First cut at faster GPU slice sum via atomics

Feature/fermtoprop

Grid/DisableWarnings.h

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

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/GridStd.h

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

Grid/Grid_Eigen_Dense.h

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

Grid/algorithms/Algorithms.h

@@ -54,7 +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));
       });
@@ -292,7 +292,6 @@ public:
 template<class Field>
 class ChebyshevLanczos : public Chebyshev<Field> {
 private:
-
   std::vector<RealD> Coeffs;
   int order;
   RealD alpha;

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

@@ -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)
   { 
@@ -182,9 +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

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -1,411 +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)
-  { 
-    GridBase *DoublePrecGrid = src_d.Grid();
-    precisionChangeWorkspace wk_f_from_d(SinglePrecGrid, DoublePrecGrid);
-    precisionChangeWorkspace wk_d_from_f(DoublePrecGrid, SinglePrecGrid);
-
-    ////////////////////////////////////////////////////////////////////////
-    // Convenience references to the info stored in "MultiShiftFunction"
-    ////////////////////////////////////////////////////////////////////////
-    int nshift = shifts.order;
-
-    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
-    std::vector<RealD> &mresidual(shifts.tolerances);
-    std::vector<RealD> alpha(nshift,1.0);
-
-    //Double precision search directions
-    FieldD p_d(DoublePrecGrid);
-    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
-
-    FieldD tmp_d(DoublePrecGrid);
-    FieldD r_d(DoublePrecGrid);
-    FieldD mmp_d(DoublePrecGrid);
-
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
-  
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
-  
-    const int       primary =0;
-  
-    //Primary shift fields CG iteration
-    RealD a,b,c,d;
-    RealD cp,bp,qq; //prev
-  
-    // Matrix mult fields
-    FieldF r_f(SinglePrecGrid);
-    FieldF p_f(SinglePrecGrid);
-    FieldF tmp_f(SinglePrecGrid);
-    FieldF mmp_f(SinglePrecGrid);
-    FieldF src_f(SinglePrecGrid);
-    precisionChange(src_f, src_d, wk_f_from_d);
-
-    // Check lightest mass
-    for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
-      converged[s]=0;
-    }
-  
-    // Wire guess to zero
-    // Residuals "r" are src
-    // First search direction "p" is also src
-    cp = norm2(src_d);
-
-    // Handle trivial case of zero src.
-    if( cp == 0. ){
-      for(int s=0;s<nshift;s++){
-	psi_d[s] = Zero();
-	IterationsToCompleteShift[s] = 1;
-	TrueResidualShift[s] = 0.;
-      }
-      return;
-    }
-
-    for(int s=0;s<nshift;s++){
-      rsq[s] = cp * mresidual[s] * mresidual[s];
-      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
-      ps_d[s] = src_d;
-    }
-    // r and p for primary
-    r_f=src_f; //residual maintained in single
-    p_f=src_f;
-    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
-  
-    //MdagM+m[0]
-    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    axpy(mmp_f,mass[0],p_f,mmp_f);
-    RealD rn = norm2(p_f);
-    d += rn*mass[0];
-
-    b = -cp /d;
-  
-    // Set up the various shift variables
-    int       iz=0;
-    z[0][1-iz] = 1.0;
-    z[0][iz]   = 1.0;
-    bs[0]      = b;
-    for(int s=1;s<nshift;s++){
-      z[s][1-iz] = 1.0;
-      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
-      bs[s]      = b*z[s][iz]; 
-    }
-  
-    // r += b[0] A.p[0]
-    // c= norm(r)
-    c=axpy_norm(r_f,b,mmp_f,r_f);
-  
-    for(int s=0;s<nshift;s++) {
-      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
-    }
-  
-    ///////////////////////////////////////
-    // Timers
-    ///////////////////////////////////////
-    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
-
-    SolverTimer.Start();
-  
-    // Iteration loop
-    int k;
-  
-    for (k=1;k<=MaxIterations;k++){    
-      a = c /cp;
-
-      //Update double precision search direction by residual
-      PrecChangeTimer.Start();
-      precisionChange(r_d, r_f, wk_d_from_f);
-      PrecChangeTimer.Stop();
-
-      AXPYTimer.Start();
-      axpy(p_d,a,p_d,r_d); 
-
-      for(int s=0;s<nshift;s++){
-	if ( ! converged[s] ) { 
-	  if (s==0){
-	    axpy(ps_d[s],a,ps_d[s],r_d);
-	  } else{
-	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
-	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
-	  }
-	}
-      }
-      AXPYTimer.Stop();
-
-      PrecChangeTimer.Start();
-      precisionChange(p_f, p_d, wk_f_from_d); //get back single prec search direction for linop
-      PrecChangeTimer.Stop();
-
-      cp=c;
-      MatrixTimer.Start();  
-      Linop_f.HermOp(p_f,mmp_f); 
-      d=real(innerProduct(p_f,mmp_f));    
-      MatrixTimer.Stop();  
-
-      AXPYTimer.Start();
-      axpy(mmp_f,mass[0],p_f,mmp_f);
-      AXPYTimer.Stop();
-      RealD rn = norm2(p_f);
-      d += rn*mass[0];
-    
-      bp=b;
-      b=-cp/d;
-    
-      // Toggle the recurrence history
-      bs[0] = b;
-      iz = 1-iz;
-      ShiftTimer.Start();
-      for(int s=1;s<nshift;s++){
-	if((!converged[s])){
-	  RealD z0 = z[s][1-iz];
-	  RealD z1 = z[s][iz];
-	  z[s][iz] = z0*z1*bp
-	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
-	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
-	}
-      }
-      ShiftTimer.Stop();
-
-      //Update double precision solutions
-      AXPYTimer.Start();
-      for(int s=0;s<nshift;s++){
-	int ss = s;
-	if( (!converged[s]) ) { 
-	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
-	}
-      }
-
-      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
-      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
-      AXPYTimer.Stop();
-
-      c = c_f;
-
-      if(k % ReliableUpdateFreq == 0){
-	//Replace r with true residual
-	MatrixTimer.Start();  
-	Linop_d.HermOp(psi_d[0],mmp_d); 
-	MatrixTimer.Stop();  
-
-	AXPYTimer.Start();
-	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
-
-	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
-	AXPYTimer.Stop();
-
-	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
-	
-	PrecChangeTimer.Start();
-	precisionChange(r_f, r_d, wk_f_from_d);
-	PrecChangeTimer.Stop();
-	c = c_d;
-      }
-    
-      // Convergence checks
-      int all_converged = 1;
-      for(int s=0;s<nshift;s++){
-      
-	if ( (!converged[s]) ){
-	  IterationsToCompleteShift[s] = k;
-	
-	  RealD css  = c * z[s][iz]* z[s][iz];
-	
-	  if(css<rsq[s]){
-	    if ( ! converged[s] )
-	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
-	    converged[s]=1;
-	  } else {
-	    all_converged=0;
-	  }
-
-	}
-      }
-
-      if ( all_converged ){
-
-	SolverTimer.Stop();
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
-      
-	// Check answers 
-	for(int s=0; s < nshift; s++) { 
-	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
-	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
-	  axpy(r_d,-alpha[s],src_d,tmp_d);
-	  RealD rn = norm2(r_d);
-	  RealD cn = norm2(src_d);
-	  TrueResidualShift[s] = std::sqrt(rn/cn);
-	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
-
-	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
-	  if(rn >= rsq[s]){
-	    CleanupTimer.Start();
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
-
-	    //Setup linear operators for final cleanup
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
-					       
-	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
-	    cg(src_d, psi_d[s]);
-	    
-	    TrueResidualShift[s] = cg.TrueResidual;
-	    CleanupTimer.Stop();
-	  }
-	}
-
-	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
-	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
-
-	IterationsToComplete = k;	
-
-	return;
-      }
-
-   
-    }
-    // ugly hack
-    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/Deflation.h

