000, 1e16, MdagM,gauss_o, rhmc.ApproxNegHalfPowerMD);
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 50000, 1e16, VdagV,gauss_o, rhmc.ApproxNegPowerMD);
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 50000, 1e16, VdagV,gauss_o, rhmc.ApproxNegHalfPowerMD);
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+  }
+}
+
+
+template<typename FermionImplPolicy>
+void checkEOFA(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA,
+	       GridCartesian* FGrid, GridParallelRNG &rng, const LatticeGaugeFieldD &latt){
+  std::cout << GridLogMessage << "Starting EOFA action/bounds check" << std::endl;
+  typename FermionImplPolicy::FermionField eta(FGrid);
+  RealD scale = std::sqrt(0.5);
+  gaussian(rng,eta); eta = eta * scale;
+
+  //Use the inbuilt check
+  EOFA.refresh(latt, eta);
+  EOFA.S(latt);
+  std::cout << GridLogMessage << "Finished EOFA upper action/bounds check" << std::endl;
+}
+
+
+template<typename FermionImplPolicy>
+class EOFAlinop: public LinearOperatorBase<typename FermionImplPolicy::FermionField>{
+  ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA;
+  LatticeGaugeFieldD &U;
+public:
+  EOFAlinop(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA, LatticeGaugeFieldD &U): EOFA(EOFA), U(U){}
+
+  typedef typename FermionImplPolicy::FermionField Field;
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); } 
+
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ assert(0); }
+  void HermOp(const Field &in, Field &out){ EOFA.Meofa(U, in, out); }
+};
+
+template<typename FermionImplPolicy>
+void upperBoundEOFA(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA,
+		    GridCartesian* FGrid, GridParallelRNG &rng, LatticeGaugeFieldD &latt){
+  std::cout << GridLogMessage << "Starting EOFA upper bound compute" << std::endl;
+  EOFAlinop<FermionImplPolicy> linop(EOFA, latt);
+  typename FermionImplPolicy::FermionField eta(FGrid);
+  gaussian(rng,eta);
+  PowerMethod<typename FermionImplPolicy::FermionField> power_method;
+  auto lambda_max = power_method(linop,eta);
+  std::cout << GridLogMessage << "Upper bound of EOFA operator " << lambda_max << std::endl;
+}
+
+//Applications of M^{-1} cost the same as M for EOFA!
+template<typename FermionImplPolicy>
+class EOFAinvLinop: public LinearOperatorBase<typename FermionImplPolicy::FermionField>{
+  ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA;
+  LatticeGaugeFieldD &U;
+public:
+  EOFAinvLinop(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA, LatticeGaugeFieldD &U): EOFA(EOFA), U(U){}
+
+  typedef typename FermionImplPolicy::FermionField Field;
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); } 
+
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ assert(0); }
+  void HermOp(const Field &in, Field &out){ EOFA.MeofaInv(U, in, out); }
+};
+
+template<typename FermionImplPolicy>
+void lowerBoundEOFA(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA,
+		    GridCartesian* FGrid, GridParallelRNG &rng, LatticeGaugeFieldD &latt){
+  std::cout << GridLogMessage << "Starting EOFA lower bound compute using power method on M^{-1}. Inverse of highest eigenvalue is the lowest eigenvalue of M" << std::endl;
+  EOFAinvLinop<FermionImplPolicy> linop(EOFA, latt);
+  typename FermionImplPolicy::FermionField eta(FGrid);
+  gaussian(rng,eta);
+  PowerMethod<typename FermionImplPolicy::FermionField> power_method;
+  auto lambda_max = power_method(linop,eta);
+  std::cout << GridLogMessage << "Lower bound of EOFA operator " << 1./lambda_max << std::endl;
+}
+
+
+NAMESPACE_BEGIN(Grid);
+
+  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;
+
+    using OperatorFunction<FieldD>::operator();
+
+    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.) 
+    { 
+    };
+
+    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);
+      assert(&(SchurOpU->_Mat)==&(LinOpD._Mat));
+
+      precisionChange(FermOpF.Umu, FermOpD.Umu);
+
+      pickCheckerboard(Even,FermOpF.UmuEven,FermOpF.Umu);
+      pickCheckerboard(Odd ,FermOpF.UmuOdd ,FermOpF.Umu);
+
+      ////////////////////////////////////////////////////////////////////////////////////
+      // Make a mixed precision conjugate gradient
+      ////////////////////////////////////////////////////////////////////////////////////
+      MixedPrecisionConjugateGradient<FieldD,FieldF> MPCG(Tolerance,MaxInnerIterations,MaxOuterIterations,SinglePrecGrid5,LinOpF,LinOpD);
+      MPCG.InnerTolerance = InnerTolerance;
+      std::cout << GridLogMessage << "Calling mixed precision Conjugate Gradient" <<std::endl;
+      MPCG(src,psi);
+    }
+  };
+
+
+
+  template<class FermionOperatorD, class FermionOperatorF, class SchurOperatorD, class  SchurOperatorF> 
+  class MixedPrecisionReliableUpdateConjugateGradientOperatorFunction : public OperatorFunction<typename FermionOperatorD::FermionField> {
+  public:
+    typedef typename FermionOperatorD::FermionField FieldD;
+    typedef typename FermionOperatorF::FermionField FieldF;
+
+    using OperatorFunction<FieldD>::operator();
+
+    RealD Tolerance;
+    Integer MaxIterations;
+
+    RealD Delta; //reliable update parameter
+
+    GridBase* SinglePrecGrid4; //Grid for single-precision fields
+    GridBase* SinglePrecGrid5; //Grid for single-precision fields
+
+    FermionOperatorF &FermOpF;
+    FermionOperatorD &FermOpD;;
+    SchurOperatorF &LinOpF;
+    SchurOperatorD &LinOpD;
+    
+    MixedPrecisionReliableUpdateConjugateGradientOperatorFunction(RealD tol, 
+								  RealD delta,
+								  Integer maxit, 
+								  GridBase* _sp_grid4, 
+								  GridBase* _sp_grid5, 
+								  FermionOperatorF &_FermOpF,
+								  FermionOperatorD &_FermOpD,
+								  SchurOperatorF   &_LinOpF,
+								  SchurOperatorD   &_LinOpD): 
+      LinOpF(_LinOpF),
+      LinOpD(_LinOpD),
+      FermOpF(_FermOpF),
+      FermOpD(_FermOpD),
+      Tolerance(tol), 
+      Delta(delta),
+      MaxIterations(maxit), 
+      SinglePrecGrid4(_sp_grid4),
+      SinglePrecGrid5(_sp_grid5)
+    { 
+    };
+
+    void operator()(LinearOperatorBase<FieldD> &LinOpU, const FieldD &src, FieldD &psi) {
+
+      std::cout << GridLogMessage << " Mixed precision reliable CG update wrapper operator() "<<std::endl;
+
+      SchurOperatorD * SchurOpU = static_cast<SchurOperatorD *>(&LinOpU);
+      assert(&(SchurOpU->_Mat)==&(LinOpD._Mat));
+
+      precisionChange(FermOpF.Umu, FermOpD.Umu);
+
+      pickCheckerboard(Even,FermOpF.UmuEven,FermOpF.Umu);
+      pickCheckerboard(Odd ,FermOpF.UmuOdd ,FermOpF.Umu);
+
+      ////////////////////////////////////////////////////////////////////////////////////
+      // Make a mixed precision conjugate gradient
+      ////////////////////////////////////////////////////////////////////////////////////
+
+      ConjugateGradientReliableUpdate<FieldD,FieldF> MPCG(Tolerance,MaxIterations,Delta,SinglePrecGrid5,LinOpF,LinOpD);
+      std::cout << GridLogMessage << "Calling mixed precision reliable update Conjugate Gradient" <<std::endl;
+      MPCG(src,psi);
+    }
+  };
+
+
+
+NAMESPACE_END(Grid);
+
+
+
+
+
+int main(int argc, char **argv) {
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+  // here make a routine to print all the relevant information on the run
+  std::cout << GridLogMessage << "Grid is setup to use " << threads << " threads" << std::endl;
+
+  std::string param_file = "params.xml";
+  bool file_load_check = false;
+
+  std::string serial_seeds = "1 2 3 4 5";
+  std::string parallel_seeds = "6 7 8 9 10";
+
+  int i=1;
+  while(i < argc){
+    std::string sarg(argv[i]);
+    if(sarg == "--param_file"){
+      assert(i!=argc-1);
+      param_file = argv[i+1];
+      i+=2;
+    }else if(sarg == "--read_check"){ //check the fields load correctly and pass checksum/plaquette repro
+      file_load_check = true;
+      i++;
+    }else if(sarg == "--set_seeds"){ //set the rng seeds. Expects two vector args, e.g.  --set_seeds 1.2.3.4 5.6.7.8
+      assert(i < argc-2);
+      std::vector<int> tmp;
+      GridCmdOptionIntVector(argv[i+1],tmp);
+      {
+	std::stringstream ss;
+	for(int j=0;j<tmp.size()-1;j++) ss << tmp[j] << " ";
+	ss << tmp.back();
+	serial_seeds = ss.str();
+      }
+      GridCmdOptionIntVector(argv[i+2],tmp);
+      {
+	std::stringstream ss;
+	for(int j=0;j<tmp.size()-1;j++) ss << tmp[j] << " ";
+	ss << tmp.back();
+	parallel_seeds = ss.str();
+      }
+      i+=3;
+      std::cout << GridLogMessage << "Set serial seeds to " << serial_seeds << std::endl;
+      std::cout << GridLogMessage << "Set parallel seeds to " << parallel_seeds << std::endl;
+      
+    }else{
+      i++;
+    }
+  }
+
+  
+  //Read the user parameters
+  EvolParameters user_params;
+  
+  if(fileExists(param_file)){
+    std::cout << GridLogMessage << " Reading " << param_file << std::endl;
+    Grid::XmlReader rd(param_file);
+    read(rd, "Params", user_params);
+  }else if(!GlobalSharedMemory::WorldRank){
+    std::cout << GridLogMessage << " File " << param_file << " does not exist" << std::endl;
+    std::cout << GridLogMessage << " Writing xml template to " << param_file << ".templ" << std::endl;
+    {
+      Grid::XmlWriter wr(param_file + ".templ");
+      write(wr, "Params", user_params);
+    }
+    std::cout << GridLogMessage << " Done" << std::endl;
+    Grid_finalize();
+    return 0;
+  }
+
+  //Check the parameters
+  if(user_params.GparityDirs.size() != Nd-1){
+    std::cerr << "Error in input parameters: expect GparityDirs to have size = " << Nd-1 << std::endl;
+    exit(1);
+  }
+  for(int i=0;i<Nd-1;i++)
+    if(user_params.GparityDirs[i] != 0 && user_params.GparityDirs[i] != 1){
+      std::cerr << "Error in input parameters: expect GparityDirs values to be 0 (periodic) or 1 (G-parity)" << std::endl;
+      exit(1);
+    }
+
+
+  typedef GparityMobiusEOFAFermionD EOFAactionD;
+  typedef GparityMobiusFermionD FermionActionD;
+  typedef typename FermionActionD::Impl_t FermionImplPolicyD;
+  typedef typename FermionActionD::FermionField FermionFieldD;
+
+  typedef GparityMobiusEOFAFermionF EOFAactionF;
+  typedef GparityMobiusFermionF FermionActionF;
+  typedef typename FermionActionF::Impl_t FermionImplPolicyF;
+  typedef typename FermionActionF::FermionField FermionFieldF;
+
+  typedef GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicyD,FermionImplPolicyF> MixedPrecRHMC;
+  typedef GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicyD> DoublePrecRHMC;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  typedef ConjugateHMCRunnerD<MinimumNorm2> HMCWrapper; //NB: This is the "Omelyan integrator"
+  MD.name    = std::string("MinimumNorm2");
+
+  // typedef ConjugateHMCRunnerD<ForceGradient> HMCWrapper;
+  // MD.name    = std::string("ForceGradient");
+  
+  MD.MDsteps = user_params.Steps;
+  MD.trajL   = user_params.TrajectoryLength;
+
+  typedef HMCWrapper::ImplPolicy GaugeImplPolicy;
+  
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = user_params.StartTrajectory;
+  HMCparams.Trajectories     = user_params.Trajectories;
+  HMCparams.NoMetropolisUntil= 0;
+  HMCparams.StartingType     = user_params.StartingType;
+  HMCparams.MetropolisTest = user_params.MetropolisTest;
+  HMCparams.MD = MD;
+  HMCWrapper TheHMC(HMCparams);
+
+  // Grid from the command line arguments --grid and --mpi
+  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
+
+  CheckpointerParameters CPparams;
+  CPparams.config_prefix = "ckpoint_lat";
+  CPparams.rng_prefix    = "ckpoint_rng";
+  CPparams.saveInterval  = user_params.SaveInterval;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  //Note that checkpointing saves the RNG state so that this initialization is required only for the very first configuration
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = serial_seeds;
+  RNGpar.parallel_seeds = parallel_seeds;
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  typedef PlaquetteMod<GaugeImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+  //aiming for ainv=1.723 GeV
+  //                                  me         bob
+  //Estimated  a(ml+mres) [40ID] = 0.001305    0.00131
+  //           a(mh+mres) [40ID] = 0.035910    0.03529
+  //Estimate Ls=12, b+c=2  mres~0.0011
+
+  //1/24/2022 initial mres measurement gives mres=0.001,  adjusted light quark mass to 0.0003 from 0.0001
+  
+  const int Ls      = 12;
+  Real beta         = 1.848;
+  Real light_mass   = 0.0003;
+  Real strange_mass = 0.0342;
+  Real pv_mass      = 1.0;
+  RealD M5  = 1.8;
+  RealD mobius_scale = 2.; //b+c
+
+  RealD mob_bmc = 1.0;
+  RealD mob_b = (mobius_scale + mob_bmc)/2.;
+  RealD mob_c = (mobius_scale - mob_bmc)/2.;
+
+  std::cout << GridLogMessage
+	    << "Ensemble parameters:" << std::endl
+	    << "Ls=" << Ls << std::endl
+	    << "beta=" << beta << std::endl
+	    << "light_mass=" << light_mass << std::endl
