Performance improve for Tesseract

Changing thread block order and adding launch_bounds

.gitignore

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

.travis.yml

@@ -1,61 +0,0 @@
-language: cpp
-
-cache:
-  directories:
-    - clang
-
-matrix:
-  include:
-    - os:        osx
-      osx_image: xcode8.3
-      compiler: clang
-      env: PREC=single
-    - os:        osx
-      osx_image: xcode8.3
-      compiler: clang
-      env: PREC=double
-      
-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 $CWD/build
-    - ../configure --enable-precision=$PREC --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

@@ -37,7 +37,9 @@ directory
 #endif
 
  //disables and intel compiler specific warning (in json.hpp)
+#ifdef __ICC
 #pragma warning disable 488  
+#endif
 
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp

Grid/GridStd.h

@@ -28,4 +28,7 @@
 ///////////////////
 #include "Config.h"
 
+#ifdef TOFU
+#undef GRID_COMMS_THREADS
+#endif
 #endif /* GRID_STD_H */

Grid/Grid_Eigen_Dense.h

@@ -34,6 +34,12 @@
 #define __SYCL__REDEFINE__
 #endif
 
+/* HIP save and restore compile environment*/
+#ifdef GRID_HIP
+#pragma push
+#pragma push_macro("__HIP_DEVICE_COMPILE__")
+#endif
+#define EIGEN_NO_HIP
 
 #include <Grid/Eigen/Dense>
 #include <Grid/Eigen/unsupported/CXX11/Tensor>
@@ -42,7 +48,7 @@
 #ifdef __NVCC__REDEFINE__
 #pragma pop_macro("__CUDACC__")
 #pragma pop_macro("__NVCC__")
-#pragma pop_macro("GRID_SIMT")
+#pragma pop_macro("__CUDA_ARCH__")
 #pragma pop
 #endif
 
@@ -52,6 +58,12 @@
 #pragma pop
 #endif
 
+/*HIP restore*/
+#ifdef __HIP__REDEFINE__
+#pragma pop_macro("__HIP_DEVICE_COMPILE__")
+#pragma pop
+#endif
+
 #if defined __GNUC__
 #pragma GCC diagnostic pop
 #endif

Grid/Makefile.am

@@ -21,6 +21,7 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
+
 all: version-cache Version.h
 
 version-cache:
@@ -53,6 +54,19 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
+extra_sources+=$(WILS_FERMION_FILES)
+extra_sources+=$(STAG_FERMION_FILES)
+if BUILD_ZMOBIUS
+  extra_sources+=$(ZWILS_FERMION_FILES)
+endif
+if BUILD_GPARITY
+  extra_sources+=$(GP_FERMION_FILES)
+endif
+if BUILD_FERMION_REPS
+  extra_sources+=$(ADJ_FERMION_FILES)
+  extra_sources+=$(TWOIND_FERMION_FILES)
+endif
+
 lib_LIBRARIES = libGrid.a
 
 CCFILES += $(extra_sources)

Grid/algorithms/CoarsenedMatrix.h

@@ -31,6 +31,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_ALGORITHM_COARSENED_MATRIX_H
 #define  GRID_ALGORITHM_COARSENED_MATRIX_H
 
+#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
 
 NAMESPACE_BEGIN(Grid);
 
@@ -59,12 +60,14 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
 class Geometry {
 public:
   int npoint;
+  int base;
   std::vector<int> directions   ;
   std::vector<int> displacements;
+  std::vector<int> points_dagger;
 
   Geometry(int _d)  {
     
-    int base = (_d==5) ? 1:0;
+    base = (_d==5) ? 1:0;
 
     // make coarse grid stencil for 4d , not 5d
     if ( _d==5 ) _d=4;
@@ -72,16 +75,51 @@ public:
     npoint = 2*_d+1;
     directions.resize(npoint);
     displacements.resize(npoint);
+    points_dagger.resize(npoint);
     for(int d=0;d<_d;d++){
       directions[d   ] = d+base;
       directions[d+_d] = d+base;
       displacements[d  ] = +1;
       displacements[d+_d]= -1;
+      points_dagger[d   ] = d+_d;
+      points_dagger[d+_d] = d;
     }
     directions   [2*_d]=0;
     displacements[2*_d]=0;
+    points_dagger[2*_d]=2*_d;
   }
 
+  int point(int dir, int disp) {
+    assert(disp == -1 || disp == 0 || disp == 1);
+    assert(base+0 <= dir && dir < base+4);
+
+    // directions faster index = new indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  1  2  3  0  1  2  3  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  2  3  4  1  2  3  4  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+
+    // displacements faster index = old indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  0  1  1  2  2  3  3  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  1  2  2  3  3  4  4  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+
+    if(dir == 0 and disp == 0)
+      return 8;
+    else // New indexing
+      return (1 - disp) / 2 * 4 + dir - base;
+    // else // Old indexing
+    //   return (4 * (dir - base) + 1 - disp) / 2;
+  }
 };
   
 template<class Fobj,class CComplex,int nbasis>
@@ -258,7 +296,7 @@ public:
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>
-class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+class CoarsenedMatrix : public CheckerBoardedSparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
 public:
     
   typedef iVector<CComplex,nbasis >           siteVector;
@@ -268,33 +306,59 @@ public:
   typedef iMatrix<CComplex,nbasis >  Cobj;
   typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
   typedef Lattice<Fobj >        FineField;
+  typedef CoarseVector FermionField;
+
+  // enrich interface, use default implementation as in FermionOperator ///////
+  void Dminus(CoarseVector const& in, CoarseVector& out) { out = in; }
+  void DminusDag(CoarseVector const& in, CoarseVector& out) { out = in; }
+  void ImportPhysicalFermionSource(CoarseVector const& input, CoarseVector& imported) { imported = input; }
+  void ImportUnphysicalFermion(CoarseVector const& input, CoarseVector& imported) { imported = input; }
+  void ExportPhysicalFermionSolution(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
+  void ExportPhysicalFermionSource(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
 
   ////////////////////
   // Data members
   ////////////////////
   Geometry         geom;
   GridBase *       _grid; 
+  GridBase*        _cbgrid;
   int hermitian;
 
   CartesianStencil<siteVector,siteVector,int> Stencil; 
+  CartesianStencil<siteVector,siteVector,int> StencilEven;
+  CartesianStencil<siteVector,siteVector,int> StencilOdd;
 
   std::vector<CoarseMatrix> A;
+  std::vector<CoarseMatrix> Aeven;
+  std::vector<CoarseMatrix> Aodd;
+
+  CoarseMatrix AselfInv;
+  CoarseMatrix AselfInvEven;
+  CoarseMatrix AselfInvOdd;
+
+  Vector<RealD> dag_factor;
 
   ///////////////////////
   // Interface
   ///////////////////////
   GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know
+  GridBase * RedBlackGrid()     { return _cbgrid; };
+
+  int ConstEE() { return 0; }
 
   void M (const CoarseVector &in, CoarseVector &out)
   {
     conformable(_grid,in.Grid());
     conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
 
     SimpleCompressor<siteVector> compressor;
 
     Stencil.HaloExchange(in,compressor);
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
     Vector<Aview> AcceleratorViewContainer;
@@ -316,14 +380,14 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom.npoint;point++){
+      for(int point=0;point<npoint;point++){
 
-	SE=Stencil.GetEntry(ptype,point,ss);
+	SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
 	if(SE->_is_local) { 
 	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
 	} else {
-	  nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
+	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
 	}
 	acceleratorSynchronise();
 
@@ -344,12 +408,74 @@ public:
       return M(in,out);
     } else {
       // corresponds to Galerkin coarsening
-      CoarseVector tmp(Grid());
-      G5C(tmp, in); 
-      M(tmp, out);
-      G5C(out, out);
+      return MdagNonHermitian(in, out);
     }
   };
+
+  void MdagNonHermitian(const CoarseVector &in, CoarseVector &out)
+  {
+    conformable(_grid,in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    SimpleCompressor<siteVector> compressor;
+
+    Stencil.HaloExchange(in,compressor);
+    autoView( in_v , in, AcceleratorRead);
+    autoView( out_v , out, AcceleratorWrite);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+
+    Vector<Aview> AcceleratorViewContainer;
+
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    Aview *Aview_p = & AcceleratorViewContainer[0];
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    int osites=Grid()->oSites();
+
+    Vector<int> points(geom.npoint, 0);
+    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, {
+      int ss = sss/nbasis;
+      int b  = sss%nbasis;
+      calcComplex res = Zero();
+      calcVector nbr;
+      int ptype;
+      StencilEntry *SE;
+
+      for(int p=0;p<npoint;p++){
+        int point = points_p[p];
+
+	SE=Stencil_v.GetEntry(ptype,point,ss);
+
+	if(SE->_is_local) {
+	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+	} else {
+	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
+	}
+	acceleratorSynchronise();
+
+	for(int bb=0;bb<nbasis;bb++) {
+	  res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+	}
+      }
+      coalescedWrite(out_v[ss](b),res);
+      });
+
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+  }
+
   void MdirComms(const CoarseVector &in)
   {
     SimpleCompressor<siteVector> compressor;
@@ -359,6 +485,7 @@ public:
   {
     conformable(_grid,in.Grid());
     conformable(_grid,out.Grid());
+    out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
     Vector<Aview> AcceleratorViewContainer;
@@ -367,6 +494,7 @@ public:
 
     autoView( out_v , out, AcceleratorWrite);
     autoView( in_v  , in, AcceleratorRead);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
 
     const int Nsimd = CComplex::Nsimd();
     typedef decltype(coalescedRead(in_v[0])) calcVector;
@@ -380,12 +508,12 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      SE=Stencil.GetEntry(ptype,point,ss);
+      SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
       if(SE->_is_local) { 
 	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
       } else {
-	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
+	nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
       }
       acceleratorSynchronise();
 
@@ -413,34 +541,7 @@ public:
 
     this->MdirComms(in);
 
-    int ndim = in.Grid()->Nd();
-
-    //////////////
-    // 4D action like wilson
-    // 0+ => 0 
-    // 0- => 1
-    // 1+ => 2 
-    // 1- => 3
-    // etc..
-    //////////////
-    // 5D action like DWF
-    // 1+ => 0 
-    // 1- => 1
-    // 2+ => 2 
-    // 2- => 3
-    // etc..
-    auto point = [dir, disp, ndim](){
-      if(dir == 0 and disp == 0)
-	return 8;
-      else if ( ndim==4 ) { 
-	return (4 * dir + 1 - disp) / 2;
-      } else { 
-	return (4 * (dir-1) + 1 - disp) / 2;
-      }
-    }();
-
-    MdirCalc(in,out,point);
-
+    MdirCalc(in,out,geom.point(dir,disp));
   };
 
   void Mdiag(const CoarseVector &in, CoarseVector &out)
@@ -449,23 +550,298 @@ public:
     MdirCalc(in, out, point); // No comms
   };
 
+  void Mooee(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerNo, InverseNo);
+  }
+
+  void MooeeInv(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerNo, InverseYes);
+  }
+
+  void MooeeDag(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerYes, InverseNo);
+  }
+
+  void MooeeInvDag(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerYes, InverseYes);
+  }
+
+  void Meooe(const CoarseVector &in, CoarseVector &out) {
+    if(in.Checkerboard() == Odd) {
+      DhopEO(in, out, DaggerNo);
+    } else {
+      DhopOE(in, out, DaggerNo);
+    }
+  }
+
+  void MeooeDag(const CoarseVector &in, CoarseVector &out) {
+    if(in.Checkerboard() == Odd) {
+      DhopEO(in, out, DaggerYes);
+    } else {
+      DhopOE(in, out, DaggerYes);
+    }
+  }
+
+  void Dhop(const CoarseVector &in, CoarseVector &out, int dag) {
+    conformable(in.Grid(), _grid); // verifies full grid
+    conformable(in.Grid(), out.Grid());
+
+    out.Checkerboard() = in.Checkerboard();
+
+    DhopInternal(Stencil, A, in, out, dag);
+  }
+
+  void DhopOE(const CoarseVector &in, CoarseVector &out, int dag) {
+    conformable(in.Grid(), _cbgrid);    // verifies half grid
+    conformable(in.Grid(), out.Grid()); // drops the cb check
+
+    assert(in.Checkerboard() == Even);
+    out.Checkerboard() = Odd;
+
+    DhopInternal(StencilEven, Aodd, in, out, dag);
+  }
+
+  void DhopEO(const CoarseVector &in, CoarseVector &out, int dag) {
+    conformable(in.Grid(), _cbgrid);    // verifies half grid
+    conformable(in.Grid(), out.Grid()); // drops the cb check
+
+    assert(in.Checkerboard() == Odd);
+    out.Checkerboard() = Even;
+
+    DhopInternal(StencilOdd, Aeven, in, out, dag);
+  }
+
+  void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) {
+    out.Checkerboard() = in.Checkerboard();
+    assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
+
+    CoarseMatrix *Aself = nullptr;
+    if(in.Grid()->_isCheckerBoarded) {
+      if(in.Checkerboard() == Odd) {
+        Aself = (inv) ? &AselfInvOdd : &Aodd[geom.npoint-1];
+        DselfInternal(StencilOdd, *Aself, in, out, dag);
+      } else {
+        Aself = (inv) ? &AselfInvEven : &Aeven[geom.npoint-1];
+        DselfInternal(StencilEven, *Aself, in, out, dag);
+      }
+    } else {
+      Aself = (inv) ? &AselfInv : &A[geom.npoint-1];
+      DselfInternal(Stencil, *Aself, in, out, dag);
+    }
+    assert(Aself != nullptr);
+  }
+
+  void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
+                       const CoarseVector &in, CoarseVector &out, int dag) {
+    int point = geom.npoint-1;
+    autoView( out_v, out, AcceleratorWrite);
+    autoView( in_v,  in,  AcceleratorRead);
+    autoView( st_v,  st,  AcceleratorRead);
+    autoView( a_v,   a,   AcceleratorRead);
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    RealD* dag_factor_p = &dag_factor[0];
+
+    if(dag) {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        SE=st_v.GetEntry(ptype,point,ss);
+
+        if(SE->_is_local) {
+          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+        } else {
+          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+        }
+        acceleratorSynchronise();
+
+        for(int bb=0;bb<nbasis;bb++) {
+          res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(a_v[ss](b,bb))*nbr(bb);
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    } else {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        SE=st_v.GetEntry(ptype,point,ss);
+
+        if(SE->_is_local) {
+          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+        } else {
+          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+        }
+        acceleratorSynchronise();
+
+        for(int bb=0;bb<nbasis;bb++) {
+          res = res + coalescedRead(a_v[ss](b,bb))*nbr(bb);
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    }
+  }
+
+  void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
+                    const CoarseVector &in, CoarseVector &out, int dag) {
+    SimpleCompressor<siteVector> compressor;
+
+    st.HaloExchange(in,compressor);
+    autoView( in_v,  in,  AcceleratorRead);
+    autoView( out_v, out, AcceleratorWrite);
+    autoView( st_v , st,  AcceleratorRead);
+    typedef LatticeView<Cobj> Aview;
+
+    // determine in what order we need the points
+    int npoint = geom.npoint-1;
+    Vector<int> points(npoint, 0);
+    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];
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    RealD* dag_factor_p = &dag_factor[0];
+
+    if(dag) {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        for(int p=0;p<npoint;p++){
+          int point = points_p[p];
+          SE=st_v.GetEntry(ptype,point,ss);
+
+          if(SE->_is_local) {
+            nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+          } else {
+            nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+          }
+          acceleratorSynchronise();
+
+          for(int bb=0;bb<nbasis;bb++) {
+            res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+          }
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    } else {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        for(int p=0;p<npoint;p++){
+          int point = points_p[p];
+          SE=st_v.GetEntry(ptype,point,ss);
+
+          if(SE->_is_local) {
+            nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+          } else {
+            nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+          }
+          acceleratorSynchronise();
+
+          for(int bb=0;bb<nbasis;bb++) {
+            res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+          }
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    }
+
+    for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose();
+  }
   
   CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
-
     _grid(&CoarseGrid),
+    _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
     Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-      A(geom.npoint,&CoarseGrid)
+    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    A(geom.npoint,&CoarseGrid),
+    Aeven(geom.npoint,_cbgrid),
+    Aodd(geom.npoint,_cbgrid),
+    AselfInv(&CoarseGrid),
+    AselfInvEven(_cbgrid),
+    AselfInvOdd(_cbgrid),
+    dag_factor(nbasis*nbasis)
   {
+    fillFactor();
   };
 
+  CoarsenedMatrix(GridCartesian &CoarseGrid, GridRedBlackCartesian &CoarseRBGrid, int hermitian_=0) 	:
+
+    _grid(&CoarseGrid),
+    _cbgrid(&CoarseRBGrid),
+    geom(CoarseGrid._ndimension),
+    hermitian(hermitian_),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    A(geom.npoint,&CoarseGrid),
+    Aeven(geom.npoint,&CoarseRBGrid),
+    Aodd(geom.npoint,&CoarseRBGrid),
+    AselfInv(&CoarseGrid),
+    AselfInvEven(&CoarseRBGrid),
+    AselfInvOdd(&CoarseRBGrid),
+    dag_factor(nbasis*nbasis)
+  {
+    fillFactor();
+  };
+
+  void fillFactor() {
+    Eigen::MatrixXd dag_factor_eigen = Eigen::MatrixXd::Ones(nbasis, nbasis);
+    if(!hermitian) {
+      const int nb = nbasis/2;
+      dag_factor_eigen.block(0,nb,nb,nb) *= -1.0;
+      dag_factor_eigen.block(nb,0,nb,nb) *= -1.0;
+    }
+
+    // GPU readable prefactor
+    thread_for(i, nbasis*nbasis, {
+      int j = i/nbasis;
+      int k = i%nbasis;
+      dag_factor[i] = dag_factor_eigen(j, k);
+    });
+  }
+
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
 		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
   {
     typedef Lattice<typename Fobj::tensor_reduced> FineComplexField;
     typedef typename Fobj::scalar_type scalar_type;
 
+    std::cout << GridLogMessage<< "CoarsenMatrix "<< std::endl;
+
     FineComplexField one(FineGrid); one=scalar_type(1.0,0.0);
     FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0);
 
@@ -496,11 +872,13 @@ public:
 
     CoarseScalar InnerProd(Grid()); 
 
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog "<< std::endl;
     // Orthogonalise the subblocks over the basis
     blockOrthogonalise(InnerProd,Subspace.subspace);
 
     // Compute the matrix elements of linop between this orthonormal
     // set of vectors.
+    std::cout << GridLogMessage<< "CoarsenMatrix masks "<< std::endl;
     int self_stencil=-1;
     for(int p=0;p<geom.npoint;p++)
     { 
@@ -539,7 +917,7 @@ public:
 
       phi=Subspace.subspace[i];
 
-      //      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl;
+      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl;
       linop.OpDirAll(phi,Mphi_p);
       linop.OpDiag  (phi,Mphi_p[geom.npoint-1]);
 
@@ -568,6 +946,18 @@ public:
 	    autoView( A_self  , A[self_stencil], AcceleratorWrite);
 
 	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
+	    if ( hermitian && (disp==-1) ) {
+	      for(int pp=0;pp<geom.npoint;pp++){// Find the opposite link and set <j|A|i> = <i|A|j>*
+		int dirp   = geom.directions[pp];
+		int dispp  = geom.displacements[pp];
+		if ( (dirp==dir) && (dispp==1) ){
+		  auto sft = conjugate(Cshift(oZProj,dir,1));
+		  autoView( sft_v    ,  sft  , AcceleratorWrite);
+		  autoView( A_pp     ,  A[pp], AcceleratorWrite);
+		  accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_pp[ss](i,j),sft_v(ss)); });
+		}
+	      }
+	    }
 
 	  }
 	}
@@ -606,28 +996,54 @@ public:
     }
     if(hermitian) {
       std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
-      ForceHermitian();
-    }
     }
 
-  void ForceHermitian(void) {
-    CoarseMatrix Diff  (Grid());
-    for(int p=0;p<geom.npoint;p++){
-      int dir   = geom.directions[p];
-      int disp  = geom.displacements[p];
-      if(disp==-1) {
-	// Find the opposite link
-	for(int pp=0;pp<geom.npoint;pp++){
-	  int dirp   = geom.directions[pp];
-	  int dispp  = geom.displacements[pp];
-	  if ( (dirp==dir) && (dispp==1) ){
-	    //	    Diff = adj(Cshift(A[p],dir,1)) - A[pp]; 
-	    //	    std::cout << GridLogMessage<<" Replacing stencil leg "<<pp<<" with leg "<<p<< " diff "<<norm2(Diff) <<std::endl;
-	    A[pp] = adj(Cshift(A[p],dir,1));
-	  }
+    InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl;
+    FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl;
   }
+
+  void InvertSelfStencilLink() {
+    std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl;
+    int localVolume = Grid()->lSites();
+
+    typedef typename Cobj::scalar_object scalar_object;
+
+    autoView(Aself_v,    A[geom.npoint-1], CpuRead);
+    autoView(AselfInv_v, AselfInv,         CpuWrite);
+    thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke
+      Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis);
+      Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis);
+
+      scalar_object selfLink    = Zero();
+      scalar_object selfLinkInv = Zero();
+
+      Coordinate lcoor;
+
+      Grid()->LocalIndexToLocalCoor(site, lcoor);
+      peekLocalSite(selfLink, Aself_v, lcoor);
+
+      for (int i = 0; i < nbasis; ++i)
+        for (int j = 0; j < nbasis; ++j)
+          selfLinkEigen(i, j) = static_cast<ComplexD>(TensorRemove(selfLink(i, j)));
+
+      selfLinkInvEigen = selfLinkEigen.inverse();
+
+      for(int i = 0; i < nbasis; ++i)
+        for(int j = 0; j < nbasis; ++j)
+          selfLinkInv(i, j) = selfLinkInvEigen(i, j);
+
+      pokeLocalSite(selfLinkInv, AselfInv_v, lcoor);
+    });
   }
+
+  void FillHalfCbs() {
+    std::cout << GridLogDebug << "CoarsenedMatrix::FillHalfCbs" << std::endl;
+    for(int p = 0; p < geom.npoint; ++p) {
+      pickCheckerboard(Even, Aeven[p], A[p]);
+      pickCheckerboard(Odd, Aodd[p], A[p]);
     }
+    pickCheckerboard(Even, AselfInvEven, AselfInv);
+    pickCheckerboard(Odd, AselfInvOdd, AselfInv);
   }
 };
 

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> 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/iterative/BiCGSTABMixedPrec.h

@@ -37,6 +37,7 @@ template<class FieldD, class FieldF, typename std::enable_if< getPrecision<Field
 class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD> 
 {
   public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -36,6 +36,7 @@ NAMESPACE_BEGIN(Grid);
     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
   public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

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)

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) 
     {
@@ -150,22 +175,27 @@ 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) {
+        for(int b=0;b<nblock;b++){
           pickCheckerboard(Odd, sol_o[b], out[b]);
+        }
       } else {
-          guess(src_o[b],sol_o[b]); 
+        guess(src_o, sol_o); 
       }
 
 	    if ( subGuess ) { 
+        for(int b=0;b<nblock;b++){
           guess_save[b] = sol_o[b];
         }
       }

Grid/allocator/AlignedAllocator.cc

@@ -1,67 +0,0 @@
-#include <Grid/GridCore.h>
-#include <fcntl.h>
-
-NAMESPACE_BEGIN(Grid);
-
-MemoryStats *MemoryProfiler::stats = nullptr;
-bool         MemoryProfiler::debug = false;
-
-void check_huge_pages(void *Buf,uint64_t BYTES)
-{
-#ifdef __linux__
-  int fd = open("/proc/self/pagemap", O_RDONLY);
-  assert(fd >= 0);
-  const int page_size = 4096;
-  uint64_t virt_pfn = (uint64_t)Buf / page_size;
-  off_t offset = sizeof(uint64_t) * virt_pfn;
-  uint64_t npages = (BYTES + page_size-1) / page_size;
-  uint64_t pagedata[npages];
-  uint64_t ret = lseek(fd, offset, SEEK_SET);
-  assert(ret == offset);
-  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
-  assert(ret == sizeof(uint64_t) * npages);
-  int nhugepages = npages / 512;
-  int n4ktotal, nnothuge;
-  n4ktotal = 0;
-  nnothuge = 0;
-  for (int i = 0; i < nhugepages; ++i) {
-    uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
-    for (int j = 0; j < 512; ++j) {
-      uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
-      ++n4ktotal;
-      if (pageaddr != baseaddr + j * page_size)
-	++nnothuge;
-    }
-  }
-  int rank = CartesianCommunicator::RankWorld();
-  printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
-#endif
-}
-
-std::string sizeString(const size_t bytes)
-{
-  constexpr unsigned int bufSize = 256;
-  const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
-  char                   buf[256];
-  size_t                 s     = 0;
-  double                 count = bytes;
-  
-  while (count >= 1024 && s < 7)
-    {
-      s++;
-      count /= 1024;
-    }
-  if (count - floor(count) == 0.0)
-    {
-      snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
-    }
-  else
-    {
-      snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
-    }
-  
-  return std::string(buf);
-}
-
-NAMESPACE_END(Grid);
-

Grid/allocator/AlignedAllocator.h

@@ -65,8 +65,7 @@ public:
     MemoryManager::CpuFree((void *)__p,bytes);
   }
 
-  // FIXME: hack for the copy constructor, eventually it must be avoided
-  //void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
+  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
   void construct(pointer __p, const _Tp& __val) { assert(0);};
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
@@ -74,6 +73,9 @@ public:
 template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
+//////////////////////////////////////////////////////////////////////////////////////
+// Unified virtual memory
+//////////////////////////////////////////////////////////////////////////////////////
 template<typename _Tp>
 class uvmAllocator {
 public: 
@@ -109,22 +111,72 @@ public:
     MemoryManager::SharedFree((void *)__p,bytes);
   }
 
-  // FIXME: hack for the copy constructor, eventually it must be avoided
   void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
-  //void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
 template<typename _Tp>  inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
 
+////////////////////////////////////////////////////////////////////////////////
+// Device memory
+////////////////////////////////////////////////////////////////////////////////
+template<typename _Tp>
+class devAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef devAllocator<_Tp1> other; };
+  devAllocator() throw() { }
+  devAllocator(const devAllocator&) throw() { }
+  template<typename _Tp1> devAllocator(const devAllocator<_Tp1>&) throw() { }
+  ~devAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+  pointer allocate(size_type __n, const void* _p= 0)
+  { 
+    size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
+    _Tp *ptr = (_Tp*) MemoryManager::AcceleratorAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
+    profilerFree(bytes);
+    MemoryManager::AcceleratorFree((void *)__p,bytes);
+  }
+  void construct(pointer __p, const _Tp& __val) { };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return false; }
+
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using commAllocator = uvmAllocator<T>;
+#ifdef ACCELERATOR_CSHIFT
+// Cshift on device
+template<class T> using cshiftAllocator = devAllocator<T>;
+#else
+// Cshift on host
+template<class T> using cshiftAllocator = std::allocator<T>;
+#endif
+
 template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
-template<class T> using commVector = std::vector<T,uvmAllocator<T> >;
-//template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
+template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
+template<class T> using commVector = std::vector<T,devAllocator<T> >;
+template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);
 

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, 2, 8, 2, 8 };
+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 ) {
@@ -136,6 +179,15 @@ void MemoryManager::Init(void)
       Ncache[SharedSmall]=Nc;
     }
   }
+
+}
+
+void MemoryManager::InitMessage(void) {
+
+#ifndef GRID_UVM
+  std::cout << GridLogMessage << "MemoryManager Cache "<< MemoryManager::DeviceMaxBytes <<" bytes "<<std::endl;
+#endif
+  
   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
 #ifdef ALLOCATION_CACHE
   std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
@@ -164,6 +216,7 @@ void MemoryManager::Init(void)
   std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_device"<<std::endl;
 #endif
 #endif
+
 }
 
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
@@ -171,13 +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
@@ -201,6 +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;
@@ -209,6 +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;
 }
@@ -218,13 +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
@@ -233,6 +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

@@ -34,8 +34,6 @@ NAMESPACE_BEGIN(Grid);
 
 // Move control to configure.ac and Config.h?
 
