Test_evec_compression changes:

Added ability to choose one of a variety of preselected basis sizes from the command line
	Fine lanczos now checks enough evecs are generated and resizes the output to Nstop and not the actual amount that converged (which can be larger)

Grid/DisableWarnings.h

@@ -34,6 +34,9 @@ directory
 
 #if defined __GNUC__ && __GNUC__>=6
 #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+#if defined __GNUC__ 
+#pragma GCC diagnostic ignored "-Wpsabi"
 #endif
 
  //disables and intel compiler specific warning (in json.hpp)
@@ -44,22 +47,14 @@ directory
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
-
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
- //disables nvcc specific warning in json.hpp
-#pragma nv_diag_suppress unsigned_compare_with_zero
-#pragma nv_diag_suppress cast_to_qualified_type
- //disables nvcc specific warning in many files
-#pragma nv_diag_suppress esa_on_defaulted_function_ignored
-#pragma nv_diag_suppress extra_semicolon
-#else
- //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
+
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon
-#endif
+
+//Eigen only
 #endif
 
 // Disable vectorisation in Eigen on the Power8/9 and PowerPC

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

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

Grid/Grid_Eigen_Dense.h

@@ -14,11 +14,7 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
-#pragma nv_diag_suppress code_is_unreachable
-#else
 #pragma diag_suppress code_is_unreachable
-#endif
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")

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/approx/Chebyshev.h

@@ -292,6 +292,7 @@ public:
 template<class Field>
 class ChebyshevLanczos : public Chebyshev<Field> {
 private:
+
   std::vector<RealD> Coeffs;
   int order;
   RealD alpha;

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;
@@ -80,6 +82,11 @@ NAMESPACE_BEGIN(Grid);
     RealD stop = src_norm * Tolerance*Tolerance;
 
     GridBase* DoublePrecGrid = src_d_in.Grid();
+
+    //Generate precision change workspaces
+    precisionChangeWorkspace wk_dp_from_sp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace wk_sp_from_dp(SinglePrecGrid, DoublePrecGrid);
+
     FieldD tmp_d(DoublePrecGrid);
     tmp_d.Checkerboard() = cb;
     
@@ -97,6 +104,7 @@ 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;
 
@@ -120,7 +128,7 @@ NAMESPACE_BEGIN(Grid);
       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d);
+      precisionChange(src_f, src_d, wk_sp_from_dp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -130,6 +138,7 @@ 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);
@@ -138,7 +147,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f);
+      precisionChange(tmp_d, sol_f, wk_dp_from_sp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);
@@ -150,6 +159,7 @@ 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

@@ -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

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -0,0 +1,411 @@
+/*************************************************************************************
+
+    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();
+    precisionChangeWorkspace wk_f_from_d(SinglePrecGrid, DoublePrecGrid);
+    precisionChangeWorkspace wk_d_from_f(DoublePrecGrid, SinglePrecGrid);
+
+    ////////////////////////////////////////////////////////////////////////
+    // 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, wk_f_from_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, wk_d_from_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, wk_f_from_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, wk_f_from_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/Deflation.h

@@ -113,43 +113,7 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
-
-  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
-    int Nevec = (int)evec_coarse.size();
-    int Nsrc = (int)src.size();
-    // make temp variables
-    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
-    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
-    //Preporcessing
-    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    guess_coarse[j] = Zero();
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockProject(src_coarse[j],src[j],subspace);
-    }
-    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
-    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
-    for (int i=0;i<Nevec;i++)
-    {
-      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
-      const CoarseField & tmp = evec_coarse[i];
-      for (int j=0;j<Nsrc;j++)
-      {
-        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
-      }
-    }
-    //postprocessing
-    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockPromote(guess_coarse[j],guess[j],subspace);
-    guess[j].Checkerboard() = src[j].Checkerboard();
-    }
-  };
-
-  };
+};
 
 
 

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,
@@ -155,6 +156,7 @@ public:
       _coarse_relax_tol(coarse_relax_tol)  
   {    };
 
+  //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);
@@ -181,8 +183,16 @@ public:
     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 +211,13 @@ 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 +295,10 @@ 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 +342,8 @@ 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 +392,31 @@ 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); 
 
     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 +427,14 @@ 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

@@ -53,11 +53,10 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  int              _processor;       // linear processor rank
-  unsigned long    _ndimension;
-  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
   Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  int              _processor;       // linear processor rank
   Coordinate _processor_coor;  // linear processor coordinate
+  unsigned long    _ndimension;
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -98,9 +97,8 @@ 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
@@ -144,16 +142,16 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,int do_xmit,
+			       int xmit_to_rank,
 			       void *recv,
-			       int recv_from_rank,int do_recv,
+			       int recv_from_rank,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,int do_xmit,
+				    int xmit_to_rank,
 				    void *recv,
-				    int recv_from_rank,int do_recv,
+				    int recv_from_rank,
 				    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,_shm_processors);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,13 +124,12 @@ 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];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -155,7 +154,6 @@ 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
@@ -337,22 +335,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest, int dox,
+						     int dest,
 						     void *recv,
-						     int from, int dor,
+						     int from,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,int dox,
+							 int dest,
 							 void *recv,
-							 int from,int dor,
+							 int from,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -372,33 +370,30 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( dox ) {
-    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 ( (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 ( (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 ( (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 ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);
-    list.resize(0);
   }
 
   return off_node_bytes;

Grid/communicator/Communicator_none.cc

@@ -45,14 +45,12 @@ 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);
@@ -113,18 +111,18 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,int dox,
+						     int xmit_to_rank,
 						     void *recv,
-						     int recv_from_rank,int dor,
+						     int recv_from_rank,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,int dox,
+							 int xmit_to_rank,
 							 void *recv,
-							 int recv_from_rank,int dor,
+							 int recv_from_rank,
 							 int bytes, int dir)
 {
   return 2.0*bytes;

Grid/communicator/SharedMemory.h

@@ -93,10 +93,9 @@ 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
-  // 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 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
   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,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   //////////////////////////////////////////////////////////////////////////////
   // 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,SHM);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
 }
 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,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -294,8 +294,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-  SHM = ShmDims;
-  
+
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -342,7 +341,7 @@ 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,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -354,8 +353,6 @@ 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

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

Grid/lattice/Lattice_crc.h

@@ -29,19 +29,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 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);

Grid/lattice/Lattice_reduction.h

@@ -142,15 +142,6 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
   return sumD_cpu(arg,osites);
 #endif  
 }
-template<class vobj>
-inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  return sumD_gpu_large(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
@@ -168,22 +159,6 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
   return ssum;
 }
 
-template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu_large(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
-#endif
-  arg.Grid()->GlobalSum(ssum);
-  return ssum;
-}
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Deterministic Reduction operations
 ////////////////////////////////////////////////////////////////////////////////////////////////////

Grid/lattice/Lattice_reduction_gpu.h

@@ -23,7 +23,7 @@ unsigned int nextPow2(Iterator x) {
 }
 
 template <class Iterator>
-int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
   
   int device;
 #ifdef GRID_CUDA
@@ -37,13 +37,13 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  /*  
+  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  */  
+  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -53,12 +53,12 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   threads = warpSize;
   if ( threads*sizeofsobj > sharedMemPerBlock ) {
     std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
-    return 0;
+    exit(EXIT_FAILURE);
   }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  return 1;
+  
 }
 
 template <class sobj, class Iterator>
@@ -198,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -207,9 +207,7 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   Integer size = osites*nsimd;
 
   Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
-
+  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
   Integer smemSize = numThreads * sizeof(sobj);
 
   Vector<sobj> buffer(numBlocks);
@@ -220,54 +218,6 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   auto result = buffer_v[0];
   return result;
 }
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<vector> buffer(osites);
-  vector *dat = (vector *)lat;
-  vector *buf = &buffer[0];
-  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,1,{
-	buf[ss] = dat[ss*words+w];
-      });
-      
-    ret_p[w] = sumD_gpu_small(tbuf,osites);
-  }
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  
-  Integer nsimd= vobj::Nsimd();
-  Integer size = osites*nsimd;
-  Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  
-  if ( ok ) {
-    ret = sumD_gpu_small(lat,osites);
-  } else {
-    ret = sumD_gpu_large(lat,osites);
-  }
-  return ret;
-}
-
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Return as same precision as input performing reduction in double precision though
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -280,13 +230,6 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
+
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_rng.h

@@ -32,8 +32,9 @@
 #include <random>
 
 #ifdef RNG_SITMO
-#include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
+#include <Grid/random/sitmo_prng_engine.hpp>
 #endif 
+#include <Grid/random/gaussian.h>
 