+	    << "strange_mass=" << strange_mass << std::endl
+	    << "mobius_scale=" << mobius_scale << std::endl;
+  
+  //Setup the Grids
+  auto UGridD   = TheHMC.Resources.GetCartesian();
+  auto UrbGridD = TheHMC.Resources.GetRBCartesian();
+  auto FGridD     = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridD);
+  auto FrbGridD   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridD);
+
+  GridCartesian* UGridF = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian* UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  auto FGridF     = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridF);
+  auto FrbGridF   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridF);
+
+  ConjugateIwasakiGaugeActionD GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeFieldD Ud(UGridD);
+  LatticeGaugeFieldF Uf(UGridF);
+ 
+  //Setup the BCs
+  FermionActionD::ImplParams Params;
+  for(int i=0;i<Nd-1;i++) Params.twists[i] = user_params.GparityDirs[i]; //G-parity directions
+  Params.twists[Nd-1] = 1; //APBC in time direction
+
+  std::vector<int> dirs4(Nd);
+  for(int i=0;i<Nd-1;i++) dirs4[i] = user_params.GparityDirs[i];
+  dirs4[Nd-1] = 0; //periodic gauge BC in time
+
+  GaugeImplPolicy::setDirections(dirs4); //gauge BC
+
+  //Run optional gauge field checksum checker and exit
+  if(file_load_check){
+    TheHMC.initializeGaugeFieldAndRNGs(Ud);
+    std::cout << GridLogMessage << " Done" << std::endl;
+    Grid_finalize();
+    return 0;
+  }
+
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1); //light quark + strange quark
+  ActionLevel<HMCWrapper::Field> Level2(4); //DSDR
+  ActionLevel<HMCWrapper::Field> Level3(2); //gauge
+
+
+  /////////////////////////////////////////////////////////////
+  // Light EOFA action
+  // have to be careful with the parameters, cf. Test_dwf_gpforce_eofa.cc
+  /////////////////////////////////////////////////////////////
+  typedef SchurDiagMooeeOperator<EOFAactionD,FermionFieldD> EOFAschuropD;
+  typedef SchurDiagMooeeOperator<EOFAactionF,FermionFieldF> EOFAschuropF;
+  typedef ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction<FermionImplPolicyD, FermionImplPolicyF> EOFAmixPrecPFaction;
+  typedef MixedPrecisionConjugateGradientOperatorFunction<EOFAactionD, EOFAactionF, EOFAschuropD, EOFAschuropF> EOFA_mxCG;
+  typedef MixedPrecisionReliableUpdateConjugateGradientOperatorFunction<EOFAactionD, EOFAactionF, EOFAschuropD, EOFAschuropF> EOFA_relupCG;
+
+
+  std::vector<RealD> eofa_light_masses = { light_mass ,  0.004,   0.016,   0.064,   0.256    };
+  std::vector<RealD> eofa_pv_masses =    { 0.004       , 0.016,   0.064,   0.256,   1.0      };
+  int n_light_hsb = 5;
+  assert(user_params.eofa_l.size() == n_light_hsb);
+  
+  EOFAmixPrecPFaction* EOFA_pfactions[n_light_hsb];
+
+  for(int i=0;i<n_light_hsb;i++){
+    RealD iml = eofa_light_masses[i];
+    RealD ipv = eofa_pv_masses[i];
+
+    EOFAactionD* LopD = new EOFAactionD(Ud, *FGridD, *FrbGridD, *UGridD, *UrbGridD, iml, iml, ipv, 0.0, -1, M5, mob_b, mob_c, Params);
+    EOFAactionF* LopF = new EOFAactionF(Uf, *FGridF, *FrbGridF, *UGridF, *UrbGridF, iml, iml, ipv, 0.0, -1, M5, mob_b, mob_c, Params);
+    EOFAactionD* RopD = new EOFAactionD(Ud, *FGridD, *FrbGridD, *UGridD, *UrbGridD, ipv, iml, ipv, -1.0, 1, M5, mob_b, mob_c, Params);
+    EOFAactionF* RopF = new EOFAactionF(Uf, *FGridF, *FrbGridF, *UGridF, *UrbGridF, ipv, iml, ipv, -1.0, 1, M5, mob_b, mob_c, Params);
+
+    EOFAschuropD* linopL_D = new EOFAschuropD(*LopD);
+    EOFAschuropD* linopR_D = new EOFAschuropD(*RopD);
+    
+    EOFAschuropF* linopL_F = new EOFAschuropF(*LopF);
+    EOFAschuropF* linopR_F = new EOFAschuropF(*RopF);
+
+#if 1
+    //Note reusing user_params.eofa_l.action(|md)_mixcg_inner_tolerance  as Delta for now
+    EOFA_relupCG* ActionMCG_L = new EOFA_relupCG(user_params.eofa_l[i].action_tolerance, user_params.eofa_l[i].action_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    EOFA_relupCG* ActionMCG_R = new EOFA_relupCG(user_params.eofa_l[i].action_tolerance, user_params.eofa_l[i].action_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+
+    EOFA_relupCG* DerivMCG_L = new EOFA_relupCG(user_params.eofa_l[i].md_tolerance, user_params.eofa_l[i].md_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    EOFA_relupCG* DerivMCG_R = new EOFA_relupCG(user_params.eofa_l[i].md_tolerance, user_params.eofa_l[i].md_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+
+#else
+    EOFA_mxCG* ActionMCG_L = new EOFA_mxCG(user_params.eofa_l[i].action_tolerance, 50000, 1000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    ActionMCG_L->InnerTolerance = user_params.eofa_l[i].action_mixcg_inner_tolerance;
+    
+    EOFA_mxCG* ActionMCG_R = new EOFA_mxCG(user_params.eofa_l[i].action_tolerance, 50000, 1000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+    ActionMCG_R->InnerTolerance = user_params.eofa_l[i].action_mixcg_inner_tolerance;
+    
+    EOFA_mxCG* DerivMCG_L = new EOFA_mxCG(user_params.eofa_l[i].md_tolerance, 50000, 1000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    DerivMCG_L->InnerTolerance = user_params.eofa_l[i].md_mixcg_inner_tolerance;
+    
+    EOFA_mxCG* DerivMCG_R = new EOFA_mxCG(user_params.eofa_l[i].md_tolerance, 50000, 1000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+    DerivMCG_R->InnerTolerance = user_params.eofa_l[i].md_mixcg_inner_tolerance;
+    
+    std::cout << GridLogMessage << "Set EOFA action solver action tolerance outer=" << ActionMCG_L->Tolerance << " inner=" << ActionMCG_L->InnerTolerance << std::endl;
+    std::cout << GridLogMessage << "Set EOFA MD solver tolerance outer=" << DerivMCG_L->Tolerance << " inner=" << DerivMCG_L->InnerTolerance << std::endl;
+#endif
+
+    EOFAmixPrecPFaction* EOFA = new EOFAmixPrecPFaction(*LopF, *RopF,
+							*LopD, *RopD, 
+							*ActionMCG_L, *ActionMCG_R, 
+							*ActionMCG_L, *ActionMCG_R, 
+							*DerivMCG_L, *DerivMCG_R, 
+							user_params.eofa_l[i].rat_params, true);
+    EOFA_pfactions[i] = EOFA;
+    Level1.push_back(EOFA);
+  }
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionActionD Numerator_sD(Ud,*FGridD,*FrbGridD,*UGridD,*UrbGridD,strange_mass,M5,mob_b,mob_c,Params);
+  FermionActionD Denominator_sD(Ud,*FGridD,*FrbGridD,*UGridD,*UrbGridD, pv_mass,M5,mob_b,mob_c,Params);
+
+  FermionActionF Numerator_sF(Uf,*FGridF,*FrbGridF,*UGridF,*UrbGridF,strange_mass,M5,mob_b,mob_c,Params);
+  FermionActionF Denominator_sF(Uf,*FGridF,*FrbGridF,*UGridF,*UrbGridF, pv_mass,M5,mob_b,mob_c,Params);
+
+  RationalActionParams rat_act_params_s;
+  rat_act_params_s.inv_pow  = 4; // (M^dag M)^{1/4}
+  rat_act_params_s.precision= 60;
+  rat_act_params_s.MaxIter  = 50000;
+  user_params.rat_quo_s.Export(rat_act_params_s);
+  std::cout << GridLogMessage << " Heavy quark bounds check every " << rat_act_params_s.BoundsCheckFreq << " trajectories (avg)" << std::endl;
+
+  //MixedPrecRHMC Quotient_s(Denominator_sD, Numerator_sD, Denominator_sF, Numerator_sF, rat_act_params_s, user_params.rat_quo_s.reliable_update_freq); 
+  DoublePrecRHMC Quotient_s(Denominator_sD, Numerator_sD, rat_act_params_s); 
+  Level1.push_back(&Quotient_s);  
+
+  ///////////////////////////////////
+  // DSDR action
+  ///////////////////////////////////
+  RealD dsdr_mass=-1.8;   
+  //Use same DSDR twists as https://arxiv.org/pdf/1208.4412.pdf
+  RealD dsdr_epsilon_f = 0.02; //numerator (in determinant)
+  RealD dsdr_epsilon_b = 0.5; 
+  GparityWilsonTMFermionD Numerator_DSDR_D(Ud, *UGridD, *UrbGridD, dsdr_mass, dsdr_epsilon_f, Params);
+  GparityWilsonTMFermionF Numerator_DSDR_F(Uf, *UGridF, *UrbGridF, dsdr_mass, dsdr_epsilon_f, Params);
+
+  GparityWilsonTMFermionD Denominator_DSDR_D(Ud, *UGridD, *UrbGridD, dsdr_mass, dsdr_epsilon_b, Params);
+  GparityWilsonTMFermionF Denominator_DSDR_F(Uf, *UGridF, *UrbGridF, dsdr_mass, dsdr_epsilon_b, Params);
+ 
+  RationalActionParams rat_act_params_DSDR;
+  rat_act_params_DSDR.inv_pow  = 2; // (M^dag M)^{1/2}
+  rat_act_params_DSDR.precision= 60;
+  rat_act_params_DSDR.MaxIter  = 50000;
+  user_params.rat_quo_DSDR.Export(rat_act_params_DSDR);
+  std::cout << GridLogMessage << "DSDR quark bounds check every " << rat_act_params_DSDR.BoundsCheckFreq << " trajectories (avg)" << std::endl;
+
+  DoublePrecRHMC Quotient_DSDR(Denominator_DSDR_D, Numerator_DSDR_D, rat_act_params_DSDR);
+  Level2.push_back(&Quotient_DSDR);
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+
+
+  //Action tuning
+  bool 
+    tune_rhmc_s=false, eigenrange_s=false, 
+    tune_rhmc_DSDR=false, eigenrange_DSDR=false, 
+    check_eofa=false, 
+    upper_bound_eofa=false, lower_bound_eofa(false);
+
+  std::string lanc_params_s;
+  std::string lanc_params_DSDR;
+  int tune_rhmc_s_action_or_md;
+  int tune_rhmc_DSDR_action_or_md;
+  int eofa_which_hsb;
+
+  for(int i=1;i<argc;i++){
+    std::string sarg(argv[i]);
+    if(sarg == "--tune_rhmc_s"){
+      assert(i < argc-1);
+      tune_rhmc_s=true;
+      tune_rhmc_s_action_or_md = std::stoi(argv[i+1]);
+    }
+    else if(sarg == "--eigenrange_s"){
+      assert(i < argc-1);
+      eigenrange_s=true;
+      lanc_params_s = argv[i+1];
+    }
+    else if(sarg == "--tune_rhmc_DSDR"){
+      assert(i < argc-1);
+      tune_rhmc_DSDR=true;
+      tune_rhmc_DSDR_action_or_md = std::stoi(argv[i+1]);
+    }
+    else if(sarg == "--eigenrange_DSDR"){
+      assert(i < argc-1);
+      eigenrange_DSDR=true;
+      lanc_params_DSDR = argv[i+1];
+    }
+    else if(sarg == "--check_eofa"){
+      assert(i < argc-1);
+      check_eofa = true;
+      eofa_which_hsb = std::stoi(argv[i+1]); //-1 indicates all hasenbusch
+      assert(eofa_which_hsb == -1 || (eofa_which_hsb >= 0 && eofa_which_hsb < n_light_hsb) );
+    }
+    else if(sarg == "--upper_bound_eofa"){
+      assert(i < argc-1);
+      upper_bound_eofa = true;
+      eofa_which_hsb = std::stoi(argv[i+1]);
+      assert(eofa_which_hsb >= 0 && eofa_which_hsb < n_light_hsb);
+    }
+    else if(sarg == "--lower_bound_eofa"){
+      assert(i < argc-1);
+      lower_bound_eofa = true;      
+      eofa_which_hsb = std::stoi(argv[i+1]);
+      assert(eofa_which_hsb >= 0 && eofa_which_hsb < n_light_hsb);
+    }
+  }
+  if(tune_rhmc_s || eigenrange_s || tune_rhmc_DSDR || eigenrange_DSDR ||check_eofa || upper_bound_eofa || lower_bound_eofa) {
+    std::cout << GridLogMessage << "Running checks" << std::endl;
+    TheHMC.initializeGaugeFieldAndRNGs(Ud);
+
+    //std::cout << GridLogMessage << "EOFA action solver action tolerance outer=" << ActionMCG_L.Tolerance << " inner=" << ActionMCG_L.InnerTolerance << std::endl;
+    //std::cout << GridLogMessage << "EOFA MD solver tolerance outer=" << DerivMCG_L.Tolerance << " inner=" << DerivMCG_L.InnerTolerance << std::endl;
+
+    if(check_eofa){
+      if(eofa_which_hsb >= 0){
+	std::cout << GridLogMessage << "Starting checking EOFA Hasenbusch " << eofa_which_hsb << std::endl;
+	checkEOFA(*EOFA_pfactions[eofa_which_hsb], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+	std::cout << GridLogMessage << "Finished checking EOFA Hasenbusch " << eofa_which_hsb << std::endl;
+      }else{
+	for(int i=0;i<n_light_hsb;i++){
+	  std::cout << GridLogMessage << "Starting checking EOFA Hasenbusch " << i << std::endl;
+	  checkEOFA(*EOFA_pfactions[i], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+	  std::cout << GridLogMessage << "Finished checking EOFA Hasenbusch " << i << std::endl;
+	}
+      }
+    }	  
+    if(upper_bound_eofa) upperBoundEOFA(*EOFA_pfactions[eofa_which_hsb], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+    if(lower_bound_eofa) lowerBoundEOFA(*EOFA_pfactions[eofa_which_hsb], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+    if(eigenrange_s) computeEigenvalues<FermionActionD, FermionFieldD>(lanc_params_s, FGridD, FrbGridD, Ud, Numerator_sD, TheHMC.Resources.GetParallelRNG());
+    if(tune_rhmc_s) checkRHMC<FermionActionD, FermionFieldD, decltype(Quotient_s)>(FGridD, FrbGridD, Ud, Numerator_sD, Denominator_sD, Quotient_s, TheHMC.Resources.GetParallelRNG(), 4, "strange",  tune_rhmc_s_action_or_md);
+    if(eigenrange_DSDR) computeEigenvalues<GparityWilsonTMFermionD, GparityWilsonTMFermionD::FermionField>(lanc_params_DSDR, UGridD, UrbGridD, Ud, Numerator_DSDR_D, TheHMC.Resources.GetParallelRNG());
+    if(tune_rhmc_DSDR) checkRHMC<GparityWilsonTMFermionD, GparityWilsonTMFermionD::FermionField, decltype(Quotient_DSDR)>(UGridD, UrbGridD, Ud, Numerator_DSDR_D, Denominator_DSDR_D, Quotient_DSDR, TheHMC.Resources.GetParallelRNG(), 2, "DSDR", tune_rhmc_DSDR_action_or_md);
+
+
+    std::cout << GridLogMessage << " Done" << std::endl;
+    Grid_finalize();
+    return 0;
+  }
+
+
+  //Run the HMC
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.Run();
+
+  std::cout << GridLogMessage << " Done" << std::endl;
+  Grid_finalize();
+  return 0;
+} // main

