First cut at faster GPU slice sum via atomics

Feature/fermtoprop

.gitignore

@@ -88,6 +88,7 @@ Thumbs.db
 # build directory #
 ###################
 build*/*
+Documentation/_build
 
 # IDE related files #
 #####################

.travis.yml

@@ -1,56 +0,0 @@
-language: cpp
-
-cache:
-  directories:
-    - clang
-
-matrix:
-  include:
-    - os:        osx
-      osx_image: xcode8.3
-      compiler: clang
-      
-before_install:
-    - export GRIDDIR=`pwd`
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]] && [ ! -e clang/bin ]; then wget $CLANG_LINK; tar -xf `basename $CLANG_LINK`; mkdir clang; mv clang+*/* clang/; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export PATH="${GRIDDIR}/clang/bin:${PATH}"; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc openssl; fi
-    
-install:
-    - export CWD=`pwd`
-    - echo $CWD
-    - export CC=$CC$VERSION
-    - export CXX=$CXX$VERSION
-    - echo $PATH
-    - which autoconf
-    - autoconf  --version
-    - which automake
-    - automake  --version
-    - which $CC
-    - $CC  --version
-    - which $CXX
-    - $CXX --version
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export EXTRACONF='--with-openssl=/usr/local/opt/openssl'; fi
-    
-script:
-    - ./bootstrap.sh
-    - mkdir build
-    - cd build
-    - mkdir lime
-    - cd lime
-    - mkdir build
-    - cd build
-    - wget http://usqcd-software.github.io/downloads/c-lime/lime-1.3.2.tar.gz
-    - tar xf lime-1.3.2.tar.gz
-    - cd lime-1.3.2
-    - ./configure --prefix=$CWD/build/lime/install
-    - make -j4
-    - make install
-    - cd $CWD/build
-    - ../configure --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
-    - make -j4 
-    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
-    - make check

Grid/DisableWarnings.h

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

Grid/GridStd.h

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

Grid/Grid_Eigen_Dense.h

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

Grid/algorithms/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;
@@ -442,6 +442,8 @@ public:
     for(int p=0; p<geom.npoint; p++)
       points[p] = geom.points_dagger[p];
 
+    auto points_p = &points[0];
+
     RealD* dag_factor_p = &dag_factor[0];
 
     accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
@@ -452,8 +454,8 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int p=0;p<geom_v.npoint;p++){
-        int point = points[p];
+      for(int p=0;p<npoint;p++){
+        int point = points_p[p];
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);
 
@@ -708,6 +710,8 @@ public:
     for(int p=0; p<npoint; p++)
       points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
 
+    auto points_p = &points[0];
+
     Vector<Aview> AcceleratorViewContainer;
     for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead));
     Aview *Aview_p = & AcceleratorViewContainer[0];
@@ -728,7 +732,7 @@ public:
         StencilEntry *SE;
 
         for(int p=0;p<npoint;p++){
-          int point = points[p];
+          int point = points_p[p];
           SE=st_v.GetEntry(ptype,point,ss);
 
           if(SE->_is_local) {
@@ -754,7 +758,7 @@ public:
         StencilEntry *SE;
 
         for(int p=0;p<npoint;p++){
-          int point = points[p];
+          int point = points_p[p];
           SE=st_v.GetEntry(ptype,point,ss);
 
           if(SE->_is_local) {

Grid/algorithms/FFT.h

@@ -136,7 +136,7 @@ public:
     flops=0;
     usec =0;
     Coordinate layout(Nd,1);
-    sgrid = new GridCartesian(dimensions,layout,processors);
+    sgrid = new GridCartesian(dimensions,layout,processors,*grid);
   };
     
   ~FFT ( void)  {
@@ -182,7 +182,7 @@ public:
     pencil_gd[dim] = G*processors[dim];
       
     // Pencil global vol LxLxGxLxL per node
-    GridCartesian pencil_g(pencil_gd,layout,processors);
+    GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid);
       
     // Construct pencils
     typedef typename vobj::scalar_object sobj;

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
   }
 };
@@ -530,6 +531,16 @@ public:
 template<class Field> class LinearFunction {
 public:
   virtual void operator() (const Field &in, Field &out) = 0;
+
+  virtual void operator() (const std::vector<Field> &in, std::vector<Field> &out)
+  {
+    assert(in.size() == out.size());
+
+    for (unsigned int i = 0; i < in.size(); ++i)
+    {
+      (*this)(in[i], out[i]);
+    }
+  }
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
@@ -575,6 +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) 
@@ -588,6 +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/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; };
 };
 
@@ -54,15 +57,24 @@ class DeflatedGuesser: public LinearFunction<Field> {
 private:
   const std::vector<Field> &evec;
   const std::vector<RealD> &eval;
+  const unsigned int       N;
 
 public:
+  using LinearFunction<Field>::operator();
 
-  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
+  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
+  : DeflatedGuesser(_evec, _eval, _evec.size())
+  {}
+
+  DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N)
+  : evec(_evec), eval(_eval), N(_N)
+  {
+    assert(evec.size()==eval.size());
+    assert(N <= evec.size());
+  } 
 
   virtual void operator()(const Field &src,Field &guess) {
     guess = Zero();
-    assert(evec.size()==eval.size());
-    auto N = evec.size();
     for (int i=0;i<N;i++) {
       const Field& tmp = evec[i];
       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
@@ -79,6 +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)
@@ -100,7 +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/algorithms/iterative/SchurRedBlack.h

@@ -132,6 +132,31 @@ namespace Grid {
       (*this)(_Matrix,in,out,guess);
     }
 
+    void RedBlackSource(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &src_o) 
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      Field tmp(grid);
+      int nblock = in.size();
+      for(int b=0;b<nblock;b++){
+	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+      }
+    }
+    // James can write his own deflated guesser
+    // with optimised code for the inner products
+    //    RedBlackSolveSplitGrid();
+    //    RedBlackSolve(_Matrix,src_o,sol_o); 
+
+    void RedBlackSolution(Matrix &_Matrix, const std::vector<Field> &in, const std::vector<Field> &sol_o, std::vector<Field> &out)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      Field tmp(grid);
+      int nblock = in.size();
+      for(int b=0;b<nblock;b++) {
+	pickCheckerboard(Even,tmp,in[b]);
+	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]);
+      }
+    }
+
     template<class Guesser>
     void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess) 
     {
@@ -150,24 +175,29 @@ namespace Grid {
       ////////////////////////////////////////////////
       // Prepare RedBlack source
       ////////////////////////////////////////////////
-      for(int b=0;b<nblock;b++){
-	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
-      }
+      RedBlackSource(_Matrix,in,src_o);
+	//      for(int b=0;b<nblock;b++){
+	//	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+	//      }
+      
       ////////////////////////////////////////////////
       // Make the guesses
       ////////////////////////////////////////////////
       if ( subGuess ) guess_save.resize(nblock,grid);
 
