Interface to query memory use

Huge Cache

Grid/DisableWarnings.h

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

Grid/GridStd.h

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

Grid/Grid_Eigen_Dense.h

@@ -14,7 +14,11 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+#pragma nv_diag_suppress code_is_unreachable
+#else
 #pragma diag_suppress code_is_unreachable
+#endif
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")

Grid/algorithms/Algorithms.h

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

Grid/algorithms/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
         });
@@ -358,7 +358,7 @@ public:
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
     autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
     Vector<Aview> AcceleratorViewContainer;
@@ -380,7 +380,7 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom_v.npoint;point++){
+      for(int point=0;point<npoint;point++){
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
@@ -424,7 +424,7 @@ public:
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
     autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
     Vector<Aview> AcceleratorViewContainer;
@@ -454,7 +454,7 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int p=0;p<geom_v.npoint;p++){
+      for(int p=0;p<npoint;p++){
         int point = points_p[p];
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);

Grid/algorithms/LinearOperator.h

@@ -52,6 +52,7 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
+  virtual ~LinearOperatorBase(){};
 };
 
 
@@ -507,7 +508,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   virtual  void MpcDag   (const Field &in, Field &out){
     Mpc(in,out);
   }
-  virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+  virtual void MpcDagMpc(const Field &in, Field &out) {
     assert(0);// Never need with staggered
   }
 };
@@ -585,6 +586,7 @@ class HermOpOperatorFunction : public OperatorFunction<Field> {
 template<typename Field>
 class PlainHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   LinearOperatorBase<Field> &_Linop;
       
   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
@@ -598,6 +600,7 @@ public:
 template<typename Field>
 class FunctionHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator(); 
   OperatorFunction<Field>   & _poly;
   LinearOperatorBase<Field> &_Linop;
       

Grid/algorithms/Preconditioner.h

@@ -30,13 +30,19 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Field> class Preconditioner :  public LinearFunction<Field> { 
+template<class Field> using Preconditioner =  LinearFunction<Field> ;
+
+/*
+template<class Field> class Preconditioner :  public LinearFunction<Field> {
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field & psi)=0;
 };
+*/
 
 template<class Field> class TrivialPrecon :  public Preconditioner<Field> { 
 public:
-  void operator()(const Field &src, Field & psi){
+  using Preconditioner<Field>::operator();
+  virtual void operator()(const Field &src, Field & psi){
     psi = src;
   }
   TrivialPrecon(void){};

Grid/algorithms/SparseMatrix.h

@@ -48,6 +48,7 @@ public:
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;
+  virtual ~SparseMatrixBase() {};
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -72,7 +73,7 @@ public:
   virtual  void MeooeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeInvDag (const Field &in, Field &out)=0;
-
+  virtual ~CheckerBoardedSparseMatrixBase() {};
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/approx/Chebyshev.h

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

Grid/algorithms/iterative/BiCGSTABMixedPrec.h

@@ -36,7 +36,8 @@ NAMESPACE_BEGIN(Grid);
 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 MixedPrecisionBiCGSTAB : public LinearFunction<FieldD> 
 {
-  public:                                                
+  public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -35,7 +35,8 @@ NAMESPACE_BEGIN(Grid);
     typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
-  public:                                                
+  public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -0,0 +1,213 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
+
+    Copyright (C) 2015
+
+    Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+
+NAMESPACE_BEGIN(Grid);
+
+//Mixed precision restarted defect correction CG
+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 MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
+public:
+  using LinearFunction<FieldD>::operator();
+  RealD   Tolerance;
+  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
+  Integer MaxInnerIterations;
+  Integer MaxOuterIterations;
+  Integer MaxPatchupIterations;
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+  LinearOperatorBase<FieldF> &Linop_f;
+  LinearOperatorBase<FieldD> &Linop_d;
+
+  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+  LinearFunction<FieldF> *guesser;
+  bool updateResidual;
+  
+  MixedPrecisionConjugateGradientBatched(RealD tol, 
+          Integer maxinnerit, 
+          Integer maxouterit, 
+          Integer maxpatchit,
+          GridBase* _sp_grid, 
+          LinearOperatorBase<FieldF> &_Linop_f, 
+          LinearOperatorBase<FieldD> &_Linop_d,
+          bool _updateResidual=true) :
+    Linop_f(_Linop_f), Linop_d(_Linop_d),
+    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
+    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
+
+  void useGuesser(LinearFunction<FieldF> &g){
+    guesser = &g;
+  }
+  
+  void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::vector<FieldD> srcs_d_in{src_d_in};
+    std::vector<FieldD> sols_d{sol_d};
+
+    (*this)(srcs_d_in,sols_d);
+
+    sol_d = sols_d[0];
+  }
+
+  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
+    assert(src_d_in.size() == sol_d.size());
+    int NBatch = src_d_in.size();
+
+    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
+
+    Integer TotalOuterIterations = 0; //Number of restarts
+    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
+    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
+  
+    GridStopWatch TotalTimer;
+    TotalTimer.Start();
+
+    GridStopWatch InnerCGtimer;
+    GridStopWatch PrecChangeTimer;
+    
+    int cb = src_d_in[0].Checkerboard();
+    
+    std::vector<RealD> src_norm;
+    std::vector<RealD> norm;
+    std::vector<RealD> stop;
+    
+    GridBase* DoublePrecGrid = src_d_in[0].Grid();
+    FieldD tmp_d(DoublePrecGrid);
+    tmp_d.Checkerboard() = cb;
+    
+    FieldD tmp2_d(DoublePrecGrid);
+    tmp2_d.Checkerboard() = cb;
+
+    std::vector<FieldD> src_d;
+    std::vector<FieldF> src_f;
+    std::vector<FieldF> sol_f;
+
+    for (int i=0; i<NBatch; i++) {
+      sol_d[i].Checkerboard() = cb;
+
+      src_norm.push_back(norm2(src_d_in[i]));
+      norm.push_back(0.);
+      stop.push_back(src_norm[i] * Tolerance*Tolerance);
+
+      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
+
+      src_f.push_back(SinglePrecGrid);
+      src_f[i].Checkerboard() = cb;
+
+      sol_f.push_back(SinglePrecGrid);
+      sol_f[i].Checkerboard() = cb;
+    }
+    
+    RealD inner_tol = InnerTolerance;
+    
+    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+    CG_f.ErrorOnNoConverge = false;
+    
+    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      std::cout << GridLogMessage << std::endl;
+      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      
+      bool allConverged = true;
+      
+      for (int i=0; i<NBatch; i++) {
+        //Compute double precision rsd and also new RHS vector.
+        Linop_d.HermOp(sol_d[i], tmp_d);
+        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        
+        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
+
+        PrecChangeTimer.Start();
+        precisionChange(src_f[i], src_d[i]);
+        PrecChangeTimer.Stop();
+        
+        sol_f[i] = Zero();
+      
+        if(norm[i] > OuterLoopNormMult * stop[i]) {
+          allConverged = false;
+        }
+      }
+      if (allConverged) break;
+
+      if (updateResidual) {
+        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
+        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
+        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+        CG_f.Tolerance = inner_tol;
+      }
+
+      //Optionally improve inner solver guess (eg using known eigenvectors)
+      if(guesser != NULL) {
+        (*guesser)(src_f, sol_f);
+      }
+
+      for (int i=0; i<NBatch; i++) {
+        //Inner CG
+        InnerCGtimer.Start();
+        CG_f(Linop_f, src_f[i], sol_f[i]);
+        InnerCGtimer.Stop();
+        TotalInnerIterations[i] += CG_f.IterationsToComplete;
+        
+        //Convert sol back to double and add to double prec solution
+        PrecChangeTimer.Start();
+        precisionChange(tmp_d, sol_f[i]);
+        PrecChangeTimer.Stop();
+        
+        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
+      }
+
+    }
+    
+    //Final trial CG
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
+    
+    for (int i=0; i<NBatch; i++) {
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
+      CG_d(Linop_d, src_d_in[i], sol_d[i]);
+      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
+    }
+
+    TotalTimer.Stop();
+
+    std::cout << GridLogMessage << std::endl;
+    for (int i=0; i<NBatch; i++) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
+    }
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/algorithms/iterative/Deflation.h

@@ -33,16 +33,19 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
 };
 template<class Field>
 class DoNothingGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) {  };
 };
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) { guess = src; };
 };
 
@@ -57,6 +60,7 @@ private:
   const unsigned int       N;
 
 public:
+  using LinearFunction<Field>::operator();
 
   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
   : DeflatedGuesser(_evec, _eval, _evec.size())
@@ -87,6 +91,7 @@ private:
   const std::vector<RealD>       &eval_coarse;
 public:
   
+  using LinearFunction<FineField>::operator();
   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
 				const std::vector<CoarseField> &_evec_coarse,
 				const std::vector<RealD>       &_eval_coarse)
@@ -108,7 +113,43 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
-};
+
+  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
+    int Nevec = (int)evec_coarse.size();
+    int Nsrc = (int)src.size();
+    // make temp variables
+    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
+    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
+    //Preporcessing
+    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    guess_coarse[j] = Zero();
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockProject(src_coarse[j],src[j],subspace);
+    }
+    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
+    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
+    for (int i=0;i<Nevec;i++)
+    {
+      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
+      const CoarseField & tmp = evec_coarse[i];
+      for (int j=0;j<Nsrc;j++)
+      {
+        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
+      }
+    }
+    //postprocessing
+    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockPromote(guess_coarse[j],guess[j],subspace);
+    guess[j].Checkerboard() = src[j].Checkerboard();
+    }
+  };
+
+  };
 
 
 

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -67,6 +67,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -97,6 +98,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
 template<class Field>
 class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
 public:                                                
-
+  using LinearFunction<Field>::operator();
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
 template<class Field>
 class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
 public:                                                
-
+  using LinearFunction<Field>::operator();
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;
@@ -119,7 +119,8 @@ public:
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
 
     RealD cp;
-    ComplexD a, b, zAz;
+    ComplexD a, b;
+    //    ComplexD zAz;
     RealD zAAz;
     ComplexD rq;
 
@@ -146,7 +147,7 @@ public:
     //////////////////////////////////
     MatTimer.Start();
     Linop.Op(psi,Az);
-    zAz = innerProduct(Az,psi);
+    //    zAz = innerProduct(Az,psi);
     zAAz= norm2(Az);
     MatTimer.Stop();
     
@@ -170,7 +171,7 @@ public:
 
     LinalgTimer.Start();
 
-    zAz = innerProduct(Az,psi);
+    //    zAz = innerProduct(Az,psi);
     zAAz= norm2(Az);
 
     //p[0],q[0],qq[0] 
@@ -212,7 +213,7 @@ public:
       MatTimer.Start();
       Linop.Op(z,Az);
       MatTimer.Stop();
-      zAz = innerProduct(Az,psi);
+      //      zAz = innerProduct(Az,psi);
       zAAz= norm2(Az);
 
       LinalgTimer.Start();

