Update to static data

Grid/DisableWarnings.h

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

Grid/GridCore.h

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

Grid/GridQCDcore.h

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

Grid/GridStd.h

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

Grid/Grid_Eigen_Dense.h

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

Grid/algorithms/Algorithms.h

@@ -54,7 +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/iterative/ConjugateGradient.h

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

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

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

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -44,7 +44,7 @@ public:
 
   using OperatorFunction<Field>::operator();
 
-  //  RealD   Tolerance;
+  RealD   Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -52,7 +52,7 @@ public:
   MultiShiftFunction shifts;
   std::vector<RealD> TrueResidualShift;
 
-  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
+  ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
     MaxIterations(maxit),
     shifts(_shifts)
   { 
@@ -84,7 +84,6 @@ public:
 
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
-    GRID_TRACE("ConjugateGradientMultiShift");
   
     GridBase *grid = src.Grid();
   
@@ -183,9 +182,6 @@ public:
     for(int s=0;s<nshift;s++) {
       axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
     }
-
-    std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
-    
   
   ///////////////////////////////////////
   // Timers
@@ -325,8 +321,8 @@ public:
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -1,410 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Christopher Kelly <ckelly@bnl.gov>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
-#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
-
-NAMESPACE_BEGIN(Grid);
-
-//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
-//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
-//Every update_freq iterations the residual is corrected in double precision. 
-    
-//For safety the a final regular CG is applied to clean up if necessary
-
-//Linop to add shift to input linop, used in cleanup CG
-namespace ConjugateGradientMultiShiftMixedPrecSupport{
-template<typename Field>
-class ShiftedLinop: public LinearOperatorBase<Field>{
-public:
-  LinearOperatorBase<Field> &linop_base;
-  RealD shift;
-
-  ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
-
-  void OpDiag (const Field &in, Field &out){ assert(0); }
-  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
-  
-  void Op     (const Field &in, Field &out){ assert(0); }
-  void AdjOp  (const Field &in, Field &out){ assert(0); }
-
-  void HermOp(const Field &in, Field &out){
-    linop_base.HermOp(in, out);
-    axpy(out, shift, in, out);
-  }    
-
-  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    HermOp(in,out);
-    ComplexD dot = innerProduct(in,out);
-    n1=real(dot);
-    n2=norm2(out);
-  }
-};
-};
-
-
-template<class FieldD, class FieldF,
-	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
-	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
-class ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
-					     public OperatorFunction<FieldD>
-{
-public:                                                
-
-  using OperatorFunction<FieldD>::operator();
-
-  RealD   Tolerance;
-  Integer MaxIterations;
-  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
-  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
-  int verbose;
-  MultiShiftFunction shifts;
-  std::vector<RealD> TrueResidualShift;
-
-  int ReliableUpdateFreq; //number of iterations between reliable updates
-
-  GridBase* SinglePrecGrid; //Grid for single-precision fields
-  LinearOperatorBase<FieldF> &Linop_f; //single precision
-
-  ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
-				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
-				       int _ReliableUpdateFreq
-				       ) : 
-    MaxIterations(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
-  { 
-    verbose=1;
-    IterationsToCompleteShift.resize(_shifts.order);
-    TrueResidualShift.resize(_shifts.order);
-  }
-
-  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
-  {
-    GridBase *grid = src.Grid();
-    int nshift = shifts.order;
-    std::vector<FieldD> results(nshift,grid);
-    (*this)(Linop,src,results,psi);
-  }
-  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
-  {
-    int nshift = shifts.order;
-
-    (*this)(Linop,src,results);
-  
-    psi = shifts.norm*src;
-    for(int i=0;i<nshift;i++){
-      psi = psi + shifts.residues[i]*results[i];
-    }
-
-    return;
-  }
-
-  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
-  { 
-    GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
-    GridBase *DoublePrecGrid = src_d.Grid();
-
-    ////////////////////////////////////////////////////////////////////////
-    // Convenience references to the info stored in "MultiShiftFunction"
-    ////////////////////////////////////////////////////////////////////////
-    int nshift = shifts.order;
-
-    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
-    std::vector<RealD> &mresidual(shifts.tolerances);
-    std::vector<RealD> alpha(nshift,1.0);
-
-    //Double precision search directions
-    FieldD p_d(DoublePrecGrid);
-    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
-
-    FieldD tmp_d(DoublePrecGrid);
-    FieldD r_d(DoublePrecGrid);
-    FieldD mmp_d(DoublePrecGrid);
-
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
-  
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
-  
-    const int       primary =0;
-  
-    //Primary shift fields CG iteration
-    RealD a,b,c,d;
-    RealD cp,bp,qq; //prev
-  
-    // Matrix mult fields
-    FieldF r_f(SinglePrecGrid);
-    FieldF p_f(SinglePrecGrid);
-    FieldF tmp_f(SinglePrecGrid);
-    FieldF mmp_f(SinglePrecGrid);
-    FieldF src_f(SinglePrecGrid);
-    precisionChange(src_f, src_d);
-
-    // Check lightest mass
-    for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
-      converged[s]=0;
-    }
-  
-    // Wire guess to zero
-    // Residuals "r" are src
-    // First search direction "p" is also src
-    cp = norm2(src_d);
-
-    // Handle trivial case of zero src.
-    if( cp == 0. ){
-      for(int s=0;s<nshift;s++){
-	psi_d[s] = Zero();
-	IterationsToCompleteShift[s] = 1;
-	TrueResidualShift[s] = 0.;
-      }
-      return;
-    }
-
-    for(int s=0;s<nshift;s++){
-      rsq[s] = cp * mresidual[s] * mresidual[s];
-      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
-      ps_d[s] = src_d;
-    }
-    // r and p for primary
-    r_f=src_f; //residual maintained in single
-    p_f=src_f;
-    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
-  
-    //MdagM+m[0]
-    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    axpy(mmp_f,mass[0],p_f,mmp_f);
-    RealD rn = norm2(p_f);
-    d += rn*mass[0];
-
-    b = -cp /d;
-  
-    // Set up the various shift variables
-    int       iz=0;
-    z[0][1-iz] = 1.0;
-    z[0][iz]   = 1.0;
-    bs[0]      = b;
-    for(int s=1;s<nshift;s++){
-      z[s][1-iz] = 1.0;
-      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
-      bs[s]      = b*z[s][iz]; 
-    }
-  
-    // r += b[0] A.p[0]
-    // c= norm(r)
-    c=axpy_norm(r_f,b,mmp_f,r_f);
-  
-    for(int s=0;s<nshift;s++) {
-      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
-    }
-  
-    ///////////////////////////////////////
-    // Timers
-    ///////////////////////////////////////
-    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
-
-    SolverTimer.Start();
-  
-    // Iteration loop
-    int k;
-  
-    for (k=1;k<=MaxIterations;k++){    
-      a = c /cp;
-
-      //Update double precision search direction by residual
-      PrecChangeTimer.Start();
-      precisionChange(r_d, r_f);
-      PrecChangeTimer.Stop();
-
-      AXPYTimer.Start();
-      axpy(p_d,a,p_d,r_d); 
-
-      for(int s=0;s<nshift;s++){
-	if ( ! converged[s] ) { 
-	  if (s==0){
-	    axpy(ps_d[s],a,ps_d[s],r_d);
-	  } else{
-	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
-	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
-	  }
-	}
-      }
-      AXPYTimer.Stop();
-
-      PrecChangeTimer.Start();
-      precisionChange(p_f, p_d); //get back single prec search direction for linop
-      PrecChangeTimer.Stop();
-
-      cp=c;
-      MatrixTimer.Start();  
-      Linop_f.HermOp(p_f,mmp_f); 
-      d=real(innerProduct(p_f,mmp_f));    
-      MatrixTimer.Stop();  
-
-      AXPYTimer.Start();
-      axpy(mmp_f,mass[0],p_f,mmp_f);
-      AXPYTimer.Stop();
-      RealD rn = norm2(p_f);
-      d += rn*mass[0];
-    
-      bp=b;
-      b=-cp/d;
-    
-      // Toggle the recurrence history
-      bs[0] = b;
-      iz = 1-iz;
-      ShiftTimer.Start();
-      for(int s=1;s<nshift;s++){
-	if((!converged[s])){
-	  RealD z0 = z[s][1-iz];
-	  RealD z1 = z[s][iz];
-	  z[s][iz] = z0*z1*bp
-	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
-	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
-	}
-      }
-      ShiftTimer.Stop();
-
-      //Update double precision solutions
-      AXPYTimer.Start();
-      for(int s=0;s<nshift;s++){
-	int ss = s;
-	if( (!converged[s]) ) { 
-	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
-	}
-      }
-
-      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
-      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
-      AXPYTimer.Stop();
-
-      c = c_f;
-
-      if(k % ReliableUpdateFreq == 0){
-	//Replace r with true residual
-	MatrixTimer.Start();  
-	Linop_d.HermOp(psi_d[0],mmp_d); 
-	MatrixTimer.Stop();  
-
-	AXPYTimer.Start();
-	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
-
-	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
-	AXPYTimer.Stop();
-
-	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
-	
-	PrecChangeTimer.Start();
-	precisionChange(r_f, r_d);
-	PrecChangeTimer.Stop();
-	c = c_d;
-      }
-    
-      // Convergence checks
-      int all_converged = 1;
-      for(int s=0;s<nshift;s++){
-      
-	if ( (!converged[s]) ){
-	  IterationsToCompleteShift[s] = k;
-	
-	  RealD css  = c * z[s][iz]* z[s][iz];
-	
-	  if(css<rsq[s]){
-	    if ( ! converged[s] )
-	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
-	    converged[s]=1;
-	  } else {
-	    all_converged=0;
-	  }
-
-	}
-      }
-
-      if ( all_converged ){
-
-	SolverTimer.Stop();
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
-      
-	// Check answers 
-	for(int s=0; s < nshift; s++) { 
-	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
-	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
-	  axpy(r_d,-alpha[s],src_d,tmp_d);
-	  RealD rn = norm2(r_d);
-	  RealD cn = norm2(src_d);
-	  TrueResidualShift[s] = std::sqrt(rn/cn);
-	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
-
-	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
-	  if(rn >= rsq[s]){
-	    CleanupTimer.Start();
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
-
-	    //Setup linear operators for final cleanup
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
-					       
-	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
-	    cg(src_d, psi_d[s]);
-	    
-	    TrueResidualShift[s] = cg.TrueResidual;
-	    CleanupTimer.Stop();
-	  }
-	}
-
-	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
-	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
-
-	IterationsToComplete = k;	
-
-	return;
-      }
-
-   
-    }
-    // ugly hack
-    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

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

Grid/algorithms/iterative/Deflation.h

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

Grid/algorithms/iterative/LocalCoherenceLanczos.h

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

Grid/algorithms/iterative/PowerMethod.h

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

Grid/allocator/MemoryManager.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, 8, 16, 8, 16 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils

Grid/allocator/MemoryManager.h

@@ -113,6 +113,11 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
+  
+  static uint64_t     DeviceAccesses;
+  static uint64_t     HostAccesses;
+  static uint64_t     DeviceAccessBytes;
+  static uint64_t     HostAccessBytes;
  
  private:
 #ifndef GRID_UVM
@@ -152,6 +157,7 @@ private:
 
   //  static void  LRUupdate(AcceleratorViewEntry &AccCache);
   static void  LRUinsert(AcceleratorViewEntry &AccCache);
+  static void  LRUinsertback(AcceleratorViewEntry &AccCache);
   static void  LRUremove(AcceleratorViewEntry &AccCache);
   
   // manage entries in the table

Grid/allocator/MemoryManagerCache.cc

@@ -3,14 +3,8 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-
+//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define MAXLINE 512
+#define dprintf(...)
-static char print_buffer [ MAXLINE ];
-
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-//#define dprintf(...) printf (__VA_ARGS__ ); fflush(stdout);
-#define dprintf(...) 
-
 
 
 ////////////////////////////////////////////////////////////
@@ -29,6 +23,11 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceAccesses;
+uint64_t  MemoryManager::HostAccesses;
+uint64_t  MemoryManager::DeviceAccessBytes;
+uint64_t  MemoryManager::HostAccessBytes;
+ 
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -92,6 +91,14 @@ void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
   AccCache.LRU_valid = 1;
   DeviceLRUBytes+=AccCache.bytes;
 }
+void  MemoryManager::LRUinsertback(AcceleratorViewEntry &AccCache)
+{
+  assert(AccCache.LRU_valid==0);
+  LRU.push_back(AccCache.CpuPtr);
+  AccCache.LRU_entry = --LRU.end();
+  AccCache.LRU_valid = 1;
+  DeviceLRUBytes+=AccCache.bytes;
+}
 void  MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
 {
   assert(AccCache.LRU_valid==1);
@@ -110,7 +117,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -118,7 +125,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -132,9 +139,10 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
+  
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
@@ -143,7 +151,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -156,7 +164,7 @@ 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);
-  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -171,7 +179,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -237,6 +245,9 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
     EntryCreate(CpuPtr,bytes,mode,hint);
   }
 
+  DeviceAccesses++;
+  DeviceAccessBytes+=bytes;
+
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
   if (!AccCache.AccPtr) {
@@ -247,7 +258,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld\n",
+    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
@@ -355,6 +366,10 @@ void MemoryManager::CpuViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock==0);
 
   AccCache.cpuLock--;
+
+  if(AccCache.cpuLock==0) {
+    LRUinsertback(AccCache);
+  }
 }
 /*
  *  Action  State   StateNext         Flush    Clone
@@ -377,6 +392,9 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
     EntryCreate(CpuPtr,bytes,mode,transient);
   }
 
+  HostAccesses++;
+  HostAccessBytes+=bytes;
+  
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
@@ -422,6 +440,12 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
 
   AccCache.transient= transient? EvictNext : 0;
 
+  // If view is opened on host remove from LRU
+  // Host close says evict next from device
+  if(AccCache.LRU_valid==1){
+    LRUremove(AccCache);
+  }
+  
   return AccCache.CpuPtr;
 }
 void  MemoryManager::NotifyDeletion(void *_ptr)

Grid/allocator/MemoryManagerShared.cc

@@ -12,6 +12,10 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceAccesses;
+uint64_t  MemoryManager::HostAccesses;
+uint64_t  MemoryManager::DeviceAccessBytes;
+uint64_t  MemoryManager::HostAccessBytes;
 
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };

Grid/communicator/Communicator_base.h

@@ -53,11 +53,10 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  int              _processor;       // linear processor rank
-  unsigned long    _ndimension;
-  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
   Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  int              _processor;       // linear processor rank
   Coordinate _processor_coor;  // linear processor coordinate
+  unsigned long    _ndimension;
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -98,16 +97,14 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
-  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                ProcessorCount(void)    ;
+  int                      ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
   static void BroadcastWorld(int root,void* data, int bytes);
-  static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
   // Reduction
@@ -145,16 +142,16 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,int do_xmit,
+			       int xmit_to_rank,
 			       void *recv,
-			       int recv_from_rank,int do_recv,
+			       int recv_from_rank,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,int do_xmit,
+				    int xmit_to_rank,
 				    void *recv,
-				    int recv_from_rank,int do_recv,
+				    int recv_from_rank,
 				    int bytes,int dir);
   
   

Grid/communicator/Communicator_mpi3.cc

@@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
   // Remap using the shared memory optimising routine
   // The remap creates a comm which must be freed
   ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,13 +124,12 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
-  Coordinate shm_processors       (_ndimension,1);
+
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
   for(int d=0;d<parent_ndimension;d++){
     parent_processor_coor[pad+d]=parent._processor_coor[d];
     parent_processors    [pad+d]=parent._processors[d];
-    shm_processors       [pad+d]=parent._shm_processors[d];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -155,7 +154,6 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     ccoor[d] = parent_processor_coor[d] % processors[d];
     scoor[d] = parent_processor_coor[d] / processors[d];
     ssize[d] = parent_processors[d]     / processors[d];
-    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@@ -337,22 +335,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest, int dox,
+						     int dest,
 						     void *recv,
-						     int from, int dor,
+						     int from,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,int dox,
+							 int dest,
 							 void *recv,
-							 int from,int dor,
+							 int from,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -372,41 +370,38 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( dor ) {
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+from*32;
-      tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-      ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+    assert(ierr==0);
-      assert(ierr==0);
+    list.push_back(rrq);
-      list.push_back(rrq);
+    off_node_bytes+=bytes;
-      off_node_bytes+=bytes;
-    }
   }
-  
+
-  if (dox) {
+  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+_processor*32;
-      tag= dir+_processor*32;
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+    assert(ierr==0);
-      assert(ierr==0);
+    list.push_back(xrq);
-      list.push_back(xrq);
+    off_node_bytes+=bytes;
-      off_node_bytes+=bytes;
+  } else {
-    } else {
+    // TODO : make a OMP loop on CPU, call threaded bcopy
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
+    assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
-    }
+    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
   }
-  
+
-  /*  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-   *    this->StencilSendToRecvFromComplete(list,dir);
+    this->StencilSendToRecvFromComplete(list,dir);
-   *    list.resize(0);
+  }
-   *  }
+
-   */
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
+  //   std::cout << "Copy Synchronised\n"<<std::endl;
   acceleratorCopySynchronise();
-  StencilBarrier();// Synch shared memory on a single nodes
 
   int nreq=list.size();
 