-#define ALLOCATION_CACHE
-#define GRID_ALLOC_ALIGN (2*1024*1024)
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
 /*Pinning pages is costly*/
@@ -84,20 +82,22 @@ 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 *AcceleratorAllocate(size_t bytes);
-  static void  AcceleratorFree    (void *ptr,size_t bytes);
   static void PrintBytes(void);
  public:
   static void Init(void);
+  static void InitMessage(void);
+  static void *AcceleratorAllocate(size_t bytes);
+  static void  AcceleratorFree    (void *ptr,size_t bytes);
   static void *SharedAllocate(size_t bytes);
   static void  SharedFree    (void *ptr,size_t bytes);
   static void *CpuAllocate(size_t bytes);
@@ -170,6 +170,7 @@ private:
 
  public:
   static void Print(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

@@ -1,11 +1,12 @@
 #include <Grid/GridCore.h>
-
 #ifndef GRID_UVM
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
+//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
 #define dprintf(...)
 
+
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
 ////////////////////////////////////////////////////////////
@@ -103,7 +104,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  //  dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -111,7 +112,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    //    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -125,7 +126,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  //  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   if(AccCache.state==AccDirty) {
@@ -136,7 +137,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    //    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -149,7 +150,7 @@ 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);
+  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -164,7 +165,7 @@ 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);
+  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -227,18 +228,24 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   // Find if present, otherwise get or force an empty
   ////////////////////////////////////////////////////////////////////////////
   if ( EntryPresent(CpuPtr)==0 ){
-    EvictVictims(bytes);
     EntryCreate(CpuPtr,bytes,mode,hint);
   }
 
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
-  
+  if (!AccCache.AccPtr) {
+    EvictVictims(bytes); 
+  } 
   assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
 
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
+    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
+		    (uint64_t)AccCache.CpuPtr,
+		    (uint64_t)CpuPtr,
+		    (uint64_t)AccCache.bytes,
+		    (uint64_t)bytes);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -285,21 +292,21 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    //    printf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("Copied 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++;
-    //    printf("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++;
-    //    printf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
@@ -361,13 +368,16 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   // Find if present, otherwise get or force an empty
   ////////////////////////////////////////////////////////////////////////////
   if ( EntryPresent(CpuPtr)==0 ){
-    EvictVictims(bytes);
     EntryCreate(CpuPtr,bytes,mode,transient);
   }
 
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
+  if (!AccCache.AccPtr) {
+     EvictVictims(bytes);
+  }
+
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
 
@@ -419,6 +429,7 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 }
 void  MemoryManager::Print(void)
 {
+  PrintBytes();
   std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
   std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
   std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
@@ -463,6 +474,32 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
   }
 }
 
+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 << "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;
+
+  } else {
+    std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl; 
+  }
+}
+
 NAMESPACE_END(Grid);
 
 #endif

Grid/allocator/MemoryManagerShared.cc

@@ -1,7 +1,6 @@
 #include <Grid/GridCore.h>
 #ifdef GRID_UVM
 
-#warning "Grid is assuming unified virtual memory address space"
 NAMESPACE_BEGIN(Grid);
 /////////////////////////////////////////////////////////////////////////////////
 // View management is 1:1 address space mapping
@@ -17,6 +16,10 @@ uint64_t  MemoryManager::DeviceToHostXfer;
 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::NotifyDeletion(void *ptr){};
 

Grid/cartesian/Cartesian_red_black.h

@@ -36,7 +36,7 @@ static const int CbBlack=1;
 static const int Even   =CbRed;
 static const int Odd    =CbBlack;
 
-accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex, Coordinate &rdim, Coordinate &chk_dim_msk)
+accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex,const Coordinate &rdim,const Coordinate &chk_dim_msk)
 {
   int nd=rdim.size();
   Coordinate coor(nd);

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

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -36,6 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+extern bool Stencil_force_mpi ;
+
 class CartesianCommunicator : public SharedMemory {
 
 public:    
@@ -108,6 +109,8 @@ public:
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
+  void GlobalMax(RealD &);
+  void GlobalMax(RealF &);
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
   void GlobalSum(RealD &);
@@ -138,21 +141,6 @@ public:
 		      int recv_from_rank,
 		      int bytes);
   
-  void SendRecvPacket(void *xmit,
-		      void *recv,
-		      int xmit_to_rank,
-		      int recv_from_rank,
-		      int bytes);
-  
-  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-			   void *xmit,
-			   int xmit_to_rank,
-			   void *recv,
-			   int recv_from_rank,
-			   int bytes);
-  
-  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
-
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,
 			       void *recv,

Grid/communicator/Communicator_mpi3.cc

@@ -43,8 +43,16 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
 
+#ifndef GRID_COMMS_THREADS
+    nCommThreads=1;
+    // wrong results here too
+    // For now: comms-overlap leads to wrong results in Benchmark_wilson even on single node MPI runs
+    // other comms schemes are ok
+    MPI_Init_thread(argc,argv,MPI_THREAD_SERIALIZED,&provided);
+#else
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+#endif
     //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
     if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
       assert(0);
@@ -267,6 +275,16 @@ void CartesianCommunicator::GlobalXOR(uint64_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalMax(float &f)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSum(float &f){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
@@ -294,60 +312,28 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int bytes)
 {
   std::vector<CommsRequest_t> reqs(0);
-  //    unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  //    unsigned long  rcrc = crc32(0L, Z_NULL, 0);
-  //    xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
-  SendToRecvFromComplete(reqs);
-  //    rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //    printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
-}
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int sender,
-					   int receiver,
-					   int bytes)
-{
-  MPI_Status stat;
-  assert(sender != receiver);
-  int tag = sender;
-  if ( _processor == sender ) {
-    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
-  }
-  if ( _processor == receiver ) { 
-    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
-  }
-}
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
+  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+
   int myrank = _processor;
   int ierr;
 
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    MPI_Request xrq;
-    MPI_Request rrq;
+  // Enforce no UVM in comms, device or host OK
+  assert(acceleratorIsCommunicable(xmit));
+  assert(acceleratorIsCommunicable(recv));
 
-    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-    
-    assert(ierr==0);
-    list.push_back(xrq);
-    list.push_back(rrq);
-  } else { 
   // Give the CPU to MPI immediately; can use threads to overlap optionally
+  //  printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
   ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
 		    recv,bytes,MPI_CHAR,from, from,
 		    communicator,MPI_STATUS_IGNORE);
   assert(ierr==0);
-  }
-}
 
+  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
+  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
+}
+// Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int dest,
 						     void *recv,
@@ -382,19 +368,28 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
+  int tag;
 
-  if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
-  if ( gdest == MPI_UNDEFINED ) {
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
+  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 ) {
@@ -403,16 +398,11 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 
   return off_node_bytes;
 }
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
-{
-  SendToRecvFromComplete(waitall);
-}
-void CartesianCommunicator::StencilBarrier(void)
-{
-  MPI_Barrier  (ShmComm);
-}
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+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;
@@ -422,6 +412,13 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
   assert(ierr==0);
   list.resize(0);
 }
+void CartesianCommunicator::StencilBarrier(void)
+{
+  MPI_Barrier  (ShmComm);
+}
+//void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+//{
+//}
 void CartesianCommunicator::Barrier(void)
 {
   int ierr = MPI_Barrier(communicator);
@@ -483,5 +480,3 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 }
 
 NAMESPACE_END(Grid);
-
-

Grid/communicator/Communicator_none.cc

@@ -67,6 +67,8 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 
 CartesianCommunicator::~CartesianCommunicator(){}
 
+void CartesianCommunicator::GlobalMax(float &){}
+void CartesianCommunicator::GlobalMax(double &){}
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}
@@ -77,15 +79,6 @@ void CartesianCommunicator::GlobalSumVector(uint64_t *,int N){}
 void CartesianCommunicator::GlobalXOR(uint32_t &){}
 void CartesianCommunicator::GlobalXOR(uint64_t &){}
 
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int xmit_to_rank,
-					   int recv_from_rank,
-					   int bytes)
-{
-  assert(0);
-}
-
 
 // Basic Halo comms primitive -- should never call in single node
 void CartesianCommunicator::SendToRecvFrom(void *xmit,
@@ -96,20 +89,6 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
-  assert(0);
-}
-
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  assert(0);
-}
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   bcopy(in,out,bytes*words);
@@ -137,10 +116,6 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int recv_from_rank,
 						     int bytes, int dir)
 {
-  std::vector<CommsRequest_t> list;
-  // Discard the "dir"
-  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
-  SendToRecvFromComplete(list);
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
@@ -150,13 +125,10 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int recv_from_rank,
 							 int bytes, int dir)
 {
-  // Discard the "dir"
-  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
   return 2.0*bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {
-  SendToRecvFromComplete(waitall);
 }
 
 void CartesianCommunicator::StencilBarrier(void){};

Grid/communicator/SharedMemory.h

@@ -102,7 +102,7 @@ public:
   ///////////////////////////////////////////////////
   static void SharedMemoryAllocate(uint64_t bytes, int flags);
   static void SharedMemoryFree(void);
-  static void SharedMemoryCopy(void *dest,const void *src,size_t bytes);
+  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
   static void SharedMemoryZero(void *dest,size_t bytes);
 
 };

Grid/communicator/SharedMemoryMPI.cc

@@ -7,6 +7,7 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christoph Lehner <christoph@lhnr.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
@@ -31,6 +32,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
+#endif
+#ifdef GRID_HIP
+#include <hip/hip_runtime_api.h>
+#endif
+#ifdef GRID_SYCl
+
 #endif
 
 NAMESPACE_BEGIN(Grid); 
@@ -47,7 +54,12 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
+#ifndef GRID_MPI3_SHM_NONE
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
+#else
+  MPI_Comm_split(comm, WorldRank, 0, &WorldShmComm);
+#endif
+
   MPI_Comm_rank(WorldShmComm     ,&WorldShmRank);
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
@@ -61,6 +73,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   WorldNodes = WorldSize/WorldShmSize;
   assert( (WorldNodes * WorldShmSize) == WorldSize );
 
+
   // FIXME: Check all WorldShmSize are the same ?
 
   /////////////////////////////////////////////////////////////////////
@@ -161,6 +174,23 @@ static inline int divides(int a,int b)
 }
 void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
 {
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
   ////////////////////////////////////////////////////////////////
   // Powers of 2,3,5 only in prime decomposition for now
   ////////////////////////////////////////////////////////////////
@@ -420,7 +450,47 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_CUDA
+#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 ; 
@@ -443,37 +513,70 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  auto err =  cudaMalloc(&ShmCommBuf, bytes);
-  if ( err !=  cudaSuccess) {
-    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
-    exit(EXIT_FAILURE);  
-  }
+  ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+    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
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   for(int r=0;r<WorldShmSize;r++){
 
+#ifndef GRID_MPI3_SHM_NONE
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
     //////////////////////////////////////////////////
-    cudaIpcMemHandle_t handle;
+    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 ) { 
-      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      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 ) { 
+      auto err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
+#endif
+#ifdef GRID_HIP
+    hipIpcMemHandle_t handle;    
+    if ( r==WorldShmRank ) { 
+      auto err = hipIpcGetMemHandle(&handle,ShmCommBuf);
+      if ( err !=  hipSuccess) {
+	std::cerr << " SharedMemoryMPI.cc hipIpcGetMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+#endif
+
     //////////////////////////////////////////////////
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
@@ -489,23 +592,68 @@ 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 ) {
-      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      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);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
+#endif
+#ifdef GRID_HIP
+    if ( r!=WorldShmRank ) { 
+      auto err = hipIpcOpenMemHandle(&thisBuf,handle,hipIpcMemLazyEnablePeerAccess);
+      if ( err !=  hipSuccess) {
+	std::cerr << " SharedMemoryMPI.cc hipIpcOpenMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+#endif
     ///////////////////////////////////////////////////////////////
     // Save a copy of the device buffers
     ///////////////////////////////////////////////////////////////
+    }
     WorldShmCommBufs[r] = thisBuf;
+#else
+    WorldShmCommBufs[r] = ShmCommBuf;
+#endif
   }
 