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

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -44,7 +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,
@@ -156,7 +155,6 @@ public:
       _coarse_relax_tol(coarse_relax_tol)  
   {    };
 
-  //evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
   int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
     CoarseField v(B);
@@ -183,16 +181,8 @@ public:
     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;
@@ -211,13 +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;
   }
@@ -295,10 +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;
@@ -342,8 +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);
@@ -392,31 +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); 
+    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);
@@ -427,14 +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.cc

@@ -40,7 +40,7 @@ void MemoryManager::PrintBytes(void)
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 8, 16, 8, 16 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils

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

@@ -3,8 +3,13 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define dprintf(...)
+
+#define MAXLINE 512
+static char print_buffer [ MAXLINE ];
+
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+//#define dprintf(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -23,6 +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
@@ -104,15 +111,17 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    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;
   EntryErase(CpuPtr);
@@ -121,26 +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);
   
-  dprintf("MemoryManager: Evict(%llx) %llx\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);
-    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    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)
 {
@@ -150,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -165,7 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -191,6 +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);
@@ -202,6 +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);
@@ -212,13 +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;
     }
   }
 }
@@ -241,11 +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 %llx %llx : %lld %lld\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);
   }
@@ -280,6 +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);
@@ -292,28 +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);
   }
 
@@ -334,10 +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)
@@ -374,9 +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
@@ -430,20 +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;
     
@@ -453,13 +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) 
@@ -473,6 +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)
 {
@@ -489,8 +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_mpi3.cc

@@ -392,9 +392,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
     acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-    this->StencilSendToRecvFromComplete(list,dir);
-  }
+  //  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  //    this->StencilSendToRecvFromComplete(list,dir);
+  //  }
 
   return off_node_bytes;
 }

Grid/lattice/Lattice_base.h

@@ -129,7 +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);
@@ -152,7 +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);
@@ -174,7 +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);
@@ -245,7 +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);
@@ -288,7 +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));
@@ -303,7 +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

@@ -29,6 +29,19 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
+{
+  auto ff = localNorm2(f);
+  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
+  typedef typename vobj::tensor_reduced normtype;
+  typedef typename normtype::scalar_object scalar;
+  std::vector<scalar> sff;
+  sliceSum(ff,sff,mu);
+  for(int t=0;t<sff.size();t++){
+    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
+  }
+}
+
 template<class vobj> uint32_t crc(Lattice<vobj> & buf)
 {
   autoView( buf_v , buf, CpuRead);

Grid/lattice/Lattice_reduction.h

@@ -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);
 
@@ -127,7 +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);
@@ -136,25 +139,50 @@ 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);
 #endif  
 }
+template<class vobj>
+inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  return sumD_gpu_large(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+  Integer osites = arg.Grid()->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);
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif  
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu(&arg_v[0],osites);
+  auto ssum= sum_gpu_large(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
   auto ssum= sum_cpu(&arg_v[0],osites);
-#endif  
+#endif
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -210,11 +238,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   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, 1,{
 	auto x_l = left_v[ss];

Grid/lattice/Lattice_reduction_gpu.h

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

Grid/lattice/Lattice_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

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

Grid/lattice/Lattice_slice_gpu.h

@@ -0,0 +1,126 @@
+NAMESPACE_BEGIN(Grid);
+
+// If NOT CUDA or HIP -- we should provide
+// -- atomicAdd(float *,float)
+// -- atomicAdd(double *,double)
+// 
+// Augment CUDA with complex atomics
+#if !defined(GRID_HIP) || !defined(GRID_CUDA)
+inline void atomicAdd(float *acc,float elem)
+{
+  *acc += elem;
+}
+inline void atomicAdd(double *acc,double elem)
+{
+  *acc += elem;
+}
+#endif
+inline void atomicAdd(ComplexD *accum,ComplexD & elem)
+{
+  double *a_p = (double *)accum;
+  double *e_p = (double *)&elem;
+  for(int w=0;w<2;w++){
+    atomicAdd(&a_p[w],e_p[w]);
+  }
+}
+inline void atomicAdd(ComplexF *accum,ComplexF & elem)
+{
+  float *a_p = (float *)accum;
+  float *e_p = (float *)&elem;
+  for(int w=0;w<2;w++){
+    atomicAdd(&a_p[w],e_p[w]);
+  }
+}
+// Augment CUDA with vobj atomics
+template<class vobj> accelerator_inline void atomicAdd(vobj *accum, vobj & elem)
+{
+  typedef typename vobj::scalar_type scalar_type;
+  scalar_type *a_p= (scalar_type *)accum;
+  scalar_type *e_p= (scalar_type *)& elem;
+  for(int w=0;w<vobj::Nsimd();w++){
+    atomicAdd(&a_p[w],e_p[w]);
+  }
+}
+// Atomics based slice sum
+template<class vobj> inline void sliceSumGpu(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_object::scalar_type scalar_type;
+  GridBase  *grid = Data.Grid();
+  assert(grid!=NULL);
+
+  const int    Nd = grid->_ndimension;
+  const int Nsimd = grid->Nsimd();
+
+  assert(orthogdim >= 0);
+  assert(orthogdim < Nd);
+
+  int fd=grid->_fdimensions[orthogdim];
+  int ld=grid->_ldimensions[orthogdim];
+  int rd=grid->_rdimensions[orthogdim];
+
+  // Move to device memory and copy in / out
+  Vector<vobj> lvSum(rd); // will locally sum vectors first
+  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
+
+  result.resize(fd); // And then global sum to return the same vector to every node 
+  for(int r=0;r<rd;r++){
+    lvSum[r]=Zero();
+  }
+
+  int e1=    grid->_slice_nblock[orthogdim];
+  int e2=    grid->_slice_block [orthogdim];
+  int stride=grid->_slice_stride[orthogdim];
+
+  // sum over reduced dimension planes, breaking out orthog dir
+  // Parallel over orthog direction
+  autoView( Data_v, Data, AcceleratorRead);
+  auto lvSum_p=&lvSum[0];
+  int ostride = grid->_ostride[orthogdim]; 
+  accelerator_for( ree,rd*e1*e2,1, {
+    int b = ree%e2;
+    int re= ree/e2;
+    int n=re%e1;
+    int r=re/e1;
+    int so=r*ostride;
+    int ss=so+n*stride+b;
+    atomicAdd(&lvSum_p[r],Data_v[ss]);
+  });
+
+  // Sum across simd lanes in the plane, breaking out orthog dir.
+  Coordinate icoor(Nd);
+
+  for(int rt=0;rt<rd;rt++){
+
+    extract(lvSum[rt],extracted);
+
+    for(int idx=0;idx<Nsimd;idx++){
+
+      grid->iCoorFromIindex(icoor,idx);
+
+      int ldx =rt+icoor[orthogdim]*rd;
+
+      lsSum[ldx]=lsSum[ldx]+extracted[idx];
+
+    }
+  }
+  
+  // sum over nodes.
+  for(int t=0;t<fd;t++){
+    int pt = t/ld; // processor plane
+    int lt = t%ld;
+    if ( pt == grid->_processor_coor[orthogdim] ) {
+      result[t]=lsSum[lt];
+    } else {
+      result[t]=Zero();
+    }
+
+  }
+  scalar_type * ptr = (scalar_type *) &result[0];
+  int words = fd*sizeof(sobj)/sizeof(scalar_type);
+  grid->GlobalSumVector(ptr, words);
+}
+
+
+NAMESPACE_END(Grid);

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;
 