@@ -443,10 +438,6 @@ int CartesianCommunicator::RankWorld(void){
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
-void CartesianCommunicator::BarrierWorld(void){
-  int ierr = MPI_Barrier(communicator_world);
-  assert(ierr==0);
-}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,

Grid/communicator/Communicator_none.cc

@@ -45,14 +45,12 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
-  _shm_processors = Coordinate(processors.size(),1);
   srank=0;
   SetCommunicator(communicator_world);
 }
 
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
-  _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
   _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
@@ -104,7 +102,6 @@ int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
-void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
@@ -114,18 +111,18 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,int dox,
+						     int xmit_to_rank,
 						     void *recv,
-						     int recv_from_rank,int dor,
+						     int recv_from_rank,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,int dox,
+							 int xmit_to_rank,
 							 void *recv,
-							 int recv_from_rank,int dor,
+							 int recv_from_rank,
 							 int bytes, int dir)
 {
   return 2.0*bytes;

Grid/communicator/SharedMemory.h

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

Grid/communicator/SharedMemoryMPI.cc

@@ -152,7 +152,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   //////////////////////////////////////////////////////////////////////////////
   // Look and see if it looks like an HPE 8600 based on hostname conventions
@@ -165,8 +165,8 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
 
-  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
 }
 static inline int divides(int a,int b)
 {
@@ -221,7 +221,7 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
     dim=(dim+1) %ndimension;
   }
 }
-void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -294,8 +294,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-  SHM = ShmDims;
+
-  
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -342,7 +341,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
 }
-void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -354,8 +353,6 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-  SHM=ShmDims;
-
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -523,7 +520,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   }
   if ( WorldRank == 0 ){
     std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
+	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
   std::cout<< "Setting up IPC"<<std::endl;

Grid/communicator/SharedMemoryNone.cc

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

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_arith.h

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

Grid/lattice/Lattice_base.h

@@ -117,7 +117,6 @@ public:
   ////////////////////////////////////////////////////////////////////////////////
   template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -141,7 +140,6 @@ public:
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -165,7 +163,6 @@ public:
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);

Grid/lattice/Lattice_crc.h

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

Grid/lattice/Lattice_peekpoke.h

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

Grid/lattice/Lattice_reduction.h

@@ -91,7 +91,10 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
   for(int i=0;i<nthread;i++){
     ssum = ssum+sumarray[i];
   } 
-  return ssum;
+  
+  typedef typename vobj::scalar_object ssobj;
+  ssobj ret = ssum;
+  return ret;
 }
 /*
 Threaded max, don't use for now
@@ -139,15 +142,6 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
   return sumD_cpu(arg,osites);
 #endif  
 }
-template<class vobj>
-inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  return sumD_gpu_large(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
@@ -165,22 +159,6 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
   return ssum;
 }
 
-template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu_large(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
-#endif
-  arg.Grid()->GlobalSum(ssum);
-  return ssum;
-}
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Deterministic Reduction operations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -229,7 +207,6 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   const uint64_t sites = grid->oSites();
   
   // Might make all code paths go this way.
-#if 0
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
@@ -239,31 +216,15 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
     autoView( right_v,right, AcceleratorRead);
 
     // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
+    accelerator_for( ss, sites, 1,{
-	auto x_l = left_v(ss);
+	auto x_l = left_v[ss];
-	auto y_l = right_v(ss);
+	auto y_l = right_v[ss];
-	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
+	inner_tmp_v[ss]=innerProductD(x_l,y_l);
     });
   }
-#else
-  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-    
-  {
-    autoView( left_v , left, AcceleratorRead);
-    autoView( right_v,right, AcceleratorRead);
 
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
-    });
-  }
-#endif
   // This is in single precision and fails some tests
-  auto anrm = sumD(inner_tmp_v,sites);  
+  auto anrm = sum(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
@@ -297,7 +258,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   conformable(x,y);
 
   typedef typename vobj::scalar_type scalar_type;
-  //  typedef typename vobj::vector_typeD vector_type;
+  typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
   GridBase *grid = x.Grid();
@@ -309,29 +270,17 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-#if 0
+
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
 
-  accelerator_for( ss, sites, nsimd,{
+  accelerator_for( ss, sites, 1,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
+      auto tmp = a*x_v[ss]+b*y_v[ss];
-      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
+      inner_tmp_v[ss]=innerProductD(tmp,tmp);
-      coalescedWrite(z_v[ss],tmp);
+      z_v[ss]=tmp;
   });
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#else
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -485,14 +434,6 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
 }
-template<class vobj> inline
-std::vector<typename vobj::scalar_object> 
-sliceSum(const Lattice<vobj> &Data,int orthogdim)
-{
-  std::vector<typename vobj::scalar_object> result;
-  sliceSum(Data,result,orthogdim);
-  return result;
-}
 
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 

Grid/lattice/Lattice_reduction_gpu.h

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

Grid/lattice/Lattice_rng.h

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

Grid/lattice/Lattice_transfer.h

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

Grid/log/Log.cc

@@ -65,37 +65,29 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
-GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
-GridLogger 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(1);
+  GridLogError.Active(0);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
-  GridLogMemory.Active(0); // at least the messages should be always on
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
-  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("Memory"))      GridLogMemory.Active(1);
+    if (logstreams[i] == std::string("Error"))       GridLogError.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("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
-    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
-    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,8 +138,7 @@ 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);
@@ -154,9 +153,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      //      stream << log.colour();
+      stream << log.colour();
-      stream <<  std::right;
       stream.flags(f);
+
       return stream;
     } else { 
       return devnull;
@@ -181,11 +180,8 @@ 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 Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/IldgIO.h

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

Grid/parallelIO/MetaData.h

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

Grid/parallelIO/NerscIO.h

@@ -9,7 +9,6 @@
     Author: Matt Spraggs <matthew.spraggs@gmail.com>
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -31,8 +30,6 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
-#include <string>
-
 NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
@@ -42,10 +39,8 @@ 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)
+  typedef Lattice<vLorentzColourMatrixD> GaugeField;
-  static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; }
 
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
@@ -150,17 +145,15 @@ public:
 
     std::string format(header.floating_point);
 
-    const int ieee32big = (format == std::string("IEEE32BIG"));
+    int ieee32big = (format == std::string("IEEE32BIG"));
-    const int ieee32    = (format == std::string("IEEE32"));
+    int ieee32    = (format == std::string("IEEE32"));
-    const int ieee64big = (format == std::string("IEEE64BIG"));
+    int ieee64big = (format == std::string("IEEE64BIG"));
-    const int ieee64    = (format == std::string("IEEE64") || \
+    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
-			   format == std::string("IEEE64LITTLE"));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     // depending on datatype, set up munger;
     // munger is a function of <floating point, Real, data_type>
-    const std::string stNC = std::to_string( Nc ) ;
+    if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
-    if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@@ -171,7 +164,7 @@ public:
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
-    } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
+    } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@@ -205,7 +198,7 @@ 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 );
       
@@ -216,29 +209,27 @@ public:
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
-					std::string ens_label = std::string("DWF"),
+					std::string ens_label = std::string("DWF"))
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
   {
-    writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
+    writeConfiguration(Umu,file,0,1,ens_label);
   }
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32,
-					std::string ens_label = std::string("DWF"),
+					std::string ens_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;
-    header.sequence_number = sequence_number;
+    ///////////////////////////////////////////
-    header.ensemble_id     = ens_id;
+    // Following should become arguments
+    ///////////////////////////////////////////
+    header.sequence_number = 1;
+    header.ensemble_id     = std::string("UKQCD");
     header.ensemble_label  = ens_label;
-    header.hdr_version     = "1.0" ;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -252,14 +243,10 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write NcxNc double
+    // Sod it -- always write 3x3 double
-    header.floating_point  = std::string("IEEE64BIG");
+    header.floating_point = std::string("IEEE64BIG");
-    const std::string stNC = std::to_string( Nc ) ;
+    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
-    if( two_row ) {
+    GaugeSimpleUnmunger<fobj3D,sobj> munge;
-      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);
@@ -267,15 +254,8 @@ public:
     grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    if( two_row ) {
+    BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
-      Gauge3x2unmunger<fobj2D,sobj> munge;
+					      nersc_csum,scidac_csuma,scidac_csumb);
-      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);
@@ -307,7 +287,8 @@ public:
     header.plaquette=0.0;
     MachineCharacteristics(header);
 
-    uint64_t offset;
+	uint64_t offset;
+  
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");
@@ -347,7 +328,7 @@ 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

@@ -27,12 +27,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
-
-#include <Grid/perfmon/Timer.h>
 #include <Grid/perfmon/PerfCount.h>
-NAMESPACE_BEGIN(Grid);
 
-GridTimePoint theProgramStart = GridClock::now();
+NAMESPACE_BEGIN(Grid);
 
 #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)

Grid/perfmon/PerfCount.h

@@ -72,9 +72,17 @@ 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

@@ -35,8 +35,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid)
 
-//typedef  std::chrono::system_clock          GridClock;
+// Dress the output; use std::chrono
-typedef  std::chrono::high_resolution_clock   GridClock;
+// 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::time_point<GridClock> GridTimePoint;
 
 typedef  std::chrono::seconds               GridSecs;
@@ -44,15 +53,6 @@ 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

@@ -1,66 +0,0 @@
-#pragma once
-#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

Grid/pugixml/pugixml.cc

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

Grid/qcd/QCD.h

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

Grid/qcd/action/ActionBase.h

@@ -40,43 +40,9 @@ class Action
 
 public:
   bool is_smeared = false;
-  RealD deriv_norm_sum;
-  RealD deriv_max_sum;
-  RealD Fdt_norm_sum;
-  RealD Fdt_max_sum;
-  int   deriv_num;
-  RealD deriv_us;
-  RealD S_us;
-  RealD refresh_us;
-  void  reset_timer(void)        {
-    deriv_us = S_us = refresh_us = 0.0;
-    deriv_num=0;
-    deriv_norm_sum = deriv_max_sum=0.0;
-    Fdt_max_sum =  Fdt_norm_sum = 0.0;
-  }
-  void  deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) {
-    deriv_max_sum+=max;
-    deriv_norm_sum+=nrm;
-    Fdt_max_sum+=Fdt_max;
-    Fdt_norm_sum+=Fdt_nrm; deriv_num++;
-  }
-  RealD deriv_max_average(void)       { return deriv_max_sum/deriv_num; };
-  RealD deriv_norm_average(void)      { return deriv_norm_sum/deriv_num; };
-  RealD Fdt_max_average(void)         { return Fdt_max_sum/deriv_num; };
-  RealD Fdt_norm_average(void)        { return Fdt_norm_sum/deriv_num; };
-  RealD deriv_timer(void)        { return deriv_us; };
-  RealD S_timer(void)            { return S_us; };
-  RealD refresh_timer(void)      { return refresh_us; };
-  void deriv_timer_start(void)   { deriv_us-=usecond(); }
-  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
-  void refresh_timer_start(void) { refresh_us-=usecond(); }
-  void refresh_timer_stop(void)  { refresh_us+=usecond(); }
-  void S_timer_start(void)       { S_us-=usecond(); }
-  void S_timer_stop(void)        { S_us+=usecond(); }
   // Heatbath?
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
-  virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
   virtual std::string action_name()    = 0;                             // return the action name
   virtual std::string LogParameters()  = 0;                             // prints action parameters

Grid/qcd/action/ActionCore.h

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

Grid/qcd/action/ActionParams.h

@@ -37,33 +37,24 @@ NAMESPACE_BEGIN(Grid);
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
   Coordinate twists;
-                     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+  GparityWilsonImplParams() : twists(Nd, 0) {};
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  GparityWilsonImplParams() : twists(Nd, 0) { dirichlet.resize(0); };
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
-    dirichlet.resize(0);
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
-    dirichlet.resize(0);
   }
 };
 
 struct StaggeredImplParams {
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  StaggeredImplParams()  {};
-  StaggeredImplParams()
-  {
-    dirichlet.resize(0);
-  };
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -72,11 +63,9 @@ struct StaggeredImplParams {
 				    RealD, hi, 
 				    int,   MaxIter, 
 				    RealD, tolerance, 
-				    RealD, mdtolerance, 
 				    int,   degree, 
 				    int,   precision,
-				    int,   BoundsCheckFreq,
+				    int,   BoundsCheckFreq);
-				    RealD, BoundsCheckTol);
     
   // MaxIter and tolerance, vectors??
     
@@ -87,62 +76,16 @@ struct StaggeredImplParams {
 				RealD tol      = 1.0e-8, 
                            	int _degree    = 10,
 				int _precision = 64,
-				int _BoundsCheckFreq=20,
+				int _BoundsCheckFreq=20)
-				RealD mdtol    = 1.0e-6,
-				double _BoundsCheckTol=1e-6)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
-        mdtolerance(mdtol),
 	degree(_degree),
         precision(_precision),
-        BoundsCheckFreq(_BoundsCheckFreq),
+        BoundsCheckFreq(_BoundsCheckFreq){};
-        BoundsCheckTol(_BoundsCheckTol){};
   };
   
-  /*Action parameters for the generalized rational action
-    The approximation is for (M^dag M)^{1/inv_pow}
-    where inv_pow is the denominator of the fractional power.
-    Default inv_pow=2 for square root, making this equivalent to 
-    the OneFlavourRational action
-  */
-    struct RationalActionParams : Serializable {
-    GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams, 
-				    int, inv_pow, 
-				    RealD, lo, //low eigenvalue bound of rational approx
-				    RealD, hi, //high eigenvalue bound of rational approx
-				    int,   MaxIter,  //maximum iterations in msCG
-				    RealD, action_tolerance,  //msCG tolerance in action evaluation
-				    int,   action_degree, //rational approx tolerance in action evaluation
-				    RealD, md_tolerance,  //msCG tolerance in MD integration
-				    int,   md_degree, //rational approx tolerance in MD integration
-				    int,   precision, //precision of floating point arithmetic
-				    int,   BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
-  // constructor 
-  RationalActionParams(int _inv_pow = 2,
-		       RealD _lo      = 0.0, 
-		       RealD _hi      = 1.0, 
-		       int _maxit     = 1000,
-		       RealD _action_tolerance      = 1.0e-8, 
-		       int _action_degree    = 10,
-		       RealD _md_tolerance      = 1.0e-8, 
-		       int _md_degree    = 10,
-		       int _precision = 64,
-		       int _BoundsCheckFreq=20)
-    : inv_pow(_inv_pow), 
-      lo(_lo),
-      hi(_hi),
-      MaxIter(_maxit),
-      action_tolerance(_action_tolerance),
-      action_degree(_action_degree),
-      md_tolerance(_md_tolerance),
-      md_degree(_md_degree),
-      precision(_precision),
-      BoundsCheckFreq(_BoundsCheckFreq){};
-  };
-
-
 NAMESPACE_END(Grid);
 
 #endif

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -68,17 +68,9 @@ public:
   ///////////////////////////////////////////////////////////////
   // Support for MADWF tricks
   ///////////////////////////////////////////////////////////////
-  RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
+  RealD Mass(void) { return mass; };
-  RealD MassPlus(void) { return mass_plus; };
-  RealD MassMinus(void) { return mass_minus; };
-
   void  SetMass(RealD _mass) { 
-    mass_plus=mass_minus=_mass; 
+    mass=_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);
@@ -116,7 +108,7 @@ public:
   void   MeooeDag5D    (const FermionField &in, FermionField &out);
 
   //    protected:
-  RealD mass_plus, mass_minus;
+  RealD mass;
 
   // Save arguments to SetCoefficientsInternal
   Vector<Coeff_t> _gamma;
@@ -183,6 +175,16 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  RealD _mass,RealD _M5,const ImplParams &p= ImplParams());
 
+  void CayleyReport(void);
+  void CayleyZeroCounters(void);
+
+  double M5Dflops;
+  double M5Dcalls;
+  double M5Dtime;
+
+  double MooeeInvFlops;
+  double MooeeInvCalls;
+  double MooeeInvTime;
 
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);