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
+#endif
+
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
@@ -633,7 +781,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -677,16 +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,const void *src,size_t bytes)
+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
@@ -705,7 +852,11 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
+#ifndef GRID_MPI3_SHM_NONE
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
+#else
+  MPI_Comm_split(comm, rank, 0, &ShmComm);
+#endif
   MPI_Comm_rank(ShmComm     ,&ShmRank);
   MPI_Comm_size(ShmComm     ,&ShmSize);
   ShmCommBufs.resize(ShmSize);
@@ -735,25 +886,18 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
 
-#ifdef GRID_IBM_SUMMIT
-  // Hide the shared memory path between sockets 
-  // if even number of nodes
-  if ( (ShmSize & 0x1)==0 ) {
-    int SocketSize = ShmSize/2;
-    int mySocket = ShmRank/SocketSize; 
+#ifdef GRID_SHM_FORCE_MPI
+  // Hide the shared memory path between ranks
+  {
     for(int r=0;r<size;r++){
-      int hisRank=ShmRanks[r];
-      if ( hisRank!= MPI_UNDEFINED ) {
-	int hisSocket=hisRank/SocketSize;
-	if ( hisSocket != mySocket ) {
+      if ( r!=rank ) {
 	ShmRanks[r] = MPI_UNDEFINED;
       }
     }
   }
-  }
 #endif
 
-  SharedMemoryTest();
+  //SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier

Grid/communicator/SharedMemoryNone.cc

@@ -29,6 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 
 NAMESPACE_BEGIN(Grid); 
+#define header "SharedMemoryNone: "
 
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@@ -55,6 +56,38 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended, use anonymous mmap
 ////////////////////////////////////////////////////////////////////////////////////////////
+#if 1
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl;
+  void * ShmCommBuf ; 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Each MPI rank should allocate our own buffer
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  ShmCommBuf = acceleratorAllocDevice(bytes);
+
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryNone.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+  if ( WorldRank == 0 ){
+    std::cout << WorldRank << header " SharedMemoryNone.cc acceleratorAllocDevice "<< bytes 
+	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+  }
+  SharedMemoryZero(ShmCommBuf,bytes);
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Loop over ranks/gpu's on our node
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  WorldShmCommBufs[0] = ShmCommBuf;
+
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+}
+#else
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -83,7 +116,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 };
-
+#endif
+void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
+{
+  acceleratorMemSet(dest,0,bytes);
+}
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+{
+  acceleratorCopyToDevice(src,dest,bytes);
+}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality

Grid/cshift/Cshift.h

@@ -52,23 +52,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<typename Op, typename T1> 
-auto Cshift(const LatticeUnaryExpression<Op,T1> &expr,int dim,int shift)
-    -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
-{
-  return Cshift(closure(expr),dim,shift);
-}
-template <class Op, class T1, class T2>
-auto Cshift(const LatticeBinaryExpression<Op,T1,T2> &expr,int dim,int shift)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> 
-{
-  return Cshift(closure(expr),dim,shift);
-}
-template <class Op, class T1, class T2, class T3>
-auto Cshift(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr,int dim,int shift)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
-				   eval(0, expr.arg2),
-				   eval(0, expr.arg3)))> 
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+auto Cshift(const Expression &expr,int dim,int shift)  -> decltype(closure(expr)) 
 {
   return Cshift(closure(expr),dim,shift);
 }

Grid/cshift/Cshift_common.h

@@ -35,7 +35,7 @@ extern Vector<std::pair<int,int> > Cshift_table;
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -73,12 +73,19 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
      }
   }
   {
-    autoView(rhs_v , rhs, AcceleratorRead);
     auto buffer_p = & buffer[0];
     auto table = &Cshift_table[0];
-    accelerator_for(i,ent,1,{
+#ifdef ACCELERATOR_CSHIFT    
+    autoView(rhs_v , rhs, AcceleratorRead);
+    accelerator_for(i,ent,vobj::Nsimd(),{
+	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
+    });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for(i,ent,{
       buffer_p[table[i].first]=rhs_v[table[i].second];
     });
+#endif
   }
 }
 
@@ -103,21 +110,36 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for2d(n,e1,b,e2,1,{
+    accelerator_for(nn,e1*e2,1,{
+	int n = nn%e1;
+	int b = nn/e1;
 	int o      =   n*n1;
 	int offset = b+n*e2;
 	
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
-  } else { 
-    autoView(rhs_v , rhs, AcceleratorRead);
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+	int o      =   n*n1;
+	int offset = b+n*e2;
 	
+	vobj temp =rhs_v[so+o+b];
+	extract<vobj>(temp,pointers,offset);
+      });
+#endif
+  } else { 
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-    accelerator_for2d(n,e1,b,e2,1,{
+#ifdef ACCELERATOR_CSHIFT    
+    autoView(rhs_v , rhs, AcceleratorRead);
+    accelerator_for(nn,e1*e2,1,{
+	int n = nn%e1;
+	int b = nn/e1;
 
 	Coordinate coor;
 
@@ -134,13 +156,33 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+
+	Coordinate coor;
+
+	int o=n*n1;
+	int oindex = o+b;
+
+       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
+
+	int ocb=1<<cb;
+	int offset = b+n*e2;
+
+	if ( ocb & cbmask ) {
+	  vobj temp =rhs_v[so+o+b];
+	  extract<vobj>(temp,pointers,offset);
+	}
+      });
+#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -182,12 +224,19 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
   }
   
   {
-    autoView( rhs_v, rhs, AcceleratorWrite);
     auto buffer_p = & buffer[0];
     auto table = &Cshift_table[0];
-    accelerator_for(i,ent,1,{
+#ifdef ACCELERATOR_CSHIFT    
+    autoView( rhs_v, rhs, AcceleratorWrite);
+    accelerator_for(i,ent,vobj::Nsimd(),{
+	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
+    });
+#else
+    autoView( rhs_v, rhs, CpuWrite);
+    thread_for(i,ent,{
       rhs_v[table[i].first]=buffer_p[table[i].second];
     });
+#endif
   }
 }
 
@@ -208,18 +257,32 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
+    int _slice_stride = rhs.Grid()->_slice_stride[dimension];
+    int _slice_block = rhs.Grid()->_slice_block[dimension];
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v , rhs, AcceleratorWrite);
-    accelerator_for2d(n,e1,b,e2,1,{
-	int o      = n*rhs.Grid()->_slice_stride[dimension];
-	int offset = b+n*rhs.Grid()->_slice_block[dimension];
+    accelerator_for(nn,e1*e2,1,{
+	int n = nn%e1;
+	int b = nn/e1;
+	int o      = n*_slice_stride;
+	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
+#else
+    autoView( rhs_v , rhs, CpuWrite);
+    thread_for2d(n,e1,b,e2,{
+	int o      = n*_slice_stride;
+	int offset = b+n*_slice_block;
+	merge(rhs_v[so+o+b],pointers,offset);
+    });
+#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
     // Test_cshift_red_black code.
-    //    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
+    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
+    assert(0); // This will fail if hit on GPU
     autoView( rhs_v, rhs, CpuWrite);
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
@@ -277,12 +340,20 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
+    auto table = &Cshift_table[0];
+#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
-    auto table = &Cshift_table[0];
-    accelerator_for(i,ent,1,{
+    accelerator_for(i,ent,vobj::Nsimd(),{
+      coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
+    });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    autoView(lhs_v , lhs, CpuWrite);
+    thread_for(i,ent,{
       lhs_v[table[i].first]=rhs_v[table[i].second];
     });
+#endif
   }
 }
 
@@ -321,12 +392,20 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
+    auto table = &Cshift_table[0];
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
-    auto table = &Cshift_table[0];
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
+#else
+    autoView( rhs_v, rhs, CpuRead);
+    autoView( lhs_v, lhs, CpuWrite);
+    thread_for(i,ent,{
+      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+    });
+#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

@@ -101,7 +101,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
 }
-
+#define ACCELERATOR_CSHIFT_NO_COPY
+#ifdef ACCELERATOR_CSHIFT_NO_COPY
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
   typedef typename vobj::vector_type vector_type;
@@ -121,8 +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];
-  commVector<vobj> send_buf(buffer_size);
-  commVector<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);
@@ -138,7 +139,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
     } else {
 
-      int words = send_buf.size();
+      int words = buffer_size;
       if (cbmask != 0x3) words=words>>1;
 
       int bytes = words * sizeof(vobj);
@@ -150,12 +151,14 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
+      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+
       grid->Barrier();
 
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
@@ -195,8 +198,15 @@ 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<commVector<scalar_object> >   send_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
-  std::vector<commVector<scalar_object> >   recv_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
+  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;
+ 
+  for(int s=0;s<Nsimd;s++){
+    send_buf_extract[s].resize(buffer_size);
+    recv_buf_extract[s].resize(buffer_size);
+  }
 
   int bytes = buffer_size*sizeof(scalar_object);
 
@@ -242,11 +252,204 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	grid->SendToRecvFrom((void *)&send_buf_extract[nbr_lane][0],
+	grid->Barrier();
+
+	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
+	recv_buf_extract_mpi = &recv_buf_extract[i][0];
+	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
 			     xmit_to_rank,
-			     (void *)&recv_buf_extract[i][0],
+			     (void *)recv_buf_extract_mpi,
 			     recv_from_rank,
 			     bytes);
+
+	grid->Barrier();
+
+	rpointers[i] = &recv_buf_extract[i][0];
+      } else { 
+	rpointers[i] = &send_buf_extract[nbr_lane][0];
+      }
+
+    }
+    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
+  }
+
+}
+#else 
+template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
+{
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  GridBase *grid=rhs.Grid();
+  Lattice<vobj> temp(rhs.Grid());
+
+  int fd              = rhs.Grid()->_fdimensions[dimension];
+  int rd              = rhs.Grid()->_rdimensions[dimension];
+  int pd              = rhs.Grid()->_processors[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
+  assert(simd_layout==1);
+  assert(comm_dim==1);
+  assert(shift>=0);
+  assert(shift<fd);
+  
+  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
+  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;
+  {
+    grid->ShmBufferFreeAll();
+    size_t bytes = buffer_size*sizeof(vobj);
+    send_buf=(vobj *)grid->ShmBufferMalloc(bytes);
+    recv_buf=(vobj *)grid->ShmBufferMalloc(bytes);
+  }
+    
+  int cb= (cbmask==0x2)? Odd : Even;
+  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
+
+  for(int x=0;x<rd;x++){       
+
+    int sx        =  (x+sshift)%rd;
+    int comm_proc = ((x+sshift)/rd)%pd;
+    
+    if (comm_proc==0) {
+
+      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
+
+    } else {
+
+      int words = buffer_size;
+      if (cbmask != 0x3) words=words>>1;
+
+      int bytes = words * sizeof(vobj);
+
+      Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
+
+      //      int rank           = grid->_processor;
+      int recv_from_rank;
+      int xmit_to_rank;
+      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+
+      grid->Barrier();
+
+      acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
+      grid->SendToRecvFrom((void *)&send_buf[0],
+			   xmit_to_rank,
+			   (void *)&recv_buf[0],
+			   recv_from_rank,
+			   bytes);
+      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
+
+      grid->Barrier();
+
+      Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
+    }
+  }
+}
+
+template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
+{
+  GridBase *grid=rhs.Grid();
+  const int Nsimd = grid->Nsimd();
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_object scalar_object;
+  typedef typename vobj::scalar_type scalar_type;
+   
+  int fd = grid->_fdimensions[dimension];
+  int rd = grid->_rdimensions[dimension];
+  int ld = grid->_ldimensions[dimension];
+  int pd = grid->_processors[dimension];
+  int simd_layout     = grid->_simd_layout[dimension];
+  int comm_dim        = grid->_processors[dimension] >1 ;
+
+  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
+  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
+  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
+
+  assert(comm_dim==1);
+  assert(simd_layout==2);
+  assert(shift>=0);
+  assert(shift<fd);
+
+  int permute_type=grid->PermuteType(dimension);
+
+  ///////////////////////////////////////////////
+  // Simd direction uses an extract/merge pair
+  ///////////////////////////////////////////////
+  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
+  //  int words = sizeof(vobj)/sizeof(vector_type);
+
+  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;
+  {
+    size_t bytes = sizeof(scalar_object)*buffer_size;
+    grid->ShmBufferFreeAll();
+    send_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
+    recv_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
+  }
+  for(int s=0;s<Nsimd;s++){
+    send_buf_extract[s].resize(buffer_size);
+    recv_buf_extract[s].resize(buffer_size);
+  }
+
+  int bytes = buffer_size*sizeof(scalar_object);
+
+  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
+  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
+
+  ///////////////////////////////////////////
+  // Work out what to send where
+  ///////////////////////////////////////////
+  int cb    = (cbmask==0x2)? Odd : Even;
+  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
+
+  // loop over outer coord planes orthog to dim
+  for(int x=0;x<rd;x++){       
+
+    // FIXME call local permute copy if none are offnode.
+    for(int i=0;i<Nsimd;i++){       
+      pointers[i] = &send_buf_extract[i][0];
+    }
+    int sx   = (x+sshift)%rd;
+    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+
+    for(int i=0;i<Nsimd;i++){
+      
+      int inner_bit = (Nsimd>>(permute_type+1));
+      int ic= (i&inner_bit)? 1:0;
+
+      int my_coor          = rd*ic + x;
+      int nbr_coor         = my_coor+sshift;
+      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
+
+      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
+      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
+      int nbr_lane = (i&(~inner_bit));
+
+      int recv_from_rank;
+      int xmit_to_rank;
+
+      if (nbr_ic) nbr_lane|=inner_bit;
+
+      assert (sx == nbr_ox);
+
+      if(nbr_proc){
+	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
+
+	grid->Barrier();
+
+	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
+	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
+			     xmit_to_rank,
+			     (void *)recv_buf_extract_mpi,
+			     recv_from_rank,
+			     bytes);
+	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
+
 	grid->Barrier();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -258,7 +461,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   }
 