Grid/allocator/MemoryManager.cc

@@ -4,106 +4,156 @@ NAMESPACE_BEGIN(Grid);
 
 /*Allocation types, saying which pointer cache should be used*/
 #define Cpu      (0)
-#define CpuSmall (1)
-#define Acc      (2)
-#define AccSmall (3)
-#define Shared   (4)
-#define SharedSmall (5)
+#define CpuHuge  (1)
+#define CpuSmall (2)
+#define Acc      (3)
+#define AccHuge  (4)
+#define AccSmall (5)
+#define Shared   (6)
+#define SharedHuge  (7)
+#define SharedSmall (8)
+#undef GRID_MM_VERBOSE 
 uint64_t total_shared;
 uint64_t total_device;
 uint64_t total_host;;
 void MemoryManager::PrintBytes(void)
 {
-  std::cout << " MemoryManager : "<<total_shared<<" shared      bytes "<<std::endl;
-  std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
-  std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<std::endl;
+  std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
+  std::cout << " MemoryManager : PrintBytes "<<std::endl;
+  std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
+  std::cout << " MemoryManager : "<<(total_shared>>20)<<" shared      Mbytes "<<std::endl;
+  std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
+  std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl;
+  uint64_t cacheBytes;
+  cacheBytes = CacheBytes[Cpu];
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
+  cacheBytes = CacheBytes[Acc];
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
+  cacheBytes = CacheBytes[Shared];
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
+  
+#ifdef GRID_CUDA
+  cuda_mem();
+#endif
+  
 }
 
+uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
+uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
+
 //////////////////////////////////////////////////////////////////////
 // Data tables for recently freed pooiniter caches
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 8, 32, 8, 32, 8, 32 };
-
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
+uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils
 //////////////////////////////////////////////////////////////////////
 void *MemoryManager::AcceleratorAllocate(size_t bytes)
 {
+  total_device+=bytes;
   void *ptr = (void *) Lookup(bytes,Acc);
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocDevice(bytes);
-    total_device+=bytes;
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"AcceleratorAllocate "<<std::endl;
+  PrintBytes();
+#endif
   return ptr;
 }
 void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes)
 {
+  total_device-=bytes;
   void *__freeme = Insert(ptr,bytes,Acc);
   if ( __freeme ) {
     acceleratorFreeDevice(__freeme);
-    total_device-=bytes;
-    //    PrintBytes();
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"AcceleratorFree "<<std::endl;
+  PrintBytes();
+#endif
 }
 void *MemoryManager::SharedAllocate(size_t bytes)
 {
+  total_shared+=bytes;
   void *ptr = (void *) Lookup(bytes,Shared);
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocShared(bytes);
-    total_shared+=bytes;
-    //    std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
-    //    PrintBytes();
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"SharedAllocate "<<std::endl;
+  PrintBytes();
+#endif
   return ptr;
 }
 void  MemoryManager::SharedFree    (void *ptr,size_t bytes)
 {
+  total_shared-=bytes;
   void *__freeme = Insert(ptr,bytes,Shared);
   if ( __freeme ) {
     acceleratorFreeShared(__freeme);
-    total_shared-=bytes;
-    //    PrintBytes();
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"SharedFree "<<std::endl;
+  PrintBytes();
+#endif
 }
 #ifdef GRID_UVM
 void *MemoryManager::CpuAllocate(size_t bytes)
 {
+  total_host+=bytes;
   void *ptr = (void *) Lookup(bytes,Cpu);
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocShared(bytes);
-    total_host+=bytes;
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuAllocate "<<std::endl;
+  PrintBytes();
+#endif
   return ptr;
 }
 void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
 {
+  total_host-=bytes;
   NotifyDeletion(_ptr);
   void *__freeme = Insert(_ptr,bytes,Cpu);
   if ( __freeme ) { 
     acceleratorFreeShared(__freeme);
-    total_host-=bytes;
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuFree "<<std::endl;
+  PrintBytes();
+#endif
 }
 #else
 void *MemoryManager::CpuAllocate(size_t bytes)
 {
+  total_host+=bytes;
   void *ptr = (void *) Lookup(bytes,Cpu);
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocCpu(bytes);
-    total_host+=bytes;
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuAllocate "<<std::endl;
+  PrintBytes();
+#endif
   return ptr;
 }
 void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
 {
+  total_host-=bytes;
   NotifyDeletion(_ptr);
   void *__freeme = Insert(_ptr,bytes,Cpu);
   if ( __freeme ) { 
     acceleratorFreeCpu(__freeme);
-    total_host-=bytes;
   }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuFree "<<std::endl;
+  PrintBytes();
+#endif
 }
 #endif
 
@@ -115,7 +165,6 @@ void MemoryManager::Init(void)
 
   char * str;
   int Nc;
-  int NcS;
   
   str= getenv("GRID_ALLOC_NCACHE_LARGE");
   if ( str ) {
@@ -127,6 +176,16 @@ void MemoryManager::Init(void)
     }
   }
 
+  str= getenv("GRID_ALLOC_NCACHE_HUGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuHuge]=Nc;
+      Ncache[AccHuge]=Nc;
+      Ncache[SharedHuge]=Nc;
+    }
+  }
+
   str= getenv("GRID_ALLOC_NCACHE_SMALL");
   if ( str ) {
     Nc = atoi(str);
@@ -147,7 +206,9 @@ void MemoryManager::InitMessage(void) {
   
   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
 #ifdef ALLOCATION_CACHE
-  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
 #endif
   
 #ifdef GRID_UVM
@@ -179,21 +240,25 @@ void MemoryManager::InitMessage(void) {
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  int cache = type + small;
-  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);  
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
+  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
 #endif
 }
 
-void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) 
+void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
 
+  if (ncache == 0) return ptr;
+
   void * ret = NULL;
   int v = -1;
 
@@ -211,6 +276,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 
   if ( entries[v].valid ) {
     ret = entries[v].address;
+    cacheBytes -= entries[v].bytes;
     entries[v].valid = 0;
     entries[v].address = NULL;
     entries[v].bytes = 0;
@@ -219,6 +285,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
   entries[v].address=ptr;
   entries[v].bytes  =bytes;
   entries[v].valid  =1;
+  cacheBytes += bytes;
 
   return ret;
 }
@@ -226,23 +293,26 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 void *MemoryManager::Lookup(size_t bytes,int type)
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  int cache = type+small;
-  return Lookup(bytes,Entries[cache],Ncache[cache]);
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
+  return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
 #endif
 }
 
-void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) 
+void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
   for(int e=0;e<ncache;e++){
     if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
       entries[e].valid = 0;
+      cacheBytes -= entries[e].bytes;
       return entries[e].address;
     }
   }

Grid/allocator/MemoryManager.h

@@ -35,6 +35,12 @@ NAMESPACE_BEGIN(Grid);
 // Move control to configure.ac and Config.h?
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
+#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
 
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
@@ -65,6 +71,21 @@ enum ViewMode {
   CpuWriteDiscard = 0x10 // same for now
 };
 
+struct MemoryStatus {
+  uint64_t     DeviceBytes;
+  uint64_t     DeviceLRUBytes;
+  uint64_t     DeviceMaxBytes;
+  uint64_t     HostToDeviceBytes;
+  uint64_t     DeviceToHostBytes;
+  uint64_t     HostToDeviceXfer;
+  uint64_t     DeviceToHostXfer;
+  uint64_t     DeviceEvictions;
+  uint64_t     DeviceDestroy;
+  uint64_t     DeviceAllocCacheBytes;
+  uint64_t     HostAllocCacheBytes;
+};
+
+
 class MemoryManager {
 private:
 
@@ -78,21 +99,23 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=6;
+  static const int NallocType=9;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
+  static uint64_t CacheBytes[NallocType];
 
   /////////////////////////////////////////////////
   // Free pool
   /////////////////////////////////////////////////
   static void *Insert(void *ptr,size_t bytes,int type) ;
   static void *Lookup(size_t bytes,int type) ;
-  static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
-  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
+  static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
+  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
+  static void Audit(std::string s);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -112,7 +135,28 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
- 
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
+  
+  static uint64_t     DeviceCacheBytes();
+  static uint64_t     HostCacheBytes();
+
+  static MemoryStatus GetFootprint(void) {
+    MemoryStatus stat;
+    stat.DeviceBytes       = DeviceBytes;
+    stat.DeviceLRUBytes    = DeviceLRUBytes;
+    stat.DeviceMaxBytes    = DeviceMaxBytes;
+    stat.HostToDeviceBytes = HostToDeviceBytes;
+    stat.DeviceToHostBytes = DeviceToHostBytes;
+    stat.HostToDeviceXfer  = HostToDeviceXfer;
+    stat.DeviceToHostXfer  = DeviceToHostXfer;
+    stat.DeviceEvictions   = DeviceEvictions;
+    stat.DeviceDestroy     = DeviceDestroy;
+    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
+    stat.HostAllocCacheBytes   = HostCacheBytes();
+    return stat;
+  };
+  
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////
@@ -169,6 +213,8 @@ private:
 
  public:
   static void Print(void);
+  static void PrintAll(void);
+  static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);
   static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);

Grid/allocator/MemoryManagerCache.cc

@@ -3,8 +3,13 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-//define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define dprintf(...)
+
+#define MAXLINE 512
+static char print_buffer [ MAXLINE ];
+
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+//#define dprintf(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -23,6 +28,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -104,15 +111,17 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    AccCache.AccPtr=(uint64_t) NULL;
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -121,26 +130,36 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove entry
-  // Cannot be locked. If allocated must be in LRU pool.
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-  assert(AccCache.accLock==0);
-  assert(AccCache.cpuLock==0);
+  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
+	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  assert(AccCache.accLock==0); // Cannot evict so logic bomb
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
+    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  uint64_t CpuPtr = AccCache.CpuPtr;
-  EntryErase(CpuPtr);
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -150,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -165,7 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -191,6 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -202,6 +222,7 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -212,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back();
+      uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
+    } else {
+      return;
     }
   }
 }
@@ -241,11 +265,12 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
+    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-		    (uint64_t)bytes);
+	            (uint64_t)bytes,
+		    (uint64_t)AccCache.accLock);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -280,6 +305,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
+    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -292,28 +318,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  // If view is opened on device remove from LRU
+  assert(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
+    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -334,10 +362,12 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
-
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
+  } else {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -374,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  if (!AccCache.AccPtr) {
-     EvictVictims(bytes);
-  }
+  // CPU doesn't need to free space
+  //  if (!AccCache.AccPtr) {
+  //    EvictVictims(bytes);
+  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -429,20 +460,29 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 }
 void  MemoryManager::Print(void)
 {
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  PrintBytes();
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -452,13 +492,13 @@ void  MemoryManager::Print(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -472,6 +512,87 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
+void MemoryManager::Audit(std::string s)
+{
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  uint64_t LockedBytes=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    assert(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    assert(AccCache.LRU_valid==1);
+    assert(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      assert(AccCache.LRU_valid==0);
+      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+		<< "\t cpuLock  " << AccCache.cpuLock
+		<< "\t accLock  " << AccCache.accLock
+		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
+    }
+
+    assert( AccCache.cpuLock== 0 ) ;
+    assert( AccCache.accLock== 0 ) ;
+  }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  assert(LruBytes1==LruBytes2);
+  assert(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  assert(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  assert(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+}
+
+void MemoryManager::PrintState(void* _CpuPtr)
+{
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+    if ( AccCache.state==EvictNext) str = std::string("EvictNext");
+
+    std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\t" << AccCache.cpuLock
+    << "\t" << AccCache.accLock
+    << "\t" << AccCache.LRU_valid<<std::endl;
+
+  } else {
+    std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl; 
+  }
+}
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryManagerShared.cc