@@ -1080,95 +1080,53 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
-class precisionChangeWorkspace{
-  std::pair<Integer,Integer>* fmap_device; //device pointer
-public:
-  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid){
-    //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
-    assert(out_grid->Nd() == in_grid->Nd());
-    for(int d=0;d<out_grid->Nd();d++){
-      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
-    }
-    int Nsimd_out = out_grid->Nsimd();
-
-    std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
-    for(int lane=0; lane < out_grid->Nsimd(); lane++)
-      out_grid->iCoorFromIindex(out_icorrs[lane], lane);
-  
-    std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
-    thread_for(out_oidx,out_grid->oSites(),{
-	Coordinate out_ocorr; 
-	out_grid->oCoorFromOindex(out_ocorr, out_oidx);
-      
-	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
-	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
-	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
-	
-	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
-	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
-	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
-	  int in_oidx = 0, in_lane = 0;
-	  for(int d=0;d<in_grid->_ndimension;d++){
-	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
-	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
-	  }
-	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
-	}
-      });
-
-    //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
-    size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
-    fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
-    acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes); 
-  }
-
-  //Prevent moving or copying
-  precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
-  precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
-  precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
-  precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
-  
-  std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
-
-  ~precisionChangeWorkspace(){
-    acceleratorFreeDevice(fmap_device);
-  }
-};
-
-
-//Convert a lattice of one precision to another. The input workspace contains the mapping data.
-template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
-  static_assert( std::is_same<typename VobjOut::DoublePrecision, typename VobjIn::DoublePrecision>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
-
-  out.Checkerboard() = in.Checkerboard();
-  constexpr int Nsimd_out = VobjOut::Nsimd();
-
-  std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
-
-  //Do the copy/precision change
-  autoView( out_v , out, AcceleratorWrite);
-  autoView( in_v , in, AcceleratorRead);
-
-  accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
-      std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
-      for(int out_lane=0; out_lane < Nsimd_out; out_lane++){      
-	int in_oidx = fmap_osite[out_lane].first;
-	int in_lane = fmap_osite[out_lane].second;
-	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
-      }
-    });
-}
-
 //Convert a Lattice from one precision to another
-//Generate the workspace in place; if multiple calls with the same mapping are performed, consider pregenerating the workspace and reusing
 template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
-  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
-  precisionChange(out, in, workspace);
-}
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+{
+  assert(out.Grid()->Nd() == in.Grid()->Nd());
+  for(int d=0;d<out.Grid()->Nd();d++){
+    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
+  }
+  out.Checkerboard() = in.Checkerboard();
+  GridBase *in_grid=in.Grid();
+  GridBase *out_grid = out.Grid();
 
+  typedef typename VobjOut::scalar_object SobjOut;
+  typedef typename VobjIn::scalar_object SobjIn;
+
+  int ndim = out.Grid()->Nd();
+  int out_nsimd = out_grid->Nsimd();
+    
+  std::vector<Coordinate > out_icoor(out_nsimd);
+      
+  for(int lane=0; lane < out_nsimd; lane++){
+    out_icoor[lane].resize(ndim);
+    out_grid->iCoorFromIindex(out_icoor[lane], lane);
+  }
+        
+  std::vector<SobjOut> in_slex_conv(in_grid->lSites());
+  unvectorizeToLexOrdArray(in_slex_conv, in);
+    
+  autoView( out_v , out, CpuWrite);
+  thread_for(out_oidx,out_grid->oSites(),{
+    Coordinate out_ocoor(ndim);
+    out_grid->oCoorFromOindex(out_ocoor, out_oidx);
+
+    ExtractPointerArray<SobjOut> ptrs(out_nsimd);      
+
+    Coordinate lcoor(out_grid->Nd());
+      
+    for(int lane=0; lane < out_nsimd; lane++){
+      for(int mu=0;mu<ndim;mu++)
+	lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
+	
+      int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
+      ptrs[lane] = &in_slex_conv[llex];
+    }
+    merge(out_v[out_oidx], ptrs, 0);
+  });
+}
 
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids

Grid/log/Log.cc

@@ -65,32 +65,40 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
+GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
+GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
 GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(0);
+  GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
+  GridLogMemory.Active(0); 
+  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
+  GridLogDslash.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
   GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
+    if (logstreams[i] == std::string("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);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
-    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
+    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("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,7 +138,8 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      stream << log.colour() <<  std::left;
+      //      stream << log.colour() <<  std::left;
+      stream <<  std::left;
       if (log.chanWidth > 0)
       {
         stream << std::setw(log.chanWidth);
@@ -153,9 +154,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      stream << log.colour();
+      //      stream << log.colour();
+      stream <<  std::right;
       stream.flags(f);
-
       return stream;
     } else { 
       return devnull;
@@ -180,9 +181,12 @@ extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
 extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
+extern GridLogger GridLogDslash;
 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/IldgIO.h

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

Grid/parallelIO/MetaData.h

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

Grid/parallelIO/NerscIO.h

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

Grid/perfmon/PerfCount.cc

@@ -27,10 +27,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
-#include <Grid/perfmon/PerfCount.h>
 
+#include <Grid/perfmon/Timer.h>
+#include <Grid/perfmon/PerfCount.h>
 NAMESPACE_BEGIN(Grid);
 
+GridTimePoint theProgramStart = GridClock::now();
+
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
 #define RawConfig(A,B) (A<<8|B)
 const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::PerformanceCounterConfigs [] = {

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>
@@ -72,17 +78,9 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
-#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
-#define __SSC_STOP  __SSC_MARK(0x110)
-#define __SSC_START __SSC_MARK(0x111)
-
 
 #else
 
-#define __SSC_MARK(mark) 
-#define __SSC_STOP  
-#define __SSC_START 
-
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

@@ -35,17 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid)
 
-// Dress the output; use std::chrono
-// C++11 time facilities better?
-inline double usecond(void) {
-  struct timeval tv;
-#ifdef TIMERS_ON
-  gettimeofday(&tv,NULL);
-#endif
-  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
-}
-
-typedef  std::chrono::system_clock          GridClock;
+//typedef  std::chrono::system_clock          GridClock;
+typedef  std::chrono::high_resolution_clock   GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
 
 typedef  std::chrono::seconds               GridSecs;
@@ -53,6 +44,15 @@ typedef  std::chrono::milliseconds          GridMillisecs;
 typedef  std::chrono::microseconds          GridUsecs;
 typedef  std::chrono::microseconds          GridTime;
 
+extern GridTimePoint theProgramStart;
+// Dress the output; use std::chrono
+// C++11 time facilities better?
+inline double usecond(void) {
+  auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart); 
+  return 1.0*usecs.count();
+}
+
+
 inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
   stream << time.count()<<" s";

