CPU open doesn't need to free space

Compact Exponential Cloverterm on GPU

Grid/GridCore.h

@@ -44,8 +44,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridStd.h>
 #include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
-#include <Grid/perfmon/Tracing.h>
-//#include <Grid/perfmon/PerfCount.h>
+#include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
 #include <Grid/allocator/Allocator.h>

Grid/GridQCDcore.h

@@ -36,7 +36,6 @@ 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/algorithms/Algorithms.h

@@ -54,7 +54,6 @@ 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/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
         });

Grid/algorithms/approx/Chebyshev.h

@@ -264,7 +264,7 @@ public:
       auto Tnp_v = Tnp->View();
       auto Tnm_v = Tnm->View();
       constexpr int Nsimd = vector_type::Nsimd();
-      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
+      accelerator_for(ss, in.Grid()->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
       });

Grid/algorithms/iterative/ConjugateGradient.h

@@ -58,7 +58,6 @@ public:
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
-    GRID_TRACE("ConjugateGradient");
     psi.Checkerboard() = src.Checkerboard();
 
     conformable(psi, src);
@@ -118,13 +117,9 @@ public:
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
 
-    RealD usecs = -usecond();
     SolverTimer.Start();
     int k;
     for (k = 1; k <= MaxIterations; k++) {
-
-      GridStopWatch IterationTimer;
-      IterationTimer.Start();
       c = cp;
 
       MatrixTimer.Start();
@@ -157,41 +152,31 @@ public:
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
-      IterationTimer.Stop();
-      if ( (k % 500) == 0 ) {
-	std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
+      std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
                 << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
-      } else { 
-	std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
-		  << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
-      }
 
       // Stopping condition
       if (cp <= rsq) {
-	usecs +=usecond();
         SolverTimer.Stop();
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
-	GridBase *grid = src.Grid();
-	RealD DwfFlops = (1452. )*grid->gSites()*4*k
-   	               + (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra
+
         RealD srcnorm = std::sqrt(norm2(src));
         RealD resnorm = std::sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
+
         std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k 
 		  << "\tComputed residual " << std::sqrt(cp / ssq)
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
-	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
-
-	std::cout << GridLogMessage << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
+        std::cout << GridLogIterative << "Time breakdown "<<std::endl;
+	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -49,7 +49,6 @@ 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;
@@ -69,7 +68,6 @@ 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;
@@ -99,7 +97,6 @@ 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;
 
@@ -133,7 +130,6 @@ 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);
@@ -154,7 +150,6 @@ NAMESPACE_BEGIN(Grid);
     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
     CG_d(Linop_d, src_d_in, sol_d);
     TotalFinalStepIterations = CG_d.IterationsToComplete;
-    TrueResidual = CG_d.TrueResidual;
 
     TotalTimer.Stop();
     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -44,7 +44,7 @@ public:
 
   using OperatorFunction<Field>::operator();
 
-  //  RealD   Tolerance;
+  RealD   Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -52,7 +52,7 @@ public:
   MultiShiftFunction shifts;
   std::vector<RealD> TrueResidualShift;
 
-  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
+  ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
     MaxIterations(maxit),
     shifts(_shifts)
   { 
@@ -84,7 +84,6 @@ public:
 
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
-    GRID_TRACE("ConjugateGradientMultiShift");
   
     GridBase *grid = src.Grid();
   
@@ -183,9 +182,6 @@ 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
@@ -325,8 +321,8 @@ public:
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -1,410 +0,0 @@
-/*************************************************************************************
-
-    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)
-  { 
-    GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
-    GridBase *DoublePrecGrid = src_d.Grid();
-
-    ////////////////////////////////////////////////////////////////////////
-    // Convenience references to the info stored in "MultiShiftFunction"
-    ////////////////////////////////////////////////////////////////////////
-    int nshift = shifts.order;
-
-    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
-    std::vector<RealD> &mresidual(shifts.tolerances);
-    std::vector<RealD> alpha(nshift,1.0);
-
-    //Double precision search directions
-    FieldD p_d(DoublePrecGrid);
-    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
-
-    FieldD tmp_d(DoublePrecGrid);
-    FieldD r_d(DoublePrecGrid);
-    FieldD mmp_d(DoublePrecGrid);
-
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
-  
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
-  
-    const int       primary =0;
-  
-    //Primary shift fields CG iteration
-    RealD a,b,c,d;
-    RealD cp,bp,qq; //prev
-  
-    // Matrix mult fields
-    FieldF r_f(SinglePrecGrid);
-    FieldF p_f(SinglePrecGrid);
-    FieldF tmp_f(SinglePrecGrid);
-    FieldF mmp_f(SinglePrecGrid);
-    FieldF src_f(SinglePrecGrid);
-    precisionChange(src_f, src_d);
-
-    // Check lightest mass
-    for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
-      converged[s]=0;
-    }
-  
-    // Wire guess to zero
-    // Residuals "r" are src
-    // First search direction "p" is also src
-    cp = norm2(src_d);
-
-    // Handle trivial case of zero src.
-    if( cp == 0. ){
-      for(int s=0;s<nshift;s++){
-	psi_d[s] = Zero();
-	IterationsToCompleteShift[s] = 1;
-	TrueResidualShift[s] = 0.;
-      }
-      return;
-    }
-
-    for(int s=0;s<nshift;s++){
-      rsq[s] = cp * mresidual[s] * mresidual[s];
-      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
-      ps_d[s] = src_d;
-    }
-    // r and p for primary
-    r_f=src_f; //residual maintained in single
-    p_f=src_f;
-    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
-  
-    //MdagM+m[0]
-    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    axpy(mmp_f,mass[0],p_f,mmp_f);
-    RealD rn = norm2(p_f);
-    d += rn*mass[0];
-
-    b = -cp /d;
-  
-    // Set up the various shift variables
-    int       iz=0;
-    z[0][1-iz] = 1.0;
-    z[0][iz]   = 1.0;
-    bs[0]      = b;
-    for(int s=1;s<nshift;s++){
-      z[s][1-iz] = 1.0;
-      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
-      bs[s]      = b*z[s][iz]; 
-    }
-  
-    // r += b[0] A.p[0]
-    // c= norm(r)
-    c=axpy_norm(r_f,b,mmp_f,r_f);
-  
-    for(int s=0;s<nshift;s++) {
-      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
-    }
-  
-    ///////////////////////////////////////
-    // Timers
-    ///////////////////////////////////////
-    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
-
-    SolverTimer.Start();
-  
-    // Iteration loop
-    int k;
-  
-    for (k=1;k<=MaxIterations;k++){    
-      a = c /cp;
-
-      //Update double precision search direction by residual
-      PrecChangeTimer.Start();
-      precisionChange(r_d, r_f);
-      PrecChangeTimer.Stop();
-
-      AXPYTimer.Start();
-      axpy(p_d,a,p_d,r_d); 
-
-      for(int s=0;s<nshift;s++){
-	if ( ! converged[s] ) { 
-	  if (s==0){
-	    axpy(ps_d[s],a,ps_d[s],r_d);
-	  } else{
-	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
-	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
-	  }
-	}
-      }
-      AXPYTimer.Stop();
-
-      PrecChangeTimer.Start();
-      precisionChange(p_f, p_d); //get back single prec search direction for linop
-      PrecChangeTimer.Stop();
-
-      cp=c;
-      MatrixTimer.Start();  
-      Linop_f.HermOp(p_f,mmp_f); 
-      d=real(innerProduct(p_f,mmp_f));    
-      MatrixTimer.Stop();  
-
-      AXPYTimer.Start();
-      axpy(mmp_f,mass[0],p_f,mmp_f);
-      AXPYTimer.Stop();
-      RealD rn = norm2(p_f);
-      d += rn*mass[0];
-    
-      bp=b;
-      b=-cp/d;
-    
-      // Toggle the recurrence history
-      bs[0] = b;
-      iz = 1-iz;
-      ShiftTimer.Start();
-      for(int s=1;s<nshift;s++){
-	if((!converged[s])){
-	  RealD z0 = z[s][1-iz];
-	  RealD z1 = z[s][iz];
-	  z[s][iz] = z0*z1*bp
-	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
-	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
-	}
-      }
-      ShiftTimer.Stop();
-
-      //Update double precision solutions
-      AXPYTimer.Start();
-      for(int s=0;s<nshift;s++){
-	int ss = s;
-	if( (!converged[s]) ) { 
-	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
-	}
-      }
-
-      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
-      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
-      AXPYTimer.Stop();
-
-      c = c_f;
-
-      if(k % ReliableUpdateFreq == 0){
-	//Replace r with true residual
-	MatrixTimer.Start();  
-	Linop_d.HermOp(psi_d[0],mmp_d); 
-	MatrixTimer.Stop();  
-
-	AXPYTimer.Start();
-	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
-
-	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
-	AXPYTimer.Stop();
-
-	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
-	
-	PrecChangeTimer.Start();
-	precisionChange(r_f, r_d);
-	PrecChangeTimer.Stop();
-	c = c_d;
-      }
-    
-      // Convergence checks
-      int all_converged = 1;
-      for(int s=0;s<nshift;s++){
-      
-	if ( (!converged[s]) ){
-	  IterationsToCompleteShift[s] = k;
-	
-	  RealD css  = c * z[s][iz]* z[s][iz];
-	
-	  if(css<rsq[s]){
-	    if ( ! converged[s] )
-	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
-	    converged[s]=1;
-	  } else {
-	    all_converged=0;
-	  }
-
-	}
-      }
-
-      if ( all_converged ){
-
-	SolverTimer.Stop();
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
-      
-	// Check answers 
-	for(int s=0; s < nshift; s++) { 
-	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
-	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
-	  axpy(r_d,-alpha[s],src_d,tmp_d);
-	  RealD rn = norm2(r_d);
-	  RealD cn = norm2(src_d);
-	  TrueResidualShift[s] = std::sqrt(rn/cn);
-	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
-
-	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
-	  if(rn >= rsq[s]){
-	    CleanupTimer.Start();
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
-
-	    //Setup linear operators for final cleanup
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
-					       
-	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
-	    cg(src_d, psi_d[s]);
-	    
-	    TrueResidualShift[s] = cg.TrueResidual;
-	    CleanupTimer.Stop();
-	  }
-	}
-
-	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
-	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
-
-	IterationsToComplete = k;	
-
-	return;
-      }
-
-   
-    }
-    // ugly hack
-    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -73,7 +73,6 @@ public:
   }
     
   void operator()(const FieldD &src, FieldD &psi) {
-    GRID_TRACE("ConjugateGradientReliableUpdate");
     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
     bool using_fallback = false;
       

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -44,7 +44,6 @@ 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,
@@ -146,24 +145,16 @@ public:
   LinearOperatorBase<FineField> &_Linop;
   RealD                             _coarse_relax_tol;
   std::vector<FineField>        &_subspace;
-
-  int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
-                                //As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
-                                //To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
-                                //out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
-                                //NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
   
   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
 					   OperatorFunction<FineField>   &smoother,
 					   LinearOperatorBase<FineField> &Linop,
 					   std::vector<FineField>        &subspace,
-					   RealD coarse_relax_tol=5.0e3,
-					   int largestEvalIdxForReport=-1) 
+					   RealD coarse_relax_tol=5.0e3) 
     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
-      _coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
+      _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);
@@ -186,26 +177,12 @@ public:
 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
 	     <<std::endl;
 
-    if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
-      std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
-      RealD tmp_eval;
-      ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
-    }
-    
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
     return conv;
   }
-
-  //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)  
+  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;
@@ -224,13 +201,13 @@ 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 << " target " << eresid*eresid
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
 	     <<std::endl;
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
     if( (vv<eresid*eresid) ) return 1;
     return 0;
   }
@@ -308,10 +285,6 @@ 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;
@@ -355,8 +328,6 @@ 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);
@@ -405,31 +376,25 @@ 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); //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
+    Chebyshev<FineField>                          Cheby(cheby_op);
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
 
-    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1); 
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
+    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);
@@ -440,14 +405,6 @@ 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

@@ -29,8 +29,6 @@ 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/allocator/MemoryManager.h

@@ -36,6 +36,11 @@ NAMESPACE_BEGIN(Grid);
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
+
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
 // Advise the LatticeAccelerator class
@@ -92,8 +97,9 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
+  static void Audit(std::string s);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -113,6 +119,8 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
  
  private:
 #ifndef GRID_UVM
@@ -170,6 +178,7 @@ private:
 
  public:
   static void Print(void);
+  static void PrintAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -8,9 +8,8 @@ NAMESPACE_BEGIN(Grid);
 static char print_buffer [ MAXLINE ];
 
 #define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-//#define dprintf(...) printf (__VA_ARGS__ ); fflush(stdout);
-#define dprintf(...) 
-
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+//#define dprintf(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -29,6 +28,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -116,8 +117,10 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t) NULL;
     dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
@@ -127,26 +130,36 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove entry
-  // Cannot be locked. If allocated must be in LRU pool.
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-  assert(AccCache.accLock==0);
-  assert(AccCache.cpuLock==0);
+  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
+	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  assert(AccCache.accLock==0); // Cannot evict so logic bomb
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
     dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  uint64_t CpuPtr = AccCache.CpuPtr;
-  EntryErase(CpuPtr);
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -197,6 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -208,6 +222,7 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -218,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back();
+      uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
+    } else {
+      return;
     }
   }
 }
@@ -247,11 +265,12 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld\n",
+    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-		    (uint64_t)bytes);
+	            (uint64_t)bytes,
+		    (uint64_t)AccCache.accLock);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -286,6 +305,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
+    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -298,28 +318,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  // If view is opened on device remove from LRU
+  assert(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
+    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -340,10 +362,12 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
-
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
+  } else {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -380,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  if (!AccCache.AccPtr) {
-     EvictVictims(bytes);
-  }
+  // CPU doesn't need to free space
+  //  if (!AccCache.AccPtr) {
+  //    EvictVictims(bytes);
+  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -436,20 +461,28 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -459,13 +492,13 @@ void  MemoryManager::Print(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -479,6 +512,61 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
+void MemoryManager::Audit(std::string s)
+{
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  uint64_t LockedBytes=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    assert(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    assert(AccCache.LRU_valid==1);
+    assert(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      assert(AccCache.LRU_valid==0);
+      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+		<< "\t cpuLock  " << AccCache.cpuLock
+		<< "\t accLock  " << AccCache.accLock
+		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
+    }
+
+    assert( AccCache.cpuLock== 0 ) ;
+    assert( AccCache.accLock== 0 ) ;
+  }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  assert(LruBytes1==LruBytes2);
+  assert(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  assert(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  assert(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+}
 
 void MemoryManager::PrintState(void* _CpuPtr)
 {
@@ -495,8 +583,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
     if ( AccCache.state==EvictNext) str = std::string("EvictNext");
 
     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
     << "\t" << AccCache.cpuLock
     << "\t" << AccCache.accLock
     << "\t" << AccCache.LRU_valid<<std::endl;

Grid/allocator/MemoryManagerShared.cc

@@ -12,7 +12,10 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
+void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
@@ -21,6 +24,7 @@ void  MemoryManager::PrintState(void* CpuPtr)
 std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
 };
 void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

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;
@@ -98,16 +97,14 @@ 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
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
   static void BroadcastWorld(int root,void* data, int bytes);
-  static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
   // Reduction
@@ -145,16 +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();
@@ -372,41 +370,38 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( dor ) {
-    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+from*32;
-      ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-      assert(ierr==0);
-      list.push_back(rrq);
-      off_node_bytes+=bytes;
-    }
+  if ( (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 (dox) {
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
-      list.push_back(xrq);
-      off_node_bytes+=bytes;
-    } else {
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
-    }
+
+  if ( (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);
-   *  }
-   */
+
+  //  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  //    this->StencilSendToRecvFromComplete(list,dir);
+  //  }
+
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
+  //   std::cout << "Copy Synchronised\n"<<std::endl;
   acceleratorCopySynchronise();
-  StencilBarrier();// Synch shared memory on a single nodes
 
   int nreq=list.size();
 
@@ -443,10 +438,6 @@ int CartesianCommunicator::RankWorld(void){
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
-void CartesianCommunicator::BarrierWorld(void){
-  int ierr = MPI_Barrier(communicator_world);
-  assert(ierr==0);
-}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,

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);
@@ -104,7 +102,6 @@ int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
-void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
@@ -114,18 +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
   ////////////////////////////////////////////////////////////////
@@ -523,7 +520,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   }
   if ( WorldRank == 0 ){
     std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
+	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
   std::cout<< "Setting up IPC"<<std::endl;

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.h

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

Grid/lattice/Lattice_arith.h

@@ -36,7 +36,6 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mult");
   ret.Checkerboard() = lhs.Checkerboard();
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
@@ -54,7 +53,6 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mac");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -72,7 +70,6 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("sub");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -89,7 +86,6 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("add");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -110,7 +106,6 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("mult");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -124,7 +119,6 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("mac");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -139,7 +133,6 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("sub");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -153,7 +146,6 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("add");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -171,7 +163,6 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mult");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -186,7 +177,6 @@ void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mac");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -201,7 +191,6 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("sub");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -215,7 +204,6 @@ void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("add");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -230,7 +218,6 @@ void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class sobj,class vobj> inline
 void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-  GRID_TRACE("axpy");
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
@@ -244,7 +231,6 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
 }
 template<class sobj,class vobj> inline
 void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
-  GRID_TRACE("axpby");
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
@@ -260,13 +246,11 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
-  GRID_TRACE("axpy_norm");
     return axpy_norm_fast(ret,a,x,y);
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
-  GRID_TRACE("axpby_norm");
     return axpby_norm_fast(ret,a,b,x,y);
 }
 

Grid/lattice/Lattice_base.h

@@ -117,7 +117,6 @@ public:
   ////////////////////////////////////////////////////////////////////////////////
   template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -130,7 +129,7 @@ public:
     
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
+    auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -141,7 +140,6 @@ public:
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -154,7 +152,7 @@ public:
 
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
+    auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -165,7 +163,6 @@ public:
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -177,7 +174,7 @@ public:
     this->checkerboard=cb;
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
+    auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -248,7 +245,7 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWrite) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
@@ -291,7 +288,7 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWriteDiscard);
+    auto me =   View(AcceleratorWrite);
     auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
@@ -306,7 +303,7 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
+    auto me =   View(AcceleratorWrite);
     auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));

Grid/lattice/Lattice_crc.h

@@ -1,55 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_crc.h
-
-    Copyright (C) 2021
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
-{
-  auto ff = localNorm2(f);
-  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
-  typedef typename vobj::tensor_reduced normtype;
-  typedef typename normtype::scalar_object scalar;
-  std::vector<scalar> sff;
-  sliceSum(ff,sff,mu);
-  for(int t=0;t<sff.size();t++){
-    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
-  }
-}
-
-template<class vobj> uint32_t crc(Lattice<vobj> & buf)
-{
-  autoView( buf_v , buf, CpuRead);
-  return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
-}
-
-#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
-
-NAMESPACE_END(Grid);
-
-