 }
-
+#endif
 NAMESPACE_END(Grid); 
 
 #endif

Grid/lattice/Lattice.h

@@ -36,7 +36,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_local.h>
 #include <Grid/lattice/Lattice_reduction.h>
 #include <Grid/lattice/Lattice_peekpoke.h>
-//#include <Grid/lattice/Lattice_reality.h>
+#include <Grid/lattice/Lattice_reality.h>
+#include <Grid/lattice/Lattice_real_imag.h>
 #include <Grid/lattice/Lattice_comparison_utils.h>
 #include <Grid/lattice/Lattice_comparison.h>
 #include <Grid/lattice/Lattice_coordinate.h>

Grid/lattice/Lattice_ET.h

@@ -42,9 +42,24 @@ NAMESPACE_BEGIN(Grid);
 ////////////////////////////////////////////////////
 // Predicated where support
 ////////////////////////////////////////////////////
+#ifdef GRID_SIMT
+// drop to scalar in SIMT; cleaner in fact
 template <class iobj, class vobj, class robj>
-accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
-                            const robj &iffalse) {
+accelerator_inline vobj predicatedWhere(const iobj &predicate, 
+					const vobj &iftrue, 
+					const robj &iffalse) 
+{
+  Integer mask = TensorRemove(predicate);
+  typename std::remove_const<vobj>::type ret= iffalse;
+  if (mask) ret=iftrue;
+  return ret;
+}
+#else
+template <class iobj, class vobj, class robj>
+accelerator_inline vobj predicatedWhere(const iobj &predicate, 
+					const vobj &iftrue, 
+					const robj &iffalse) 
+{
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
@@ -68,6 +83,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftru
   merge(ret, falsevals);
   return ret;
 }
+#endif
 
 /////////////////////////////////////////////////////
 //Specialization of getVectorType for lattices
@@ -81,32 +97,62 @@ struct getVectorType<Lattice<T> >{
 //--  recursive evaluation of expressions; --
 // handle leaves of syntax tree
 ///////////////////////////////////////////////////
-template<class sobj> accelerator_inline 
+template<class sobj,
+  typename std::enable_if<!is_lattice<sobj>::value&&!is_lattice_expr<sobj>::value,sobj>::type * = nullptr> 
+accelerator_inline 
 sobj eval(const uint64_t ss, const sobj &arg)
 {
   return arg;
 }
-
 template <class lobj> accelerator_inline 
-const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg) 
+auto eval(const uint64_t ss, const LatticeView<lobj> &arg) -> decltype(arg(ss))
+{
+  return arg(ss);
+}
+
+////////////////////////////////////////////
+//--  recursive evaluation of expressions; --
+// whole vector return, used only for expression return type inference
+///////////////////////////////////////////////////
+template<class sobj> accelerator_inline 
+sobj vecEval(const uint64_t ss, const sobj &arg)
+{
+  return arg;
+}
+template <class lobj> accelerator_inline 
+const lobj & vecEval(const uint64_t ss, const LatticeView<lobj> &arg) 
 {
   return arg[ss];
 }
 
-// What needs this?
-// Cannot be legal on accelerator
-// Comparison must convert
-#if 1
-template <class lobj> accelerator_inline 
-const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg) 
-{
-  auto view = arg.View(AcceleratorRead);
-  return view[ss];
-}
-#endif
-
 ///////////////////////////////////////////////////
 // handle nodes in syntax tree- eval one operand
+// vecEval needed (but never called as all expressions offloaded) to infer the return type
+// in SIMT contexts of closure.
+///////////////////////////////////////////////////
+template <typename Op, typename T1> accelerator_inline 
+auto vecEval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
+  -> decltype(expr.op.func( vecEval(ss, expr.arg1)))
+{
+  return expr.op.func( vecEval(ss, expr.arg1) );
+}
+// vecEval two operands
+template <typename Op, typename T1, typename T2> accelerator_inline
+auto vecEval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
+  -> decltype(expr.op.func( vecEval(ss,expr.arg1),vecEval(ss,expr.arg2)))
+{
+  return expr.op.func( vecEval(ss,expr.arg1), vecEval(ss,expr.arg2) );
+}
+// vecEval three operands
+template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
+auto vecEval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
+  -> decltype(expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3)))
+{
+  return expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3));
+}
+
+///////////////////////////////////////////////////
+// handle nodes in syntax tree- eval one operand coalesced
 ///////////////////////////////////////////////////
 template <typename Op, typename T1> accelerator_inline 
 auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
@@ -114,23 +160,41 @@ auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)
 {
   return expr.op.func( eval(ss, expr.arg1) );
 }
-///////////////////////
 // eval two operands
-///////////////////////
 template <typename Op, typename T1, typename T2> accelerator_inline
 auto eval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
   -> decltype(expr.op.func( eval(ss,expr.arg1),eval(ss,expr.arg2)))
 {
   return expr.op.func( eval(ss,expr.arg1), eval(ss,expr.arg2) );
 }
-///////////////////////
 // eval three operands
-///////////////////////
 template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
 auto eval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
-  -> decltype(expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3)))
+  -> decltype(expr.op.func(eval(ss, expr.arg1), 
+			   eval(ss, expr.arg2), 
+			   eval(ss, expr.arg3)))
 {
-  return expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3));
+#ifdef GRID_SIMT
+  // Handles Nsimd (vInteger) != Nsimd(ComplexD)
+  typedef decltype(vecEval(ss, expr.arg2)) rvobj;
+  typedef typename std::remove_reference<rvobj>::type vobj;
+
+  const int Nsimd = vobj::vector_type::Nsimd();
+
+  auto vpred = vecEval(ss,expr.arg1);
+
+  ExtractBuffer<Integer> mask(Nsimd);
+  extract<vInteger, Integer>(TensorRemove(vpred), mask);
+
+  int s = acceleratorSIMTlane(Nsimd);
+  return expr.op.func(mask[s],
+		      eval(ss, expr.arg2), 
+		      eval(ss, expr.arg3));
+#else
+  return expr.op.func(eval(ss, expr.arg1),
+		      eval(ss, expr.arg2), 
+		      eval(ss, expr.arg3));
+#endif
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -228,7 +292,7 @@ template <typename Op, typename T1, typename T2> inline
 void ExpressionViewOpen(LatticeBinaryExpression<Op, T1, T2> &expr) 
 {
   ExpressionViewOpen(expr.arg1);  // recurse AST
-  ExpressionViewOpen(expr.arg2);  // recurse AST
+  ExpressionViewOpen(expr.arg2);  // rrecurse AST
 }
 template <typename Op, typename T1, typename T2, typename T3>
 inline void ExpressionViewOpen(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
@@ -272,28 +336,20 @@ inline void ExpressionViewClose(LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
 // Unary operators and funcs
 ////////////////////////////////////////////
 #define GridUnopClass(name, ret)					\
-  template <class arg>							\
   struct name {								\
-    static auto accelerator_inline func(const arg a) -> decltype(ret) { return ret; } \
+    template<class _arg> static auto accelerator_inline func(const _arg a) -> decltype(ret) { return ret; } \
   };
 
 GridUnopClass(UnarySub, -a);
 GridUnopClass(UnaryNot, Not(a));
-GridUnopClass(UnaryAdj, adj(a));
-GridUnopClass(UnaryConj, conjugate(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
-GridUnopClass(UnaryReal, real(a));
-GridUnopClass(UnaryImag, imag(a));
-GridUnopClass(UnaryToReal, toReal(a));
-GridUnopClass(UnaryToComplex, toComplex(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
 GridUnopClass(UnarySqrt, sqrt(a));
-GridUnopClass(UnaryRsqrt, rsqrt(a));
 GridUnopClass(UnarySin, sin(a));
 GridUnopClass(UnaryCos, cos(a));
 GridUnopClass(UnaryAsin, asin(a));
@@ -305,10 +361,10 @@ GridUnopClass(UnaryExp, exp(a));
 // Binary operators
 ////////////////////////////////////////////
 #define GridBinOpClass(name, combination)			\
-  template <class left, class right>				\
   struct name {							\
+    template <class _left, class _right>			\
     static auto accelerator_inline				\
-    func(const left &lhs, const right &rhs)			\
+    func(const _left &lhs, const _right &rhs)			\
       -> decltype(combination) const				\
     {								\
       return combination;					\
@@ -328,10 +384,10 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
 // Trinary conditional op
 ////////////////////////////////////////////////////
 #define GridTrinOpClass(name, combination)				\
-  template <class predicate, class left, class right>			\
   struct name {								\
+    template <class _predicate,class _left, class _right>		\
     static auto accelerator_inline					\
-    func(const predicate &pred, const left &lhs, const right &rhs)	\
+    func(const _predicate &pred, const _left &lhs, const _right &rhs)	\
       -> decltype(combination) const					\
     {									\
       return combination;						\
@@ -339,17 +395,17 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
   };
 
 GridTrinOpClass(TrinaryWhere,
-		(predicatedWhere<predicate, 
-		 typename std::remove_reference<left>::type,
-		 typename std::remove_reference<right>::type>(pred, lhs,rhs)));
+		(predicatedWhere<
+		 typename std::remove_reference<_predicate>::type, 
+		 typename std::remove_reference<_left>::type,
+		 typename std::remove_reference<_right>::type>(pred, lhs,rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
 ////////////////////////////////////////////
-
-#define GRID_UNOP(name)   name<decltype(eval(0, arg))>
-#define GRID_BINOP(name)  name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
-#define GRID_TRINOP(name) name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
+#define GRID_UNOP(name)   name
+#define GRID_BINOP(name)  name
+#define GRID_TRINOP(name) name
 
 #define GRID_DEF_UNOP(op, name)						\
   template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
@@ -395,22 +451,17 @@ GridTrinOpClass(TrinaryWhere,
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
-GRID_DEF_UNOP(adj, UnaryAdj);
-GRID_DEF_UNOP(conjugate, UnaryConj);
+//GRID_DEF_UNOP(adj, UnaryAdj);
+//GRID_DEF_UNOP(conjugate, UnaryConj);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
-GRID_DEF_UNOP(real, UnaryReal);
-GRID_DEF_UNOP(imag, UnaryImag);
-GRID_DEF_UNOP(toReal, UnaryToReal);
-GRID_DEF_UNOP(toComplex, UnaryToComplex);
 GRID_DEF_UNOP(timesI, UnaryTimesI);
 GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
 GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the
                                // abs-fabs-dabs-labs thing
 GRID_DEF_UNOP(sqrt, UnarySqrt);
-GRID_DEF_UNOP(rsqrt, UnaryRsqrt);
 GRID_DEF_UNOP(sin, UnarySin);
 GRID_DEF_UNOP(cos, UnaryCos);
 GRID_DEF_UNOP(asin, UnaryAsin);
@@ -435,29 +486,36 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type >
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
-				   eval(0, expr.arg2),
-				   eval(0, expr.arg3)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+				   vecEval(0, expr.arg2),
+				   vecEval(0, expr.arg3)))>::type >
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1),
-				eval(0, expr.arg2),
-				eval(0, expr.arg3)))>  ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+				vecEval(0, expr.arg2),
+			        vecEval(0, expr.arg3)))>::type >  ret(expr);
   return ret;
 }
+#define EXPRESSION_CLOSURE(function)					\
+  template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> \
+    auto function(Expression &expr) -> decltype(function(closure(expr))) \
+  {									\
+    return function(closure(expr));					\
+  }
+
 
 #undef GRID_UNOP
 #undef GRID_BINOP

Grid/lattice/Lattice_arith.h

@@ -60,9 +60,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   autoView( lhs_v , lhs, AcceleratorRead);
   autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     auto rhs_t=rhs_v(ss);
+    auto tmp  =ret_v(ss);
     mac(&tmp,&lhs_t,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
   });
@@ -124,7 +124,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     auto lhs_t=lhs_v(ss);
     mac(&tmp,&lhs_t,&rhs);
     coalescedWrite(ret_v[ss],tmp);
@@ -182,7 +182,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     auto rhs_t=rhs_v(ss);
     mac(&tmp,&lhs,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
@@ -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
@@ -123,9 +130,9 @@ public:
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
     auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),1,{
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
-      vstream(me[ss],tmp);
+      coalescedWrite(me[ss],tmp);
     });
     me.ViewClose();
     ExpressionViewClose(exprCopy);
@@ -146,9 +153,9 @@ public:
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
     auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),1,{
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
-      vstream(me[ss],tmp);
+      coalescedWrite(me[ss],tmp);
     });
     me.ViewClose();
     ExpressionViewClose(exprCopy);
@@ -168,9 +175,9 @@ public:
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
     auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),1,{
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
-      vstream(me[ss],tmp);
+      coalescedWrite(me[ss],tmp);
     });
     me.ViewClose();
     ExpressionViewClose(exprCopy);

Grid/lattice/Lattice_basis.h

@@ -54,13 +54,34 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   typedef decltype(basis[0].View(AcceleratorRead)) View;
 
   Vector<View> basis_v; basis_v.reserve(basis.size());
+  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
+  typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
   GridBase* grid = basis[0].Grid();
       
   for(int k=0;k<basis.size();k++){
     basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
+#if ( (!defined(GRID_CUDA)) )
+  int max_threads = thread_max();
+  Vector < vobj > Bt(Nm * max_threads);
+  thread_region
+    {
+      vobj* B = &Bt[Nm * thread_num()];
+      thread_for_in_region(ss, grid->oSites(),{
+	  for(int j=j0; j<j1; ++j) B[j]=0.;
       