@@ -12,11 +12,19 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
+void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
+void  MemoryManager::PrintState(void* CpuPtr)
+{
+std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
+};
 void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_mpi3.cc

@@ -388,18 +388,21 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
     // TODO : make a OMP loop on CPU, call threaded bcopy
     void *shm = (void *) this->ShmBufferTranslate(dest,recv);
     assert(shm!=NULL);
+    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
     acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
-    acceleratorCopySynchronise(); // MPI prob slower
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-    this->StencilSendToRecvFromComplete(list,dir);
-  }
+  //  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  //    this->StencilSendToRecvFromComplete(list,dir);
+  //  }
 
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
+  //   std::cout << "Copy Synchronised\n"<<std::endl;
+  acceleratorCopySynchronise();
+
   int nreq=list.size();
 
   if (nreq==0) return;

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_base.h

@@ -88,6 +88,13 @@ public:
     LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
     accessor.ViewClose();
   }
+
+  // Helper function to print the state of this object in the AccCache
+  void PrintCacheState(void)
+  {
+    MemoryManager::PrintState(this->_odata);
+  }
+
   /////////////////////////////////////////////////////////////////////////////////
   // Return a view object that may be dereferenced in site loops.
   // The view is trivially copy constructible and may be copied to an accelerator device
@@ -281,8 +288,8 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
@@ -296,8 +303,8 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });

Grid/lattice/Lattice_crc.h

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

Grid/lattice/Lattice_reduction.h

@@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
+#if defined(GRID_SYCL)
+#include <Grid/lattice/Lattice_reduction_sycl.h>
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -127,7 +130,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -136,25 +139,50 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
 #endif  
 }
+template<class vobj>
+inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  return sumD_gpu_large(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+  Integer osites = arg.Grid()->oSites();
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  typename vobj::scalar_object ssum;
+  autoView( arg_v, arg, AcceleratorRead);
+  ssum= sum_gpu(&arg_v[0],osites);
+#else
+  autoView(arg_v, arg, CpuRead);
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif  
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu(&arg_v[0],osites);
+  auto ssum= sum_gpu_large(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
   auto ssum= sum_cpu(&arg_v[0],osites);
-#endif  
+#endif
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -210,11 +238,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-    
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
-
+    // This code could read coalesce
     // GPU - SIMT lane compliance...
     accelerator_for( ss, sites, 1,{
 	auto x_l = left_v[ss];

Grid/lattice/Lattice_reduction_gpu.h

@@ -23,7 +23,7 @@ unsigned int nextPow2(Iterator x) {
 }
 
 template <class Iterator>
-void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
   
   int device;
 #ifdef GRID_CUDA
@@ -37,14 +37,13 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  
+  /*  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
-  std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << warpSize << 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);
@@ -52,10 +51,14 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   
   // let the number of threads in a block be a multiple of 2, starting from warpSize
   threads = warpSize;
+  if ( threads*sizeofsobj > sharedMemPerBlock ) {
+    std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
+    return 0;
+  }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  
+  return 1;
 }
 
 template <class sobj, class Iterator>
@@ -195,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -204,7 +207,9 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   Integer size = osites*nsimd;
 
   Integer numThreads, numBlocks;
-  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  assert(ok);
+
   Integer smemSize = numThreads * sizeof(sobj);
 
   Vector<sobj> buffer(numBlocks);
@@ -215,6 +220,54 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   auto result = buffer_v[0];
   return result;
 }
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobj;
+  sobj ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<vector> buffer(osites);
+  vector *dat = (vector *)lat;
+  vector *buf = &buffer[0];
+  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,1,{
+	buf[ss] = dat[ss*words+w];
+      });
+      
+    ret_p[w] = sumD_gpu_small(tbuf,osites);
+  }
+  return ret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobj;
+  sobj ret;
+  
+  Integer nsimd= vobj::Nsimd();
+  Integer size = osites*nsimd;
+  Integer numThreads, numBlocks;
+  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  
+  if ( ok ) {
+    ret = sumD_gpu_small(lat,osites);
+  } else {
+    ret = sumD_gpu_large(lat,osites);
+  }
+  return ret;
+}
+
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Return as same precision as input performing reduction in double precision though
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -227,6 +280,13 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
-
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu_large(lat,osites);
+  return result;
+}
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_sycl.h

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

Grid/lattice/Lattice_transfer.h

@@ -85,6 +85,76 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
   });
 }
 
+template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
+{
+  half.Checkerboard() = cb;
+  autoView(half_v, half, AcceleratorWrite);
+  autoView(full_v, full, AcceleratorRead);
+  Coordinate rdim_full             = full.Grid()->_rdimensions;
+  Coordinate rdim_half             = half.Grid()->_rdimensions;
+  unsigned long ndim_half          = half.Grid()->_ndimension;
+  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
+  Coordinate ostride_half          = half.Grid()->_ostride;
+  accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{
+    
+    Coordinate coor;
+    int cbos;
+    int linear=0;
+
+    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
+    assert(coor.size()==ndim_half);
+
+    for(int d=0;d<ndim_half;d++){ 
+      if(checker_dim_mask_half[d]) linear += coor[d];
+    }
+    cbos = (linear&0x1);
+
+    if (cbos==cb) {
+      int ssh=0;
+      for(int d=0;d<ndim_half;d++) {
+        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
+        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
+      }
+      coalescedWrite(half_v[ssh],full_v(ss));
+    }
+  });
+}
+template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
+{
+  int cb = half.Checkerboard();
+  autoView(half_v , half, AcceleratorRead);
+  autoView(full_v , full, AcceleratorWrite);
+  Coordinate rdim_full             = full.Grid()->_rdimensions;
+  Coordinate rdim_half             = half.Grid()->_rdimensions;
+  unsigned long ndim_half          = half.Grid()->_ndimension;
+  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
+  Coordinate ostride_half          = half.Grid()->_ostride;
+  accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{
+
+    Coordinate coor;
+    int cbos;
+    int linear=0;
+  
+    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
+    assert(coor.size()==ndim_half);
+
+    for(int d=0;d<ndim_half;d++){ 
+      if(checker_dim_mask_half[d]) linear += coor[d];
+    }
+    cbos = (linear&0x1);
+
+    if (cbos==cb) {
+      int ssh=0;
+      for(int d=0;d<ndim_half;d++){
+        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
+        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
+      }
+      coalescedWrite(full_v[ss],half_v(ssh));
+    }
+
+  });
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Flexible Type Conversion for internal promotion to double as well as graceful
 // treatment of scalar-compatible types
@@ -218,7 +288,36 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
   }
 }
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               const std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = fineData.size();
+  assert(coarseData.size() == NBatch);
 
+  GridBase * fine  = fineData[0].Grid();
+  GridBase * coarse= coarseData[0].Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  std::vector<Lattice<vobj>> fineDataCopy = fineData;
+
+  autoView(ip_, ip, AcceleratorWrite);
+  for(int v=0;v<nbasis;v++) {
+    for (int k=0; k<NBatch; k++) {
+      autoView( coarseData_ , coarseData[k], AcceleratorWrite);
+      blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
+      accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+        convertType(coarseData_[sc](v),ip_[sc]);
+      });
+
+      // improve numerical stability of projection
+      // |fine> = |fine> - <basis|fine> |basis>
+      ip=-ip;
+      blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]); 
+    }
+  }
+}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -520,6 +619,26 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 }
 #endif
 
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = coarseData.size();
+  assert(fineData.size() == NBatch);
+
+  GridBase * fine   = fineData[0].Grid();
+  GridBase * coarse = coarseData[0].Grid();
+  for (int k=0; k<NBatch; k++)
+    fineData[k]=Zero();
+  for (int i=0;i<nbasis;i++) {
+    for (int k=0; k<NBatch; k++) {
+      Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
+      blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
+    }
+  }
+}
+
 // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
 // Simd layouts need not match since we use peek/poke Local
 template<class vobj,class vvobj>

Grid/log/Log.cc

@@ -65,29 +65,40 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
+GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
+GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(0);
+  GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
+  GridLogMemory.Active(0); 
+  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
+  GridLogDslash.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
+  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
+    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
+    if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,7 +138,8 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      stream << log.colour() <<  std::left;
+      //      stream << log.colour() <<  std::left;
+      stream <<  std::left;
       if (log.chanWidth > 0)
       {
         stream << std::setw(log.chanWidth);
@@ -153,9 +154,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      stream << log.colour();
+      //      stream << log.colour();
+      stream <<  std::right;
       stream.flags(f);
-
       return stream;
     } else { 
       return devnull;
@@ -180,8 +181,12 @@ extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
 extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
+extern GridLogger GridLogDslash;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogHMC;
+extern GridLogger GridLogMemory;
+extern GridLogger GridLogTracing;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/IldgIO.h

@@ -31,6 +31,7 @@ directory
 #include <fstream>
 #include <iomanip>
 #include <iostream>
+#include <string>
 #include <map>
 
 #include <pwd.h>
@@ -576,6 +577,8 @@ class ScidacReader : public GridLimeReader {
     std::string rec_name(ILDG_BINARY_DATA);
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
       if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+  // in principle should do the line below, but that breaks backard compatibility with old data
+  // skipPastObjectRecord(std::string(GRID_FIELD_NORM));
 	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
 	return;
       }
@@ -652,7 +655,8 @@ class IldgWriter : public ScidacWriter {
     // Fill ILDG header data struct
     //////////////////////////////////////////////////////
     ildgFormat ildgfmt ;
-    ildgfmt.field     = std::string("su3gauge");
+    const std::string stNC = std::to_string( Nc ) ;
+    ildgfmt.field          = std::string("su"+stNC+"gauge");
 
     if ( format == std::string("IEEE32BIG") ) { 
       ildgfmt.precision = 32;
@@ -869,7 +873,8 @@ class IldgReader : public GridLimeReader {
     } else { 
 
       assert(found_ildgFormat);
-      assert ( ildgFormat_.field == std::string("su3gauge") );
+      const std::string stNC = std::to_string( Nc ) ;
+      assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
 
       ///////////////////////////////////////////////////////////////////////////////////////
       // Populate our Grid metadata as best we can
@@ -877,7 +882,7 @@ class IldgReader : public GridLimeReader {
 
       std::ostringstream vers; vers << ildgFormat_.version;
       FieldMetaData_.hdr_version = vers.str();
-      FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
+      FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
 