Grid/lattice/Lattice_peekpoke.h

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

Grid/lattice/Lattice_reduction.h

@@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
+#if defined(GRID_SYCL)
+#include <Grid/lattice/Lattice_reduction_sycl.h>
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -91,7 +94,10 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
   for(int i=0;i<nthread;i++){
     ssum = ssum+sumarray[i];
   } 
-  return ssum;
+  
+  typedef typename vobj::scalar_object ssobj;
+  ssobj ret = ssum;
+  return ret;
 }
 /*
 Threaded max, don't use for now
@@ -124,7 +130,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -133,7 +139,7 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -142,7 +148,7 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu_large(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -152,13 +158,13 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(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(&arg_v[0],osites);
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  typename vobj::scalar_object ssum;
+  autoView( arg_v, arg, AcceleratorRead);
+  ssum= sum_gpu(&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);
@@ -168,7 +174,7 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 template<class vobj>
 inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
   auto ssum= sum_gpu_large(&arg_v[0],osites);
@@ -229,41 +235,23 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   const uint64_t sites = grid->oSites();
   
   // Might make all code paths go this way.
-#if 0
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-    
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
-
+    // This code could read coalesce
     // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
+    accelerator_for( ss, sites, 1,{
+	auto x_l = left_v[ss];
+	auto y_l = right_v[ss];
+	inner_tmp_v[ss]=innerProductD(x_l,y_l);
     });
   }
-#else
-  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-    
-  {
-    autoView( left_v , left, AcceleratorRead);
-    autoView( right_v,right, AcceleratorRead);
 
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
-    });
-  }
-#endif
   // This is in single precision and fails some tests
-  auto anrm = sumD(inner_tmp_v,sites);  
+  auto anrm = sum(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
@@ -297,7 +285,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   conformable(x,y);
 
   typedef typename vobj::scalar_type scalar_type;
-  //  typedef typename vobj::vector_typeD vector_type;
+  typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
   GridBase *grid = x.Grid();
@@ -309,29 +297,17 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-#if 0
+
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
 
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
+  accelerator_for( ss, sites, 1,{
+      auto tmp = a*x_v[ss]+b*y_v[ss];
+      inner_tmp_v[ss]=innerProductD(tmp,tmp);
+      z_v[ss]=tmp;
   });
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#else
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -485,14 +461,6 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
 }
-template<class vobj> inline
-std::vector<typename vobj::scalar_object> 
-sliceSum(const Lattice<vobj> &Data,int orthogdim)
-{
-  std::vector<typename vobj::scalar_object> result;
-  sliceSum(Data,result,orthogdim);
-  return result;
-}
 
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 

Grid/lattice/Lattice_reduction_sycl.h

@@ -0,0 +1,125 @@
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Possibly promote to double and sum
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_objectD sobjD;
+  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
+  sobj identity; zeroit(identity);
+  sobj ret ; 
+
+  Integer nsimd= vobj::Nsimd();
+  
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{osites},
+		      Reduction,
+		      [=] (cl::sycl::id<1> item, auto &sum) {
+      auto osite   = item[0];
+      sum +=Reduce(lat[osite]);
+     });
+   });
+  theGridAccelerator->wait();
+  ret = mysum[0];
+  free(mysum,*theGridAccelerator);
+  sobjD dret; convertType(dret,ret);
+  return dret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_tensor(lat,osites);
+}
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Return as same precision as input performing reduction in double precision though
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu(lat,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);
+
+/*
+template<class Double> Double svm_reduce(Double *vec,uint64_t L)
+{
+  Double sumResult; zeroit(sumResult);
+  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
+  Double identity;  zeroit(identity);
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{L},
+		      Reduction,
+		      [=] (cl::sycl::id<1> index, auto &sum) {
+	 sum +=vec[index];
+     });
+   });
+  theGridAccelerator->wait();
+  Double ret = d_sum[0];
+  free(d_sum,*theGridAccelerator);
+  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
+  return ret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_type  scalar;
+
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobjD;
+
+  sobjD ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int nsimd = vobj::Nsimd();
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<scalar> buffer(osites*nsimd);
+  scalar *buf = &buffer[0];
+  vector *dat = (vector *)lat;
+
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,nsimd,{
+	int lane = acceleratorSIMTlane(nsimd);
+	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
+    });
+    //Precision change at this point is to late to gain precision
+    ret_p[w] = svm_reduce(buf,nsimd*osites);
+  }
+  return ret;
+}
+*/

Grid/lattice/Lattice_rng.h

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

Grid/lattice/Lattice_transfer.h

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

Grid/log/Log.cc

@@ -66,6 +66,7 @@ GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
 GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
+GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
 GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
@@ -77,7 +78,8 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
-  GridLogMemory.Active(0); // at least the messages should be always on
+  GridLogMemory.Active(0); 
+  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
@@ -87,6 +89,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
+    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
     if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
@@ -94,8 +97,8 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
     if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
-    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
-    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
+    if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -186,6 +186,7 @@ extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
 extern GridLogger GridLogHMC;
 extern GridLogger GridLogMemory;
+extern GridLogger GridLogTracing;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/NerscIO.h

@@ -42,10 +42,8 @@ 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; }
+  typedef Lattice<vLorentzColourMatrixD> GaugeField;
 
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
@@ -205,7 +203,7 @@ public:
       std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
       exit(0);
     }
-    if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
+    assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
     assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
     assert(nersc_csum == header.checksum );
       

Grid/perfmon/PerfCount.h

@@ -30,6 +30,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
+
+#ifndef __SSC_START
+#define __SSC_START
+#define __SSC_STOP
+#endif
+
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>

Grid/perfmon/Tracing.h

@@ -1,4 +1,7 @@
 #pragma once
+
+NAMESPACE_BEGIN(Grid);
+
 #ifdef GRID_TRACING_NVTX
 #include <nvToolsExt.h>
 class GridTracer {
@@ -64,3 +67,4 @@ inline void traceStop(int ID) {  }
 #else
 #define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
 #endif
+NAMESPACE_END(Grid);

Grid/qcd/QCD.h

@@ -63,7 +63,6 @@ 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;
@@ -88,8 +87,6 @@ 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 
@@ -113,10 +110,8 @@ 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;
@@ -181,16 +176,6 @@ 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;
@@ -235,16 +220,6 @@ 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

@@ -40,43 +40,9 @@ class Action
 
 public:
   bool is_smeared = false;
-  RealD deriv_norm_sum;
-  RealD deriv_max_sum;
-  RealD Fdt_norm_sum;
-  RealD Fdt_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;
-    Fdt_max_sum =  Fdt_norm_sum = 0.0;
-  }
-  void  deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) {
-    deriv_max_sum+=max;
-    deriv_norm_sum+=nrm;
-    Fdt_max_sum+=Fdt_max;
-    Fdt_norm_sum+=Fdt_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 Fdt_max_average(void)         { return Fdt_max_sum/deriv_num; };
-  RealD Fdt_norm_average(void)        { return Fdt_norm_sum/deriv_num; };
-  RealD deriv_timer(void)        { return deriv_us; };
-  RealD S_timer(void)            { return S_us; };
-  RealD refresh_timer(void)      { return refresh_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
-  virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
   virtual std::string action_name()    = 0;                             // return the action name
   virtual std::string LogParameters()  = 0;                             // prints action parameters

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

@@ -37,33 +37,24 @@ NAMESPACE_BEGIN(Grid);
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
   Coordinate twists;
-                     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  GparityWilsonImplParams() : twists(Nd, 0) { dirichlet.resize(0); };
+  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(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(0);
   }
 };
 
 struct StaggeredImplParams {
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  StaggeredImplParams()
-  {
-    dirichlet.resize(0);
-  };
+  StaggeredImplParams()  {};
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -72,11 +63,9 @@ struct StaggeredImplParams {
 				    RealD, hi, 
 				    int,   MaxIter, 
 				    RealD, tolerance, 
-				    RealD, mdtolerance, 
 				    int,   degree, 
 				    int,   precision,
-				    int,   BoundsCheckFreq,
-				    RealD, BoundsCheckTol);
+				    int,   BoundsCheckFreq);
     
   // MaxIter and tolerance, vectors??
     
@@ -87,62 +76,16 @@ struct StaggeredImplParams {
 				RealD tol      = 1.0e-8, 
                            	int _degree    = 10,
 				int _precision = 64,
-				int _BoundsCheckFreq=20,
-				RealD mdtol    = 1.0e-6,
-				double _BoundsCheckTol=1e-6)
+				int _BoundsCheckFreq=20)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
-        mdtolerance(mdtol),
 	degree(_degree),
         precision(_precision),
-        BoundsCheckFreq(_BoundsCheckFreq),
-        BoundsCheckTol(_BoundsCheckTol){};
+        BoundsCheckFreq(_BoundsCheckFreq){};
   };
   
-  /*Action parameters for the generalized rational action
-    The approximation is for (M^dag M)^{1/inv_pow}
-    where inv_pow is the denominator of the fractional power.
-    Default inv_pow=2 for square root, making this equivalent to 
-    the OneFlavourRational action
-  */
-    struct RationalActionParams : Serializable {
-    GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams, 
-				    int, inv_pow, 
-				    RealD, lo, //low eigenvalue bound of rational approx
-				    RealD, hi, //high eigenvalue bound of rational approx
-				    int,   MaxIter,  //maximum iterations in msCG
-				    RealD, action_tolerance,  //msCG tolerance in action evaluation
-				    int,   action_degree, //rational approx tolerance in action evaluation
-				    RealD, md_tolerance,  //msCG tolerance in MD integration
-				    int,   md_degree, //rational approx tolerance in MD integration
-				    int,   precision, //precision of floating point arithmetic
-				    int,   BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
-  // constructor 
-  RationalActionParams(int _inv_pow = 2,
-		       RealD _lo      = 0.0, 
-		       RealD _hi      = 1.0, 
-		       int _maxit     = 1000,
-		       RealD _action_tolerance      = 1.0e-8, 
-		       int _action_degree    = 10,
-		       RealD _md_tolerance      = 1.0e-8, 
-		       int _md_degree    = 10,
-		       int _precision = 64,
-		       int _BoundsCheckFreq=20)
-    : inv_pow(_inv_pow), 
-      lo(_lo),
-      hi(_hi),
-      MaxIter(_maxit),
-      action_tolerance(_action_tolerance),
-      action_degree(_action_degree),
-      md_tolerance(_md_tolerance),
-      md_degree(_md_degree),
-      precision(_precision),
-      BoundsCheckFreq(_BoundsCheckFreq){};
-  };
-
-
 NAMESPACE_END(Grid);
 
 #endif

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -71,7 +71,6 @@ public:
   RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
   RealD MassPlus(void) { return mass_plus; };
   RealD MassMinus(void) { return mass_minus; };
-
   void  SetMass(RealD _mass) { 
     mass_plus=mass_minus=_mass; 
     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
@@ -183,6 +182,16 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  RealD _mass,RealD _M5,const ImplParams &p= ImplParams());
 
+  void CayleyReport(void);
+  void CayleyZeroCounters(void);
+
+  double M5Dflops;
+  double M5Dcalls;
+  double M5Dtime;
+
+  double MooeeInvFlops;
+  double MooeeInvCalls;
+  double MooeeInvTime;
 
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);

Grid/qcd/action/fermion/CloverHelpers.h

@@ -204,15 +204,18 @@ public:
   typedef WilsonCloverHelpers<Impl> Helpers;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
     Clover += diag_mass;
   }
 
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
-                          CloverTriangleField& Triangle,
-                          RealD csw_t, RealD diag_mass) {
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
 
-    // Do nothing
+    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
   }
 
   // TODO: implement Cmunu for better performances with compact layout, but don't do it
@@ -237,9 +240,17 @@ public:
   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
+  // 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) {
@@ -254,175 +265,62 @@ public:
     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 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 ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
 
-  	  typedef iMatrix<ComplexD,6> mat;
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
 
-  	  RealD qn[6];
-  	  RealD qnold[6];
-  	  RealD p[5];
-  	  RealD trA2, trA3, trA4;
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
 
-  	  mat A2, A3, A4, A5;
-  	  A2 = alpha * alpha * arg * arg;
-  	  A3 = alpha * arg * A2;
-  	  A4 = A2 * A2;
-  	  A5 = A2 * A3;
+    Clover *= (1.0/diag_mass);
 
-  	  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
+    // 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++){
+    for (int i=1; i<=NMAX; i++)
       cn[i] = cn[i-1] / RealD(i);
-    }
 
-      // Taken over from Daniel's implementation
-      conformable(Diagonal, Triangle);
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
 
-      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;
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
 
-      autoView(diagonal_v,  Diagonal,  CpuRead);
-      autoView(triangle_v,  Triangle,  CpuRead);
-      autoView(diagonalExp_v, Diagonal, CpuWrite);
-      autoView(triangleExp_v, Triangle, CpuWrite);
+    Clover = ExpClover * diag_mass;
 
-      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
 
-    	  mat srcCloverOpUL(0.0); // upper left block
-    	  mat srcCloverOpLR(0.0); // lower right block
-    	  mat ExpCloverOp;
+    CloverInv = ExpClover * (1.0/diag_mass);
 
-        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 void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
+
+    if (fixedBoundaries)
+    {
+      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+    }
+    else
+    {
+      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
+    }
+  }
 
   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
     assert(0);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -225,7 +225,7 @@ public:
   RealD csw_t;
   RealD cF;
 
-  bool open_boundaries;
+  bool fixedBoundaries;
 
   CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
   CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;

Grid/qcd/action/fermion/FermionOperator.h

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

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -30,18 +30,6 @@ 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:
@@ -125,7 +113,7 @@ public:
     || ((distance== 1)&&(icoor[direction]==1))
     || ((distance==-1)&&(icoor[direction]==0));
 
-    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
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
 
     //Apply the links
     int f_upper = permute_lane ? 1 : 0;
@@ -151,10 +139,10 @@ public:
     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code
     assert((sl == 1) || (sl == 2));
 
-    //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 ( 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);
@@ -209,19 +197,6 @@ 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);
@@ -232,19 +207,14 @@ public:
     GaugeLinkField Uconj(GaugeGrid);
    
     Lattice<iScalar<vInteger> > coor(GaugeGrid);
-
-    //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);
-      }
+        
+    for(int mu=0;mu<Nd;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] ) { 
@@ -259,7 +229,7 @@ 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
@@ -290,38 +260,6 @@ 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) {
@@ -360,48 +298,28 @@ 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];
-    
-    {
-      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);
-  	});
-    }
-  }
-
-
   
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
 
+    int Ls = Btilde.Grid()->_fdimensions[0];
+        
+    GaugeLinkField tmp(mat.Grid());
+    tmp = Zero();
+    {
+      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));
+	  }
+	});
+    }
+    PokeIndex<LorentzIndex>(mat, tmp, mu);
+    return;
+  }
   
 };
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -47,6 +47,18 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -52,6 +52,18 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -47,6 +47,18 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -117,19 +117,19 @@ public:
     typedef decltype(coalescedRead(*in))    sobj;
     typedef decltype(coalescedRead(*out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
-    const vobj *vp1 = &in[m];
-    const vobj *vp = (lane&mask) ? vp1:vp0;
-    auto sa = coalescedRead(*vp,j0);
-    auto sb = coalescedRead(*vp,j1);
+    const vobj *vp0 = &in[k];  // out0[j] = merge low bit of type from in[k] and in[m] 
+    const vobj *vp1 = &in[m];  // out1[j] = merge hi  bit of type from in[k] and in[m]
+    const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
+    auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
+    auto sb = coalescedRead(*vp,j1); // lane to read for out 1
     hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);
-    projector::Proj(psb,sb,mu,dag);
+    projector::Proj(psa,sa,mu,dag);  // spin project the result0
+    projector::Proj(psb,sb,mu,dag);  // spin project the result1
     coalescedWrite(out0[j],psa);
     coalescedWrite(out1[j],psb);
 #else
@@ -294,7 +294,11 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  //  Vector<int> surface_list;
+  void ZeroCountersi(void)  {  }
+  void Reporti(int calls)  {  }
+
+  std::vector<int> surface_list;
+
   WilsonStencil(GridBase *grid,
 		int npoints,
 		int checkerboard,
@@ -302,11 +306,11 @@ public:
 		const std::vector<int> &distances,Parameters p)  
     : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
   { 
-    //    surface_list.resize(0);
+    ZeroCountersi();
+    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
 
-  /*
   void BuildSurfaceList(int Ls,int vol4){
 
     // find same node for SHM
@@ -327,8 +331,7 @@ public:
       }
     }
   }
-  */
-  
+
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
   {
@@ -395,6 +398,7 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
+    accelerator_barrier();
   }
 
 };

Grid/qcd/action/fermion/WilsonFermion.h

@@ -74,6 +74,20 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
+
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,8 +75,19 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
-  int Dirichlet;
-  Coordinate Block; 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
@@ -162,10 +173,7 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-
-  virtual void DirichletBlock(const Coordinate & block)
-  {
-  }
+    
   // Constructors
   /*
     WilsonFermion5D(int simd, 

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

@@ -152,6 +152,58 @@ void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi
   }
 }
 
+template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
+{
+  this->Report();
+  Coordinate latt = GridDefaultLatt();          
+  RealD volume = this->Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP     = this->_FourDimGrid->_Nprocessors;
+  if ( M5Dcalls > 0 ) {
+    std::cout << GridLogMessage << "#### M5D calls report " << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D Number of M5D Calls     : " << M5Dcalls   << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls       : " << M5Dtime / M5Dcalls << " us" << std::endl;
+
+    // Flops = 10.0*(Nc*Ns) *Ls*vol
+    RealD mflops = 10.0*(Nc*Ns)*volume*M5Dcalls/M5Dtime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+
+    // Bytes = sizeof(Real) * (Nc*Ns*Nreim) * Ls * vol * (read+write) (/2 for red black counting)
+    // read = 2 ( psi[ss+s+1] and psi[ss+s-1] count as 1 )
+    // write = 1
+    RealD Gbytes = sizeof(Real) * (Nc*Ns*2) * volume * 3 /2. * 1.e-9;
+    std::cout << GridLogMessage << "Average bandwidth (GB/s)                 : " << Gbytes/M5Dtime*M5Dcalls*1.e6 << std::endl;
+  }
+
+  if ( MooeeInvCalls > 0 ) {
+
+    std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls     : " << MooeeInvCalls   << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls            : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
+#ifdef GRID_CUDA
+    RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+#else
+    // Flops = MADD * Ls *Ls *4dvol * spin/colour/complex
+    RealD mflops = 2.0*24*this->Ls*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+#endif
+  }
+
+}
+template<class Impl> void CayleyFermion5D<Impl>::CayleyZeroCounters(void)
+{
+  this->ZeroCounters();
+  M5Dflops=0;
+  M5Dcalls=0;
+  M5Dtime=0;
+  MooeeInvFlops=0;
+  MooeeInvCalls=0;
+  MooeeInvTime=0;
+}
+
 template<class Impl>  
 void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
@@ -594,6 +646,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
   assert(mass_plus == mass_minus);
   RealD mass = mass_plus;
   