 #if defined(RNG_SITMO)
 #define RNG_FAST_DISCARD
@@ -142,8 +143,8 @@ public:
 
   std::vector<RngEngine>                             _generators;
   std::vector<std::uniform_real_distribution<RealD> > _uniform;
-  std::vector<std::normal_distribution<RealD> >       _gaussian;
-  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
+  std::vector<Grid::gaussian_distribution<RealD> >       _gaussian;
+  //  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
   std::vector<std::uniform_int_distribution<uint32_t> > _uid;
 
   ///////////////////////
@@ -243,8 +244,8 @@ public:
   GridSerialRNG() : GridRNGbase() {
     _generators.resize(1);
     _uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
-    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
+    _gaussian.resize(1,gaussian_distribution<RealD>(0.0,1.0) );
+    //    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(1,std::uniform_int_distribution<uint32_t>() );
   }
 
@@ -357,8 +358,8 @@ public:
 
     _generators.resize(_vol);
     _uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
-    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
+    _gaussian.resize(_vol,gaussian_distribution<RealD>(0.0,1.0) );
+    //    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
   }
 
@@ -515,11 +516,11 @@ public:
 
 template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l)   { rng.fill(l,rng._uniform);  }
 template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
-template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
+//template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
 
 template <class sobj> inline void random(GridSerialRNG &rng,sobj &l)   { rng.fill(l,rng._uniform  ); }
 template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
-template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
+//template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
 
 NAMESPACE_END(Grid);
 #endif

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;
 
@@ -1080,54 +1080,96 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//Convert a Lattice from one precision to another
-template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
-{
-  assert(out.Grid()->Nd() == in.Grid()->Nd());
-  for(int d=0;d<out.Grid()->Nd();d++){
-    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
-  }
-  out.Checkerboard() = in.Checkerboard();
-  GridBase *in_grid=in.Grid();
-  GridBase *out_grid = out.Grid();
-
-  typedef typename VobjOut::scalar_object SobjOut;
-  typedef typename VobjIn::scalar_object SobjIn;
-
-  int ndim = out.Grid()->Nd();
-  int out_nsimd = out_grid->Nsimd();
-    
-  std::vector<Coordinate > out_icoor(out_nsimd);
-      
-  for(int lane=0; lane < out_nsimd; lane++){
-    out_icoor[lane].resize(ndim);
-    out_grid->iCoorFromIindex(out_icoor[lane], lane);
-  }
-        
-  std::vector<SobjOut> in_slex_conv(in_grid->lSites());
-  unvectorizeToLexOrdArray(in_slex_conv, in);
-    
-  autoView( out_v , out, CpuWrite);
-  thread_for(out_oidx,out_grid->oSites(),{
-    Coordinate out_ocoor(ndim);
-    out_grid->oCoorFromOindex(out_ocoor, out_oidx);
-
-    ExtractPointerArray<SobjOut> ptrs(out_nsimd);      
-
-    Coordinate lcoor(out_grid->Nd());
-      
-    for(int lane=0; lane < out_nsimd; lane++){
-      for(int mu=0;mu<ndim;mu++)
-	lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
-	
-      int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
-      ptrs[lane] = &in_slex_conv[llex];
+//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
+class precisionChangeWorkspace{
+  std::pair<Integer,Integer>* fmap_device; //device pointer
+public:
+  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid){
+    //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
+    assert(out_grid->Nd() == in_grid->Nd());
+    for(int d=0;d<out_grid->Nd();d++){
+      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
     }
-    merge(out_v[out_oidx], ptrs, 0);
-  });
+    int Nsimd_out = out_grid->Nsimd();
+
+    std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
+    for(int lane=0; lane < out_grid->Nsimd(); lane++)
+      out_grid->iCoorFromIindex(out_icorrs[lane], lane);
+  
+    std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
+    thread_for(out_oidx,out_grid->oSites(),{
+	Coordinate out_ocorr; 
+	out_grid->oCoorFromOindex(out_ocorr, out_oidx);
+      
+	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
+	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
+	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
+	
+	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
+	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
+	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
+	  int in_oidx = 0, in_lane = 0;
+	  for(int d=0;d<in_grid->_ndimension;d++){
+	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
+	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
+	  }
+	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
+	}
+      });
+
+    //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
+    size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
+    fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
+    acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes); 
+  }
+
+  //Prevent moving or copying
+  precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
+  precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
+  
+  std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
+
+  ~precisionChangeWorkspace(){
+    acceleratorFreeDevice(fmap_device);
+  }
+};
+
+
+//Convert a lattice of one precision to another. The input workspace contains the mapping data.
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
+  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
+
+  out.Checkerboard() = in.Checkerboard();
+  constexpr int Nsimd_out = VobjOut::Nsimd();
+
+  std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
+
+  //Do the copy/precision change
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( in_v , in, AcceleratorRead);
+
+  accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
+      std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
+      for(int out_lane=0; out_lane < Nsimd_out; out_lane++){      
+	int in_oidx = fmap_osite[out_lane].first;
+	int in_lane = fmap_osite[out_lane].second;
+	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
+      }
+    });
 }
 
+//Convert a Lattice from one precision to another
+//Generate the workspace in place; if multiple calls with the same mapping are performed, consider pregenerating the workspace and reusing
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
+  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
+  precisionChange(out, in, workspace);
+}
+
+
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////

Grid/parallelIO/IldgIO.h

@@ -31,7 +31,6 @@ directory
 #include <fstream>
 #include <iomanip>
 #include <iostream>
-#include <string>
 #include <map>
 
 #include <pwd.h>
@@ -655,8 +654,7 @@ class IldgWriter : public ScidacWriter {
     // Fill ILDG header data struct
     //////////////////////////////////////////////////////
     ildgFormat ildgfmt ;
-    const std::string stNC = std::to_string( Nc ) ;
-    ildgfmt.field          = std::string("su"+stNC+"gauge");
+    ildgfmt.field     = std::string("su3gauge");
 
     if ( format == std::string("IEEE32BIG") ) { 
       ildgfmt.precision = 32;
@@ -873,8 +871,7 @@ class IldgReader : public GridLimeReader {
     } else { 
 
       assert(found_ildgFormat);
-      const std::string stNC = std::to_string( Nc ) ;
-      assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
+      assert ( ildgFormat_.field == std::string("su3gauge") );
 
       ///////////////////////////////////////////////////////////////////////////////////////
       // Populate our Grid metadata as best we can
@@ -882,7 +879,7 @@ class IldgReader : public GridLimeReader {
 
       std::ostringstream vers; vers << ildgFormat_.version;
       FieldMetaData_.hdr_version = vers.str();
-      FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
+      FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
 
       FieldMetaData_.nd=4;
       FieldMetaData_.dimension.resize(4);

Grid/parallelIO/MetaData.h

@@ -6,8 +6,8 @@
 
     Copyright (C) 2015
 
+
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -182,8 +182,8 @@ class GaugeStatistics
 public:
   void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
   {
-    header.link_trace = WilsonLoops<Impl>::linkTrace(data);
-    header.plaquette  = WilsonLoops<Impl>::avgPlaquette(data);
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
   }
 };
 typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
@@ -203,24 +203,20 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
 //////////////////////////////////////////////////////////////////////
 inline void reconstruct3(LorentzColourMatrix & cm)
 {
-  assert( Nc < 4 && Nc > 1 ) ;
+  const int x=0;
+  const int y=1;
+  const int z=2;
   for(int mu=0;mu<Nd;mu++){
-    #if Nc == 2
-      cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ;
-      cm(mu)()(1,1) =  adj(cm(mu)()(0,x)) ;
-    #else
-      const int x=0 , y=1 , z=2 ; // a little disinenuous labelling
-      cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
-      cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
-      cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
-    #endif
+    cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
+    cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
+    cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
   }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Some data types for intermediate storage
 ////////////////////////////////////////////////////////////////////////////////
-template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, Nc-1>, Nd >;
+template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
 
 typedef iLorentzColour2x3<Complex>  LorentzColour2x3;
 typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
@@ -282,6 +278,7 @@ struct GaugeSimpleMunger{
 
 template <class fobj, class sobj>
 struct GaugeSimpleUnmunger {
+
   void operator()(sobj &in, fobj &out) {
     for (int mu = 0; mu < Nd; mu++) {
       for (int i = 0; i < Nc; i++) {
@@ -320,8 +317,8 @@ template<class fobj,class sobj>
 struct Gauge3x2munger{
   void operator() (fobj &in,sobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<Nc-1;i++){
-	for(int j=0;j<Nc;j++){
+      for(int i=0;i<2;i++){
+	for(int j=0;j<3;j++){
 	  out(mu)()(i,j) = in(mu)(i)(j);
 	}}
     }
@@ -333,8 +330,8 @@ template<class fobj,class sobj>
 struct Gauge3x2unmunger{
   void operator() (sobj &in,fobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<Nc-1;i++){
-	for(int j=0;j<Nc;j++){
+      for(int i=0;i<2;i++){
+	for(int j=0;j<3;j++){
 	  out(mu)(i)(j) = in(mu)()(i,j);
 	}}
     }