+	  for(int j=j0; j<j1; ++j){
+	    for(int k=k0; k<k1; ++k){
+	      B[j] +=Qt(j,k) * basis_v[k][ss];
+	    }
+	  }
+	  for(int j=j0; j<j1; ++j){
+	    basis_v[j][ss] = B[j];
+	  }
+	});
+    }
+#else
   View *basis_vp = &basis_v[0];
 
   int nrot = j1-j0;
@@ -70,14 +91,12 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   uint64_t oSites   =grid->oSites();
   uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
 
-  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
-
   Vector <vobj> Bt(siteBlock * nrot); 
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  Vector<double> Qt_jv(Nm*Nm);
-  double *Qt_p = & Qt_jv[0];
+  Vector<Coeff_t> Qt_jv(Nm*Nm);
+  Coeff_t *Qt_p = & Qt_jv[0];
   thread_for(i,Nm*Nm,{
       int j = i/Nm;
       int k = i%Nm;
@@ -106,7 +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]));
 	}
       });
 
@@ -115,9 +134,10 @@ 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
 
   for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
 }
@@ -141,11 +161,13 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
   double * Qt_j = & Qt_jv[0];
   for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
+  auto basis_vp=& basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    auto B=coalescedRead(zz);
+    vobj zzz=Zero();
+    auto B=coalescedRead(zzz);
     for(int k=k0; k<k1; ++k){
-      B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
+      B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]);
     }
     coalescedWrite(result_v[ss], B);
   });

Grid/lattice/Lattice_comparison.h

@@ -42,34 +42,6 @@ NAMESPACE_BEGIN(Grid);
 
 typedef iScalar<vInteger> vPredicate ;
 
-/*
-template <class iobj, class vobj, class robj> accelerator_inline 
-vobj predicatedWhere(const iobj &predicate, const vobj &iftrue, const robj &iffalse) 
-{
-  typename std::remove_const<vobj>::type ret;
-
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
-  const int Nsimd = vobj::vector_type::Nsimd();
-
-  ExtractBuffer<Integer> mask(Nsimd);
-  ExtractBuffer<scalar_object> truevals(Nsimd);
-  ExtractBuffer<scalar_object> falsevals(Nsimd);
-
-  extract(iftrue, truevals);
-  extract(iffalse, falsevals);
-  extract<vInteger, Integer>(TensorRemove(predicate), mask);
-
-  for (int s = 0; s < Nsimd; s++) {
-    if (mask[s]) falsevals[s] = truevals[s];
-  }
-
-  merge(ret, falsevals);
-  return ret;
-}
-*/
 //////////////////////////////////////////////////////////////////////////
 // compare lattice to lattice
 //////////////////////////////////////////////////////////////////////////

Grid/lattice/Lattice_peekpoke.h

@@ -182,6 +182,14 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
       
   return;
 };
+template<class vobj,class sobj>
+inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
+{
+  autoView(lv,l,CpuRead);
+  peekLocalSite(s,lv,site);
+  return;
+};
+
 // Must be CPU write view
 template<class vobj,class sobj>
 inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
@@ -210,6 +218,14 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
   return;
 };
 
+template<class vobj,class sobj>
+inline void pokeLocalSite(const sobj &s, Lattice<vobj> &l,Coordinate &site)
+{
+  autoView(lv,l,CpuWrite);
+  pokeLocalSite(s,lv,site);
+  return;
+};
+
 NAMESPACE_END(Grid);
 #endif
 

Grid/lattice/Lattice_real_imag.h

@@ -0,0 +1,79 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/Lattice_reality.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: neo <cossu@post.kek.jp>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_LATTICE_REAL_IMAG_H
+#define GRID_LATTICE_REAL_IMAG_H
+
+
+// FIXME .. this is the sector of the code 
+// I am most worried about the directions
+// The choice of burying complex in the SIMD
+// is making the use of "real" and "imag" very cumbersome
+
+NAMESPACE_BEGIN(Grid);
+
+template<class vobj> inline Lattice<vobj> real(const Lattice<vobj> &lhs){
+  Lattice<vobj> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
+  ret.Checkerboard()=lhs.Checkerboard();
+  accelerator_for( ss, lhs_v.size(), 1, {
+    ret_v[ss] =real(lhs_v[ss]);
+  });
+  return ret;
+};
+template<class vobj> inline Lattice<vobj> imag(const Lattice<vobj> &lhs){
+  Lattice<vobj> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
+  ret.Checkerboard()=lhs.Checkerboard();
+  accelerator_for( ss, lhs_v.size(), 1, {
+    ret_v[ss] =imag(lhs_v[ss]);
+  });
+  return ret;
+};
+
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+  auto real(const Expression &expr) -> decltype(real(closure(expr)))		
+{									
+  return real(closure(expr));					
+}
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+  auto imag(const Expression &expr) -> decltype(imag(closure(expr)))		
+{									
+  return imag(closure(expr));					
+}
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/lattice/Lattice_reality.h

@@ -45,8 +45,8 @@ template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
   autoView( ret_v, ret, AcceleratorWrite);
 
   ret.Checkerboard()=lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
-    coalescedWrite(ret_v[ss], adj(lhs_v(ss)));
+  accelerator_for( ss, lhs_v.size(), 1, {
+     ret_v[ss] = adj(lhs_v[ss]);
   });
   return ret;
 };
@@ -64,6 +64,53 @@ template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
   return ret;
 };
 
+template<class vobj> inline Lattice<typename vobj::Complexified> toComplex(const Lattice<vobj> &lhs){
+  Lattice<typename vobj::Complexified> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
+  ret.Checkerboard() = lhs.Checkerboard();
+  accelerator_for( ss, lhs_v.size(), 1, {
+    ret_v[ss] = toComplex(lhs_v[ss]);
+  });
+  return ret;
+};
+template<class vobj> inline Lattice<typename vobj::Realified> toReal(const Lattice<vobj> &lhs){
+  Lattice<typename vobj::Realified> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
+  ret.Checkerboard() = lhs.Checkerboard();
+  accelerator_for( ss, lhs_v.size(), 1, {
+    ret_v[ss] = toReal(lhs_v[ss]);
+  });
+  return ret;
+};
+
+
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+auto toComplex(const Expression &expr)  -> decltype(closure(expr)) 
+{
+  return toComplex(closure(expr));
+}
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+auto toReal(const Expression &expr)  -> decltype(closure(expr)) 
+{
+  return toReal(closure(expr));
+}
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+auto adj(const Expression &expr)  -> decltype(closure(expr)) 
+{
+  return adj(closure(expr));
+}
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+auto conjugate(const Expression &expr)  -> decltype(closure(expr)) 
+{
+  return conjugate(closure(expr));
+}
+
 NAMESPACE_END(Grid);
 
 #endif

Grid/lattice/Lattice_reduction.h

@@ -96,8 +96,34 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
   ssobj ret = ssum;
   return ret;
 }
+/*
+Threaded max, don't use for now
+template<class Double>
+inline Double max(const Double *arg, Integer osites)
+{
+  //  const int Nsimd = vobj::Nsimd();
+  const int nthread = GridThread::GetThreads();
 
+  std::vector<Double> maxarray(nthread);
   
+  thread_for(thr,nthread, {
+    int nwork, mywork, myoff;
+    nwork = osites;
+    GridThread::GetWork(nwork,thr,mywork,myoff);
+    Double max=arg[0];
+    for(int ss=myoff;ss<mywork+myoff; ss++){
+      if( arg[ss] > max ) max = arg[ss];
+    }
+    maxarray[thr]=max;
+  });
+  
+  Double tmax=maxarray[0];
+  for(int i=0;i<nthread;i++){
+    if (maxarray[i]>tmax) tmax = maxarray[i];
+  } 
+  return tmax;
+}
+*/
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
@@ -141,6 +167,32 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
   return real(nrm); 
 }
 
+//The global maximum of the site norm2
+template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
+{
+  typedef typename vobj::tensor_reduced vscalar;  //iScalar<iScalar<.... <vPODtype> > >
+  typedef typename vscalar::scalar_object  scalar;   //iScalar<iScalar<.... <PODtype> > >
+
+  Lattice<vscalar> inner = localNorm2(arg);
+
+  auto grid = arg.Grid();
+
+  RealD max;
+  for(int l=0;l<grid->lSites();l++){
+    Coordinate coor;
+    scalar val;
+    RealD r;
+    grid->LocalIndexToLocalCoor(l,coor);
+    peekLocalSite(val,inner,coor);
+    r=real(TensorRemove(val));
+    if( (l==0) || (r>max)){
+      max=r;
+    }
+  }
+  grid->GlobalMax(max);
+  return max;
+}
+
 // Double inner product
 template<class vobj>
 inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
@@ -309,6 +361,7 @@ 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);
 
@@ -367,20 +420,19 @@ 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

@@ -2,12 +2,13 @@ NAMESPACE_BEGIN(Grid);
 
 #ifdef GRID_HIP
 extern hipDeviceProp_t *gpu_props;
+#define WARP_SIZE 64
 #endif
 #ifdef GRID_CUDA
 extern cudaDeviceProp *gpu_props;
+#define WARP_SIZE 32
 #endif
 
-#define WARP_SIZE 32
 __device__ unsigned int retirementCount = 0;
 
 template <class Iterator>
@@ -41,7 +42,6 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   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) {
@@ -51,6 +51,10 @@ 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;
+    exit(EXIT_FAILURE);
+  }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
@@ -64,7 +68,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
   
   // cannot use overloaded operators for sobj as they are not volatile-qualified
   memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
-  __syncwarp();
+  acceleratorSynchronise();
   
   const Iterator VEC = WARP_SIZE;
   const Iterator vid = tid & (VEC-1);
@@ -78,9 +82,9 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
       beta += temp;
       memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
     }
-    __syncwarp();
+    acceleratorSynchronise();
   }
-  __syncthreads();
+  acceleratorSynchroniseAll();
   
   if (threadIdx.x == 0) {
     beta  = Zero();
@@ -90,7 +94,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
     }
     memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
   }
-  __syncthreads();
+  acceleratorSynchroniseAll();
 }
 
 

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
@@ -97,6 +167,20 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> &
   out = in;
 }
 
+template<typename T>
+accelerator_inline EnableIf<isGridFundamental<T>> convertType(T & out, const T & in) {
+  out = in;
+}
+
+// This would allow for conversions between GridFundamental types, but is not strictly needed as yet
+/*template<typename T1, typename T2>
+accelerator_inline typename std::enable_if<isGridFundamental<T1>::value && isGridFundamental<T2>::value>::type
+// Or to make this very broad, conversions between anything that's not a GridTensor could be allowed
+//accelerator_inline typename std::enable_if<!isGridTensor<T1>::value && !isGridTensor<T2>::value>::type
+convertType(T1 & out, const T2 & in) {
+  out = in;
+}*/
+
 #ifdef GRID_SIMT
 accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
   ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
@@ -117,18 +201,18 @@ accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
   Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
 }
 
-template<typename T1,typename T2,int N>
-  accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in);
-template<typename T1,typename T2,int N>
-  accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in);
+template<typename T1,typename T2>
+accelerator_inline void convertType(iScalar<T1> & out, const iScalar<T2> & in) {
+  convertType(out._internal,in._internal);
+}
 
-template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
-accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
+template<typename T1,typename T2>
+accelerator_inline NotEnableIf<isGridScalar<T1>> convertType(T1 & out, const iScalar<T2> & in) {
   convertType(out,in._internal);
 }
 
 template<typename T1,typename T2>
-accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
+accelerator_inline NotEnableIf<isGridScalar<T2>> convertType(iScalar<T1> & out, const T2 & in) {
   convertType(out._internal,in);
 }
 
@@ -145,11 +229,6 @@ accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & i
     convertType(out._internal[i],in._internal[i]);
 }
 
-template<typename T, typename std::enable_if<isGridFundamental<T>::value, T>::type* = nullptr>
-accelerator_inline void convertType(T & out, const T & in) {
-  out = in;
-}
-
 template<typename T1,typename T2>
 accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
   autoView( out_v , out,AcceleratorWrite);
@@ -240,6 +319,8 @@ template<class vobj,class vobj2,class CComplex>
   autoView( fineX_  , fineX, AcceleratorRead);
   autoView( fineY_  , fineY, AcceleratorRead);
   autoView( coarseA_, coarseA, AcceleratorRead);
+  Coordinate fine_rdimensions = fine->_rdimensions;
+  Coordinate coarse_rdimensions = coarse->_rdimensions;
 
   accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
 
@@ -247,9 +328,9 @@ template<class vobj,class vobj2,class CComplex>
       Coordinate coor_c(_ndimension);
       Coordinate coor_f(_ndimension);
 
-      Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
+      Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
       for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-      Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+      Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
 
       // z = A x + y
 #ifdef GRID_SIMT
@@ -353,12 +434,21 @@ 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();
+      Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
+
+      vobj cd = zz;
       
       for(int sb=0;sb<blockVol;sb++){
 
@@ -367,11 +457,13 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 	Coordinate coor_f(_ndimension);
 	Lexicographic::CoorFromIndex(coor_b,sb,block_r);               // Block sub coordinate
 	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);
+	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
 
-	coarseData_[sc]=coarseData_[sc]+fineData_[sf];
+	cd=cd+fineData_p[sf];
       }
 
+      coarseData_p[sc] = cd;
+
     });
   return;
 }

Grid/lattice/Lattice_view.h

@@ -67,8 +67,13 @@ public:
   accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
 #endif
 