-      for(int b=0;b<nblock;b++){
-        if(useSolnAsInitGuess) {
+      
+      if(useSolnAsInitGuess) {
+        for(int b=0;b<nblock;b++){
           pickCheckerboard(Odd, sol_o[b], out[b]);
-        } else {
-          guess(src_o[b],sol_o[b]); 
         }
+      } else {
+        guess(src_o, sol_o); 
+      }
 
-	if ( subGuess ) { 
-	  guess_save[b] = sol_o[b];
-	}
+	    if ( subGuess ) { 
+        for(int b=0;b<nblock;b++){
+          guess_save[b] = sol_o[b];
+        }
       }
       //////////////////////////////////////////////////////////////
       // Call the block solver

Grid/allocator/MemoryManager.cc

@@ -9,14 +9,30 @@ NAMESPACE_BEGIN(Grid);
 #define AccSmall (3)
 #define Shared   (4)
 #define SharedSmall (5)
+#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
+  
 }
 
 //////////////////////////////////////////////////////////////////////
@@ -24,86 +40,114 @@ void MemoryManager::PrintBytes(void)
 //////////////////////////////////////////////////////////////////////
 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, 8, 8, 16, 8, 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 +159,6 @@ void MemoryManager::Init(void)
 
   char * str;
   int Nc;
-  int NcS;
   
   str= getenv("GRID_ALLOC_NCACHE_LARGE");
   if ( str ) {
@@ -181,13 +224,13 @@ 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]);  
+  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
@@ -211,6 +254,7 @@ 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 +263,7 @@ 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;
 }
@@ -228,13 +273,13 @@ 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]);
+  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
@@ -243,6 +288,7 @@ void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncach
   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

@@ -36,6 +36,11 @@ NAMESPACE_BEGIN(Grid);
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
+
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
 // Advise the LatticeAccelerator class
@@ -82,17 +87,19 @@ private:
   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);
@@ -112,6 +119,8 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
  
  private:
 #ifndef GRID_UVM
@@ -169,6 +178,8 @@ 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_base.cc

@@ -33,6 +33,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+bool Stencil_force_mpi = true;
+
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////

Grid/communicator/Communicator_base.h

@@ -35,6 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+extern bool Stencil_force_mpi ;
+
 class CartesianCommunicator : public SharedMemory {
 
 public:    

Grid/communicator/Communicator_mpi3.cc

@@ -370,7 +370,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( gfrom ==MPI_UNDEFINED) {
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
     tag= dir+from*32;
     ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
     assert(ierr==0);
@@ -378,22 +378,31 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
     off_node_bytes+=bytes;
   }
 
-  if ( gdest == MPI_UNDEFINED ) {
+  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
     tag= dir+_processor*32;
     ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
+  } else {
+    // TODO : make a OMP loop on CPU, call threaded bcopy
+    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+    assert(shm!=NULL);
+    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
+    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-    this->StencilSendToRecvFromComplete(list,dir);
-  }
+  //  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/communicator/SharedMemoryMPI.cc

@@ -35,6 +35,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #endif
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
+#endif
+#ifdef GRID_SYCl
+
 #endif
 
 NAMESPACE_BEGIN(Grid); 
@@ -70,6 +73,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   WorldNodes = WorldSize/WorldShmSize;
   assert( (WorldNodes * WorldShmSize) == WorldSize );
 
+
   // FIXME: Check all WorldShmSize are the same ?
 
   /////////////////////////////////////////////////////////////////////
@@ -446,7 +450,47 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#if defined(GRID_CUDA) ||defined(GRID_HIP)
+#if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
+
+//if defined(GRID_SYCL)
+#if 0
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  void * ShmCommBuf ; 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the pointer array for shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Each MPI rank should allocate our own buffer
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  ShmCommBuf = acceleratorAllocDevice(bytes);
+
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+
+  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+
+  SharedMemoryZero(ShmCommBuf,bytes);
+
+  assert(WorldShmSize == 1);
+  for(int r=0;r<WorldShmSize;r++){
+    WorldShmCommBufs[r] = ShmCommBuf;
+  }
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+}
+#endif
+
+#if defined(GRID_CUDA) ||defined(GRID_HIP) ||defined(GRID_SYCL)  
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -470,18 +514,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   ShmCommBuf = acceleratorAllocDevice(bytes);
-
   if (ShmCommBuf == (void *)NULL ) {
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
-  //  if ( WorldRank == 0 ){
-  if ( 1 ){
+  if ( WorldRank == 0 ){
     std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
-
+  std::cout<< "Setting up IPC"<<std::endl;
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -491,6 +533,29 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
     //////////////////////////////////////////////////
+    void * thisBuf = ShmCommBuf;
+    if(!Stencil_force_mpi) {
+#ifdef GRID_SYCL_LEVEL_ZERO_IPC
+    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+      
+    ze_ipc_mem_handle_t ihandle;
+    clone_mem_t handle;
+
+    if ( r==WorldShmRank ) { 
+      auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
+      if ( err != ZE_RESULT_SUCCESS ) {
+	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	exit(EXIT_FAILURE);
+      } else {
+	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+      }
+      memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
+      handle.pid = getpid();
+    }
+#endif
 #ifdef GRID_CUDA
     cudaIpcMemHandle_t handle;
     if ( r==WorldShmRank ) { 
@@ -511,6 +576,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       }
     }
 #endif
+
     //////////////////////////////////////////////////
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
@@ -526,7 +592,35 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     ///////////////////////////////////////////////////////////////
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
-    void * thisBuf = ShmCommBuf;
+
+#ifdef GRID_SYCL_LEVEL_ZERO_IPC
+    if ( r!=WorldShmRank ) {
+      thisBuf = nullptr;
+      std::cout<<"mapping seeking remote pid/fd "
+	       <<handle.pid<<"/"
+	       <<handle.fd<<std::endl;
+
+      int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
+      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
+      //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
+      int myfd  = syscall(438,pidfd,handle.fd,0);
+
+      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
+      
+      memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
+
+      auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
+      if ( err != ZE_RESULT_SUCCESS ) {
+	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	exit(EXIT_FAILURE);
+      } else {
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl;
+      }
+      assert(thisBuf!=nullptr);
+    }
+#endif
 #ifdef GRID_CUDA
     if ( r!=WorldShmRank ) { 
       auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
@@ -548,6 +642,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     ///////////////////////////////////////////////////////////////
     // Save a copy of the device buffers
     ///////////////////////////////////////////////////////////////
+    }
     WorldShmCommBufs[r] = thisBuf;
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
@@ -557,6 +652,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
+#endif
+
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
@@ -727,16 +824,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 /////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
-#ifdef GRID_CUDA
-  cudaMemset(dest,0,bytes);
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  acceleratorMemSet(dest,0,bytes);
 #else
   bzero(dest,bytes);
 #endif
 }
 void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
 {
-#ifdef GRID_CUDA
-  cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  acceleratorCopyToDevice(src,dest,bytes);
 #else   
   bcopy(src,dest,bytes);
 #endif
@@ -800,7 +897,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
 #endif
 