+#if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -712,7 +765,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
     else          q_out +=     C;
     
   }
-
+#endif
 }
 
 template <class Impl>
@@ -779,6 +832,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
+#if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
@@ -898,6 +952,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     InsertSlice(L_Q, q_out, s , 0);
   }
+#endif
 }
 #undef Pp
 #undef Pm

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

@@ -63,18 +63,23 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
 
   // 10 = 3 complex mult + 2 complex add
   // Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
-  uint64_t nloop = grid->oSites();
+  M5Dcalls++;
+  M5Dtime-=usecond();
+
+  uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t s = sss%Ls;
-    uint64_t ss= sss-s;
+    uint64_t ss= sss*Ls;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1, tmp2;
-    uint64_t idx_u = ss+((s+1)%Ls);
-    uint64_t idx_l = ss+((s+Ls-1)%Ls);
-    spProj5m(tmp1,psi(idx_u));
-    spProj5p(tmp2,psi(idx_l));
-    coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+    for(int s=0;s<Ls;s++){
+      uint64_t idx_u = ss+((s+1)%Ls);
+      uint64_t idx_l = ss+((s+Ls-1)%Ls);
+      spProj5m(tmp1,psi(idx_u));
+      spProj5p(tmp2,psi(idx_l));
+      coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+    }
   });
+  M5Dtime+=usecond();
 }
 
 template<class Impl>  
@@ -100,18 +105,23 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   int Ls=this->Ls;
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
-  uint64_t nloop = grid->oSites();
+  M5Dcalls++;
+  M5Dtime-=usecond();
+
+  uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t s = sss%Ls;
-    uint64_t ss= sss-s;
+    uint64_t ss=sss*Ls;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1,tmp2;
-    uint64_t idx_u = ss+((s+1)%Ls);
-    uint64_t idx_l = ss+((s+Ls-1)%Ls);
-    spProj5p(tmp1,psi(idx_u));
-    spProj5m(tmp2,psi(idx_l));
-    coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+    for(int s=0;s<Ls;s++){
+      uint64_t idx_u = ss+((s+1)%Ls);
+      uint64_t idx_l = ss+((s+Ls-1)%Ls);
+      spProj5p(tmp1,psi(idx_u));
+      spProj5m(tmp2,psi(idx_l));
+      coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+    }
   });
+  M5Dtime+=usecond();
 }
 
 template<class Impl>
@@ -132,6 +142,8 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
   auto pleem = & leem[0];
   auto pueem = & ueem[0];
 
+  MooeeInvCalls++;
+  MooeeInvTime-=usecond();
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -168,6 +180,8 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
       coalescedWrite(chi[ss+s],res);
     }
   });
+
+  MooeeInvTime+=usecond();
   
 }
 
@@ -190,6 +204,10 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
 
   assert(psi.Checkerboard() == psi.Checkerboard());
 
+  MooeeInvCalls++;
+  MooeeInvTime-=usecond();
+
+
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -226,6 +244,7 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
       coalescedWrite(chi[ss+s],res);
     }
   });
+  MooeeInvTime+=usecond();
 
 }
 

Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h

@@ -94,6 +94,10 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
       d_p[ss] = diag[s];
     }}
 
+
+  M5Dcalls++;
+  M5Dtime-=usecond();
+
   assert(Nc==3);
 
   thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
@@ -194,6 +198,7 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
     }
 #endif
   });
+  M5Dtime+=usecond();
 }
 
 template<class Impl>  
@@ -237,6 +242,8 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
       d_p[ss] = diag[s];
     }}
 
+  M5Dcalls++;
+  M5Dtime-=usecond();
   thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
 #if 0
     alignas(64) SiteHalfSpinor hp;
@@ -332,6 +339,7 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
     }
 #endif
   });
+  M5Dtime+=usecond();
 }
 
 
@@ -805,6 +813,9 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
   }
   assert(_Matp->size()==Ls*LLs);
 
+  MooeeInvCalls++;
+  MooeeInvTime-=usecond();
+
   if ( switcheroo<Coeff_t>::iscomplex() ) {
     thread_loop( (auto site=0;site<vol;site++),{
       MooeeInternalZAsm(psi,chi,LLs,site,*_Matp,*_Matm);
@@ -814,7 +825,7 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
       MooeeInternalAsm(psi,chi,LLs,site,*_Matp,*_Matm);
     });
   }
-
+  MooeeInvTime+=usecond();
 }
 
 NAMESPACE_END(Grid);

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

@@ -48,7 +48,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , csw_r(_csw_r)
   , csw_t(_csw_t)
   , cF(_cF)
-  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
   , Diagonal(&Fgrid),        Triangle(&Fgrid)
   , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
   , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
@@ -67,7 +67,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries) {
+  if (fixedBoundaries) {
     this->BoundaryMaskEven.Checkerboard() = Even;
     this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
@@ -77,31 +77,31 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->open_boundaries) ApplyBoundaryMask(out);
+  if(this->fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
-  if(this->open_boundaries) {
+  if(this->fixedBoundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
@@ -112,7 +112,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in,
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -121,19 +121,19 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& i
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -147,7 +147,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -166,7 +166,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -186,7 +186,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!open_boundaries); // TODO check for changes required for open bc
+  assert(!fixedBoundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -305,6 +305,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   GridBase* grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
   CloverField TmpOriginal(grid);
+  CloverField TmpInverse(grid);
 
   // Compute the field strength terms mu>nu
   double t2 = usecond();
@@ -324,24 +325,27 @@ 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);
-  
-  // Convert the data layout of the clover term
+
+  // Instantiate the clover term
+  // - In case of the standard clover the mass term is added
+  // - In case of the exponential clover the clover term is exponentiated
   double t4 = usecond();
+  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
+
+  // Convert the data layout of the clover term
+  double t5 = 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
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
   double t6 = usecond();
-  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+  if(fixedBoundaries) 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)
+  // Invert the Clover term
+  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
+  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
+  // TODO: For now this inversion is explictly done on the CPU
   double t7 = usecond();
-  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
 
   // Fill the remaining clover fields
   double t8 = usecond();
@@ -362,10 +366,10 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   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 << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
   std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
   std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
 }

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

@@ -54,6 +54,8 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   auto pupper = &upper[0];
   auto plower = &lower[0];
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
   
   auto nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -69,6 +71,7 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -88,6 +91,8 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
   auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
 
   auto nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -103,6 +108,7 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -121,6 +127,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
   auto pleem = & this->leem[0];
   auto pueem = & this->ueem[0];
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
   uint64_t nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -156,6 +164,7 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
       coalescedWrite(chi[ss+s],res);
     }
   });
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -176,6 +185,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
 
   assert(psi.Checkerboard() == psi.Checkerboard());
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
   auto nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -212,6 +223,7 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
     }
   });
 
+  this->MooeeInvTime += usecond();
 }
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h

@@ -298,33 +298,45 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   int LLs = in.Grid()->_rdimensions[0];
   int len =  U.Grid()->oSites();
 
+  DhopFaceTime-=usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
 
+  DhopCommTime -=usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
   st.CommunicateBegin(requests);
 
   //  st.HaloExchangeOptGather(in,compressor); // Wilson compressor
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+  DhopFaceTime+=usecond();
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // Remove explicit thread mapping introduced for OPA reasons.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime-=usecond();
   {
     int interior=1;
     int exterior=0;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime+=usecond();
 
+  DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
   st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
+  DhopComputeTime2-=usecond();
   {
     int interior=0;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime2+=usecond();
 }
 
 template<class Impl>
@@ -335,14 +347,22 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   Compressor compressor;
   int LLs = in.Grid()->_rdimensions[0];
 
+ //double t1=usecond();
+  DhopTotalTime -= usecond();
+  DhopCommTime -= usecond();
   st.HaloExchange(in,compressor);
+  DhopCommTime += usecond();
   
+  DhopComputeTime -= usecond();
   // Dhop takes the 4d grid from U, and makes a 5d index for fermion
   {
     int interior=1;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
+
 }
 /*CHANGE END*/
 
@@ -351,6 +371,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=1;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -362,6 +383,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=1;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -373,6 +395,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(),FermionGrid()); // verifies full grid
   conformable(in.Grid(),out.Grid());
 
@@ -381,6 +404,58 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
   DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
 }
 
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::Report(void) 
+{
+  Coordinate latt = GridDefaultLatt();          
+  RealD volume = Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _FourDimGrid->_Nprocessors;
+  RealD NN = _FourDimGrid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _FourDimGrid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime    = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
 /////////////////////////////////////////////////////////////////////////
 // Implement the general interface. Here we use SAME mass on all slices
 /////////////////////////////////////////////////////////////////////////

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

@@ -334,6 +334,7 @@ void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionF
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -345,6 +346,7 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=1;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -357,6 +359,7 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=1;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -415,33 +418,47 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   Compressor compressor; 
   int len =  U.Grid()->oSites();
 
+  DhopTotalTime   -= usecond();
+
+  DhopFaceTime    -= usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime    += usecond();
 
+  DhopCommTime -=usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
   st.CommunicateBegin(requests);
 
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // Removed explicit thread comms
   //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime    -= usecond();
   {
     int interior=1;
     int exterior=0;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime    += usecond();
 
   st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
   // First to enter, last to leave timing
+  DhopFaceTime    -= usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime    -= usecond();
 
+  DhopComputeTime2    -= usecond();
   {
     int interior=0;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime2    += usecond();
 }
 
 
@@ -454,16 +471,78 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
 {
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
+  DhopTotalTime   -= usecond();
+
+  DhopCommTime    -= usecond();
   Compressor compressor;
   st.HaloExchange(in, compressor);
+  DhopCommTime    += usecond();
 
+  DhopComputeTime -= usecond();
   {
     int interior=1;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
 };
 
+  ////////////////////////////////////////////////////////////////
+  // Reporting
+  ////////////////////////////////////////////////////////////////
+template<class Impl>
+void ImprovedStaggeredFermion<Impl>::Report(void) 
+{
+  Coordinate latt = _grid->GlobalDimensions();
+  RealD volume = 1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _grid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void ImprovedStaggeredFermion<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime   = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
+
 //////////////////////////////////////////////////////// 
 // Conserved current - not yet implemented.
 ////////////////////////////////////////////////////////

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

@@ -55,6 +55,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
   auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -70,6 +73,7 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -95,6 +99,9 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
   auto pshift_coeffs = &shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -115,6 +122,7 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -135,6 +143,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
   auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(), {
     uint64_t ss = sss*Ls;
@@ -150,6 +161,8 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
       coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
     }
   });
+
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -173,6 +186,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
   auto pshift_coeffs = &shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   auto pm = this->pm;
 
   int nloop = grid->oSites()/Ls;
@@ -201,6 +217,7 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -220,6 +237,9 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
 
   if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -257,6 +277,7 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
     }
   });
    
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -276,6 +297,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   auto pueem= & this->ueem[0];
   auto pMooeeInv_shift_lc   = &MooeeInv_shift_lc[0];
   auto pMooeeInv_shift_norm = &MooeeInv_shift_norm[0];
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -320,6 +343,7 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
       }
   });
 
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -339,6 +363,9 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   auto pleem= & this->leem[0];
   auto pueem= & this->ueem[0];
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -375,6 +402,7 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
       coalescedWrite(chi[ss+s],res);
     }
   });
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -395,6 +423,9 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
   auto pMooeeInvDag_shift_lc   = &MooeeInvDag_shift_lc[0];
   auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
       uint64_t ss=sss*Ls;
@@ -438,6 +469,7 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
       }
   });
 
+  this->MooeeInvTime += usecond();
 }
 
 NAMESPACE_END(Grid);

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

@@ -263,6 +263,7 @@ void NaiveStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionFiel
 template <class Impl>
 void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -274,6 +275,7 @@ void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out
 template <class Impl>
 void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=1;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -286,6 +288,7 @@ void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &o
 template <class Impl>
 void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=1;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -342,33 +345,47 @@ void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, L
   Compressor compressor; 
   int len =  U.Grid()->oSites();
 
+  DhopTotalTime   -= usecond();
+
+  DhopFaceTime    -= usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime    += usecond();
 
+  DhopCommTime -=usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
   st.CommunicateBegin(requests);
 
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // Removed explicit thread comms
   //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime    -= usecond();
   {
     int interior=1;
     int exterior=0;
     Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
+  DhopComputeTime    += usecond();
 
   st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
   // First to enter, last to leave timing
+  DhopFaceTime    -= usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime    -= usecond();
 
+  DhopComputeTime2    -= usecond();
   {
     int interior=0;
     int exterior=1;
     Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
+  DhopComputeTime2    += usecond();
 }
 
 template <class Impl>
@@ -379,16 +396,78 @@ void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Lebes
 {
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
+  DhopTotalTime   -= usecond();
+
+  DhopCommTime    -= usecond();
   Compressor compressor;
   st.HaloExchange(in, compressor);
+  DhopCommTime    += usecond();
 
+  DhopComputeTime -= usecond();
   {
     int interior=1;
     int exterior=1;
     Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
 };
 
+  ////////////////////////////////////////////////////////////////
+  // Reporting
+  ////////////////////////////////////////////////////////////////
+template<class Impl>
+void NaiveStaggeredFermion<Impl>::Report(void) 
+{
+  Coordinate latt = _grid->GlobalDimensions();
+  RealD volume = 1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "NaiveStaggeredFermion Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _grid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "NaiveStaggeredFermion Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "NaiveStaggeredFermion StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "NaiveStaggeredFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void NaiveStaggeredFermion<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime   = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
+
 //////////////////////////////////////////////////////// 
 // Conserved current - not yet implemented.
 ////////////////////////////////////////////////////////

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);
@@ -92,17 +91,6 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
     assert(FourDimRedBlackGrid._simd_layout[d]  ==FourDimGrid._simd_layout[d]);
   }
 
-  if ( p.dirichlet.size() == Nd+1) {
-    Coordinate block = p.dirichlet;
-    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
-      Dirichlet = 1;
-      Block = block;
-    }
-  } else {
-    Coordinate block(Nd+1,0);
-    Block = block;
-  }
-
   if (Impl::LsVectorised) { 
 
     int nsimd = Simd::Nsimd();
@@ -141,20 +129,95 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
    //                       <<" " << StencilEven.surface_list.size()<<std::endl;
 
 }
+     
+template<class Impl>
+void WilsonFermion5D<Impl>::Report(void)
+{
+  RealD NP     = _FourDimGrid->_Nprocessors;
+  RealD NN     = _FourDimGrid->NodeCount();
+  RealD volume = Ls;  
+  Coordinate latt = _FourDimGrid->GlobalDimensions();
+  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Number of DhopEO Calls   : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D TotalTime   /Calls        : " << DhopTotalTime   / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D CommTime    /Calls        : " << DhopCommTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D FaceTime    /Calls        : " << DhopFaceTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime1/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime2/Calls        : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
+
+    // Average the compute time
+    _FourDimGrid->GlobalSum(DhopComputeTime);
+    DhopComputeTime/=NP;
+    RealD mflops = 1344*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+
+    RealD Fullmflops = 1344*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+    
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NP << std::endl;
+
+    RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NP << std::endl;  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion5D Stencil"    <<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion5D StencilOdd" <<std::endl;  StencilOdd.Report();
+  }
+  if ( DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion5D Stencil     Reporti()"    <<std::endl;  Stencil.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion5D StencilEven Reporti()"<<std::endl;  StencilEven.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion5D StencilOdd  Reporti()" <<std::endl;  StencilOdd.Reporti(DhopCalls);
+  }
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopComputeTime2= 0;
+  DhopFaceTime    = 0;
+  DhopTotalTime   = 0;
+
+  DerivCalls       = 0;
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+  Stencil.ZeroCountersi();
+  StencilEven.ZeroCountersi();
+  StencilOdd.ZeroCountersi();
+}
+
 
 template<class Impl>
 void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
-  if ( Dirichlet ) {
-    std::cout << GridLogDslash << " Dirichlet BCs 5d " <<Block<<std::endl;
-    Coordinate GaugeBlock(Nd);
-    for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
-    std::cout << GridLogDslash << " 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);
@@ -196,6 +259,7 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
 					  const FermionField &B,
 					  int dag)
 {
+  DerivCalls++;
   assert((dag==DaggerNo) ||(dag==DaggerYes));
 
   conformable(st.Grid(),A.Grid());
@@ -206,12 +270,15 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
   FermionField Btilde(B.Grid());
   FermionField Atilde(B.Grid());
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B,compressor);
+  DerivCommTime+=usecond();
 
   Atilde=A;
   int LLs = B.Grid()->_rdimensions[0];
 
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma if dag
@@ -223,6 +290,8 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
     // Call the single hop
     ////////////////////////
 
+    DerivDhopComputeTime -= usecond();
+
     int Usites = U.Grid()->oSites();
 
     Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, Usites, B, Btilde, mu,gamma);
@@ -230,8 +299,10 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
     ////////////////////////////
     // spin trace outer product
     ////////////////////////////
+    DerivDhopComputeTime += usecond();
     Impl::InsertForce5D(mat, Btilde, Atilde, mu);
   }
+  DerivComputeTime += usecond();
 }
 
 template<class Impl>
@@ -289,10 +360,12 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
                                          DoubledGaugeField & U,
                                          const FermionField &in, FermionField &out,int dag)
 {
+  DhopTotalTime-=usecond();
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else 
     DhopInternalSerialComms(st,lo,U,in,out,dag);
+  DhopTotalTime+=usecond();
 }
 
 
@@ -301,7 +374,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
 							DoubledGaugeField & U,
 							const FermionField &in, FermionField &out,int dag)
 {
-  GRID_TRACE("DhopInternalOverlappedComms");
   Compressor compressor(dag);
 
   int LLs = in.Grid()->_rdimensions[0];
@@ -310,58 +382,53 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   /////////////////////////////
   // Start comms  // Gather intranode and extra node differentiated??
   /////////////////////////////
-  {
-    GRID_TRACE("Gather");
-    st.HaloExchangeOptGather(in,compressor);
-    accelerator_barrier();
-  }
-  
+  DhopFaceTime-=usecond();
+  st.HaloExchangeOptGather(in,compressor);
+  DhopFaceTime+=usecond();
+
+  DhopCommTime -=usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
-  auto id=traceStart("Communicate overlapped");
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
-  {
-    GRID_TRACE("MergeSHM");
-    st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
-  }
+  DhopFaceTime-=usecond();
+  st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+  DhopFaceTime+=usecond();
       
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
+  DhopComputeTime-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagInterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   } else {
-    GRID_TRACE("DhopInterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   }
+  DhopComputeTime+=usecond();
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
-  traceStop(id);
+  DhopCommTime   +=usecond();
 