HMC/Mobius2p1fIDSDRGparityEOFA_48ID.cc

@@ -0,0 +1,873 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./HMC/Mobius2p1fIDSDRGparityEOFA.cc
+
+Copyright (C) 2015-2016
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <pabobyle@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 */
+#include <Grid/Grid.h>
+
+using namespace Grid;
+
+//Production binary for the 40ID G-parity ensemble
+
+struct RatQuoParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(RatQuoParameters,
+				  double, bnd_lo,
+				  double, bnd_hi,
+				  Integer, action_degree,
+				  double, action_tolerance,
+				  Integer, md_degree,
+				  double, md_tolerance,
+				  Integer, reliable_update_freq,
+				  Integer, bnd_check_freq);
+  RatQuoParameters() { 
+    bnd_lo = 1e-2;
+    bnd_hi = 30;
+    action_degree = 10;
+    action_tolerance = 1e-10;
+    md_degree = 10;
+    md_tolerance = 1e-8;
+    bnd_check_freq = 20;
+    reliable_update_freq = 50;
+  }
+
+  void Export(RationalActionParams &into) const{
+    into.lo = bnd_lo;
+    into.hi = bnd_hi;
+    into.action_degree = action_degree;
+    into.action_tolerance = action_tolerance;
+    into.md_degree = md_degree;
+    into.md_tolerance = md_tolerance;
+    into.BoundsCheckFreq = bnd_check_freq;
+  }
+};
+
+struct EOFAparameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(EOFAparameters,
+				  OneFlavourRationalParams, rat_params,
+				  double, action_tolerance,
+				  double, action_mixcg_inner_tolerance,
+				  double, md_tolerance,
+				  double, md_mixcg_inner_tolerance);
+
+  EOFAparameters() { 
+    action_mixcg_inner_tolerance = 1e-8;
+    action_tolerance = 1e-10;
+    md_tolerance = 1e-8;
+    md_mixcg_inner_tolerance = 1e-8;
+
+    rat_params.lo = 1.0;
+    rat_params.hi = 25.0;
+    rat_params.MaxIter  = 10000;
+    rat_params.tolerance= 1.0e-9;
+    rat_params.degree   = 14;
+    rat_params.precision= 50;
+  }
+};
+
+struct EvolParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(EvolParameters,
+                                  Integer, StartTrajectory,
+                                  Integer, Trajectories,
+				  Integer, SaveInterval,
+				  Integer, Steps,
+				  RealD, TrajectoryLength,
+                                  bool, MetropolisTest,
+				  std::string, StartingType,
+				  std::vector<Integer>, GparityDirs,
+				  std::vector<EOFAparameters>, eofa_l,
+				  RatQuoParameters, rat_quo_s,
+				  RatQuoParameters, rat_quo_DSDR);
+
+  EvolParameters() {
+    //For initial thermalization; afterwards user should switch Metropolis on and use StartingType=CheckpointStart
+    MetropolisTest    = false;
+    StartTrajectory   = 0;
+    Trajectories      = 50;
+    SaveInterval = 5;
+    StartingType      = "ColdStart";
+    GparityDirs.resize(3, 1); //1 for G-parity, 0 for periodic
+    Steps = 5;
+    TrajectoryLength = 1.0;
+  }
+};
+
+bool fileExists(const std::string &fn){
+  std::ifstream f(fn);
+  return f.good();
+}
+
+
+
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+				  double, alpha,
+				  double, beta,
+				  double, mu,
+				  int, ord,
+				  int, n_stop,
+				  int, n_want,
+				  int, n_use,
+				  double, tolerance);
+
+  LanczosParameters() {
+    alpha = 35;
+    beta = 5;
+    mu = 0;
+    ord = 100;
+    n_stop = 10;
+    n_want = 10;
+    n_use = 15;
+    tolerance = 1e-6;
+  }
+};
+
+
+
+template<typename FermionActionD, typename FermionFieldD>
+void computeEigenvalues(std::string param_file,
+			GridCartesian* Grid, GridRedBlackCartesian* rbGrid, const LatticeGaugeFieldD &latt,  //expect lattice to have been initialized to something
+			FermionActionD &action, GridParallelRNG &rng){
+  
+  LanczosParameters params;
+  if(fileExists(param_file)){
+    std::cout << GridLogMessage << " Reading " << param_file << std::endl;
+    Grid::XmlReader rd(param_file);
+    read(rd, "LanczosParameters", params);
+  }else if(!GlobalSharedMemory::WorldRank){
+    std::cout << GridLogMessage << " File " << param_file << " does not exist" << std::endl;
+    std::cout << GridLogMessage << " Writing xml template to " << param_file << ".templ" << std::endl;
+    Grid::XmlWriter wr(param_file + ".templ");
+    write(wr, "LanczosParameters", params);
+  }
+
+  FermionFieldD gauss_o(rbGrid);
+  FermionFieldD gauss(Grid);
+  gaussian(rng, gauss);
+  pickCheckerboard(Odd, gauss_o, gauss);
+
+  action.ImportGauge(latt);
+
+  SchurDiagMooeeOperator<FermionActionD, FermionFieldD> hermop(action);
+  PlainHermOp<FermionFieldD> hermop_wrap(hermop);
+  //ChebyshevLanczos<FermionFieldD> Cheb(params.alpha, params.beta, params.mu, params.ord);
+  assert(params.mu == 0.0);
+
+  Chebyshev<FermionFieldD> Cheb(params.beta*params.beta, params.alpha*params.alpha, params.ord+1);
+  FunctionHermOp<FermionFieldD> Cheb_wrap(Cheb, hermop);
+
+  std::cout << "IRL: alpha=" << params.alpha << " beta=" << params.beta << " mu=" << params.mu << " ord=" << params.ord << std::endl;
+  ImplicitlyRestartedLanczos<FermionFieldD> IRL(Cheb_wrap, hermop_wrap, params.n_stop, params.n_want, params.n_use, params.tolerance, 10000);
+
+  std::vector<RealD> eval(params.n_use);
+  std::vector<FermionFieldD> evec(params.n_use, rbGrid);
+  int Nconv;
+  IRL.calc(eval, evec, gauss_o, Nconv);
+
+  std::cout << "Eigenvalues:" << std::endl;
+  for(int i=0;i<params.n_want;i++){
+    std::cout << i << " " << eval[i] << std::endl;
+  }
+}
+
+
+//Check the quality of the RHMC approx
+//action_or_md toggles checking the action (0), MD (1) or both (2) setups
+template<typename FermionActionD, typename FermionFieldD, typename RHMCtype>
+void checkRHMC(GridCartesian* Grid, GridRedBlackCartesian* rbGrid, const LatticeGaugeFieldD &latt,  //expect lattice to have been initialized to something
+	       FermionActionD &numOp, FermionActionD &denOp, RHMCtype &rhmc, GridParallelRNG &rng,
+	       int inv_pow, const std::string &quark_descr, int action_or_md){
+  assert(action_or_md == 0 || action_or_md == 1 || action_or_md == 2);
+  
+  FermionFieldD gauss_o(rbGrid);
+  FermionFieldD gauss(Grid);
+  gaussian(rng, gauss);
+  pickCheckerboard(Odd, gauss_o, gauss);
+
+  numOp.ImportGauge(latt);
+  denOp.ImportGauge(latt);
+
+  typedef typename FermionActionD::Impl_t FermionImplPolicyD;
+  SchurDifferentiableOperator<FermionImplPolicyD> MdagM(numOp);
+  SchurDifferentiableOperator<FermionImplPolicyD> VdagV(denOp);
+
+  PowerMethod<FermionFieldD> power_method;
+  RealD lambda_max;
+
+  std::cout << "Starting: Get RHMC high bound approx for " << quark_descr << " numerator" << std::endl;
+
+  lambda_max = power_method(MdagM,gauss_o);
+  std::cout << GridLogMessage << "Got lambda_max "<<lambda_max<<std::endl;
+
+  std::cout << "Starting: Get RHMC high bound approx for " << quark_descr << " denominator" << std::endl;
+  lambda_max = power_method(VdagV,gauss_o);
+  std::cout << GridLogMessage << "Got lambda_max "<<lambda_max<<std::endl;
+
+  if(action_or_md == 0 || action_or_md == 2){
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegPowerAction); //use large tolerance to prevent exit on fail; we are trying to tune here!
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegHalfPowerAction);
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegPowerAction);
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegHalfPowerAction);
+    std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+  }
+
+  std::cout << "-------------------------------------------------------------------------------" << std::endl;
+
+  if(action_or_md == 1 || action_or_md == 2){
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegPowerMD); 
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegHalfPowerMD);
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+    InversePowerBoundsCheck(inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegPowerMD);
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
+
+    std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+    InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegHalfPowerMD);
+    std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
+  }
+}
+
+
+template<typename FermionImplPolicy>
+void checkEOFA(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA,
+	       GridCartesian* FGrid, GridParallelRNG &rng, const LatticeGaugeFieldD &latt){
+  std::cout << GridLogMessage << "Starting EOFA action/bounds check" << std::endl;
+  typename FermionImplPolicy::FermionField eta(FGrid);
+  RealD scale = std::sqrt(0.5);
+  gaussian(rng,eta); eta = eta * scale;
+
+  //Use the inbuilt check
+  EOFA.refresh(latt, eta);
+  EOFA.S(latt);
+  std::cout << GridLogMessage << "Finished EOFA upper action/bounds check" << std::endl;
+}
+
+
+template<typename FermionImplPolicy>
+class EOFAlinop: public LinearOperatorBase<typename FermionImplPolicy::FermionField>{
+  ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA;
+  LatticeGaugeFieldD &U;
+public:
+  EOFAlinop(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA, LatticeGaugeFieldD &U): EOFA(EOFA), U(U){}
+
+  typedef typename FermionImplPolicy::FermionField Field;
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); } 
+
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ assert(0); }
+  void HermOp(const Field &in, Field &out){ EOFA.Meofa(U, in, out); }
+};
+
+template<typename FermionImplPolicy>
+void upperBoundEOFA(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA,
+		    GridCartesian* FGrid, GridParallelRNG &rng, LatticeGaugeFieldD &latt){
+  std::cout << GridLogMessage << "Starting EOFA upper bound compute" << std::endl;
+  EOFAlinop<FermionImplPolicy> linop(EOFA, latt);
+  typename FermionImplPolicy::FermionField eta(FGrid);
+  gaussian(rng,eta);
+  PowerMethod<typename FermionImplPolicy::FermionField> power_method;
+  auto lambda_max = power_method(linop,eta);
+  std::cout << GridLogMessage << "Upper bound of EOFA operator " << lambda_max << std::endl;
+}
+
+//Applications of M^{-1} cost the same as M for EOFA!
+template<typename FermionImplPolicy>
+class EOFAinvLinop: public LinearOperatorBase<typename FermionImplPolicy::FermionField>{
+  ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA;
+  LatticeGaugeFieldD &U;
+public:
+  EOFAinvLinop(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA, LatticeGaugeFieldD &U): EOFA(EOFA), U(U){}
+
+  typedef typename FermionImplPolicy::FermionField Field;
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); } 
+
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ assert(0); }
+  void HermOp(const Field &in, Field &out){ EOFA.MeofaInv(U, in, out); }
+};
+
+template<typename FermionImplPolicy>
+void lowerBoundEOFA(ExactOneFlavourRatioPseudoFermionAction<FermionImplPolicy> &EOFA,
+		    GridCartesian* FGrid, GridParallelRNG &rng, LatticeGaugeFieldD &latt){
+  std::cout << GridLogMessage << "Starting EOFA lower bound compute using power method on M^{-1}. Inverse of highest eigenvalue is the lowest eigenvalue of M" << std::endl;
+  EOFAinvLinop<FermionImplPolicy> linop(EOFA, latt);
+  typename FermionImplPolicy::FermionField eta(FGrid);
+  gaussian(rng,eta);
+  PowerMethod<typename FermionImplPolicy::FermionField> power_method;
+  auto lambda_max = power_method(linop,eta);
+  std::cout << GridLogMessage << "Lower bound of EOFA operator " << 1./lambda_max << std::endl;
+}
+
+
+NAMESPACE_BEGIN(Grid);
+
+  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;
+
+    using OperatorFunction<FieldD>::operator();
+
+    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.) 
+    { 
+    };
+
+    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);
+      assert(&(SchurOpU->_Mat)==&(LinOpD._Mat));
+
+      precisionChange(FermOpF.Umu, FermOpD.Umu);
+
+      pickCheckerboard(Even,FermOpF.UmuEven,FermOpF.Umu);
+      pickCheckerboard(Odd ,FermOpF.UmuOdd ,FermOpF.Umu);
+
+      ////////////////////////////////////////////////////////////////////////////////////
+      // Make a mixed precision conjugate gradient
+      ////////////////////////////////////////////////////////////////////////////////////
+      MixedPrecisionConjugateGradient<FieldD,FieldF> MPCG(Tolerance,MaxInnerIterations,MaxOuterIterations,SinglePrecGrid5,LinOpF,LinOpD);
+      MPCG.InnerTolerance = InnerTolerance;
+      std::cout << GridLogMessage << "Calling mixed precision Conjugate Gradient" <<std::endl;
+      MPCG(src,psi);
+    }
+  };
+
+
+  template<class FermionOperatorD, class FermionOperatorF, class SchurOperatorD, class  SchurOperatorF> 
+  class MixedPrecisionReliableUpdateConjugateGradientOperatorFunction : public OperatorFunction<typename FermionOperatorD::FermionField> {
+  public:
+    typedef typename FermionOperatorD::FermionField FieldD;
+    typedef typename FermionOperatorF::FermionField FieldF;
+
+    using OperatorFunction<FieldD>::operator();
+
+    RealD Tolerance;
+    Integer MaxIterations;
+
+    RealD Delta; //reliable update parameter
+
+    GridBase* SinglePrecGrid4; //Grid for single-precision fields
+    GridBase* SinglePrecGrid5; //Grid for single-precision fields
+
+    FermionOperatorF &FermOpF;
+    FermionOperatorD &FermOpD;;
+    SchurOperatorF &LinOpF;
+    SchurOperatorD &LinOpD;
+    
+    MixedPrecisionReliableUpdateConjugateGradientOperatorFunction(RealD tol, 
+								  RealD delta,
+								  Integer maxit, 
+								  GridBase* _sp_grid4, 
+								  GridBase* _sp_grid5, 
+								  FermionOperatorF &_FermOpF,
+								  FermionOperatorD &_FermOpD,
+								  SchurOperatorF   &_LinOpF,
+								  SchurOperatorD   &_LinOpD): 
+      LinOpF(_LinOpF),
+      LinOpD(_LinOpD),
+      FermOpF(_FermOpF),
+      FermOpD(_FermOpD),
+      Tolerance(tol), 
+      Delta(delta),
+      MaxIterations(maxit), 
+      SinglePrecGrid4(_sp_grid4),
+      SinglePrecGrid5(_sp_grid5)
+    { 
+    };
+
+    void operator()(LinearOperatorBase<FieldD> &LinOpU, const FieldD &src, FieldD &psi) {
+
+      std::cout << GridLogMessage << " Mixed precision reliable CG update wrapper operator() "<<std::endl;
+
+      SchurOperatorD * SchurOpU = static_cast<SchurOperatorD *>(&LinOpU);
+      assert(&(SchurOpU->_Mat)==&(LinOpD._Mat));
+
+      precisionChange(FermOpF.Umu, FermOpD.Umu);
+
+      pickCheckerboard(Even,FermOpF.UmuEven,FermOpF.Umu);
+      pickCheckerboard(Odd ,FermOpF.UmuOdd ,FermOpF.Umu);
+
+      ////////////////////////////////////////////////////////////////////////////////////
+      // Make a mixed precision conjugate gradient
+      ////////////////////////////////////////////////////////////////////////////////////
+
+      ConjugateGradientReliableUpdate<FieldD,FieldF> MPCG(Tolerance,MaxIterations,Delta,SinglePrecGrid5,LinOpF,LinOpD);
+      std::cout << GridLogMessage << "Calling mixed precision reliable update Conjugate Gradient" <<std::endl;
+      MPCG(src,psi);
+    }
+  };
+
+
+
+NAMESPACE_END(Grid);
+
+
+
+
+
+int main(int argc, char **argv) {
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+  // here make a routine to print all the relevant information on the run
+  std::cout << GridLogMessage << "Grid is setup to use " << threads << " threads" << std::endl;
+
+  std::string param_file = "params.xml";
+  bool file_load_check = false;
+  for(int i=1;i<argc;i++){
+    std::string sarg(argv[i]);
+    if(sarg == "--param_file"){
+      assert(i!=argc-1);
+      param_file = argv[i+1];
+    }else if(sarg == "--read_check"){ //check the fields load correctly and pass checksum/plaquette repro
+      file_load_check = true;
+    }
+  }
+
+  //Read the user parameters
+  EvolParameters user_params;
+  
+  if(fileExists(param_file)){
+    std::cout << GridLogMessage << " Reading " << param_file << std::endl;
+    Grid::XmlReader rd(param_file);
+    read(rd, "Params", user_params);
+  }else if(!GlobalSharedMemory::WorldRank){
+    std::cout << GridLogMessage << " File " << param_file << " does not exist" << std::endl;
+    std::cout << GridLogMessage << " Writing xml template to " << param_file << ".templ" << std::endl;
+    {
+      Grid::XmlWriter wr(param_file + ".templ");