-  SharedMemoryTest();
+  //SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier

Grid/cshift/Cshift_mpi.h

@@ -122,8 +122,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  cshiftVector<vobj> send_buf(buffer_size);
-  cshiftVector<vobj> recv_buf(buffer_size);
+  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
+  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
     
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
@@ -198,8 +198,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd);
-  std::vector<cshiftVector<scalar_object> >  recv_buf_extract(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
  
@@ -294,8 +294,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  cshiftVector<vobj> send_buf_v(buffer_size);
-  cshiftVector<vobj> recv_buf_v(buffer_size);
+  static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
+  static cshiftVector<vobj> recv_buf_v; recv_buf_v.resize(buffer_size);
   vobj *send_buf;
   vobj *recv_buf;
   {
@@ -381,8 +381,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd);
-  std::vector<cshiftVector<scalar_object> >  recv_buf_extract(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
   {

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_arith.h

@@ -225,7 +225,7 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
   autoView( x_v , x, AcceleratorRead);
   autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-    auto tmp = a*x_v(ss)+y_v(ss);
+    auto tmp = a*coalescedRead(x_v[ss])+coalescedRead(y_v[ss]);
     coalescedWrite(ret_v[ss],tmp);
   });
 }

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
@@ -122,7 +129,7 @@ public:
     
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
+    auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -145,7 +152,7 @@ public:
 
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
+    auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -167,7 +174,7 @@ public:
     this->checkerboard=cb;
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
+    auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -238,7 +245,7 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWrite) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
@@ -281,7 +288,7 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWriteDiscard);
+    auto me =   View(AcceleratorWrite);
     auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
@@ -296,7 +303,7 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
+    auto me =   View(AcceleratorWrite);
     auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));

Grid/lattice/Lattice_basis.h

@@ -125,7 +125,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 
 	for(int k=k0; k<k1; ++k){
 	  auto tmp = coalescedRead(Bp[ss*nrot+j]);
-	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss]));
+	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_vp[k][sss]));
 	}
       });
 
@@ -134,7 +134,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	int jj  =j0+j;
 	int ss =sj/nrot;
 	int sss=ss+s;
-	coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j]));
+	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
       });
   }
 #endif

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];
@@ -361,6 +388,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   // But easily avoided by using double precision fields
   ///////////////////////////////////////////////////////
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_object::scalar_type scalar_type;
   GridBase  *grid = Data.Grid();
   assert(grid!=NULL);
 
@@ -419,20 +447,19 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   }
   
   // sum over nodes.
-  sobj gsum;
   for(int t=0;t<fd;t++){
     int pt = t/ld; // processor plane
     int lt = t%ld;
     if ( pt == grid->_processor_coor[orthogdim] ) {
-      gsum=lsSum[lt];
+      result[t]=lsSum[lt];
     } else {
-      gsum=Zero();
+      result[t]=Zero();
     }
 
-    grid->GlobalSum(gsum);
-
-    result[t]=gsum;
   }
+  scalar_type * ptr = (scalar_type *) &result[0];
+  int words = fd*sizeof(sobj)/sizeof(scalar_type);
+  grid->GlobalSumVector(ptr, words);
 }
 
 template<class vobj>

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_slice_gpu.h

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

Grid/lattice/Lattice_transfer.h

@@ -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
@@ -364,16 +434,22 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   autoView( coarseData_ , coarseData, AcceleratorWrite);
   autoView( fineData_   , fineData, AcceleratorRead);
 
+  auto coarseData_p = &coarseData_[0];
+  auto fineData_p = &fineData_[0];
+  
   Coordinate fine_rdimensions = fine->_rdimensions;
   Coordinate coarse_rdimensions = coarse->_rdimensions;
+
+  vobj zz = Zero();
   
   accelerator_for(sc,coarse->oSites(),1,{
 
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
-      coarseData_[sc]=Zero();
 
+      vobj cd = zz;
+      
       for(int sb=0;sb<blockVol;sb++){
 
 	int sf;
@@ -383,9 +459,11 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
 	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
 
-	coarseData_[sc]=coarseData_[sc]+fineData_[sf];
+	cd=cd+fineData_p[sf];
       }
 
+      coarseData_p[sc] = cd;
+
     });
   return;
 }

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
@@ -128,7 +128,7 @@ inline void MachineCharacteristics(FieldMetaData &header)
   std::time_t t = std::time(nullptr);
   std::tm tm_ = *std::localtime(&t);
   std::ostringstream oss; 
-  //      oss << std::put_time(&tm_, "%c %Z");
+  oss << std::put_time(&tm_, "%c %Z");
   header.creation_date = oss.str();
   header.archive_date  = header.creation_date;
 
@@ -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,
@@ -205,22 +210,33 @@ public:
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
+  // Preferred interface
+  template<class GaugeStats=PeriodicGaugeStatistics>
+  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
+					std::string file, 
+					std::string ens_label = std::string("DWF"),
+					std::string ens_id = std::string("UKQCD"),
+					unsigned int sequence_number = 1)
+  {
+    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)
+					int bits32,
+					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     = "UKQCD";
-    header.ensemble_label  = "DWF";
+    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;
@@ -232,12 +248,16 @@ public:
     GaugeStats Stats; Stats(Umu,header);
     MachineCharacteristics(header);
 
-	uint64_t offset;
+    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);
@@ -245,8 +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);
@@ -278,8 +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");
@@ -319,7 +345,7 @@ public:
 
     GridBase *grid = parallel.Grid();
 
-	uint64_t offset = readHeader(file,grid,header);
+    uint64_t offset = readHeader(file,grid,header);
 
     FieldMetaData clone(header);
 