Grid/perfmon/Tracing.h

@@ -0,0 +1,70 @@
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+#ifdef GRID_TRACING_NVTX
+#include <nvToolsExt.h>
+class GridTracer {
+public:
+  GridTracer(const char* name) {
+    nvtxRangePushA(name);
+  }
+  ~GridTracer() {
+    nvtxRangePop();
+  }
+};
+inline void tracePush(const char *name) { nvtxRangePushA(name); }
+inline void tracePop(const char *name) { nvtxRangePop(); }
+inline int  traceStart(const char *name) {  }
+inline void traceStop(int ID) {  }
+#endif
+
+#ifdef GRID_TRACING_ROCTX
+#include <roctracer/roctx.h>
+class GridTracer {
+ public:
+  GridTracer(const char* name) {
+    roctxRangePushA(name);
+    std::cout << "roctxRangePush "<<name<<std::endl;
+  }
+  ~GridTracer() {
+    roctxRangePop();
+    std::cout << "roctxRangePop "<<std::endl;
+  }
+};
+inline void tracePush(const char *name) { roctxRangePushA(name); }
+inline void tracePop(const char *name) { roctxRangePop(); }
+inline int  traceStart(const char *name) { roctxRangeStart(name); }
+inline void traceStop(int ID) { roctxRangeStop(ID); }
+#endif
+
+#ifdef GRID_TRACING_TIMER
+class GridTracer {
+ public:
+  const char *name;
+  double elapsed;
+  GridTracer(const char* _name) {
+    name = _name;
+    elapsed=-usecond();
+  }
+  ~GridTracer() {
+    elapsed+=usecond();
+    std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl;
+  }
+};
+inline void tracePush(const char *name) {  }
+inline void tracePop(const char *name) {  }
+inline int  traceStart(const char *name) { return 0; }
+inline void traceStop(int ID) {  }
+#endif
+
+#ifdef GRID_TRACING_NONE
+#define GRID_TRACE(name) 
+inline void tracePush(const char *name) {  }
+inline void tracePop(const char *name) {  }
+inline int  traceStart(const char *name) { return 0;  }
+inline void traceStop(int ID) {  }
+#else
+#define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
+#endif
+NAMESPACE_END(Grid);

Grid/pugixml/pugixml.cc

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

Grid/qcd/QCD.h

@@ -63,7 +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.
@@ -476,9 +451,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
+#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuWrite);
+  autoView(f_v,f,CpuRead);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -490,12 +476,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
+#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuRead);
+  autoView(f_v,f,CpuWrite);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -507,6 +504,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/ActionParams.h

@@ -36,8 +36,7 @@ NAMESPACE_BEGIN(Grid);
 
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
-  Coordinate twists; //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
-                     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+  Coordinate twists;
   GparityWilsonImplParams() : twists(Nd, 0) {};
 };
   
@@ -66,8 +65,7 @@ struct StaggeredImplParams {
 				    RealD, tolerance, 
 				    int,   degree, 
 				    int,   precision,
-				    int,   BoundsCheckFreq,
-				    RealD, BoundsCheckTol);
+				    int,   BoundsCheckFreq);
     
   // MaxIter and tolerance, vectors??
     
@@ -78,61 +76,15 @@ struct StaggeredImplParams {
 				RealD tol      = 1.0e-8, 
                            	int _degree    = 10,
 				int _precision = 64,
-				int _BoundsCheckFreq=20,
-				double _BoundsCheckTol=1e-6)
+				int _BoundsCheckFreq=20)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
 	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);
 

Grid/qcd/action/fermion/CayleyFermion5D.h

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

Grid/qcd/action/fermion/CloverHelpers.h

@@ -0,0 +1,333 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
+
+    Copyright (C) 2017 - 2022
+
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Mattia Bruno <mattia.bruno@cern.ch>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#pragma once
+
+#include <Grid/Grid.h>
+#include <Grid/qcd/spin/Dirac.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+
+////////////////////////////////////////////
+// Standard Clover
+//   (4+m0) + csw * clover_term
+// Exp Clover
+//   (4+m0) * exp(csw/(4+m0) clover_term)
+//   = (4+m0) + csw * clover_term + ...
+////////////////////////////////////////////
+
+NAMESPACE_BEGIN(Grid);
+
+
+//////////////////////////////////
+// Generic Standard Clover
+//////////////////////////////////
+
+template<class Impl>
+class CloverHelpers: public WilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+
+  typedef WilsonCloverHelpers<Impl> Helpers;
+
+  static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
+    GridBase *grid = CloverTerm.Grid();
+    CloverTerm += diag_mass;
+
+    int lvol = grid->lSites();
+    int DimRep = Impl::Dimension;
+    {
+      autoView(CTv,CloverTerm,CpuRead);
+      autoView(CTIv,CloverTermInv,CpuWrite);
+      thread_for(site, lvol, {
+        Coordinate lcoor;
+        grid->LocalIndexToLocalCoor(site, lcoor);
+        Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+        Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+        typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
+        peekLocalSite(Qx, CTv, lcoor);
+
+        for (int j = 0; j < Ns; j++)
+          for (int k = 0; k < Ns; k++)
+            for (int a = 0; a < DimRep; a++)
+              for (int b = 0; b < DimRep; b++){
+                auto zz =  Qx()(j, k)(a, b);
+                EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+              }
+
+        EigenInvCloverOp = EigenCloverOp.inverse();
+        for (int j = 0; j < Ns; j++)
+          for (int k = 0; k < Ns; k++)
+            for (int a = 0; a < DimRep; a++)
+              for (int b = 0; b < DimRep; b++)
+                Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+               pokeLocalSite(Qxinv, CTIv, lcoor);
+      });
+    }
+  }
+
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    return Helpers::Cmunu(U, lambda, mu, nu);
+  }
+
+};
+
+
+//////////////////////////////////
+// Generic Exp Clover
+//////////////////////////////////
+
+template<class Impl>
+class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+
+  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+  typedef WilsonCloverHelpers<Impl> Helpers;
+
+  // Can this be avoided?
+  static void IdentityTimesC(const CloverField& in, RealD c) {
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
+  }
+
+  static int getNMAX(RealD prec, RealD R) {
+    /* compute stop condition for exponential */
+    int NMAX=1;
+    RealD cond=R*R/2.;
+
+    while (cond*std::exp(R)>prec) {
+      NMAX++;
+      cond*=R/(double)(NMAX+1);
+    }
+    return NMAX;
+  }
+
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+
+  static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
+
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
+
+    Clover *= (1.0/diag_mass);
+
+    // Taylor expansion, slow but generic
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+    // qN = cN
+    // qn = cn + qn+1 X
+    std::vector<RealD> cn(NMAX+1);
+    cn[0] = 1.0;
+    for (int i=1; i<=NMAX; i++)
+      cn[i] = cn[i-1] / RealD(i);
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
+
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
+
+    Clover = ExpClover * diag_mass;
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
+
+    CloverInv = ExpClover * (1.0/diag_mass);
+
+  }
+
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    assert(0);
+    return lambda;
+  }
+
+};
+
+
+//////////////////////////////////
+// Compact Standard Clover
+//////////////////////////////////
+
+
+template<class Impl>
+class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
+                            public WilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  typedef WilsonCloverHelpers<Impl> Helpers;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+    Clover += diag_mass;
+  }
+
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
+
+    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+  }
+
+  // TODO: implement Cmunu for better performances with compact layout, but don't do it
+  // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    return Helpers::Cmunu(U, lambda, mu, nu);
+  }
+};
+
+//////////////////////////////////
+// Compact Exp Clover
+//////////////////////////////////
+
+template<class Impl>
+class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  // Can this be avoided?
+  static void IdentityTimesC(const CloverField& in, RealD c) {
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
+  }
+
+  static int getNMAX(RealD prec, RealD R) {
+    /* compute stop condition for exponential */
+    int NMAX=1;
+    RealD cond=R*R/2.;
+
+    while (cond*std::exp(R)>prec) {
+      NMAX++;
+      cond*=R/(double)(NMAX+1);
+    }
+    return NMAX;
+  }
+
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
+
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
+
+    Clover *= (1.0/diag_mass);
+
+    // Taylor expansion, slow but generic
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+    // qN = cN
+    // qn = cn + qn+1 X
+    std::vector<RealD> cn(NMAX+1);
+    cn[0] = 1.0;
+    for (int i=1; i<=NMAX; i++)
+      cn[i] = cn[i-1] / RealD(i);
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
+
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
+
+    Clover = ExpClover * diag_mass;
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
+
+    CloverInv = ExpClover * (1.0/diag_mass);
+
+  }
+
+  static 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);
+    return lambda;
+  }
+
+};
+
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -31,6 +31,7 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -85,7 +86,7 @@ NAMESPACE_BEGIN(Grid);
 //           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
 //                                                                 =  84 complex words per site
 