+      write(wr, "Params", user_params);
+    }
+    std::cout << GridLogMessage << " Done" << std::endl;
+    Grid_finalize();
+    return 0;
+  }
+
+  //Check the parameters
+  if(user_params.GparityDirs.size() != Nd-1){
+    std::cerr << "Error in input parameters: expect GparityDirs to have size = " << Nd-1 << std::endl;
+    exit(1);
+  }
+  for(int i=0;i<Nd-1;i++)
+    if(user_params.GparityDirs[i] != 0 && user_params.GparityDirs[i] != 1){
+      std::cerr << "Error in input parameters: expect GparityDirs values to be 0 (periodic) or 1 (G-parity)" << std::endl;
+      exit(1);
+    }
+
+
+  typedef GparityMobiusEOFAFermionD EOFAactionD;
+  typedef GparityMobiusFermionD FermionActionD;
+  typedef typename FermionActionD::Impl_t FermionImplPolicyD;
+  typedef typename FermionActionD::FermionField FermionFieldD;
+
+  typedef GparityMobiusEOFAFermionF EOFAactionF;
+  typedef GparityMobiusFermionF FermionActionF;
+  typedef typename FermionActionF::Impl_t FermionImplPolicyF;
+  typedef typename FermionActionF::FermionField FermionFieldF;
+
+  typedef GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicyD,FermionImplPolicyF> MixedPrecRHMC;
+  typedef GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicyD> DoublePrecRHMC;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  typedef ConjugateHMCRunnerD<MinimumNorm2> HMCWrapper; //NB: This is the "Omelyan integrator"
+  typedef HMCWrapper::ImplPolicy GaugeImplPolicy;
+  MD.name    = std::string("MinimumNorm2");
+  MD.MDsteps = user_params.Steps;
+  MD.trajL   = user_params.TrajectoryLength;
+
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = user_params.StartTrajectory;
+  HMCparams.Trajectories     = user_params.Trajectories;
+  HMCparams.NoMetropolisUntil= 0;
+  HMCparams.StartingType     = user_params.StartingType;
+  HMCparams.MetropolisTest = user_params.MetropolisTest;
+  HMCparams.MD = MD;
+  HMCWrapper TheHMC(HMCparams);
+
+  // Grid from the command line arguments --grid and --mpi
+  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
+
+  CheckpointerParameters CPparams;
+  CPparams.config_prefix = "ckpoint_lat";
+  CPparams.rng_prefix    = "ckpoint_rng";
+  CPparams.saveInterval  = user_params.SaveInterval;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  //Note that checkpointing saves the RNG state so that this initialization is required only for the very first configuration
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = "1 2 3 4 5";
+  RNGpar.parallel_seeds = "6 7 8 9 10";
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  typedef PlaquetteMod<GaugeImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+
+  //aiming for ainv=2.068             me          Bob
+  //Estimated  a(ml+mres) [48ID] = 0.001048    0.00104 
+  //           a(mh+mres) [48ID] = 0.028847    0.02805
+  //Estimate Ls=12, b+c=2  mres~0.0003
+
+  const int Ls      = 12;
+  Real beta         = 1.946;
+  Real light_mass   = 0.00074;   //0.00104 - mres_approx;
+  Real strange_mass = 0.02775;    //0.02805 - mres_approx
+  Real pv_mass      = 1.0;
+  RealD M5  = 1.8;
+  RealD mobius_scale = 2.; //b+c
+
+  RealD mob_bmc = 1.0;
+  RealD mob_b = (mobius_scale + mob_bmc)/2.;
+  RealD mob_c = (mobius_scale - mob_bmc)/2.;
+
+  //Setup the Grids
+  auto UGridD   = TheHMC.Resources.GetCartesian();
+  auto UrbGridD = TheHMC.Resources.GetRBCartesian();
+  auto FGridD     = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridD);
+  auto FrbGridD   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridD);
+
+  GridCartesian* UGridF = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian* UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  auto FGridF     = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridF);
+  auto FrbGridF   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridF);
+
+  ConjugateIwasakiGaugeActionD GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeFieldD Ud(UGridD);
+  LatticeGaugeFieldF Uf(UGridF);
+ 
+  //Setup the BCs
+  FermionActionD::ImplParams Params;
+  for(int i=0;i<Nd-1;i++) Params.twists[i] = user_params.GparityDirs[i]; //G-parity directions
+  Params.twists[Nd-1] = 1; //APBC in time direction
+
+  std::vector<int> dirs4(Nd);
+  for(int i=0;i<Nd-1;i++) dirs4[i] = user_params.GparityDirs[i];
+  dirs4[Nd-1] = 0; //periodic gauge BC in time
+
+  GaugeImplPolicy::setDirections(dirs4); //gauge BC
+
+  //Run optional gauge field checksum checker and exit
+  if(file_load_check){
+    TheHMC.initializeGaugeFieldAndRNGs(Ud);
+    std::cout << GridLogMessage << " Done" << std::endl;
+    Grid_finalize();
+    return 0;
+  }
+
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1); //light quark + strange quark
+  ActionLevel<HMCWrapper::Field> Level2(4); //DSDR
+  ActionLevel<HMCWrapper::Field> Level3(2); //gauge
+
+
+  /////////////////////////////////////////////////////////////
+  // Light EOFA action
+  // have to be careful with the parameters, cf. Test_dwf_gpforce_eofa.cc
+  /////////////////////////////////////////////////////////////
+  typedef SchurDiagMooeeOperator<EOFAactionD,FermionFieldD> EOFAschuropD;
+  typedef SchurDiagMooeeOperator<EOFAactionF,FermionFieldF> EOFAschuropF;
+  typedef ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction<FermionImplPolicyD, FermionImplPolicyF> EOFAmixPrecPFaction;
+  typedef MixedPrecisionConjugateGradientOperatorFunction<EOFAactionD, EOFAactionF, EOFAschuropD, EOFAschuropF> EOFA_mxCG;
+  typedef MixedPrecisionReliableUpdateConjugateGradientOperatorFunction<EOFAactionD, EOFAactionF, EOFAschuropD, EOFAschuropF> EOFA_relupCG;
+  
+  std::vector<RealD> eofa_light_masses = { light_mass ,  0.004,   0.016,   0.064,   0.256    };
+  std::vector<RealD> eofa_pv_masses =    { 0.004       , 0.016,   0.064,   0.256,   1.0      };
+  int n_light_hsb = 5;
+  assert(user_params.eofa_l.size() == n_light_hsb);
+  
+  EOFAmixPrecPFaction* EOFA_pfactions[n_light_hsb];
+
+  for(int i=0;i<n_light_hsb;i++){
+    RealD iml = eofa_light_masses[i];
+    RealD ipv = eofa_pv_masses[i];
+
+    EOFAactionD* LopD = new EOFAactionD(Ud, *FGridD, *FrbGridD, *UGridD, *UrbGridD, iml, iml, ipv, 0.0, -1, M5, mob_b, mob_c, Params);
+    EOFAactionF* LopF = new EOFAactionF(Uf, *FGridF, *FrbGridF, *UGridF, *UrbGridF, iml, iml, ipv, 0.0, -1, M5, mob_b, mob_c, Params);
+    EOFAactionD* RopD = new EOFAactionD(Ud, *FGridD, *FrbGridD, *UGridD, *UrbGridD, ipv, iml, ipv, -1.0, 1, M5, mob_b, mob_c, Params);
+    EOFAactionF* RopF = new EOFAactionF(Uf, *FGridF, *FrbGridF, *UGridF, *UrbGridF, ipv, iml, ipv, -1.0, 1, M5, mob_b, mob_c, Params);
+
+    EOFAschuropD* linopL_D = new EOFAschuropD(*LopD);
+    EOFAschuropD* linopR_D = new EOFAschuropD(*RopD);
+    
+    EOFAschuropF* linopL_F = new EOFAschuropF(*LopF);
+    EOFAschuropF* linopR_F = new EOFAschuropF(*RopF);
+
+#if 1
+    //Note reusing user_params.eofa_l.action(|md)_mixcg_inner_tolerance  as Delta for now
+    EOFA_relupCG* ActionMCG_L = new EOFA_relupCG(user_params.eofa_l[i].action_tolerance, user_params.eofa_l[i].action_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    EOFA_relupCG* ActionMCG_R = new EOFA_relupCG(user_params.eofa_l[i].action_tolerance, user_params.eofa_l[i].action_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+
+    EOFA_relupCG* DerivMCG_L = new EOFA_relupCG(user_params.eofa_l[i].md_tolerance, user_params.eofa_l[i].md_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    EOFA_relupCG* DerivMCG_R = new EOFA_relupCG(user_params.eofa_l[i].md_tolerance, user_params.eofa_l[i].md_mixcg_inner_tolerance, 50000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+
+#else
+    
+    EOFA_mxCG* ActionMCG_L = new EOFA_mxCG(user_params.eofa_l[i].action_tolerance, 10000, 1000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    ActionMCG_L->InnerTolerance = user_params.eofa_l[i].action_mixcg_inner_tolerance;
+    
+    EOFA_mxCG* ActionMCG_R = new EOFA_mxCG(user_params.eofa_l[i].action_tolerance, 10000, 1000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+    ActionMCG_R->InnerTolerance = user_params.eofa_l[i].action_mixcg_inner_tolerance;
+    
+    EOFA_mxCG* DerivMCG_L = new EOFA_mxCG(user_params.eofa_l[i].md_tolerance, 10000, 1000, UGridF, FrbGridF, *LopF, *LopD, *linopL_F, *linopL_D);
+    DerivMCG_L->InnerTolerance = user_params.eofa_l[i].md_mixcg_inner_tolerance;
+    
+    EOFA_mxCG* DerivMCG_R = new EOFA_mxCG(user_params.eofa_l[i].md_tolerance, 10000, 1000, UGridF, FrbGridF, *RopF, *RopD, *linopR_F, *linopR_D);
+    DerivMCG_R->InnerTolerance = user_params.eofa_l[i].md_mixcg_inner_tolerance;
+    
+    std::cout << GridLogMessage << "Set EOFA action solver action tolerance outer=" << ActionMCG_L->Tolerance << " inner=" << ActionMCG_L->InnerTolerance << std::endl;
+    std::cout << GridLogMessage << "Set EOFA MD solver tolerance outer=" << DerivMCG_L->Tolerance << " inner=" << DerivMCG_L->InnerTolerance << std::endl;
+#endif
+
+    
+    EOFAmixPrecPFaction* EOFA = new EOFAmixPrecPFaction(*LopF, *RopF,
+							*LopD, *RopD, 
+							*ActionMCG_L, *ActionMCG_R, 
+							*ActionMCG_L, *ActionMCG_R, 
+							*DerivMCG_L, *DerivMCG_R, 
+							user_params.eofa_l[i].rat_params, true);
+    EOFA_pfactions[i] = EOFA;
+    Level1.push_back(EOFA);
+  }
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionActionD Numerator_sD(Ud,*FGridD,*FrbGridD,*UGridD,*UrbGridD,strange_mass,M5,mob_b,mob_c,Params);
+  FermionActionD Denominator_sD(Ud,*FGridD,*FrbGridD,*UGridD,*UrbGridD, pv_mass,M5,mob_b,mob_c,Params);
+
+  FermionActionF Numerator_sF(Uf,*FGridF,*FrbGridF,*UGridF,*UrbGridF,strange_mass,M5,mob_b,mob_c,Params);
+  FermionActionF Denominator_sF(Uf,*FGridF,*FrbGridF,*UGridF,*UrbGridF, pv_mass,M5,mob_b,mob_c,Params);
+
+  RationalActionParams rat_act_params_s;
+  rat_act_params_s.inv_pow  = 4; // (M^dag M)^{1/4}
+  rat_act_params_s.precision= 60;
+  rat_act_params_s.MaxIter  = 10000;
+  user_params.rat_quo_s.Export(rat_act_params_s);
+  std::cout << GridLogMessage << " Heavy quark bounds check every " << rat_act_params_s.BoundsCheckFreq << " trajectories (avg)" << std::endl;
+
+  //MixedPrecRHMC Quotient_s(Denominator_sD, Numerator_sD, Denominator_sF, Numerator_sF, rat_act_params_s, user_params.rat_quo_s.reliable_update_freq); 
+  DoublePrecRHMC Quotient_s(Denominator_sD, Numerator_sD, rat_act_params_s); 
+  Level1.push_back(&Quotient_s);  
+
+  ///////////////////////////////////
+  // DSDR action
+  ///////////////////////////////////
+  RealD dsdr_mass=-1.8;   
+  //Use same DSDR twists as https://arxiv.org/pdf/1208.4412.pdf
+  RealD dsdr_epsilon_f = 0.02; //numerator (in determinant)
+  RealD dsdr_epsilon_b = 0.5; 
+  GparityWilsonTMFermionD Numerator_DSDR_D(Ud, *UGridD, *UrbGridD, dsdr_mass, dsdr_epsilon_f, Params);
+  GparityWilsonTMFermionF Numerator_DSDR_F(Uf, *UGridF, *UrbGridF, dsdr_mass, dsdr_epsilon_f, Params);
+
+  GparityWilsonTMFermionD Denominator_DSDR_D(Ud, *UGridD, *UrbGridD, dsdr_mass, dsdr_epsilon_b, Params);
+  GparityWilsonTMFermionF Denominator_DSDR_F(Uf, *UGridF, *UrbGridF, dsdr_mass, dsdr_epsilon_b, Params);
+ 
+  RationalActionParams rat_act_params_DSDR;
+  rat_act_params_DSDR.inv_pow  = 2; // (M^dag M)^{1/2}
+  rat_act_params_DSDR.precision= 60;
+  rat_act_params_DSDR.MaxIter  = 10000;
+  user_params.rat_quo_DSDR.Export(rat_act_params_DSDR);
+  std::cout << GridLogMessage << "DSDR quark bounds check every " << rat_act_params_DSDR.BoundsCheckFreq << " trajectories (avg)" << std::endl;
+
+  DoublePrecRHMC Quotient_DSDR(Denominator_DSDR_D, Numerator_DSDR_D, rat_act_params_DSDR);
+  Level2.push_back(&Quotient_DSDR);
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+
+
+  //Action tuning
+  bool 
+    tune_rhmc_s=false, eigenrange_s=false, 
+    tune_rhmc_DSDR=false, eigenrange_DSDR=false, 
+    check_eofa=false, 
+    upper_bound_eofa=false, lower_bound_eofa(false);
+
+  std::string lanc_params_s;
+  std::string lanc_params_DSDR;
+  int tune_rhmc_s_action_or_md;
+  int tune_rhmc_DSDR_action_or_md;
+  int eofa_which_hsb;
+
+  for(int i=1;i<argc;i++){
+    std::string sarg(argv[i]);
+    if(sarg == "--tune_rhmc_s"){
+      assert(i < argc-1);
+      tune_rhmc_s=true;
+      tune_rhmc_s_action_or_md = std::stoi(argv[i+1]);
+    }
+    else if(sarg == "--eigenrange_s"){
+      assert(i < argc-1);
+      eigenrange_s=true;
+      lanc_params_s = argv[i+1];
+    }
+    else if(sarg == "--tune_rhmc_DSDR"){
+      assert(i < argc-1);
+      tune_rhmc_DSDR=true;
+      tune_rhmc_DSDR_action_or_md = std::stoi(argv[i+1]);
+    }
+    else if(sarg == "--eigenrange_DSDR"){
+      assert(i < argc-1);
+      eigenrange_DSDR=true;
+      lanc_params_DSDR = argv[i+1];
+    }
+    else if(sarg == "--check_eofa"){
+      assert(i < argc-1);
+      check_eofa = true;
+      eofa_which_hsb = std::stoi(argv[i+1]); //-1 indicates all hasenbusch
+      assert(eofa_which_hsb == -1 || (eofa_which_hsb >= 0 && eofa_which_hsb < n_light_hsb) );
+    }
+    else if(sarg == "--upper_bound_eofa"){
+      assert(i < argc-1);
+      upper_bound_eofa = true;
+      eofa_which_hsb = std::stoi(argv[i+1]);
+      assert(eofa_which_hsb >= 0 && eofa_which_hsb < n_light_hsb);
+    }
+    else if(sarg == "--lower_bound_eofa"){
+      assert(i < argc-1);
+      lower_bound_eofa = true;      
+      eofa_which_hsb = std::stoi(argv[i+1]);
+      assert(eofa_which_hsb >= 0 && eofa_which_hsb < n_light_hsb);
+    }
+  }
+  if(tune_rhmc_s || eigenrange_s || tune_rhmc_DSDR || eigenrange_DSDR ||check_eofa || upper_bound_eofa || lower_bound_eofa) {
+    std::cout << GridLogMessage << "Running checks" << std::endl;
+    TheHMC.initializeGaugeFieldAndRNGs(Ud);
+
+    //std::cout << GridLogMessage << "EOFA action solver action tolerance outer=" << ActionMCG_L.Tolerance << " inner=" << ActionMCG_L.InnerTolerance << std::endl;
+    //std::cout << GridLogMessage << "EOFA MD solver tolerance outer=" << DerivMCG_L.Tolerance << " inner=" << DerivMCG_L.InnerTolerance << std::endl;
+
+
+    if(check_eofa){
+      if(eofa_which_hsb >= 0){
+	std::cout << GridLogMessage << "Starting checking EOFA Hasenbusch " << eofa_which_hsb << std::endl;
+	checkEOFA(*EOFA_pfactions[eofa_which_hsb], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+	std::cout << GridLogMessage << "Finished checking EOFA Hasenbusch " << eofa_which_hsb << std::endl;
+      }else{
+	for(int i=0;i<n_light_hsb;i++){
+	  std::cout << GridLogMessage << "Starting checking EOFA Hasenbusch " << i << std::endl;
+	  checkEOFA(*EOFA_pfactions[i], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+	  std::cout << GridLogMessage << "Finished checking EOFA Hasenbusch " << i << std::endl;
+	}
+      }
+    }	  
+    if(upper_bound_eofa) upperBoundEOFA(*EOFA_pfactions[eofa_which_hsb], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+    if(lower_bound_eofa) lowerBoundEOFA(*EOFA_pfactions[eofa_which_hsb], FGridD, TheHMC.Resources.GetParallelRNG(), Ud);
+    if(eigenrange_s) computeEigenvalues<FermionActionD, FermionFieldD>(lanc_params_s, FGridD, FrbGridD, Ud, Numerator_sD, TheHMC.Resources.GetParallelRNG());
+    if(tune_rhmc_s) checkRHMC<FermionActionD, FermionFieldD, decltype(Quotient_s)>(FGridD, FrbGridD, Ud, Numerator_sD, Denominator_sD, Quotient_s, TheHMC.Resources.GetParallelRNG(), 4, "strange",  tune_rhmc_s_action_or_md);
+    if(eigenrange_DSDR) computeEigenvalues<GparityWilsonTMFermionD, GparityWilsonTMFermionD::FermionField>(lanc_params_DSDR, UGridD, UrbGridD, Ud, Numerator_DSDR_D, TheHMC.Resources.GetParallelRNG());
+    if(tune_rhmc_DSDR) checkRHMC<GparityWilsonTMFermionD, GparityWilsonTMFermionD::FermionField, decltype(Quotient_DSDR)>(UGridD, UrbGridD, Ud, Numerator_DSDR_D, Denominator_DSDR_D, Quotient_DSDR, TheHMC.Resources.GetParallelRNG(), 2, "DSDR", tune_rhmc_DSDR_action_or_md);
+
+
+    std::cout << GridLogMessage << " Done" << std::endl;
+    Grid_finalize();
+    return 0;
+  }
+
+
+  //Run the HMC
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.Run();
+
+  std::cout << GridLogMessage << " Done" << std::endl;
+  Grid_finalize();
+  return 0;
+} // main