Grid/parallelIO/NerscIO.h

@@ -9,7 +9,6 @@
     Author: Matt Spraggs <matthew.spraggs@gmail.com>
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -31,8 +30,6 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
-#include <string>
-
 NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
@@ -42,9 +39,11 @@ 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);
   }
@@ -148,17 +147,15 @@ public:
 
     std::string format(header.floating_point);
 
-    const int ieee32big = (format == std::string("IEEE32BIG"));
-    const int ieee32    = (format == std::string("IEEE32"));
-    const int ieee64big = (format == std::string("IEEE64BIG"));
-    const int ieee64    = (format == std::string("IEEE64") || \
-			   format == std::string("IEEE64LITTLE"));
+    int ieee32big = (format == std::string("IEEE32BIG"));
+    int ieee32    = (format == std::string("IEEE32"));
+    int ieee64big = (format == std::string("IEEE64BIG"));
+    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     // depending on datatype, set up munger;
     // munger is a function of <floating point, Real, data_type>
-    const std::string stNC = std::to_string( Nc ) ;
-    if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
+    if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@@ -169,7 +166,7 @@ public:
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
-    } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
+    } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@@ -203,7 +200,7 @@ 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 );
       
@@ -214,29 +211,27 @@ public:
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
-					std::string ens_label = std::string("DWF"),
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
+					std::string ens_label = std::string("DWF"))
   {
-    writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
+    writeConfiguration(Umu,file,0,1,ens_label);
   }
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32,
-					std::string ens_label = std::string("DWF"),
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
+					std::string ens_label = std::string("DWF"))
   {
     typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
-    header.sequence_number = sequence_number;
-    header.ensemble_id     = ens_id;
+    ///////////////////////////////////////////
+    // Following should become arguments
+    ///////////////////////////////////////////
+    header.sequence_number = 1;
+    header.ensemble_id     = std::string("UKQCD");
     header.ensemble_label  = ens_label;
-    header.hdr_version     = "1.0" ;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -250,14 +245,10 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write NcxNc double
-    header.floating_point  = std::string("IEEE64BIG");
-    const std::string stNC = std::to_string( Nc ) ;
-    if( two_row ) {
-      header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
-    } else {
-      header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
-    }
+    // Sod it -- always write 3x3 double
+    header.floating_point = std::string("IEEE64BIG");
+    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
+    GaugeSimpleUnmunger<fobj3D,sobj> munge;
     if ( grid->IsBoss() ) { 
       truncate(file);
       offset = writeHeader(header,file);
@@ -265,15 +256,8 @@ public:
     grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    if( two_row ) {
-      Gauge3x2unmunger<fobj2D,sobj> munge;
-      BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
-						nersc_csum,scidac_csuma,scidac_csumb);
-    } else {
-      GaugeSimpleUnmunger<fobj3D,sobj> munge;
-      BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
-						nersc_csum,scidac_csuma,scidac_csumb);
-    }
+    BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
+					      nersc_csum,scidac_csuma,scidac_csumb);
     header.checksum = nersc_csum;
     if ( grid->IsBoss() ) { 
       writeHeader(header,file);
@@ -305,7 +289,8 @@ public:
     header.plaquette=0.0;
     MachineCharacteristics(header);
 
-    uint64_t offset;
+	uint64_t offset;
+  
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");
@@ -345,7 +330,7 @@ public:
 
     GridBase *grid = parallel.Grid();
 
-    uint64_t offset = readHeader(file,grid,header);
+	uint64_t offset = readHeader(file,grid,header);
 
     FieldMetaData clone(header);
 

Grid/pugixml/pugixml.cc

@@ -16,12 +16,8 @@
 
 #ifdef __NVCC__
 #pragma push
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
-#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
-#else
 #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
 #endif
-#endif
 
 #include "pugixml.h"
 

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

@@ -40,29 +40,6 @@ 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

@@ -37,10 +37,6 @@ 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

@@ -36,33 +36,26 @@ NAMESPACE_BEGIN(Grid);
 
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
-  Coordinate twists;
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  GparityWilsonImplParams() : twists(Nd, 0), dirichlet(Nd, 0) {};
+  Coordinate twists; //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
+  GparityWilsonImplParams() : twists(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 {
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  StaggeredImplParams()
-  {
-    dirichlet.resize(Nd,0);
-  };
+  StaggeredImplParams()  {};
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -73,7 +66,8 @@ struct StaggeredImplParams {
 				    RealD, tolerance, 
 				    int,   degree, 
 				    int,   precision,
-				    int,   BoundsCheckFreq);
+				    int,   BoundsCheckFreq,
+				    RealD, BoundsCheckTol);
     
   // MaxIter and tolerance, vectors??
     
@@ -84,15 +78,61 @@ struct StaggeredImplParams {
 				RealD tol      = 1.0e-8, 
                            	int _degree    = 10,
 				int _precision = 64,
-				int _BoundsCheckFreq=20)
+				int _BoundsCheckFreq=20,
+				double _BoundsCheckTol=1e-6)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
 	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);
 

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -68,16 +68,9 @@ public:
   ///////////////////////////////////////////////////////////////
   // Support for MADWF tricks
   ///////////////////////////////////////////////////////////////
-  RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
-  RealD MassPlus(void) { return mass_plus; };
-  RealD MassMinus(void) { return mass_minus; };
+  RealD Mass(void) { return mass; };
   void  SetMass(RealD _mass) { 
-    mass_plus=mass_minus=_mass; 
-    SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
-  } ;
-  void  SetMass(RealD _mass_plus, RealD _mass_minus) { 
-    mass_plus=_mass_plus;
-    mass_minus=_mass_minus;
+    mass=_mass; 
     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
   } ;
   void  P(const FermionField &psi, FermionField &chi);
@@ -115,7 +108,7 @@ public:
   void   MeooeDag5D    (const FermionField &in, FermionField &out);
 
   //    protected:
-  RealD mass_plus, mass_minus;
+  RealD mass;
 
   // Save arguments to SetCoefficientsInternal
   Vector<Coeff_t> _gamma;