-template<class Impl>
+template<class Impl, class CloverHelpers>
 class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
                                    public WilsonCloverHelpers<Impl>,
                                    public CompactWilsonCloverHelpers<Impl> {
@@ -224,7 +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/Fermion.h

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

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/WilsonCloverFermion.h

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

Grid/qcd/action/fermion/WilsonCloverHelpers.h

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

Grid/qcd/action/fermion/WilsonCloverTypes.h

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

Grid/qcd/action/fermion/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

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

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

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

@@ -32,22 +32,23 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 
+
 NAMESPACE_BEGIN(Grid);
-template<class Impl>
-CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
-                                                             GridCartesian& Fgrid,
-                                                             GridRedBlackCartesian& Hgrid,
-                                                             const RealD _mass,
-                                                             const RealD _csw_r,
-                                                             const RealD _csw_t,
-                                                             const RealD _cF,
-                                                             const WilsonAnisotropyCoefficients& clover_anisotropy,
-                                                             const ImplParams& impl_p)
+template<class Impl, class CloverHelpers>
+CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
+                                                                            GridCartesian& Fgrid,
+                                                                            GridRedBlackCartesian& Hgrid,
+                                                                            const RealD _mass,
+                                                                            const RealD _csw_r,
+                                                                            const RealD _csw_t,
+                                                                            const RealD _cF,
+                                                                            const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                                                            const ImplParams& impl_p)
   : WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
   , 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)
@@ -58,80 +59,85 @@ CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
   , BoundaryMask(&Fgrid)
   , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
 {
+  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
+
   csw_r *= 0.5;
   csw_t *= 0.5;
   if (clover_anisotropy.isAnisotropic)
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries)
+  if (fixedBoundaries) {
+    this->BoundaryMaskEven.Checkerboard() = Even;
+    this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
+  }
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
+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>
-void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+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>
-void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+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>
-void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+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>
-void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& 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);
   }
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   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>
-void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   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>
-void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& 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>
-void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& 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>
-void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
@@ -141,16 +147,16 @@ void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionFiel
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
   Mooee(in, out); // blocks are hermitian
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
@@ -160,27 +166,27 @@ void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionF
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
   MooeeInv(in, out); // blocks are hermitian
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
   DhopDir(in, out, dir, disp);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
   DhopDirAll(in, out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!open_boundaries); // TODO check for changes required for open bc
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+  assert(!fixedBoundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -251,7 +257,7 @@ void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionFi
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);   // checked
+      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
       count++;
     }
 
@@ -261,18 +267,18 @@ void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionFi
   force += clover_force;
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&        in,
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
                     FermionField&              out,
                     const CloverDiagonalField& diagonal,
                     const CloverTriangleField& triangle) {
@@ -285,8 +291,8 @@ void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&
   CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
   // NOTE: parts copied from original implementation
 
   // Import gauge into base class
@@ -299,6 +305,7 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   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();
@@ -318,22 +325,30 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  TmpOriginal += this->diag_mass;
+
+  // 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 t4 = usecond();
+  double t5 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Possible modify the boundary values
-  double t5 = usecond();
-  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
-
-  // Invert the clover term in the improved layout
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
   double t6 = usecond();
-  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+
+  // 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();
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
 
   // Fill the remaining clover fields
-  double t7 = usecond();
+  double t8 = usecond();
   pickCheckerboard(Even, DiagonalEven,    Diagonal);
   pickCheckerboard(Even, TriangleEven,    Triangle);
   pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
@@ -344,20 +359,19 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
 
   // Report timings
-  double t8 = usecond();
-#if 0
-  std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
-            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
-            << ", allocations = "               << (t2 - t1) / 1e6
-            << ", field strength = "            << (t3 - t2) / 1e6
-            << ", fill clover = "               << (t4 - t3) / 1e6
-            << ", convert = "                   << (t5 - t4) / 1e6
-            << ", boundaries = "                << (t6 - t5) / 1e6
-            << ", inversions = "                << (t7 - t6) / 1e6
-            << ", pick cbs = "                  << (t8 - t7) / 1e6
-            << ", total = "                     << (t8 - t0) / 1e6
-            << std::endl;
-#endif
+  double t9 = usecond();
+
+  std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
+  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "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;
 }
 
 NAMESPACE_END(Grid);

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

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

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

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

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

@@ -498,6 +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;}
@@ -505,11 +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/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master

@@ -9,6 +9,7 @@
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Mattia Bruno <mattia.bruno@cern.ch>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -32,10 +33,12 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class CompactWilsonCloverFermion<IMPLEMENTATION>; 
+template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>; 
+template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>; 
 
 NAMESPACE_END(Grid);

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -8,7 +8,8 @@
 
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
-
+    Author: Mattia Bruno <mattia.bruno@cern.ch>
+    
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
     the Free Software Foundation; either version 2 of the License, or
@@ -31,10 +32,12 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class WilsonCloverFermion<IMPLEMENTATION>; 
+template class WilsonCloverFermion<IMPLEMENTATION, CloverHelpers<IMPLEMENTATION>>; 
+template class WilsonCloverFermion<IMPLEMENTATION, ExpCloverHelpers<IMPLEMENTATION>>; 
 
 NAMESPACE_END(Grid);

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

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

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

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

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

Grid/qcd/action/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

@@ -40,66 +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/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] ); }
@@ -195,16 +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);
+          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];
@@ -223,16 +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]);
+          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];
@@ -242,119 +205,49 @@ NAMESPACE_BEGIN(Grid);
         Phi = Phi + tmp[1];
 
         // Reset shift coefficients for energy and force evals
-        //Lop.RefreshShiftCoefficients(0.0);
-        //Rop.RefreshShiftCoefficients(-1.0);
-	heatbathRefreshShiftCoefficients(0, 0.0);
-	heatbathRefreshShiftCoefficients(1, -1.0);
-
-	//Mark that the next call to S is the first after refresh
-	initial_action = true;
-
+        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)
       {
@@ -378,7 +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]);
@@ -388,26 +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;
       };
 