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
-  {
-    GRID_TRACE("Merge");
-    st.CommsMerge(compressor);
-  }
-  
+  DhopFaceTime-=usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
+  DhopComputeTime2-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagExterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   } else {
-    GRID_TRACE("DhopExterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   }
+  DhopComputeTime2+=usecond();
 }
 
 
@@ -371,30 +438,29 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOr
 						    const FermionField &in, 
 						    FermionField &out,int dag)
 {
-  GRID_TRACE("DhopInternalSerialComms");
   Compressor compressor(dag);
 
   int LLs = in.Grid()->_rdimensions[0];
-
-  {
-    GRID_TRACE("HaloExchange");
-    st.HaloExchangeOpt(in,compressor);
-  }
   
+  DhopCommTime-=usecond();
+  st.HaloExchangeOpt(in,compressor);
+  DhopCommTime+=usecond();
+  
+  DhopComputeTime-=usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDag");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   } else {
-    GRID_TRACE("Dhop");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   }
+  DhopComputeTime+=usecond();
 }
 
 
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -406,6 +472,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -417,6 +484,7 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
 template<class Impl>
 void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(),FermionGrid()); // verifies full grid
   conformable(in.Grid(),out.Grid());
 
@@ -471,17 +539,12 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
   LatComplex    sk(_grid);  sk = Zero();
   LatComplex    sk2(_grid); sk2= Zero();
   LatComplex    W(_grid); W= Zero();
+  LatComplex    a(_grid); a= Zero();
   LatComplex    one  (_grid); one = ScalComplex(1.0,0.0);
   LatComplex 	cosha(_grid);
   LatComplex 	kmu(_grid);
   LatComplex 	Wea(_grid);
   LatComplex 	Wema(_grid);
-  LatComplex 	ea(_grid);
-  LatComplex 	ema(_grid);
-  LatComplex 	eaLs(_grid);
-  LatComplex 	emaLs(_grid);
-  LatComplex 	ea2Ls(_grid);
-  LatComplex 	ema2Ls(_grid);
   LatComplex 	sinha(_grid);
   LatComplex 	sinhaLs(_grid);
   LatComplex 	coshaLs(_grid);
@@ -516,29 +579,39 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
   ////////////////////////////////////////////
   cosha = (one + W*W + sk) / (abs(W)*2.0);
 
-  ea = (cosha + sqrt(cosha*cosha-one));
-  ema= (cosha - sqrt(cosha*cosha-one));
-  eaLs = pow(ea,Ls);
-  emaLs= pow(ema,Ls);
-  ea2Ls = pow(ea,2.0*Ls);
-  ema2Ls= pow(ema,2.0*Ls);
-  Wea= abs(W) * ea;
-  Wema= abs(W) * ema;
-  //  a=log(ea);
-  
-  sinha = 0.5*(ea - ema);
-  sinhaLs = 0.5*(eaLs-emaLs);
-  coshaLs = 0.5*(eaLs+emaLs);
+  // FIXME Need a Lattice acosh
+
+  {
+    autoView(cosha_v,cosha,CpuRead);
+    autoView(a_v,a,CpuWrite);
+    for(int idx=0;idx<_grid->lSites();idx++){
+      Coordinate lcoor(Nd);
+      Tcomplex cc;
+      //    RealD sgn;
+      _grid->LocalIndexToLocalCoor(idx,lcoor);
+      peekLocalSite(cc,cosha_v,lcoor);
+      assert((double)real(cc)>=1.0);
+      assert(fabs((double)imag(cc))<=1.0e-15);
+      cc = ScalComplex(::acosh(real(cc)),0.0);
+      pokeLocalSite(cc,a_v,lcoor);
+    }
+  }
+
+  Wea = ( exp( a) * abs(W)  );
+  Wema= ( exp(-a) * abs(W)  );
+  sinha = 0.5*(exp( a) - exp(-a));
+  sinhaLs = 0.5*(exp( a*Ls) - exp(-a*Ls));
+  coshaLs = 0.5*(exp( a*Ls) + exp(-a*Ls));
 
   A = one / (abs(W) * sinha * 2.0) * one / (sinhaLs * 2.0);
-  F = eaLs * (one - Wea + (Wema - one) * mass*mass);
-  F = F + emaLs * (Wema - one + (one - Wea) * mass*mass);
+  F = exp( a*Ls) * (one - Wea + (Wema - one) * mass*mass);
+  F = F + exp(-a*Ls) * (Wema - one + (one - Wea) * mass*mass);
   F = F - abs(W) * sinha * 4.0 * mass;
 
-  Bpp =  (A/F) * (ema2Ls - one) * (one - Wema) * (one - mass*mass * one);
-  Bmm =  (A/F) * (one - ea2Ls)  * (one - Wea) * (one - mass*mass * one);
-  App =  (A/F) * (ema2Ls - one) * ema * (ema - abs(W)) * (one - mass*mass * one);
-  Amm =  (A/F) * (one - ea2Ls)  * ea  * (ea  - abs(W)) * (one - mass*mass * one);
+  Bpp =  (A/F) * (exp(-a*Ls*2.0) - one) * (one - Wema) * (one - mass*mass * one);
+  Bmm =  (A/F) * (one - exp(a*Ls*2.0)) * (one - Wea) * (one - mass*mass * one);
+  App =  (A/F) * (exp(-a*Ls*2.0) - one) * exp(-a) * (exp(-a) - abs(W)) * (one - mass*mass * one);
+  Amm =  (A/F) * (one - exp(a*Ls*2.0)) * exp(a) * (exp(a) - abs(W)) * (one - mass*mass * one);
   ABpm = (A/F) * abs(W) * sinha * 2.0  * (one + mass * coshaLs * 2.0 + mass*mass * one);
 
   //P+ source, P- source
@@ -561,29 +634,29 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
       buf1_4d = Zero();
       ExtractSlice(buf1_4d, PRsource, (tt-1), 0);
       //G(s,t)
-      bufR_4d = bufR_4d + A * eaLs * pow(ema,f) * signW * buf1_4d + A * emaLs * pow(ea,f) * signW * buf1_4d;
+      bufR_4d = bufR_4d + A * exp(a*Ls) * exp(-a*f) * signW * buf1_4d + A * exp(-a*Ls) * exp(a*f) * signW * buf1_4d;
       //A++*exp(a(s+t))
-      bufR_4d = bufR_4d + App * pow(ea,ss) * pow(ea,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + App * exp(a*ss) * exp(a*tt) * signW * buf1_4d ;
       //A+-*exp(a(s-t))
-      bufR_4d = bufR_4d + ABpm * pow(ea,ss) * pow(ema,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + ABpm * exp(a*ss) * exp(-a*tt) * signW * buf1_4d ;
       //A-+*exp(a(-s+t))
-      bufR_4d = bufR_4d + ABpm * pow(ema,ss) * pow(ea,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + ABpm * exp(-a*ss) * exp(a*tt) * signW * buf1_4d ;
       //A--*exp(a(-s-t))
-      bufR_4d = bufR_4d + Amm * pow(ema,ss) * pow(ema,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + Amm * exp(-a*ss) * exp(-a*tt) * signW * buf1_4d ;
 
       //GL
       buf2_4d = Zero();
       ExtractSlice(buf2_4d, PLsource, (tt-1), 0);
       //G(s,t)
-      bufL_4d = bufL_4d + A * eaLs * pow(ema,f) * signW * buf2_4d + A * emaLs * pow(ea,f) * signW * buf2_4d;
+      bufL_4d = bufL_4d + A * exp(a*Ls) * exp(-a*f) * signW * buf2_4d + A * exp(-a*Ls) * exp(a*f) * signW * buf2_4d;
       //B++*exp(a(s+t))
-      bufL_4d = bufL_4d + Bpp * pow(ea,ss) * pow(ea,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + Bpp * exp(a*ss) * exp(a*tt) * signW * buf2_4d ;
       //B+-*exp(a(s-t))
-      bufL_4d = bufL_4d + ABpm * pow(ea,ss) * pow(ema,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + ABpm * exp(a*ss) * exp(-a*tt) * signW * buf2_4d ;
       //B-+*exp(a(-s+t))
-      bufL_4d = bufL_4d + ABpm * pow(ema,ss) * pow(ea,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + ABpm * exp(-a*ss) * exp(a*tt) * signW * buf2_4d ;
       //B--*exp(a(-s-t))
-      bufL_4d = bufL_4d + Bmm * pow(ema,ss) * pow(ema,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + Bmm * exp(-a*ss) * exp(-a*tt) * signW * buf2_4d ;
     }
     InsertSlice(bufR_4d, GR, (ss-1), 0);
     InsertSlice(bufL_4d, GL, (ss-1), 0);
@@ -702,12 +775,28 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
   W = one - M5 + sk2;
 
   ////////////////////////////////////////////
-  // Cosh alpha -> exp(+/- alpha)
+  // Cosh alpha -> alpha
   ////////////////////////////////////////////
   cosha =  (one + W*W + sk) / (abs(W)*2.0);
 
-  Wea = abs(W)*(cosha + sqrt(cosha*cosha-one));
-  Wema= abs(W)*(cosha - sqrt(cosha*cosha-one));
+  // FIXME Need a Lattice acosh
+  {
+  autoView(cosha_v,cosha,CpuRead);
+  autoView(a_v,a,CpuWrite);
+  for(int idx=0;idx<_grid->lSites();idx++){
+    Coordinate lcoor(Nd);
+    Tcomplex cc;
+    //    RealD sgn;
+    _grid->LocalIndexToLocalCoor(idx,lcoor);
+    peekLocalSite(cc,cosha_v,lcoor);
+    assert((double)real(cc)>=1.0);
+    assert(fabs((double)imag(cc))<=1.0e-15);
+    cc = ScalComplex(::acosh(real(cc)),0.0);
+    pokeLocalSite(cc,a_v,lcoor);
+  }}
+  
+  Wea = ( exp( a) * abs(W)  );
+  Wema= ( exp(-a) * abs(W)  );
   
   num   = num + ( one - Wema ) * mass * in;
   denom= ( Wea - one ) + mass*mass * (one - Wema); 

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

@@ -76,6 +76,91 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
   StencilOdd.BuildSurfaceList(1,vol4);
 }
 
+template<class Impl>
+void WilsonFermion<Impl>::Report(void)
+{
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+  RealD volume = 1;
+  Coordinate latt = _grid->GlobalDimensions();
+  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of DhopEO Calls   : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion TotalTime   /Calls        : " << DhopTotalTime   / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime    /Calls        : " << DhopCommTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion FaceTime    /Calls        : " << DhopFaceTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime1/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime2/Calls        : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
+
+    // Average the compute time
+    _grid->GlobalSum(DhopComputeTime);
+    DhopComputeTime/=NP;
+    RealD mflops = 1320*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+
+    RealD Fullmflops = 1320*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+
+    // how to count flops here?
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call               ? : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node      ? : " << mflops/NP << std::endl;
+
+    // how to count flops here?
+    RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)        ? : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full) ? : " << Fullmflops/NP << std::endl;  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil"    <<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+  }
+  if ( DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil     Reporti()"    <<std::endl;  Stencil.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilEven Reporti()"<<std::endl;  StencilEven.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd  Reporti()" <<std::endl;  StencilOdd.Reporti(DhopCalls);
+  }
+}
+
+template<class Impl>
+void WilsonFermion<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0; // ok
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopComputeTime2= 0;
+  DhopFaceTime    = 0;
+  DhopTotalTime   = 0;
+
+  DerivCalls       = 0; // ok
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+  Stencil.ZeroCountersi();
+  StencilEven.ZeroCountersi();
+  StencilOdd.ZeroCountersi();
+}
+
+
 template <class Impl>
 void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
@@ -235,6 +320,7 @@ template <class Impl>
 void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
                                         GaugeField &mat, const FermionField &A,
                                         const FermionField &B, int dag) {
+  DerivCalls++;
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -243,8 +329,11 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
   FermionField Atilde(B.Grid());
   Atilde = A;
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B, compressor);
+  DerivCommTime+=usecond();
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma (1+g)<->(1-g) if dag
@@ -252,6 +341,7 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     int gamma = mu;
     if (!dag) gamma += Nd;
 
+    DerivDhopComputeTime -= usecond();
     int Ls=1;
     Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, B.Grid()->oSites(), B, Btilde, mu, gamma);
 
@@ -259,7 +349,9 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     // spin trace outer product
     //////////////////////////////////////////////////
     Impl::InsertForce4D(mat, Btilde, Atilde, mu);
+    DerivDhopComputeTime += usecond();
   }
+  DerivComputeTime += usecond();
 }
 
 template <class Impl>
@@ -306,6 +398,7 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
 template <class Impl>
 void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -317,6 +410,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 template <class Impl>
 void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -329,6 +423,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
 template <class Impl>
 void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -393,12 +488,14 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
+  DhopTotalTime-=usecond();
 #ifdef GRID_OMP
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
 #endif
     DhopInternalSerial(st,lo,U,in,out,dag);
+  DhopTotalTime+=usecond();
 }
 
 template <class Impl>
@@ -407,7 +504,6 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
 						      const FermionField &in,
 						      FermionField &out, int dag)
 {
-  GRID_TRACE("DhopOverlapped");
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -418,55 +514,53 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
   /////////////////////////////
   std::vector<std::vector<CommsRequest_t> > requests;
   st.Prepare();
-  {
-    GRID_TRACE("Gather");
-    st.HaloGather(in,compressor);
-  }
+  DhopFaceTime-=usecond();
+  st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
 
-  tracePush("Communication");
+  DhopCommTime -=usecond();
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
-  {
-    GRID_TRACE("MergeSHM");
-    st.CommsMergeSHM(compressor);
-  }
+  DhopFaceTime-=usecond();
+  st.CommsMergeSHM(compressor);
+  DhopFaceTime+=usecond();
 
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt;
+  DhopComputeTime-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagInterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   } else {
-    GRID_TRACE("DhopInterior");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   }
+  DhopComputeTime+=usecond();
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
-  tracePop("Communication");
+  DhopCommTime   +=usecond();
+
+  DhopFaceTime-=usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
-  {
-    GRID_TRACE("Merge");
-    st.CommsMerge(compressor);
-  }
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
 
+  DhopComputeTime2-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagExterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   } else {
-    GRID_TRACE("DhopExterior");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   }
+  DhopComputeTime2+=usecond();
 };
 
 
@@ -476,22 +570,20 @@ void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
-  GRID_TRACE("DhopSerial");
   assert((dag == DaggerNo) || (dag == DaggerYes));
   Compressor compressor(dag);
-  {
-    GRID_TRACE("HaloExchange");
-    st.HaloExchange(in, compressor);
-  }
+  DhopCommTime-=usecond();
+  st.HaloExchange(in, compressor);
+  DhopCommTime+=usecond();
 
+  DhopComputeTime-=usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDag");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   } else {
-    GRID_TRACE("Dhop");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   }
+  DhopComputeTime+=usecond();
 };
 /*Change ends */
 

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

@@ -72,15 +72,20 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   if (SE->_is_local) {						\
     int perm= SE->_permute;					\
     auto tmp = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);	\
-    spProj(chi,tmp);							\
-    Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);			\
-    Recon(result, Uchi);						\
-  }									\
+    spProj(chi,tmp);						\
+  } else if ( st.same_node[Dir] ) {				\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
+  }								\
+  acceleratorSynchronise();						\
+  if (SE->_is_local || st.same_node[Dir] ) {			\
+    Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
+    Recon(result, Uchi);					\
+  }								\
   acceleratorSynchronise();
 
 #define GENERIC_STENCIL_LEG_EXT(Dir,spProj,Recon)		\
   SE = st.GetEntry(ptype, Dir, sF);				\
-  if (!SE->_is_local ) {		\
+  if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
     auto chi = coalescedRead(buf[SE->_offset],lane);		\
     Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
     Recon(result, Uchi);					\
@@ -411,6 +416,19 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
+#define KERNEL_CALL_TMP(A) \
+  const uint64_t    NN = Nsite*Ls;					\
+  auto U_p = & U_v[0];							\
+  auto in_p = & in_v[0];						\
+  auto out_p = & out_v[0];						\
+  auto st_p = st_v._entries_p;						\
+  auto st_perm = st_v._permute_type;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_perm,st_p,U_p,buf,sF,sU,in_p,out_p);	\
+    });									\
+  accelerator_barrier();
 
 #define KERNEL_CALLNB(A)						\
   const uint64_t    NN = Nsite*Ls;					\
@@ -422,16 +440,6 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
-#define KERNEL_CALL_EXT(A)						\
-  const uint64_t    NN = Nsite*Ls;					\
-  const uint64_t    sz = st.surface_list.size();			\
-  auto ptr = &st.surface_list[0];					\
-  accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
-      int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
-      WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
-
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
     int sU = ss;							\
@@ -452,7 +460,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
 #ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteSycl);    return; }
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
 #else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
 #endif     
@@ -490,6 +498,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
+     acceleratorFenceComputeStream();
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
@@ -497,11 +506,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
+     acceleratorFenceComputeStream();
    }
    assert(0 && " Kernel optimisation case not covered ");
   }

Grid/qcd/action/filters/DDHMCFilter.h

@@ -1,115 +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);
-	}
-	if ( Width==4) { 
-	  U    = where(mod(coor,B1)==Integer(B1-4),zzz,U);
-	  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); 
-	  U    = where(mod(coor,B1)==Integer(3)   ,zzz,U); 
-	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
-	  U_mu = where(mod(coor,B1)==Integer(B1-5),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

@@ -69,11 +69,6 @@ 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; }
 };
 
@@ -115,11 +110,6 @@ 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)
   {
@@ -139,13 +129,6 @@ 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);
@@ -155,27 +138,6 @@ 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/gauge/Photon.h

@@ -49,7 +49,7 @@ NAMESPACE_BEGIN(Grid);
     
     typedef Lattice<SiteLink>  LinkField;
     typedef Lattice<SiteField> Field;
-    typedef Field              ComplexField;
+    typedef LinkField          ComplexField;
   };
   
   typedef QedGImpl<vComplex> QedGImplR;

Grid/qcd/action/pseudofermion/Bounds.h

@@ -13,31 +13,6 @@ NAMESPACE_BEGIN(Grid);
       std::cout << GridLogMessage << "Pseudofermion action lamda_max "<<lambda_max<<"( bound "<<hi<<")"<<std::endl;
       assert( (lambda_max < hi) && " High Bounds Check on operator failed" );
     }
-
-     template<class Field> void ChebyBoundsCheck(LinearOperatorBase<Field> &HermOp,
-						 Field &GaussNoise,
-						 RealD lo,RealD hi) 
-    {
-      int orderfilter = 1000;
-      Chebyshev<Field> Cheb(lo,hi,orderfilter);
-
-      GridBase *FermionGrid = GaussNoise.Grid();
-
-      Field X(FermionGrid);
-      Field Z(FermionGrid);
-
-      X=GaussNoise;
-      RealD Nx = norm2(X);
-      Cheb(HermOp,X,Z);
-      RealD Nz = norm2(Z);
-
-      std::cout << "************************* "<<std::endl;
-      std::cout << " noise                    = "<<Nx<<std::endl;
-      std::cout << " Cheb x noise             = "<<Nz<<std::endl;
-      std::cout << " Ratio                    = "<<Nz/Nx<<std::endl;
-      std::cout << "************************* "<<std::endl;
-      assert( ((Nz/Nx)<1.0) && " ChebyBoundsCheck ");
-    }
       
     template<class Field> void InverseSqrtBoundsCheck(int MaxIter,double tol,
 						       LinearOperatorBase<Field> &HermOp,
@@ -65,65 +40,13 @@ NAMESPACE_BEGIN(Grid);
       X=X-Y;
       RealD Nd = norm2(X);
       std::cout << "************************* "<<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 << " 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 << "************************* "<<std::endl;
       assert( (std::sqrt(Nd/Nx)<tol) && " InverseSqrtBoundsCheck ");
     }
 