Grid/qcd/action/fermion/CloverHelpers.h

@@ -1,435 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
-
-    Copyright (C) 2017 - 2022
-
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
-
-    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
-
-#include <Grid/Grid.h>
-#include <Grid/qcd/spin/Dirac.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-
-////////////////////////////////////////////
-// Standard Clover
-//   (4+m0) + csw * clover_term
-// Exp Clover
-//   (4+m0) * exp(csw/(4+m0) clover_term)
-//   = (4+m0) + csw * clover_term + ...
-////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-
-//////////////////////////////////
-// Generic Standard Clover
-//////////////////////////////////
-
-template<class Impl>
-class CloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
-    GridBase *grid = CloverTerm.Grid();
-    CloverTerm += diag_mass;
-
-    int lvol = grid->lSites();
-    int DimRep = Impl::Dimension;
-    {
-      autoView(CTv,CloverTerm,CpuRead);
-      autoView(CTIv,CloverTermInv,CpuWrite);
-      thread_for(site, lvol, {
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-        Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
-        peekLocalSite(Qx, CTv, lcoor);
-
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++){
-                auto zz =  Qx()(j, k)(a, b);
-                EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
-              }
-
-        EigenInvCloverOp = EigenCloverOp.inverse();
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++)
-                Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-               pokeLocalSite(Qxinv, CTIv, lcoor);
-      });
-    }
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-
-};
-
-
-//////////////////////////////////
-// Generic Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  // Can this be avoided?
-  static void IdentityTimesC(const CloverField& in, RealD c) {
-    int DimRep = Impl::Dimension;
-
-    autoView(in_v, in, AcceleratorWrite);
-
-    accelerator_for(ss, in.Grid()->oSites(), 1, {
-      for (int sa=0; sa<Ns; sa++)
-        for (int ca=0; ca<DimRep; ca++)
-          in_v[ss]()(sa,sa)(ca,ca) = c;
-    });
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-    GridBase* grid = Clover.Grid();
-    CloverField ExpClover(grid);
-
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-
-    Clover *= (1.0/diag_mass);
-
-    // Taylor expansion, slow but generic
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-    // qN = cN
-    // qn = cn + qn+1 X
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
-      cn[i] = cn[i-1] / RealD(i);
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * Clover + cn[i];
-
-    // prepare inverse
-    CloverInv = (-1.0)*Clover;
-
-    Clover = ExpClover * diag_mass;
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * CloverInv + cn[i];
-
-    CloverInv = ExpClover * (1.0/diag_mass);
-
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-//////////////////////////////////
-// Compact Standard Clover
-//////////////////////////////////
-
-
-template<class Impl>
-class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
-                            public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-    Clover += diag_mass;
-  }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
-                          CloverTriangleField& Triangle,
-                          RealD csw_t, RealD diag_mass) {
-
-    // Do nothing
-  }
-
-  // TODO: implement Cmunu for better performances with compact layout, but don't do it
-  // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-};
-
-//////////////////////////////////
-// Compact Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-    // do nothing!
-    // mass term is multiplied to exp(Clover) below
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
-
-  	  typedef iMatrix<ComplexD,6> mat;
-
-  	  RealD qn[6];
-  	  RealD qnold[6];
-  	  RealD p[5];
-  	  RealD trA2, trA3, trA4;
-
-  	  mat A2, A3, A4, A5;
-  	  A2 = alpha * alpha * arg * arg;
-  	  A3 = alpha * arg * A2;
-  	  A4 = A2 * A2;
-  	  A5 = A2 * A3;
-
-  	  trA2 = toReal( trace(A2) );
-  	  trA3 = toReal( trace(A3) );
-  	  trA4 = toReal( trace(A4));
-
-  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
-  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
-  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
-  	  p[3] = trA3 / 3.0;
-  	  p[4] = 0.5 * trA2;
-
-  	  qnold[0] = cN[Niter];
-  	  qnold[1] = 0.0;
-  	  qnold[2] = 0.0;
-  	  qnold[3] = 0.0;
-  	  qnold[4] = 0.0;
-  	  qnold[5] = 0.0;
-
-  	  for(int i = Niter-1; i >= 0; i--)
-  	  {
-  	   qn[0] = p[0] * qnold[5] + cN[i];
-  	   qn[1] = p[1] * qnold[5] + qnold[0];
-  	   qn[2] = p[2] * qnold[5] + qnold[1];
-  	   qn[3] = p[3] * qnold[5] + qnold[2];
-  	   qn[4] = p[4] * qnold[5] + qnold[3];
-  	   qn[5] = qnold[4];
-
-  	   qnold[0] = qn[0];
-  	   qnold[1] = qn[1];
-  	   qnold[2] = qn[2];
-  	   qnold[3] = qn[3];
-  	   qnold[4] = qn[4];
-  	   qnold[5] = qn[5];
-  	  }
-
-  	  mat unit(1.0);
-
-  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
-
-    }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Diagonal.Grid();
-    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
-
-    //
-    // Implementation completely in Daniel's layout
-    //
-
-    // Taylor expansion with Cayley-Hamilton recursion
-    // underlying Horner scheme as above
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++){
-      cn[i] = cn[i-1] / RealD(i);
-    }
-
-      // Taken over from Daniel's implementation
-      conformable(Diagonal, Triangle);
-
-      long lsites = grid->lSites();
-    {
-      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
-      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-      typedef iMatrix<ComplexD,6> mat;
-
-      autoView(diagonal_v,  Diagonal,  CpuRead);
-      autoView(triangle_v,  Triangle,  CpuRead);
-      autoView(diagonalExp_v, Diagonal, CpuWrite);
-      autoView(triangleExp_v, Triangle, CpuWrite);
-
-      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
-
-    	  mat srcCloverOpUL(0.0); // upper left block
-    	  mat srcCloverOpLR(0.0); // lower right block
-    	  mat ExpCloverOp;
-
-        scalar_object_diagonal diagonal_tmp     = Zero();
-        scalar_object_diagonal diagonal_exp_tmp = Zero();
-        scalar_object_triangle triangle_tmp     = Zero();
-        scalar_object_triangle triangle_exp_tmp = Zero();
-
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-
-        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-        peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-        int block;
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-        		if (i == j){
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-        		}
-        		else{
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-        		}
-        	}
-        }
-        block = 1;
-        for(int i = 0; i < 6; i++){
-          	for(int j = 0; j < 6; j++){
-           		if (i == j){
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-           		}
-           		else{
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-           		}
-            }
-        }
-
-        // exp(Clover)
-
-        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-            	if (i == j){
-            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-            	}
-            	else if(i < j){
-            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-            	}
-           	}
-        }
-
-        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 1;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-              	if (i == j){
-              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-               	}
-               	else if(i < j){
-               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-               	}
-            }
-        }
-
-        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
-        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
-      });
-    }
-
-    Diagonal *= diag_mass;
-    Triangle *= diag_mass;
-  }
-
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -31,7 +31,6 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -86,7 +85,7 @@ NAMESPACE_BEGIN(Grid);
 //           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
 //                                                                 =  84 complex words per site
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
 class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
                                    public WilsonCloverHelpers<Impl>,
                                    public CompactWilsonCloverHelpers<Impl> {

Grid/qcd/action/fermion/Fermion.h

@@ -138,52 +138,38 @@ typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
 // Clover fermions
-template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
-template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
+typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
+typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
-typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
-typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
+typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
+typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
 
-typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
-typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
-typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
 
-typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
-typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
-typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
-
-typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
-
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Compact Clover fermions
-template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
-template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
+typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
+typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
 
-typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
-typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
-typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
+typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
+typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
 
-typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
-typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
-typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
 
-typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
-typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
-typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
-
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
-
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Domain Wall fermions
 typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;

Grid/qcd/action/fermion/FermionOperator.h

@@ -49,8 +49,6 @@ public:
 
   virtual FermionField &tmp(void) = 0;
 
-  virtual void DirichletBlock(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/WilsonCloverFermion.h

@@ -32,7 +32,6 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -52,7 +51,7 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
 class WilsonCloverFermion : public WilsonFermion<Impl>,
                             public WilsonCloverHelpers<Impl>
 {

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -209,8 +209,6 @@ public:
 };
 
 
-////////////////////////////////////////////////////////
-
 template<class Impl> class CompactWilsonCloverHelpers {
 public:
 

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -47,6 +47,8 @@ class CompactWilsonCloverTypes {
 public:
   INHERIT_IMPL_TYPES(Impl);
 
+  static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
+
   static constexpr int Nred      = Nc * Nhs;        // 6
   static constexpr int Nblock    = Nhs;             // 2
   static constexpr int Ndiagonal = Nred;            // 6

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,10 +75,6 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
-  int Dirichlet;
-  Coordinate Block; 
-
-  /********** Deprecate timers **********/
   void Report(void);
   void ZeroCounters(void);
   double DhopCalls;
@@ -177,18 +173,7 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-
-  virtual void DirichletBlock(Coordinate & block)
-  {
-    assert(block.size()==Nd+1);
-    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
-      Dirichlet = 1;
-      Block = block;
-      Stencil.DirichletBlock(block); 
-      StencilEven.DirichletBlock(block); 
-      StencilOdd.DirichletBlock(block);
-    }
-  }
+    
   // Constructors
   /*
     WilsonFermion5D(int simd, 

Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h

@@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FiveDimRedBlackGrid,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
-  mass_plus(_mass), mass_minus(_mass)
+  mass(_mass)
 { 
 }
 
@@ -209,8 +209,8 @@ void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
   Vector<Coeff_t> diag (Ls,1.0);
-  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
-  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -220,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
   Vector<Coeff_t> diag = bs;
   Vector<Coeff_t> upper= cs;
   Vector<Coeff_t> lower= cs; 
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
 }
 // FIXME Redunant with the above routine; check this and eliminate
@@ -235,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -250,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &
     upper[i]=-cee[i];
     lower[i]=-cee[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -266,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &
     // Assemble the 5d matrix
     if ( s==0 ) {
       upper[s] = -cee[s+1] ;
-      lower[s] = mass_minus*cee[Ls-1];
+      lower[s] = mass*cee[Ls-1];
     } else if ( s==(Ls-1)) { 
-      upper[s] = mass_plus*cee[0];
+      upper[s] = mass*cee[0];
       lower[s] = -cee[s-1];
     } else {
       upper[s]=-cee[s+1];
@@ -291,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
   Vector<Coeff_t> diag(Ls,1.0);
   Vector<Coeff_t> upper(Ls,-1.0);
   Vector<Coeff_t> lower(Ls,-1.0);
-  upper[Ls-1]=-mass_plus*upper[Ls-1];
-  lower[0]   =-mass_minus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
 }
 