       FieldMetaData_.nd=4;
       FieldMetaData_.dimension.resize(4);

Grid/parallelIO/MetaData.h

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

Grid/parallelIO/NerscIO.h

@@ -9,6 +9,7 @@
     Author: Matt Spraggs <matthew.spraggs@gmail.com>
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -30,6 +31,8 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
+#include <string>
+
 NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
@@ -145,15 +148,17 @@ public:
 
     std::string format(header.floating_point);
 
-    int ieee32big = (format == std::string("IEEE32BIG"));
-    int ieee32    = (format == std::string("IEEE32"));
-    int ieee64big = (format == std::string("IEEE64BIG"));
-    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
+    const int ieee32big = (format == std::string("IEEE32BIG"));
+    const int ieee32    = (format == std::string("IEEE32"));
+    const int ieee64big = (format == std::string("IEEE64BIG"));
+    const int ieee64    = (format == std::string("IEEE64") || \
+			   format == std::string("IEEE64LITTLE"));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     // depending on datatype, set up munger;
     // munger is a function of <floating point, Real, data_type>
-    if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
+    const std::string stNC = std::to_string( Nc ) ;
+    if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@@ -164,7 +169,7 @@ public:
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
-    } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
+    } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@@ -209,27 +214,29 @@ public:
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
-					std::string ens_label = std::string("DWF"))
+					std::string ens_label = std::string("DWF"),
+					std::string ens_id = std::string("UKQCD"),
+					unsigned int sequence_number = 1)
   {
-    writeConfiguration(Umu,file,0,1,ens_label);
+    writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
   }
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32,
-					std::string ens_label = std::string("DWF"))
+					std::string ens_label = std::string("DWF"),
+					std::string ens_id = std::string("UKQCD"),
+					unsigned int sequence_number = 1)
   {
     typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
-    ///////////////////////////////////////////
-    // Following should become arguments
-    ///////////////////////////////////////////
-    header.sequence_number = 1;
-    header.ensemble_id     = std::string("UKQCD");
+    header.sequence_number = sequence_number;
+    header.ensemble_id     = ens_id;
     header.ensemble_label  = ens_label;
+    header.hdr_version     = "1.0" ;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -243,10 +250,14 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write 3x3 double
-    header.floating_point = std::string("IEEE64BIG");
-    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
-    GaugeSimpleUnmunger<fobj3D,sobj> munge;
+    // Sod it -- always write NcxNc double
+    header.floating_point  = std::string("IEEE64BIG");
+    const std::string stNC = std::to_string( Nc ) ;
+    if( two_row ) {
+      header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
+    } else {
+      header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
+    }
     if ( grid->IsBoss() ) { 
       truncate(file);
       offset = writeHeader(header,file);
@@ -254,8 +265,15 @@ public:
     grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
-					      nersc_csum,scidac_csuma,scidac_csumb);
+    if( two_row ) {
+      Gauge3x2unmunger<fobj2D,sobj> munge;
+      BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
+						nersc_csum,scidac_csuma,scidac_csumb);
+    } else {
+      GaugeSimpleUnmunger<fobj3D,sobj> munge;
+      BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
+						nersc_csum,scidac_csuma,scidac_csumb);
+    }
     header.checksum = nersc_csum;
     if ( grid->IsBoss() ) { 
       writeHeader(header,file);
@@ -287,8 +305,7 @@ public:
     header.plaquette=0.0;
     MachineCharacteristics(header);
 
-	uint64_t offset;
-  
+    uint64_t offset;
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");
@@ -328,7 +345,7 @@ public:
 
     GridBase *grid = parallel.Grid();
 
-	uint64_t offset = readHeader(file,grid,header);
+    uint64_t offset = readHeader(file,grid,header);
 
     FieldMetaData clone(header);
 

Grid/perfmon/PerfCount.cc

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

Grid/perfmon/PerfCount.h

@@ -30,6 +30,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
+
+#ifndef __SSC_START
+#define __SSC_START
+#define __SSC_STOP
+#endif
+
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>
@@ -72,17 +78,9 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
-#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
-#define __SSC_STOP  __SSC_MARK(0x110)
-#define __SSC_START __SSC_MARK(0x111)
-
 
 #else
 
-#define __SSC_MARK(mark) 
-#define __SSC_STOP  
-#define __SSC_START 
-
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

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

Grid/perfmon/Tracing.h

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

Grid/pugixml/pugixml.cc

@@ -16,8 +16,12 @@
 
 #ifdef __NVCC__
 #pragma push
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+#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

@@ -451,9 +451,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
+#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuWrite);
+  autoView(f_v,f,CpuRead);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -465,12 +476,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
+#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuRead);
+  autoView(f_v,f,CpuWrite);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -482,6 +504,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/fermion/CayleyFermion5D.h

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

Grid/qcd/action/fermion/CloverHelpers.h

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

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -0,0 +1,241 @@
+/*************************************************************************************
+
+    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 fixedBoundaries;
+
+  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

@@ -53,6 +53,7 @@ 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);
@@ -137,21 +138,52 @@ typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
 // Clover fermions
-typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
-typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
-typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
+template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
+template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
 
-typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
-typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
-typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
+typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
+typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
+typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
+typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
+typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
 
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
+typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
+
+typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
+typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
+typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
+
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+
+// Compact Clover fermions
+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/WilsonCloverFermion.h

@@ -4,10 +4,11 @@
 
     Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
 
-    Copyright (C) 2017
+    Copyright (C) 2017 - 2022
 
     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
@@ -29,7 +30,9 @@
 
 #pragma once
 
-#include <Grid/Grid.h>
+#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -49,19 +52,16 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template <class Impl>
-class WilsonCloverFermion : public WilsonFermion<Impl>
+template<class Impl, class CloverHelpers>
+class WilsonCloverFermion : public WilsonFermion<Impl>,
+                            public WilsonCloverHelpers<Impl>
 {
 public:
-  // Types definitions
   INHERIT_IMPL_TYPES(Impl);
-  template <typename vtype>
-  using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef iImplClover<Simd> SiteCloverType;
-  typedef Lattice<SiteCloverType> CloverFieldType;
+  INHERIT_CLOVER_TYPES(Impl);
 
-public:
-  typedef WilsonFermion<Impl> WilsonBase;
+  typedef WilsonFermion<Impl>       WilsonBase;
+  typedef WilsonCloverHelpers<Impl> Helpers;
 
   virtual int    ConstEE(void)     { return 0; };
   virtual void Instantiatable(void){};
@@ -72,42 +72,7 @@ public:
                       const RealD _csw_r = 0.0,
                       const RealD _csw_t = 0.0,
                       const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
-                      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);
-  }
+                      const ImplParams &impl_p = ImplParams());
 
   virtual void M(const FermionField &in, FermionField &out);
   virtual void Mdag(const FermionField &in, FermionField &out);
@@ -124,250 +89,21 @@ public:
   void ImportGauge(const GaugeField &_Umu);
 
   // Derivative parts unpreconditioned pseudofermions
-  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());
+  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
 
-    // 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:
+public:
   // 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
-  CloverFieldType CloverTerm, CloverTermInv;                 // Clover term
-  CloverFieldType CloverTermEven, CloverTermOdd;             // Clover term EO
-  CloverFieldType CloverTermInvEven, CloverTermInvOdd;       // Clover term Inv EO
-  CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term 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;
-  }
+  CloverField CloverTerm, CloverTermInv;                     // Clover term
+  CloverField CloverTermEven, CloverTermOdd;                 // Clover term EO
+  CloverField CloverTermInvEven, CloverTermInvOdd;           // Clover term Inv EO
+  CloverField CloverTermDagEven, CloverTermDagOdd;           // Clover term Dag EO
+  CloverField CloverTermInvDagEven, CloverTermInvDagOdd;     // Clover term Inv Dag EO
 };
+
 NAMESPACE_END(Grid);
 
 

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -0,0 +1,763 @@
+/*************************************************************************************
+
+    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

@@ -0,0 +1,90 @@
+/*************************************************************************************
+
+    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

@@ -117,19 +117,19 @@ public:
     typedef decltype(coalescedRead(*in))    sobj;
     typedef decltype(coalescedRead(*out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
-    const vobj *vp1 = &in[m];
-    const vobj *vp = (lane&mask) ? vp1:vp0;
-    auto sa = coalescedRead(*vp,j0);
-    auto sb = coalescedRead(*vp,j1);
+    const vobj *vp0 = &in[k];  // out0[j] = merge low bit of type from in[k] and in[m] 
+    const vobj *vp1 = &in[m];  // out1[j] = merge hi  bit of type from in[k] and in[m]
+    const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
+    auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
+    auto sb = coalescedRead(*vp,j1); // lane to read for out 1
     hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);
-    projector::Proj(psb,sb,mu,dag);
+    projector::Proj(psa,sa,mu,dag);  // spin project the result0
+    projector::Proj(psb,sb,mu,dag);  // spin project the result1
     coalescedWrite(out0[j],psa);
     coalescedWrite(out1[j],psb);
 #else

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

@@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FiveDimRedBlackGrid,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
-  mass(_mass)
+  mass_plus(_mass), mass_minus(_mass)
 { 
 }
 
@@ -209,8 +209,8 @@ void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
   Vector<Coeff_t> diag (Ls,1.0);
-  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
-  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -220,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
   Vector<Coeff_t> diag = bs;
   Vector<Coeff_t> upper= cs;
   Vector<Coeff_t> lower= cs; 
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
 }
 // FIXME Redunant with the above routine; check this and eliminate
@@ -235,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
   }
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -250,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &
     upper[i]=-cee[i];
     lower[i]=-cee[i];
   }
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -266,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &
     // Assemble the 5d matrix
     if ( s==0 ) {
       upper[s] = -cee[s+1] ;
-      lower[s] = mass*cee[Ls-1];
+      lower[s] = mass_minus*cee[Ls-1];
     } else if ( s==(Ls-1)) { 
-      upper[s] = mass*cee[0];
+      upper[s] = mass_plus*cee[0];
       lower[s] = -cee[s-1];
     } else {
       upper[s]=-cee[s+1];
@@ -291,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
   Vector<Coeff_t> diag(Ls,1.0);
   Vector<Coeff_t> upper(Ls,-1.0);
   Vector<Coeff_t> lower(Ls,-1.0);
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_plus*upper[Ls-1];
+  lower[0]   =-mass_minus*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
 }
 
@@ -307,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
       upper[s] = cs[s+1];
-      lower[s] =-mass*cs[Ls-1];
+      lower[s] =-mass_minus*cs[Ls-1];
     } else if ( s==(Ls-1) ) { 
-      upper[s] =-mass*cs[0];
+      upper[s] =-mass_plus*cs[0];
       lower[s] = cs[s-1];
     } else { 
       upper[s] = cs[s+1];
@@ -552,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
-      leem[i]=mass*cee[Ls-1]/bee[0];
+      leem[i]=mass_minus*cee[Ls-1]/bee[0];
       for(int j=0;j<i;j++) {
 	assert(bee[j+1]!=Coeff_t(0.0));
 	leem[i]*= aee[j]/bee[j+1];
@@ -560,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
-      ueem[i]=mass;
+      ueem[i]=mass_plus;
       for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
       ueem[i]*= aee[0]/bee[0];
       