-    /* For a HermOp = M^dag M, check the approximation of  HermOp^{-1/inv_pow}
-       by computing   |X -    HermOp * [ Hermop^{-1/inv_pow} ]^{inv_pow} X|  < tol  
-       for noise X (aka GaussNoise).
-       ApproxNegPow should be the rational approximation for   X^{-1/inv_pow}
-    */
-    template<class Field> void InversePowerBoundsCheck(int inv_pow,
-						       int MaxIter,double tol,
-						       LinearOperatorBase<Field> &HermOp,
-						       Field &GaussNoise,
-						       MultiShiftFunction &ApproxNegPow) 
-    {
-      GridBase *FermionGrid = GaussNoise.Grid();
-
-      Field X(FermionGrid);
-      Field Y(FermionGrid);
-      Field Z(FermionGrid);
-
-      Field tmp1(FermionGrid), tmp2(FermionGrid);
-
-      X=GaussNoise;
-      RealD Nx = norm2(X);
-
-      ConjugateGradientMultiShift<Field> msCG(MaxIter,ApproxNegPow);
-
-      tmp1 = X;
-      
-      Field* in = &tmp1;
-      Field* out = &tmp2;
-      for(int i=0;i<inv_pow;i++){ //apply  [ Hermop^{-1/inv_pow}  ]^{inv_pow} X =   HermOp^{-1} X
-	msCG(HermOp, *in, *out); //backwards conventions!
-	if(i!=inv_pow-1) std::swap(in, out);
-      }
-      Z = *out;
-
-      RealD Nz = norm2(Z);
-
-      HermOp.HermOp(Z,Y);
-      RealD Ny = norm2(Y);
-
-      X=X-Y;
-      RealD Nd = norm2(X);
-      std::cout << "************************* "<<std::endl;
-      std::cout << " | noise |^2                         = "<<Nx<<std::endl;
-      std::cout << " | (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2        = "<<Nz<<std::endl;
-      std::cout << " | MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2   = "<<Ny<<std::endl;
-      std::cout << " | noise - MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2  = "<<Nd<<std::endl;
-      std::cout << " | noise - MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |/| noise |  = "<<std::sqrt(Nd/Nx)<<std::endl;
-      std::cout << "************************* "<<std::endl;
-      assert( (std::sqrt(Nd/Nx)<tol) && " InversePowerBoundsCheck ");
-    }
-
 NAMESPACE_END(Grid);
 

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion.h

@@ -1,163 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundaryBoson.h
-
-    Copyright (C) 2021
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-///////////////////////////////////////
-// Two flavour ratio
-///////////////////////////////////////
-template<class ImplD,class ImplF>
-class DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion : public Action<typename ImplD::GaugeField> {
-public:
-  INHERIT_IMPL_TYPES(ImplD);
-
-private:
-  SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
-  RealD InnerStoppingCondition;
-  RealD ActionStoppingCondition;
-  RealD DerivativeStoppingCondition;
-  FermionField Phi; // the pseudo fermion field for this trajectory
-public:
-  DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_NumOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
-    : NumOp(_NumOp), 
-      DerivativeStoppingCondition(_DerivativeTol),
-      ActionStoppingCondition(_ActionTol),
-      InnerStoppingCondition(_InnerTol),
-      Phi(_NumOp.FermionGrid()) {};
-
-  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion";}
-
-  virtual std::string LogParameters(){
-    std::stringstream sstream;
-    return sstream.str();
-  }  
-  
-  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
-  {
-    // P(phi) = e^{- phi^dag P^dag P phi}
-    //
-    // NumOp == P
-    //
-    // Take phi = P^{-1} eta  ; eta = P Phi
-    //
-    // P(eta) = e^{- eta^dag eta}
-    //
-    // e^{x^2/2 sig^2} => sig^2 = 0.5.
-    // 
-    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
-    //
-    RealD scale = std::sqrt(0.5);
-
-    NumOp.tolinner=InnerStoppingCondition;
-    NumOp.tol=ActionStoppingCondition;
-    NumOp.ImportGauge(U);
-
-    FermionField eta(NumOp.FermionGrid());
-
-    gaussian(pRNG,eta);    eta=eta*scale;
-    
-    NumOp.ProjectBoundaryBar(eta);
-    //DumpSliceNorm("eta",eta);
-    NumOp.RInv(eta,Phi);
-
-    //DumpSliceNorm("Phi",Phi);
-
-  };
-
-  //////////////////////////////////////////////////////
-  // S = phi^dag Pdag P phi
-  //////////////////////////////////////////////////////
-  virtual RealD S(const GaugeField &U) {
-
-    NumOp.tolinner=InnerStoppingCondition;
-    NumOp.tol=ActionStoppingCondition;
-    NumOp.ImportGauge(U);
-
-    FermionField Y(NumOp.FermionGrid());
-
-    NumOp.R(Phi,Y);
-
-    RealD action = norm2(Y);
-
-    return action;
-  };
-
-  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
-  {
-    NumOp.tolinner=InnerStoppingCondition;
-    NumOp.tol=DerivativeStoppingCondition;
-    NumOp.ImportGauge(U);
-
-    GridBase *fgrid = NumOp.FermionGrid();
-    GridBase *ugrid = NumOp.GaugeGrid();
-
-    FermionField  X(fgrid);
-    FermionField  Y(fgrid);
-    FermionField  tmp(fgrid);
-
-    GaugeField   force(ugrid);	
-
-    FermionField DobiDdbPhi(fgrid);      // Vector A in my notes
-    FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
-    FermionField DoidP_Phi(fgrid);    // Vector E in my notes
-    FermionField DobidDddDoidP_Phi(fgrid);    // Vector F in my notes
-    
-    FermionField P_Phi(fgrid);
-    
-    // P term
-    NumOp.dBoundaryBar(Phi,tmp);
-    NumOp.dOmegaBarInv(tmp,DobiDdbPhi);        // Vector A
-    NumOp.dBoundary(DobiDdbPhi,tmp);
-    NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi);      // Vector B
-    P_Phi  = Phi - DoiDdDobiDdbPhi;
-    NumOp.ProjectBoundaryBar(P_Phi);
-    
-    // P^dag P term
-    NumOp.dOmegaDagInv(P_Phi,DoidP_Phi); // Vector E
-    NumOp.dBoundaryDag(DoidP_Phi,tmp);
-    NumOp.dOmegaBarDagInv(tmp,DobidDddDoidP_Phi);   // Vector F
-    NumOp.dBoundaryBarDag(DobidDddDoidP_Phi,tmp);
-
-    X = DobiDdbPhi;
-    Y = DobidDddDoidP_Phi;
-    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=force;
-    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
-
-    X = DoiDdDobiDdbPhi;
-    Y = DoidP_Phi;
-    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
-    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
-
-    dSdU *= -1.0;
-
-  };
-};
-
-NAMESPACE_END(Grid);
-

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourPseudoFermion.h

@@ -1,158 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.h
-
-    Copyright (C) 2021
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-///////////////////////////////////////
-// Two flavour ratio
-///////////////////////////////////////
-template<class ImplD,class ImplF>
-class DomainDecomposedBoundaryTwoFlavourPseudoFermion : public Action<typename ImplD::GaugeField> {
-public:
-  INHERIT_IMPL_TYPES(ImplD);
-
-private:
-  SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
-  RealD ActionStoppingCondition;
-  RealD DerivativeStoppingCondition;
-  RealD InnerStoppingCondition;
-
-  FermionField Phi; // the pseudo fermion field for this trajectory
-
-  RealD refresh_action;
-public:
-  DomainDecomposedBoundaryTwoFlavourPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_DenOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol = 1.0e-6 )
-    : DenOp(_DenOp),
-      DerivativeStoppingCondition(_DerivativeTol),
-      ActionStoppingCondition(_ActionTol),
-      InnerStoppingCondition(_InnerTol),
-      Phi(_DenOp.FermionGrid()) {};
-      
-  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourPseudoFermion";}
-
- 
-  virtual std::string LogParameters(){
-    std::stringstream sstream;
-    return sstream.str();
-  }  
-  
-  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
-  {
-    // P(phi) = e^{- phi^dag Rdag^-1 R^-1 phi}
-    //
-    // DenOp == R
-    //
-    // Take phi = R eta  ; eta = R^-1 Phi
-    //
-    // P(eta) = e^{- eta^dag eta}
-    //
-    // e^{x^2/2 sig^2} => sig^2 = 0.5.
-    // 
-    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
-    //
-    RealD scale = std::sqrt(0.5);
-
-    DenOp.tolinner=InnerStoppingCondition;
-    DenOp.tol     =ActionStoppingCondition;
-    DenOp.ImportGauge(U);
-
-    FermionField eta(DenOp.FermionGrid());
-
-    gaussian(pRNG,eta);    eta=eta*scale;
-    
-    DenOp.ProjectBoundaryBar(eta);
-    DenOp.R(eta,Phi);
-    //DumpSliceNorm("Phi",Phi);
-    refresh_action = norm2(eta);
-  };
-
-  //////////////////////////////////////////////////////
-  // S = phi^dag Rdag^-1 R^-1 phi
-  //////////////////////////////////////////////////////
-  virtual RealD S(const GaugeField &U) {
-
-    DenOp.tolinner=InnerStoppingCondition;
-    DenOp.tol=ActionStoppingCondition;
-    DenOp.ImportGauge(U);
-
-    FermionField X(DenOp.FermionGrid());
-
-    DenOp.RInv(Phi,X);
-
-    RealD action = norm2(X);
-
-    return action;
-  };
-
-  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
-  {
-    DenOp.tolinner=InnerStoppingCondition;
-    DenOp.tol=DerivativeStoppingCondition;
-    DenOp.ImportGauge(U);
-
-    GridBase *fgrid = DenOp.FermionGrid();
-    GridBase *ugrid = DenOp.GaugeGrid();
-
-    FermionField  X(fgrid);
-    FermionField  Y(fgrid);
-    FermionField  tmp(fgrid);
-
-    GaugeField   force(ugrid);	
-
-    FermionField DiDdb_Phi(fgrid);      // Vector C in my notes
-    FermionField DidRinv_Phi(fgrid);    // Vector D in my notes
-    FermionField Rinv_Phi(fgrid);
-
-//   FermionField RinvDagRinv_Phi(fgrid);
-//   FermionField DdbdDidRinv_Phi(fgrid);
-
-    // R^-1 term
-    DenOp.dBoundaryBar(Phi,tmp);
-    DenOp.Dinverse(tmp,DiDdb_Phi);            // Vector C
-    Rinv_Phi = Phi - DiDdb_Phi;
-    DenOp.ProjectBoundaryBar(Rinv_Phi); 
- 
-    // R^-dagger R^-1 term
-    DenOp.DinverseDag(Rinv_Phi,DidRinv_Phi); // Vector D
-/*
-    DenOp.dBoundaryBarDag(DidRinv_Phi,DdbdDidRinv_Phi);
-    RinvDagRinv_Phi = Rinv_Phi - DdbdDidRinv_Phi;
-    DenOp.ProjectBoundaryBar(RinvDagRinv_Phi);
-*/
-    X = DiDdb_Phi;
-    Y = DidRinv_Phi;
-    DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=force;
-    DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
-    DumpSliceNorm("force",dSdU);
-    dSdU *= -1.0;
-  };
-};
-
-NAMESPACE_END(Grid);
-

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion.h

@@ -1,237 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.h
-
-    Copyright (C) 2021
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-///////////////////////////////////////
-// Two flavour ratio
-///////////////////////////////////////
-template<class ImplD,class ImplF>
-class DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion : public Action<typename ImplD::GaugeField> {
-public:
-  INHERIT_IMPL_TYPES(ImplD);
-
-private:
-  SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
-  SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
-
-  RealD InnerStoppingCondition;
-  RealD ActionStoppingCondition;
-  RealD DerivativeStoppingCondition;
-  
-  FermionField Phi; // the pseudo fermion field for this trajectory
-
-public:
-  DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_NumOp, 
-						       SchurFactoredFermionOperator<ImplD,ImplF>  &_DenOp,
-						       RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
-    : NumOp(_NumOp), DenOp(_DenOp),
-      Phi(_NumOp.PeriodicFermOpD.FermionGrid()),
-      InnerStoppingCondition(_InnerTol),
-      DerivativeStoppingCondition(_DerivativeTol),
-      ActionStoppingCondition(_ActionTol)
-  {};
-      
-  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion";}
- 
-  virtual std::string LogParameters(){
-    std::stringstream sstream;
-    return sstream.str();
-  }  
-  
-  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
-  {
-    NumOp.ImportGauge(U);
-    DenOp.ImportGauge(U);
-
-    FermionField eta(NumOp.PeriodicFermOpD.FermionGrid());
-    FermionField tmp(NumOp.PeriodicFermOpD.FermionGrid());
-
-    // P(phi) = e^{- phi^dag P^dag Rdag^-1 R^-1 P phi}
-    //
-    // NumOp == P
-    // DenOp == R
-    //
-    // Take phi = P^{-1} R eta  ; eta = R^-1 P Phi
-    //
-    // P(eta) = e^{- eta^dag eta}
-    //
-    // e^{x^2/2 sig^2} => sig^2 = 0.5.
-    // 
-    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
-    //
-    RealD scale = std::sqrt(0.5);
-
-    gaussian(pRNG,eta);    eta=eta*scale;
-    
-    NumOp.ProjectBoundaryBar(eta);
-    NumOp.tolinner=InnerStoppingCondition;
-    DenOp.tolinner=InnerStoppingCondition;
-    DenOp.tol = ActionStoppingCondition;
-    NumOp.tol = ActionStoppingCondition;
-    DenOp.R(eta,tmp);
-    NumOp.RInv(tmp,Phi);
-    DumpSliceNorm("Phi",Phi);
-
-  };
-
-  //////////////////////////////////////////////////////
-  // S = phi^dag Pdag Rdag^-1 R^-1 P phi
-  //////////////////////////////////////////////////////
-  virtual RealD S(const GaugeField &U) {
-
-    NumOp.ImportGauge(U);
-    DenOp.ImportGauge(U);
-
-    FermionField X(NumOp.PeriodicFermOpD.FermionGrid());
-    FermionField Y(NumOp.PeriodicFermOpD.FermionGrid());
-
-    NumOp.tolinner=InnerStoppingCondition;
-    DenOp.tolinner=InnerStoppingCondition;
-    DenOp.tol = ActionStoppingCondition;
-    NumOp.tol = ActionStoppingCondition;
-    NumOp.R(Phi,Y);
-    DenOp.RInv(Y,X);
-
-    RealD action = norm2(X);
-    //    std::cout << " DD boundary action is " <<action<<std::endl;
-
-    return action;
-  };
-
-  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
-  {
-    NumOp.ImportGauge(U);
-    DenOp.ImportGauge(U);
-
-    GridBase *fgrid = NumOp.PeriodicFermOpD.FermionGrid();
-    GridBase *ugrid = NumOp.PeriodicFermOpD.GaugeGrid();
-
-    FermionField  X(fgrid);
-    FermionField  Y(fgrid);
-    FermionField  tmp(fgrid);
-
-    GaugeField   force(ugrid);	
-
-    FermionField DobiDdbPhi(fgrid);      // Vector A in my notes
-    FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
-    FermionField DiDdbP_Phi(fgrid);      // Vector C in my notes
-    FermionField DidRinvP_Phi(fgrid);    // Vector D in my notes
-    FermionField DdbdDidRinvP_Phi(fgrid);
-    FermionField DoidRinvDagRinvP_Phi(fgrid);    // Vector E in my notes
-    FermionField DobidDddDoidRinvDagRinvP_Phi(fgrid);    // Vector F in my notes
-    
-    FermionField P_Phi(fgrid);
-    FermionField RinvP_Phi(fgrid);
-    FermionField RinvDagRinvP_Phi(fgrid);
-    FermionField PdagRinvDagRinvP_Phi(fgrid);
-
-    //    RealD action = S(U);
-    NumOp.tolinner=InnerStoppingCondition;
-    DenOp.tolinner=InnerStoppingCondition;
-    DenOp.tol = DerivativeStoppingCondition;
-    NumOp.tol = DerivativeStoppingCondition;
-    
-    // P term
-    NumOp.dBoundaryBar(Phi,tmp);
-    NumOp.dOmegaBarInv(tmp,DobiDdbPhi);        // Vector A
-    NumOp.dBoundary(DobiDdbPhi,tmp);
-    NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi);      // Vector B
-    P_Phi  = Phi - DoiDdDobiDdbPhi;
-    NumOp.ProjectBoundaryBar(P_Phi);
-
-    // R^-1 P term
-    DenOp.dBoundaryBar(P_Phi,tmp);
-    DenOp.Dinverse(tmp,DiDdbP_Phi);            // Vector C
-    RinvP_Phi = P_Phi - DiDdbP_Phi;
-    DenOp.ProjectBoundaryBar(RinvP_Phi); // Correct to here
-
- 
-    // R^-dagger R^-1 P term
-    DenOp.DinverseDag(RinvP_Phi,DidRinvP_Phi); // Vector D
-    DenOp.dBoundaryBarDag(DidRinvP_Phi,DdbdDidRinvP_Phi);
-    RinvDagRinvP_Phi = RinvP_Phi - DdbdDidRinvP_Phi;
-    DenOp.ProjectBoundaryBar(RinvDagRinvP_Phi);
-
-    
-    // P^dag R^-dagger R^-1 P term
-    NumOp.dOmegaDagInv(RinvDagRinvP_Phi,DoidRinvDagRinvP_Phi); // Vector E
-    NumOp.dBoundaryDag(DoidRinvDagRinvP_Phi,tmp);
-    NumOp.dOmegaBarDagInv(tmp,DobidDddDoidRinvDagRinvP_Phi);   // Vector F
-    NumOp.dBoundaryBarDag(DobidDddDoidRinvDagRinvP_Phi,tmp);
-    PdagRinvDagRinvP_Phi = RinvDagRinvP_Phi- tmp;
-    NumOp.ProjectBoundaryBar(PdagRinvDagRinvP_Phi);
-
-    /*
-    std::cout << "S eval  "<< action << std::endl;
-    std::cout << "S - IP1 "<< innerProduct(Phi,PdagRinvDagRinvP_Phi) << std::endl;
-    std::cout << "S - IP2 "<< norm2(RinvP_Phi) << std::endl;
-
-    NumOp.R(Phi,tmp);
-    tmp = tmp - P_Phi;
-    std::cout << "diff1 "<<norm2(tmp) <<std::endl;
-    
-    
-    DenOp.RInv(P_Phi,tmp);
-    tmp = tmp - RinvP_Phi;
-    std::cout << "diff2 "<<norm2(tmp) <<std::endl;
-
-    DenOp.RDagInv(RinvP_Phi,tmp);
-    tmp  = tmp - RinvDagRinvP_Phi;
-    std::cout << "diff3 "<<norm2(tmp) <<std::endl;
-
-    DenOp.RDag(RinvDagRinvP_Phi,tmp);
-    tmp  = tmp - PdagRinvDagRinvP_Phi;
-    std::cout << "diff4 "<<norm2(tmp) <<std::endl;
-    */
-    
-    dSdU=Zero();
-
-    X = DobiDdbPhi;
-    Y = DobidDddDoidRinvDagRinvP_Phi;
-    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
-    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
-
-    X = DoiDdDobiDdbPhi;
-    Y = DoidRinvDagRinvP_Phi;
-    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
-    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
-
-    X = DiDdbP_Phi;
-    Y = DidRinvP_Phi;
-    DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
-    DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
-
-    dSdU *= -1.0;
-
-  };
-};
-
-NAMESPACE_END(Grid);
-

Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h

@@ -44,10 +44,6 @@ 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>
   {
@@ -61,60 +57,37 @@ NAMESPACE_BEGIN(Grid);
       bool use_heatbath_forecasting;
       AbstractEOFAFermion<Impl>& Lop; // the basic LH operator
       AbstractEOFAFermion<Impl>& Rop; // the basic RH operator
-      SchurRedBlackDiagMooeeSolve<FermionField> SolverHBL;
-      SchurRedBlackDiagMooeeSolve<FermionField> SolverHBR;
+      SchurRedBlackDiagMooeeSolve<FermionField> SolverHB;
       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,CG,p,use_fc) {};
-
-      //Use the same solver for L,R in the heatbath but different solvers elsewhere
+	: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,CG,CG,CG,CG,CG,p,use_fc) {};
+	
       ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop, 
 					      AbstractEOFAFermion<Impl>& _Rop,
-					      OperatorFunction<FermionField>& HeatbathCG,
-					      OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR, 
-					      OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR, 
-					      Params& p, 
-					      bool use_fc=false)
-	: 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>& HeatbathCG, 
 					      OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR, 
 					      OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR, 
 					      Params& p, 
 					      bool use_fc=false) : 
         Lop(_Lop), 
 	Rop(_Rop), 
-	SolverHBL(HeatbathCGL,false,true), SolverHBR(HeatbathCGR,false,true),
+	SolverHB(HeatbathCG,false,true),
 	SolverL(ActionCGL, false, true), SolverR(ActionCGR, false, true), 
 	DerivativeSolverL(DerivCGL, false, true), DerivativeSolverR(DerivCGR, false, true), 
 	Phi(_Lop.FermionGrid()), 
 	param(p), 
-	use_heatbath_forecasting(use_fc),
-	initial_action(false)
+        use_heatbath_forecasting(use_fc)
       {
         AlgRemez remez(param.lo, param.hi, param.precision);
 
@@ -124,8 +97,6 @@ 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() {
@@ -146,19 +117,6 @@ 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
@@ -166,10 +124,12 @@ 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
       //
-     void refresh(const GaugeField &U, const FermionField &eta) {
+      virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
+      {
         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());
@@ -180,6 +140,11 @@ 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] ); }
@@ -195,15 +160,15 @@ NAMESPACE_BEGIN(Grid);
         tmp[1] = Zero();
         for(int k=0; k<param.degree; ++k){
           gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
-          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);
-            SolverHBL(Lop, CG_src, CG_soln);
+            SolverHB(Lop, CG_src, CG_soln);
             prev_solns.push_back(CG_soln);
           } else {
             CG_soln = Zero(); // Just use zero as the initial guess
-	    SolverHBL(Lop, CG_src, CG_soln);
+            SolverHB(Lop, CG_src, CG_soln);
           }
           Lop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
           tmp[1] = tmp[1] + ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Lop.k ) * tmp[0];
@@ -222,15 +187,15 @@ NAMESPACE_BEGIN(Grid);
         if(use_heatbath_forecasting){ prev_solns.clear(); } // empirically, LH solns don't help for RH solves
         for(int k=0; k<param.degree; ++k){
           gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
-	  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);
-            SolverHBR(Rop, CG_src, CG_soln);
+            SolverHB(Rop, CG_src, CG_soln);
             prev_solns.push_back(CG_soln);
           } else {
             CG_soln = Zero();
-            SolverHBR(Rop, CG_src, CG_soln);
+            SolverHB(Rop, CG_src, CG_soln);
           }
           Rop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
           tmp[1] = tmp[1] - ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Rop.k ) * tmp[0];
@@ -240,117 +205,49 @@ NAMESPACE_BEGIN(Grid);
         Phi = Phi + tmp[1];
 
         // Reset shift coefficients for energy and force evals
-	heatbathRefreshShiftCoefficients(0, 0.0);
-	heatbathRefreshShiftCoefficients(1, -1.0);
-
-	//Mark that the next call to S is the first after refresh
-	initial_action = true;
-
+        Lop.RefreshShiftCoefficients(0.0);
+        Rop.RefreshShiftCoefficients(-1.0);
 
 	// Bounds check
 	RealD EtaDagEta = norm2(eta);
-	norm2_eta = EtaDagEta;
-
 	//	RealD PhiDagMPhi= norm2(eta);
 
       };
 
-      void Meofa(const GaugeField& U,const FermionField &in, FermionField & out) 
+      void Meofa(const GaugeField& U,const FermionField &phi, FermionField & Mphi) 
       {
+#if 0
         Lop.ImportGauge(U);
         Rop.ImportGauge(U);
 
-        FermionField spProj_in(Lop.FermionGrid());
+        FermionField spProj_Phi(Lop.FermionGrid());
+	FermionField mPhi(Lop.FermionGrid());
         std::vector<FermionField> tmp(2, Lop.FermionGrid());
-	out = in;
+	mPhi = phi;
 	
         // LH term: S = S - k <\Phi| P_{-} \Omega_{-}^{\dagger} H(mf)^{-1} \Omega_{-} P_{-} |\Phi>
-        spProj(in, spProj_in, -1, Lop.Ls);
-        Lop.Omega(spProj_in, tmp[0], -1, 0);
+        spProj(Phi, spProj_Phi, -1, Lop.Ls);
+        Lop.Omega(spProj_Phi, tmp[0], -1, 0);
         G5R5(tmp[1], tmp[0]);
         tmp[0] = Zero();
         SolverL(Lop, tmp[1], tmp[0]);
         Lop.Dtilde(tmp[0], tmp[1]); // We actually solved Cayley preconditioned system: transform back
         Lop.Omega(tmp[1], tmp[0], -1, 1);
-	spProj(tmp[0], tmp[1], -1, Lop.Ls);
-
-	out = out -  Lop.k * tmp[1];
+	mPhi = mPhi -  Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
 
         // RH term: S = S + k <\Phi| P_{+} \Omega_{+}^{\dagger} ( H(mb)
-        //               - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} |\Phi>
-        spProj(in, spProj_in, 1, Rop.Ls);
-        Rop.Omega(spProj_in, tmp[0], 1, 0);
+        //               - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{-} P_{-} |\Phi>
+        spProj(Phi, spProj_Phi, 1, Rop.Ls);
+        Rop.Omega(spProj_Phi, tmp[0], 1, 0);
         G5R5(tmp[1], tmp[0]);
         tmp[0] = Zero();
         SolverR(Rop, tmp[1], tmp[0]);
         Rop.Dtilde(tmp[0], tmp[1]);
         Rop.Omega(tmp[1], tmp[0], 1, 1);
-	spProj(tmp[0], tmp[1], 1, Rop.Ls);
-
-        out = out + Rop.k * tmp[1];
+        action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real();
+#endif
       }
 
-      //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)
       {
@@ -374,7 +271,7 @@ 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]);
@@ -384,26 +281,6 @@ 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;
       };
 
@@ -452,40 +329,6 @@ 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

@@ -1,372 +0,0 @@
-    /*************************************************************************************
-
-    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

@@ -1,93 +0,0 @@
-    /*************************************************************************************
-
-    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

@@ -40,61 +40,249 @@ NAMESPACE_BEGIN(Grid);
     // Here N/D \sim R_{-1/2} ~ (M^dagM)^{-1/2}  
   
     template<class Impl>
-    class OneFlavourEvenOddRatioRationalPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<Impl> {
+    class OneFlavourEvenOddRatioRationalPseudoFermionAction : public Action<typename Impl::GaugeField> {
     public:
+
+      INHERIT_IMPL_TYPES(Impl);
+
       typedef OneFlavourRationalParams Params;
+      Params param;
+
+      MultiShiftFunction PowerHalf   ;
+      MultiShiftFunction PowerNegHalf;
+      MultiShiftFunction PowerQuarter;
+      MultiShiftFunction PowerNegQuarter;
+
     private:
-      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;
-      }
+     
+      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
 
     public:
+
       OneFlavourEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
-							FermionOperator<Impl>  &_DenOp, 
-							const Params & p
-							) : 
-	GeneralEvenOddRatioRationalPseudoFermionAction<Impl>(_NumOp, _DenOp, transcribe(p)){}
+					    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);
 
-      virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
-    };
+	// 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);
 
-    template<class Impl,class ImplF>
-    class OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction : public GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF> {
-    public:
-      typedef OneFlavourRationalParams Params;
-    private:
-      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;
+	// 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) {
 
-    public:
-      OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
-								 FermionOperator<Impl>  &_DenOp, 
-								 FermionOperator<ImplF>  &_NumOpF, 
-								 FermionOperator<ImplF>  &_DenOpF, 
-								 const Params & p, Integer ReliableUpdateFreq
-							) : 
-	GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF>(_NumOp, _DenOp,_NumOpF, _DenOpF, transcribe(p),ReliableUpdateFreq){}
+	// 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).
 
-      virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
+	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);
+
+      };
     };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h

@@ -49,12 +49,10 @@ NAMESPACE_BEGIN(Grid);
       Params param;
 
       MultiShiftFunction PowerHalf   ;
-      MultiShiftFunction PowerQuarter;
       MultiShiftFunction PowerNegHalf;
+      MultiShiftFunction PowerQuarter;
       MultiShiftFunction PowerNegQuarter;
 
-      MultiShiftFunction MDPowerQuarter;
-      MultiShiftFunction MDPowerNegHalf;
     private:
      
       FermionOperator<Impl> & NumOp;// the basic operator
@@ -75,22 +73,15 @@ NAMESPACE_BEGIN(Grid);
 	remez.generateApprox(param.degree,1,2);
 	PowerHalf.Init(remez,param.tolerance,false);
 	PowerNegHalf.Init(remez,param.tolerance,true);
-	MDPowerNegHalf.Init(remez,param.mdtolerance,true);
 
 	// MdagM^(+- 1/4)
 	std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/4)"<<std::endl;
 	remez.generateApprox(param.degree,1,4);
    	PowerQuarter.Init(remez,param.tolerance,false);
-   	MDPowerQuarter.Init(remez,param.mdtolerance,false);
 	PowerNegQuarter.Init(remez,param.tolerance,true);
       };
 
-      virtual std::string action_name(){
-	std::stringstream sstream;
-	sstream<<"OneFlavourRatioRationalPseudoFermionAction("
-	       <<DenOp.Mass()<<") / det("<<NumOp.Mass()<<")";
-	return sstream.str();
-      }
+      virtual std::string action_name(){return "OneFlavourRatioRationalPseudoFermionAction";}
       
       virtual std::string LogParameters(){
 	std::stringstream sstream;
@@ -213,8 +204,8 @@ NAMESPACE_BEGIN(Grid);
 
       virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
 
-	const int n_f  = MDPowerNegHalf.poles.size();
-	const int n_pv = MDPowerQuarter.poles.size();
+	const int n_f  = PowerNegHalf.poles.size();
+	const int n_pv = PowerQuarter.poles.size();
 
 	std::vector<FermionField> MpvPhi_k     (n_pv,NumOp.FermionGrid());
 	std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionGrid());
@@ -233,8 +224,8 @@ NAMESPACE_BEGIN(Grid);
 	MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagM(DenOp);
 	MdagMLinearOperator<FermionOperator<Impl> ,FermionField> VdagV(NumOp);
 
-	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,MDPowerQuarter);
-	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,MDPowerNegHalf);
+	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
+	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
 
 	msCG_V(VdagV,Phi,MpvPhi_k,MpvPhi);
 	msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
@@ -253,7 +244,7 @@ NAMESPACE_BEGIN(Grid);
 
 	//(1)
 	for(int k=0;k<n_f;k++){
-	  ak = MDPowerNegHalf.residues[k];
+	  ak = PowerNegHalf.residues[k];
 	  DenOp.M(MfMpvPhi_k[k],Y);
 	  DenOp.MDeriv(tmp , MfMpvPhi_k[k], Y,DaggerYes );  dSdU=dSdU+ak*tmp;
 	  DenOp.MDeriv(tmp , Y, MfMpvPhi_k[k], DaggerNo );  dSdU=dSdU+ak*tmp;
@@ -263,7 +254,7 @@ NAMESPACE_BEGIN(Grid);
 	//(3)
 	for(int k=0;k<n_pv;k++){
 
-          ak = MDPowerQuarter.residues[k];
+          ak = PowerQuarter.residues[k];
 	  
 	  NumOp.M(MpvPhi_k[k],Y);
 	  NumOp.MDeriv(tmp,MpvMfMpvPhi_k[k],Y,DaggerYes); dSdU=dSdU+ak*tmp;

Grid/qcd/action/pseudofermion/PseudoFermion.h

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

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -75,22 +75,24 @@ NAMESPACE_BEGIN(Grid);
           conformable(_NumOp.GaugeRedBlackGrid(), _DenOp.GaugeRedBlackGrid());
         };
 
-      virtual std::string action_name(){
-	std::stringstream sstream;
-	sstream<<"TwoFlavourEvenOddRatioPseudoFermionAction det("<<DenOp.Mass()<<") / det("<<NumOp.Mass()<<")";
-	return sstream.str();
-      }
+      virtual std::string action_name(){return "TwoFlavourEvenOddRatioPseudoFermionAction";}
 
       virtual std::string LogParameters(){
 	std::stringstream sstream;
-	sstream<< GridLogMessage << "["<<action_name()<<"] -- No further parameters "<<std::endl;
+	sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
 	return sstream.str();
       } 
 
       
-      const FermionField &getPhiOdd() const{ return PhiOdd; }
-
       virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
+
+        // 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
+        //
         // P(eta_o) = e^{- eta_o^dag eta_o}
         //
         // e^{x^2/2 sig^2} => sig^2 = 0.5.
@@ -98,22 +100,12 @@ NAMESPACE_BEGIN(Grid);
         RealD scale = std::sqrt(0.5);
 
         FermionField eta    (NumOp.FermionGrid());
-        gaussian(pRNG,eta); eta = eta * scale;
-
-	refresh(U,eta);
-      }
-
-      void refresh(const GaugeField &U, const FermionField &eta) {
-
-        // P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
-        //
-        // NumOp == V
-        // DenOp == M
-        //
         FermionField etaOdd (NumOp.FermionRedBlackGrid());
         FermionField etaEven(NumOp.FermionRedBlackGrid());
         FermionField tmp    (NumOp.FermionRedBlackGrid());
 
+        gaussian(pRNG,eta);
+
         pickCheckerboard(Even,etaEven,eta);
         pickCheckerboard(Odd,etaOdd,eta);
 
@@ -132,6 +124,10 @@ NAMESPACE_BEGIN(Grid);
         // Even det factors
         DenOp.MooeeDag(etaEven,tmp);
         NumOp.MooeeInvDag(tmp,PhiEven);
+
+        PhiOdd =PhiOdd*scale;
+        PhiEven=PhiEven*scale;
+        
       };
 