@@ -307,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
       upper[s] = cs[s+1];
-      lower[s] =-mass_minus*cs[Ls-1];
+      lower[s] =-mass*cs[Ls-1];
     } else if ( s==(Ls-1) ) { 
-      upper[s] =-mass_plus*cs[0];
+      upper[s] =-mass*cs[0];
       lower[s] = cs[s-1];
     } else { 
       upper[s] = cs[s+1];
@@ -552,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
-      leem[i]=mass_minus*cee[Ls-1]/bee[0];
+      leem[i]=mass*cee[Ls-1]/bee[0];
       for(int j=0;j<i;j++) {
 	assert(bee[j+1]!=Coeff_t(0.0));
 	leem[i]*= aee[j]/bee[j+1];
@@ -560,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
-      ueem[i]=mass_plus;
+      ueem[i]=mass;
       for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
       ueem[i]*= aee[0]/bee[0];
       
@@ -573,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
   }
 	
   { 
-    Coeff_t delta_d=mass_minus*cee[Ls-1];
+    Coeff_t delta_d=mass*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
       assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
@@ -642,10 +642,6 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
-  
 #if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
@@ -781,8 +777,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   assert(mu>=0);
   assert(mu<Nd);
 
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
 
 #if 0
   int tshift = (mu == Nd-1) ? 1 : 0;

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

@@ -32,18 +32,17 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 
-
 NAMESPACE_BEGIN(Grid);
-template<class Impl, class CloverHelpers>
-CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
-                                                                            GridCartesian& Fgrid,
-                                                                            GridRedBlackCartesian& Hgrid,
-                                                                            const RealD _mass,
-                                                                            const RealD _csw_r,
-                                                                            const RealD _csw_t,
-                                                                            const RealD _cF,
-                                                                            const WilsonAnisotropyCoefficients& clover_anisotropy,
-                                                                            const ImplParams& impl_p)
+template<class Impl>
+CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
+                                                             GridCartesian& Fgrid,
+                                                             GridRedBlackCartesian& Hgrid,
+                                                             const RealD _mass,
+                                                             const RealD _csw_r,
+                                                             const RealD _csw_t,
+                                                             const RealD _cF,
+                                                             const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                                             const ImplParams& impl_p)
   : WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
   , csw_r(_csw_r)
   , csw_t(_csw_t)
@@ -59,55 +58,50 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , BoundaryMask(&Fgrid)
   , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
 {
-  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
-
   csw_r *= 0.5;
   csw_t *= 0.5;
   if (clover_anisotropy.isAnisotropic)
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries) {
-    this->BoundaryMaskEven.Checkerboard() = Even;
-    this->BoundaryMaskOdd.Checkerboard() = Odd;
+  if (open_boundaries)
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
-  }
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
   if(this->open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
   if(this->open_boundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
@@ -115,8 +109,8 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in,
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
@@ -124,20 +118,20 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& i
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
@@ -150,13 +144,13 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
   Mooee(in, out); // blocks are hermitian
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
@@ -169,23 +163,23 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
   MooeeInv(in, out); // blocks are hermitian
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
   DhopDir(in, out, dir, disp);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
   DhopDirAll(in, out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
   assert(!open_boundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
@@ -257,7 +251,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force,
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);   // checked
       count++;
     }
 
@@ -267,18 +261,18 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force,
   force += clover_force;
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&        in,
                     FermionField&              out,
                     const CloverDiagonalField& diagonal,
                     const CloverTriangleField& triangle) {
@@ -291,8 +285,8 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const Fermio
   CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   // NOTE: parts copied from original implementation
 
   // Import gauge into base class
@@ -324,27 +318,22 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  // Handle mass term based on clover policy
-  CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
-  
+  TmpOriginal += this->diag_mass;
+
   // Convert the data layout of the clover term
   double t4 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Exponentiate the clover (nothing happens in case of the standard clover)
-  double t5 = usecond();
-  CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
-
   // Possible modify the boundary values
-  double t6 = usecond();
+  double t5 = usecond();
   if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
-  double t7 = usecond();
+  // Invert the clover term in the improved layout
+  double t6 = usecond();
   CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
 
   // Fill the remaining clover fields
-  double t8 = usecond();
+  double t7 = usecond();
   pickCheckerboard(Even, DiagonalEven,    Diagonal);
   pickCheckerboard(Even, TriangleEven,    Triangle);
   pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
@@ -355,19 +344,20 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
 
   // Report timings
-  double t9 = usecond();
-
-  std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "convert =                    " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "exponentiation =             " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "boundaries =                 " << (t7 - t6) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "inversions =                 " << (t8 - t7) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
+  double t8 = usecond();
+#if 0
+  std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
+            << ", allocations = "               << (t2 - t1) / 1e6
+            << ", field strength = "            << (t3 - t2) / 1e6
+            << ", fill clover = "               << (t4 - t3) / 1e6
+            << ", convert = "                   << (t5 - t4) / 1e6
+            << ", boundaries = "                << (t6 - t5) / 1e6
+            << ", inversions = "                << (t7 - t6) / 1e6
+            << ", pick cbs = "                  << (t8 - t7) / 1e6
+            << ", total = "                     << (t8 - t0) / 1e6
+            << std::endl;
+#endif
 }
 
 NAMESPACE_END(Grid);

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

@@ -34,8 +34,8 @@
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl, class CloverHelpers>
-WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField&                         _Umu,
+template<class Impl>
+WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField&                         _Umu,
                                                GridCartesian&                      Fgrid,
                                                GridRedBlackCartesian&              Hgrid,
                                                const RealD                         _mass,
@@ -74,8 +74,8 @@ WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField&
 }
 
 // *NOT* EO
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -89,8 +89,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, Fermion
   out += temp;
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -104,8 +104,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, Ferm
   out += temp;
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
+template <class Impl>
+void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
@@ -131,11 +131,47 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
   CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
   CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
   CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
-   
+  CloverTerm += diag_mass;
+
   double t4 = usecond();
-  CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
+  int lvol = _Umu.Grid()->lSites();
+  int DimRep = Impl::Dimension;
 
   double t5 = usecond();
+  {
+    autoView(CTv,CloverTerm,CpuRead);
+    autoView(CTIv,CloverTermInv,CpuWrite);
+    thread_for(site, lvol, {
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(site, lcoor);
+      Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
+      peekLocalSite(Qx, CTv, lcoor);
+      //if (csw!=0){
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++){
+	      auto zz =  Qx()(j, k)(a, b);
+	      EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+	    }
+      //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
+      
+      EigenInvCloverOp = EigenCloverOp.inverse();
+      //std::cout << EigenInvCloverOp << std::endl;
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++)
+	      Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+      //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
+      //  }
+      pokeLocalSite(Qxinv, CTIv, lcoor);
+    });
+  }
+
+  double t6 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@@ -148,44 +184,48 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
 
   pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
   pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
-  double t6 = usecond();
+  double t7 = usecond();
 
-  std::cout << GridLogDebug << "WilsonCloverFermion::ImportGauge timings:" << std::endl;
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "instantiation =              " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "pick cbs =                   " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "total =                      " << (t6 - t0) / 1e6 << std::endl;
+#if 0
+  std::cout << GridLogMessage << "WilsonCloverFermion::ImportGauge timings:"
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
+            << ", allocations = "               << (t2 - t1) / 1e6
+            << ", field strength = "            << (t3 - t2) / 1e6
+            << ", fill clover = "               << (t4 - t3) / 1e6
+            << ", misc = "                      << (t5 - t4) / 1e6
+            << ", inversions = "                << (t6 - t5) / 1e6
+            << ", pick cbs = "                  << (t7 - t6) / 1e6
+            << ", total = "                     << (t7 - t0) / 1e6
+            << std::endl;
+#endif
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseNo);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseNo);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseYes);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseYes);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
 {
   out.Checkerboard() = in.Checkerboard();
   CloverField *Clover;
@@ -238,8 +278,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField
 } // MooeeInternal
 
 // Derivative parts unpreconditioned pseudofermions
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
 {
   conformable(X.Grid(), Y.Grid());
   conformable(X.Grid(), force.Grid());
@@ -309,7 +349,7 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const F
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);                   // checked
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);                   // checked
       count++;
     }
 
@@ -320,15 +360,15 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const F
 }
 
 // Derivative parts
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
 {
   assert(0);
 }
 
 // Derivative parts
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   assert(0); // not implemented yet
 }