HMC/Mobius2p1f_DD_RHMC.cc

@@ -0,0 +1,267 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_hmc_EODWFRatio.cc
+
+Copyright (C) 2015-2016
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Guido Cossu <guido.cossu@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 */
+#include <Grid/Grid.h>
+
+int main(int argc, char **argv) {
+  using namespace Grid;
+
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+
+   // Typedefs to simplify notation
+  typedef WilsonImplR FermionImplPolicy;
+  typedef MobiusFermionR FermionAction;
+  typedef typename FermionAction::FermionField FermionField;
+
+  typedef Grid::XmlReader       Serialiser;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  //  typedef GenericHMCRunner<LeapFrog> HMCWrapper;
+  //  MD.name    = std::string("Leap Frog");
+  //  typedef GenericHMCRunner<ForceGradient> HMCWrapper;
+  //  MD.name    = std::string("Force Gradient");
+  typedef GenericHMCRunner<MinimumNorm2> HMCWrapper;
+  MD.name    = std::string("MinimumNorm2");
+  MD.MDsteps =  4;
+  MD.trajL   = 1.0;
+
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = 17;
+  HMCparams.Trajectories     = 200;
+  HMCparams.NoMetropolisUntil=  0;
+  // "[HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+  //  HMCparams.StartingType     =std::string("ColdStart");
+  HMCparams.StartingType     =std::string("CheckpointStart");
+  HMCparams.MD = MD;
+  HMCWrapper TheHMC(HMCparams);
+
+  // Grid from the command line arguments --grid and --mpi
+  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
+
+  CheckpointerParameters CPparams;
+  CPparams.config_prefix = "ckpoint_DDHMC_lat";
+  CPparams.rng_prefix    = "ckpoint_DDHMC_rng";
+  CPparams.saveInterval  = 1;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = "1 2 3 4 5";
+  RNGpar.parallel_seeds = "6 7 8 9 10";
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  // Construct observables
+  // here there is too much indirection
+  typedef PlaquetteMod<HMCWrapper::ImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+
+  const int Ls      = 16;
+  RealD M5  = 1.8;
+  RealD b   = 1.0;
+  RealD c   = 0.0;
+  Real beta         = 2.13;
+  Real light_mass   = 0.01;
+  Real strange_mass = 0.04;
+  Real pv_mass      = 1.0;
+  std::vector<Real> hasenbusch({ light_mass, 0.04, 0.25, 0.4, 0.7 , pv_mass });
+
+  // FIXME:
+  // Same in MC and MD
+  // Need to mix precision too
+  OneFlavourRationalParams SFRp;
+  SFRp.lo       = 4.0e-3;
+  SFRp.hi       = 30.0;
+  SFRp.MaxIter  = 10000;
+  SFRp.tolerance= 1.0e-8;
+  SFRp.mdtolerance= 1.0e-5;
+  SFRp.degree   = 16;
+  SFRp.precision= 50;
+  SFRp.BoundsCheckFreq=5;
+
+  OneFlavourRationalParams OFRp;
+  OFRp.lo       = 1.0e-4;
+  OFRp.hi       = 30.0;
+  OFRp.MaxIter  = 10000;
+  OFRp.tolerance= 1.0e-8;
+  OFRp.mdtolerance= 1.0e-5;
+  OFRp.degree   = 16;
+  OFRp.precision= 50;
+  OFRp.BoundsCheckFreq=5;
+
+  auto GridPtr   = TheHMC.Resources.GetCartesian();
+  auto GridRBPtr = TheHMC.Resources.GetRBCartesian();
+
+  ////////////////////////////////////////////////////////////////
+  // Domain decomposed
+  ////////////////////////////////////////////////////////////////
+  Coordinate latt4  = GridPtr->GlobalDimensions();
+  Coordinate mpi    = GridPtr->ProcessorGrid();
+  Coordinate shm;
+
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  
+  Coordinate CommDim(Nd);
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+
+  Coordinate Dirichlet(Nd+1,0);
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Coordinate Block4(Nd);
+  Block4[0] = Dirichlet[1];
+  Block4[1] = Dirichlet[2];
+  Block4[2] = Dirichlet[3];
+  Block4[3] = Dirichlet[4];
+  int Width=3;
+  TheHMC.Resources.SetMomentumFilter(new DDHMCFilter<WilsonImplR::Field>(Block4,Width));
+
+  //////////////////////////
+  // Fermion Grid
+  //////////////////////////
+  auto FGrid     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr);
+  auto FrbGrid   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr);
+
+  IwasakiGaugeActionR GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeField U(GridPtr);
+
+  // These lines are unecessary if BC are all periodic
+  std::vector<Complex> boundary = {1,1,1,-1};
+  FermionAction::ImplParams Params(boundary);
+
+  double StoppingCondition = 1e-8;
+  double MDStoppingCondition = 1e-6;
+  double MaxCGIterations = 30000;
+  ConjugateGradient<FermionField>  CG(StoppingCondition,MaxCGIterations);
+  ConjugateGradient<FermionField>  MDCG(MDStoppingCondition,MaxCGIterations);
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1);
+  ActionLevel<HMCWrapper::Field> Level2(4);
+  ActionLevel<HMCWrapper::Field> Level3(8);
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionAction StrangeOp (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params);
+  FermionAction StrangePauliVillarsOp(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, Params);
+
+  FermionAction StrangeOpDir (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params);
+  FermionAction StrangePauliVillarsOpDir(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, Params);
+  StrangeOpDir.DirichletBlock(Dirichlet);  
+  StrangePauliVillarsOpDir.DirichletBlock(Dirichlet);  
+  
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionBdy(StrangeOpDir,StrangeOp,SFRp);
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionLocal(StrangePauliVillarsOpDir,StrangeOpDir,SFRp);
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionPVBdy(StrangePauliVillarsOp,StrangePauliVillarsOpDir,SFRp);
+  Level1.push_back(&StrangePseudoFermionBdy);
+  Level2.push_back(&StrangePseudoFermionLocal);
+  Level1.push_back(&StrangePseudoFermionPVBdy);
+
+  ////////////////////////////////////
+  // up down action
+  ////////////////////////////////////
+  std::vector<Real> light_den;
+  std::vector<Real> light_num;
+  std::vector<int> dirichlet_den;
+  std::vector<int> dirichlet_num;
+
+  int n_hasenbusch = hasenbusch.size();
+  light_den.push_back(light_mass);  dirichlet_den.push_back(0);
+  for(int h=0;h<n_hasenbusch;h++){
+    light_den.push_back(hasenbusch[h]); dirichlet_den.push_back(1);
+  }
+
+  for(int h=0;h<n_hasenbusch;h++){
+    light_num.push_back(hasenbusch[h]); dirichlet_num.push_back(1);
+  }
+  light_num.push_back(pv_mass);  dirichlet_num.push_back(0);
+
+  std::vector<FermionAction *> Numerators;
+  std::vector<FermionAction *> Denominators;
+  std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
+  std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
+  
+  for(int h=0;h<n_hasenbusch+1;h++){
+    std::cout << GridLogMessage
+	      << " 2f quotient Action ";
+    std::cout << "det D("<<light_den[h]<<")";
+    if ( dirichlet_den[h] ) std::cout << "^dirichlet    ";
+    std::cout << "/ det D("<<light_num[h]<<")";
+    if ( dirichlet_num[h] ) std::cout << "^dirichlet    ";
+    std::cout << std::endl;
+    
+    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));
+    if(h!=0) {
+      Quotients.push_back   (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],MDCG,CG));
+    } else {
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+    }
+    if ( dirichlet_den[h]==1) Denominators[h]->DirichletBlock(Dirichlet);
+    if ( dirichlet_num[h]==1) Numerators[h]->DirichletBlock(Dirichlet);
+  }
+
+  int nquo=Quotients.size();
+  Level1.push_back(Bdys[0]);
+  Level1.push_back(Bdys[1]);
+  for(int h=0;h<nquo-1;h++){
+    Level2.push_back(Quotients[h]);
+  }
+  Level2.push_back(Quotients[nquo-1]);
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+
+  /////////////////////////////////////////////////////////////
+
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.ReadCommandLine(argc,argv);  // params on CML or from param file
+  TheHMC.Run();  // no smearing
+
+  Grid_finalize();
+} // main
+
+
+