       //////////////////////////////////////////////////////

Grid/qcd/action/pseudofermion/TwoFlavourRatioEO4DPseudoFermion.h

@@ -1,203 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/qcd/action/pseudofermion/TwoFlavourRatio.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-///////////////////////////////////////
-// Two flavour ratio
-///////////////////////////////////////
-template<class Impl>
-class TwoFlavourRatioEO4DPseudoFermionAction : public Action<typename Impl::GaugeField> {
-public:
-  INHERIT_IMPL_TYPES(Impl);
-
-private:
-  typedef FermionOperator<Impl> FermOp;
-  FermionOperator<Impl> & NumOp;// the basic operator
-  FermionOperator<Impl> & DenOp;// the basic operator
-
-  OperatorFunction<FermionField> &DerivativeSolver;
-  OperatorFunction<FermionField> &DerivativeDagSolver;
-  OperatorFunction<FermionField> &ActionSolver;
-  OperatorFunction<FermionField> &HeatbathSolver;
-
-  FermionField phi4; // the pseudo fermion field for this trajectory
-
-public:
-  TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
-					 FermionOperator<Impl>  &_DenOp, 
-					 OperatorFunction<FermionField> & DS,
-					 OperatorFunction<FermionField> & AS ) : 
-    TwoFlavourRatioEO4DPseudoFermionAction(_NumOp,_DenOp, DS,DS,AS,AS) {};
-  TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
-					 FermionOperator<Impl>  &_DenOp, 
-					 OperatorFunction<FermionField> & DS,
-					 OperatorFunction<FermionField> & DDS,
-					 OperatorFunction<FermionField> & AS,
-					 OperatorFunction<FermionField> & HS
-				       ) : NumOp(_NumOp),
-					   DenOp(_DenOp),
-					   DerivativeSolver(DS),
-					   DerivativeDagSolver(DDS),
-					   ActionSolver(AS),
-					   HeatbathSolver(HS),
-					   phi4(_NumOp.GaugeGrid())
-  {};
-      
-  virtual std::string action_name(){return "TwoFlavourRatioEO4DPseudoFermionAction";}
-
-  virtual std::string LogParameters(){
-    std::stringstream sstream;
-    sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
-    return sstream.str();
-  }  
-      
-  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
-
-    // P(phi) = e^{- phi^dag (V^dag M^-dag)_11  (M^-1 V)_11 phi}
-    //
-    // NumOp == V
-    // DenOp == M
-    //
-    // Take phi = (V^{-1} M)_11 eta  ; eta = (M^{-1} V)_11 Phi
-    //
-    // P(eta) = e^{- eta^dag eta}
-    //
-    // e^{x^2/2 sig^2} => sig^2 = 0.5.
-    // 
-    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
-    //
-    RealD scale = std::sqrt(0.5);
-
-    FermionField eta4(NumOp.GaugeGrid());
-    FermionField eta5(NumOp.FermionGrid());
-    FermionField tmp(NumOp.FermionGrid());
-    FermionField phi5(NumOp.FermionGrid());
-
-    gaussian(pRNG,eta4);
-    NumOp.ImportFourDimPseudoFermion(eta4,eta5);
-    NumOp.ImportGauge(U);
-    DenOp.ImportGauge(U);
-
-    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(HeatbathSolver);
-
-    DenOp.M(eta5,tmp);               // M eta
-    PrecSolve(NumOp,tmp,phi5);  // phi = V^-1 M eta
-    phi5=phi5*scale;
-    std::cout << GridLogMessage << "4d pf refresh "<< norm2(phi5)<<"\n";
-    // Project to 4d
-    NumOp.ExportFourDimPseudoFermion(phi5,phi4);
-      
-  };
-
-  //////////////////////////////////////////////////////
-  // S = phi^dag (V^dag M^-dag)_11  (M^-1 V)_11 phi
-  //////////////////////////////////////////////////////
-  virtual RealD S(const GaugeField &U) {
-
-    NumOp.ImportGauge(U);
-    DenOp.ImportGauge(U);
-
-    FermionField Y4(NumOp.GaugeGrid());
-    FermionField X(NumOp.FermionGrid());
-    FermionField Y(NumOp.FermionGrid());
-    FermionField phi5(NumOp.FermionGrid());
-	
-    MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagMOp(DenOp);
-    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(ActionSolver);
-
-    NumOp.ImportFourDimPseudoFermion(phi4,phi5);
-    NumOp.M(phi5,X);              // X= V phi
-    PrecSolve(DenOp,X,Y);    // Y= (MdagM)^-1 Mdag Vdag phi = M^-1 V phi
-    NumOp.ExportFourDimPseudoFermion(Y,Y4);
-
-    RealD action = norm2(Y4);
-
-    return action;
-  };
-
-  //////////////////////////////////////////////////////
-  // dS/du = 2 Re phi^dag (V^dag M^-dag)_11  (M^-1 d V)_11  phi
-  //       - 2 Re phi^dag (dV^dag M^-dag)_11  (M^-1 dM M^-1 V)_11  phi
-  //////////////////////////////////////////////////////
-  virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
-
-    NumOp.ImportGauge(U);
-    DenOp.ImportGauge(U);
-
-    FermionField  X(NumOp.FermionGrid());
-    FermionField  Y(NumOp.FermionGrid());
-    FermionField       phi(NumOp.FermionGrid());
-    FermionField      Vphi(NumOp.FermionGrid());
-    FermionField  MinvVphi(NumOp.FermionGrid());
-    FermionField      tmp4(NumOp.GaugeGrid());
-    FermionField  MdagInvMinvVphi(NumOp.FermionGrid());
-
-    GaugeField   force(NumOp.GaugeGrid());	
-
-    //Y=V phi
-    //X = (Mdag V phi
-    //Y = (Mdag M)^-1 Mdag V phi = M^-1 V Phi
-    NumOp.ImportFourDimPseudoFermion(phi4,phi);
-    NumOp.M(phi,Vphi);               //  V phi
-    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(DerivativeSolver);
-    PrecSolve(DenOp,Vphi,MinvVphi);// M^-1 V phi
-    std::cout << GridLogMessage << "4d deriv solve "<< norm2(MinvVphi)<<"\n";
-
-    // Projects onto the physical space and back
-    NumOp.ExportFourDimPseudoFermion(MinvVphi,tmp4);
-    NumOp.ImportFourDimPseudoFermion(tmp4,Y);
-
-    SchurRedBlackDiagMooeeDagSolve<FermionField> PrecDagSolve(DerivativeDagSolver);
-    // X = proj M^-dag V phi
-    // Need an adjoint solve
-    PrecDagSolve(DenOp,Y,MdagInvMinvVphi);
-    std::cout << GridLogMessage << "4d deriv solve dag "<< norm2(MdagInvMinvVphi)<<"\n";
-    
-    // phi^dag (Vdag Mdag^-1) (M^-1 dV)  phi
-    NumOp.MDeriv(force ,MdagInvMinvVphi , phi, DaggerNo );  dSdU=force;
-  
-    // phi^dag (dVdag Mdag^-1) (M^-1 V)  phi
-    NumOp.MDeriv(force , phi, MdagInvMinvVphi ,DaggerYes  );  dSdU=dSdU+force;
-
-    //    - 2 Re phi^dag (dV^dag M^-dag)_11  (M^-1 dM M^-1 V)_11  phi
-    DenOp.MDeriv(force,MdagInvMinvVphi,MinvVphi,DaggerNo);   dSdU=dSdU-force;
-    DenOp.MDeriv(force,MinvVphi,MdagInvMinvVphi,DaggerYes);  dSdU=dSdU-force;
-
-    dSdU *= -1.0; 
-    //dSdU = - Ta(dSdU);
-    
-  };
-};
-
-NAMESPACE_END(Grid);
-
-

Grid/qcd/gparity/Gparity.h

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

Grid/qcd/gparity/GparityFlavour.cc

@@ -1,34 +0,0 @@
-#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

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

Grid/qcd/hmc/GenericHMCrunner.h

@@ -129,10 +129,18 @@ public:
     Runner(S);
   }
 
-  //Use the checkpointer to initialize the RNGs and the gauge field, writing the resulting gauge field into U.
-  //This is called automatically by Run but may be useful elsewhere, e.g. for integrator tuning experiments
-  void initializeGaugeFieldAndRNGs(Field &U){
-    if(!Resources.haveRNGs()) Resources.AddRNGs();
+  //////////////////////////////////////////////////////////////////
+
+private:
+  template <class SmearingPolicy>
+  void Runner(SmearingPolicy &Smearing) {
+    auto UGrid = Resources.GetCartesian();
+    Resources.AddRNGs();
+    Field U(UGrid);
+
+    // Can move this outside?
+    typedef IntegratorType<SmearingPolicy> TheIntegrator;
+    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
 
     if (Parameters.StartingType == "HotStart") {
       // Hot start
@@ -151,43 +159,14 @@ 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, CheckpointStartReseed]\n";
+	<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart]\n";
       exit(1);
     }
-  }
-
-
-
-  //////////////////////////////////////////////////////////////////
-
-private:
-  template <class SmearingPolicy>
-  void Runner(SmearingPolicy &Smearing) {
-    auto UGrid = Resources.GetCartesian();
-    Field U(UGrid);
-
-    initializeGaugeFieldAndRNGs(U);
-
-    typedef IntegratorType<SmearingPolicy> TheIntegrator;
-    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
-
-    // Sets the momentum filter
-    MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
 
     Smearing.set_Field(U);
 

Grid/qcd/hmc/HMC.h

@@ -34,7 +34,6 @@ directory
 			    * @brief Classes for Hybrid Monte Carlo update
 			    *
 			    * @author Guido Cossu
-			    * @author Peter Boyle
 			    */
 			   //--------------------------------------------------------------------
 #pragma once
@@ -53,7 +52,6 @@ struct HMCparameters: Serializable {
                                   Integer, Trajectories, /* @brief Number of sweeps in this run */
                                   bool, MetropolisTest,
                                   Integer, NoMetropolisUntil,
-				  bool, PerformRandomShift, /* @brief Randomly shift the gauge configuration at the start of a trajectory */
                                   std::string, StartingType,
                                   IntegratorParameters, MD)
 
@@ -64,7 +62,6 @@ struct HMCparameters: Serializable {
     StartTrajectory   = 0;
     Trajectories      = 10;
     StartingType      = "HotStart";
-    PerformRandomShift = true;
     /////////////////////////////////
   }
 
@@ -85,7 +82,6 @@ struct HMCparameters: Serializable {
     std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
     std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
     std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
-    std::cout << GridLogMessage << "[HMC parameters] Doing random shift      : " << std::boolalpha << PerformRandomShift << "\n";
     std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
     MD.print_parameters();
   }
@@ -98,7 +94,6 @@ private:
   const HMCparameters Params;
 
   typedef typename IntegratorType::Field Field;
-  typedef typename IntegratorType::FieldImplementation FieldImplementation;
   typedef std::vector< HmcObservable<Field> * > ObsListType;
 
   //pass these from the resource manager
@@ -120,17 +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 << GridLogMessage
+              << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "exp(-dH) = " << prob
+              << "  Random = " << rn_test << "\n";
+    std::cout << GridLogMessage
+              << "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 << GridLogMessage << "Metropolis_test -- ACCEPTED\n";
+      std::cout << GridLogMessage
+                << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
-      std::cout << GridLogHMC << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogHMC << "--------------------------------------------------\n";
+      std::cout << GridLogMessage << "Metropolis_test -- REJECTED\n";
+      std::cout << GridLogMessage
+                << "--------------------------------------------------\n";
       return false;
     }
   }
@@ -139,79 +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();
-
-    if(Params.PerformRandomShift){
-      //////////////////////////////////////////////////////////////////////////////////////////////////////
-      // Mainly for DDHMC perform a random translation of U modulo volume
-      //////////////////////////////////////////////////////////////////////////////////////////////////////
-      std::cout << GridLogMessage << "--------------------------------------------------\n";
-      std::cout << GridLogMessage << "Random shifting gauge field by [";
-
-      std::vector<typename FieldImplementation::GaugeLinkField> Umu(Grid->Nd(), U.Grid());
-      for(int mu=0;mu<Grid->Nd();mu++) Umu[mu] = PeekIndex<LorentzIndex>(U, mu);
-
-      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";
-      
-	//shift all fields together in a way that respects the gauge BCs
-	for(int mu=0; mu < Grid->Nd(); mu++)
-	  Umu[mu] = FieldImplementation::CshiftLink(Umu[mu],d,shift);
-      }
-
-      for(int mu=0;mu<Grid->Nd();mu++) PokeIndex<LorentzIndex>(U,Umu[mu],mu);
-      
-      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 << GridLogMessage << "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;
@@ -223,16 +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 << GridLogMessage << "Total H after trajectory  = " << H1
+	      << "  dH = " << H1 - H0 << "\n";
     std::cout.precision(current_precision);
     
     return (H1 - H0);
   }
+  
+
+  
 
 public:
   /////////////////////////////////////////
@@ -254,13 +195,10 @@ 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";
-
+      std::cout << GridLogMessage << "-- # Trajectory = " << traj << "\n";
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
-      	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
+      	std::cout << GridLogMessage << "-- Thermalization" << std::endl;
       }
       
       double t0=usecond();
@@ -269,19 +207,20 @@ public:
       DeltaH = evolve_hmc_step(Ucopy);
       // Metropolis-Hastings test
       bool accept = true;
-      if (Params.MetropolisTest && traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
+      if (traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
         accept = metropolis_test(DeltaH);
       } else {
-      	std::cout << GridLogHMC << "Skipping Metropolis test" << std::endl;
+      	std::cout << GridLogMessage << "Skipping Metropolis test" << std::endl;
       }
 
       if (accept)
         Ucur = Ucopy; 
       
+     
+      
       double t1=usecond();
-      std::cout << GridLogHMC << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
+      std::cout << GridLogMessage << "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;
@@ -289,7 +228,7 @@ public:
       	std::cout << GridLogDebug << "Observables pointer " << Observables[obs] << std::endl;
         Observables[obs]->TrajectoryComplete(traj + 1, Ucur, sRNG, pRNG);
       }
-      std::cout << GridLogHMC << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
+      std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
     }
   }
 

Grid/qcd/hmc/HMCModules.h

@@ -80,9 +80,7 @@ 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;
@@ -243,9 +226,6 @@ public:
   //////////////////////////////////////////////////////
   // Random number generators
   //////////////////////////////////////////////////////
-  
-  //Return true if the RNG objects have been instantiated
-  bool haveRNGs() const{ return have_RNG; }
 
   void AddRNGs(std::string s = "") {
     // Couple the RNGs to the GridModule tagged by s

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,6 +33,7 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
+#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -63,10 +64,9 @@ public:
 };
 
 /*! @brief Class for Molecular Dynamics management */
-template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy>
+template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
-  typedef FieldImplementation_ FieldImplementation;
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 
@@ -119,59 +119,36 @@ 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();
-
-      //      DumpSliceNorm("force ",force,Nd-1);
-      MomFilter->applyFilter(force);
-      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<< std::endl;
-      DumpSliceNorm("force filtered ",force,Nd-1);
-      
-      Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
-      Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      Real force_max   = std::sqrt(maxLocalNorm2(force));
-      Real impulse_max = force_max * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      as[level].actions.at(a)->deriv_log(force_abs,force_max,impulse_abs,impulse_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 force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
+      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_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) 
@@ -185,12 +162,8 @@ 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);
@@ -233,68 +206,6 @@ 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()
-		  <<" Fdt max  "  << as[level].actions.at(actionID)->Fdt_max_average()
-		  <<" Fdt norm "  << as[level].actions.at(actionID)->Fdt_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;
@@ -313,6 +224,7 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
+
   }
 
   void reverse_momenta()
@@ -337,19 +249,15 @@ public:
   void refresh(Field& U,  GridSerialRNG & sRNG, GridParallelRNG& pRNG) 
   {
     assert(P.Grid() == U.Grid());
-    std::cout << GridLogIntegrator << "Integrator refresh" << std::endl;
+    std::cout << GridLogIntegrator << "Integrator refresh\n";
 
-    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
@@ -359,19 +267,15 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
-	auto name = as[level].actions.at(actionID)->action_name();
-        std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
-
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
-	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
-	as[level].actions.at(actionID)->refresh_timer_stop();
       }
 
       // Refresh the higher representation actions
       as[level].apply(refresh_hireps, Representations, sRNG, pRNG);
     }
 
+    MomFilter->applyFilter(P);
   }
 
   // to be used by the actionlevel class to iterate
@@ -406,9 +310,7 @@ 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/hmc/integrators/Integrator_algorithm.h

@@ -92,11 +92,10 @@ NAMESPACE_BEGIN(Grid);
  *  P 1/2                            P 1/2
  */
 
-template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
-class LeapFrog : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
+template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
+class LeapFrog : public Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy> 
 {
 public:
-  typedef FieldImplementation_ FieldImplementation;
   typedef LeapFrog<FieldImplementation, SmearingPolicy, RepresentationPolicy> Algorithm;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
@@ -136,14 +135,13 @@ public:
   }
 };
 
-template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
-class MinimumNorm2 : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
+template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
+class MinimumNorm2 : public Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy> 
 {
 private:
   const RealD lambda = 0.1931833275037836;
 
 public:
-  typedef FieldImplementation_ FieldImplementation;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
   MinimumNorm2(GridBase* grid, IntegratorParameters Par, ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm)
@@ -194,8 +192,8 @@ public:
   }
 };
 
-template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
-class ForceGradient : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
+template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
+class ForceGradient : public Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy> 
 {
 private:
   const RealD lambda = 1.0 / 6.0;
@@ -204,7 +202,6 @@ private:
   const RealD theta = 0.0;
 
 public:
-  typedef FieldImplementation_ FieldImplementation;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
   // Looks like dH scales as dt^4. tested wilson/wilson 2 level.

Grid/qcd/hmc/integrators/MomentumFilter.h

@@ -37,8 +37,7 @@ NAMESPACE_BEGIN(Grid);
 
 template<typename MomentaField>
 struct MomentumFilterBase{
-  virtual void applyFilter(MomentaField &P) const = 0;
-  virtual ~MomentumFilterBase(){};
+  virtual void applyFilter(MomentaField &P) const;
 };
 
 //Do nothing
@@ -84,6 +83,7 @@ struct MomentumFilterApplyPhase: public MomentumFilterBase<MomentaField>{
     
   }
 
+
 };
 
 

Grid/qcd/observables/topological_charge.h

@@ -31,16 +31,15 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
-
 struct TopologySmearingParameters : Serializable {
   GRID_SERIALIZABLE_CLASS_MEMBERS(TopologySmearingParameters,
+				  int, steps,
+				  float, step_size,
 				  int, meas_interval,
-				  float, init_step_size,
-				  float, maxTau,
-				  float, tolerance);
+				  float, maxTau);
 
- TopologySmearingParameters(float ss = 0.0f, int mi = 0, float mT = 0.0f, float tol = 1e-4):
-  init_step_size(ss), meas_interval(mi), maxTau(mT), tolerance(tol){}
+  TopologySmearingParameters(int s = 0, float ss = 0.0f, int mi = 0, float mT = 0.0f):
+    steps(s), step_size(ss), meas_interval(mi), maxTau(mT){}
 
   template < class ReaderClass >
   TopologySmearingParameters(Reader<ReaderClass>& Reader){
@@ -98,9 +97,9 @@ public:
         
       if (Pars.do_smearing){
 	// using wilson flow by default here
-	WilsonFlowAdaptive<PeriodicGimplR> WF(Pars.Smearing.init_step_size, Pars.Smearing.maxTau, Pars.Smearing.tolerance, Pars.Smearing.meas_interval);
-	WF.smear(Usmear, U);
-	Real T0   = WF.energyDensityPlaquette(Pars.Smearing.maxTau, Usmear);
+	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);
 	std::cout << GridLogMessage << std::setprecision(std::numeric_limits<Real>::digits10 + 1)
 		  << "T0                : [ " << traj << " ] "<< T0 << std::endl;
       }

Grid/qcd/smearing/WilsonFlow.h

@@ -7,7 +7,6 @@ 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
@@ -33,318 +32,177 @@ directory
 NAMESPACE_BEGIN(Grid);
 
 template <class Gimpl>
-class WilsonFlowBase: 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
+class WilsonFlow: public Smear<Gimpl>{
+  unsigned int Nstep;
+  unsigned int measure_interval;
+  mutable RealD epsilon, taus;
 
-protected:
-  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); }
+
 public:
   INHERIT_GIMPL_TYPES(Gimpl)
 
-  explicit WilsonFlowBase(unsigned int 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
-    setDefaultMeasurements(meas_interval);
+    assert(epsilon > 0.0);
+    LogMessage();
   }
-    
-  void resetActions(){ functions.clear(); }
 
-  void addMeasurement(int meas_interval, FunctionType meas){ functions.push_back({meas_interval, meas}); }
+  void LogMessage() {
+    std::cout << GridLogMessage
+	      << "[WilsonFlow] Nstep   : " << Nstep << std::endl;
+    std::cout << GridLogMessage
+	      << "[WilsonFlow] epsilon : " << epsilon << std::endl;
+    std::cout << GridLogMessage
+	      << "[WilsonFlow] full trajectory : " << Nstep * epsilon << std::endl;
+  }
 
-  //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);
+  virtual void smear(GaugeField&, const GaugeField&) const;
 
-  void derivative(GaugeField&, const GaugeField&, const GaugeField&) const override{
+  virtual void derivative(GaugeField&, const GaugeField&, const GaugeField&) const {
     assert(0);
     // undefined for WilsonFlow
   }
 