@@ -456,40 +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

@@ -40,31 +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)){}
 
-      virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
+      OneFlavourEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
+					    FermionOperator<Impl>  &_DenOp, 
+					    Params & p
+					    ) : 
+      NumOp(_NumOp), 
+      DenOp(_DenOp), 
+      PhiOdd (_NumOp.FermionRedBlackGrid()),
+      PhiEven(_NumOp.FermionRedBlackGrid()),
+      param(p) 
+      {
+	AlgRemez remez(param.lo,param.hi,param.precision);
+
+	// MdagM^(+- 1/2)
+	std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/2)"<<std::endl;
+	remez.generateApprox(param.degree,1,2);
+	PowerHalf.Init(remez,param.tolerance,false);
+	PowerNegHalf.Init(remez,param.tolerance,true);
+
+	// MdagM^(+- 1/4)
+	std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/4)"<<std::endl;
+	remez.generateApprox(param.degree,1,4);
+   	PowerQuarter.Init(remez,param.tolerance,false);
+	PowerNegQuarter.Init(remez,param.tolerance,true);
+      };
+
+      virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}
+
+      virtual std::string LogParameters(){
+	std::stringstream sstream;
+	sstream << GridLogMessage << "["<<action_name()<<"] Low            :" << param.lo <<  std::endl;
+	sstream << GridLogMessage << "["<<action_name()<<"] High           :" << param.hi <<  std::endl;
+	sstream << GridLogMessage << "["<<action_name()<<"] Max iterations :" << param.MaxIter <<  std::endl;
+	sstream << GridLogMessage << "["<<action_name()<<"] Tolerance      :" << param.tolerance <<  std::endl;
+	sstream << GridLogMessage << "["<<action_name()<<"] Degree         :" << param.degree <<  std::endl;
+	sstream << GridLogMessage << "["<<action_name()<<"] Precision      :" << param.precision <<  std::endl;
+	return sstream.str();
+      }
+      
+      
+      virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
+
+	// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi       
+	//
+	// P(phi) = e^{- phi^dag (VdagV)^1/4 (MdagM)^-1/2 (VdagV)^1/4 phi}
+	//        = e^{- phi^dag  (VdagV)^1/4 (MdagM)^-1/4 (MdagM)^-1/4  (VdagV)^1/4 phi}
+	//
+	// Phi =  (VdagV)^-1/4 Mdag^{1/4} eta 
+	//
+	// P(eta) = e^{- eta^dag eta}
+	//
+	// e^{x^2/2 sig^2} => sig^2 = 0.5.
+	// 
+	// So eta should be of width sig = 1/sqrt(2).
+
+	RealD scale = std::sqrt(0.5);
+
+	FermionField eta(NumOp.FermionGrid());
+	FermionField etaOdd (NumOp.FermionRedBlackGrid());
+	FermionField etaEven(NumOp.FermionRedBlackGrid());
+	FermionField     tmp(NumOp.FermionRedBlackGrid());
+
+	gaussian(pRNG,eta);	eta=eta*scale;
+
+	pickCheckerboard(Even,etaEven,eta);
+	pickCheckerboard(Odd,etaOdd,eta);
+
+	NumOp.ImportGauge(U);
+	DenOp.ImportGauge(U);
+
+
+	// MdagM^1/4 eta
+	SchurDifferentiableOperator<Impl> MdagM(DenOp);
+	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerQuarter);
+	msCG_M(MdagM,etaOdd,tmp);
+
+	// VdagV^-1/4 MdagM^1/4 eta
+	SchurDifferentiableOperator<Impl> VdagV(NumOp);
+	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerNegQuarter);
+	msCG_V(VdagV,tmp,PhiOdd);
+
+	assert(NumOp.ConstEE() == 1);
+	assert(DenOp.ConstEE() == 1);
+	PhiEven = Zero();
+	
+      };
+
+      //////////////////////////////////////////////////////
+      // S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi       
+      //////////////////////////////////////////////////////
+      virtual RealD S(const GaugeField &U) {
+
+	NumOp.ImportGauge(U);
+	DenOp.ImportGauge(U);
+
+	FermionField X(NumOp.FermionRedBlackGrid());
+	FermionField Y(NumOp.FermionRedBlackGrid());
+
+	// VdagV^1/4 Phi
+	SchurDifferentiableOperator<Impl> VdagV(NumOp);
+	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
+	msCG_V(VdagV,PhiOdd,X);
+
+	// MdagM^-1/4 VdagV^1/4 Phi
+	SchurDifferentiableOperator<Impl> MdagM(DenOp);
+	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegQuarter);
+	msCG_M(MdagM,X,Y);
+
+	// Randomly apply rational bounds checks.
+	auto grid = NumOp.FermionGrid();
+        auto r=rand();
+        grid->Broadcast(0,r);
+        if ( (r%param.BoundsCheckFreq)==0 ) { 
+	  FermionField gauss(NumOp.FermionRedBlackGrid());
+	  gauss = PhiOdd;
+	  HighBoundCheck(MdagM,gauss,param.hi);
+	  InverseSqrtBoundsCheck(param.MaxIter,param.tolerance*100,MdagM,gauss,PowerNegHalf);
+	}
+
+	//  Phidag VdagV^1/4 MdagM^-1/4  MdagM^-1/4 VdagV^1/4 Phi
+	RealD action = norm2(Y);
+
+	return action;
+      };
+
+      // S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi       
+      //
+      // Here, M is some 5D operator and V is the Pauli-Villars field
+      // N and D makeup the rat. poly of the M term and P and & makeup the rat.poly of the denom term
+      //
+      // Need  
+      // dS_f/dU =  chi^dag d[P/Q]  N/D   P/Q  chi 
+      //         +  chi^dag   P/Q d[N/D]  P/Q  chi 
+      //         +  chi^dag   P/Q   N/D d[P/Q] chi 
+      //
+      // P/Q is expressed as partial fraction expansion: 
+      // 
+      //           a0 + \sum_k ak/(V^dagV + bk) 
+      //  
+      // d[P/Q] is then  
+      //
+      //          \sum_k -ak [V^dagV+bk]^{-1}  [ dV^dag V + V^dag dV ] [V^dag V + bk]^{-1} 
+      //  
+      // and similar for N/D. 
+      // 
+      // Need   
+      //       MpvPhi_k   = [Vdag V + bk]^{-1} chi  
+      //       MpvPhi     = {a0 +  \sum_k ak [Vdag V + bk]^{-1} }chi   
+      //   
+      //       MfMpvPhi_k = [MdagM+bk]^{-1} MpvPhi  
+      //       MfMpvPhi   = {a0 +  \sum_k ak [Mdag M + bk]^{-1} } MpvPhi
+      // 
+      //       MpvMfMpvPhi_k = [Vdag V + bk]^{-1} MfMpvchi   
+      //  
+
+      virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
+
+	const int n_f  = PowerNegHalf.poles.size();
+	const int n_pv = PowerQuarter.poles.size();
+
+	std::vector<FermionField> MpvPhi_k     (n_pv,NumOp.FermionRedBlackGrid());
+	std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionRedBlackGrid());
+	std::vector<FermionField> MfMpvPhi_k   (n_f ,NumOp.FermionRedBlackGrid());
+
+	FermionField      MpvPhi(NumOp.FermionRedBlackGrid());
+	FermionField    MfMpvPhi(NumOp.FermionRedBlackGrid());
+	FermionField MpvMfMpvPhi(NumOp.FermionRedBlackGrid());
+	FermionField           Y(NumOp.FermionRedBlackGrid());
+
+	GaugeField   tmp(NumOp.GaugeGrid());
+
+	NumOp.ImportGauge(U);
+	DenOp.ImportGauge(U);
+
+	SchurDifferentiableOperator<Impl> VdagV(NumOp);
+	SchurDifferentiableOperator<Impl> MdagM(DenOp);
+
+	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
+	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
+
+	msCG_V(VdagV,PhiOdd,MpvPhi_k,MpvPhi);
+	msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
+	msCG_V(VdagV,MfMpvPhi,MpvMfMpvPhi_k,MpvMfMpvPhi);
+
+	RealD ak;
+
+	dSdU = Zero();
+
+	// With these building blocks  
+	//  
+	//       dS/dU = 
+	//                 \sum_k -ak MfMpvPhi_k^dag      [ dM^dag M + M^dag dM ] MfMpvPhi_k         (1)
+	//             +   \sum_k -ak MpvMfMpvPhi_k^\dag  [ dV^dag V + V^dag dV ] MpvPhi_k           (2)
+	//                        -ak MpvPhi_k^dag        [ dV^dag V + V^dag dV ] MpvMfMpvPhi_k      (3)
+
+	//(1)
+	for(int k=0;k<n_f;k++){
+	  ak = PowerNegHalf.residues[k];
+	  MdagM.Mpc(MfMpvPhi_k[k],Y);
+	  MdagM.MpcDagDeriv(tmp , MfMpvPhi_k[k], Y );  dSdU=dSdU+ak*tmp;
+	  MdagM.MpcDeriv(tmp , Y, MfMpvPhi_k[k] );  dSdU=dSdU+ak*tmp;
+	}
+	
+	//(2)
+	//(3)
+	for(int k=0;k<n_pv;k++){
+
+          ak = PowerQuarter.residues[k];
+	  
+	  VdagV.Mpc(MpvPhi_k[k],Y);
+	  VdagV.MpcDagDeriv(tmp,MpvMfMpvPhi_k[k],Y); dSdU=dSdU+ak*tmp;
+	  VdagV.MpcDeriv   (tmp,Y,MpvMfMpvPhi_k[k]);  dSdU=dSdU+ak*tmp;     
+	  
+	  VdagV.Mpc(MpvMfMpvPhi_k[k],Y);                // V as we take Ydag 
+	  VdagV.MpcDeriv   (tmp,Y, MpvPhi_k[k]); dSdU=dSdU+ak*tmp;
+	  VdagV.MpcDagDeriv(tmp,MpvPhi_k[k], Y); dSdU=dSdU+ak*tmp;
+
+	}
+
+	//dSdU = Ta(dSdU);
+
+      };
     };
 