@@ -573,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
   }
 	
   { 
-    Coeff_t delta_d=mass*cee[Ls-1];
+    Coeff_t delta_d=mass_minus*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
       assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
@@ -642,6 +642,10 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
+
+  assert(mass_plus == mass_minus);
+  RealD mass = mass_plus;
+  
 #if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
@@ -777,6 +781,8 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   assert(mu>=0);
   assert(mu<Nd);
 
+  assert(mass_plus == mass_minus);
+  RealD mass = mass_plus;
 
 #if 0
   int tshift = (mu == Nd-1) ? 1 : 0;
@@ -828,6 +834,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
 #if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
+  unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
   ////////////////////////////////////////////////
@@ -880,7 +887,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
-  RealD sign = 1; // sign flip for vector/tadpole
+  RealD sign = 1.0; // sign flip for vector/tadpole
   if ( curr_type == Current::Axial ) {
     for(int s=0;s<Ls/2;s++){
       G_s[s] = -1.0;
@@ -890,7 +897,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
     auto b=this->_b;
     auto c=this->_c;
     if ( b == 1 && c == 0 ) {
-      sign = -1;    
+      sign = -1.0;    
     }
     else {
       std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
@@ -934,7 +941,13 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
     tmp    = Cshift(tmp,mu,-1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
     tmp = -G_s[s]*( Utmp + gmu*Utmp );
-    tmp    = where((lcoor>=tmin+tshift),tmp,zz); // Mask the time 
+    // 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);
+    }
     L_Q   += where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
 
     InsertSlice(L_Q, q_out, s , 0);

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

@@ -0,0 +1,377 @@
+/*************************************************************************************
+
+    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)
+  , fixedBoundaries(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 (fixedBoundaries) {
+    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(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::DhopOE(in, out, dag);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::DhopEO(in, out, dag);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+  WilsonBase::DhopDir(in, out, dir, disp);
+  if(this->fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+  WilsonBase::DhopDirAll(in, out);
+  if(this->fixedBoundaries) {
+    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(fixedBoundaries) 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(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
+  WilsonBase::Meooe(in, out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
+  WilsonBase::MeooeDag(in, out);
+  if(fixedBoundaries) 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(fixedBoundaries) 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(fixedBoundaries) 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(!fixedBoundaries); // 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);
+  CloverField TmpInverse(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;
+
+  // Instantiate the clover term
+  // - In case of the standard clover the mass term is added
+  // - In case of the exponential clover the clover term is exponentiated
+  double t4 = usecond();
+  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
+
+  // Convert the data layout of the clover term
+  double t5 = usecond();
+  CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
+
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
+  double t6 = usecond();
+  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+
+  // Invert the Clover term
+  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
+  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
+  // TODO: For now this inversion is explictly done on the CPU
+  double t7 = usecond();
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
+
+  // Fill the remaining clover fields
+  double 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 << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
+}
+
+NAMESPACE_END(Grid);

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

@@ -2,12 +2,13 @@
 
     Grid physics library, www.github.com/paboyle/Grid
 
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
+    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
 
-    Copyright (C) 2017
+    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
@@ -33,9 +34,48 @@
 
 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>
-void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -49,8 +89,8 @@ void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
   out += temp;
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -64,13 +104,16 @@ void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
   out += temp;
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
 {
+  double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
+  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);
@@ -79,52 +122,20 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   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  = fillCloverYZ(Bx) * csw_r;
-  CloverTerm += fillCloverXZ(By) * csw_r;
-  CloverTerm += fillCloverXY(Bz) * csw_r;
-  CloverTerm += fillCloverXT(Ex) * csw_t;
-  CloverTerm += fillCloverYT(Ey) * csw_t;
-  CloverTerm += fillCloverZT(Ez) * csw_t;
-  CloverTerm += 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);
-    });
-  }
+  CloverTerm  = Helpers::fillCloverYZ(Bx) * csw_r;
+  CloverTerm += Helpers::fillCloverXZ(By) * csw_r;
+  CloverTerm += Helpers::fillCloverXY(Bz) * csw_r;
+  CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
+  CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
+  CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
+   
+  double t4 = usecond();
+  CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
 
+  double t5 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@@ -137,37 +148,47 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 
   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>
-void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseNo);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseNo);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseYes);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseYes);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
 {
   out.Checkerboard() = in.Checkerboard();
-  CloverFieldType *Clover;
+  CloverField *Clover;
   assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
 
   if (dag)
@@ -182,12 +203,12 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
       {
         Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
       }
-      out = *Clover * in;
+      Helpers::multCloverField(out, *Clover, in);
     }
     else
     {
       Clover = (inv) ? &CloverTermInv : &CloverTerm;
-      out = adj(*Clover) * in;
+      Helpers::multCloverField(out, *Clover, in); // don't bother with adj, hermitian anyway
     }
   }
   else
@@ -205,29 +226,109 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
         //  std::cout << "Calling clover term Even" << std::endl;
         Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
       }
-      out = *Clover * in;
+      Helpers::multCloverField(out, *Clover, in);
       //  std::cout << GridLogMessage << "*Clover.Checkerboard() "  << (*Clover).Checkerboard() << std::endl;
     }
     else
     {
       Clover = (inv) ? &CloverTermInv : &CloverTerm;
-      out = *Clover * in;
+      Helpers::multCloverField(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>
-void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
 {
   assert(0);
 }
 
 // Derivative parts
-template <class Impl>
-void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   assert(0); // not implemented yet
 }

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

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

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

@@ -77,23 +77,23 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define REGISTER
 
 #ifdef GRID_SIMT
-#define LOAD_CHIMU(ptype)		\
+#define LOAD_CHIMU(Ptype)		\
   {const SiteSpinor & ref (in[offset]);	\
-    Chimu_00=coalescedReadPermute<ptype>(ref()(0)(0),perm,lane);	\
-    Chimu_01=coalescedReadPermute<ptype>(ref()(0)(1),perm,lane);		\
-    Chimu_02=coalescedReadPermute<ptype>(ref()(0)(2),perm,lane);		\
-    Chimu_10=coalescedReadPermute<ptype>(ref()(1)(0),perm,lane);		\
-    Chimu_11=coalescedReadPermute<ptype>(ref()(1)(1),perm,lane);		\
-    Chimu_12=coalescedReadPermute<ptype>(ref()(1)(2),perm,lane);		\
-    Chimu_20=coalescedReadPermute<ptype>(ref()(2)(0),perm,lane);		\
-    Chimu_21=coalescedReadPermute<ptype>(ref()(2)(1),perm,lane);		\
-    Chimu_22=coalescedReadPermute<ptype>(ref()(2)(2),perm,lane);		\
-    Chimu_30=coalescedReadPermute<ptype>(ref()(3)(0),perm,lane);		\
-    Chimu_31=coalescedReadPermute<ptype>(ref()(3)(1),perm,lane);		\
-    Chimu_32=coalescedReadPermute<ptype>(ref()(3)(2),perm,lane);	}
+    Chimu_00=coalescedReadPermute<Ptype>(ref()(0)(0),perm,lane);	\
+    Chimu_01=coalescedReadPermute<Ptype>(ref()(0)(1),perm,lane);		\
+    Chimu_02=coalescedReadPermute<Ptype>(ref()(0)(2),perm,lane);		\
+    Chimu_10=coalescedReadPermute<Ptype>(ref()(1)(0),perm,lane);		\
+    Chimu_11=coalescedReadPermute<Ptype>(ref()(1)(1),perm,lane);		\
+    Chimu_12=coalescedReadPermute<Ptype>(ref()(1)(2),perm,lane);		\
+    Chimu_20=coalescedReadPermute<Ptype>(ref()(2)(0),perm,lane);		\
+    Chimu_21=coalescedReadPermute<Ptype>(ref()(2)(1),perm,lane);		\
+    Chimu_22=coalescedReadPermute<Ptype>(ref()(2)(2),perm,lane);		\
+    Chimu_30=coalescedReadPermute<Ptype>(ref()(3)(0),perm,lane);		\
+    Chimu_31=coalescedReadPermute<Ptype>(ref()(3)(1),perm,lane);		\
+    Chimu_32=coalescedReadPermute<Ptype>(ref()(3)(2),perm,lane);	}
 #define PERMUTE_DIR(dir) ;
 #else
-#define LOAD_CHIMU(ptype)		\
+#define LOAD_CHIMU(Ptype)		\
   {const SiteSpinor & ref (in[offset]);	\
     Chimu_00=ref()(0)(0);\
     Chimu_01=ref()(0)(1);\
@@ -109,12 +109,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     Chimu_32=ref()(3)(2);}
 
 #define PERMUTE_DIR(dir)			\
-  permute##dir(Chi_00,Chi_00);	\
-      permute##dir(Chi_01,Chi_01);\
-      permute##dir(Chi_02,Chi_02);\
-      permute##dir(Chi_10,Chi_10);	\
-      permute##dir(Chi_11,Chi_11);\
-      permute##dir(Chi_12,Chi_12);
+  permute##dir(Chi_00,Chi_00);			\
+  permute##dir(Chi_01,Chi_01);			\
+  permute##dir(Chi_02,Chi_02);			\
+  permute##dir(Chi_10,Chi_10);			\
+  permute##dir(Chi_11,Chi_11);			\
+  permute##dir(Chi_12,Chi_12);
 
 #endif
 
@@ -371,88 +371,91 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   result_32-= UChi_12;
 
 #define HAND_STENCIL_LEGB(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
-  if ( local ) {				\
-    LOAD_CHIMU(PERM);				\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else {					\
-    LOAD_CHI;					\
-  }						\
-  acceleratorSynchronise();			\
-  MULT_2SPIN(DIR);				\
-  RECON;					
+  {int ptype;					\
+   SE=st.GetEntry(ptype,DIR,ss);		\
+   auto offset = SE->_offset;			\
+   auto local  = SE->_is_local;			\
+   auto perm   = SE->_permute;			\
+   if ( local ) {				\
+     LOAD_CHIMU(PERM);				\
+     PROJ;					\
+     if ( perm) {				\
+       PERMUTE_DIR(PERM);			\
+     }						\
+   } else {					\
+     LOAD_CHI;					\
+   }						\
+   acceleratorSynchronise();			\
+   MULT_2SPIN(DIR);				\
+   RECON;					}
 