Grid/perfmon/PerfCount.cc

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

Grid/perfmon/PerfCount.h

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

Grid/perfmon/Timer.h

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

Grid/perfmon/Tracing.h

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

Grid/pugixml/pugixml.cc

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

Grid/qcd/QCD.h

@@ -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);
@@ -115,9 +116,9 @@ typedef WilsonFermion<WilsonImplR> WilsonFermionR;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
-typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
-typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
-typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
+//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
+//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
+//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
 
 typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
@@ -137,62 +138,93 @@ 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;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
 
-typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
-typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
-typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
+//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
+//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
+//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
 
 typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
-typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
-typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
+//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
+//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
+//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
 
 typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
-typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
-typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
+//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
+//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
+//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
 
 typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
-typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
-typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
+//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
+//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
+//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
 
 typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
-typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
-typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
+//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
+//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
+//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
 
 // Ls vectorised
 typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
@@ -235,49 +267,49 @@ typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 
-typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
-typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
-typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
+//typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
+//typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
+//typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
 
 typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
-typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
-typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
+//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
+//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
+//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
 
 typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
-typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
-typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
+//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
+//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
+//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
 
 typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
-typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
-typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
+//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
+//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
+//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
 
 typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
-typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
-typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
+//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
+//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
+//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
 
 typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 
-typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
-typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
-typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
+//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
+//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
+//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
 
 typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
@@ -291,12 +323,6 @@ typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
 typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
 typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
 
-#ifndef GRID_CUDA
-typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR;
-typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF;
-typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD;
-#endif
-
 NAMESPACE_END(Grid);
 
 ////////////////////

Grid/qcd/action/fermion/FermionOperatorImpl.h

@@ -183,7 +183,8 @@ NAMESPACE_CHECK(ImplStaggered);
 /////////////////////////////////////////////////////////////////////////////
 // Single flavour one component spinors with colour index. 5d vec
 /////////////////////////////////////////////////////////////////////////////
-#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
-NAMESPACE_CHECK(ImplStaggered5dVec);  
+// Deprecate Vec5d
+//#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
+//NAMESPACE_CHECK(ImplStaggered5dVec);  
 
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -327,8 +327,8 @@ typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffReal> Gparit
 typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffReal> GparityWilsonImplF;  // Float
 typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffReal> GparityWilsonImplD;  // Double
  
-typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
-typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
-typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
+//typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
+//typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
+//typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/StaggeredImpl.h

@@ -72,19 +72,23 @@ public:
     
   StaggeredImpl(const ImplParams &p = ImplParams()) : Params(p){};
       
-  static accelerator_inline void multLink(SiteSpinor &phi,
+  template<class _Spinor>
+  static accelerator_inline void multLink(_Spinor &phi,
 		       const SiteDoubledGaugeField &U,
-		       const SiteSpinor &chi,
+		       const _Spinor &chi,
 		       int mu)
   {
-    mult(&phi(), &U(mu), &chi());
+    auto UU = coalescedRead(U(mu));
+    mult(&phi(), &UU, &chi());
   }
-  static accelerator_inline void multLinkAdd(SiteSpinor &phi,
+  template<class _Spinor>
+  static accelerator_inline void multLinkAdd(_Spinor &phi,
 			  const SiteDoubledGaugeField &U,
-			  const SiteSpinor &chi,
+			  const _Spinor &chi,
 			  int mu)
   {
-    mac(&phi(), &U(mu), &chi());
+    auto UU = coalescedRead(U(mu));
+    mac(&phi(), &UU, &chi());
   }
       
   template <class ref>

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

@@ -68,11 +68,12 @@ public:
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  template<class _SiteHalfSpinor, class _SiteSpinor>
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
-    _SiteHalfSpinor tmp;
-    projector::Proj(tmp,in,mu,dag);
-    vstream(buf[o],tmp);
+  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
+    typedef decltype(coalescedRead(buf)) sobj;
+    sobj sp;
+    auto sin = coalescedRead(in);
+    projector::Proj(sp,sin,mu,dag);
+    coalescedWrite(buf,sp);
   }
 
   /*****************************************************/
@@ -82,13 +83,18 @@ public:
 				   const SiteHalfSpinor * __restrict__ vp0,
 				   const SiteHalfSpinor * __restrict__ vp1,
 				   Integer type,Integer o) const {
+#ifdef GRID_SIMT
+    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+#else
     SiteHalfSpinor tmp1;
     SiteHalfSpinor tmp2;
     exchange(tmp1,tmp2,vp0[o],vp1[o],type);
     vstream(mp[2*o  ],tmp1);
     vstream(mp[2*o+1],tmp2);
+#endif
   }
 
+
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
@@ -105,6 +111,28 @@ public:
 					   const SiteSpinor * __restrict__ in,
 					   Integer j,Integer k, Integer m,Integer type) const
   {
+#ifdef GRID_SIMT
+    typedef SiteSpinor vobj;
+    typedef SiteHalfSpinor hvobj;
+    typedef decltype(coalescedRead(*in))    sobj;
+    typedef decltype(coalescedRead(*out0)) hsobj;
+
+    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];  // 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);  // spin project the result0
+    projector::Proj(psb,sb,mu,dag);  // spin project the result1
+    coalescedWrite(out0[j],psa);
+    coalescedWrite(out1[j],psb);
+#else
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
     projector::Proj(temp1,in[k],mu,dag);
@@ -112,6 +140,7 @@ public:
     exchange(temp3,temp4,temp1,temp2,type);
     vstream(out0[j],temp3);
     vstream(out1[j],temp4);
+#endif
   }
 
   /*****************************************************/
@@ -121,6 +150,7 @@ public:
 
 };
 
+#if 0
 template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
 class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
 				typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
@@ -149,13 +179,23 @@ public:
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  template<class _SiteHalfSpinor, class _SiteSpinor>
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
-    _SiteHalfSpinor hsp;
+  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
+    SiteHalfSpinor hsp;
     SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
     projector::Proj(hsp,in,mu,dag);
     precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
   }
+  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
+#ifdef GRID_SIMT
+    typedef decltype(coalescedRead(buf)) sobj;
+    sobj sp;
+    auto sin = coalescedRead(in);
+    projector::Proj(sp,sin,mu,dag);
+    coalescedWrite(buf,sp);
+#else
+    projector::Proj(buf,in,mu,dag);
+#endif
+  }
 