Grid/qcd/action/fermion/CloverHelpers.h

@@ -1,435 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
-
-    Copyright (C) 2017 - 2022
-
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-/*  END LEGAL */
-
-#pragma once
-
-#include <Grid/Grid.h>
-#include <Grid/qcd/spin/Dirac.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-
-////////////////////////////////////////////
-// Standard Clover
-//   (4+m0) + csw * clover_term
-// Exp Clover
-//   (4+m0) * exp(csw/(4+m0) clover_term)
-//   = (4+m0) + csw * clover_term + ...
-////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-
-//////////////////////////////////
-// Generic Standard Clover
-//////////////////////////////////
-
-template<class Impl>
-class CloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
-    GridBase *grid = CloverTerm.Grid();
-    CloverTerm += diag_mass;
-
-    int lvol = grid->lSites();
-    int DimRep = Impl::Dimension;
-    {
-      autoView(CTv,CloverTerm,CpuRead);
-      autoView(CTIv,CloverTermInv,CpuWrite);
-      thread_for(site, lvol, {
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-        Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
-        peekLocalSite(Qx, CTv, lcoor);
-
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++){
-                auto zz =  Qx()(j, k)(a, b);
-                EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
-              }
-
-        EigenInvCloverOp = EigenCloverOp.inverse();
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++)
-                Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-               pokeLocalSite(Qxinv, CTIv, lcoor);
-      });
-    }
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-
-};
-
-
-//////////////////////////////////
-// Generic Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  // Can this be avoided?
-  static void IdentityTimesC(const CloverField& in, RealD c) {
-    int DimRep = Impl::Dimension;
-
-    autoView(in_v, in, AcceleratorWrite);
-
-    accelerator_for(ss, in.Grid()->oSites(), 1, {
-      for (int sa=0; sa<Ns; sa++)
-        for (int ca=0; ca<DimRep; ca++)
-          in_v[ss]()(sa,sa)(ca,ca) = c;
-    });
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-    GridBase* grid = Clover.Grid();
-    CloverField ExpClover(grid);
-
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-
-    Clover *= (1.0/diag_mass);
-
-    // Taylor expansion, slow but generic
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-    // qN = cN
-    // qn = cn + qn+1 X
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
-      cn[i] = cn[i-1] / RealD(i);
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * Clover + cn[i];
-
-    // prepare inverse
-    CloverInv = (-1.0)*Clover;
-
-    Clover = ExpClover * diag_mass;
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * CloverInv + cn[i];
-
-    CloverInv = ExpClover * (1.0/diag_mass);
-
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-//////////////////////////////////
-// Compact Standard Clover
-//////////////////////////////////
-
-
-template<class Impl>
-class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
-                            public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-    Clover += diag_mass;
-  }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
-                          CloverTriangleField& Triangle,
-                          RealD csw_t, RealD diag_mass) {
-
-    // Do nothing
-  }
-
-  // TODO: implement Cmunu for better performances with compact layout, but don't do it
-  // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-};
-
-//////////////////////////////////
-// Compact Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-    // do nothing!
-    // mass term is multiplied to exp(Clover) below
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
-
-  	  typedef iMatrix<ComplexD,6> mat;
-
-  	  RealD qn[6];
-  	  RealD qnold[6];
-  	  RealD p[5];
-  	  RealD trA2, trA3, trA4;
-
-  	  mat A2, A3, A4, A5;
-  	  A2 = alpha * alpha * arg * arg;
-  	  A3 = alpha * arg * A2;
-  	  A4 = A2 * A2;
-  	  A5 = A2 * A3;
-
-  	  trA2 = toReal( trace(A2) );
-  	  trA3 = toReal( trace(A3) );
-  	  trA4 = toReal( trace(A4));
-
-  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
-  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
-  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
-  	  p[3] = trA3 / 3.0;
-  	  p[4] = 0.5 * trA2;
-
-  	  qnold[0] = cN[Niter];
-  	  qnold[1] = 0.0;
-  	  qnold[2] = 0.0;
-  	  qnold[3] = 0.0;
-  	  qnold[4] = 0.0;
-  	  qnold[5] = 0.0;
-
-  	  for(int i = Niter-1; i >= 0; i--)
-  	  {
-  	   qn[0] = p[0] * qnold[5] + cN[i];
-  	   qn[1] = p[1] * qnold[5] + qnold[0];
-  	   qn[2] = p[2] * qnold[5] + qnold[1];
-  	   qn[3] = p[3] * qnold[5] + qnold[2];
-  	   qn[4] = p[4] * qnold[5] + qnold[3];
-  	   qn[5] = qnold[4];
-
-  	   qnold[0] = qn[0];
-  	   qnold[1] = qn[1];
-  	   qnold[2] = qn[2];
-  	   qnold[3] = qn[3];
-  	   qnold[4] = qn[4];
-  	   qnold[5] = qn[5];
-  	  }
-
-  	  mat unit(1.0);
-
-  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
-
-    }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Diagonal.Grid();
-    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
-
-    //
-    // Implementation completely in Daniel's layout
-    //
-
-    // Taylor expansion with Cayley-Hamilton recursion
-    // underlying Horner scheme as above
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++){
-      cn[i] = cn[i-1] / RealD(i);
-    }
-
-      // Taken over from Daniel's implementation
-      conformable(Diagonal, Triangle);
-
-      long lsites = grid->lSites();
-    {
-      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
-      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-      typedef iMatrix<ComplexD,6> mat;
-
-      autoView(diagonal_v,  Diagonal,  CpuRead);
-      autoView(triangle_v,  Triangle,  CpuRead);
-      autoView(diagonalExp_v, Diagonal, CpuWrite);
-      autoView(triangleExp_v, Triangle, CpuWrite);
-
-      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
-
-    	  mat srcCloverOpUL(0.0); // upper left block
-    	  mat srcCloverOpLR(0.0); // lower right block
-    	  mat ExpCloverOp;
-
-        scalar_object_diagonal diagonal_tmp     = Zero();
-        scalar_object_diagonal diagonal_exp_tmp = Zero();
-        scalar_object_triangle triangle_tmp     = Zero();
-        scalar_object_triangle triangle_exp_tmp = Zero();
-
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-
-        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-        peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-        int block;
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-        		if (i == j){
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-        		}
-        		else{
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-        		}
-        	}
-        }
-        block = 1;
-        for(int i = 0; i < 6; i++){
-          	for(int j = 0; j < 6; j++){
-           		if (i == j){
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-           		}
-           		else{
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-           		}
-            }
-        }
-
-        // exp(Clover)
-
-        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-            	if (i == j){
-            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-            	}
-            	else if(i < j){
-            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-            	}
-           	}
-        }
-
-        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 1;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-              	if (i == j){
-              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-               	}
-               	else if(i < j){
-               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-               	}
-            }
-        }
-
-        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
-        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
-      });
-    }
-
-    Diagonal *= diag_mass;
-    Triangle *= diag_mass;
-  }
-
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -1,241 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h
-
-    Copyright (C) 2020 - 2022
-
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Nils Meyer <nils.meyer@ur.de>
-
-    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/qcd/action/fermion/WilsonCloverTypes.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
-
-NAMESPACE_BEGIN(Grid);
-
-// see Grid/qcd/action/fermion/WilsonCloverFermion.h for description
-//
-// Modifications done here:
-//
-// Original: clover term = 12x12 matrix per site
-//
-// But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running)
-// Sufficient to store/transfer only the real parts of the diagonal and one triangular part
-// 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered
-//
-// Here: Above but diagonal as complex numbers, i.e., need to store/transfer
-// 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words
-//
-// Words per site and improvement compared to original (combined with the input and output spinors):
-//
-// - Original: 2*12 + 12*12 = 168 words -> 1.00 x less
-// - Minimal:  2*12 + 36    =  60 words -> 2.80 x less
-// - Here:     2*12 + 42    =  66 words -> 2.55 x less
-//
-// These improvements directly translate to wall-clock time
-//
-// Data layout:
-//
-// - diagonal and triangle part as separate lattice fields,
-//   this was faster than as 1 combined field on all tested machines
-// - diagonal: as expected
-// - triangle: store upper right triangle in row major order
-// - graphical:
-//        0  1  2  3  4
-//           5  6  7  8
-//              9 10 11 = upper right triangle indices
-//                12 13
-//                   14
-//     0
-//        1
-//           2
-//              3       = diagonal indices
-//                 4
-//                    5
-//     0
-//     1  5
-//     2  6  9          = lower left triangle indices
-//     3  7 10 12
-//     4  8 11 13 14
-//
-// Impact on total memory consumption:
-// - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site
-// - Here:     (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts =  24 complex words per site
-//           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
-//                                                                 =  84 complex words per site
-
-template<class Impl, class CloverHelpers>
-class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
-                                   public WilsonCloverHelpers<Impl>,
-                                   public CompactWilsonCloverHelpers<Impl> {
-  /////////////////////////////////////////////
-  // Sizes
-  /////////////////////////////////////////////
-
-public:
-
-  INHERIT_COMPACT_CLOVER_SIZES(Impl);
-
-  /////////////////////////////////////////////
-  // Type definitions
-  /////////////////////////////////////////////
-
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  typedef WilsonFermion<Impl>              WilsonBase;
-  typedef WilsonCloverHelpers<Impl>        Helpers;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  /////////////////////////////////////////////
-  // Constructors
-  /////////////////////////////////////////////
-
-public:
-
-  CompactWilsonCloverFermion(GaugeField& _Umu,
-			    GridCartesian& Fgrid,
-			    GridRedBlackCartesian& Hgrid,
-			    const RealD _mass,
-			    const RealD _csw_r = 0.0,
-			    const RealD _csw_t = 0.0,
-			    const RealD _cF = 1.0,
-			    const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
-			    const ImplParams& impl_p = ImplParams());
-
-  /////////////////////////////////////////////
-  // Member functions (implementing interface)
-  /////////////////////////////////////////////
-
-public:
-
-  virtual void Instantiatable() {};
-  int          ConstEE()     override { return 0; };
-  int          isTrivialEE() override { return 0; };
-
-  void Dhop(const FermionField& in, FermionField& out, int dag) override;
-
-  void DhopOE(const FermionField& in, FermionField& out, int dag) override;
-
-  void DhopEO(const FermionField& in, FermionField& out, int dag) override;
-
-  void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
-
-  void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
-
-  void M(const FermionField& in, FermionField& out) override;
-
-  void Mdag(const FermionField& in, FermionField& out) override;
-
-  void Meooe(const FermionField& in, FermionField& out) override;
-
-  void MeooeDag(const FermionField& in, FermionField& out) override;
-
-  void Mooee(const FermionField& in, FermionField& out) override;
-
-  void MooeeDag(const FermionField& in, FermionField& out) override;
-
-  void MooeeInv(const FermionField& in, FermionField& out) override;
-
-  void MooeeInvDag(const FermionField& in, FermionField& out) override;
-
-  void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
-
-  void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
-
-  void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
-
-  void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
-
-  void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
-
-  /////////////////////////////////////////////
-  // Member functions (internals)
-  /////////////////////////////////////////////
-
-  void MooeeInternal(const FermionField&        in,
-                     FermionField&              out,
-                     const CloverDiagonalField& diagonal,
-                     const CloverTriangleField& triangle);
-
-  /////////////////////////////////////////////
-  // Helpers
-  /////////////////////////////////////////////
-
-  void ImportGauge(const GaugeField& _Umu) override;
-
-  /////////////////////////////////////////////
-  // Helpers
-  /////////////////////////////////////////////
-
-private:
-
-  template<class Field>
-  const MaskField* getCorrectMaskField(const Field &in) const {
-    if(in.Grid()->_isCheckerBoarded) {
-      if(in.Checkerboard() == Odd) {
-        return &this->BoundaryMaskOdd;
-      } else {
-        return &this->BoundaryMaskEven;
-      }
-    } else {
-      return &this->BoundaryMask;
-    }
-  }
-
-  template<class Field>
-  void ApplyBoundaryMask(Field& f) {
-    const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
-    assert(m != nullptr);
-    CompactHelpers::ApplyBoundaryMask(f, *m);
-  }
-
-  /////////////////////////////////////////////
-  // Member Data
-  /////////////////////////////////////////////
-
-public:
-
-  RealD csw_r;
-  RealD csw_t;
-  RealD cF;
-
-  bool open_boundaries;
-
-  CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
-  CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
-
-  CloverTriangleField Triangle,    TriangleEven,    TriangleOdd;
-  CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
-
-  FermionField Tmp;
-
-  MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
-};
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -53,7 +53,6 @@ NAMESPACE_CHECK(Wilson);
 #include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(WilsonTM);
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
-#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
 NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
@@ -138,52 +137,21 @@ typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
 // Clover fermions
-template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
+typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
-template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
+typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
+typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
-typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
-typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
+typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
 
-typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
 
-typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
-
-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
-template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
-template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
-
-typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
-typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
-typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
-
-typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
-typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
-typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
-
-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/FermionOperator.h

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

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -30,18 +30,6 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
-/*
-  Policy implementation for G-parity boundary conditions
-
-  Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular
-  field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first
-  with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field
-  to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to
-  either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor.
-
-  Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
-  mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
- */
 template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal>
 class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > {
 public:
@@ -125,7 +113,7 @@ public:
     || ((distance== 1)&&(icoor[direction]==1))
     || ((distance==-1)&&(icoor[direction]==0));
 
-    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
 
     //Apply the links
     int f_upper = permute_lane ? 1 : 0;
@@ -151,10 +139,10 @@ public:
     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code
     assert((sl == 1) || (sl == 2));
 
-    //If this site is an global boundary site, perform the G-parity flavor twist
+    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
-    if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) {
+
       if ( sl == 2 ) {
-	//Only do the twist for lanes on the edge of the physical node
+       
 	ExtractBuffer<sobj> vals(Nsimd);
 
 	extract(chi,vals);
@@ -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. 
+    for(int mu=0;mu<Nd;mu++){
-    //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++){
+      LatticeCoordinate(coor,mu);
-
-      if( Params.twists[mu] ){
-	LatticeCoordinate(coor,mu);
-      }
           
       U     = PeekIndex<LorentzIndex>(Umu,mu);
       Uconj = conjugate(U);
      
-      // Implement the isospin rotation sign on the boundary between f=1 and f=0
       // This phase could come from a simple bc 1,1,-1,1 ..
       int neglink = GaugeGrid->GlobalDimensions()[mu]-1;
       if ( Params.twists[mu] ) { 
@@ -259,7 +229,7 @@ public:
 	thread_foreach(ss,U_v,{
 	    Uds_v[ss](0)(mu) = U_v[ss]();
 	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
-	});
+	  });
       }
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
@@ -290,38 +260,6 @@ public:
         });
       }
     }
-
-    { //periodic / antiperiodic temporal BCs
-      int mu = Nd-1;
-      int L   = GaugeGrid->GlobalDimensions()[mu];
-      int Lmu = L - 1;
-
-      LatticeCoordinate(coor, mu);
-
-      U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links
-      
-      GaugeLinkField *Upoke = &U;
-
-      if(Params.twists[mu]){ //antiperiodic
-	Utmp =  where(coor == Lmu, -U, U);
-	Upoke = &Utmp;
-      }
-    
-      Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links      
-      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu);
-
-      //Get the barrel-shifted field
-      Utmp = adj(Cshift(U, mu, -1)); //is a forward shift!
-      Upoke = &Utmp;
-
-      if(Params.twists[mu]){
-	U = where(coor == 0, -Utmp, Utmp);  //boundary phase
-	Upoke = &U;
-      }
-      
-      Uconj = conjugate(*Upoke);
-      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4);
-    }
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
@@ -360,48 +298,28 @@ public:
   inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
     assert(0);
   }
- 
-  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
-    int Ls=Btilde.Grid()->_fdimensions[0];
-    
-    {
-      GridBase *GaugeGrid = mat.Grid();
-      Lattice<iScalar<vInteger> > coor(GaugeGrid);
-
-      if( Params.twists[mu] ){
-	LatticeCoordinate(coor,mu);
-      }
-
-      autoView( mat_v , mat, AcceleratorWrite);
-      autoView( Btilde_v , Btilde, AcceleratorRead);
-      autoView( Atilde_v , Atilde, AcceleratorRead);
-      accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{	  
-  	  int sU=sss;
-  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
-  	  ColorMatrixType sum;
-  	  zeroit(sum);
-  	  for(int s=0;s<Ls;s++){
-  	    int sF = s+Ls*sU;
-  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
-	      //Flavor 0
-  	      auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector
-  	      auto aa = coalescedRead(Atilde_v[sF](0)(spn) );
-  	      sum = sum + outerProduct(bb,aa);
-
-  	      //Flavor 1
-  	      bb = coalescedRead(Btilde_v[sF](1)(spn) );
-  	      aa = coalescedRead(Atilde_v[sF](1)(spn) );
-  	      sum = sum + conjugate(outerProduct(bb,aa));
-  	    }
-  	  }	    
-  	  coalescedWrite(mat_v[sU](mu)(), sum);
-  	});
-    }
-  }
-
-
   