+#if 1
+  //  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
+  accelerator_inline vobj       & operator[](size_t i) const { return this->_odata[i]; };
+#else
   accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
   accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
+#endif
   
   accelerator_inline uint64_t begin(void) const { return 0;};
   accelerator_inline uint64_t end(void)   const { return this->_odata_size; };

Grid/lattice/Lattice_where.h

@@ -43,7 +43,7 @@ inline void whereWolf(Lattice<vobj> &ret,const Lattice<iobj> &predicate,Lattice<
   conformable(iftrue,predicate);
   conformable(iftrue,ret);
 
-  GridBase *grid=iftrue._grid;
+  GridBase *grid=iftrue.Grid();
 
   typedef typename vobj::scalar_object scalar_object;
   typedef typename vobj::scalar_type scalar_type;
@@ -52,22 +52,23 @@ inline void whereWolf(Lattice<vobj> &ret,const Lattice<iobj> &predicate,Lattice<
 
   const int Nsimd = grid->Nsimd();
 
-  std::vector<Integer> mask(Nsimd);
-  std::vector<scalar_object> truevals (Nsimd);
-  std::vector<scalar_object> falsevals(Nsimd);
-
-  parallel_for(int ss=0;ss<iftrue._grid->oSites(); ss++){
-
-    extract(iftrue._odata[ss]   ,truevals);
-    extract(iffalse._odata[ss]  ,falsevals);
-    extract<vInteger,Integer>(TensorRemove(predicate._odata[ss]),mask);
-
-    for(int s=0;s<Nsimd;s++){
-      if (mask[s]) falsevals[s]=truevals[s];
-    }
-
-    merge(ret._odata[ss],falsevals);
+  autoView(iftrue_v,iftrue,CpuRead);
+  autoView(iffalse_v,iffalse,CpuRead);
+  autoView(predicate_v,predicate,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  Integer NN= grid->oSites();
+  thread_for(ss,NN,{
+    Integer mask;
+    scalar_object trueval;
+    scalar_object falseval;
+    for(int l=0;l<Nsimd;l++){
+      trueval =extractLane(l,iftrue_v[ss]);
+      falseval=extractLane(l,iffalse_v[ss]);
+      mask    =extractLane(l,predicate_v[ss]);
+      if (mask) falseval=trueval;
+      insertLane(l,ret_v[ss],falseval);
     }
+  });
 }
 
 template<class vobj,class iobj>
@@ -76,9 +77,9 @@ inline Lattice<vobj> whereWolf(const Lattice<iobj> &predicate,Lattice<vobj> &ift
   conformable(iftrue,iffalse);
   conformable(iftrue,predicate);
 
-  Lattice<vobj> ret(iftrue._grid);
+  Lattice<vobj> ret(iftrue.Grid());
 
-  where(ret,predicate,iftrue,iffalse);
+  whereWolf(ret,predicate,iftrue,iffalse);
 
   return ret;
 }

Grid/log/Log.h

@@ -130,6 +130,8 @@ public:
   friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
     if ( log.active ) {
+      std::ios_base::fmtflags f(stream.flags());
+
       stream << log.background()<<  std::left;
       if (log.topWidth > 0)
       {
@@ -152,6 +154,8 @@ public:
 	       << now	       << log.background() << " : " ;
       }
       stream << log.colour();
+      stream.flags(f);
+
       return stream;
     } else { 
       return devnull;

Grid/parallelIO/BinaryIO.cc

@@ -1,3 +1,4 @@
 #include <Grid/GridCore.h>
 
 int                    Grid::BinaryIO::latticeWriteMaxRetry = -1;
+Grid::BinaryIO::IoPerf Grid::BinaryIO::lastPerf;

Grid/parallelIO/BinaryIO.h

@@ -79,6 +79,13 @@ inline void removeWhitespace(std::string &key)
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 class BinaryIO {
  public:
+  struct IoPerf
+  {
+    uint64_t size{0},time{0};
+    double   mbytesPerSecond{0.};
+  };
+
+  static IoPerf lastPerf;
   static int latticeWriteMaxRetry;
 
   /////////////////////////////////////////////////////////////////////////////
@@ -502,12 +509,15 @@ class BinaryIO {
       timer.Stop();
     }
     
+    lastPerf.size            = sizeof(fobj)*iodata.size()*nrank;
+    lastPerf.time            = timer.useconds();
+    lastPerf.mbytesPerSecond = lastPerf.size/1024./1024./(lastPerf.time/1.0e6);
     std::cout<<GridLogMessage<<"IOobject: ";
     if ( control & BINARYIO_READ) std::cout << " read  ";
     else                          std::cout << " write ";
     uint64_t bytes = sizeof(fobj)*iodata.size()*nrank;
-    std::cout<< bytes <<" bytes in "<<timer.Elapsed() <<" "
-	     << (double)bytes/ (double)timer.useconds() <<" MB/s "<<std::endl;
+    std::cout<< lastPerf.size <<" bytes in "<< timer.Elapsed() <<" "
+	     << lastPerf.mbytesPerSecond <<" MB/s "<<std::endl;
 
     std::cout<<GridLogMessage<<"IOobject: endian and checksum overhead "<<bstimer.Elapsed()  <<std::endl;
 
@@ -663,10 +673,15 @@ class BinaryIO {
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     timer.Start();
-    thread_for(lidx,lsites,{
+    thread_for(lidx,lsites,{  // FIX ME, suboptimal implementation
       std::vector<RngStateType> tmp(RngStateCount);
       std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
-      parallel_rng.SetState(tmp,lidx);
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(lidx, lcoor);
+      int o_idx=grid->oIndex(lcoor);
+      int i_idx=grid->iIndex(lcoor);
+      int gidx=parallel_rng.generator_idx(o_idx,i_idx);
+      parallel_rng.SetState(tmp,gidx);
       });
     timer.Stop();
 
@@ -723,7 +738,12 @@ class BinaryIO {
     std::vector<RNGstate> iodata(lsites);
     thread_for(lidx,lsites,{
       std::vector<RngStateType> tmp(RngStateCount);
-      parallel_rng.GetState(tmp,lidx);
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(lidx, lcoor);
+      int o_idx=grid->oIndex(lcoor);
+      int i_idx=grid->iIndex(lcoor);
+      int gidx=parallel_rng.generator_idx(o_idx,i_idx);
+      parallel_rng.GetState(tmp,gidx);
       std::copy(tmp.begin(),tmp.end(),iodata[lidx].begin());
     });
     timer.Stop();

Grid/parallelIO/IldgIO.h

@@ -576,6 +576,8 @@ 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;
       }
@@ -619,12 +621,12 @@ class IldgWriter : public ScidacWriter {
   // Don't require scidac records EXCEPT checksum
   // Use Grid MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
-  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
+  template <class stats = PeriodicGaugeStatistics>
+  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
-    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef Lattice<vLorentzColourMatrixD> GaugeField;
+    typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -636,6 +638,9 @@ class IldgWriter : public ScidacWriter {
 
     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
 
+    stats Stats;
+    Stats(Umu,header);
+    
     std::string format = header.floating_point;
     header.ensemble_id    = description;
     header.ensemble_label = description;
@@ -705,10 +710,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
-  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
+  template <class stats = PeriodicGaugeStatistics>
+  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
 
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef Lattice<vLorentzColourMatrixD > GaugeField;
     typedef typename GaugeField::vector_object  vobj;
     typedef typename vobj::scalar_object sobj;
 
@@ -921,7 +926,8 @@ class IldgReader : public GridLimeReader {
 
     if ( found_FieldMetaData || found_usqcdInfo ) {
       FieldMetaData checker;
-      GaugeStatistics(Umu,checker);
+      stats Stats;
+      Stats(Umu,checker);
       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;

Grid/parallelIO/MetaData.h

@@ -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;
 
@@ -176,29 +176,18 @@ template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMet
   GridMetaData(grid,header); 
   MachineCharacteristics(header);
 }
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
+template<class Impl>
+class GaugeStatistics
 {
-  // How to convert data precision etc...
-  header.link_trace=WilsonLoops<PeriodicGimplF>::linkTrace(data);
-  header.plaquette =WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
-}
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
+public:
+  void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
   {
-  // How to convert data precision etc...
-  header.link_trace=WilsonLoops<PeriodicGimplD>::linkTrace(data);
-  header.plaquette =WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
-}
-template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header)
-{
-   
-  GridBase *grid = field.Grid();
-  std::string format = getFormatString<vLorentzColourMatrixF>();
-  header.floating_point = format;
-  header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
-  GridMetaData(grid,header); 
-  GaugeStatistics(field,header);
-  MachineCharacteristics(header);
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
   }
+};
+typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
+typedef GaugeStatistics<ConjugateGimplD> ConjugateGaugeStatistics;
 template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
 {
   GridBase *grid = field.Grid();
@@ -206,7 +195,6 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
   header.floating_point = format;
   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
   GridMetaData(grid,header); 
-  GaugeStatistics(field,header);
   MachineCharacteristics(header);
 }
 

Grid/parallelIO/NerscIO.h

@@ -40,6 +40,8 @@ using namespace Grid;
 class NerscIO : public BinaryIO { 
 public:
 
+  typedef Lattice<vLorentzColourMatrixD> GaugeField;
+
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -129,12 +131,12 @@ public:
   // Now the meat: the object readers
   /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-  template<class vsimd>
-  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+  template<class GaugeStats=PeriodicGaugeStatistics>
+  static inline void readConfiguration(GaugeField &Umu,
 				       FieldMetaData& header,
-				       std::string file)
+				       std::string file,
+				       GaugeStats GaugeStatisticsCalculator=GaugeStats())
   {
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
     GridBase *grid = Umu.Grid();
     uint64_t offset = readHeader(file,Umu.Grid(),header);
@@ -153,23 +155,23 @@ public:
     // munger is a function of <floating point, Real, data_type>
     if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F> 
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D> 
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3D> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
     } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixD>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
@@ -177,7 +179,7 @@ public:
       assert(0);
     }
 
-    GaugeStatistics(Umu,clone);
+    GaugeStats Stats; Stats(Umu,clone);
 
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl;
@@ -203,15 +205,22 @@ public:
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
-  template<class vsimd>
-  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+  // Preferred interface
+  template<class GaugeStats=PeriodicGaugeStatistics>
+  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
+					std::string file, 
+					std::string ens_label = std::string("DWF"))
+  {
+    writeConfiguration(Umu,file,0,1,ens_label);
+  }
+  template<class GaugeStats=PeriodicGaugeStatistics>
+  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
-					int bits32)
+					int bits32,
+					std::string ens_label = std::string("DWF"))
   {
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
-
-    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
@@ -219,8 +228,8 @@ public:
     // Following should become arguments
     ///////////////////////////////////////////
     header.sequence_number = 1;
-    header.ensemble_id     = "UKQCD";
-    header.ensemble_label  = "DWF";
+    header.ensemble_id     = std::string("UKQCD");
+    header.ensemble_label  = ens_label;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -229,7 +238,7 @@ public:
 
     GridMetaData(grid,header);
     assert(header.nd==4);
-    GaugeStatistics(Umu,header);
+    GaugeStats Stats; Stats(Umu,header);
     MachineCharacteristics(header);
 
     uint64_t offset;

Grid/parallelIO/OpenQcdIO.h

@@ -154,7 +154,7 @@ public:
     grid->Barrier(); timer.Stop();
     std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: redistribute overhead " << timer.Elapsed() << std::endl;
 
-    GaugeStatistics(Umu, clone);
+    PeriodicGaugeStatistics Stats; Stats(Umu, clone);
 
     RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
 

Grid/parallelIO/OpenQcdIOChromaReference.h

@@ -208,7 +208,7 @@ public:
 
     FieldMetaData clone(header);
 
-    GaugeStatistics(Umu, clone);
+    PeriodicGaugeStatistics Stats; Stats(Umu, clone);
 
     RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
 

Grid/qcd/QCD.h

@@ -47,7 +47,7 @@ static constexpr int Ym = 5;
 static constexpr int Zm = 6;
 static constexpr int Tm = 7;
 
-static constexpr int Nc=3;
+static constexpr int Nc=Config_Nc;
 static constexpr int Ns=4;
 static constexpr int Nd=4;
 static constexpr int Nhs=2; // half spinor
@@ -80,6 +80,13 @@ template<typename T> struct isSpinor {
 template <typename T> using IfSpinor    = Invoke<std::enable_if< isSpinor<T>::value,int> > ;
 template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ;
 
+const int CoarseIndex = 4;
+template<typename T> struct isCoarsened {
+   static constexpr bool value = (CoarseIndex<=T::TensorLevel);
+};
+template <typename T> using IfCoarsened    = Invoke<std::enable_if< isCoarsened<T>::value,int> > ;
+template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ;
+
 // ChrisK very keen to add extra space for Gparity doubling.
 //
 // Also add domain wall index, in a way where Wilson operator 

Grid/qcd/action/ActionBase.h

@@ -41,7 +41,7 @@ class Action
 public:
   bool is_smeared = false;
   // Heatbath?
-  virtual void refresh(const GaugeField& U, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
+  virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
   virtual std::string action_name()    = 0;                             // return the action name