HMC/Mobius2p1f_DD_RHMC_96I.cc

@@ -0,0 +1,419 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_hmc_EODWFRatio.cc
+
+Copyright (C) 2015-2016
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Guido Cossu <guido.cossu@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 */
+#include <Grid/Grid.h>
+
+int main(int argc, char **argv) {
+  using namespace Grid;
+
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+
+   // Typedefs to simplify notation
+  typedef WilsonImplR FermionImplPolicy;
+  typedef MobiusFermionR FermionAction;
+  typedef typename FermionAction::FermionField FermionField;
+
+  typedef Grid::XmlReader       Serialiser;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  //  typedef GenericHMCRunner<LeapFrog> HMCWrapper;
+  //  MD.name    = std::string("Leap Frog");
+  //  typedef GenericHMCRunner<ForceGradient> HMCWrapper;
+  //  MD.name    = std::string("Force Gradient");
+  typedef GenericHMCRunner<MinimumNorm2> HMCWrapper;
+  MD.name    = std::string("MinimumNorm2");
+  MD.MDsteps =  6;
+  MD.trajL   = 1.0;
+
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = 1077;
+  HMCparams.Trajectories     = 1;
+  HMCparams.NoMetropolisUntil=  0;
+  // "[HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+  //  HMCparams.StartingType     =std::string("ColdStart");
+  HMCparams.StartingType     =std::string("CheckpointStart");
+  HMCparams.MD = MD;
+  HMCWrapper TheHMC(HMCparams);
+
+  // Grid from the command line arguments --grid and --mpi
+  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
+
+  CheckpointerParameters CPparams;
+  CPparams.config_prefix = "ckpoint_DDHMC_lat";
+  CPparams.rng_prefix    = "ckpoint_DDHMC_rng";
+  CPparams.saveInterval  = 1;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = "1 2 3 4 5";
+  RNGpar.parallel_seeds = "6 7 8 9 10";
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  // Construct observables
+  // here there is too much indirection
+  typedef PlaquetteMod<HMCWrapper::ImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+
+  const int Ls      = 12;
+  RealD M5  = 1.8;
+  RealD b   = 1.5;
+  RealD c   = 0.5;
+  //  Real beta         = 2.31;
+  //  Real light_mass   = 5.4e-4;
+  Real beta         = 2.13;
+  Real light_mass   = 7.8e-4;
+  Real strange_mass = 0.02132;
+  Real pv_mass      = 1.0;
+  //  std::vector<Real> hasenbusch({ light_mass, 3.8e-3, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass });
+  std::vector<Real> hasenbusch({ light_mass, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass });
+
+  // FIXME:
+  // Same in MC and MD
+  // Need to mix precision too
+  OneFlavourRationalParams SFRp; // Strange
+  SFRp.lo       = 4.0e-3;
+  SFRp.hi       = 90.0;
+  SFRp.MaxIter  = 60000;
+  SFRp.tolerance= 1.0e-8;
+  SFRp.mdtolerance= 1.0e-4;
+  SFRp.degree   = 12;
+  SFRp.precision= 50;
+  SFRp.BoundsCheckFreq=0;
+
+  OneFlavourRationalParams OFRp; // Up/down
+  OFRp.lo       = 2.0e-5;
+  OFRp.hi       = 90.0;
+  OFRp.MaxIter  = 60000;
+  OFRp.tolerance= 1.0e-7;
+  OFRp.mdtolerance= 1.0e-4;
+  //  OFRp.degree   = 20; converges
+  //  OFRp.degree   = 16;
+  OFRp.degree   = 12;
+  OFRp.precision= 80;
+  OFRp.BoundsCheckFreq=0;
+
+  auto GridPtr   = TheHMC.Resources.GetCartesian();
+  auto GridRBPtr = TheHMC.Resources.GetRBCartesian();
+
+  ////////////////////////////////////////////////////////////////
+  // Domain decomposed
+  ////////////////////////////////////////////////////////////////
+  Coordinate latt4  = GridPtr->GlobalDimensions();
+  Coordinate mpi    = GridPtr->ProcessorGrid();
+  Coordinate shm;
+
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  
+  Coordinate CommDim(Nd);
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+
+  Coordinate NonDirichlet(Nd+1,0);
+  Coordinate Dirichlet(Nd+1,0);
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Coordinate Block4(Nd);
+  //  Block4[0] = Dirichlet[1];
+  //  Block4[1] = Dirichlet[2];
+  //  Block4[2] = Dirichlet[3];
+  Block4[0] = 0;
+  Block4[1] = 0;
+  Block4[2] = 0;
+  Block4[3] = Dirichlet[4];
+
+  int Width=3;
+  TheHMC.Resources.SetMomentumFilter(new DDHMCFilter<WilsonImplR::Field>(Block4,Width));
+
+  //////////////////////////
+  // Fermion Grid
+  //////////////////////////
+  auto FGrid     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr);
+  auto FrbGrid   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr);
+
+  IwasakiGaugeActionR GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeField U(GridPtr);
+
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.ReadCommandLine(argc,argv);  // params on CML or from param file
+  TheHMC.initializeGaugeFieldAndRNGs(U);
+
+
+  // These lines are unecessary if BC are all periodic
+  std::vector<Complex> boundary = {1,1,1,-1};
+  FermionAction::ImplParams Params(boundary);
+  Params.dirichlet=NonDirichlet;
+  FermionAction::ImplParams ParamsDir(boundary);
+  ParamsDir.dirichlet=Dirichlet;
+
+  //  double StoppingCondition = 1e-14;
+  //  double MDStoppingCondition = 1e-9;
+  double StoppingCondition = 1e-8;
+  double MDStoppingCondition = 1e-6;
+  double MaxCGIterations = 300000;
+  ConjugateGradient<FermionField>  CG(StoppingCondition,MaxCGIterations);
+  ConjugateGradient<FermionField>  MDCG(MDStoppingCondition,MaxCGIterations);
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1);
+  ActionLevel<HMCWrapper::Field> Level2(4);
+  ActionLevel<HMCWrapper::Field> Level3(8);
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionAction StrangeOp (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params);
+  FermionAction StrangePauliVillarsOp(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, Params);
+
+  FermionAction StrangeOpDir (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, ParamsDir);
+  FermionAction StrangePauliVillarsOpDir(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, ParamsDir);
+  
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionBdy(StrangeOpDir,StrangeOp,SFRp);
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionLocal(StrangePauliVillarsOpDir,StrangeOpDir,SFRp);
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionPVBdy(StrangePauliVillarsOp,StrangePauliVillarsOpDir,SFRp);
+  Level1.push_back(&StrangePseudoFermionBdy);
+  Level2.push_back(&StrangePseudoFermionLocal);
+  Level1.push_back(&StrangePseudoFermionPVBdy);
+
+  ////////////////////////////////////
+  // up down action
+  ////////////////////////////////////
+  std::vector<Real> light_den;
+  std::vector<Real> light_num;
+  std::vector<int> dirichlet_den;
+  std::vector<int> dirichlet_num;
+
+  int n_hasenbusch = hasenbusch.size();
+  light_den.push_back(light_mass);  dirichlet_den.push_back(0);
+  for(int h=0;h<n_hasenbusch;h++){
+    light_den.push_back(hasenbusch[h]); dirichlet_den.push_back(1);
+  }
+
+  for(int h=0;h<n_hasenbusch;h++){
+    light_num.push_back(hasenbusch[h]); dirichlet_num.push_back(1);
+  }
+  light_num.push_back(pv_mass);  dirichlet_num.push_back(0);
+
+  std::vector<FermionAction *> Numerators;
+  std::vector<FermionAction *> Denominators;
+  std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
+  std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
+  
+  for(int h=0;h<n_hasenbusch+1;h++){
+    std::cout << GridLogMessage
+	      << " 2f quotient Action ";
+    std::cout << "det D("<<light_den[h]<<")";
+    if ( dirichlet_den[h] ) std::cout << "^dirichlet    ";
+    std::cout << "/ det D("<<light_num[h]<<")";
+    if ( dirichlet_num[h] ) std::cout << "^dirichlet    ";
+    std::cout << std::endl;
+
+    FermionAction::ImplParams ParamsNum(boundary);
+    FermionAction::ImplParams ParamsDen(boundary);
+    
+    if ( dirichlet_num[h]==1) ParamsNum.dirichlet = Dirichlet;
+    else                      ParamsNum.dirichlet = NonDirichlet;
+    Numerators.push_back  (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, ParamsNum));
+
+    if ( dirichlet_den[h]==1) ParamsDen.dirichlet = Dirichlet;
+    else                      ParamsDen.dirichlet = NonDirichlet;
+    Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, ParamsDen));
+    
+    if(h!=0) {
+      Quotients.push_back   (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],MDCG,CG));
+    } else {
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+    }
+  }
+
+  int nquo=Quotients.size();
+  Level1.push_back(Bdys[0]);
+  Level1.push_back(Bdys[1]);
+  for(int h=0;h<nquo-1;h++){
+    Level2.push_back(Quotients[h]);
+  }
+  Level2.push_back(Quotients[nquo-1]);
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+  /////////////////////////////////////////////////////////////
+
+  if(1){
+    // TODO:
+    // i)  Break high bound, how rapidly does it break? Tune this test.
+    // ii) Break low bound, how rapidly?
+    // iii) Run lanczos
+    // iv)  Have CG return spectral range estimate
+    FermionField vec(StrangeOp.FermionRedBlackGrid());
+    FermionField res(StrangeOp.FermionRedBlackGrid());
+    vec = 1; // Fill with any old junk
+
+    std::cout << "Bounds check on strange operator mass "<< StrangeOp.Mass()<<std::endl;
+    SchurDifferentiableOperator<FermionImplPolicy> SdagS(StrangeOp);
+    HighBoundCheck(SdagS,vec,SFRp.hi);
+    ChebyBoundsCheck(SdagS,vec,SFRp.lo,SFRp.hi);
+    std::cout << "Strange inversion"<<std::endl;
+    res=Zero();
+    //    MDCG(SdagS,vec,res);
+
+
+    std::cout << "Bounds check on light quark operator mass "<< Denominators[0]->Mass() <<std::endl;
+    SchurDifferentiableOperator<FermionImplPolicy> UdagU(*Denominators[0]);
+    HighBoundCheck(UdagU,vec,OFRp.hi);
+    ChebyBoundsCheck(UdagU,vec,OFRp.lo,OFRp.hi);
+    std::cout << "light inversion"<<std::endl;
+    res=Zero();
+    //    MDCG(UdagU,vec,res);
+
+
+    std::cout << "Bounds check on strange dirichlet operator mass "<< StrangeOpDir.Mass()<<std::endl;
+    SchurDifferentiableOperator<FermionImplPolicy> SddagSd(StrangeOpDir);
+    HighBoundCheck(SddagSd,vec,OFRp.hi);
+    ChebyBoundsCheck(SddagSd,vec,OFRp.lo,OFRp.hi);
+    std::cout << "strange dirichlet inversion"<<std::endl;
+    res=Zero();
+    //    MDCG(SddagSd,vec,res);
+
+    std::cout << "Bounds check on light dirichlet operator mass "<< Numerators[0]->Mass()<<std::endl;
+    SchurDifferentiableOperator<FermionImplPolicy> UddagUd(*Numerators[0]);
+    HighBoundCheck(UddagUd,vec,OFRp.hi);
+    ChebyBoundsCheck(UddagUd,vec,OFRp.lo,OFRp.hi);
+    std::cout << "light dirichlet inversion"<<std::endl;
+    res=Zero();
+    //MDCG(UddagUd,vec,res);
+
+    
+    auto grid4= GridPtr;
+    auto rbgrid4= GridRBPtr;
+    auto rbgrid = StrangeOp.FermionRedBlackGrid();
+    auto grid = StrangeOp.FermionGrid();
+    if(1){
+    const int Nstop = 5;
+    const int Nk = 20;
+    const int Np = 20;
+    const int Nm = Nk+Np;
+    const int MaxIt= 10000;
+    int Nconv;
+    RealD resid = 1.0e-5;
+    if(0)
+    {
+      int order = 501;
+      RealD bound = 5.0e-4;
+      std::cout << GridLogMessage << " Lanczos for dirichlet bound " << bound<<" order "<< order<<std::endl;
+      Chebyshev<FermionField> Cheby(bound,90.,order);
+      FunctionHermOp<FermionField> OpCheby(Cheby,UddagUd);
+      PlainHermOp<FermionField> Op     (UddagUd);
+      ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby,Op,Nstop,Nk,Nm,resid,MaxIt);
+      std::vector<RealD>          eval(Nm);
+      std::vector<FermionField> evec(Nm,rbgrid);
+      FermionField    src(rbgrid);src = 1.0;
+      IRL.calc(eval,evec,src,Nconv);
+      
+      FermionField tmp(rbgrid);
+      FermionField ftmp(grid);
+      FermionField ftmp4(grid4);
+      for(int ev=0;ev<evec.size();ev++){
+	Gamma GT(Gamma::Algebra::GammaT);
+	std::cout << " evec " << ev << std::endl;
+	tmp = evec[ev] + GT*evec[ev];
+	DumpSliceNorm(" 1+gammaT ",tmp,Nd);
+	tmp = evec[ev] - GT*evec[ev];
+	DumpSliceNorm(" 1-gammaT ",tmp,Nd);
+      }
+      for(int e=0;e<10;e++){
+	std::cout << " Dirichlet evec "<<e<<std::endl;
+	tmp = evec[e];
+	for(int s=0;s<Ls;s++){
+	  ftmp=Zero();
+	  setCheckerboard(ftmp,tmp);
+	  ExtractSlice(ftmp4,ftmp,s,0);
+	  std::cout << "s-slice "<<s<< " evec[0] " << std::endl;
+	  DumpSliceNorm(" s-slice ",ftmp4,Nd-1);
+	}
+      }
+    }
+    if(1)
+    {
+      int order = 2001;
+      RealD bound = 6.0e-5;
+      std::cout << GridLogMessage << " Lanczos for full operator  bound " << bound<<" order "<< order<<std::endl;
+      Chebyshev<FermionField> Cheby(bound,90.,order);
+      FunctionHermOp<FermionField> OpCheby(Cheby,UdagU);
+      PlainHermOp<FermionField> Op     (UdagU);
+      ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby,Op,Nstop,Nk,Nm,resid,MaxIt);
+      std::vector<RealD>          eval(Nm);
+      std::vector<FermionField> evec(Nm,rbgrid);
+      FermionField    src(rbgrid); src = 1.0;
+      IRL.calc(eval,evec,src,Nconv);
+
+      FermionField tmp(rbgrid);
+      FermionField ftmp(grid);
+      FermionField ftmp4(grid4);
+      for(int e=0;e<evec.size();e++){
+	std::cout << " Full evec "<<e<<std::endl;
+	tmp = evec[e];
+	for(int s=0;s<Ls;s++){
+	  ftmp=Zero();
+	  setCheckerboard(ftmp,tmp);
+	  ExtractSlice(ftmp4,ftmp,s,0);
+	  std::cout << "s-slice "<<s<< " evec[0] " << std::endl;
+	  DumpSliceNorm(" s-slice ",ftmp4,Nd-1);
+	}
+      }
+
+    }
+    Grid_finalize();
+    std::cout << " All done "<<std::endl;
+    exit(EXIT_SUCCESS);
+    }
+  }
+
+  TheHMC.Run();  // no smearing
+
+  Grid_finalize();
+} // main
+
+
+