   /*****************************************************/
   /* Exchange includes precision change if mpi data is not same */
@@ -203,6 +243,7 @@ public:
   accelerator_inline bool DecompressionStep(void) const { return true; }
 
 };
+#endif
 
 #define DECLARE_PROJ(Projector,Compressor,spProj)			\
   class Projector {							\
@@ -253,33 +294,8 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  double timer0;
-  double timer1;
-  double timer2;
-  double timer3;
-  double timer4;
-  double timer5;
-  double timer6;
-  uint64_t callsi;
-  void ZeroCountersi(void)
-  {
-    timer0=0;
-    timer1=0;
-    timer2=0;
-    timer3=0;
-    timer4=0;
-    timer5=0;
-    timer6=0;
-    callsi=0;
-  }
-  void Reporti(int calls)
-  {
-    if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
-    if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate)   " <<timer1/calls <<std::endl;
-    if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge )   " <<timer2/calls <<std::endl;
-    if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
-    if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
-  }
+  void ZeroCountersi(void)  {  }
+  void Reporti(int calls)  {  }
 
   std::vector<int> surface_list;
 
@@ -321,26 +337,18 @@ public:
   {
     std::vector<std::vector<CommsRequest_t> > reqs;
     this->HaloExchangeOptGather(source,compress);
-    double t1=usecond();
     // Asynchronous MPI calls multidirectional, Isend etc...
     // Non-overlapped directions within a thread. Asynchronous calls except MPI3, threaded up to comm threads ways.
     this->Communicate();
-    double t2=usecond(); timer1 += t2-t1;
     this->CommsMerge(compress);
-    double t3=usecond(); timer2 += t3-t2;
     this->CommsMergeSHM(compress);
-    double t4=usecond(); timer3 += t4-t3;
   }
   
   template <class compressor>
   void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress) 
   {
     this->Prepare();
-    double t0=usecond();
     this->HaloGatherOpt(source,compress);
-    double t1=usecond();
-    timer0 += t1-t0;
-    callsi++;
   }
 
   template <class compressor>
@@ -352,12 +360,9 @@ public:
     typedef typename compressor::SiteHalfSpinor     SiteHalfSpinor;
     typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
 
-    this->mpi3synctime_g-=usecond();
     this->_grid->StencilBarrier();
-    this->mpi3synctime_g+=usecond();
 
     assert(source.Grid()==this->_grid);
-    this->halogtime-=usecond();
     
     this->u_comm_offset=0;
       
@@ -393,7 +398,6 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
-    this->halogtime+=usecond();
     accelerator_barrier();
   }
 

Grid/qcd/action/fermion/WilsonImpl.h

@@ -184,18 +184,22 @@ public:
       mat = TraceIndex<SpinIndex>(P); 
     }
       
-    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
+    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds)
+    {
       for (int mu = 0; mu < Nd; mu++)
       mat[mu] = PeekIndex<LorentzIndex>(Uds, mu);
     }
 
-
-  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
-      
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu)
+  {
+#undef USE_OLD_INSERT_FORCE    
     int Ls=Btilde.Grid()->_fdimensions[0];
+    autoView( mat_v , mat, AcceleratorWrite);
+#ifdef USE_OLD_INSERT_FORCE    
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
     {
+      const int Nsimd = SiteSpinor::Nsimd();
       autoView( tmp_v , tmp, AcceleratorWrite);
       autoView( Btilde_v , Btilde, AcceleratorRead);
       autoView( Atilde_v , Atilde, AcceleratorRead);
@@ -208,6 +212,29 @@ public:
 	});
     }
     PokeIndex<LorentzIndex>(mat,tmp,mu);
+#else
+    {
+      const int Nsimd = SiteSpinor::Nsimd();
+      autoView( Btilde_v , Btilde, AcceleratorRead);
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,mat.Grid()->oSites(),Nsimd,{
+	  int sU=sss;
+  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
+  	  ColorMatrixType sum;
+	  zeroit(sum);  
+	  for(int s=0;s<Ls;s++){
+	    int sF = s+Ls*sU;
+  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
+  	      auto bb = coalescedRead(Btilde_v[sF]()(spn) ); //color vector
+  	      auto aa = coalescedRead(Atilde_v[sF]()(spn) );
+	      auto op = outerProduct(bb,aa);
+  	      sum = sum + op;
+	    }
+	  }
+  	  coalescedWrite(mat_v[sU](mu)(), sum);
+      });
+    }
+#endif    
   }
 };
 
@@ -216,17 +243,17 @@ typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
 
-typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
-typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
-typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
+//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
+//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
+//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
 
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
 
-typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
-typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
-typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
+//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
+//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
+//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
  
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float

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
   ////////////////////////////////////////////////
@@ -880,17 +887,29 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
+  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;
     }
   }
+  else if ( curr_type == Current::Tadpole ) {
+    auto b=this->_b;
+    auto c=this->_c;
+    if ( b == 1 && c == 0 ) {
+      sign = -1.0;    
+    }
+    else {
+      std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
+      assert(b==1 && c==0);
+    }
+  }
 
   for(int s=0;s<Ls;s++){
 
     int sp = (s+1)%Ls;
-    int sr = Ls-1-s;
-    int srp= (sr+1)%Ls;
+    //    int sr = Ls-1-s;
+    //    int srp= (sr+1)%Ls;
 
     // Mobius parameters
     auto b=this->bs[s];
@@ -907,7 +926,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     tmp    = Cshift(tmp,mu,1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu);
-    tmp    = G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
+    tmp    = sign*G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
     tmp    = where((lcoor>=tmin),tmp,zz); // Mask the time 
     L_Q    = where((lcoor<=tmax),tmp,zz); // Position of current complicated
 
@@ -922,7 +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/StaggeredKernelsAsm.h

@@ -680,7 +680,8 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilView &st,
   gauge2 =(uint64_t)&UU[sU]( Z );				\
   gauge3 =(uint64_t)&UU[sU]( T ); 
   
-
+#undef STAG_VEC5D
+#ifdef STAG_VEC5D
   // This is the single precision 5th direction vectorised kernel
 #include <Grid/simd/Intel512single.h>
 template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
@@ -790,7 +791,7 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView
 #endif
 }
    