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

@@ -60,8 +60,7 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   UmuOdd (_FourDimRedBlackGrid),
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid),
-  _tmp(&FiveDimRedBlackGrid),
-  Dirichlet(0)
+  _tmp(&FiveDimRedBlackGrid)
 {
   // some assertions
   assert(FiveDimGrid._ndimension==5);
@@ -219,14 +218,6 @@ 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/WilsonFermionImplementation.h

@@ -4,13 +4,12 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
 
-Copyright (C) 2022
+Copyright (C) 2015
 
 Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Fabian Joswig <fabian.joswig@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
@@ -600,47 +599,11 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
                                                    Current curr_type,
                                                    unsigned int mu)
 {
-  if(curr_type != Current::Vector)
-  {
-    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
-    exit(1);
-  }
-
   Gamma g5(Gamma::Algebra::Gamma5);
   conformable(_grid, q_in_1.Grid());
   conformable(_grid, q_in_2.Grid());
   conformable(_grid, q_out.Grid());
-  auto UGrid= this->GaugeGrid();
-
-  PropagatorField tmp_shifted(UGrid);
-  PropagatorField g5Lg5(UGrid);
-  PropagatorField R(UGrid);
-  PropagatorField gmuR(UGrid);
-
-    Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT,
-  };
-  Gamma gmu=Gamma(Gmu[mu]);
-
-  g5Lg5=g5*q_in_1*g5;
-  tmp_shifted=Cshift(q_in_2,mu,1);
-  Impl::multLinkField(R,this->Umu,tmp_shifted,mu);
-  gmuR=gmu*R;
-
-  q_out=adj(g5Lg5)*R;
-  q_out-=adj(g5Lg5)*gmuR;
-
-  tmp_shifted=Cshift(q_in_1,mu,1);
-  Impl::multLinkField(g5Lg5,this->Umu,tmp_shifted,mu);
-  g5Lg5=g5*g5Lg5*g5;
-  R=q_in_2;
-  gmuR=gmu*R;
-
-  q_out-=adj(g5Lg5)*R;
-  q_out-=adj(g5Lg5)*gmuR;
+  assert(0);
 }
 
 
@@ -654,51 +617,9 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
                                               unsigned int tmax,
 					      ComplexField &lattice_cmplx)
 {
-  if(curr_type != Current::Vector)
-  {
-    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
-    exit(1);
-  }
-
-  int tshift = (mu == Nd-1) ? 1 : 0;
-  unsigned int LLt    = GridDefaultLatt()[Tp];
   conformable(_grid, q_in.Grid());
   conformable(_grid, q_out.Grid());
-  auto UGrid= this->GaugeGrid();
-
-  PropagatorField tmp(UGrid);
-  PropagatorField Utmp(UGrid);
-  PropagatorField L(UGrid);
-  PropagatorField zz (UGrid);
-  zz=Zero();
-  LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
-
-    Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT,
-  };
-  Gamma gmu=Gamma(Gmu[mu]);
-
-  tmp = Cshift(q_in,mu,1);
-  Impl::multLinkField(Utmp,this->Umu,tmp,mu);
-  tmp = ( Utmp*lattice_cmplx - gmu*Utmp*lattice_cmplx ); // Forward hop
-  tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
-  q_out = where((lcoor<=tmax),tmp,zz); // Position of current complicated
-
-  tmp = q_in *lattice_cmplx;
-  tmp = Cshift(tmp,mu,-1);
-  Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
-  tmp = -( Utmp + gmu*Utmp );
-  // Mask the time
-  if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
-    unsigned int t0 = 0;
-    tmp = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
-  } else {
-    tmp = where((lcoor>=tmin+tshift),tmp,zz);
-  }
-  q_out+= where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
+  assert(0);
 }
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master

@@ -9,7 +9,6 @@
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
 
     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
@@ -33,12 +32,10 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>; 
-template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>; 
+template class CompactWilsonCloverFermion<IMPLEMENTATION>; 
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonCloverFermionInstantiation.cc.master

@@ -8,8 +8,7 @@
 
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
-    
+
     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
@@ -32,12 +31,10 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class WilsonCloverFermion<IMPLEMENTATION, CloverHelpers<IMPLEMENTATION>>; 
-template class WilsonCloverFermion<IMPLEMENTATION, ExpCloverHelpers<IMPLEMENTATION>>; 
+template class WilsonCloverFermion<IMPLEMENTATION>; 
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+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/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/generate_instantiations.sh

@@ -18,10 +18,6 @@ WILSON_IMPL_LIST=" \
 	   GparityWilsonImplF \
 	   GparityWilsonImplD "
 
-COMPACT_WILSON_IMPL_LIST=" \
-	   WilsonImplF \
-	   WilsonImplD "
-
 DWF_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
@@ -44,23 +40,13 @@ EOF
 
 done
 
-CC_LIST="WilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
+CC_LIST="WilsonCloverFermionInstantiation CompactWilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
 
 for impl in $WILSON_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
-done
-done
-
-CC_LIST="CompactWilsonCloverFermionInstantiation"
-
-for impl in $COMPACT_WILSON_IMPL_LIST
-do
-for f in $CC_LIST
-do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 
@@ -77,14 +63,14 @@ for impl in $DWF_IMPL_LIST $GDWF_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 
 # overwrite the .cc file in Gparity directories
 for impl in $GDWF_IMPL_LIST
 do
-  ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
+  ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc 
 done
 
 
@@ -98,7 +84,7 @@ for impl in $STAG_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 

Grid/qcd/action/filters/DDHMCFilter.h

@@ -1,102 +0,0 @@
-/*************************************************************************************
-
-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

@@ -1,71 +0,0 @@
-/*************************************************************************************
-
-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/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

@@ -40,13 +40,66 @@ 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/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 +195,16 @@ 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);
+	  //Lop.RefreshShiftCoefficients(-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 +223,16 @@ 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]);
+          //Rop.RefreshShiftCoefficients(-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 +242,119 @@ NAMESPACE_BEGIN(Grid);
         Phi = Phi + tmp[1];
 
         // Reset shift coefficients for energy and force evals
-        Lop.RefreshShiftCoefficients(0.0);
-        Rop.RefreshShiftCoefficients(-1.0);
+        //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 +378,7 @@ 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 +388,26 @@ 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 +456,40 @@ 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/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

@@ -83,16 +83,10 @@ NAMESPACE_BEGIN(Grid);
 	return sstream.str();
       } 
 
-      
-      virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
+      //Access the fermion field
+      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 +94,22 @@ 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
+        //
+        // Take phi_o = Vpcdag^{-1} Mpcdag eta_o  ; eta_o = Mpcdag^{-1} Vpcdag Phi
         FermionField etaOdd (NumOp.FermionRedBlackGrid());
         FermionField etaEven(NumOp.FermionRedBlackGrid());
         FermionField tmp    (NumOp.FermionRedBlackGrid());
 
-        gaussian(pRNG,eta);
-
         pickCheckerboard(Even,etaEven,eta);
         pickCheckerboard(Odd,etaOdd,eta);
 
@@ -125,8 +129,8 @@ NAMESPACE_BEGIN(Grid);
         DenOp.MooeeDag(etaEven,tmp);
         NumOp.MooeeInvDag(tmp,PhiEven);
 
-        PhiOdd =PhiOdd*scale;
-        PhiEven=PhiEven*scale;
+        //PhiOdd =PhiOdd*scale;
+        //PhiEven=PhiEven*scale;
         
       };
 

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

@@ -151,12 +151,22 @@ public:
       Resources.GetCheckPointer()->CheckpointRestore(Parameters.StartTrajectory, U,
 						     Resources.GetSerialRNG(),
 						     Resources.GetParallelRNG());
+    } else if (Parameters.StartingType == "CheckpointStartReseed") {
+      // Same as CheckpointRestart but reseed the RNGs using the fixed integer seeding used for ColdStart and HotStart
+      // Useful for creating new evolution streams from an existing stream
+      
+      // WARNING: Unfortunately because the checkpointer doesn't presently allow us to separately restore the RNG and gauge fields we have to load
+      // an existing RNG checkpoint first; make sure one is available and named correctly
+      Resources.GetCheckPointer()->CheckpointRestore(Parameters.StartTrajectory, U,
+						     Resources.GetSerialRNG(),
+						     Resources.GetParallelRNG());
+      Resources.SeedFixedIntegers();      
     } else {
       // others
       std::cout << GridLogError << "Unrecognized StartingType\n";
       std::cout
 	<< GridLogError
-	<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+	<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart, CheckpointStartReseed]\n";
       exit(1);
     }
   }
@@ -176,9 +186,6 @@ private:
     typedef IntegratorType<SmearingPolicy> TheIntegrator;
     TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
 