-  //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);
+  void smear_adaptive(GaugeField&, const GaugeField&, RealD maxTau);
+  RealD energyDensityPlaquette(unsigned int step, const GaugeField& U) const;
+  RealD energyDensityPlaquette(const GaugeField& U) const;
 };
 
-//Basic iterative Wilson flow
-template <class Gimpl>
-class WilsonFlow: public WilsonFlowBase<Gimpl>{
-private:
-  int Nstep; //number of steps
-  RealD epsilon;  //step size
-
-  //Evolve the gauge field by 1 step of size eps and update tau
-  void evolve_step(typename Gimpl::GaugeField &U, RealD &tau) const;
-
-public:
-  INHERIT_GIMPL_TYPES(Gimpl)
-
-  //Integrate the Wilson flow for Nstep steps of size epsilon
-  WilsonFlow(const RealD epsilon, const int Nstep, unsigned int meas_interval = 1): WilsonFlowBase<Gimpl>(meas_interval), Nstep(Nstep), epsilon(epsilon){}
-
-  void smear(GaugeField& out, const GaugeField& in) const override;
-};
-
-//Wilson flow with adaptive step size
-template <class Gimpl>
-class WilsonFlowAdaptive: public WilsonFlowBase<Gimpl>{
-private:
-  RealD init_epsilon; //initial step size
-  RealD maxTau; //integrate to t=maxTau
-  RealD tolerance; //integration error tolerance
-
-  //Evolve the gauge field by 1 step and update tau and the current time step eps
-  //
-  //If the step size eps is too large that a significant integration error results,
-  //the gauge field (U) and tau will not be updated and the function will return 0; eps will be adjusted to a smaller
-  //value for the next iteration.
-  //
-  //For a successful integration step the function will return 1
-  int evolve_step_adaptive(typename Gimpl::GaugeField&U, RealD &tau, RealD &eps) const;
-
-public:
-  INHERIT_GIMPL_TYPES(Gimpl)
-
-  WilsonFlowAdaptive(const RealD init_epsilon, const RealD maxTau, const RealD tolerance, unsigned int meas_interval = 1): 
-  WilsonFlowBase<Gimpl>(meas_interval), init_epsilon(init_epsilon), maxTau(maxTau), tolerance(tolerance){}
-
-  void smear(GaugeField& out, const GaugeField& in) const override;
-};
 
 ////////////////////////////////////////////////////////////////////////////////
 // Implementations
 ////////////////////////////////////////////////////////////////////////////////
 template <class Gimpl>
-RealD WilsonFlowBase<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 WilsonFlowBase<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> WilsonFlowBase<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>
-std::vector<RealD> WilsonFlowBase<Gimpl>::flowMeasureEnergyDensityPlaquette(const GaugeField& U, int measure_interval){
-  GaugeField V(U);
-  return flowMeasureEnergyDensityPlaquette(V,U, measure_interval);
-}
-
-template <class Gimpl>
-std::vector<RealD> WilsonFlowBase<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> WilsonFlowBase<Gimpl>::flowMeasureEnergyDensityCloverleaf(const GaugeField& U, int measure_interval){
-  GaugeField V(U);
-  return flowMeasureEnergyDensityCloverleaf(V,U, measure_interval);
-}
-
-template <class Gimpl>
-void WilsonFlowBase<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;
-    });
-}
-
-
-
-template <class Gimpl>
-void WilsonFlow<Gimpl>::evolve_step(typename Gimpl::GaugeField &U, RealD &tau) const{
+void WilsonFlow<Gimpl>::evolve_step(typename Gimpl::GaugeField &U) const{
   GaugeField Z(U.Grid());
   GaugeField tmp(U.Grid());
-  this->SG.deriv(U, Z);
+  SG.deriv(U, Z);
   Z *= 0.25;                                  // Z0 = 1/4 * F(U)
   Gimpl::update_field(Z, U, -2.0*epsilon);    // U = W1 = exp(ep*Z0)*W0
 
   Z *= -17.0/8.0;
-  this->SG.deriv(U, tmp); Z += tmp;                 // -17/32*Z0 +Z1
+  SG.deriv(U, tmp); Z += tmp;                 // -17/32*Z0 +Z1
   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
 
   Z *= -4.0/3.0;
-  this->SG.deriv(U, tmp); Z += tmp;                 // 4/3*(17/36*Z0 -8/9*Z1) +Z2
+  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>::smear(GaugeField& out, const GaugeField& in) const{
-  std::cout << GridLogMessage
-	    << "[WilsonFlow] Nstep   : " << Nstep << std::endl;
-  std::cout << GridLogMessage
-	    << "[WilsonFlow] epsilon : " << epsilon << std::endl;
-  std::cout << GridLogMessage
-	    << "[WilsonFlow] full trajectory : " << Nstep * epsilon << std::endl;
+void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, RealD maxTau) {
+  if (maxTau - taus < epsilon){
+    epsilon = maxTau-taus;
+  }
+  //std::cout << GridLogMessage << "Integration epsilon : " << epsilon << std::endl;
+  GaugeField Z(U.Grid());
+  GaugeField Zprime(U.Grid());
+  GaugeField tmp(U.Grid()), Uprime(U.Grid());
+  Uprime = U;
+  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
 
+  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
+    
+
+  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
+
+  // Ramos 
+  Gimpl::update_field(Zprime, Uprime, -2.0*epsilon); // 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;
+  //std::cout << GridLogMessage << "Adjusting integration step with distance: " << diff << std::endl;
+    
+  epsilon = epsilon*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();
+}
+
+template <class Gimpl>
+RealD WilsonFlow<Gimpl>::energyDensityPlaquette(const GaugeField& U) const {
+  return 2.0 * taus * taus * SG.S(U)/U.Grid()->gSites();
+}
+
+
+//#define WF_TIMING 
+
+
+
+template <class Gimpl>
+void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const {
   out = in;
-  RealD taus = 0.;
-  for (unsigned int step = 1; step <= Nstep; step++) { //step indicates the number of smearing steps applied at the time of measurement
+  for (unsigned int step = 1; step <= Nstep; step++) {
     auto start = std::chrono::high_resolution_clock::now();
-    evolve_step(out, taus);
+    evolve_step(out);
     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
-    //Perform measurements
-    for(auto const &meas : this->functions)
-      if( step % meas.first == 0 ) meas.second(step,taus,out);
+    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;
+    }
   }
 }
 
-
-
 template <class Gimpl>
-int WilsonFlowAdaptive<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, RealD &tau, RealD &eps) const{
-  if (maxTau - tau < eps){
-    eps = maxTau-tau;
-  }
-  //std::cout << GridLogMessage << "Integration epsilon : " << epsilon << std::endl;
-  GaugeField Z(U.Grid());
-  GaugeField Zprime(U.Grid());
-  GaugeField tmp(U.Grid()), Uprime(U.Grid()), Usave(U.Grid());
-  Uprime = U;
-  Usave = U;
-
-  this->SG.deriv(U, Z);
-  Zprime = -Z;
-  Z *= 0.25;                                  // Z0 = 1/4 * F(U)
-  Gimpl::update_field(Z, U, -2.0*eps);    // U = W1 = exp(ep*Z0)*W0
-
-  Z *= -17.0/8.0;
-  this->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*eps);    // U_= W2 = exp(ep*Z)*W1
-    
-
-  Z *= -4.0/3.0;
-  this->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*eps);    // V(t+e) = exp(ep*Z)*W2
-
-  // Ramos arXiv:1301.4388
-  Gimpl::update_field(Zprime, Uprime, -2.0*eps); // V'(t+e) = exp(ep*Z')*W0
-
-  // Compute distance using Ramos' definition
-  GaugeField diffU = U - Uprime;
-  RealD max_dist = 0;
-
-  for(int mu=0;mu<Nd;mu++){
-    typename Gimpl::GaugeLinkField diffU_mu = PeekIndex<LorentzIndex>(diffU, mu);
-    RealD dist_mu = sqrt( maxLocalNorm2(diffU_mu) ) /Nc/Nc; //maximize over sites
-    max_dist = std::max(max_dist, dist_mu); //maximize over mu
-  }
-  
-  int ret;
-  if(max_dist < tolerance) {
-    tau += eps;
-    ret = 1;
-  } else {
-    U = Usave;
-    ret = 0;
-  }
-  eps = eps*0.95*std::pow(tolerance/max_dist,1./3.);
-  std::cout << GridLogMessage << "Adaptive smearing : Distance: "<< max_dist <<" Step successful: " << ret << " New epsilon: " << eps << std::endl; 
-
-  return ret;
-}
-
-template <class Gimpl>
-void WilsonFlowAdaptive<Gimpl>::smear(GaugeField& out, const GaugeField& in) const{
-  std::cout << GridLogMessage
-	    << "[WilsonFlow] initial epsilon : " << init_epsilon << std::endl;
-  std::cout << GridLogMessage
-	    << "[WilsonFlow] full trajectory : " << maxTau << std::endl;
-  std::cout << GridLogMessage
-	    << "[WilsonFlow] tolerance   : " << tolerance << std::endl;
+void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, RealD maxTau){
   out = in;
-  RealD taus = 0.;
-  RealD eps = init_epsilon;
+  taus = epsilon;
   unsigned int step = 0;
   do{
-    int step_success = evolve_step_adaptive(out, taus, eps); 
-    step += step_success; //step will not be incremented if the integration step fails
-
-    //Perform measurements
-    if(step_success)
-      for(auto const &meas : this->functions)
-	if( step % meas.first == 0 ) meas.second(step,taus,out);
+    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;
+    }
   } while (taus < maxTau);
+
+
+
 }
 
-
-
 NAMESPACE_END(Grid);
 

Grid/qcd/utils/CovariantCshift.h

@@ -88,12 +88,6 @@ 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);
-  }
 
 }
 
@@ -164,9 +158,6 @@ 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();
@@ -185,9 +176,6 @@ 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)
   {
@@ -220,47 +208,6 @@ 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
-  //shift = 2
-  //Out(x) = U_\mu(x+2\hat\mu)  | x_\mu < L-2
-  //       = U*_\mu(1)  | x_\mu == L-1
-  //       = U*_\mu(0)  | x_\mu == L-2
-  //shift = -2
-  //Out(x) = U_\mu(x-2mu)  | x_\mu > 1
-  //       = U*_\mu(L-2)  | x_\mu == 0
-  //       = U*_\mu(L-1)  | x_\mu == 1
-  //etc
-  template<class gauge> Lattice<gauge>
-  CshiftLink(const Lattice<gauge> &Link, int mu, int shift)
-  {
-    GridBase *grid = Link.Grid();
-    int Lmu = grid->GlobalDimensions()[mu];
-    assert(abs(shift) < Lmu && "Invalid shift value");
-
-    Lattice<iScalar<vInteger>> coor(grid);
-    LatticeCoordinate(coor, mu);
-
-    Lattice<gauge> tmp(grid);
-    if(shift > 0){
-      tmp = Cshift(Link, mu, shift);
-      tmp = where(coor >= Lmu-shift, conjugate(tmp), tmp);
-      return tmp;
-    }else if(shift < 0){
-      tmp = Link;
-      tmp = where(coor >= Lmu+shift, conjugate(tmp), tmp);
-      return Cshift(tmp, mu, shift);
-    }
-    
-    //shift == 0
-    return Link;
-  }
-
 }
 
 

Grid/qcd/utils/GaugeFix.h

@@ -40,45 +40,27 @@ public:
   typedef typename Gimpl::GaugeLinkField GaugeMat;
   typedef typename Gimpl::GaugeField GaugeLorentz;
 
-  //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 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;
+    }
   }
-  
-  //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);
-
+  static void DmuAmu(const std::vector<GaugeMat> &A,GaugeMat &dmuAmu,int orthog) {
     dmuAmu=Zero();
     for(int mu=0;mu<Nd;mu++){
       if ( mu != orthog ) {
-	//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;
+	dmuAmu = dmuAmu + A[mu] - Cshift(A[mu],mu,-1);
       }
     }
   }  
 
-  //Fix the gauge field Umu
-  //0 < alpha < 1 is related to the step size, cf https://arxiv.org/pdf/1405.5812.pdf
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
     SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
   }
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
-  //Fix the gauge field Umu and also return the gauge transformation from the original gauge field, xform
+  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) {
 
     GridBase *grid = Umu.Grid();
 
@@ -141,25 +123,28 @@ public:
       }
     }
     std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
-    if (err_on_no_converge)
-      assert(0 && "Gauge fixing did not converge within the specified number of iterations");
+    if (err_on_no_converge) assert(0);
   };
-  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);
-    ExpiAlphaDmuAmu(U,g,alpha,dmuAmu,orthog);
+
+    GaugeLinkToLieAlgebraField(U,A);
+    ExpiAlphaDmuAmu(A,g,alpha,dmuAmu,orthog);
+
 
     Real vol = grid->gSites();
     Real trG = TensorRemove(sum(trace(g))).real()/vol/Nc;
 
     xform = g*xform ;
-    SU<Nc>::GaugeTransform<Gimpl>(U,g);
+    SU<Nc>::GaugeTransform(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();
 
@@ -174,7 +159,11 @@ public:
 
     GaugeMat g(grid);
     GaugeMat dmuAmu_p(grid);
-    DmuAmu(U,dmuAmu,orthog);
+    std::vector<GaugeMat> A(Nd,grid);
+
+    GaugeLinkToLieAlgebraField(U,A);
+
+    DmuAmu(A,dmuAmu,orthog);
 
     std::vector<int> mask(Nd,1);
     for(int mu=0;mu<Nd;mu++) if (mu==orthog) mask[mu]=0;
@@ -218,16 +207,16 @@ public:
     Real trG = TensorRemove(sum(trace(g))).real()/vol/Nc;
 
     xform = g*xform ;
-    SU<Nc>::GaugeTransform<Gimpl>(U,g);
+    SU<Nc>::GaugeTransform(U,g);
 
     return trG;
   }
 
-  static void ExpiAlphaDmuAmu(const std::vector<GaugeMat> &U,GaugeMat &g, Real alpha, GaugeMat &dmuAmu,int orthog) {
+  static void ExpiAlphaDmuAmu(const std::vector<GaugeMat> &A,GaugeMat &g,Real & alpha, GaugeMat &dmuAmu,int orthog) {
     GridBase *grid = g.Grid();
     Complex cialpha(0.0,-alpha);
     GaugeMat ciadmam(grid);
-    DmuAmu(U,dmuAmu,orthog);
+    DmuAmu(A,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 Gimpl>
-  static void GaugeTransform(typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
+  template<typename GaugeField,typename GaugeMat>
+  static void GaugeTransform( GaugeField &Umu, GaugeMat &g){
     GridBase *grid = Umu.Grid();
     conformable(grid,g.Grid());
 
-    typename Gimpl::GaugeLinkField U(grid);
-    typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
+    GaugeMat U(grid);
+    GaugeMat ag(grid); ag = adj(g);
 
     for(int mu=0;mu<Nd;mu++){
       U= PeekIndex<LorentzIndex>(Umu,mu);
-      U = g*U*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
+      U = g*U*Cshift(ag, mu, 1);
       PokeIndex<LorentzIndex>(Umu,U,mu);
     }
   }
-  template<typename Gimpl>
-  static void GaugeTransform( std::vector<typename Gimpl::GaugeLinkField> &U, typename Gimpl::GaugeLinkField &g){
+  template<typename GaugeMat>
+  static void GaugeTransform( std::vector<GaugeMat> &U, GaugeMat &g){
     GridBase *grid = g.Grid();
-    typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
+    GaugeMat ag(grid); ag = adj(g);
     for(int mu=0;mu<Nd;mu++){
-      U[mu] = g*U[mu]*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
+      U[mu] = g*U[mu]*Cshift(ag, mu, 1);
     }
   }
-  template<typename Gimpl>
-  static void RandomGaugeTransform(GridParallelRNG &pRNG, typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
+  template<typename GaugeField,typename GaugeMat>
+  static void RandomGaugeTransform(GridParallelRNG &pRNG, GaugeField &Umu, GaugeMat &g){
     LieRandomize(pRNG,g,1.0);
-    GaugeTransform<Gimpl>(Umu,g);
+    GaugeTransform(Umu,g);
   }
 
   // Projects the algebra components a lattice matrix (of dimension ncol*ncol -1 )

Grid/qcd/utils/WilsonLoops.h

@@ -125,57 +125,6 @@ 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
   //////////////////////////////////////////////////
@@ -413,11 +362,11 @@ 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 + Gimpl::CshiftLink(vu, mu, -1));
+      FS = (u*v + Cshift(vu, mu, -1));
       FS = 0.125*(FS - adj(FS));
   }
 
-  static Real TopologicalCharge(const GaugeLorentz &U){
+  static Real TopologicalCharge(GaugeLorentz &U){
     // 4d topological charge
     assert(Nd==4);
     // Bx = -iF(y,z), By = -iF(z,y), Bz = -iF(x,y)
@@ -440,203 +389,6 @@ 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/serialisation/JSON_IO.cc

@@ -26,7 +26,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef GRID_HIP
 
 NAMESPACE_BEGIN(Grid);
 
@@ -82,7 +82,7 @@ void JSONWriter::writeDefault(const std::string &s,	const std::string &x)
   if (s.size())
     ss_ << "\""<< s << "\" : \"" << os.str() << "\" ," ;
   else
-    ss_ << os.str() << " ," ;
+    ss_ << "\""<< os.str() << "\" ," ;
 }
 
 // Reader implementation ///////////////////////////////////////////////////////

Grid/serialisation/JSON_IO.h

@@ -54,7 +54,7 @@ namespace Grid
     void pop(void);
     template <typename U>
     void writeDefault(const std::string &s, const U &x);
-#ifdef __NVCC__
+#if defined(GRID_CUDA) || defined(GRID_HIP)
     void writeDefault(const std::string &s, const Grid::ComplexD &x) 
     { 
       std::complex<double> z(real(x),imag(x));
@@ -101,7 +101,7 @@ namespace Grid
     void readDefault(const std::string &s, std::vector<U> &output);
     template <typename U, typename P>
     void readDefault(const std::string &s, std::pair<U,P> &output);
-#ifdef __NVCC__
+#if defined(GRID_CUDA) || defined(GRID_HIP)
     void readDefault(const std::string &s, ComplexD &output)
     { 
       std::complex<double> z;

Grid/serialisation/Serialisation.h

@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include "BinaryIO.h"
 #include "TextIO.h"
 #include "XmlIO.h"
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef GRID_HIP
 #include "JSON_IO.h"
 #endif
 

Grid/simd/Grid_sse4.h

@@ -35,7 +35,7 @@ Author: neo <cossu@post.kek.jp>
 */
 // Time-stamp: <2015-06-16 23:27:54 neo>
 //----------------------------------------------------------------------
-#include <immintrin.h>
+
 #include <pmmintrin.h>
 
 NAMESPACE_BEGIN(Grid);

Grid/stencil/SimpleCompressor.h

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