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
 
+    int Ls = Btilde.Grid()->_fdimensions[0];
+        
+    GaugeLinkField tmp(mat.Grid());
+    tmp = Zero();
+    {
+      autoView( tmp_v , tmp, CpuWrite);
+      autoView( Atilde_v , Atilde, CpuRead);
+      autoView( Btilde_v , Btilde, CpuRead);
+      thread_for(ss,tmp.Grid()->oSites(),{
+	  for (int s = 0; s < Ls; s++) {
+	    int sF = s + Ls * ss;
+	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+	  }
+	});
+    }
+    PokeIndex<LorentzIndex>(mat, tmp, mu);
+    return;
+  }
   
 };
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

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

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

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

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

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

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -4,11 +4,10 @@
 
     Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
 
-    Copyright (C) 2017 - 2022
+    Copyright (C) 2017
 
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: David Preti <>
-    Author: Daniel Richtmann <daniel.richtmann@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
@@ -30,9 +29,7 @@
 
 #pragma once
 
-#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/Grid.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -52,16 +49,19 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-class WilsonCloverFermion : public WilsonFermion<Impl>,
+class WilsonCloverFermion : public WilsonFermion<Impl>
-                            public WilsonCloverHelpers<Impl>
 {
 public:
+  // Types definitions
   INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
+  template <typename vtype>
+  using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+  typedef iImplClover<Simd> SiteCloverType;
+  typedef Lattice<SiteCloverType> CloverFieldType;
 
-  typedef WilsonFermion<Impl>       WilsonBase;
+public:
-  typedef WilsonCloverHelpers<Impl> Helpers;
+  typedef WilsonFermion<Impl> WilsonBase;
 
   virtual int    ConstEE(void)     { return 0; };
   virtual void Instantiatable(void){};
@@ -72,7 +72,42 @@ public:
                       const RealD _csw_r = 0.0,
                       const RealD _csw_t = 0.0,
                       const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
-                      const ImplParams &impl_p = ImplParams());
+                      const ImplParams &impl_p = ImplParams()) : WilsonFermion<Impl>(_Umu,
+                                                                                     Fgrid,
+                                                                                     Hgrid,
+                                                                                     _mass, impl_p, clover_anisotropy),
+                                                                 CloverTerm(&Fgrid),
+                                                                 CloverTermInv(&Fgrid),
+                                                                 CloverTermEven(&Hgrid),
+                                                                 CloverTermOdd(&Hgrid),
+                                                                 CloverTermInvEven(&Hgrid),
+                                                                 CloverTermInvOdd(&Hgrid),
+                                                                 CloverTermDagEven(&Hgrid),
+                                                                 CloverTermDagOdd(&Hgrid),
+                                                                 CloverTermInvDagEven(&Hgrid),
+                                                                 CloverTermInvDagOdd(&Hgrid)
+  {
+    assert(Nd == 4); // require 4 dimensions
+
+    if (clover_anisotropy.isAnisotropic)
+    {
+      csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
+      diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
+    }
+    else
+    {
+      csw_r = _csw_r * 0.5;
+      diag_mass = 4.0 + _mass;
+    }
+    csw_t = _csw_t * 0.5;
+
+    if (csw_r == 0)
+      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
+    if (csw_t == 0)
+      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
+
+    ImportGauge(_Umu);
+  }
 
   virtual void M(const FermionField &in, FermionField &out);
   virtual void Mdag(const FermionField &in, FermionField &out);
@@ -89,21 +124,250 @@ public:
   void ImportGauge(const GaugeField &_Umu);
 
   // Derivative parts unpreconditioned pseudofermions
-  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
+  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
+  {
+    conformable(X.Grid(), Y.Grid());
+    conformable(X.Grid(), force.Grid());
+    GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
+    GaugeField clover_force(force.Grid());
+    PropagatorField Lambda(force.Grid());
 
-public:
+    // Guido: Here we are hitting some performance issues:
+    // need to extract the components of the DoubledGaugeField
+    // for each call
+    // Possible solution
+    // Create a vector object to store them? (cons: wasting space)
+    std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
+
+    Impl::extractLinkField(U, this->Umu);
+
+    force = Zero();
+    // Derivative of the Wilson hopping term
+    this->DhopDeriv(force, X, Y, dag);
+
+    ///////////////////////////////////////////////////////////
+    // Clover term derivative
+    ///////////////////////////////////////////////////////////
+    Impl::outerProductImpl(Lambda, X, Y);
+    //std::cout << "Lambda:" << Lambda << std::endl;
+
+    Gamma::Algebra sigma[] = {
+        Gamma::Algebra::SigmaXY,
+        Gamma::Algebra::SigmaXZ,
+        Gamma::Algebra::SigmaXT,
+        Gamma::Algebra::MinusSigmaXY,
+        Gamma::Algebra::SigmaYZ,
+        Gamma::Algebra::SigmaYT,
+        Gamma::Algebra::MinusSigmaXZ,
+        Gamma::Algebra::MinusSigmaYZ,
+        Gamma::Algebra::SigmaZT,
+        Gamma::Algebra::MinusSigmaXT,
+        Gamma::Algebra::MinusSigmaYT,
+        Gamma::Algebra::MinusSigmaZT};
+
+    /*
+      sigma_{\mu \nu}=
+      | 0         sigma[0]  sigma[1]  sigma[2] |
+      | sigma[3]    0       sigma[4]  sigma[5] |
+      | sigma[6]  sigma[7]     0      sigma[8] |
+      | sigma[9]  sigma[10] sigma[11]   0      |
+    */
+
+    int count = 0;
+    clover_force = Zero();
+    for (int mu = 0; mu < 4; mu++)
+    {
+      force_mu = Zero();
+      for (int nu = 0; nu < 4; nu++)
+      {
+        if (mu == nu)
+        continue;
+        
+        RealD factor;
+        if (nu == 4 || mu == 4)
+        {
+          factor = 2.0 * csw_t;
+        }
+        else
+        {
+          factor = 2.0 * csw_r;
+        }
+        PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
+        Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
+        force_mu -= factor*Cmunu(U, lambda, mu, nu);                   // checked
+        count++;
+      }
+
+      pokeLorentz(clover_force, U[mu] * force_mu, mu);
+    }
+    //clover_force *= csw;
+    force += clover_force;
+  }
+
+  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
+  GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
+  {
+    conformable(lambda.Grid(), U[0].Grid());
+    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
+    // insertion in upper staple
+    // please check redundancy of shift operations
+
+    // C1+
+    tmp = lambda * U[nu];
+    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C2+
+    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C3+
+    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
+
+    // C4+
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
+
+    // insertion in lower staple
+    // C1-
+    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C2-
+    tmp = adj(lambda) * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C3-
+    tmp = lambda * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
+
+    // C4-
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
+
+    return out;
+  }
+
+protected:
   // here fixing the 4 dimensions, make it more general?
 
   RealD csw_r;                                               // Clover coefficient - spatial
   RealD csw_t;                                               // Clover coefficient - temporal
   RealD diag_mass;                                           // Mass term
-  CloverField CloverTerm, CloverTermInv;                     // Clover term
+  CloverFieldType CloverTerm, CloverTermInv;                 // Clover term
-  CloverField CloverTermEven, CloverTermOdd;                 // Clover term EO
+  CloverFieldType CloverTermEven, CloverTermOdd;             // Clover term EO
-  CloverField CloverTermInvEven, CloverTermInvOdd;           // Clover term Inv EO
+  CloverFieldType CloverTermInvEven, CloverTermInvOdd;       // Clover term Inv EO
-  CloverField CloverTermDagEven, CloverTermDagOdd;           // Clover term Dag EO
+  CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term Dag EO
-  CloverField CloverTermInvDagEven, CloverTermInvDagOdd;     // Clover term Inv Dag EO
+  CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
-};
 
+ public:
+  // eventually these can be compressed into 6x6 blocks instead of the 12x12
+  // using the DeGrand-Rossi basis for the gamma matrices
+  CloverFieldType fillCloverYZ(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
+    {
+      T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
+      T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
+      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverXZ(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+    
+    autoView(T_v, T,AcceleratorWrite);
+    autoView(F_v, F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
+    {
+      T_v[i]()(0, 1) = -F_v[i]()();
+      T_v[i]()(1, 0) = F_v[i]()();
+      T_v[i]()(2, 3) = -F_v[i]()();
+      T_v[i]()(3, 2) = F_v[i]()();
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverXY(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
+    {
+      T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
+      T_v[i]()(1, 1) = timesI(F_v[i]()());
+      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 3) = timesI(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverXT(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+
+    autoView( T_v , T, AcceleratorWrite);
+    autoView( F_v , F, AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
+    {
+      T_v[i]()(0, 1) = timesI(F_v[i]()());
+      T_v[i]()(1, 0) = timesI(F_v[i]()());
+      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverYT(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+    
+    autoView( T_v ,T,AcceleratorWrite);
+    autoView( F_v ,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
+    {
+      T_v[i]()(0, 1) = -(F_v[i]()());
+      T_v[i]()(1, 0) = (F_v[i]()());
+      T_v[i]()(2, 3) = (F_v[i]()());
+      T_v[i]()(3, 2) = -(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverZT(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+
+    T = Zero();
+
+    autoView( T_v , T,AcceleratorWrite);
+    autoView( F_v , F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
+    {
+      T_v[i]()(0, 0) = timesI(F_v[i]()());
+      T_v[i]()(1, 1) = timesMinusI(F_v[i]()());
+      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 3) = timesI(F_v[i]()());
+    });
+
+    return T;
+  }
+};
 NAMESPACE_END(Grid);
 
 