-#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
-  SE=&st_p[DIR+8*ss];				\
-  ptype=st_perm[DIR];				\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
-  if ( local ) {				\
-    LOAD_CHIMU(PERM);				\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else {					\
-    LOAD_CHI;					\
-  }						\
-  acceleratorSynchronise();			\
-  MULT_2SPIN(DIR);				\
-  RECON;					
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)		\
+  { SE=&st_p[DIR+8*ss];						\
+  auto ptype=st_perm[DIR];					\
+  auto offset = SE->_offset;					\
+  auto local  = SE->_is_local;					\
+  auto perm   = SE->_permute;					\
+  if ( local ) {						\
+    LOAD_CHIMU(PERM);						\
+    PROJ;							\
+    if ( perm) {						\
+      PERMUTE_DIR(PERM);					\
+    }								\
+  } else {							\
+    LOAD_CHI;							\
+  }								\
+  acceleratorSynchronise();					\
+  MULT_2SPIN(DIR);						\
+  RECON;					}
 
 #define HAND_STENCIL_LEGA(PROJ,PERM,DIR,RECON)				\
-  SE=&st_p[DIR+8*ss];							\
-  ptype=st_perm[DIR];							\
- /*SE=st.GetEntry(ptype,DIR,ss);*/					\
-  offset = SE->_offset;				\
-  perm   = SE->_permute;			\
-  LOAD_CHIMU(PERM);				\
-  PROJ;						\
-  MULT_2SPIN(DIR);				\
-  RECON;					
+  { SE=&st_p[DIR+8*ss];							\
+    auto ptype=st_perm[DIR];						\
+    /*SE=st.GetEntry(ptype,DIR,ss);*/					\
+    auto offset = SE->_offset;						\
+    auto perm   = SE->_permute;						\
+    LOAD_CHIMU(PERM);							\
+    PROJ;								\
+    MULT_2SPIN(DIR);							\
+    RECON;					}
 
 #define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
-  if ( local ) {				\
-    LOAD_CHIMU(PERM);				\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else if ( st.same_node[DIR] ) {		\
-    LOAD_CHI;					\
-  }						\
-  acceleratorSynchronise();			\
-  if (local || st.same_node[DIR] ) {		\
-    MULT_2SPIN(DIR);				\
-    RECON;					\
-  }						\
-  acceleratorSynchronise();			
+  { int ptype;						\
+  SE=st.GetEntry(ptype,DIR,ss);				\
+  auto offset = SE->_offset;					\
+  auto local  = SE->_is_local;					\
+  auto perm   = SE->_permute;					\
+  if ( local ) {						\
+    LOAD_CHIMU(PERM);						\
+    PROJ;							\
+    if ( perm) {						\
+      PERMUTE_DIR(PERM);					\
+    }								\
+  } else if ( st.same_node[DIR] ) {				\
+    LOAD_CHI;							\
+  }								\
+  acceleratorSynchronise();					\
+  if (local || st.same_node[DIR] ) {				\
+    MULT_2SPIN(DIR);						\
+    RECON;							\
+  }								\
+  acceleratorSynchronise();			}
 
 #define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {	\
-    LOAD_CHI;					\
-    MULT_2SPIN(DIR);				\
-    RECON;					\
-    nmu++;					\
-  }						\
-  acceleratorSynchronise();			
+  { int ptype;						\
+  SE=st.GetEntry(ptype,DIR,ss);				\
+  auto offset = SE->_offset;				\
+  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {		\
+    LOAD_CHI;						\
+    MULT_2SPIN(DIR);					\
+    RECON;						\
+    nmu++;						\
+  }							\
+  acceleratorSynchronise();			}
 
-#define HAND_RESULT(ss)				\
-  {						\
-    SiteSpinor & ref (out[ss]);			\
+#define HAND_RESULT(ss)					\
+  {							\
+    SiteSpinor & ref (out[ss]);				\
     coalescedWrite(ref()(0)(0),result_00,lane);		\
     coalescedWrite(ref()(0)(1),result_01,lane);		\
     coalescedWrite(ref()(0)(2),result_02,lane);		\
@@ -563,7 +566,6 @@ WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p,
 
   HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
   HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
   HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
@@ -593,9 +595,7 @@ WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,Site
 
   HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
-
   HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
   HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
   HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
@@ -623,8 +623,6 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView
   HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
-  int offset,local,perm, ptype;
-  
   HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
   HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
   HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
@@ -640,8 +638,8 @@ template<class Impl>  accelerator_inline void
 WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -652,7 +650,6 @@ WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,Si
 
   HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
   ZERO_RESULT;
   HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
@@ -670,8 +667,8 @@ template<class Impl> accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
   typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
@@ -682,7 +679,6 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldVi
   HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
-  int offset,local,perm, ptype;
   ZERO_RESULT;
   HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
   HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
@@ -699,8 +695,8 @@ template<class Impl>  accelerator_inline void
 WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -711,7 +707,7 @@ WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,Si
 
   HAND_DECLARATIONS(Simt);
 
-  int offset, ptype;
+  //  int offset, ptype;
   StencilEntry *SE;
   int nmu=0;
   ZERO_RESULT;
@@ -730,8 +726,8 @@ template<class Impl>  accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
   typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
@@ -742,7 +738,7 @@ void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldVi
   HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
-  int offset, ptype;
+  //  int offset, ptype;
   int nmu=0;
   ZERO_RESULT;
   HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);

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

@@ -498,6 +498,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
+     acceleratorFenceComputeStream();
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
@@ -505,11 +506,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
+     acceleratorFenceComputeStream();
    }
    assert(0 && " Kernel optimisation case not covered ");
   }

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

@@ -0,0 +1,44 @@
+/*************************************************************************************
+
+    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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Grid/qcd/action/gauge/Photon.h

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

Grid/qcd/hmc/UsingHMC.md

@@ -1,61 +1,63 @@
-Using HMC in Grid version 0.5.1
+# Using HMC in Grid
 
-These are the instructions to use the Generalised HMC on Grid version 0.5.1.
-Disclaimer: GRID is still under active development so any information here can be changed in future releases.
+These are the instructions to use the Generalised HMC on Grid as of commit `749b802`.
+Disclaimer: Grid is still under active development so any information here can be changed in future releases.
 
 
-Command line options
-===================
-(relevant file GenericHMCrunner.h)
+## Command line options
+
+(relevant file `GenericHMCrunner.h`)
 The initial configuration can be changed at the command line using 
---StartType <your choice>
-valid choices, one among these
-HotStart, ColdStart, TepidStart, CheckpointStart
-default: HotStart
+`--StartingType STARTING_TYPE`, where `STARTING_TYPE` is one of
+`HotStart`, `ColdStart`, `TepidStart`, and `CheckpointStart`.
+Default: `--StartingType HotStart`
 
-example
-./My_hmc_exec  --StartType HotStart
+Example:
+```
+./My_hmc_exec  --StartingType HotStart
+```
 
-The CheckpointStart option uses the prefix for the configurations and rng seed files defined in your executable and the initial configuration is specified by
---StartTrajectory <integer>
-default: 0
+The `CheckpointStart` option uses the prefix for the configurations and rng seed files defined in your executable and the initial configuration is specified by
+`--StartingTrajectory STARTING_TRAJECTORY`, where `STARTING_TRAJECTORY` is an integer.
+Default: `--StartingTrajectory 0`
 
 The number of trajectories for a specific run are specified at command line by
---Trajectories <integer>
-default: 1
+`--Trajectories TRAJECTORIES`, where `TRAJECTORIES` is an integer.
+Default: `--Trajectories 1`
 
 The number of thermalization steps (i.e. steps when the Metropolis acceptance check is turned off) is specified by
---Thermalizations <integer>
-default: 10
-
+`--Thermalizations THERMALIZATIONS`, where `THERMALIZATIONS` is an integer.
+Default: `--Thermalizations 10`
 
 Any other parameter is defined in the source for the executable.
 
-HMC controls
-===========
+## HMC controls
 
 The lines 
 
+```
   std::vector<int> SerSeed({1, 2, 3, 4, 5});
   std::vector<int> ParSeed({6, 7, 8, 9, 10});
+```
 
 define the seeds for the serial and the parallel RNG.
 
 The line 
 
+```
   TheHMC.MDparameters.set(20, 1.0);// MDsteps, traj length
+```
 
 declares the number of molecular dynamics steps and the total trajectory length.
 
 
-Actions
-======
+## Actions
 
-Action names are defined in the file
-lib/qcd/Actions.h
+Action names are defined in the directory `Grid/qcd/action`.
 
-Gauge actions list:
+Gauge actions list (from `Grid/qcd/action/gauge/Gauge.h`):
 
+```
 WilsonGaugeActionR;
 WilsonGaugeActionF;
 WilsonGaugeActionD;
@@ -68,8 +70,9 @@ IwasakiGaugeActionD;
 SymanzikGaugeActionR;
 SymanzikGaugeActionF;
 SymanzikGaugeActionD;
+```
 
-
+```
 ConjugateWilsonGaugeActionR;
 ConjugateWilsonGaugeActionF;
 ConjugateWilsonGaugeActionD;
@@ -82,26 +85,23 @@ ConjugateIwasakiGaugeActionD;
 ConjugateSymanzikGaugeActionR;
 ConjugateSymanzikGaugeActionF;
 ConjugateSymanzikGaugeActionD;
+```
 
+Each of these action accepts one single parameter at creation time (beta).
+Example for creating a Symanzik action with beta=4.0
 
+```
+  SymanzikGaugeActionR(4.0)
+```
+
+Scalar actions list (from `Grid/qcd/action/scalar/Scalar.h`):
+
+```
 ScalarActionR;
 ScalarActionF;
 ScalarActionD;
+```
 
-
-each of these action accept one single parameter at creation time (beta).
-Example for creating a Symanzik action with beta=4.0
-
-	SymanzikGaugeActionR(4.0)
-
-The suffixes R,F,D in the action names refer to the Real
-(the precision is defined at compile time by the --enable-precision flag in the configure),
-Float and Double, that force the precision of the action to be 32, 64 bit respectively.
-
-
-
-
-
-
-
-
+The suffixes `R`, `F`, `D` in the action names refer to the `Real`
+(the precision is defined at compile time by the `--enable-precision` flag in the configure),
+`Float` and `Double`, that force the precision of the action to be 32, 64 bit respectively.