 NAMESPACE_END(Grid);

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

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

Grid/qcd/gparity/Gparity.h

@@ -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
@@ -151,40 +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);
 
     Smearing.set_Field(U);
 

Grid/qcd/hmc/HMC.h

@@ -115,21 +115,21 @@ private:
 
     random(sRNG, rn_test);
 
-    std::cout << GridLogHMC
+    std::cout << GridLogMessage
               << "--------------------------------------------------\n";
-    std::cout << GridLogHMC << "exp(-dH) = " << prob
+    std::cout << GridLogMessage << "exp(-dH) = " << prob
               << "  Random = " << rn_test << "\n";
-    std::cout << GridLogHMC
+    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
+      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
+      std::cout << GridLogMessage << "Metropolis_test -- REJECTED\n";
+      std::cout << GridLogMessage
                 << "--------------------------------------------------\n";
       return false;
     }
@@ -145,7 +145,7 @@ private:
 
     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);
 
     TheIntegrator.integrate(U);
@@ -165,7 +165,7 @@ private:
 
 
     std::cout.precision(15);
-    std::cout << GridLogHMC << "Total H after trajectory  = " << H1
+    std::cout << GridLogMessage << "Total H after trajectory  = " << H1
 	      << "  dH = " << H1 - H0 << "\n";
     std::cout.precision(current_precision);
     
@@ -196,9 +196,9 @@ 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();
@@ -207,10 +207,10 @@ 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)
@@ -219,7 +219,7 @@ public:
      
       
       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;
 
 
       for (int obs = 0; obs < Observables.size(); obs++) {
@@ -228,7 +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

@@ -226,9 +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

@@ -136,14 +136,8 @@ protected:
       if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
-
-      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
-      Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      Real max_force_abs = std::sqrt(maxLocalNorm2(force));
-      Real max_impulse_abs = max_force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << " Max force: " << max_force_abs << " Time step: " << ep << " Impulse average: " << impulse_abs << " Max impulse: " << max_impulse_abs << std::endl;
+      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;
@@ -255,19 +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
@@ -277,7 +267,6 @@ 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
-	std::cout << GridLogIntegrator << "Refreshing integrator level " << level << " index " << actionID << std::endl;
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
       }

Grid/qcd/observables/topological_charge.h

@@ -99,7 +99,7 @@ public:
 	// using wilson flow by default here
 	WilsonFlow<PeriodicGimplR> WF(Pars.Smearing.steps, Pars.Smearing.step_size, Pars.Smearing.meas_interval);
 	WF.smear_adaptive(Usmear, U, Pars.Smearing.maxTau);
-	Real T0   = WF.energyDensityPlaquette(Pars.Smearing.maxTau, Usmear);
+	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
@@ -34,44 +33,28 @@ NAMESPACE_BEGIN(Grid);
 
 template <class Gimpl>
 class WilsonFlow: public Smear<Gimpl>{
-public:
-  //Store generic measurements to take during smearing process using std::function
-  typedef std::function<void(int, RealD, const typename Gimpl::GaugeField &)> FunctionType;  //int: step,  RealD: flow time,  GaugeField : the gauge field
-  
-private:
   unsigned int Nstep;
-  RealD epsilon; //for regular smearing this is the time step, for adaptive it is the initial time step
- 
-  std::vector< std::pair<int, FunctionType> > functions; //The int maps to the measurement frequency
+  unsigned int measure_interval;
+  mutable RealD epsilon, taus;
+
 
   mutable WilsonGaugeAction<Gimpl> SG;
 
-  //Evolve the gauge field by 1 step and update tau
-  void evolve_step(typename Gimpl::GaugeField &U, RealD &tau) const;
-  //Evolve the gauge field by 1 step and update tau and the current time step eps
-  void evolve_step_adaptive(typename Gimpl::GaugeField&U, RealD &tau, RealD &eps, RealD maxTau) const;
+  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)
 
-  void resetActions(){ functions.clear(); }
-
-  void addMeasurement(int meas_interval, FunctionType meas){ functions.push_back({meas_interval, meas}); }
-
-  //Set the class to perform the default measurements: 
-  //the plaquette energy density every step
-  //the plaquette topological charge every 'topq_meas_interval' steps
-  //and output to stdout
-  void setDefaultMeasurements(int topq_meas_interval = 1);
-
   explicit WilsonFlow(unsigned int Nstep, RealD epsilon, unsigned int interval = 1):
   Nstep(Nstep),
     epsilon(epsilon),
+    measure_interval(interval),
     SG(WilsonGaugeAction<Gimpl>(3.0)) {
     // WilsonGaugeAction with beta 3.0
     assert(epsilon > 0.0);
     LogMessage();
-    setDefaultMeasurements(interval);
   }
 
   void LogMessage() {
@@ -90,29 +73,9 @@ public:
     // undefined for WilsonFlow
   }
 
-  void smear_adaptive(GaugeField&, const GaugeField&, RealD maxTau) const;
-
-  //Compute t^2 <E(t)> for time t from the plaquette
-  static RealD energyDensityPlaquette(const RealD t, const GaugeField& U);
-
-  //Compute t^2 <E(t)> for time t from the 1x1 cloverleaf form
-  //t is the Wilson flow time
-  static RealD energyDensityCloverleaf(const RealD t, const GaugeField& U);
-  
-  //Evolve the gauge field by Nstep steps of epsilon and return the energy density computed every interval steps
-  //The smeared field is output as V
-  std::vector<RealD> flowMeasureEnergyDensityPlaquette(GaugeField &V, const GaugeField& U, int measure_interval = 1);
-
-  //Version that does not return the smeared field
-  std::vector<RealD> flowMeasureEnergyDensityPlaquette(const GaugeField& U, int measure_interval = 1);
-
-
-  //Evolve the gauge field by Nstep steps of epsilon and return the Cloverleaf energy density computed every interval steps
-  //The smeared field is output as V
-  std::vector<RealD> flowMeasureEnergyDensityCloverleaf(GaugeField &V, const GaugeField& U, int measure_interval = 1);
-
-  //Version that does not return the smeared field
-  std::vector<RealD> flowMeasureEnergyDensityCloverleaf(const GaugeField& U, int measure_interval = 1);
+  void smear_adaptive(GaugeField&, const GaugeField&, RealD maxTau);
+  RealD energyDensityPlaquette(unsigned int step, const GaugeField& U) const;
+  RealD energyDensityPlaquette(const GaugeField& U) const;
 };
 