-   
+#endif   
 
 
 #define PERMUTE_DIR3 __asm__ (	\

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

@@ -32,25 +32,50 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-#define LOAD_CHI(b)		\
+#ifdef GRID_SIMT
+
+#define LOAD_CHI(ptype,b)			\
+  const SiteSpinor & ref (b[offset]);				\
+  Chi_0=coalescedReadPermute<ptype>(ref()()(0),perm,lane);	\
+  Chi_1=coalescedReadPermute<ptype>(ref()()(1),perm,lane);	\
+  Chi_2=coalescedReadPermute<ptype>(ref()()(2),perm,lane);
+
+#define LOAD_CHI_COMMS(b)		\
   const SiteSpinor & ref (b[offset]);	\
-    Chi_0=ref()()(0);\
-    Chi_1=ref()()(1);\
-    Chi_2=ref()()(2);
+  Chi_0=coalescedRead(ref()()(0),lane);	\
+  Chi_1=coalescedRead(ref()()(1),lane);	\
+  Chi_2=coalescedRead(ref()()(2),lane);
+
+#define PERMUTE_DIR(dir)	;
+#else
+#define LOAD_CHI(ptype,b)      LOAD_CHI_COMMS(b)
+
+#define LOAD_CHI_COMMS(b)		\
+  const SiteSpinor & ref (b[offset]);	\
+  Chi_0=ref()()(0);			\
+  Chi_1=ref()()(1);			\
+  Chi_2=ref()()(2);
+
+#define PERMUTE_DIR(dir)			\
+  permute##dir(Chi_0,Chi_0);			\
+  permute##dir(Chi_1,Chi_1);			\
+  permute##dir(Chi_2,Chi_2);
+
+#endif
 
 
 // To splat or not to splat depends on the implementation
 #define MULT(A,UChi)				\
   auto & ref(U[sU](A));			\
-   Impl::loadLinkElement(U_00,ref()(0,0));      \
-   Impl::loadLinkElement(U_10,ref()(1,0));      \
-   Impl::loadLinkElement(U_20,ref()(2,0));      \
-   Impl::loadLinkElement(U_01,ref()(0,1));      \
-   Impl::loadLinkElement(U_11,ref()(1,1));      \
-   Impl::loadLinkElement(U_21,ref()(2,1));      \
-   Impl::loadLinkElement(U_02,ref()(0,2));     \
-   Impl::loadLinkElement(U_12,ref()(1,2));     \
-   Impl::loadLinkElement(U_22,ref()(2,2));     \
+    U_00=coalescedRead(ref()(0,0),lane);				\
+    U_10=coalescedRead(ref()(1,0),lane);				\
+    U_20=coalescedRead(ref()(2,0),lane);				\
+    U_01=coalescedRead(ref()(0,1),lane);				\
+    U_11=coalescedRead(ref()(1,1),lane);				\
+    U_21=coalescedRead(ref()(2,1),lane);				\
+    U_02=coalescedRead(ref()(0,2),lane);				\
+    U_12=coalescedRead(ref()(1,2),lane);				\
+    U_22=coalescedRead(ref()(2,2),lane);				\
     UChi ## _0  = U_00*Chi_0;	       \
     UChi ## _1  = U_10*Chi_0;\
     UChi ## _2  = U_20*Chi_0;\
@@ -63,15 +88,15 @@ NAMESPACE_BEGIN(Grid);
 
 #define MULT_ADD(U,A,UChi)			\
   auto & ref(U[sU](A));			\
-   Impl::loadLinkElement(U_00,ref()(0,0));      \
-   Impl::loadLinkElement(U_10,ref()(1,0));      \
-   Impl::loadLinkElement(U_20,ref()(2,0));      \
-   Impl::loadLinkElement(U_01,ref()(0,1));      \
-   Impl::loadLinkElement(U_11,ref()(1,1));      \
-   Impl::loadLinkElement(U_21,ref()(2,1));      \
-   Impl::loadLinkElement(U_02,ref()(0,2));     \
-   Impl::loadLinkElement(U_12,ref()(1,2));     \
-   Impl::loadLinkElement(U_22,ref()(2,2));     \
+    U_00=coalescedRead(ref()(0,0),lane);				\
+    U_10=coalescedRead(ref()(1,0),lane);				\
+    U_20=coalescedRead(ref()(2,0),lane);				\
+    U_01=coalescedRead(ref()(0,1),lane);				\
+    U_11=coalescedRead(ref()(1,1),lane);				\
+    U_21=coalescedRead(ref()(2,1),lane);				\
+    U_02=coalescedRead(ref()(0,2),lane);				\
+    U_12=coalescedRead(ref()(1,2),lane);				\
+    U_22=coalescedRead(ref()(2,2),lane);				\
     UChi ## _0 += U_00*Chi_0;	       \
     UChi ## _1 += U_10*Chi_0;\
     UChi ## _2 += U_20*Chi_0;\
@@ -83,24 +108,18 @@ NAMESPACE_BEGIN(Grid);
     UChi ## _2 += U_22*Chi_2;
 
 
-#define PERMUTE_DIR(dir)			\
-  permute##dir(Chi_0,Chi_0);			\
-  permute##dir(Chi_1,Chi_1);			\
-  permute##dir(Chi_2,Chi_2);
-
-
 #define HAND_STENCIL_LEG_BASE(Dir,Perm,skew)	\
   SE=st.GetEntry(ptype,Dir+skew,sF);	\
   offset = SE->_offset;			\
   local  = SE->_is_local;		\
   perm   = SE->_permute;		\
   if ( local ) {						\
-    LOAD_CHI(in);					\
+    LOAD_CHI(Perm,in);						\
     if ( perm) {						\
       PERMUTE_DIR(Perm);					\
     }								\
   } else {							\
-    LOAD_CHI(buf);						\
+    LOAD_CHI_COMMS(buf);					\
   }								
 
 #define HAND_STENCIL_LEG_BEGIN(Dir,Perm,skew,even)		\
@@ -116,19 +135,18 @@ NAMESPACE_BEGIN(Grid);
   }
 
 
-
 #define HAND_STENCIL_LEG_INT(U,Dir,Perm,skew,even)	\
   SE=st.GetEntry(ptype,Dir+skew,sF);			\
   offset = SE->_offset;					\
   local  = SE->_is_local;				\
   perm   = SE->_permute;				\
   if ( local ) {					\
-    LOAD_CHI(in);				\
+    LOAD_CHI(Perm,in);					\
     if ( perm) {					\
       PERMUTE_DIR(Perm);				\
     }							\
   } else if ( st.same_node[Dir] ) {			\
-    LOAD_CHI(buf);					\
+    LOAD_CHI_COMMS(buf);				\
   }							\
   if (local || st.same_node[Dir] ) {		\
     MULT_ADD(U,Dir,even);				\
@@ -140,10 +158,32 @@ NAMESPACE_BEGIN(Grid);
   local  = SE->_is_local;				\
   if ((!local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
-    { LOAD_CHI(buf);	  }					\
+    { LOAD_CHI_COMMS(buf);	  }				\
     { MULT_ADD(U,Dir,even); }					\
   }								
 