HMC/Mobius2p1f_DD_RHMC_96I_mixed.cc

@@ -0,0 +1,444 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_hmc_EODWFRatio.cc
+
+Copyright (C) 2015-2016
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Guido Cossu <guido.cossu@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 */
+#include <Grid/Grid.h>
+
+NAMESPACE_BEGIN(Grid);
+
+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;
+
+    using OperatorFunction<FieldD>::operator();
+
+    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);
+
+      ////////////////////////////////////////////////////////////////////////////////////
+      // 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);
+    }
+  };
+
+NAMESPACE_END(Grid);
+
+
+int main(int argc, char **argv) {
+  using namespace Grid;
+
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+
+   // Typedefs to simplify notation
+  typedef WilsonImplR FermionImplPolicy;
+  typedef WilsonImplF FermionImplPolicyF;
+  typedef MobiusFermionR FermionAction;
+  typedef MobiusFermionF FermionActionF;
+  typedef typename FermionAction::FermionField FermionField;
+  typedef typename FermionActionF::FermionField FermionFieldF;
+
+  typedef Grid::XmlReader       Serialiser;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  //  typedef GenericHMCRunner<LeapFrog> HMCWrapper;
+  //  MD.name    = std::string("Leap Frog");
+  //  typedef GenericHMCRunner<ForceGradient> HMCWrapper;
+  //  MD.name    = std::string("Force Gradient");
+  typedef GenericHMCRunner<MinimumNorm2> HMCWrapper;
+  MD.name    = std::string("MinimumNorm2");
+  MD.MDsteps =  4;
+  MD.trajL   = 1.0;
+
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = 1077;
+  HMCparams.Trajectories     = 1;
+  HMCparams.NoMetropolisUntil=  0;
+  // "[HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+  //  HMCparams.StartingType     =std::string("ColdStart");
+  HMCparams.StartingType     =std::string("CheckpointStart");
+  HMCparams.MD = MD;
+  HMCWrapper TheHMC(HMCparams);
+
+  // Grid from the command line arguments --grid and --mpi
+  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
+
+  CheckpointerParameters CPparams;
+  CPparams.config_prefix = "ckpoint_DDHMC_lat";
+  CPparams.rng_prefix    = "ckpoint_DDHMC_rng";
+  CPparams.saveInterval  = 1;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = "1 2 3 4 5";
+  RNGpar.parallel_seeds = "6 7 8 9 10";
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  // Construct observables
+  // here there is too much indirection
+  typedef PlaquetteMod<HMCWrapper::ImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+
+  const int Ls      = 12;
+  RealD M5  = 1.8;
+  RealD b   = 1.5;
+  RealD c   = 0.5;
+  Real beta         = 2.31;
+  //  Real light_mass   = 5.4e-4;
+  Real light_mass   = 7.8e-4;
+  Real strange_mass = 0.02132;
+  Real pv_mass      = 1.0;
+  std::vector<Real> hasenbusch({ light_mass, 3.8e-3, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass });
+
+  // FIXME:
+  // Same in MC and MD
+  // Need to mix precision too
+  OneFlavourRationalParams SFRp; // Strange
+  SFRp.lo       = 4.0e-3;
+  SFRp.hi       = 90.0;
+  SFRp.MaxIter  = 60000;
+  SFRp.tolerance= 1.0e-8;
+  SFRp.mdtolerance= 1.0e-6;
+  SFRp.degree   = 12;
+  SFRp.precision= 50;
+  SFRp.BoundsCheckFreq=0;
+
+  OneFlavourRationalParams OFRp; // Up/down
+  OFRp.lo       = 2.0e-5;
+  OFRp.hi       = 90.0;
+  OFRp.MaxIter  = 60000;
+  OFRp.tolerance= 1.0e-8;
+  OFRp.mdtolerance= 1.0e-6;
+  //  OFRp.degree   = 20; converges
+  //  OFRp.degree   = 16;
+  OFRp.degree   = 12;
+  OFRp.precision= 80;
+  OFRp.BoundsCheckFreq=0;
+
+  auto GridPtr   = TheHMC.Resources.GetCartesian();
+  auto GridRBPtr = TheHMC.Resources.GetRBCartesian();
+
+  typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
+  typedef SchurDiagMooeeOperator<FermionAction ,FermionField > LinearOperatorD;
+  typedef MixedPrecisionConjugateGradientOperatorFunction<MobiusFermionD,MobiusFermionF,LinearOperatorD,LinearOperatorF> MxPCG;
+
+  ////////////////////////////////////////////////////////////////
+  // Domain decomposed
+  ////////////////////////////////////////////////////////////////
+  Coordinate latt4  = GridPtr->GlobalDimensions();
+  Coordinate mpi    = GridPtr->ProcessorGrid();
+  Coordinate shm;
+
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  
+  Coordinate CommDim(Nd);
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+
+  Coordinate NonDirichlet(Nd+1,0);
+  Coordinate Dirichlet(Nd+1,0);
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Coordinate Block4(Nd);
+  Block4[0] = Dirichlet[1];
+  Block4[1] = Dirichlet[2];
+  Block4[2] = Dirichlet[3];
+  Block4[3] = Dirichlet[4];
+
+  int Width=3;
+  TheHMC.Resources.SetMomentumFilter(new DDHMCFilter<WilsonImplR::Field>(Block4,Width));
+
+  //////////////////////////
+  // Fermion Grids
+  //////////////////////////
+  auto FGrid     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr);
+  auto FrbGrid   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr);
+
+  Coordinate simdF = GridDefaultSimd(Nd,vComplexF::Nsimd());
+  auto GridPtrF   = SpaceTimeGrid::makeFourDimGrid(latt4,simdF,mpi);
+  auto GridRBPtrF = SpaceTimeGrid::makeFourDimRedBlackGrid(GridPtrF);
+  auto FGridF     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtrF);
+  auto FrbGridF   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtrF);
+
+  IwasakiGaugeActionR GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeField U(GridPtr);
+  LatticeGaugeFieldF UF(GridPtrF);
+
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.ReadCommandLine(argc,argv);  // params on CML or from param file
+  TheHMC.initializeGaugeFieldAndRNGs(U);
+
+
+  // These lines are unecessary if BC are all periodic
+  std::vector<Complex> boundary = {1,1,1,-1};
+  FermionAction::ImplParams Params(boundary);
+  Params.dirichlet=NonDirichlet;
+  FermionAction::ImplParams ParamsDir(boundary);
+  ParamsDir.dirichlet=Dirichlet;
+
+  //  double StoppingCondition = 1e-14;
+  //  double MDStoppingCondition = 1e-9;
+  double StoppingCondition = 1e-10;
+  double MDStoppingCondition = 1e-7;
+  double MDStoppingConditionLoose = 1e-6;
+  double MaxCGIterations = 300000;
+  ConjugateGradient<FermionField>  CG(StoppingCondition,MaxCGIterations);
+  ConjugateGradient<FermionField>  MDCG(MDStoppingCondition,MaxCGIterations);
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1);
+  ActionLevel<HMCWrapper::Field> Level2(4);
+  ActionLevel<HMCWrapper::Field> Level3(8);
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionAction StrangeOp (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params);
+  FermionAction StrangePauliVillarsOp(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, Params);
+
+  FermionAction StrangeOpDir (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, ParamsDir);
+  FermionAction StrangePauliVillarsOpDir(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, ParamsDir);
+  
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionBdy(StrangeOpDir,StrangeOp,SFRp);
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionLocal(StrangePauliVillarsOpDir,StrangeOpDir,SFRp);
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionPVBdy(StrangePauliVillarsOp,StrangePauliVillarsOpDir,SFRp);
+  Level1.push_back(&StrangePseudoFermionBdy);
+  Level2.push_back(&StrangePseudoFermionLocal);
+  Level1.push_back(&StrangePseudoFermionPVBdy);
+
+  ////////////////////////////////////
+  // up down action
+  ////////////////////////////////////
+  std::vector<Real> light_den;
+  std::vector<Real> light_num;
+  std::vector<int> dirichlet_den;
+  std::vector<int> dirichlet_num;
+
+  int n_hasenbusch = hasenbusch.size();
+  light_den.push_back(light_mass);  dirichlet_den.push_back(0);
+  for(int h=0;h<n_hasenbusch;h++){
+    light_den.push_back(hasenbusch[h]); dirichlet_den.push_back(1);
+  }
+
+  for(int h=0;h<n_hasenbusch;h++){
+    light_num.push_back(hasenbusch[h]); dirichlet_num.push_back(1);
+  }
+  light_num.push_back(pv_mass);  dirichlet_num.push_back(0);
+
+  std::vector<FermionAction *> Numerators;
+  std::vector<FermionAction *> Denominators;
+  std::vector<FermionActionF *> DenominatorsF;
+  std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
+  std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
+  std::vector<MxPCG *> ActionMPCG;
+  std::vector<MxPCG *> MPCG;
+
+  typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
+  typedef SchurDiagMooeeOperator<FermionAction ,FermionField > LinearOperatorD;
+  std::vector<LinearOperatorD *> LinOpD;
+  std::vector<LinearOperatorF *> LinOpF; 
+  
+  for(int h=0;h<n_hasenbusch+1;h++){
+    std::cout << GridLogMessage
+	      << " 2f quotient Action ";
+    std::cout << "det D("<<light_den[h]<<")";
+    if ( dirichlet_den[h] ) std::cout << "^dirichlet    ";
+    std::cout << "/ det D("<<light_num[h]<<")";
+    if ( dirichlet_num[h] ) std::cout << "^dirichlet    ";
+    std::cout << std::endl;
+
+    FermionAction::ImplParams ParamsNum(boundary);
+    FermionAction::ImplParams ParamsDen(boundary);
+    FermionActionF::ImplParams ParamsDenF(boundary);
+    
+    if ( dirichlet_num[h]==1) ParamsNum.dirichlet = Dirichlet;
+    else                      ParamsNum.dirichlet = NonDirichlet;
+    Numerators.push_back  (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, ParamsNum));
+
+    if ( dirichlet_den[h]==1) ParamsDen.dirichlet = Dirichlet;
+    else                      ParamsDen.dirichlet = NonDirichlet;
+
+    Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, ParamsDen));
+
+    ParamsDenF.dirichlet = ParamsDen.dirichlet;
+    DenominatorsF.push_back(new FermionActionF(UF,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_den[h],M5,b,c, ParamsDenF));
+
+    LinOpD.push_back(new LinearOperatorD(*Denominators[h]));
+    LinOpF.push_back(new LinearOperatorF(*DenominatorsF[h]));
+
+    double conv  = MDStoppingCondition;
+    if (h<3) conv= MDStoppingConditionLoose; // Relax on first two hasenbusch factors
+    const int MX_inner = 5000;
+    MPCG.push_back(new MxPCG(conv,
+			     MX_inner,
+			     MaxCGIterations,
+			     GridPtrF,
+			     FrbGridF,
+			     *DenominatorsF[h],*Denominators[h],
+			     *LinOpF[h], *LinOpD[h]) );
+
+    ActionMPCG.push_back(new MxPCG(StoppingCondition,
+				   MX_inner,
+				   MaxCGIterations,
+				   GridPtrF,
+				   FrbGridF,
+				   *DenominatorsF[h],*Denominators[h],
+				   *LinOpF[h], *LinOpD[h]) );
+
+    
+    if(h!=0) {
+      //      Quotients.push_back (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],MDCG,CG));
+      Quotients.push_back (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],*MPCG[h],*ActionMPCG[h],CG));
+    } else {
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+    }
+  }
+
+  int nquo=Quotients.size();
+  Level1.push_back(Bdys[0]);
+  Level1.push_back(Bdys[1]);
+  for(int h=0;h<nquo-1;h++){
+    Level2.push_back(Quotients[h]);
+  }
+  Level2.push_back(Quotients[nquo-1]);
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+  /////////////////////////////////////////////////////////////
+
+  TheHMC.Run();  // no smearing
+
+  Grid_finalize();
+} // main
+
+
+

HMC/RNGstate.cc

@@ -0,0 +1,53 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file:
+
+Copyright (C) 2015-2016
+
+Author: Peter Boyle <pabobyle@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 */
+#include <Grid/Grid.h>
+
+int main(int argc, char **argv)
+{
+  using namespace Grid;
+
+  Grid_init(&argc, &argv);
+
+  Coordinate latt4  = GridDefaultLatt();
+  Coordinate mpi    = GridDefaultMpi();
+  Coordinate simd   = GridDefaultSimd(Nd,vComplexD::Nsimd());
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,simd,mpi);
+
+  GridSerialRNG   sRNG;         sRNG.SeedUniqueString(std::string("The Serial RNG"));
+  GridParallelRNG pRNG(UGrid);  pRNG.SeedUniqueString(std::string("The 4D RNG"));
+
+  std::string rngfile("ckpoint_rng.0");
+  NerscIO::writeRNGState(sRNG, pRNG, rngfile);
+  
+  Grid_finalize();
+}
+
+
+

benchmarks/Benchmark_comms.cc

@@ -217,9 +217,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu);
 	
 	    comm_proc = mpi_layout[mu]-1;
@@ -228,9 +228,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu+4][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu+4);
 	  
 	  }
@@ -309,9 +309,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu);
 	    Grid.StencilSendToRecvFromComplete(requests,mu);
 	    requests.resize(0);
@@ -322,9 +322,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu+4][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu+4);
 	    Grid.StencilSendToRecvFromComplete(requests,mu+4);
 	    requests.resize(0);
@@ -411,8 +411,8 @@ int main (int argc, char ** argv)
 	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	    }
             int tid = omp_get_thread_num();
-	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
-					       (void *)&rbuf[dir][0], recv_from_rank, bytes,tid);
+	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,1,
+					       (void *)&rbuf[dir][0], recv_from_rank,1, bytes,tid);
 
 	    thread_critical { dbytes+=tbytes; }
 	  }

benchmarks/Benchmark_dwf.cc

@@ -191,9 +191,7 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
     for(int i=0;i<ncall;i++){
-      __SSC_START;
       Dw.Dhop(src,result,0);
-      __SSC_STOP;
     }
     double t1=usecond();
     FGrid->Barrier();

benchmarks/Benchmark_dwf_fp32.cc

@@ -32,18 +32,18 @@
 using namespace std;
 using namespace Grid;
 
-template<class d>
-struct scal {
-  d internal;
+////////////////////////
+/// Move to domains ////
+////////////////////////
+
+Gamma::Algebra Gmu [] = {
+			 Gamma::Algebra::GammaX,
+			 Gamma::Algebra::GammaY,
+			 Gamma::Algebra::GammaZ,
+			 Gamma::Algebra::GammaT
 };
 
-  Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT
-  };
-
+void Benchmark(int Ls, Coordinate Dirichlet);
 
 int main (int argc, char ** argv)
 {
@@ -52,24 +52,82 @@ int main (int argc, char ** argv)
 
   int threads = GridThread::GetThreads();
 
-  Coordinate latt4 = GridDefaultLatt();
   int Ls=16;
-  for(int i=0;i<argc;i++)
+  for(int i=0;i<argc;i++) {
     if(std::string(argv[i]) == "-Ls"){
       std::stringstream ss(argv[i+1]); ss >> Ls;
     }
+  }
 
+  //////////////////
+  // With comms
+  //////////////////
+  Coordinate Dirichlet(Nd+1,0);
+
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing with full communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  
+  Benchmark(Ls,Dirichlet);
+
+  //////////////////
+  // Domain decomposed
+  //////////////////
+  Coordinate latt4  = GridDefaultLatt();
+  Coordinate mpi    = GridDefaultMpi();
+  Coordinate CommDim(Nd);
+  Coordinate shm;
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+
+
+  //////////////////////
+  // Node level
+  //////////////////////
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing without internode communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+  Dirichlet[0] = 0;
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Benchmark(Ls,Dirichlet);
+
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing without intranode communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+
+  for(int d=0;d<Nd;d++) CommDim[d]= mpi[d]>1 ? 1 : 0;
+  Dirichlet[0] = 0;
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3];
+  
+  Benchmark(Ls,Dirichlet);
+
+  Grid_finalize();
+  exit(0);
+}
+void Benchmark(int Ls, Coordinate Dirichlet)
+{
+  Coordinate latt4 = GridDefaultLatt();
   GridLogLayout();
 
   long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
 
-
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
   GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
   GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
   GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
@@ -80,9 +138,9 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
   GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
+
   std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
   GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
-  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
 
   LatticeFermionF src   (FGrid); random(RNG5,src);
 #if 0
@@ -100,7 +158,6 @@ int main (int argc, char ** argv)
   src = src*N2;
 #endif
 
-
   LatticeFermionF result(FGrid); result=Zero();
   LatticeFermionF    ref(FGrid);    ref=Zero();
   LatticeFermionF    tmp(FGrid);
@@ -108,29 +165,31 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeFieldF Umu(UGrid);
+  LatticeGaugeFieldF UmuCopy(UGrid);
   SU<Nc>::HotConfiguration(RNG4,Umu);
+  UmuCopy=Umu;
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
-#if 0
-  Umu=1.0;
-  for(int mu=0;mu<Nd;mu++){
-    LatticeColourMatrixF ttmp(UGrid);
-    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
-    //    if (mu !=2 ) ttmp = 0;
-    //    ttmp = ttmp* pow(10.0,mu);
-    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
-  }
-  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
-#endif
 
+  ////////////////////////////////////
+  // Apply BCs
+  ////////////////////////////////////
+  Coordinate Block(4);
+  for(int d=0;d<4;d++)  Block[d]= Dirichlet[d+1];
+
+  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block5 " << Dirichlet << std::endl;
+  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block4 " << Block << std::endl;
+
+  DirichletFilter<LatticeGaugeFieldF> Filter(Block);
+  Filter.applyFilter(Umu);
+  
   ////////////////////////////////////
   // Naive wilson implementation
   ////////////////////////////////////
-  // replicate across fifth dimension
-  //  LatticeGaugeFieldF Umu5d(FGrid);
   std::vector<LatticeColourMatrixF> U(4,UGrid);
   for(int mu=0;mu<Nd;mu++){
     U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
   }
+
   std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
   if (1)
@@ -190,19 +249,20 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
-  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionF::ImplParams p;
+  p.dirichlet=Dirichlet;
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,p);
+  Dw.ImportGauge(Umu);
+  
   int ncall =300;
 
   if (1) {
     FGrid->Barrier();
-    Dw.ZeroCounters();
     Dw.Dhop(src,result,0);
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
     for(int i=0;i<ncall;i++){
-      __SSC_START;
       Dw.Dhop(src,result,0);
-      __SSC_STOP;
     }
     double t1=usecond();
     FGrid->Barrier();
@@ -220,29 +280,20 @@ int main (int argc, char ** argv)
     double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
 
     std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
-    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
-    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
+    //    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
+    //    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
     err = ref-result;
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-    //exit(0);
 
     if(( norm2(err)>1.0e-4) ) {
-      /*
-      std::cout << "RESULT\n " << result<<std::endl;
-      std::cout << "REF   \n " << ref   <<std::endl;
-      std::cout << "ERR   \n " << err   <<std::endl;
-      */
       std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
       FGrid->Barrier();
       exit(-1);
     }
     assert (norm2(err)< 1.0e-4 );
-    Dw.Report();
   }
 
   if (1)
@@ -286,21 +337,20 @@ int main (int argc, char ** argv)
     }
     ref = -0.5*ref;
   }
-  //  dump=1;
-  Dw.Dhop(src,result,1);
+
+  Dw.Dhop(src,result,DaggerYes);
+
+  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
   std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
+  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
+
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
   std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
   std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result;
   std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-/*
-	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
-	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
-	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
-*/
-  }
+  assert((norm2(err)<1.0e-4));
+  
   LatticeFermionF src_e (FrbGrid);
   LatticeFermionF src_o (FrbGrid);
   LatticeFermionF r_e   (FrbGrid);
@@ -330,7 +380,6 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
   {
-    Dw.ZeroCounters();
     FGrid->Barrier();
     Dw.DhopEO(src_o,r_e,DaggerNo);
     double t0=usecond();
@@ -352,7 +401,6 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
-    Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
   Dw.DhopOE(src_e,r_o,DaggerNo);
@@ -367,13 +415,7 @@ int main (int argc, char ** argv)
 
   err = r_eo-result;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-    /*
-	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
-	std::cout<< "Deo REF\n " <<result  << std::endl;
-	std::cout<< "Deo ERR   \n " << err <<std::endl;
-    */
-  }
+  assert(norm2(err)<1.0e-4);
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
@@ -382,6 +424,4 @@ int main (int argc, char ** argv)
 
   assert(norm2(src_e)<1.0e-4);
   assert(norm2(src_o)<1.0e-4);
-  Grid_finalize();
-  exit(0);
 }

configure.ac

@@ -394,11 +394,10 @@ case ${CXXTEST} in
     fi
     ;;
   hipcc)
-#    CXXFLAGS="$CXXFLAGS -Xcompiler -fno-strict-aliasing --expt-extended-lambda --expt-relaxed-constexpr"
     CXXFLAGS="$CXXFLAGS -fno-strict-aliasing"
     CXXLD=${CXX}
     if test $ac_openmp = yes; then
-       CXXFLAGS="$CXXFLAGS -Xcompiler -fopenmp"
+       CXXFLAGS="$CXXFLAGS -fopenmp"
     fi
     ;;
   dpcpp)
@@ -557,16 +556,19 @@ esac
 AC_ARG_ENABLE([setdevice],[AC_HELP_STRING([--enable-setdevice | --disable-setdevice],
               [Set GPU to rank in node with cudaSetDevice or similar])],[ac_SETDEVICE=${enable_SETDEVICE}],[ac_SETDEVICE=no])
 case ${ac_SETDEVICE} in
-    yes);;
-    no)
+    yes)
+	echo ENABLE SET DEVICE
+    ;;
+    *)
      AC_DEFINE([GRID_DEFAULT_GPU],[1],[GRID_DEFAULT_GPU] )
+   	echo DISABLE SET DEVICE
     ;;
 esac
 