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -1,763 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverHelpers.h
-
-    Copyright (C) 2021 - 2022
-
-    Author: Daniel Richtmann <daniel.richtmann@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
-    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
-
-// Helper routines that implement common clover functionality
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Impl> class WilsonCloverHelpers {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
-  {
-    conformable(lambda.Grid(), U[0].Grid());
-    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
-    // insertion in upper staple
-    // please check redundancy of shift operations
-
-    // C1+
-    tmp = lambda * U[nu];
-    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C2+
-    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C3+
-    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
-
-    // C4+
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
-
-    // insertion in lower staple
-    // C1-
-    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C2-
-    tmp = adj(lambda) * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C3-
-    tmp = lambda * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
-
-    // C4-
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
-
-    return out;
-  }
-
-  static CloverField fillCloverYZ(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-    autoView(T_v,T,AcceleratorWrite);
-    autoView(F_v,F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverXZ(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-    
-    autoView(T_v, T,AcceleratorWrite);
-    autoView(F_v, F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-F_v[i]()()));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead(F_v[i]()()));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead(-F_v[i]()()));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(F_v[i]()()));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverXY(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-
-    autoView(T_v,T,AcceleratorWrite);
-    autoView(F_v,F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverXT(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-
-    autoView( T_v , T, AcceleratorWrite);
-    autoView( F_v , F, AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverYT(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-    
-    autoView( T_v ,T,AcceleratorWrite);
-    autoView( F_v ,F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead((F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead((F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(-(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverZT(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-
-    T = Zero();
-
-    autoView( T_v , T,AcceleratorWrite);
-    autoView( F_v , F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  template<class _Spinor>
-  static accelerator_inline void multClover(_Spinor& phi, const SiteClover& C, const _Spinor& chi) {
-    auto CC = coalescedRead(C);
-    mult(&phi, &CC, &chi);
-  }
-
-  template<class _SpinorField>
-  inline void multCloverField(_SpinorField& out, const CloverField& C, const _SpinorField& phi) {
-    const int Nsimd = SiteSpinor::Nsimd();
-    autoView(out_v, out, AcceleratorWrite);
-    autoView(phi_v, phi, AcceleratorRead);
-    autoView(C_v,   C,   AcceleratorRead);
-    typedef decltype(coalescedRead(out_v[0])) calcSpinor;
-    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
-      calcSpinor tmp;
-      multClover(tmp,C_v[sss],phi_v(sss));
-      coalescedWrite(out_v[sss],tmp);
-    });
-  }
-};
-
-
-////////////////////////////////////////////////////////
-
-template<class Impl> class CompactWilsonCloverHelpers {
-public:
-
-  INHERIT_COMPACT_CLOVER_SIZES(Impl);
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  #if 0
-  static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
-    assert(i != j);
-    if(i < j) {
-      return triangle()(block)(triangle_index(i, j));
-    } else { // i > j
-      return conjugate(triangle()(block)(triangle_index(i, j)));
-    }
-  }
-  #else
-  template<typename vobj>
-  static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
-    assert(i != j);
-    if(i < j) {
-      return triangle()(block)(triangle_index(i, j));
-    } else { // i > j
-      return conjugate(triangle()(block)(triangle_index(i, j)));
-    }
-  }
-  #endif
-
-  static accelerator_inline int triangle_index(int i, int j) {
-    if(i == j)
-      return 0;
-    else if(i < j)
-      return Nred * (Nred - 1) / 2 - (Nred - i) * (Nred - i - 1) / 2 + j - i - 1;
-    else // i > j
-      return Nred * (Nred - 1) / 2 - (Nred - j) * (Nred - j - 1) / 2 + i - j - 1;
-  }
-
-  static void MooeeKernel_gpu(int                        Nsite,
-                              int                        Ls,
-                              const FermionField&        in,
-                              FermionField&              out,
-                              const CloverDiagonalField& diagonal,
-                              const CloverTriangleField& triangle) {
-    autoView(diagonal_v, diagonal, AcceleratorRead);
-    autoView(triangle_v, triangle, AcceleratorRead);
-    autoView(in_v,       in,       AcceleratorRead);
-    autoView(out_v,      out,      AcceleratorWrite);
-
-    typedef decltype(coalescedRead(out_v[0])) CalcSpinor;
-
-    const uint64_t NN = Nsite * Ls;
-
-    accelerator_for(ss, NN, Simd::Nsimd(), {
-      int sF = ss;
-      int sU = ss/Ls;
-      CalcSpinor res;
-      CalcSpinor in_t = in_v(sF);
-      auto diagonal_t = diagonal_v(sU);
-      auto triangle_t = triangle_v(sU);
-      for(int block=0; block<Nhs; block++) {
-        int s_start = block*Nhs;
-        for(int i=0; i<Nred; i++) {
-          int si = s_start + i/Nc, ci = i%Nc;
-          res()(si)(ci) = diagonal_t()(block)(i) * in_t()(si)(ci);
-          for(int j=0; j<Nred; j++) {
-            if (j == i) continue;
-            int sj = s_start + j/Nc, cj = j%Nc;
-            res()(si)(ci) = res()(si)(ci) + triangle_elem(triangle_t, block, i, j) * in_t()(sj)(cj);
-          };
-        };
-      };
-      coalescedWrite(out_v[sF], res);
-    });
-  }
-
-  static void MooeeKernel_cpu(int                        Nsite,
-                              int                        Ls,
-                              const FermionField&        in,
-                              FermionField&              out,
-                              const CloverDiagonalField& diagonal,
-                              const CloverTriangleField& triangle) {
-    autoView(diagonal_v, diagonal, CpuRead);
-    autoView(triangle_v, triangle, CpuRead);
-    autoView(in_v,       in,       CpuRead);
-    autoView(out_v,      out,      CpuWrite);
-
-    typedef SiteSpinor CalcSpinor;
-
-#if defined(A64FX) || defined(A64FXFIXEDSIZE)
-#define PREFETCH_CLOVER(BASE) {                                     \
-    uint64_t base;                                                  \
-    int pf_dist_L1 = 1;                                             \
-    int pf_dist_L2 = -5; /* -> penalty -> disable */                \
-                                                                    \
-    if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {           \
-      base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);               \
-      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL1STRM); \
-    }                                                               \
-                                                                    \
-    if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {           \
-      base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);               \
-      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL2STRM); \
-    }                                                               \
-  }
-// TODO: Implement/generalize this for other architectures
-// I played around a bit on KNL (see below) but didn't bring anything
-// #elif defined(AVX512)
-// #define PREFETCH_CLOVER(BASE) {                              \
-//     uint64_t base;                                           \
-//     int pf_dist_L1 = 1;                                      \
-//     int pf_dist_L2 = +4;                                     \
-//                                                              \
-//     if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {    \
-//       base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);        \
-//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T0); \
-//     }                                                        \
-//                                                              \
-//     if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {    \
-//       base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);        \
-//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T1); \
-//     }                                                        \
-//   }
-#else
-#define PREFETCH_CLOVER(BASE)
-#endif
-
-    const uint64_t NN = Nsite * Ls;
-
-    thread_for(ss, NN, {
-      int sF = ss;
-      int sU = ss/Ls;
-      CalcSpinor res;
-      CalcSpinor in_t = in_v[sF];
-      auto diag_t     = diagonal_v[sU]; // "diag" instead of "diagonal" here to make code below easier to read
-      auto triangle_t = triangle_v[sU];
-
-      // upper half
-      PREFETCH_CLOVER(0);
-
-      auto in_cc_0_0 = conjugate(in_t()(0)(0)); // Nils: reduces number
-      auto in_cc_0_1 = conjugate(in_t()(0)(1)); // of conjugates from
-      auto in_cc_0_2 = conjugate(in_t()(0)(2)); // 30 to 20
-      auto in_cc_1_0 = conjugate(in_t()(1)(0));
-      auto in_cc_1_1 = conjugate(in_t()(1)(1));
-
-      res()(0)(0) =               diag_t()(0)( 0) * in_t()(0)(0)
-                  +           triangle_t()(0)( 0) * in_t()(0)(1)
-                  +           triangle_t()(0)( 1) * in_t()(0)(2)
-                  +           triangle_t()(0)( 2) * in_t()(1)(0)
-                  +           triangle_t()(0)( 3) * in_t()(1)(1)
-                  +           triangle_t()(0)( 4) * in_t()(1)(2);
-
-      res()(0)(1) =           triangle_t()(0)( 0) * in_cc_0_0;
-      res()(0)(1) =               diag_t()(0)( 1) * in_t()(0)(1)
-                  +           triangle_t()(0)( 5) * in_t()(0)(2)
-                  +           triangle_t()(0)( 6) * in_t()(1)(0)
-                  +           triangle_t()(0)( 7) * in_t()(1)(1)
-                  +           triangle_t()(0)( 8) * in_t()(1)(2)
-                  + conjugate(       res()(0)( 1));
-
-      res()(0)(2) =           triangle_t()(0)( 1) * in_cc_0_0
-                  +           triangle_t()(0)( 5) * in_cc_0_1;
-      res()(0)(2) =               diag_t()(0)( 2) * in_t()(0)(2)
-                  +           triangle_t()(0)( 9) * in_t()(1)(0)
-                  +           triangle_t()(0)(10) * in_t()(1)(1)
-                  +           triangle_t()(0)(11) * in_t()(1)(2)
-                  + conjugate(       res()(0)( 2));
-
-      res()(1)(0) =           triangle_t()(0)( 2) * in_cc_0_0
-                  +           triangle_t()(0)( 6) * in_cc_0_1
-                  +           triangle_t()(0)( 9) * in_cc_0_2;
-      res()(1)(0) =               diag_t()(0)( 3) * in_t()(1)(0)
-                  +           triangle_t()(0)(12) * in_t()(1)(1)
-                  +           triangle_t()(0)(13) * in_t()(1)(2)
-                  + conjugate(       res()(1)( 0));
-
-      res()(1)(1) =           triangle_t()(0)( 3) * in_cc_0_0
-                  +           triangle_t()(0)( 7) * in_cc_0_1
-                  +           triangle_t()(0)(10) * in_cc_0_2
-                  +           triangle_t()(0)(12) * in_cc_1_0;
-      res()(1)(1) =               diag_t()(0)( 4) * in_t()(1)(1)
-                  +           triangle_t()(0)(14) * in_t()(1)(2)
-                  + conjugate(       res()(1)( 1));
-
-      res()(1)(2) =           triangle_t()(0)( 4) * in_cc_0_0
-                  +           triangle_t()(0)( 8) * in_cc_0_1
-                  +           triangle_t()(0)(11) * in_cc_0_2
-                  +           triangle_t()(0)(13) * in_cc_1_0
-                  +           triangle_t()(0)(14) * in_cc_1_1;
-      res()(1)(2) =               diag_t()(0)( 5) * in_t()(1)(2)
-                  + conjugate(       res()(1)( 2));
-
-      vstream(out_v[sF]()(0)(0), res()(0)(0));
-      vstream(out_v[sF]()(0)(1), res()(0)(1));
-      vstream(out_v[sF]()(0)(2), res()(0)(2));
-      vstream(out_v[sF]()(1)(0), res()(1)(0));
-      vstream(out_v[sF]()(1)(1), res()(1)(1));
-      vstream(out_v[sF]()(1)(2), res()(1)(2));
-
-      // lower half
-      PREFETCH_CLOVER(1);
-
-      auto in_cc_2_0 = conjugate(in_t()(2)(0));
-      auto in_cc_2_1 = conjugate(in_t()(2)(1));
-      auto in_cc_2_2 = conjugate(in_t()(2)(2));
-      auto in_cc_3_0 = conjugate(in_t()(3)(0));
-      auto in_cc_3_1 = conjugate(in_t()(3)(1));
-
-      res()(2)(0) =               diag_t()(1)( 0) * in_t()(2)(0)
-                  +           triangle_t()(1)( 0) * in_t()(2)(1)
-                  +           triangle_t()(1)( 1) * in_t()(2)(2)
-                  +           triangle_t()(1)( 2) * in_t()(3)(0)
-                  +           triangle_t()(1)( 3) * in_t()(3)(1)
-                  +           triangle_t()(1)( 4) * in_t()(3)(2);
-
-      res()(2)(1) =           triangle_t()(1)( 0) * in_cc_2_0;
-      res()(2)(1) =               diag_t()(1)( 1) * in_t()(2)(1)
-                  +           triangle_t()(1)( 5) * in_t()(2)(2)
-                  +           triangle_t()(1)( 6) * in_t()(3)(0)
-                  +           triangle_t()(1)( 7) * in_t()(3)(1)
-                  +           triangle_t()(1)( 8) * in_t()(3)(2)
-                  + conjugate(       res()(2)( 1));
-
-      res()(2)(2) =           triangle_t()(1)( 1) * in_cc_2_0
-                  +           triangle_t()(1)( 5) * in_cc_2_1;
-      res()(2)(2) =               diag_t()(1)( 2) * in_t()(2)(2)
-                  +           triangle_t()(1)( 9) * in_t()(3)(0)
-                  +           triangle_t()(1)(10) * in_t()(3)(1)
-                  +           triangle_t()(1)(11) * in_t()(3)(2)
-                  + conjugate(       res()(2)( 2));
-
-      res()(3)(0) =           triangle_t()(1)( 2) * in_cc_2_0
-                  +           triangle_t()(1)( 6) * in_cc_2_1
-                  +           triangle_t()(1)( 9) * in_cc_2_2;
-      res()(3)(0) =               diag_t()(1)( 3) * in_t()(3)(0)
-                  +           triangle_t()(1)(12) * in_t()(3)(1)
-                  +           triangle_t()(1)(13) * in_t()(3)(2)
-                  + conjugate(       res()(3)( 0));
-
-      res()(3)(1) =           triangle_t()(1)( 3) * in_cc_2_0
-                  +           triangle_t()(1)( 7) * in_cc_2_1
-                  +           triangle_t()(1)(10) * in_cc_2_2
-                  +           triangle_t()(1)(12) * in_cc_3_0;
-      res()(3)(1) =               diag_t()(1)( 4) * in_t()(3)(1)
-                  +           triangle_t()(1)(14) * in_t()(3)(2)
-                  + conjugate(       res()(3)( 1));
-
-      res()(3)(2) =           triangle_t()(1)( 4) * in_cc_2_0
-                  +           triangle_t()(1)( 8) * in_cc_2_1
-                  +           triangle_t()(1)(11) * in_cc_2_2
-                  +           triangle_t()(1)(13) * in_cc_3_0
-                  +           triangle_t()(1)(14) * in_cc_3_1;
-      res()(3)(2) =               diag_t()(1)( 5) * in_t()(3)(2)
-                  + conjugate(       res()(3)( 2));
-
-      vstream(out_v[sF]()(2)(0), res()(2)(0));
-      vstream(out_v[sF]()(2)(1), res()(2)(1));
-      vstream(out_v[sF]()(2)(2), res()(2)(2));
-      vstream(out_v[sF]()(3)(0), res()(3)(0));
-      vstream(out_v[sF]()(3)(1), res()(3)(1));
-      vstream(out_v[sF]()(3)(2), res()(3)(2));
-    });
-  }
-
-  static void MooeeKernel(int                        Nsite,
-                          int                        Ls,
-                          const FermionField&        in,
-                          FermionField&              out,
-                          const CloverDiagonalField& diagonal,
-                          const CloverTriangleField& triangle) {
-#if defined(GRID_CUDA) || defined(GRID_HIP)
-    MooeeKernel_gpu(Nsite, Ls, in, out, diagonal, triangle);
-#else
-    MooeeKernel_cpu(Nsite, Ls, in, out, diagonal, triangle);
-#endif
-  }
-
-  static void Invert(const CloverDiagonalField& diagonal,
-                     const CloverTriangleField& triangle,
-                     CloverDiagonalField&       diagonalInv,
-                     CloverTriangleField&       triangleInv) {
-    conformable(diagonal, diagonalInv);
-    conformable(triangle, triangleInv);
-    conformable(diagonal, triangle);
-
-    diagonalInv.Checkerboard() = diagonal.Checkerboard();
-    triangleInv.Checkerboard() = triangle.Checkerboard();
-
-    GridBase* grid = diagonal.Grid();
-
-    long lsites = grid->lSites();
-
-    typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
-    typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-
-    autoView(diagonal_v,  diagonal,  CpuRead);
-    autoView(triangle_v,  triangle,  CpuRead);
-    autoView(diagonalInv_v, diagonalInv, CpuWrite);
-    autoView(triangleInv_v, triangleInv, CpuWrite);
-
-    thread_for(site, lsites, { // NOTE: Not on GPU because of Eigen & (peek/poke)LocalSite
-      Eigen::MatrixXcd clover_inv_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
-      Eigen::MatrixXcd clover_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
-
-      scalar_object_diagonal diagonal_tmp     = Zero();
-      scalar_object_diagonal diagonal_inv_tmp = Zero();
-      scalar_object_triangle triangle_tmp     = Zero();
-      scalar_object_triangle triangle_inv_tmp = Zero();
-
-      Coordinate lcoor;
-      grid->LocalIndexToLocalCoor(site, lcoor);
-
-      peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-      peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-      // TODO: can we save time here by inverting the two 6x6 hermitian matrices separately?
-      for (long s_row=0;s_row<Ns;s_row++) {
-        for (long s_col=0;s_col<Ns;s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for (long c_row=0;c_row<Nc;c_row++) {
-            for (long c_col=0;c_col<Nc;c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-              else
-                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(triangle_elem(triangle_tmp, block, i, j)));
-            }
-          }
-        }
-      }
-
-      clover_inv_eigen = clover_eigen.inverse();
-
-      for (long s_row=0;s_row<Ns;s_row++) {
-        for (long s_col=0;s_col<Ns;s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for (long c_row=0;c_row<Nc;c_row++) {
-            for (long c_col=0;c_col<Nc;c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                diagonal_inv_tmp()(block)(i) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
-              else if(i < j)
-                triangle_inv_tmp()(block)(triangle_index(i, j)) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
-              else
-                continue;
-            }
-          }
-        }
-      }
-
-      pokeLocalSite(diagonal_inv_tmp, diagonalInv_v, lcoor);
-      pokeLocalSite(triangle_inv_tmp, triangleInv_v, lcoor);
-    });
-  }
-
-  static void ConvertLayout(const CloverField&   full,
-                            CloverDiagonalField& diagonal,
-                            CloverTriangleField& triangle) {
-    conformable(full, diagonal);
-    conformable(full, triangle);
-
-    diagonal.Checkerboard() = full.Checkerboard();
-    triangle.Checkerboard() = full.Checkerboard();
-
-    autoView(full_v,     full,     AcceleratorRead);
-    autoView(diagonal_v, diagonal, AcceleratorWrite);
-    autoView(triangle_v, triangle, AcceleratorWrite);
-
-    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
-    accelerator_for(ss, full.Grid()->oSites(), 1, {
-      for(int s_row = 0; s_row < Ns; s_row++) {
-        for(int s_col = 0; s_col < Ns; s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for(int c_row = 0; c_row < Nc; c_row++) {
-            for(int c_col = 0; c_col < Nc; c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                diagonal_v[ss]()(block)(i) = full_v[ss]()(s_row, s_col)(c_row, c_col);
-              else if(i < j)
-                triangle_v[ss]()(block)(triangle_index(i, j)) = full_v[ss]()(s_row, s_col)(c_row, c_col);
-              else
-                continue;
-            }
-          }
-        }
-      }
-    });
-  }
-
-
-  static void ConvertLayout(const CloverDiagonalField& diagonal,
-                            const CloverTriangleField& triangle,
-                            CloverField&               full) {
-    conformable(full, diagonal);
-    conformable(full, triangle);
-
-    full.Checkerboard() = diagonal.Checkerboard();
-
-    full = Zero();
-
-    autoView(diagonal_v, diagonal, AcceleratorRead);
-    autoView(triangle_v, triangle, AcceleratorRead);
-    autoView(full_v,     full,     AcceleratorWrite);
-
-    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
-    accelerator_for(ss, full.Grid()->oSites(), 1, {
-      for(int s_row = 0; s_row < Ns; s_row++) {
-        for(int s_col = 0; s_col < Ns; s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for(int c_row = 0; c_row < Nc; c_row++) {
-            for(int c_col = 0; c_col < Nc; c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                full_v[ss]()(s_row, s_col)(c_row, c_col) = diagonal_v[ss]()(block)(i);
-              else
-                full_v[ss]()(s_row, s_col)(c_row, c_col) = triangle_elem(triangle_v[ss], block, i, j);
-            }
-          }
-        }
-      }
-    });
-  }
-
-  static void ModifyBoundaries(CloverDiagonalField& diagonal, CloverTriangleField& triangle, RealD csw_t, RealD cF, RealD diag_mass) {
-    // Checks/grid
-    double t0 = usecond();
-    conformable(diagonal, triangle);
-    GridBase* grid = diagonal.Grid();
-
-    // Determine the boundary coordinates/sites
-    double t1 = usecond();
-    int t_dir = Nd - 1;
-    Lattice<iScalar<vInteger>> t_coor(grid);
-    LatticeCoordinate(t_coor, t_dir);
-    int T = grid->GlobalDimensions()[t_dir];
-
-    // Set off-diagonal parts at boundary to zero -- OK
-    double t2 = usecond();
-    CloverTriangleField zeroTriangle(grid);
-    zeroTriangle.Checkerboard() = triangle.Checkerboard();
-    zeroTriangle = Zero();
-    triangle = where(t_coor == 0,   zeroTriangle, triangle);
-    triangle = where(t_coor == T-1, zeroTriangle, triangle);
-
-    // Set diagonal to unity (scaled correctly) -- OK
-    double t3 = usecond();
-    CloverDiagonalField tmp(grid);
-    tmp.Checkerboard() = diagonal.Checkerboard();
-    tmp                = -1.0 * csw_t + diag_mass;
-    diagonal           = where(t_coor == 0,   tmp, diagonal);
-    diagonal           = where(t_coor == T-1, tmp, diagonal);
-
-    // Correct values next to boundary
-    double t4 = usecond();
-    if(cF != 1.0) {
-      tmp = cF - 1.0;
-      tmp += diagonal;
-      diagonal = where(t_coor == 1,   tmp, diagonal);
-      diagonal = where(t_coor == T-2, tmp, diagonal);
-    }
-
-    // Report timings
-    double t5 = usecond();
-#if 0
-    std::cout << GridLogMessage << "CompactWilsonCloverHelpers::ModifyBoundaries timings:"
-              << " checks = "          << (t1 - t0) / 1e6
-              << ", coordinate = "     << (t2 - t1) / 1e6
-              << ", off-diag zero = "  << (t3 - t2) / 1e6
-              << ", diagonal unity = " << (t4 - t3) / 1e6
-              << ", near-boundary = "  << (t5 - t4) / 1e6
-              << ", total = "          << (t5 - t0) / 1e6
-              << std::endl;
-#endif
-  }
-
-  template<class Field, class Mask>
-  static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
-    conformable(f, m);
-    auto grid  = f.Grid();
-    const uint32_t Nsite = grid->oSites();
-    const uint32_t Nsimd = grid->Nsimd();
-    autoView(f_v, f, AcceleratorWrite);
-    autoView(m_v, m, AcceleratorRead);
-    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
-    accelerator_for(ss, Nsite, Nsimd, {
-      coalescedWrite(f_v[ss], m_v(ss) * f_v(ss));
-    });
-  }
-
-  template<class MaskField>
-  static void SetupMasks(MaskField& full, MaskField& even, MaskField& odd) {
-    assert(even.Grid()->_isCheckerBoarded && even.Checkerboard() == Even);
-    assert(odd.Grid()->_isCheckerBoarded  && odd.Checkerboard()  == Odd);
-    assert(!full.Grid()->_isCheckerBoarded);
-
-    GridBase* grid = full.Grid();
-    int t_dir = Nd-1;
-    Lattice<iScalar<vInteger>> t_coor(grid);
-    LatticeCoordinate(t_coor, t_dir);
-    int T = grid->GlobalDimensions()[t_dir];
-
-    MaskField zeroMask(grid); zeroMask = Zero();
-    full = 1.0;
-    full = where(t_coor == 0,   zeroMask, full);
-    full = where(t_coor == T-1, zeroMask, full);
-
-    pickCheckerboard(Even, even, full);
-    pickCheckerboard(Odd,  odd,  full);
-  }
-};
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -1,90 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverTypes.h
-
-    Copyright (C) 2021 - 2022
-
-    Author: Daniel Richtmann <daniel.richtmann@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
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Impl>
-class WilsonCloverTypes {
-public:
-  INHERIT_IMPL_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-
-  typedef iImplClover<Simd> SiteClover;
-
-  typedef Lattice<SiteClover> CloverField;
-};
-
-template<class Impl>
-class CompactWilsonCloverTypes {
-public:
-  INHERIT_IMPL_TYPES(Impl);
-
-  static constexpr int Nred      = Nc * Nhs;        // 6
-  static constexpr int Nblock    = Nhs;             // 2
-  static constexpr int Ndiagonal = Nred;            // 6
-  static constexpr int Ntriangle = (Nred - 1) * Nc; // 15
-
-  template<typename vtype> using iImplCloverDiagonal = iScalar<iVector<iVector<vtype, Ndiagonal>, Nblock>>;
-  template<typename vtype> using iImplCloverTriangle = iScalar<iVector<iVector<vtype, Ntriangle>, Nblock>>;
-
-  typedef iImplCloverDiagonal<Simd> SiteCloverDiagonal;
-  typedef iImplCloverTriangle<Simd> SiteCloverTriangle;
-  typedef iSinglet<Simd>            SiteMask;
-
-  typedef Lattice<SiteCloverDiagonal> CloverDiagonalField;
-  typedef Lattice<SiteCloverTriangle> CloverTriangleField;
-  typedef Lattice<SiteMask>           MaskField;
-};
-
-#define INHERIT_CLOVER_TYPES(Impl)                                 \
-  typedef typename WilsonCloverTypes<Impl>::SiteClover SiteClover; \
-  typedef typename WilsonCloverTypes<Impl>::CloverField CloverField;
-
-#define INHERIT_COMPACT_CLOVER_TYPES(Impl) \
-  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverDiagonal  SiteCloverDiagonal; \
-  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverTriangle  SiteCloverTriangle; \
-  typedef typename CompactWilsonCloverTypes<Impl>::SiteMask            SiteMask; \
-  typedef typename CompactWilsonCloverTypes<Impl>::CloverDiagonalField CloverDiagonalField; \
-  typedef typename CompactWilsonCloverTypes<Impl>::CloverTriangleField CloverTriangleField; \
-  typedef typename CompactWilsonCloverTypes<Impl>::MaskField           MaskField; \
-  /* ugly duplication but needed inside functionality classes */ \
-  template<typename vtype> using iImplCloverDiagonal = \
-    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ndiagonal>, CompactWilsonCloverTypes<Impl>::Nblock>>; \
-  template<typename vtype> using iImplCloverTriangle = \
-    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ntriangle>, CompactWilsonCloverTypes<Impl>::Nblock>>;
-
-#define INHERIT_COMPACT_CLOVER_SIZES(Impl)                                    \
-  static constexpr int Nred      = CompactWilsonCloverTypes<Impl>::Nred;      \
-  static constexpr int Nblock    = CompactWilsonCloverTypes<Impl>::Nblock;    \
-  static constexpr int Ndiagonal = CompactWilsonCloverTypes<Impl>::Ndiagonal; \
-  static constexpr int Ntriangle = CompactWilsonCloverTypes<Impl>::Ntriangle;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -294,7 +294,11 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  //  Vector<int> surface_list;
+  void ZeroCountersi(void)  {  }
+  void Reporti(int calls)  {  }
+
+  std::vector<int> surface_list;
+
   WilsonStencil(GridBase *grid,
 		int npoints,
 		int checkerboard,
@@ -302,11 +306,11 @@ public:
 		const std::vector<int> &distances,Parameters p)  
     : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
   { 
-    //    surface_list.resize(0);
+    ZeroCountersi();
+    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
 