+#define HAND_DECLARATIONS(Simd) \
+  Simd even_0;			\
+  Simd even_1;			\
+  Simd even_2;			\
+  Simd odd_0;			\
+  Simd odd_1;			\
+  Simd odd_2;		        \
+		      		\
+  Simd Chi_0;			\
+  Simd Chi_1;			\
+  Simd Chi_2;			\
+				\
+  Simd U_00;			\
+  Simd U_10;			\
+  Simd U_20;			\
+  Simd U_01;			\
+  Simd U_11;			\
+  Simd U_21;			\
+  Simd U_02;			\
+  Simd U_12;			\
+  Simd U_22;			
+  
 
 template <class Impl>
 template <int Naik> accelerator_inline
@@ -155,28 +195,14 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
 
-  Simd Chi_0;    // two spinor; 6 regs
-  Simd Chi_1;
-  Simd Chi_2;
-  
-  Simd U_00;  // two rows of U matrix
-  Simd U_10;
-  Simd U_20;  
-  Simd U_01;
-  Simd U_11;
-  Simd U_21;  // 2 reg left.
-  Simd U_02;
-  Simd U_12;
-  Simd U_22; 
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  SiteSpinor result;
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
+  calcSiteSpinor result;
   int offset,local,perm, ptype;
 
   StencilEntry *SE;
@@ -215,7 +241,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
       result()()(1) = even_1 + odd_1;
       result()()(2) = even_2 + odd_2;
     }
-    vstream(out[sF],result);
+    coalescedWrite(out[sF],result);
   }
 }
 
@@ -230,28 +256,13 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  Simd Chi_0;    // two spinor; 6 regs
-  Simd Chi_1;
-  Simd Chi_2;
-  
-  Simd U_00;  // two rows of U matrix
-  Simd U_10;
-  Simd U_20;  
-  Simd U_01;
-  Simd U_11;
-  Simd U_21;  // 2 reg left.
-  Simd U_02;
-  Simd U_12;
-  Simd U_22; 
-
-  SiteSpinor result;
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
+  calcSiteSpinor result;
   int offset, ptype, local, perm;
 
   StencilEntry *SE;
@@ -261,8 +272,8 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
   //    int sF=s+LLs*sU;
   {
 
-    even_0 = Zero();    even_1 = Zero();    even_2 = Zero();
-     odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
+    zeroit(even_0);    zeroit(even_1);    zeroit(even_2);
+    zeroit(odd_0);    zeroit(odd_1);    zeroit(odd_2);
 
     skew = 0;
     HAND_STENCIL_LEG_INT(U,Xp,3,skew,even);  
@@ -294,7 +305,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
       result()()(1) = even_1 + odd_1;
       result()()(2) = even_2 + odd_2;
     }
-    vstream(out[sF],result);
+    coalescedWrite(out[sF],result);
   }
 }
 
@@ -309,28 +320,13 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  Simd Chi_0;    // two spinor; 6 regs
-  Simd Chi_1;
-  Simd Chi_2;
-  
-  Simd U_00;  // two rows of U matrix
-  Simd U_10;
-  Simd U_20;  
-  Simd U_01;
-  Simd U_11;
-  Simd U_21;  // 2 reg left.
-  Simd U_02;
-  Simd U_12;
-  Simd U_22; 
-
-  SiteSpinor result;
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
+  calcSiteSpinor result;
   int offset, ptype, local;
 
   StencilEntry *SE;
@@ -340,8 +336,8 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
   //    int sF=s+LLs*sU;
   {
 
-    even_0 = Zero();    even_1 = Zero();    even_2 = Zero();
-     odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
+    zeroit(even_0);    zeroit(even_1);    zeroit(even_2);
+    zeroit(odd_0);    zeroit(odd_1);    zeroit(odd_2);
     int nmu=0;
     skew = 0;
     HAND_STENCIL_LEG_EXT(U,Xp,3,skew,even);  
@@ -374,7 +370,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
 	result()()(1) = even_1 + odd_1;
 	result()()(2) = even_2 + odd_2;
       }
-      out[sF] = out[sF] + result;
+      coalescedWrite(out[sF] , out(sF)+ result);
     }
   }
 }
@@ -397,6 +393,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
 						     const FermionFieldView &in, FermionFieldView &out, int dag); \
 */
 #undef LOAD_CHI
+#undef HAND_DECLARATIONS
 
 NAMESPACE_END(Grid);
 

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

@@ -35,39 +35,32 @@ NAMESPACE_BEGIN(Grid);
 #define GENERIC_STENCIL_LEG(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
-      chi_p = χ						\
-      permute(chi,  in[SE->_offset], ptype);			\
-    } else {							\
-      chi_p = &in[SE->_offset];					\
-    }								\
+    int perm= SE->_permute;						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
   } else {							\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  multLink(Uchi, U[sU], *chi_p, Dir);			
+  acceleratorSynchronise();					\
+  multLink(Uchi, U[sU], chi, Dir);			
 