 #########################################################
 ######################  Shared memory intranode #########
 #########################################################
-AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmopen|shmget|hugetlbfs|shmnone|nvlink|no],
+AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmopen|shmget|hugetlbfs|shmnone|nvlink|no|none],
               [Select SHM allocation technique])],[ac_SHM=${enable_shm}],[ac_SHM=no])
 
 case ${ac_SHM} in
@@ -586,7 +588,7 @@ case ${ac_SHM} in
      AC_DEFINE([GRID_MPI3_SHMGET],[1],[GRID_MPI3_SHMGET] )
      ;;
 
-     shmnone | no)
+     shmnone | no | none)
      AC_DEFINE([GRID_MPI3_SHM_NONE],[1],[GRID_MPI3_SHM_NONE] )
      ;;
 

examples/Example_Laplacian.cc

@@ -93,14 +93,14 @@ template<class Field> class FreeLaplacianStencil : public SparseMatrixBase<Field
 {
 public:
   typedef typename Field::vector_object siteObject;
-  typedef CartesianStencil<siteObject, siteObject, int> StencilImpl;
+  typedef CartesianStencil<siteObject, siteObject, SimpleStencilParams> StencilImpl;
 
   GridBase *grid;
   StencilImpl Stencil;
   SimpleCompressor<siteObject> Compressor;
   
   FreeLaplacianStencil(GridBase *_grid)
-    : Stencil    (_grid,6,Even,directions,displacements,0), grid(_grid)
+    : Stencil    (_grid,6,Even,directions,displacements,SimpleStencilParams()), grid(_grid)
   {  };
   
   virtual GridBase *Grid(void) { return grid; };
@@ -168,7 +168,8 @@ public:
   typedef iImplDoubledGaugeField<Simd> SiteDoubledGaugeField;
   typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
 
-  typedef CartesianStencil<siteObject, siteObject, int> StencilImpl;
+  typedef CartesianStencil<siteObject, siteObject,SimpleStencilParams> StencilImpl;
+  SimpleStencilParams p;
 
   GridBase *grid;
   StencilImpl Stencil;
@@ -177,7 +178,7 @@ public:
   CovariantLaplacianStencil(GaugeField &Umu)
     :
       grid(Umu.Grid()),
-      Stencil    (grid,6,Even,directions,displacements,0),
+      Stencil    (grid,6,Even,directions,displacements,p),
       Uds(grid)
   {
     for (int mu = 0; mu < Nd; mu++) {
@@ -324,7 +325,7 @@ int main(int argc, char ** argv)
 
   U_GT = U;
   // Make a random xform to teh gauge field
-  SU<Nc>::RandomGaugeTransform(RNG,U_GT,g); // Unit gauge
+  SU<Nc>::RandomGaugeTransform<PeriodicGimplR>(RNG,U_GT,g); // Unit gauge
 
   Field in_GT(&Grid); 
   Field out_GT(&Grid);

systems/Crusher/comms.slurm

@@ -0,0 +1,26 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -p ecp
+#SBATCH -J comms
+#SBATCH -o comms.%J
+#SBATCH -e comms.%J
+#SBATCH -N 1
+#SBATCH -n 2
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+export OMP_NUM_THREADS=8
+
+AT=8
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads ${AT} --grid 64.64.32.32 --mpi 2.1.1.1 "
+srun -n2 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_comms_host_device $PARAMS
+

systems/Crusher/config-command

@@ -5,8 +5,11 @@
 --enable-gen-simd-width=64 \
 --enable-simd=GPU \
 --disable-fermion-reps \
+--with-gmp=$OLCF_GMP_ROOT \
+--with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 --disable-gparity \
 CXX=hipcc MPICXX=mpicxx \
-CXXFLAGS="-fPIC -I/opt/rocm-4.5.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
- LDFLAGS=" -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa "
-HIPFLAGS = --amdgpu-target=gfx90a
+CXXFLAGS="-fPIC -I{$ROCM_PATH}/include/ -std=c++14 -I${MPICH_DIR}/include " \
+ LDFLAGS=" -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64 "
+
+

systems/Crusher/dwf.slurm

@@ -3,28 +3,30 @@
 #SBATCH -A LGT104
 #SBATCH -t 01:00:00
 ##SBATCH -U openmpThu
-##SBATCH -p ecp
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
 #SBATCH -N 1
-#SBATCH -n 1
-#SBATCH --exclusive  
+#SBATCH -n 8
+#SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
 
 DIR=.
-module list
-#export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+source sourceme.sh
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
 export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-#export MPICH_SMP_SINGLE_COPY_MODE=NONE
-#export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export OMP_NUM_THREADS=1
 
-AT=8
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
 
-PARAMS=" --accelerator-threads ${AT} --grid 24.24.24.24 --shm-mpi 0 --mpi 1.1.1.1"
+echo working directory
+pwd
 
-srun --gpus-per-task 1 -n1 ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+PARAMS=" --accelerator-threads 8 --grid 32.32.32.32 --mpi 1.1.1.1 --comms-sequential --shm 2048 --shm-mpi 0"
+srun --gpus-per-task 1 -n1  ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+
+PARAMS=" --accelerator-threads 8 --grid 64.64.64.32 --mpi 2.2.2.1 --comms-sequential --shm 2048 --shm-mpi 0"
+srun --gpus-per-task 1 -n8  ./benchmarks/Benchmark_dwf_fp32 $PARAMS
 
 

systems/Crusher/dwf4.slurm

@@ -7,21 +7,19 @@
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
 #SBATCH -N 1
-#SBATCH -n 4
-#SBATCH --exclusive
+#SBATCH -n 2
+#SBATCH --gpu-bind=map_gpu:0,1
 
 DIR=.
-module list
+source setup.sh
+
+export MPICH_OFI_NIC_POLICY=GPU
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-export MPICH_SMP_SINGLE_COPY_MODE=NONE
-#export MPICH_SMP_SINGLE_COPY_MODE=CMA
-export OMP_NUM_THREADS=4
+export OMP_NUM_THREADS=16
 
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads 8 --grid 32.32.64.64 --mpi 1.1.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
 
-srun --gpus-per-task 1 -n4 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+srun --gpus-per-task 1 -N1 -n2  ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 16.16.32.64 --shm-mpi 1 --shm 2048 --comms-sequential --accelerator-threads 8
 
 

systems/Crusher/dwf8.slurm

@@ -8,20 +8,21 @@
 #SBATCH -e DWF.%J
 #SBATCH -N 1
 #SBATCH -n 8
-#SBATCH --exclusive
+##SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,6,7,4,5
 
 DIR=.
-module list
+source setup.sh
+
+export MPICH_OFI_NIC_POLICY=GPU
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-#export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
-export OMP_NUM_THREADS=1
+#export MPICH_SMP_SINGLE_COPY_MODE=NONE
+export OMP_NUM_THREADS=16
 
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads 8 --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
-
-srun --gpus-per-task 1 -n8 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
 
+srun --gpus-per-task 1 -N1 -n8  ./benchmarks/Benchmark_comms_host_device --mpi 2.2.2.1 --shm-mpi 1 --shm 2048 --comms-sequential --accelerator-threads 8
 

systems/Crusher/mpiwrapper.sh

@@ -1,10 +1,11 @@
 #!/bin/bash
 
 lrank=$SLURM_LOCALID
+lgpu=(0 1 2 3 7 6 5 4)
 
-export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
+export ROCR_VISIBLE_DEVICES=${lgpu[$lrank]}
 
-echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES "
 
 $*
 

systems/Crusher/sourceme.sh

@@ -1,5 +1,9 @@
 module load PrgEnv-gnu
-module load rocm/4.5.0
+module load rocm/5.1.0
+module load cray-mpich/8.1.16
 module load gmp
-module load cray-fftw
+#module load cray-fftw
 module load craype-accel-amd-gfx90a
+export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
+#Hack for lib
+export LD_LIBRARY_PATH=`pwd`:$LD_LIBRARY_PATH

systems/Perlmutter/config-command

@@ -1,9 +1,14 @@
+DIR=`pwd`
+PREFIX=$DIR/../Prequisites/install/
 ../../configure \
     --enable-comms=mpi \
     --enable-simd=GPU \
     --enable-shm=nvlink \
     --enable-gen-simd-width=64 \
     --enable-accelerator=cuda \
+    --enable-setdevice \
+    --disable-accelerator-cshift \
+    --with-gmp=$PREFIX \
     --disable-fermion-reps \
     --disable-unified \
     --disable-gparity \

systems/Perlmutter/dwf4.slurm

@@ -1,24 +1,27 @@
 #!/bin/bash
-#SBATCH -A mp13
+#SBATCH -A m3886_g
 #SBATCH -C gpu
-#SBATCH -q regular
+#SBATCH -q debug
 #SBATCH -t 0:20:00
-#SBATCH -n 16
-#SBATCH --ntasks-per-node=4
 #SBATCH -c 32
-#SBATCH --exclusive
+#SBATCH -N 1
+#SBATCH -n 4
+#SBATCH --ntasks-per-node=4
 #SBATCH --gpus-per-task=1
-#SBATCH --gpu-bind=map_gpu:0,1,2,3
+#SBATCH --exclusive
+#SBATCH --gpu-bind=none
 
 export SLURM_CPU_BIND="cores"
-export MPICH_RDMA_ENABLED_CUDA=1
 export MPICH_GPU_SUPPORT_ENABLED=1
-srun ./benchmarks/Benchmark_comms_host_device --mpi 2.2.2.2  --accelerator-threads 8 > comms.4node
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
 
-OPT="--comms-overlap --comms-concurrent --shm-mpi 0"
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 64.64.64.64 --accelerator-threads 8 --shm 2048 $OPT > dwf.64.64.64.64.4node.opt0
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 48.48.48.48 --accelerator-threads 8 --shm 2048 $OPT > dwf.48.48.48.48.4node.opt0
+OPT="--comms-sequential --shm-mpi 1"
+VOL=64.64.64.64
+srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
+#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.1.1.4 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
+#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.8 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
 
-OPT="--comms-overlap --comms-concurrent --shm-mpi 1"
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 64.64.64.64 --accelerator-threads 8 --shm 2048 $OPT > dwf.64.64.64.64.4node.opt1
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 48.48.48.48 --accelerator-threads 8 --shm 2048 $OPT > dwf.48.48.48.48.4node.opt1

systems/Perlmutter/sourceme.sh

@@ -1,4 +1,4 @@
 
 export CRAY_ACCEL_TARGET=nvidia80
 
-module load PrgEnv-gnu cpe-cuda cuda
+module load PrgEnv-gnu cpe-cuda cudatoolkit/11.4

systems/Spock/config-command

@@ -6,6 +6,8 @@
 --enable-simd=GPU \
 --disable-fermion-reps \
 --disable-gparity \
+--with-gmp=$OLCF_GMP_ROOT \
+--with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 CXX=hipcc MPICXX=mpicxx \
 CXXFLAGS="-fPIC -I/opt/rocm-4.3.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
 --prefix=/ccs/home/chulwoo/Grid \

systems/Spock/dwf8.slurm

@@ -1,8 +1,7 @@
 #!/bin/bash
 # Begin LSF Directives
 #SBATCH -A LGT104
-#SBATCH -t 01:00:00
-##SBATCH -U openmpThu
+#SBATCH -t 3:00:00
 #SBATCH -p ecp
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
@@ -14,13 +13,12 @@ DIR=.
 module list
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-export MPICH_SMP_SINGLE_COPY_MODE=NONE
-#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_SMP_SINGLE_COPY_MODE=CMA
+
 export OMP_NUM_THREADS=8
 
 AT=8
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads ${AT} --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
-srun -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+PARAMS=" --accelerator-threads ${AT} --grid 16.16.16.48 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+srun -N2 -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./HMC/Mobius2p1f_DD_RHMC $PARAMS
 

systems/Spock/sourceme.sh

@@ -1,5 +1,9 @@
+module load emacs
 module load PrgEnv-gnu
-module load rocm/4.3.0
+module load rocm/4.5.0
 module load gmp
 module load cray-fftw
 module load craype-accel-amd-gfx908
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+export LD_LIBRARY_PATH=/opt/cray/pe/gcc/mpfr/3.1.4/lib/:$LD_LIBRARY_PATH

systems/Summit/config-command

@@ -2,11 +2,12 @@
 	      --enable-simd=GPU \
 	      --enable-gen-simd-width=32 \
 	      --enable-unified=no \
-	       --enable-shm=nvlink \
+	       --enable-shm=no \
 	       --disable-gparity \
-	       --enable-setdevice \
+	       --disable-setdevice \
 	       --disable-fermion-reps \
 	       --enable-accelerator=cuda \
+	       --enable-accelerator-cshift \
 	       --prefix /ccs/home/paboyle/prefix \
 	       CXX=nvcc \
 	       LDFLAGS=-L/ccs/home/paboyle/prefix/lib/ \

systems/Summit/dwf16.lsf

@@ -1,25 +1,39 @@
 #!/bin/bash
 #BSUB -P LGT104
-#BSUB -W 2:00
+#BSUB -W 0:20
 #BSUB -nnodes 16
 #BSUB -J DWF
 
+
 export OMP_NUM_THREADS=6
 export PAMI_IBV_ADAPTER_AFFINITY=1
 export PAMI_ENABLE_STRIPING=1
-export OPT="--comms-concurrent --comms-overlap "
 
-APP="./benchmarks/Benchmark_comms_host_device  --mpi 4.4.4.3 "
-jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > comms.16node.log
+DIR=.
+source sourceme.sh
 
-APP="./benchmarks/Benchmark_dwf_fp32 --grid 96.96.96.72 --mpi 4.4.4.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
-jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.16node.24.log
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
 
-APP="./benchmarks/Benchmark_dwf_fp32 --grid 128.128.128.96 --mpi 4.4.4.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
-jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.16node.32.log
+VOLS=( 32.32.32.16 32.32.32.64 64.32.32.64 64.32.64.64 64.64.64.64 64.64.64.128  64.64.64.256  64.64.64.512 128.64.64.64.512)
+MPI=( 1.1.1.1      1.1.1.4     2.1.1.4         2.1.2.4     2.2.2.4      2.2.2.8      2.2.2.16      2.2.2.32 4.4.2.32 )
+RANKS=(     1            4           8              16          32          64            128           256 1024)
+NODES=(     1            1           2               4           8           16            32            64  128)
+INTS=(      0            1           2               3           4            5             6             7    8)
 
+for i in 5
+do
+    vol=${VOLS[$i]} 
+    nodes=${NODES[$i]} 
+    mpi=${MPI[$i]} 
+    ranks=${RANKS[$i]} 
 
+    JSRUN="jsrun --nrs $nodes -a4 -g4 -c42 -dpacked -b packed:10 --latency_priority gpu-cpu --smpiargs=-gpu"
 
+    PARAMS=" --accelerator-threads 8 --grid $vol --mpi $mpi --comms-sequential --shm 2048 --shm-mpi 0"
+    $JSRUN ./benchmarks/Benchmark_dwf_fp32 $PARAMS > run.v${vol}.n${nodes}.m${mpi}.seq.ker
 
+    PARAMS=" --accelerator-threads 8 --grid $vol --mpi $mpi --comms-overlap --shm 2048 --shm-mpi 0"
+    $JSRUN ./benchmarks/Benchmark_dwf_fp32 $PARAMS > run.v${vol}.n${nodes}.m${mpi}.over.ker
 
+done