-  /*
   void BuildSurfaceList(int Ls,int vol4){
 
     // find same node for SHM
@@ -327,8 +331,7 @@ public:
       }
     }
   }
-  */
+
-  
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
   {
@@ -395,6 +398,7 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
+    accelerator_barrier();
   }
 
 };

Grid/qcd/action/fermion/WilsonFermion.h

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

Grid/qcd/action/fermion/WilsonFermion5D.h

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

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

@@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FiveDimRedBlackGrid,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
-  mass_plus(_mass), mass_minus(_mass)
+  mass(_mass)
 { 
 }
 
@@ -152,13 +152,65 @@ void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi
   }
 }
 
+template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
+{
+  this->Report();
+  Coordinate latt = GridDefaultLatt();          
+  RealD volume = this->Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP     = this->_FourDimGrid->_Nprocessors;
+  if ( M5Dcalls > 0 ) {
+    std::cout << GridLogMessage << "#### M5D calls report " << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D Number of M5D Calls     : " << M5Dcalls   << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls       : " << M5Dtime / M5Dcalls << " us" << std::endl;
+
+    // Flops = 10.0*(Nc*Ns) *Ls*vol
+    RealD mflops = 10.0*(Nc*Ns)*volume*M5Dcalls/M5Dtime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+
+    // Bytes = sizeof(Real) * (Nc*Ns*Nreim) * Ls * vol * (read+write) (/2 for red black counting)
+    // read = 2 ( psi[ss+s+1] and psi[ss+s-1] count as 1 )
+    // write = 1
+    RealD Gbytes = sizeof(Real) * (Nc*Ns*2) * volume * 3 /2. * 1.e-9;
+    std::cout << GridLogMessage << "Average bandwidth (GB/s)                 : " << Gbytes/M5Dtime*M5Dcalls*1.e6 << std::endl;
+  }
+
+  if ( MooeeInvCalls > 0 ) {
+
+    std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls     : " << MooeeInvCalls   << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls            : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
+#ifdef GRID_CUDA
+    RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+#else
+    // Flops = MADD * Ls *Ls *4dvol * spin/colour/complex
+    RealD mflops = 2.0*24*this->Ls*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+#endif
+  }
+
+}
+template<class Impl> void CayleyFermion5D<Impl>::CayleyZeroCounters(void)
+{
+  this->ZeroCounters();
+  M5Dflops=0;
+  M5Dcalls=0;
+  M5Dtime=0;
+  MooeeInvFlops=0;
+  MooeeInvCalls=0;
+  MooeeInvTime=0;
+}
+
 template<class Impl>  
 void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
   Vector<Coeff_t> diag (Ls,1.0);
-  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
-  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -168,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
   Vector<Coeff_t> diag = bs;
   Vector<Coeff_t> upper= cs;
   Vector<Coeff_t> lower= cs; 
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
 }
 // FIXME Redunant with the above routine; check this and eliminate
@@ -183,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -198,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &
     upper[i]=-cee[i];
     lower[i]=-cee[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -214,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &
     // Assemble the 5d matrix
     if ( s==0 ) {
       upper[s] = -cee[s+1] ;
-      lower[s] = mass_minus*cee[Ls-1];
+      lower[s] = mass*cee[Ls-1];
     } else if ( s==(Ls-1)) { 
-      upper[s] = mass_plus*cee[0];
+      upper[s] = mass*cee[0];
       lower[s] = -cee[s-1];
     } else {
       upper[s]=-cee[s+1];
@@ -239,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
   Vector<Coeff_t> diag(Ls,1.0);
   Vector<Coeff_t> upper(Ls,-1.0);
   Vector<Coeff_t> lower(Ls,-1.0);
-  upper[Ls-1]=-mass_plus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_minus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
 }
 
@@ -255,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
       upper[s] = cs[s+1];
-      lower[s] =-mass_minus*cs[Ls-1];
+      lower[s] =-mass*cs[Ls-1];
     } else if ( s==(Ls-1) ) { 
-      upper[s] =-mass_plus*cs[0];
+      upper[s] =-mass*cs[0];
       lower[s] = cs[s-1];
     } else { 
       upper[s] = cs[s+1];
@@ -500,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
-      leem[i]=mass_minus*cee[Ls-1]/bee[0];
+      leem[i]=mass*cee[Ls-1]/bee[0];
       for(int j=0;j<i;j++) {
 	assert(bee[j+1]!=Coeff_t(0.0));
 	leem[i]*= aee[j]/bee[j+1];
@@ -508,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
-      ueem[i]=mass_plus;
+      ueem[i]=mass;
       for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
       ueem[i]*= aee[0]/bee[0];
       
@@ -521,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
   }
 	
   { 
-    Coeff_t delta_d=mass_minus*cee[Ls-1];
+    Coeff_t delta_d=mass*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
       assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
@@ -590,10 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-
+#if (!defined(GRID_HIP))
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
-  
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -712,7 +761,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
     else          q_out +=     C;
     
   }
-
+#endif
 }
 
 template <class Impl>
@@ -728,8 +777,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   assert(mu>=0);
   assert(mu<Nd);
 
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
 
 #if 0
   int tshift = (mu == Nd-1) ? 1 : 0;
@@ -779,6 +826,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
+#if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
@@ -898,6 +946,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     InsertSlice(L_Q, q_out, s , 0);
   }
+#endif
 }
 #undef Pp
 #undef Pm

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

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

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

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

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

@@ -1,373 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermionImplementation.h
-
-    Copyright (C) 2017 - 2022
-
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Guido Cossu <guido.cossu@ed.ac.uk>
-    Author: Daniel Richtmann <daniel.richtmann@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
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-/*  END LEGAL */
-
-#include <Grid/Grid.h>
-#include <Grid/qcd/spin/Dirac.h>
-#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
-
-
-NAMESPACE_BEGIN(Grid);
-template<class Impl, class CloverHelpers>
-CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
-                                                                            GridCartesian& Fgrid,
-                                                                            GridRedBlackCartesian& Hgrid,
-                                                                            const RealD _mass,
-                                                                            const RealD _csw_r,
-                                                                            const RealD _csw_t,
-                                                                            const RealD _cF,
-                                                                            const WilsonAnisotropyCoefficients& clover_anisotropy,
-                                                                            const ImplParams& impl_p)
-  : 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)
-  , Diagonal(&Fgrid),        Triangle(&Fgrid)
-  , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
-  , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
-  , DiagonalInv(&Fgrid),     TriangleInv(&Fgrid)
-  , DiagonalInvEven(&Hgrid), TriangleInvEven(&Hgrid)
-  , DiagonalInvOdd(&Hgrid),  TriangleInvOdd(&Hgrid)
-  , Tmp(&Fgrid)
-  , BoundaryMask(&Fgrid)
-  , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
-{
-  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
-
-  csw_r *= 0.5;
-  csw_t *= 0.5;
-  if (clover_anisotropy.isAnisotropic)
-    csw_r /= clover_anisotropy.xi_0;
-
-  ImportGauge(_Umu);
-  if (open_boundaries) {
-    this->BoundaryMaskEven.Checkerboard() = Even;
-    this->BoundaryMaskOdd.Checkerboard() = Odd;
-    CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
-  }
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
-  WilsonBase::Dhop(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
-  WilsonBase::DhopOE(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
-  WilsonBase::DhopEO(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
-  WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->open_boundaries) ApplyBoundaryMask(out);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
-  WilsonBase::DhopDirAll(in, out);
-  if(this->open_boundaries) {
-    for(auto& o : out) ApplyBoundaryMask(o);
-  }
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
-  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);
-}
-
-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);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
-  WilsonBase::Meooe(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
-  WilsonBase::MeooeDag(in, out);
-  if(open_boundaries) ApplyBoundaryMask(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);
-    } else {
-      MooeeInternal(in, out, DiagonalEven, TriangleEven);
-    }
-  } else {
-    MooeeInternal(in, out, Diagonal, Triangle);
-  }
-  if(open_boundaries) ApplyBoundaryMask(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, 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);
-    } else {
-      MooeeInternal(in, out, DiagonalInvEven, TriangleInvEven);
-    }
-  } else {
-    MooeeInternal(in, out, DiagonalInv, TriangleInv);
-  }
-  if(open_boundaries) ApplyBoundaryMask(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, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
-  DhopDir(in, out, dir, disp);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
-  DhopDirAll(in, out);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!open_boundaries); // TODO check for changes required for open bc
-
-  // NOTE: code copied from original clover term
-  conformable(X.Grid(), Y.Grid());
-  conformable(X.Grid(), force.Grid());
-  GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
-  GaugeField clover_force(force.Grid());
-  PropagatorField Lambda(force.Grid());
-
-  // Guido: Here we are hitting some performance issues:
-  // need to extract the components of the DoubledGaugeField
-  // for each call
-  // Possible solution
-  // Create a vector object to store them? (cons: wasting space)
-  std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
-
-  Impl::extractLinkField(U, this->Umu);
-
-  force = Zero();
-  // Derivative of the Wilson hopping term
-  this->DhopDeriv(force, X, Y, dag);
-
-  ///////////////////////////////////////////////////////////
-  // Clover term derivative
-  ///////////////////////////////////////////////////////////
-  Impl::outerProductImpl(Lambda, X, Y);
-  //std::cout << "Lambda:" << Lambda << std::endl;
-
-  Gamma::Algebra sigma[] = {
-      Gamma::Algebra::SigmaXY,
-      Gamma::Algebra::SigmaXZ,
-      Gamma::Algebra::SigmaXT,
-      Gamma::Algebra::MinusSigmaXY,
-      Gamma::Algebra::SigmaYZ,
-      Gamma::Algebra::SigmaYT,
-      Gamma::Algebra::MinusSigmaXZ,
-      Gamma::Algebra::MinusSigmaYZ,
-      Gamma::Algebra::SigmaZT,
-      Gamma::Algebra::MinusSigmaXT,
-      Gamma::Algebra::MinusSigmaYT,
-      Gamma::Algebra::MinusSigmaZT};
-
-  /*
-    sigma_{\mu \nu}=
-    | 0         sigma[0]  sigma[1]  sigma[2] |
-    | sigma[3]    0       sigma[4]  sigma[5] |
-    | sigma[6]  sigma[7]     0      sigma[8] |
-    | sigma[9]  sigma[10] sigma[11]   0      |
-  */
-
-  int count = 0;
-  clover_force = Zero();
-  for (int mu = 0; mu < 4; mu++)
-  {
-    force_mu = Zero();
-    for (int nu = 0; nu < 4; nu++)
-    {
-      if (mu == nu)
-        continue;
-
-      RealD factor;
-      if (nu == 4 || mu == 4)
-      {
-        factor = 2.0 * csw_t;
-      }
-      else
-      {
-        factor = 2.0 * csw_r;
-      }
-      PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
-      Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
-      count++;
-    }
-
-    pokeLorentz(clover_force, U[mu] * force_mu, mu);
-  }
-  //clover_force *= csw;
-  force += clover_force;
-}
-
-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, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
-  assert(0);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
-                    FermionField&              out,
-                    const CloverDiagonalField& diagonal,
-                    const CloverTriangleField& triangle) {
-  assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
-  out.Checkerboard() = in.Checkerboard();
-  conformable(in, out);
-  conformable(in, diagonal);
-  conformable(in, triangle);
-
-  CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
-}
-
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
-  // NOTE: parts copied from original implementation
-
-  // Import gauge into base class
-  double t0 = usecond();
-  WilsonBase::ImportGauge(_Umu); // NOTE: called here and in wilson constructor -> performed twice, but can't avoid that
-
-  // Initialize temporary variables
-  double t1 = usecond();
-  conformable(_Umu.Grid(), this->GaugeGrid());
-  GridBase* grid = _Umu.Grid();
-  typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
-  CloverField TmpOriginal(grid);
-
-  // Compute the field strength terms mu>nu
-  double t2 = usecond();
-  WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
-  WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
-  WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
-  WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
-  WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
-  WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
-
-  // Compute the Clover Operator acting on Colour and Spin
-  // multiply here by the clover coefficients for the anisotropy
-  double t3 = usecond();
-  TmpOriginal  = Helpers::fillCloverYZ(Bx) * csw_r;
-  TmpOriginal += Helpers::fillCloverXZ(By) * csw_r;
-  TmpOriginal += Helpers::fillCloverXY(Bz) * csw_r;
-  TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
-  TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
-  TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  // Handle mass term based on clover policy
-  CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
-  
-  // Convert the data layout of the clover term
-  double t4 = usecond();
-  CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
-
-  // Exponentiate the clover (nothing happens in case of the standard clover)
-  double t5 = usecond();
-  CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
-
-  // Possible modify the boundary values
-  double t6 = usecond();
-  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
-
-  // Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
-  double t7 = usecond();
-  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
-
-  // Fill the remaining clover fields
-  double t8 = usecond();
-  pickCheckerboard(Even, DiagonalEven,    Diagonal);
-  pickCheckerboard(Even, TriangleEven,    Triangle);
-  pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
-  pickCheckerboard(Odd,  TriangleOdd,     Triangle);
-  pickCheckerboard(Even, DiagonalInvEven, DiagonalInv);
-  pickCheckerboard(Even, TriangleInvEven, TriangleInv);
-  pickCheckerboard(Odd,  DiagonalInvOdd,  DiagonalInv);
-  pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
-
-  // Report timings
-  double t9 = usecond();
-
-  std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "convert =                    " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "exponentiation =             " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "boundaries =                 " << (t7 - t6) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "inversions =                 " << (t8 - t7) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
-}
-
-NAMESPACE_END(Grid);