Grid/qcd/utils/GaugeFix.h

@@ -55,12 +55,12 @@ public:
     }
   }  
 
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
-    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog);
+    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
   }
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
 
     GridBase *grid = Umu.Grid();
 
@@ -122,6 +122,8 @@ public:
 
       }
     }
+    std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
+    if (err_on_no_converge) assert(0);
   };
   static Real SteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform,Real & alpha, GaugeMat & dmuAmu,int orthog) {
     GridBase *grid = U[0].Grid();

Grid/qcd/utils/WilsonLoops.h

@@ -125,7 +125,6 @@ public:
     return sumplaq / vol / faces / Nc; // Nd , Nc dependent... FIXME
   }
 
-
   //////////////////////////////////////////////////
   // average over all x,y,z the temporal loop
   //////////////////////////////////////////////////
@@ -165,7 +164,7 @@ public:
 
     double vol = Umu.Grid()->gSites();
 
-    return p.real() / vol / 4.0 / 3.0;
+    return p.real() / vol / (4.0 * Nc ) ;
   };
 
   //////////////////////////////////////////////////

Grid/serialisation/JSON_IO.cc

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

Grid/serialisation/JSON_IO.h

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

Grid/serialisation/Serialisation.h

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

Grid/stencil/Stencil.h

@@ -80,11 +80,14 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
 ///////////////////////////////////////////////////////////////////
 template<class cobj,class vobj,class compressor>
 void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
-				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
+				 commVector<cobj *> pointers,
+				 int dimension,int plane,
+				 int cbmask,compressor &compress,int type) __attribute__((noinline));
 
 template<class cobj,class vobj,class compressor>
-void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
+				 const Lattice<vobj> &rhs,
+				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -92,14 +95,15 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const La
   int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
   auto rhs_v = rhs.View(AcceleratorRead);
+  auto rhs_p = &rhs_v[0];
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
   accelerator_forNB(j, num, vobj::Nsimd(), {
-      compress.CompressExchange(p0,p1, &rhs_v[0], j,
-			      so+tp[2*j  ].second,
-			      so+tp[2*j+1].second,
-			      type);
+      compress.CompressExchange(p0,p1, rhs_p, j,
+				so+tp[2*j  ].second,
+				so+tp[2*j+1].second,
+				type);
   });
   rhs_v.ViewClose();
 }
@@ -230,8 +234,8 @@ public:
   };
   struct Merge {
     cobj * mpointer;
-    Vector<scalar_object *> rpointers;
-    Vector<cobj *> vpointers;
+    //    std::vector<scalar_object *> rpointers;
+    std::vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
   };
@@ -322,8 +326,8 @@ public:
     int simd_layout     = _grid->_simd_layout[dimension];
     int comm_dim        = _grid->_processors[dimension] >1 ;
 
-    int recv_from_rank;
-    int xmit_to_rank;
+    //    int recv_from_rank;
+    //    int xmit_to_rank;
 
     if ( ! comm_dim ) return 1;
     if ( displacement == 0 ) return 1;
@@ -406,6 +410,7 @@ public:
       comms_bytes+=bytes;
       shm_bytes  +=2*Packets[i].bytes-bytes;
     }
+    _grid->StencilBarrier();// Synch shared memory on a single nodes
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
@@ -420,7 +425,7 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
+    if ( 0 ){
       thread_region {
 	// must be called in parallel region
 	int mythread  = thread_num();
@@ -569,7 +574,7 @@ public:
     d.buffer_size = buffer_size;
     dv.push_back(d);
   }
-  void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
+  void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
     Merge m;
     m.type     = type;
     m.mpointer = merge_p;
@@ -582,6 +587,7 @@ public:
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
     mpi3synctime-=usecond();
+    accelerator_barrier();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
     mpi3synctime+=usecond();
     shmmergetime-=usecond();
@@ -1114,8 +1120,8 @@ public:
     int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    Vector<cobj *> rpointers(maxl);
-    Vector<cobj *> spointers(maxl);
+    std::vector<cobj *> rpointers(maxl);
+    std::vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where

Grid/tensors/Tensor_traits.h

@@ -47,20 +47,20 @@ NAMESPACE_BEGIN(Grid);
   class TypePair {
   public:
     T _internal[2];
-    TypePair<T>& operator=(const Grid::Zero& o) {
+    accelerator TypePair<T>& operator=(const Grid::Zero& o) {
       _internal[0] = Zero();
       _internal[1] = Zero();
       return *this;
     }
 
-    TypePair<T> operator+(const TypePair<T>& o) const {
+    accelerator TypePair<T> operator+(const TypePair<T>& o) const {
       TypePair<T> r;
       r._internal[0] = _internal[0] + o._internal[0];
       r._internal[1] = _internal[1] + o._internal[1];
       return r;
     }
 
-    TypePair<T>& operator+=(const TypePair<T>& o) {
+    accelerator TypePair<T>& operator+=(const TypePair<T>& o) {
       _internal[0] += o._internal[0];
       _internal[1] += o._internal[1];
       return *this;

Grid/threads/Accelerator.cc

@@ -74,29 +74,43 @@ void acceleratorInit(void)
       //      GPU_PROP(singleToDoublePrecisionPerfRatio);
     }
   }
+
   MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
 #undef GPU_PROP_FMT    
 #undef GPU_PROP
 
 #ifdef GRID_DEFAULT_GPU
+  int device = 0;
   // IBM Jsrun makes cuda Device numbering screwy and not match rank
   if ( world_rank == 0 ) {
     printf("AcceleratorCudaInit: using default device \n");
-    printf("AcceleratorCudaInit: assume user either uses a) IBM jsrun, or \n");
+    printf("AcceleratorCudaInit: assume user either uses\n");
+    printf("AcceleratorCudaInit: a) IBM jsrun, or \n");
     printf("AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
     printf("AcceleratorCudaInit: Configure options --enable-setdevice=no \n");
   }
 #else
+  int device = rank;
   printf("AcceleratorCudaInit: rank %d setting device to node rank %d\n",world_rank,rank);
   printf("AcceleratorCudaInit: Configure options --enable-setdevice=yes \n");
-  cudaSetDevice(rank);
 #endif
+
+  cudaSetDevice(device);
+  cudaStreamCreate(&copyStream);
+  const int len=64;
+  char busid[len];
+  if( rank == world_rank ) { 
+    cudaDeviceGetPCIBusId(busid, len, device);
+    printf("local rank %d device %d bus id: %s\n", rank, device, busid);
+  }
+
   if ( world_rank == 0 )  printf("AcceleratorCudaInit: ================================================\n");
 }
 #endif
 
 #ifdef GRID_HIP
 hipDeviceProp_t *gpu_props;
+hipStream_t copyStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -154,16 +168,25 @@ void acceleratorInit(void)
 #ifdef GRID_DEFAULT_GPU
   if ( world_rank == 0 ) {
     printf("AcceleratorHipInit: using default device \n");
-    printf("AcceleratorHipInit: assume user either uses a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
-    printf("AcceleratorHipInit: Configure options --enable-summit, --enable-select-gpu=no \n");
+    printf("AcceleratorHipInit: assume user or srun sets ROCR_VISIBLE_DEVICES and numa binding \n");
+    printf("AcceleratorHipInit: Configure options --enable-setdevice=no \n");
   }
+  int device = 0;
 #else
   if ( world_rank == 0 ) {
     printf("AcceleratorHipInit: rank %d setting device to node rank %d\n",world_rank,rank);
-    printf("AcceleratorHipInit: Configure options --enable-select-gpu=yes \n");
+    printf("AcceleratorHipInit: Configure options --enable-setdevice=yes \n");
   }
-  hipSetDevice(rank);
+  int device = rank;
 #endif
+  hipSetDevice(device);
+  hipStreamCreate(&copyStream);
+  const int len=64;
+  char busid[len];
+  if( rank == world_rank ) { 
+    hipDeviceGetPCIBusId(busid, len, device);
+    printf("local rank %d device %d bus id: %s\n", rank, device, busid);
+  }
   if ( world_rank == 0 )  printf("AcceleratorHipInit: ================================================\n");
 }
 #endif
@@ -172,12 +195,15 @@ void acceleratorInit(void)
 #ifdef GRID_SYCL
 
 cl::sycl::queue *theGridAccelerator;
+cl::sycl::queue *theCopyAccelerator;
 void acceleratorInit(void)
 {
   int nDevices = 1;
   cl::sycl::gpu_selector selector;
   cl::sycl::device selectedDevice { selector };
   theGridAccelerator = new sycl::queue (selectedDevice);
+  //  theCopyAccelerator = new sycl::queue (selectedDevice);
+  theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
 
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
   zeInit(0);

Grid/threads/Accelerator.h

@@ -95,6 +95,7 @@ void     acceleratorInit(void);
 //////////////////////////////////////////////
 
 #ifdef GRID_CUDA
+
 #include <cuda.h>
 
 #ifdef __CUDA_ARCH__
@@ -115,6 +116,14 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #endif
 } // CUDA specific
 
+inline void cuda_mem(void)
+{
+  size_t free_t,total_t,used_t;
+  cudaMemGetInfo(&free_t,&total_t);
+  used_t=total_t-free_t;
+  std::cout << " MemoryManager : GPU used "<<used_t<<" free "<<free_t<< " total "<<total_t<<std::endl;
+}
+
 #define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
   {									\
     int nt=acceleratorThreads();					\
@@ -221,6 +230,7 @@ inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes
   cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
 }
 inline void acceleratorCopySynchronise(void) { cudaStreamSynchronize(copyStream); };
+
 inline int  acceleratorIsCommunicable(void *ptr)
 {
   //  int uvm=0;
@@ -237,7 +247,6 @@ inline int  acceleratorIsCommunicable(void *ptr)
 //////////////////////////////////////////////
 // SyCL acceleration
 //////////////////////////////////////////////
-
 #ifdef GRID_SYCL
 NAMESPACE_END(Grid);
 #include <CL/sycl.hpp>
@@ -252,6 +261,7 @@ NAMESPACE_END(Grid);
 NAMESPACE_BEGIN(Grid);
 
 extern cl::sycl::queue *theGridAccelerator;
+extern cl::sycl::queue *theCopyAccelerator;
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define GRID_SIMT
@@ -279,7 +289,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
       cgh.parallel_for(					\
       cl::sycl::nd_range<3>(global,local), \
       [=] (cl::sycl::nd_item<3> item) /*mutable*/     \
-      [[intel::reqd_sub_group_size(8)]]	      \
+      [[intel::reqd_sub_group_size(16)]]	      \
       {						      \
       auto iter1    = item.get_global_id(0);	      \
       auto iter2    = item.get_global_id(1);	      \
@@ -288,19 +298,19 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
      });	   			              \
     });
 
-#define accelerator_barrier(dummy) theGridAccelerator->wait();
+#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {
-  theGridAccelerator->memcpy(to,from,bytes);
-}
-inline void acceleratorCopySynchronise(void) {  theGridAccelerator->wait(); }
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
+
+inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {  theCopyAccelerator->memcpy(to,from,bytes);}
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
+
 inline int  acceleratorIsCommunicable(void *ptr)
 {
 #if 0
@@ -328,10 +338,11 @@ NAMESPACE_BEGIN(Grid);
 #define accelerator        __host__ __device__
 #define accelerator_inline __host__ __device__ inline
 
+extern hipStream_t copyStream;
 /*These routines define mapping from thread grid to loop & vector lane indexing */
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
-  return hipThreadIdx_z; 
+  return hipThreadIdx_x; 
 #else
   return 0;
 #endif
@@ -345,19 +356,41 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
       { __VA_ARGS__;}							\
     };									\
     int nt=acceleratorThreads();					\
-    dim3 hip_threads(nt,1,nsimd);					\
-    dim3 hip_blocks ((num1+nt-1)/nt,num2,1);				\
-    hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
-		       0,0,						\
-		       num1,num2,nsimd,lambda);				\
+    dim3 hip_threads(nsimd, nt, 1);					 \
+    dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
+    if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
+      hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads,		\
+            0,0,						\
+            num1,num2,nsimd, lambda);				\
+    } else { \
+      hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
+            0,0,						\
+            num1,num2,nsimd, lambda);				\
+    } \
   }
 