-    // Sets the momentum filter
-    MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
-
     Smearing.set_Field(U);
 
     HybridMonteCarlo<TheIntegrator> HMC(Parameters, MDynamics,

Grid/qcd/hmc/HMC.h

@@ -34,7 +34,6 @@ directory
 			    * @brief Classes for Hybrid Monte Carlo update
 			    *
 			    * @author Guido Cossu
-			    * @author Peter Boyle
 			    */
 			   //--------------------------------------------------------------------
 #pragma once
@@ -116,17 +115,22 @@ private:
 
     random(sRNG, rn_test);
 
-    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";
+    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 << GridLogHMC << "Metropolis_test -- ACCEPTED\n";
-      std::cout << GridLogHMC << "--------------------------------------------------\n";
+      std::cout << GridLogHMC
+                << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
       std::cout << GridLogHMC << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogHMC << "--------------------------------------------------\n";
+      std::cout << GridLogHMC
+                << "--------------------------------------------------\n";
       return false;
     }
   }
@@ -135,68 +139,19 @@ private:
   // Evolution
   /////////////////////////////////////////////////////////
   RealD evolve_hmc_step(Field &U) {
+    TheIntegrator.refresh(U, sRNG, pRNG);  // set U and initialize P and phi's
 
-    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";
+    RealD H0 = TheIntegrator.S(U);  // initial state action
 
     std::streamsize current_precision = std::cout.precision();
     std::cout.precision(15);
     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";
 
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // updated state action
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Compute final action";
-    RealD H1 = TheIntegrator.S(U);  
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    RealD H1 = TheIntegrator.S(U);  // updated state action
 
-
-    
     ///////////////////////////////////////////////////////////
     if(0){
       std::cout << "------------------------- Reversibility test" << std::endl;
@@ -208,16 +163,17 @@ private:
     }
     ///////////////////////////////////////////////////////////
 
+
     std::cout.precision(15);
-
-    std::cout << GridLogHMC << "--------------------------------------------------\n";
-    std::cout << GridLogHMC << "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.precision(current_precision);
     
     return (H1 - H0);
   }
+  
+
+  
 
 public:
   /////////////////////////////////////////
@@ -239,11 +195,8 @@ 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 << GridLogHMC << "-- # Trajectory = " << traj << "\n";
-
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
       	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
       }
@@ -263,10 +216,11 @@ public:
       if (accept)
         Ucur = Ucopy; 
       
+     
+      
       double t1=usecond();
       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;

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

@@ -72,8 +72,6 @@ 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;
@@ -82,9 +80,6 @@ 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;
 
@@ -95,7 +90,6 @@ 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
@@ -107,7 +101,7 @@ class HMCResourceManager {
 
 
 public:
-  HMCResourceManager() : have_RNG(false), have_CheckPointer(false), have_Filter(false) {}
+  HMCResourceManager() : have_RNG(false), have_CheckPointer(false) {}
 
   template <class ReaderClass, class vector_type = vComplex >
   void initialize(ReaderClass &Read){
@@ -135,7 +129,6 @@ 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...
@@ -215,16 +208,6 @@ 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;

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,6 +33,7 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
+#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -66,7 +67,6 @@ public:
 template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
-
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 
@@ -119,58 +119,42 @@ 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();
 
-      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 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;
+      Real max_force_abs = std::sqrt(maxLocalNorm2(force));
+      Real max_impulse_abs = max_force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
 
+      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << " Max force: " << max_force_abs << " Time step: " << ep << " Impulse average: " << impulse_abs << " Max impulse: " << max_impulse_abs << 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) 
@@ -184,12 +168,8 @@ 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(MomFiltered, U, ep);
+    FieldImplementation::update_field(Mom, U, ep);
 
     // Update the smeared fields, can be implemented as observer
     Smearer.set_Field(U);
@@ -232,66 +212,6 @@ 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;
@@ -310,6 +230,7 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
+
   }
 
   void reverse_momenta()
@@ -334,15 +255,19 @@ 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
@@ -352,19 +277,16 @@ 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;
-
+	std::cout << GridLogIntegrator << "Refreshing integrator level " << level << " index " << actionID << 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
@@ -399,9 +321,7 @@ 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 +40,46 @@ 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;
       }
     }
   }  
 
-  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) {
+  //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) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
-    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
+    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog);
   }
-  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
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform, Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
 
     GridBase *grid = Umu.Grid();
 
@@ -122,29 +141,24 @@ public:
 
       }
     }
-    std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
-    if (err_on_no_converge) assert(0);
+    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 +173,7 @@ 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 +217,16 @@ 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
   //////////////////////////////////////////////////
@@ -164,7 +215,7 @@ public:
 
     double vol = Umu.Grid()->gSites();
 
-    return p.real() / vol / (4.0 * Nc ) ;
+    return p.real() / vol / 4.0 / 3.0;
   };
 
   //////////////////////////////////////////////////
@@ -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/random/gaussian.h

@@ -0,0 +1,200 @@
+// -*- C++ -*-
+//===--------------------------- random -----------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+// Peter Boyle: Taken from libc++ in Clang/LLVM.
+// Reason is that libstdc++ and clang differ in their return order in the normal_distribution / box mueller type step.
+// standardise on one and call it "gaussian_distribution".
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <cmath>
+#include <type_traits>
+#include <initializer_list>
+#include <limits>
+#include <algorithm>
+#include <numeric>
+#include <vector>
+#include <string>
+#include <istream>
+#include <ostream>
+#include <random>
+
+// normal_distribution -> gaussian distribution
+namespace Grid {
+
+template<class _RealType = double>
+class  gaussian_distribution
+{
+public:
+    // types
+    typedef _RealType result_type;
+
+    class param_type
+    {
+        result_type __mean_;
+        result_type __stddev_;
+    public:
+        typedef gaussian_distribution distribution_type;
+
+        strong_inline
+        explicit param_type(result_type __mean = 0, result_type __stddev = 1)
+            : __mean_(__mean), __stddev_(__stddev) {}
+
+        strong_inline
+        result_type mean() const {return __mean_;}
+        strong_inline
+        result_type stddev() const {return __stddev_;}
+
+        friend strong_inline
+            bool operator==(const param_type& __x, const param_type& __y)
+            {return __x.__mean_ == __y.__mean_ && __x.__stddev_ == __y.__stddev_;}
+        friend strong_inline
+            bool operator!=(const param_type& __x, const param_type& __y)
+            {return !(__x == __y);}
+    };
+
+private:
+    param_type __p_;
+    result_type _V_;
+    bool _V_hot_;
+
+public:
+    // constructors and reset functions
+    strong_inline
+    explicit gaussian_distribution(result_type __mean = 0, result_type __stddev = 1)
+        : __p_(param_type(__mean, __stddev)), _V_hot_(false) {}
+    strong_inline
+    explicit gaussian_distribution(const param_type& __p)
+        : __p_(__p), _V_hot_(false) {}
+    strong_inline
+    void reset() {_V_hot_ = false;}
+
+    // generating functions
+    template<class _URNG>
+        strong_inline
+        result_type operator()(_URNG& __g)
+        {return (*this)(__g, __p_);}
+    template<class _URNG> result_type operator()(_URNG& __g, const param_type& __p);
+
+    // property functions
+    strong_inline
+    result_type mean() const {return __p_.mean();}
+    strong_inline
+    result_type stddev() const {return __p_.stddev();}
+
+    strong_inline
+    param_type param() const {return __p_;}
+    strong_inline
+    void param(const param_type& __p) {__p_ = __p;}
+
+    strong_inline
+    result_type min() const {return -std::numeric_limits<result_type>::infinity();}
+    strong_inline
+    result_type max() const {return std::numeric_limits<result_type>::infinity();}
+
+    friend strong_inline
+        bool operator==(const gaussian_distribution& __x,
+                        const gaussian_distribution& __y)
+        {return __x.__p_ == __y.__p_ && __x._V_hot_ == __y._V_hot_ &&
+                (!__x._V_hot_ || __x._V_ == __y._V_);}
+    friend strong_inline
+        bool operator!=(const gaussian_distribution& __x,
+                        const gaussian_distribution& __y)
+        {return !(__x == __y);}
+
+    template <class _CharT, class _Traits, class _RT>
+    friend
+    std::basic_ostream<_CharT, _Traits>&
+    operator<<(std::basic_ostream<_CharT, _Traits>& __os,
+               const gaussian_distribution<_RT>& __x);
+
+    template <class _CharT, class _Traits, class _RT>
+    friend
+    std::basic_istream<_CharT, _Traits>&
+    operator>>(std::basic_istream<_CharT, _Traits>& __is,
+               gaussian_distribution<_RT>& __x);
+};
+
+template <class _RealType>
+template<class _URNG>
+_RealType
+gaussian_distribution<_RealType>::operator()(_URNG& __g, const param_type& __p)
+{
+    result_type _Up;
+    if (_V_hot_)
+    {
+        _V_hot_ = false;
+        _Up = _V_;
+    }
+    else
+    {
+        std::uniform_real_distribution<result_type> _Uni(-1, 1);
+        result_type __u;
+        result_type __v;
+        result_type __s;
+        do
+        {
+            __u = _Uni(__g);
+            __v = _Uni(__g);
+            __s = __u * __u + __v * __v;
+        } while (__s > 1 || __s == 0);
+        result_type _Fp = std::sqrt(-2 * std::log(__s) / __s);
+        _V_ = __v * _Fp;
+        _V_hot_ = true;
+        _Up = __u * _Fp;
+    }
+    return _Up * __p.stddev() + __p.mean();
+}
+
+template <class _CharT, class _Traits, class _RT>
+std::basic_ostream<_CharT, _Traits>&
+operator<<(std::basic_ostream<_CharT, _Traits>& __os,
+           const gaussian_distribution<_RT>& __x)
+{
+    auto __save_flags = __os.flags();
+    __os.flags(std::ios_base::dec | std::ios_base::left | std::ios_base::fixed |
+               std::ios_base::scientific);
+    _CharT __sp = __os.widen(' ');
+    __os.fill(__sp);
+    __os << __x.mean() << __sp << __x.stddev() << __sp << __x._V_hot_;
+    if (__x._V_hot_)
+        __os << __sp << __x._V_;
+    __os.flags(__save_flags);
+    return __os;
+}
+
+template <class _CharT, class _Traits, class _RT>
+std::basic_istream<_CharT, _Traits>&
+operator>>(std::basic_istream<_CharT, _Traits>& __is,
+           gaussian_distribution<_RT>& __x)
+{
+    typedef gaussian_distribution<_RT> _Eng;
+    typedef typename _Eng::result_type result_type;
+    typedef typename _Eng::param_type param_type;
+    auto __save_flags = __is.flags();
+    __is.flags(std::ios_base::dec | std::ios_base::skipws);
+    result_type __mean;
+    result_type __stddev;
+    result_type _Vp = 0;
+    bool _V_hot = false;
+    __is >> __mean >> __stddev >> _V_hot;
+    if (_V_hot)
+        __is >> _Vp;
+    if (!__is.fail())
+    {
+        __x.param(param_type(__mean, __stddev));
+        __x._V_hot_ = _V_hot;
+        __x._V_ = _Vp;
+    }
+    __is.flags(__save_flags);
+    return __is;
+}
+}