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

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

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

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

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

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

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

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

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

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

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

@@ -2,13 +2,12 @@
 
     Grid physics library, www.github.com/paboyle/Grid
 
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
+    Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
 
-    Copyright (C) 2017 - 2022
+    Copyright (C) 2017
 
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
-    Author: Daniel Richtmann <daniel.richtmann@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
@@ -34,48 +33,9 @@
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl, class CloverHelpers>
-WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField&                         _Umu,
-                                               GridCartesian&                      Fgrid,
-                                               GridRedBlackCartesian&              Hgrid,
-                                               const RealD                         _mass,
-                                               const RealD                         _csw_r,
-                                               const RealD                         _csw_t,
-                                               const WilsonAnisotropyCoefficients& clover_anisotropy,
-                                               const ImplParams&                   impl_p)
-  : WilsonFermion<Impl>(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
-  , CloverTerm(&Fgrid)
-  , CloverTermInv(&Fgrid)
-  , CloverTermEven(&Hgrid)
-  , CloverTermOdd(&Hgrid)
-  , CloverTermInvEven(&Hgrid)
-  , CloverTermInvOdd(&Hgrid)
-  , CloverTermDagEven(&Hgrid)
-  , CloverTermDagOdd(&Hgrid)
-  , CloverTermInvDagEven(&Hgrid)
-  , CloverTermInvDagOdd(&Hgrid) {
-  assert(Nd == 4); // require 4 dimensions
-
-  if(clover_anisotropy.isAnisotropic) {
-    csw_r     = _csw_r * 0.5 / clover_anisotropy.xi_0;
-    diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
-  } else {
-    csw_r     = _csw_r * 0.5;
-    diag_mass = 4.0 + _mass;
-  }
-  csw_t = _csw_t * 0.5;
-
-  if(csw_r == 0)
-    std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
-  if(csw_t == 0)
-    std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
-
-  ImportGauge(_Umu);
-}
-
 // *NOT* EO
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
+void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -89,8 +49,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, Fermion
   out += temp;
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
+void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -104,16 +64,13 @@ void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, Ferm
   out += temp;
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
+void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
-  double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
-  double t1 = usecond();
   GridBase *grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
 
-  double t2 = usecond();
   // Compute the field strength terms mu>nu
   WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
   WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
@@ -122,20 +79,52 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
   WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
   WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
 
-  double t3 = usecond();
   // Compute the Clover Operator acting on Colour and Spin
   // multiply here by the clover coefficients for the anisotropy
-  CloverTerm  = Helpers::fillCloverYZ(Bx) * csw_r;
+  CloverTerm  = fillCloverYZ(Bx) * csw_r;
-  CloverTerm += Helpers::fillCloverXZ(By) * csw_r;
+  CloverTerm += fillCloverXZ(By) * csw_r;
-  CloverTerm += Helpers::fillCloverXY(Bz) * csw_r;
+  CloverTerm += fillCloverXY(Bz) * csw_r;
-  CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
+  CloverTerm += fillCloverXT(Ex) * csw_t;
-  CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
+  CloverTerm += fillCloverYT(Ey) * csw_t;
-  CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
+  CloverTerm += fillCloverZT(Ez) * csw_t;
-   
+  CloverTerm += diag_mass;
-  double t4 = usecond();
+
-  CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
+  int lvol = _Umu.Grid()->lSites();
+  int DimRep = Impl::Dimension;
+
+  {
+    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 SiteCloverType::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 t5 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@@ -148,47 +137,37 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
 
   pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
   pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
-  double t6 = usecond();
-
-  std::cout << GridLogDebug << "WilsonCloverFermion::ImportGauge timings:" << std::endl;
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "instantiation =              " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "pick cbs =                   " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "total =                      " << (t6 - t0) / 1e6 << std::endl;
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
+void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseNo);
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
+void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseNo);
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
+void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseYes);
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
+void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseYes);
 }
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
+void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
 {
   out.Checkerboard() = in.Checkerboard();
-  CloverField *Clover;
+  CloverFieldType *Clover;
   assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
 
   if (dag)
@@ -203,12 +182,12 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField
       {
         Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
       }
-      Helpers::multCloverField(out, *Clover, in);
+      out = *Clover * in;
     }
     else
     {
       Clover = (inv) ? &CloverTermInv : &CloverTerm;
-      Helpers::multCloverField(out, *Clover, in); // don't bother with adj, hermitian anyway
+      out = adj(*Clover) * in;
     }
   }
   else
@@ -226,109 +205,29 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField
         //  std::cout << "Calling clover term Even" << std::endl;
         Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
       }
-      Helpers::multCloverField(out, *Clover, in);
+      out = *Clover * in;
       //  std::cout << GridLogMessage << "*Clover.Checkerboard() "  << (*Clover).Checkerboard() << std::endl;
     }
     else
     {
       Clover = (inv) ? &CloverTermInv : &CloverTerm;
-      Helpers::multCloverField(out, *Clover, in);
+      out = *Clover * in;
     }
   }
+
 } // MooeeInternal
 
-// Derivative parts unpreconditioned pseudofermions
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
-{
-  conformable(X.Grid(), Y.Grid());
-  conformable(X.Grid(), force.Grid());
-  GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
-  GaugeField clover_force(force.Grid());
-  PropagatorField Lambda(force.Grid());
-
-  // Guido: Here we are hitting some performance issues:
-  // need to extract the components of the DoubledGaugeField
-  // for each call
-  // Possible solution
-  // Create a vector object to store them? (cons: wasting space)
-  std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
-
-  Impl::extractLinkField(U, this->Umu);
-
-  force = Zero();
-  // Derivative of the Wilson hopping term
-  this->DhopDeriv(force, X, Y, dag);
-
-  ///////////////////////////////////////////////////////////
-  // Clover term derivative
-  ///////////////////////////////////////////////////////////
-  Impl::outerProductImpl(Lambda, X, Y);
-  //std::cout << "Lambda:" << Lambda << std::endl;
-
-  Gamma::Algebra sigma[] = {
-      Gamma::Algebra::SigmaXY,
-      Gamma::Algebra::SigmaXZ,
-      Gamma::Algebra::SigmaXT,
-      Gamma::Algebra::MinusSigmaXY,
-      Gamma::Algebra::SigmaYZ,
-      Gamma::Algebra::SigmaYT,
-      Gamma::Algebra::MinusSigmaXZ,
-      Gamma::Algebra::MinusSigmaYZ,
-      Gamma::Algebra::SigmaZT,
-      Gamma::Algebra::MinusSigmaXT,
-      Gamma::Algebra::MinusSigmaYT,
-      Gamma::Algebra::MinusSigmaZT};
-
-  /*
-    sigma_{\mu \nu}=
-    | 0         sigma[0]  sigma[1]  sigma[2] |
-    | sigma[3]    0       sigma[4]  sigma[5] |
-    | sigma[6]  sigma[7]     0      sigma[8] |
-    | sigma[9]  sigma[10] sigma[11]   0      |
-  */
-
-  int count = 0;
-  clover_force = Zero();
-  for (int mu = 0; mu < 4; mu++)
-  {
-    force_mu = Zero();
-    for (int nu = 0; nu < 4; nu++)
-    {
-      if (mu == nu)
-      continue;
-
-      RealD factor;
-      if (nu == 4 || mu == 4)
-      {
-        factor = 2.0 * csw_t;
-      }
-      else
-      {
-        factor = 2.0 * csw_r;
-      }
-      PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
-      Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);                   // checked
-      count++;
-    }
-
-    pokeLorentz(clover_force, U[mu] * force_mu, mu);
-  }
-  //clover_force *= csw;
-  force += clover_force;
-}
 
 // Derivative parts
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
+void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
 {
   assert(0);
 }
 
 // Derivative parts
-template<class Impl, class CloverHelpers>
+template <class Impl>
-void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
+void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   assert(0); // not implemented yet
 }

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

@@ -60,8 +60,7 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   UmuOdd (_FourDimRedBlackGrid),
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid),
-  _tmp(&FiveDimRedBlackGrid),
+  _tmp(&FiveDimRedBlackGrid)
-  Dirichlet(0)
 {
   // some assertions
   assert(FiveDimGrid._ndimension==5);
@@ -92,17 +91,6 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
     assert(FourDimRedBlackGrid._simd_layout[d]  ==FourDimGrid._simd_layout[d]);
   }
 
-  if ( p.dirichlet.size() == Nd+1) {
-    Coordinate block = p.dirichlet;
-    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
-      Dirichlet = 1;
-      Block = block;
-    }
-  } else {
-    Coordinate block(Nd+1,0);
-    Block = block;
-  }
-
   if (Impl::LsVectorised) { 
 
     int nsimd = Simd::Nsimd();
@@ -141,20 +129,95 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
    //                       <<" " << StencilEven.surface_list.size()<<std::endl;
 
 }
+     
+template<class Impl>
+void WilsonFermion5D<Impl>::Report(void)
+{
+  RealD NP     = _FourDimGrid->_Nprocessors;
+  RealD NN     = _FourDimGrid->NodeCount();
+  RealD volume = Ls;  
+  Coordinate latt = _FourDimGrid->GlobalDimensions();
+  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Number of DhopEO Calls   : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D TotalTime   /Calls        : " << DhopTotalTime   / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D CommTime    /Calls        : " << DhopCommTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D FaceTime    /Calls        : " << DhopFaceTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime1/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime2/Calls        : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
+
+    // Average the compute time
+    _FourDimGrid->GlobalSum(DhopComputeTime);
+    DhopComputeTime/=NP;
+    RealD mflops = 1344*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+
+    RealD Fullmflops = 1344*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+    
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NP << std::endl;
+
+    RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NP << std::endl;  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion5D Stencil"    <<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion5D StencilOdd" <<std::endl;  StencilOdd.Report();
+  }
+  if ( DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion5D Stencil     Reporti()"    <<std::endl;  Stencil.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion5D StencilEven Reporti()"<<std::endl;  StencilEven.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion5D StencilOdd  Reporti()" <<std::endl;  StencilOdd.Reporti(DhopCalls);
+  }
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopComputeTime2= 0;
+  DhopFaceTime    = 0;
+  DhopTotalTime   = 0;
+
+  DerivCalls       = 0;
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+  Stencil.ZeroCountersi();
+  StencilEven.ZeroCountersi();
+  StencilOdd.ZeroCountersi();
+}
+
 
 template<class Impl>
 void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
-  if ( Dirichlet ) {
-    std::cout << GridLogDslash << " Dirichlet BCs 5d " <<Block<<std::endl;
-    Coordinate GaugeBlock(Nd);
-    for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
-    std::cout << GridLogDslash << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
-    DirichletFilter<GaugeField> Filter(GaugeBlock);
-    Filter.applyFilter(HUmu);
-  }
   Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
   pickCheckerboard(Even,UmuEven,Umu);
   pickCheckerboard(Odd ,UmuOdd,Umu);
@@ -196,6 +259,7 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
 					  const FermionField &B,
 					  int dag)
 {
+  DerivCalls++;
   assert((dag==DaggerNo) ||(dag==DaggerYes));
 
   conformable(st.Grid(),A.Grid());
@@ -206,12 +270,15 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
   FermionField Btilde(B.Grid());
   FermionField Atilde(B.Grid());
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B,compressor);
+  DerivCommTime+=usecond();
 
   Atilde=A;
   int LLs = B.Grid()->_rdimensions[0];
 
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma if dag
@@ -223,6 +290,8 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
     // Call the single hop
     ////////////////////////
 
+    DerivDhopComputeTime -= usecond();
+
     int Usites = U.Grid()->oSites();
 
     Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, Usites, B, Btilde, mu,gamma);
@@ -230,8 +299,10 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
     ////////////////////////////
     // spin trace outer product
     ////////////////////////////
+    DerivDhopComputeTime += usecond();
     Impl::InsertForce5D(mat, Btilde, Atilde, mu);
   }
+  DerivComputeTime += usecond();
 }
 
 template<class Impl>
@@ -289,10 +360,12 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
                                          DoubledGaugeField & U,
                                          const FermionField &in, FermionField &out,int dag)
 {
+  DhopTotalTime-=usecond();
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else 
     DhopInternalSerialComms(st,lo,U,in,out,dag);
+  DhopTotalTime+=usecond();
 }
 
 
@@ -301,7 +374,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
 							DoubledGaugeField & U,
 							const FermionField &in, FermionField &out,int dag)
 {
-  GRID_TRACE("DhopInternalOverlappedComms");
   Compressor compressor(dag);
 
   int LLs = in.Grid()->_rdimensions[0];
@@ -310,58 +382,53 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   /////////////////////////////
   // Start comms  // Gather intranode and extra node differentiated??
   /////////////////////////////
-  {
+  DhopFaceTime-=usecond();
-    GRID_TRACE("Gather");
+  st.HaloExchangeOptGather(in,compressor);
-    st.HaloExchangeOptGather(in,compressor);
+  DhopFaceTime+=usecond();
-    accelerator_barrier();
+
-  }
+  DhopCommTime -=usecond();
-  
   std::vector<std::vector<CommsRequest_t> > requests;
-  auto id=traceStart("Communicate overlapped");
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
-  {
+  DhopFaceTime-=usecond();
-    GRID_TRACE("MergeSHM");
+  st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
-    st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+  DhopFaceTime+=usecond();
-  }
       
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
+  DhopComputeTime-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagInterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   } else {
-    GRID_TRACE("DhopInterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   }
+  DhopComputeTime+=usecond();
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
-  traceStop(id);
+  DhopCommTime   +=usecond();
 
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
-  {
+  DhopFaceTime-=usecond();
-    GRID_TRACE("Merge");
+  st.CommsMerge(compressor);
-    st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
-  }
-  
 
+  DhopComputeTime2-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagExterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   } else {
-    GRID_TRACE("DhopExterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   }
+  DhopComputeTime2+=usecond();
 }
 
 
@@ -371,30 +438,29 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOr
 						    const FermionField &in, 
 						    FermionField &out,int dag)
 {
-  GRID_TRACE("DhopInternalSerialComms");
   Compressor compressor(dag);
 
   int LLs = in.Grid()->_rdimensions[0];
-
-  {
-    GRID_TRACE("HaloExchange");
-    st.HaloExchangeOpt(in,compressor);
-  }
   
+  DhopCommTime-=usecond();
+  st.HaloExchangeOpt(in,compressor);
+  DhopCommTime+=usecond();
+  
+  DhopComputeTime-=usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDag");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   } else {
-    GRID_TRACE("Dhop");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   }
+  DhopComputeTime+=usecond();
 }
 
 
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -406,6 +472,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -417,6 +484,7 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
 template<class Impl>
 void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(),FermionGrid()); // verifies full grid
   conformable(in.Grid(),out.Grid());
 
@@ -471,17 +539,12 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
   LatComplex    sk(_grid);  sk = Zero();
   LatComplex    sk2(_grid); sk2= Zero();
   LatComplex    W(_grid); W= Zero();
+  LatComplex    a(_grid); a= Zero();
   LatComplex    one  (_grid); one = ScalComplex(1.0,0.0);
   LatComplex 	cosha(_grid);
   LatComplex 	kmu(_grid);
   LatComplex 	Wea(_grid);
   LatComplex 	Wema(_grid);
-  LatComplex 	ea(_grid);
-  LatComplex 	ema(_grid);
-  LatComplex 	eaLs(_grid);
-  LatComplex 	emaLs(_grid);
-  LatComplex 	ea2Ls(_grid);
-  LatComplex 	ema2Ls(_grid);
   LatComplex 	sinha(_grid);
   LatComplex 	sinhaLs(_grid);
   LatComplex 	coshaLs(_grid);
@@ -516,29 +579,39 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
   ////////////////////////////////////////////
   cosha = (one + W*W + sk) / (abs(W)*2.0);
 