 #define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
-      chi_p = χ						\
-      permute(chi,  in[SE->_offset], ptype);			\
-    } else {							\
-      chi_p = &in[SE->_offset];					\
-    }								\
+    int perm= SE->_permute;						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
   } else if ( st.same_node[Dir] ) {				\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);                 \
   }								\
   if (SE->_is_local || st.same_node[Dir] ) {			\
-    multLink(Uchi, U[sU], *chi_p, Dir);				\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 #define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
-    chi_p = &buf[SE->_offset];					\
-    multLink(Uchi, U[sU], *chi_p, Dir);				\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 template <class Impl>
@@ -84,12 +77,14 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
 					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
-  const SiteSpinor *chi_p;
-  SiteSpinor chi;
-  SiteSpinor Uchi;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //
@@ -118,7 +113,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
     if ( dag ) { 
       Uchi = - Uchi;
     } 
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -130,13 +125,16 @@ template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 						SiteSpinor *buf, int sF, int sU, 
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
-  const SiteSpinor *chi_p;
-  SiteSpinor chi;
-  SiteSpinor Uchi;
+						const FermionFieldView &in, FermionFieldView &out,int dag)
+{
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //    int sF=LLs*sU+s;
@@ -165,7 +163,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
     if ( dag ) {
       Uchi = - Uchi;
     }
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -178,14 +176,17 @@ template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 						SiteSpinor *buf, int sF, int sU,
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
-  const SiteSpinor *chi_p;
-  //  SiteSpinor chi;
-  SiteSpinor Uchi;
+						const FermionFieldView &in, FermionFieldView &out,int dag)
+{
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int nmu=0;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //    int sF=LLs*sU+s;
@@ -211,11 +212,12 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
     GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
     }
-    if ( nmu ) { 
-      if ( dag ) { 
-	out[sF] = out[sF] - Uchi;
+    if ( nmu ) {
+      auto _out = coalescedRead(out[sF],lane);
+      if ( dag ) {
+	coalescedWrite(out[sF], _out-Uchi,lane);
       } else { 
-	out[sF] = out[sF] + Uchi;
+	coalescedWrite(out[sF], _out+Uchi,lane);
       }
     }
   }
@@ -261,6 +263,8 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo,
   GridBase *FGrid=in.Grid();  
   GridBase *UGrid=U.Grid();  
   typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
   autoView( UUU_v , UUU, AcceleratorRead);
   autoView( U_v   ,   U, AcceleratorRead);
   autoView( in_v  ,  in, AcceleratorRead);
@@ -301,6 +305,8 @@ void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo,
   GridBase *FGrid=in.Grid();  
   GridBase *UGrid=U.Grid();  
   typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
   autoView( UUU_v ,   U, AcceleratorRead);
   autoView( U_v   ,   U, AcceleratorRead);
   autoView( in_v  ,  in, AcceleratorRead);

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/WilsonKernelsAsmA64FX.h

@@ -73,17 +73,17 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -102,17 +102,17 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -131,17 +131,17 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 
@@ -165,17 +165,17 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -194,17 +194,17 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -223,17 +223,17 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 
@@ -280,17 +280,17 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -309,17 +309,17 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -338,17 +338,17 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 /////////////////////////////////////////////////////////////////
@@ -371,17 +371,17 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -400,17 +400,17 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -429,17 +429,17 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+// template<> void
+// WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
 
 
 

Grid/qcd/action/fermion/implementation/WilsonKernelsAsmAvx512.h

@@ -74,15 +74,15 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
-template<> void 
-WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -97,15 +97,15 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
-template<> void 
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 
 #undef INTERIOR_AND_EXTERIOR
@@ -121,15 +121,15 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
-template<> void 
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
       
 /////////////////////////////////////////////////////////////////
 // XYZT vectorised, dag Kernel, single
@@ -148,15 +148,15 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
-template<> void 
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -171,15 +171,15 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
-template<> void 
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -194,15 +194,15 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
-template<> void 
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-				    
-template<> void 
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
 #undef MAYBEPERM
 #undef MULT_2SPIN
@@ -228,14 +228,14 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSite(StencilView &st, DoubledGaugeF
 							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -249,14 +249,14 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteInt(StencilView &st, DoubledGau
 							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -273,15 +273,15 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteExt(StencilView &st, DoubledGau
 							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
-template<> void 
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-				    
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
 /////////////////////////////////////////////////////////////////
 // Ls vectorised, dag Kernel, single
@@ -299,14 +299,14 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDag(StencilView &st, DoubledGau
 							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -320,14 +320,14 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagInt(StencilView &st, Doubled
 							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -341,14 +341,14 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, Doubled
 							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #endif  // VEC 5D
 
@@ -392,14 +392,14 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -413,14 +413,14 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -434,14 +434,14 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
       
-template<> void 
-WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
       
 /////////////////////////////////////////////////////////////////
 // XYZT vectorised, dag Kernel, single
@@ -459,14 +459,14 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldVi
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -480,14 +480,14 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -501,14 +501,14 @@ WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFiel
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
-template<> void 
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
 #undef MAYBEPERM
 #undef MULT_2SPIN
@@ -533,14 +533,14 @@ WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSite(StencilView &st, DoubledGaugeF
 							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -554,14 +554,14 @@ WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteInt(StencilView &st, DoubledGau
 							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -577,14 +577,14 @@ WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteExt(StencilView &st, DoubledGau
 							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
-template<> void 
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 				    
 /////////////////////////////////////////////////////////////////
 // Ls vectorised, dag Kernel, single
@@ -602,14 +602,14 @@ WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDag(StencilView &st, DoubledGau
 							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
@@ -623,14 +623,14 @@ WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDagInt(StencilView &st, Doubled
 							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
@@ -645,14 +645,14 @@ WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDagExt(StencilView &st, Doubled
 							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-template<> void 
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//template<> void
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #endif  // VEC 5D
 

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/GparityWilsonImplDF/CayleyFermion5DInstantiationGparityWilsonImplDF.cc

@@ -1 +0,0 @@
-../CayleyFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/ContinuedFractionFermion5DInstantiationGparityWilsonImplDF.cc

@@ -1 +0,0 @@
-../ContinuedFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/DomainWallEOFAFermionInstantiationGparityWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/MobiusEOFAFermionInstantiationGparityWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/PartialFractionFermion5DInstantiationGparityWilsonImplDF.cc

@@ -1 +0,0 @@
-../PartialFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/WilsonCloverFermionInstantiationGparityWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/WilsonFermion5DInstantiationGparityWilsonImplDF.cc

@@ -1 +0,0 @@
-../WilsonFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/WilsonFermionInstantiationGparityWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/WilsonKernelsInstantiationGparityWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/WilsonTMFermionInstantiationGparityWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/impl.h

@@ -1 +0,0 @@
-#define IMPLEMENTATION GparityWilsonImplDF

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/CayleyFermion5DInstantiationGparityWilsonImplFH.cc

@@ -1 +0,0 @@
-../CayleyFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/ContinuedFractionFermion5DInstantiationGparityWilsonImplFH.cc

@@ -1 +0,0 @@
-../ContinuedFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/DomainWallEOFAFermionInstantiationGparityWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/MobiusEOFAFermionInstantiationGparityWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/PartialFractionFermion5DInstantiationGparityWilsonImplFH.cc

@@ -1 +0,0 @@
-../PartialFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/WilsonCloverFermionInstantiationGparityWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/WilsonFermion5DInstantiationGparityWilsonImplFH.cc

@@ -1 +0,0 @@
-../WilsonFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/WilsonFermionInstantiationGparityWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/WilsonKernelsInstantiationGparityWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/WilsonTMFermionInstantiationGparityWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/GparityWilsonImplFH/impl.h

@@ -1 +0,0 @@
-#define IMPLEMENTATION GparityWilsonImplFH

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);