Grid/stencil/SimpleCompressor.h

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

Grid/stencil/Stencil.h

@@ -133,8 +133,6 @@ class CartesianStencilAccelerator {
   int           _osites;
   StencilVector _directions;
   StencilVector _distances;
-  StencilVector _comms_send;
-  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -228,8 +226,6 @@ public:
     void * recv_buf;
     Integer to_rank;
     Integer from_rank;
-    Integer do_send;
-    Integer do_recv;
     Integer bytes;
   };
   struct Merge {
@@ -244,20 +240,7 @@ 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;
@@ -288,8 +271,7 @@ 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
   ///////////////////////////////////////////////////////////
@@ -302,6 +284,29 @@ 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
   ////////////////////////////////////////
@@ -328,12 +333,11 @@ 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
@@ -342,29 +346,65 @@ 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++){
-      _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);
+      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;
     }
   }
 
@@ -373,6 +413,7 @@ public:
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
+    commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
@@ -380,27 +421,28 @@ public:
   void Communicate(void)
   {
     if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
-      /////////////////////////////////////////////////////////
-      // 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
-      /////////////////////////////////////////////////////////
+      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
       std::vector<std::vector<CommsRequest_t> > reqs;
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
@@ -442,23 +484,31 @@ public:
       sshift[1] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Odd);
       if ( sshift[0] == sshift[1] ) {
 	if (splice_dim) {
-	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx,point);
+	  splicetime-=usecond();
+	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
 	  is_same_node = is_same_node && tmp;
+	  splicetime+=usecond();
 	} else {
-	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx,point);
+	  nosplicetime-=usecond();
+	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
 	  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,point);
-	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx,point);
+	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx);
+	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx);
 	  is_same_node = is_same_node && tmp1 && tmp2;
+	  splicetime+=usecond();
 	} else {
-	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx,point);
-	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx,point);
+	  nosplicetime-=usecond();
+	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
+	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
 	  is_same_node = is_same_node && tmp1 && tmp2;
+	  nosplicetime+=usecond();
 	}
       }
     }
@@ -468,10 +518,13 @@ 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;
 
@@ -485,6 +538,7 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
+    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -497,70 +551,14 @@ public:
     Mergers.resize(0);
     MergersSHM.resize(0);
     Packets.resize(0);
-    CopyReceiveBuffers.resize(0);
-    CachedTransfers.resize(0);
+    calls++;
   }
-  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){
+  void AddPacket(void *xmit,void * rcv, Integer to,Integer from,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);
   }
@@ -580,17 +578,22 @@ 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)
-  {
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
+
+    
+    mergetime-=usecond();
     for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
@@ -600,7 +603,9 @@ 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;
@@ -608,6 +613,7 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
+    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -644,54 +650,18 @@ public:
       }
     }
   }
-  /// 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 );
-      
-	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;
@@ -705,12 +675,8 @@ 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);
 
@@ -727,16 +693,15 @@ 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);
@@ -745,6 +710,7 @@ 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
@@ -845,7 +811,6 @@ 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];
@@ -902,32 +867,25 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-
-      // Wrap locally dirichlet support case OR node local
-      if ( offnode==0 ) {
+      if (!offnode) {
 
 	int permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-	
+
       } else {
 
-	if ( comms_recv==0 ) {
-
-	  int permute_slice=1;
-	  CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-
-	} else { 
-
-	  ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
-
-	}
-
-      }
-      
-      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
+
       }
     }
   }
@@ -1026,14 +984,11 @@ public:
   }
 
   template<class compressor>
-  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx, int point)
+  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
   {
     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());
 
@@ -1056,93 +1011,78 @@ 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
-	int comm_off = u_comm_offset;
+	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 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());
 
-	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 *recv_buf;
+	if ( compress.DecompressionStep() ) {
+	  recv_buf=u_simd_recv_buf[0];
+	} else {
+	  recv_buf=this->u_recv_buf_p;
 	}
 
-	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,bytes,cbmask);
-	if ( (!duplicate) ) { // Force comms for now
+	cobj *send_buf;
+	send_buf = this->u_send_buf_p; // Gather locally, must send
 
-	  ///////////////////////////////////////////////////////////
-	  // 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],
+	////////////////////////////////////////////////////////
+	// 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],
 			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 point)
+  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
   {
     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];
@@ -1157,6 +1097,7 @@ public:
 
 
     int permute_type=_grid->PermuteType(dimension);
+    //    std::cout << "SimdNew permute type "<<permute_type<<std::endl;
 
     ///////////////////////////////////////////////
     // Simd direction uses an extract/merge pair
@@ -1190,9 +1131,8 @@ 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][comm_off];
+	  spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
 	}
 
 	int sx   = (x+sshift)%rd;
@@ -1201,17 +1141,18 @@ 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,comm_off,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]));
 	}
+	gathermtime-=usecond();
 
-	if ( comms_send )
-	  Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
+	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
 
@@ -1228,8 +1169,8 @@ public:
 	  int nbr_plane = nbr_ic;
 	  assert (sx == nbr_ox);
 
-	  auto rp = &u_simd_recv_buf[i        ][comm_off];
-	  auto sp = &u_simd_send_buf[nbr_plane][comm_off];
+	  auto rp = &u_simd_recv_buf[i        ][u_comm_offset];
+	  auto sp = &u_simd_send_buf[nbr_plane][u_comm_offset];
 
 	  if(nbr_proc){
 
@@ -1240,13 +1181,8 @@ public:
 
 	    rpointers[i] = rp;
 
-	    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);
-	    }
+	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
+
 
 	  } else {
 
@@ -1255,12 +1191,9 @@ public:
 	  }
 	}
 
-	if ( comms_recv ) {
-	  AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
-	}
+	AddMerge(&this->u_recv_buf_p[u_comm_offset],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_ACCELERATED
+#ifndef GRID_CUDA
 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 )
 {