+
 template<typename lambda>  __global__
+__launch_bounds__(64,1)
+void LambdaApply64(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
+{
+  // Following the same scheme as CUDA for now
+  uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
+  uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
+  uint64_t z = threadIdx.x;
+  if ( (x < numx) && (y<numy) && (z<numz) ) {
+    Lambda(x,y,z);
+  }
+}
+
+template<typename lambda>  __global__
+__launch_bounds__(1024,1)
 void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 {
-  uint64_t x = hipThreadIdx_x + hipBlockDim_x*hipBlockIdx_x;
-  uint64_t y = hipThreadIdx_y + hipBlockDim_y*hipBlockIdx_y;
-  uint64_t z = hipThreadIdx_z ;//+ hipBlockDim_z*hipBlockIdx_z;
+  // Following the same scheme as CUDA for now
+  uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
+  uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
+  uint64_t z = threadIdx.x;
   if ( (x < numx) && (y<numy) && (z<numz) ) {
     Lambda(x,y,z);
   }
@@ -402,10 +435,16 @@ inline void acceleratorFreeShared(void *ptr){ hipFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ hipFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
-inline void acceleratorCopySynchronise(void) {  }
+//inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
+//inline void acceleratorCopySynchronise(void) {  }
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(base,value,bytes);}
 
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
+{
+  hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToDevice,copyStream);
+}
+inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
+
 #endif
 
 //////////////////////////////////////////////
@@ -442,9 +481,10 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 #define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) thread_for2d(iter1,num1,iter2,num2,{ __VA_ARGS__ });
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
+
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes);}
 inline void acceleratorCopySynchronise(void) {};
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
@@ -471,23 +511,26 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
 inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 #endif
 
+//////////////////////////////////////////////
+// Fencing needed ONLY for SYCL
+//////////////////////////////////////////////
 
+#ifdef GRID_SYCL
+inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
+#else
+// Ordering within a stream guaranteed on Nvidia & AMD
+inline void acceleratorFenceComputeStream(void){ };
+#endif
 
 ///////////////////////////////////////////////////
 // Synchronise across local threads for divergence resynch
 ///////////////////////////////////////////////////
-accelerator_inline void acceleratorSynchronise(void) 
+accelerator_inline void acceleratorSynchronise(void)  // Only Nvidia needs 
 {
 #ifdef GRID_SIMT
 #ifdef GRID_CUDA
   __syncwarp();
 #endif
-#ifdef GRID_SYCL
-  //cl::sycl::detail::workGroupBarrier();
-#endif
-#ifdef GRID_HIP
-  __syncthreads();
-#endif
 #endif
   return;
 }

Grid/threads/Threads.h

@@ -72,3 +72,20 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define thread_region                                       DO_PRAGMA(omp parallel)
 #define thread_critical                                     DO_PRAGMA(omp critical)
 
+#ifdef GRID_OMP
+inline void thread_bcopy(void *from, void *to,size_t bytes)
+{
+  uint64_t *ufrom = (uint64_t *)from;
+  uint64_t *uto   = (uint64_t *)to;
+  assert(bytes%8==0);
+  uint64_t words=bytes/8;
+  thread_for(w,words,{
+      uto[w] = ufrom[w];
+  });
+}
+#else
+inline void thread_bcopy(void *from, void *to,size_t bytes)
+{
+  bcopy(from,to,bytes);
+}
+#endif

Grid/util/Coordinate.h

@@ -88,7 +88,7 @@ public:
 // Coordinate class, maxdims = 8 for now.
 ////////////////////////////////////////////////////////////////
 #define GRID_MAX_LATTICE_DIMENSION (8)
-#define GRID_MAX_SIMD              (16)
+#define GRID_MAX_SIMD              (32)
 
 static constexpr int MaxDims = GRID_MAX_LATTICE_DIMENSION;
 

Grid/util/Init.cc

@@ -167,6 +167,13 @@ void GridCmdOptionInt(std::string &str,int & val)
   return;
 }
 
+void GridCmdOptionFloat(std::string &str,float & val)
+{
+  std::stringstream ss(str);
+  ss>>val;
+  return;
+}
+
 
 void GridParseLayout(char **argv,int argc,
 		     Coordinate &latt_c,
@@ -527,6 +534,7 @@ void Grid_init(int *argc,char ***argv)
 void Grid_finalize(void)
 {
 #if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
+  MPI_Barrier(MPI_COMM_WORLD);
   MPI_Finalize();
   Grid_unquiesce_nodes();
 #endif

Grid/util/Init.h

@@ -57,6 +57,7 @@ void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec);
 template<class VectorInt>
 void GridCmdOptionIntVector(const std::string &str,VectorInt & vec);
 void GridCmdOptionInt(std::string &str,int & val);
+void GridCmdOptionFloat(std::string &str,float & val);
 
 
 void GridParseLayout(char **argv,int argc,

Grid/util/Sha.h

@@ -27,6 +27,7 @@
     /*  END LEGAL */
 extern "C" {
 #include <openssl/sha.h>
+#include <openssl/evp.h>
 }
 #ifdef USE_IPP
 #include "ipp.h"
@@ -70,10 +71,8 @@ public:
   static inline std::vector<unsigned char> sha256(const void *data,size_t bytes)
   {
     std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
-    SHA256_CTX sha256;
-    SHA256_Init  (&sha256);
-    SHA256_Update(&sha256, data,bytes);
-    SHA256_Final (&hash[0], &sha256);
+    auto digest = EVP_get_digestbyname("SHA256");
+    EVP_Digest(data, bytes, &hash[0], NULL, digest, NULL);
     return hash;
   }
   static inline std::vector<int> sha256_seeds(const std::string &s)

README.md

@@ -148,7 +148,7 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-mkl[=<path>]`: use Intel MKL for FFT (and LAPACK if enabled) routines. A UNIX prefix containing the library can be specified (optional).
 - `--enable-numa`: enable NUMA first touch optimisation
 - `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
-- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
+- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 64 bytes).
 - `--enable-comms=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.

benchmarks/Benchmark_IO.cc

@@ -137,7 +137,7 @@ int main (int argc, char ** argv)
 
   Eigen::MatrixXd mean(nVol, 4), stdDev(nVol, 4), rob(nVol, 4);
   Eigen::VectorXd avMean(4), avStdDev(4), avRob(4);
-  double          n = BENCH_IO_NPASS;
+  //  double          n = BENCH_IO_NPASS;
 
   stats(mean, stdDev, perf);
   stats(avMean, avStdDev, avPerf);
@@ -164,7 +164,7 @@ int main (int argc, char ** argv)
                 mean(volInd(l), gWrite), stdDev(volInd(l), gWrite));
   }
   MSG << std::endl;
-  MSG << "Robustness of individual results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
+  MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)" << std::endl;
   MSG << std::endl;
   grid_printf("%4s %12s %12s %12s %12s\n",
               "L", "std read", "std write", "Grid read", "Grid write");
@@ -185,7 +185,7 @@ int main (int argc, char ** argv)
               avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite),
               avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
   MSG << std::endl;
-  MSG << "Robustness of volume-averaged results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
+  MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)" << std::endl;
   MSG << std::endl;
   grid_printf("%12s %12s %12s %12s\n",
               "std read", "std write", "Grid read", "Grid write");

benchmarks/Benchmark_ITT.cc

@@ -142,7 +142,7 @@ public:
 	  //	  bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
 	}
 
-	int ncomm;
+	//	int ncomm;
 	double dbytes;
 
         for(int dir=0;dir<8;dir++) {
@@ -290,7 +290,7 @@ public:
       LatticeSU4 z(&Grid); z=Zero();
       LatticeSU4 x(&Grid); x=Zero();
       LatticeSU4 y(&Grid); y=Zero();
-      double a=2.0;
+      //      double a=2.0;
 
       uint64_t Nloop=NLOOP;
 

benchmarks/Benchmark_comms_host_device.cc

@@ -72,7 +72,7 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
   std::vector<double> t_time(Nloop);
-  time_statistics timestat;
+  //  time_statistics timestat;
 
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from host memory "<<std::endl;

benchmarks/Benchmark_dwf_fp32.cc

@@ -126,19 +126,10 @@ int main (int argc, char ** argv)
   // Naive wilson implementation
   ////////////////////////////////////
   // replicate across fifth dimension
-  LatticeGaugeFieldF Umu5d(FGrid);
-  std::vector<LatticeColourMatrixF> U(4,FGrid);
-  {
-    autoView( Umu5d_v, Umu5d, CpuWrite);
-    autoView( Umu_v  , Umu  , CpuRead);
-    for(int ss=0;ss<Umu.Grid()->oSites();ss++){
-      for(int s=0;s<Ls;s++){
-	Umu5d_v[Ls*ss+s] = Umu_v[ss];
-      }
-    }
-  }
+  //  LatticeGaugeFieldF Umu5d(FGrid);
+  std::vector<LatticeColourMatrixF> U(4,UGrid);
   for(int mu=0;mu<Nd;mu++){
-    U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
   }
   std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
@@ -147,10 +138,28 @@ int main (int argc, char ** argv)
     ref = Zero();
     for(int mu=0;mu<Nd;mu++){
 
-      tmp = U[mu]*Cshift(src,mu+1,1);
+      tmp = Cshift(src,mu+1,1);
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = U_v[ss]*tmp_v[Ls*ss+s];
+	  }
+	}
+      }
       ref=ref + tmp - Gamma(Gmu[mu])*tmp;
 
-      tmp =adj(U[mu])*src;
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	autoView( src_v, src    , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
+	  }
+	}
+      }
       tmp =Cshift(tmp,mu+1,-1);
       ref=ref + tmp + Gamma(Gmu[mu])*tmp;
     }
@@ -182,7 +191,7 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
   DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  int ncall =3000;
+  int ncall =300;
 
   if (1) {
     FGrid->Barrier();
@@ -242,16 +251,30 @@ int main (int argc, char ** argv)
     for(int mu=0;mu<Nd;mu++){
 
       //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
-      tmp = U[mu]*Cshift(src,mu+1,1);
+      tmp = Cshift(src,mu+1,1);
       {
 	autoView( ref_v, ref, CpuWrite);
 	autoView( tmp_v, tmp, CpuRead);
-	for(int i=0;i<ref_v.size();i++){
-	  ref_v[i]+= tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]; ;
+	autoView( U_v  , U[mu]  , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    int i=s+Ls*ss;
+	    ref_v[i]+= U_v[ss]*(tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]); ;
+	  }
 	}
       }
-
-      tmp =adj(U[mu])*src;
+      
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	autoView( src_v, src    , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
+	  }
+	}
+      }
+      //      tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu+1,-1);
       {
 	autoView( ref_v, ref, CpuWrite);