 
@@ -120,7 +83,7 @@ public:
 // Implementations
 ////////////////////////////////////////////////////////////////////////////////
 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());
   SG.deriv(U, Z);
@@ -136,13 +99,12 @@ void WilsonFlow<Gimpl>::evolve_step(typename Gimpl::GaugeField &U, RealD &tau) c
   SG.deriv(U, tmp); Z += tmp;                 // 4/3*(17/36*Z0 -8/9*Z1) +Z2
   Z *= 3.0/4.0;                               // Z = 17/36*Z0 -8/9*Z1 +3/4*Z2
   Gimpl::update_field(Z, U, -2.0*epsilon);    // V(t+e) = exp(ep*Z)*W2
-  tau += epsilon;
 }
 
 template <class Gimpl>
-void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, RealD &tau, RealD &eps, RealD maxTau) const{
-  if (maxTau - tau < eps){
-    eps = maxTau-tau;
+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());
@@ -152,151 +114,95 @@ void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, Real
   SG.deriv(U, Z);
   Zprime = -Z;
   Z *= 0.25;                                  // Z0 = 1/4 * F(U)
-  Gimpl::update_field(Z, U, -2.0*eps);    // U = W1 = exp(ep*Z0)*W0
+  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*eps);    // U_= W2 = exp(ep*Z)*W1
+  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*eps);    // V(t+e) = exp(ep*Z)*W2
+  Gimpl::update_field(Z, U, -2.0*epsilon);    // V(t+e) = exp(ep*Z)*W2
 
   // Ramos 
-  Gimpl::update_field(Zprime, Uprime, -2.0*eps); // V'(t+e) = exp(ep*Z')*W0
+  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
     
-  tau += eps;
+  taus += epsilon;
   //std::cout << GridLogMessage << "Adjusting integration step with distance: " << diff << std::endl;
     
-  eps = eps*0.95*std::pow(1e-4/diff,1./3.);
+  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(const RealD t, const GaugeField& U){
-  static WilsonGaugeAction<Gimpl> SG(3.0);
-  return 2.0 * t * t * SG.S(U)/U.Grid()->gSites();
-}
-
-//Compute t^2 <E(t)> for time from the 1x1 cloverleaf form
-template <class Gimpl>
-RealD WilsonFlow<Gimpl>::energyDensityCloverleaf(const RealD t, const GaugeField& U){
-  typedef typename Gimpl::GaugeLinkField GaugeMat;
-  typedef typename Gimpl::GaugeField GaugeLorentz;
-
-  assert(Nd == 4);
-  //E = 1/2 tr( F_munu F_munu )
-  //However as  F_numu = -F_munu, only need to sum the trace of the squares of the following 6 field strengths:
-  //F_01 F_02 F_03   F_12 F_13  F_23
-  GaugeMat F(U.Grid());
-  LatticeComplexD R(U.Grid());
-  R = Zero();
-  
-  for(int mu=0;mu<3;mu++){
-    for(int nu=mu+1;nu<4;nu++){
-      WilsonLoops<Gimpl>::FieldStrength(F, U, mu, nu);
-      R = R + trace(F*F);
-    }
-  }
-  ComplexD out = sum(R);
-  out = t*t*out / RealD(U.Grid()->gSites());
-  return -real(out); //minus sign necessary for +ve energy
-}
-
-
-template <class Gimpl>
-std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityPlaquette(GaugeField &V, const GaugeField& U, int measure_interval){
-  std::vector<RealD> out;
-  resetActions();
-  addMeasurement(measure_interval, [&out](int step, RealD t, const typename Gimpl::GaugeField &U){ 
-      std::cout << GridLogMessage << "[WilsonFlow] Computing plaquette energy density for step " << step << std::endl;
-      out.push_back( energyDensityPlaquette(t,U) );
-    });      
-  smear(V,U);
-  return out;
+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>
-std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityPlaquette(const GaugeField& U, int measure_interval){
-  GaugeField V(U);
-  return flowMeasureEnergyDensityPlaquette(V,U, measure_interval);
+RealD WilsonFlow<Gimpl>::energyDensityPlaquette(const GaugeField& U) const {
+  return 2.0 * taus * taus * SG.S(U)/U.Grid()->gSites();
 }
 
-template <class Gimpl>
-std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityCloverleaf(GaugeField &V, const GaugeField& U, int measure_interval){
-  std::vector<RealD> out;
-  resetActions();
-  addMeasurement(measure_interval, [&out](int step, RealD t, const typename Gimpl::GaugeField &U){ 
-      std::cout << GridLogMessage << "[WilsonFlow] Computing Cloverleaf energy density for step " << step << std::endl;
-      out.push_back( energyDensityCloverleaf(t,U) );
-    });      
-  smear(V,U);
-  return out;
-}
-
-template <class Gimpl>
-std::vector<RealD> WilsonFlow<Gimpl>::flowMeasureEnergyDensityCloverleaf(const GaugeField& U, int measure_interval){
-  GaugeField V(U);
-  return flowMeasureEnergyDensityCloverleaf(V,U, measure_interval);
-}
-
-
 
 //#define WF_TIMING 
+
+
+
 template <class Gimpl>
-void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const{
+void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const {
   out = in;
-  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 : 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>
-void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, RealD maxTau) const{
+void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, RealD maxTau){
   out = in;
-  RealD taus = 0.;
-  RealD eps = epsilon;
+  taus = epsilon;
   unsigned int step = 0;
   do{
     step++;
     //std::cout << GridLogMessage << "Evolution time :"<< taus << std::endl;
-    evolve_step_adaptive(out, taus, eps, maxTau);
-    //Perform measurements
-    for(auto const &meas : functions)
-      if( step % meas.first == 0 ) meas.second(step,taus,out);
+    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);
-}
 
-template <class Gimpl>
-void WilsonFlow<Gimpl>::setDefaultMeasurements(int topq_meas_interval){
-  addMeasurement(1, [](int step, RealD t, const typename Gimpl::GaugeField &U){
-      std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "  << step << "  " << t << "  " << energyDensityPlaquette(t,U) << std::endl;
-    });
-  addMeasurement(topq_meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
-      std::cout << GridLogMessage << "[WilsonFlow] Top. charge           : "  << step << "  " << WilsonLoops<Gimpl>::TopologicalCharge(U) << std::endl;
-    });
-}
 
 
+}
+
 NAMESPACE_END(Grid);
 

Grid/qcd/utils/CovariantCshift.h

@@ -88,12 +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)
   {
@@ -220,35 +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
-  template<class gauge> Lattice<gauge>
-  CshiftLink(const Lattice<gauge> &Link, int mu, int shift)
-  {
-    GridBase *grid = Link.Grid();
-    int Lmu = grid->GlobalDimensions()[mu] - 1;
-
-    Lattice<iScalar<vInteger>> coor(grid);
-    LatticeCoordinate(coor, mu);
-
-    Lattice<gauge> tmp(grid);
-    if(shift == 1){
-      tmp = Cshift(Link, mu, 1);
-      tmp = where(coor == Lmu, conjugate(tmp), tmp);
-      return tmp;
-    }else if(shift == -1){
-      tmp = Link;
-      tmp = where(coor == Lmu, conjugate(tmp), tmp);
-      return Cshift(tmp, mu, -1);
-    }else assert(0 && "Invalid shift value");
-    return tmp; //shuts up the compiler fussing about the return type
-  }
-
 }
 
 

Grid/qcd/utils/GaugeFix.h

@@ -40,46 +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) {
+  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);
+    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
   }
-
-  //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) {
+  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();
 
@@ -141,24 +122,29 @@ public:
 
       }
     }
-    assert(0 && "Gauge fixing did not converge within the specified number of iterations");
+    std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
+    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();
 
@@ -173,7 +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;
@@ -217,16 +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
   //////////////////////////////////////////////////
@@ -215,7 +164,7 @@ public:
 
     double vol = Umu.Grid()->gSites();
 
-    return p.real() / vol / 4.0 / 3.0;
+    return p.real() / vol / (4.0 * Nc ) ;
   };
 
   //////////////////////////////////////////////////
@@ -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/random/gaussian.h

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