-  ea = (cosha + sqrt(cosha*cosha-one));
+  // FIXME Need a Lattice acosh
-  ema= (cosha - sqrt(cosha*cosha-one));
+
-  eaLs = pow(ea,Ls);
+  {
-  emaLs= pow(ema,Ls);
+    autoView(cosha_v,cosha,CpuRead);
-  ea2Ls = pow(ea,2.0*Ls);
+    autoView(a_v,a,CpuWrite);
-  ema2Ls= pow(ema,2.0*Ls);
+    for(int idx=0;idx<_grid->lSites();idx++){
-  Wea= abs(W) * ea;
+      Coordinate lcoor(Nd);
-  Wema= abs(W) * ema;
+      Tcomplex cc;
-  //  a=log(ea);
+      //    RealD sgn;
-  
+      _grid->LocalIndexToLocalCoor(idx,lcoor);
-  sinha = 0.5*(ea - ema);
+      peekLocalSite(cc,cosha_v,lcoor);
-  sinhaLs = 0.5*(eaLs-emaLs);
+      assert((double)real(cc)>=1.0);
-  coshaLs = 0.5*(eaLs+emaLs);
+      assert(fabs((double)imag(cc))<=1.0e-15);
+      cc = ScalComplex(::acosh(real(cc)),0.0);
+      pokeLocalSite(cc,a_v,lcoor);
+    }
+  }
+
+  Wea = ( exp( a) * abs(W)  );
+  Wema= ( exp(-a) * abs(W)  );
+  sinha = 0.5*(exp( a) - exp(-a));
+  sinhaLs = 0.5*(exp( a*Ls) - exp(-a*Ls));
+  coshaLs = 0.5*(exp( a*Ls) + exp(-a*Ls));
 
   A = one / (abs(W) * sinha * 2.0) * one / (sinhaLs * 2.0);
-  F = eaLs * (one - Wea + (Wema - one) * mass*mass);
+  F = exp( a*Ls) * (one - Wea + (Wema - one) * mass*mass);
-  F = F + emaLs * (Wema - one + (one - Wea) * mass*mass);
+  F = F + exp(-a*Ls) * (Wema - one + (one - Wea) * mass*mass);
   F = F - abs(W) * sinha * 4.0 * mass;
 
-  Bpp =  (A/F) * (ema2Ls - one) * (one - Wema) * (one - mass*mass * one);
+  Bpp =  (A/F) * (exp(-a*Ls*2.0) - one) * (one - Wema) * (one - mass*mass * one);
-  Bmm =  (A/F) * (one - ea2Ls)  * (one - Wea) * (one - mass*mass * one);
+  Bmm =  (A/F) * (one - exp(a*Ls*2.0)) * (one - Wea) * (one - mass*mass * one);
-  App =  (A/F) * (ema2Ls - one) * ema * (ema - abs(W)) * (one - mass*mass * one);
+  App =  (A/F) * (exp(-a*Ls*2.0) - one) * exp(-a) * (exp(-a) - abs(W)) * (one - mass*mass * one);
-  Amm =  (A/F) * (one - ea2Ls)  * ea  * (ea  - abs(W)) * (one - mass*mass * one);
+  Amm =  (A/F) * (one - exp(a*Ls*2.0)) * exp(a) * (exp(a) - abs(W)) * (one - mass*mass * one);
   ABpm = (A/F) * abs(W) * sinha * 2.0  * (one + mass * coshaLs * 2.0 + mass*mass * one);
 
   //P+ source, P- source
@@ -561,29 +634,29 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
       buf1_4d = Zero();
       ExtractSlice(buf1_4d, PRsource, (tt-1), 0);
       //G(s,t)
-      bufR_4d = bufR_4d + A * eaLs * pow(ema,f) * signW * buf1_4d + A * emaLs * pow(ea,f) * signW * buf1_4d;
+      bufR_4d = bufR_4d + A * exp(a*Ls) * exp(-a*f) * signW * buf1_4d + A * exp(-a*Ls) * exp(a*f) * signW * buf1_4d;
       //A++*exp(a(s+t))
-      bufR_4d = bufR_4d + App * pow(ea,ss) * pow(ea,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + App * exp(a*ss) * exp(a*tt) * signW * buf1_4d ;
       //A+-*exp(a(s-t))
-      bufR_4d = bufR_4d + ABpm * pow(ea,ss) * pow(ema,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + ABpm * exp(a*ss) * exp(-a*tt) * signW * buf1_4d ;
       //A-+*exp(a(-s+t))
-      bufR_4d = bufR_4d + ABpm * pow(ema,ss) * pow(ea,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + ABpm * exp(-a*ss) * exp(a*tt) * signW * buf1_4d ;
       //A--*exp(a(-s-t))
-      bufR_4d = bufR_4d + Amm * pow(ema,ss) * pow(ema,tt) * signW * buf1_4d ;
+      bufR_4d = bufR_4d + Amm * exp(-a*ss) * exp(-a*tt) * signW * buf1_4d ;
 
       //GL
       buf2_4d = Zero();
       ExtractSlice(buf2_4d, PLsource, (tt-1), 0);
       //G(s,t)
-      bufL_4d = bufL_4d + A * eaLs * pow(ema,f) * signW * buf2_4d + A * emaLs * pow(ea,f) * signW * buf2_4d;
+      bufL_4d = bufL_4d + A * exp(a*Ls) * exp(-a*f) * signW * buf2_4d + A * exp(-a*Ls) * exp(a*f) * signW * buf2_4d;
       //B++*exp(a(s+t))
-      bufL_4d = bufL_4d + Bpp * pow(ea,ss) * pow(ea,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + Bpp * exp(a*ss) * exp(a*tt) * signW * buf2_4d ;
       //B+-*exp(a(s-t))
-      bufL_4d = bufL_4d + ABpm * pow(ea,ss) * pow(ema,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + ABpm * exp(a*ss) * exp(-a*tt) * signW * buf2_4d ;
       //B-+*exp(a(-s+t))
-      bufL_4d = bufL_4d + ABpm * pow(ema,ss) * pow(ea,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + ABpm * exp(-a*ss) * exp(a*tt) * signW * buf2_4d ;
       //B--*exp(a(-s-t))
-      bufL_4d = bufL_4d + Bmm * pow(ema,ss) * pow(ema,tt) * signW * buf2_4d ;
+      bufL_4d = bufL_4d + Bmm * exp(-a*ss) * exp(-a*tt) * signW * buf2_4d ;
     }
     InsertSlice(bufR_4d, GR, (ss-1), 0);
     InsertSlice(bufL_4d, GL, (ss-1), 0);
@@ -702,12 +775,28 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
   W = one - M5 + sk2;
 
   ////////////////////////////////////////////
-  // Cosh alpha -> exp(+/- alpha)
+  // Cosh alpha -> alpha
   ////////////////////////////////////////////
   cosha =  (one + W*W + sk) / (abs(W)*2.0);
 
-  Wea = abs(W)*(cosha + sqrt(cosha*cosha-one));
+  // FIXME Need a Lattice acosh
-  Wema= abs(W)*(cosha - sqrt(cosha*cosha-one));
+  {
+  autoView(cosha_v,cosha,CpuRead);
+  autoView(a_v,a,CpuWrite);
+  for(int idx=0;idx<_grid->lSites();idx++){
+    Coordinate lcoor(Nd);
+    Tcomplex cc;
+    //    RealD sgn;
+    _grid->LocalIndexToLocalCoor(idx,lcoor);
+    peekLocalSite(cc,cosha_v,lcoor);
+    assert((double)real(cc)>=1.0);
+    assert(fabs((double)imag(cc))<=1.0e-15);
+    cc = ScalComplex(::acosh(real(cc)),0.0);
+    pokeLocalSite(cc,a_v,lcoor);
+  }}
+  
+  Wea = ( exp( a) * abs(W)  );
+  Wema= ( exp(-a) * abs(W)  );
   
   num   = num + ( one - Wema ) * mass * in;
   denom= ( Wea - one ) + mass*mass * (one - Wema); 

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

@@ -4,13 +4,12 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
 
-Copyright (C) 2022
+Copyright (C) 2015
 
 Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Fabian Joswig <fabian.joswig@ed.ac.uk>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -76,6 +75,91 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
   StencilOdd.BuildSurfaceList(1,vol4);
 }
 
+template<class Impl>
+void WilsonFermion<Impl>::Report(void)
+{
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+  RealD volume = 1;
+  Coordinate latt = _grid->GlobalDimensions();
+  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of DhopEO Calls   : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion TotalTime   /Calls        : " << DhopTotalTime   / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime    /Calls        : " << DhopCommTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion FaceTime    /Calls        : " << DhopFaceTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime1/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime2/Calls        : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
+
+    // Average the compute time
+    _grid->GlobalSum(DhopComputeTime);
+    DhopComputeTime/=NP;
+    RealD mflops = 1320*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+
+    RealD Fullmflops = 1320*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+
+    // how to count flops here?
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call               ? : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node      ? : " << mflops/NP << std::endl;
+
+    // how to count flops here?
+    RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)        ? : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full) ? : " << Fullmflops/NP << std::endl;  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil"    <<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+  }
+  if ( DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil     Reporti()"    <<std::endl;  Stencil.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilEven Reporti()"<<std::endl;  StencilEven.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd  Reporti()" <<std::endl;  StencilOdd.Reporti(DhopCalls);
+  }
+}
+
+template<class Impl>
+void WilsonFermion<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0; // ok
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopComputeTime2= 0;
+  DhopFaceTime    = 0;
+  DhopTotalTime   = 0;
+
+  DerivCalls       = 0; // ok
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+  Stencil.ZeroCountersi();
+  StencilEven.ZeroCountersi();
+  StencilOdd.ZeroCountersi();
+}
+
+
 template <class Impl>
 void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
@@ -235,6 +319,7 @@ template <class Impl>
 void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
                                         GaugeField &mat, const FermionField &A,
                                         const FermionField &B, int dag) {
+  DerivCalls++;
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -243,8 +328,11 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
   FermionField Atilde(B.Grid());
   Atilde = A;
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B, compressor);
+  DerivCommTime+=usecond();
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma (1+g)<->(1-g) if dag
@@ -252,6 +340,7 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     int gamma = mu;
     if (!dag) gamma += Nd;
 
+    DerivDhopComputeTime -= usecond();
     int Ls=1;
     Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, B.Grid()->oSites(), B, Btilde, mu, gamma);
 
@@ -259,7 +348,9 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     // spin trace outer product
     //////////////////////////////////////////////////
     Impl::InsertForce4D(mat, Btilde, Atilde, mu);
+    DerivDhopComputeTime += usecond();
   }
+  DerivComputeTime += usecond();
 }
 
 template <class Impl>
@@ -306,6 +397,7 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
 template <class Impl>
 void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -317,6 +409,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 template <class Impl>
 void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -329,6 +422,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
 template <class Impl>
 void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -393,12 +487,14 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
+  DhopTotalTime-=usecond();
 #ifdef GRID_OMP
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
 #endif
     DhopInternalSerial(st,lo,U,in,out,dag);
+  DhopTotalTime+=usecond();
 }
 
 template <class Impl>
@@ -407,7 +503,6 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
 						      const FermionField &in,
 						      FermionField &out, int dag)
 {
-  GRID_TRACE("DhopOverlapped");
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -418,55 +513,53 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
   /////////////////////////////
   std::vector<std::vector<CommsRequest_t> > requests;
   st.Prepare();
-  {
+  DhopFaceTime-=usecond();
-    GRID_TRACE("Gather");
+  st.HaloGather(in,compressor);
-    st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
-  }
 
-  tracePush("Communication");
+  DhopCommTime -=usecond();
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
-  {
+  DhopFaceTime-=usecond();
-    GRID_TRACE("MergeSHM");
+  st.CommsMergeSHM(compressor);
-    st.CommsMergeSHM(compressor);
+  DhopFaceTime+=usecond();
-  }
 
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt;
+  DhopComputeTime-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagInterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   } else {
-    GRID_TRACE("DhopInterior");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   }
+  DhopComputeTime+=usecond();
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
-  tracePop("Communication");
+  DhopCommTime   +=usecond();
+
+  DhopFaceTime-=usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
-  {
-    GRID_TRACE("Merge");
-    st.CommsMerge(compressor);
-  }
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
 
+  DhopComputeTime2-=usecond();
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDagExterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   } else {
-    GRID_TRACE("DhopExterior");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   }
+  DhopComputeTime2+=usecond();
 };
 
 
@@ -476,22 +569,20 @@ void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
-  GRID_TRACE("DhopSerial");
   assert((dag == DaggerNo) || (dag == DaggerYes));
   Compressor compressor(dag);
-  {
+  DhopCommTime-=usecond();
-    GRID_TRACE("HaloExchange");
+  st.HaloExchange(in, compressor);
-    st.HaloExchange(in, compressor);
+  DhopCommTime+=usecond();
-  }
 
+  DhopComputeTime-=usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
-    GRID_TRACE("DhopDag");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   } else {
-    GRID_TRACE("Dhop");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   }
+  DhopComputeTime+=usecond();
 };
 /*Change ends */
 
@@ -508,47 +599,11 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
                                                    Current curr_type,
                                                    unsigned int mu)
 {
-  if(curr_type != Current::Vector)
-  {
-    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
-    exit(1);
-  }
-
   Gamma g5(Gamma::Algebra::Gamma5);
   conformable(_grid, q_in_1.Grid());
   conformable(_grid, q_in_2.Grid());
   conformable(_grid, q_out.Grid());
-  auto UGrid= this->GaugeGrid();
+  assert(0);
-
-  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;
 }
 
 
@@ -562,51 +617,9 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
                                               unsigned int tmax,
 					      ComplexField &lattice_cmplx)
 {
-  if(curr_type != Current::Vector)
-  {
-    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
-    exit(1);
-  }
-
-  int tshift = (mu == Nd-1) ? 1 : 0;
-  unsigned int LLt    = GridDefaultLatt()[Tp];
   conformable(_grid, q_in.Grid());
   conformable(_grid, q_out.Grid());
-  auto UGrid= this->GaugeGrid();
+  assert(0);
-
-  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

@@ -72,15 +72,20 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   if (SE->_is_local) {						\
     int perm= SE->_permute;					\
     auto tmp = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);	\
-    spProj(chi,tmp);							\
+    spProj(chi,tmp);						\
-    Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);			\
+  } else if ( st.same_node[Dir] ) {				\
-    Recon(result, Uchi);						\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
-  }									\
+  }								\
+  acceleratorSynchronise();						\
+  if (SE->_is_local || st.same_node[Dir] ) {			\
+    Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
+    Recon(result, Uchi);					\
+  }								\
   acceleratorSynchronise();
 
 #define GENERIC_STENCIL_LEG_EXT(Dir,spProj,Recon)		\
   SE = st.GetEntry(ptype, Dir, sF);				\
-  if (!SE->_is_local ) {		\
+  if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
     auto chi = coalescedRead(buf[SE->_offset],lane);		\
     Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
     Recon(result, Uchi);					\
@@ -411,6 +416,19 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
+#define KERNEL_CALL_TMP(A) \
+  const uint64_t    NN = Nsite*Ls;					\
+  auto U_p = & U_v[0];							\
+  auto in_p = & in_v[0];						\
+  auto out_p = & out_v[0];						\
+  auto st_p = st_v._entries_p;						\
+  auto st_perm = st_v._permute_type;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_perm,st_p,U_p,buf,sF,sU,in_p,out_p);	\
+    });									\
+  accelerator_barrier();
 
 #define KERNEL_CALLNB(A)						\
   const uint64_t    NN = Nsite*Ls;					\
@@ -422,16 +440,6 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
-#define KERNEL_CALL_EXT(A)						\
-  const uint64_t    NN = Nsite*Ls;					\
-  const uint64_t    sz = st.surface_list.size();			\
-  auto ptr = &st.surface_list[0];					\
-  accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
-      int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
-      WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
-
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
     int sU = ss;							\
@@ -452,7 +460,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
 #ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteSycl);    return; }
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
 #else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
 #endif     

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

@@ -1,44 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/ qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master
-
-    Copyright (C) 2017 - 2022
-
-    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
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-/*  END LEGAL */
-
-#include <Grid/Grid.h>
-#include <Grid/qcd/spin/Dirac.h>
-#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
-#include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>; 
-template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>; 
-
-NAMESPACE_END(Grid);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -18,10 +18,6 @@ WILSON_IMPL_LIST=" \
 	   GparityWilsonImplF \
 	   GparityWilsonImplD "
 
-COMPACT_WILSON_IMPL_LIST=" \
-	   WilsonImplF \
-	   WilsonImplD "
-
 DWF_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
@@ -50,17 +46,7 @@ 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
 
@@ -77,14 +63,14 @@ for impl in $DWF_IMPL_LIST $GDWF_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 
 # overwrite the .cc file in Gparity directories
 for impl in $GDWF_IMPL_LIST
 do
-  ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
+  ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc 
 done
 
 
@@ -98,7 +84,7 @@ for impl in $STAG_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done