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -0,0 +1,240 @@
+/*************************************************************************************
+
+    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>
+
+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 CompactWilsonCloverFermion : public WilsonFermion<Impl>,
+                                   public WilsonCloverHelpers<Impl>,
+                                   public CompactWilsonCloverHelpers<Impl> {
+  /////////////////////////////////////////////
+  // Sizes
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_COMPACT_CLOVER_SIZES(Impl);
+
+  /////////////////////////////////////////////
+  // Type definitions
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  typedef WilsonFermion<Impl>              WilsonBase;
+  typedef WilsonCloverHelpers<Impl>        Helpers;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  /////////////////////////////////////////////
+  // Constructors
+  /////////////////////////////////////////////
+
+public:
+
+  CompactWilsonCloverFermion(GaugeField& _Umu,
+			    GridCartesian& Fgrid,
+			    GridRedBlackCartesian& Hgrid,
+			    const RealD _mass,
+			    const RealD _csw_r = 0.0,
+			    const RealD _csw_t = 0.0,
+			    const RealD _cF = 1.0,
+			    const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
+			    const ImplParams& impl_p = ImplParams());
+
+  /////////////////////////////////////////////
+  // Member functions (implementing interface)
+  /////////////////////////////////////////////
+
+public:
+
+  virtual void Instantiatable() {};
+  int          ConstEE()     override { return 0; };
+  int          isTrivialEE() override { return 0; };
+
+  void Dhop(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopOE(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopEO(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
+
+  void M(const FermionField& in, FermionField& out) override;
+
+  void Mdag(const FermionField& in, FermionField& out) override;
+
+  void Meooe(const FermionField& in, FermionField& out) override;
+
+  void MeooeDag(const FermionField& in, FermionField& out) override;
+
+  void Mooee(const FermionField& in, FermionField& out) override;
+
+  void MooeeDag(const FermionField& in, FermionField& out) override;
+
+  void MooeeInv(const FermionField& in, FermionField& out) override;
+
+  void MooeeInvDag(const FermionField& in, FermionField& out) override;
+
+  void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
+
+  void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
+
+  void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  /////////////////////////////////////////////
+  // Member functions (internals)
+  /////////////////////////////////////////////
+
+  void MooeeInternal(const FermionField&        in,
+                     FermionField&              out,
+                     const CloverDiagonalField& diagonal,
+                     const CloverTriangleField& triangle);
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+  void ImportGauge(const GaugeField& _Umu) override;
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+private:
+
+  template<class Field>
+  const MaskField* getCorrectMaskField(const Field &in) const {
+    if(in.Grid()->_isCheckerBoarded) {
+      if(in.Checkerboard() == Odd) {
+        return &this->BoundaryMaskOdd;
+      } else {
+        return &this->BoundaryMaskEven;
+      }
+    } else {
+      return &this->BoundaryMask;
+    }
+  }
+
+  template<class Field>
+  void ApplyBoundaryMask(Field& f) {
+    const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
+    assert(m != nullptr);
+    CompactHelpers::ApplyBoundaryMask(f, *m);
+  }
+
+  /////////////////////////////////////////////
+  // Member Data
+  /////////////////////////////////////////////
+
+public:
+
+  RealD csw_r;
+  RealD csw_t;
+  RealD cF;
+
+  bool open_boundaries;
+
+  CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
+  CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
+
+  CloverTriangleField Triangle,    TriangleEven,    TriangleOdd;
+  CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
+
+  FermionField Tmp;
+
+  MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -53,6 +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;
@@ -153,46 +154,63 @@ typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoInd
 typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
 typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
+// Compact Clover fermions
+typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
+typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
+
+typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
+typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
+
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
+
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef CompactWilsonCloverFermion<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 +253,49 @@ 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 +309,6 @@ 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

@@ -153,8 +153,8 @@ public:
   typedef typename Impl::StencilImpl             StencilImpl;		\
   typedef typename Impl::ImplParams               ImplParams;	        \
   typedef typename Impl::StencilImpl::View_type  StencilView;		\
-  typedef typename ViewMap<FermionField>::Type      FermionFieldView;	\
-  typedef typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
+  typedef const typename ViewMap<FermionField>::Type      FermionFieldView;	\
+  typedef const typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
 
 #define INHERIT_IMPL_TYPES(Base)		\
   INHERIT_GIMPL_TYPES(Base)			\
@@ -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

@@ -97,42 +97,30 @@ public:
     Coordinate icoor;
 
 #ifdef GRID_SIMT
-    _Spinor tmp;
-
     const int Nsimd =SiteDoubledGaugeField::Nsimd();
     int s = acceleratorSIMTlane(Nsimd);
     St.iCoorFromIindex(icoor,s);
 
     int mmu = mu % Nd;
-    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
 
+    auto UU0=coalescedRead(U(0)(mu));
+    auto UU1=coalescedRead(U(1)(mu));
+    
+    //Decide whether we do a G-parity flavor twist
+    //Note: this assumes (but does not check) that sl==1 || sl==2 i.e. max 2 SIMD lanes in G-parity dir
+    //It also assumes (but does not check) that abs(distance) == 1
     int permute_lane = (sl==1) 
     || ((distance== 1)&&(icoor[direction]==1))
     || ((distance==-1)&&(icoor[direction]==0));
 
-      if ( permute_lane ) { 
-	tmp(0) = chi(1);
-	tmp(1) = chi(0);
-      } else {
-	tmp(0) = chi(0);
-	tmp(1) = chi(1);
-      }
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
 
-      auto UU0=coalescedRead(U(0)(mu));
-      auto UU1=coalescedRead(U(1)(mu));
+    //Apply the links
+    int f_upper = permute_lane ? 1 : 0;
+    int f_lower = !f_upper;
 
-      mult(&phi(0),&UU0,&tmp(0));
-      mult(&phi(1),&UU1,&tmp(1));
-
-    } else {
-
-      auto UU0=coalescedRead(U(0)(mu));
-      auto UU1=coalescedRead(U(1)(mu));
-
-      mult(&phi(0),&UU0,&chi(0));
-      mult(&phi(1),&UU1,&chi(1));
-
-    }
+    mult(&phi(0),&UU0,&chi(f_upper));
+    mult(&phi(1),&UU1,&chi(f_lower));
 
 #else
     typedef _Spinor vobj;
@@ -339,8 +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/ImprovedStaggeredFermion5D.h

@@ -208,7 +208,7 @@ public:
   LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
-  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
+  //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
     
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities

Grid/qcd/action/fermion/MADWF.h

@@ -85,7 +85,7 @@ class MADWF
       maxiter     =_maxiter;
     };
    
-  void operator() (const FermionFieldo &src4,FermionFieldo &sol5)
+  void operator() (const FermionFieldo &src,FermionFieldo &sol5)
   {
     std::cout << GridLogMessage<< " ************************************************" << std::endl;
     std::cout << GridLogMessage<< "  MADWF-like algorithm                           " << std::endl;
@@ -114,8 +114,16 @@ class MADWF
     ///////////////////////////////////////
     //Import source, include Dminus factors
     ///////////////////////////////////////
-    Mato.ImportPhysicalFermionSource(src4,b); 
-    std::cout << GridLogMessage << " src4 " <<norm2(src4)<<std::endl;
+    GridBase *src_grid = src.Grid();
+
+    assert( (src_grid == Mato.GaugeGrid()) || (src_grid == Mato.FermionGrid()));
+
+    if ( src_grid == Mato.GaugeGrid() ) {
+      Mato.ImportPhysicalFermionSource(src,b);
+    } else {
+      b=src;
+    }
+    std::cout << GridLogMessage << " src " <<norm2(src)<<std::endl;
     std::cout << GridLogMessage << " b    " <<norm2(b)<<std::endl;
 
     defect = b;

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/StaggeredKernels.h

@@ -63,17 +63,20 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    ///////////////////////////////////////////////////////////////////////////////////////
    // Generic Nc kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   template<int Naik> accelerator_inline
+   template<int Naik> 
+   static accelerator_inline
    void DhopSiteGeneric(StencilView &st, 
 			DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   template<int Naik> accelerator_inline
+   
+   template<int Naik> static accelerator_inline
    void DhopSiteGenericInt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
-   template<int Naik> accelerator_inline
+   
+   template<int Naik> static accelerator_inline
    void DhopSiteGenericExt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 			   SiteSpinor * buf, int LLs, int sU, 
@@ -82,17 +85,20 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    ///////////////////////////////////////////////////////////////////////////////////////
    // Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   template<int Naik> accelerator_inline
+   
+   template<int Naik> static accelerator_inline
    void DhopSiteHand(StencilView &st, 
 		     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		     SiteSpinor * buf, int LLs, int sU, 
 		     const FermionFieldView &in, FermionFieldView &out,int dag);
-   template<int Naik> accelerator_inline
+   
+   template<int Naik> static accelerator_inline
    void DhopSiteHandInt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   template<int Naik> accelerator_inline
+   
+   template<int Naik> static accelerator_inline
    void DhopSiteHandExt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
@@ -101,6 +107,7 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    ///////////////////////////////////////////////////////////////////////////////////////
    // Asm Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
+   
    void DhopSiteAsm(StencilView &st, 
 		    DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		    SiteSpinor * buf, int LLs, int sU, 

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 +30,8 @@
 
 #pragma once
 
-#include <Grid/Grid.h>
+#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -50,18 +52,15 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////
 
 template <class Impl>
-class WilsonCloverFermion : public WilsonFermion<Impl>
+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 WilsonCloverHelpers<Impl> Helpers;
 
   virtual int    ConstEE(void)     { return 0; };
   virtual void Instantiatable(void){};
@@ -72,42 +71,7 @@ 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 +88,21 @@ 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;
-  }
-
-private:
+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 +1,761 @@
+/*************************************************************************************
+
+    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 +1,92 @@
+/*************************************************************************************
+
+    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_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
+
+  static constexpr int Nred      = Nc * Nhs;        // 6
+  static constexpr int Nblock    = Nhs;             // 2
+  static constexpr int Ndiagonal = Nred;            // 6
+  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

@@ -61,18 +61,19 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) {
+  accelerator_inline int CommDatumSize(void) const {
     return sizeof(SiteHalfCommSpinor);
   }
 
   /*****************************************************/
   /* 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) {
-    _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);
   }
 
   /*****************************************************/
@@ -81,19 +82,24 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
 				   const SiteHalfSpinor * __restrict__ vp0,
 				   const SiteHalfSpinor * __restrict__ vp1,
-				   Integer type,Integer o){
+				   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 */
   /*****************************************************/
   accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
-				     SiteHalfSpinor * __restrict__ in, Integer o) {    
+				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
     assert(0);
   }
 
@@ -103,8 +109,30 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
 					   SiteHalfSpinor * __restrict__ out1,
 					   const SiteSpinor * __restrict__ in,
-					   Integer j,Integer k, Integer m,Integer type)
+					   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;
+
+    unsigned int Nsimd = vobj::Nsimd();
+    unsigned int mask = Nsimd >> (type + 1);
+    int lane = acceleratorSIMTlane(Nsimd);
+    int j0 = lane &(~mask); // inner coor zero
+    int j1 = lane |(mask) ; // inner coor one
+    const vobj *vp0 = &in[k];
+    const vobj *vp1 = &in[m];
+    const vobj *vp = (lane&mask) ? vp1:vp0;
+    auto sa = coalescedRead(*vp,j0);
+    auto sb = coalescedRead(*vp,j1);
+    hsobj psa, psb;
+    projector::Proj(psa,sa,mu,dag);
+    projector::Proj(psb,sb,mu,dag);
+    coalescedWrite(out0[j],psa);
+    coalescedWrite(out1[j],psb);
+#else
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
     projector::Proj(temp1,in[k],mu,dag);
@@ -112,15 +140,17 @@ public:
     exchange(temp3,temp4,temp1,temp2,type);
     vstream(out0[j],temp3);
     vstream(out1[j],temp4);
+#endif
   }
 
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) { return false; }
+  accelerator_inline bool DecompressionStep(void) const { return false; }
 
 };
 
+#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 >
@@ -142,20 +172,30 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) {
+  accelerator_inline int CommDatumSize(void) const {
     return sizeof(SiteHalfCommSpinor);
   }
 
   /*****************************************************/
   /* 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) {
-    _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 */
@@ -163,7 +203,7 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
                        SiteHalfSpinor *vp0,
                        SiteHalfSpinor *vp1,
-		       Integer type,Integer o){
+		       Integer type,Integer o) const {
     SiteHalfSpinor vt0,vt1;
     SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
     SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
@@ -175,7 +215,7 @@ public:
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o){
+  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
     SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
     precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
   }
@@ -186,7 +226,7 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
 			       SiteHalfSpinor *out1,
 			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type){
+			       Integer j,Integer k, Integer m,Integer type) const {
     SiteHalfSpinor temp1, temp2,temp3,temp4;
     SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
     SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
@@ -200,9 +240,10 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) { return true; }
+  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/WilsonFermion.h

@@ -74,6 +74,20 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
+
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent
@@ -196,5 +210,3 @@ typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
 NAMESPACE_END(Grid);
-
-

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -215,7 +215,7 @@ public:
   LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
-  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
+  //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
 
 
 };

Grid/qcd/action/fermion/WilsonImpl.h

@@ -72,7 +72,7 @@ public:
   typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
   typedef WilsonImplParams ImplParams;
   typedef WilsonStencil<SiteSpinor, SiteHalfSpinor,ImplParams> StencilImpl;
-  typedef typename StencilImpl::View_type StencilView;
+  typedef const typename StencilImpl::View_type StencilView;
     
   ImplParams Params;
 
@@ -106,11 +106,15 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
+    const int Nsimd = SiteHalfSpinor::Nsimd();
     autoView( out_v, out, AcceleratorWrite);
     autoView( phi_v, phi, AcceleratorRead);
     autoView( Umu_v, Umu, AcceleratorRead);
-    accelerator_for(sss,out.Grid()->oSites(),1,{
-	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
+    typedef decltype(coalescedRead(out_v[0]))   calcSpinor;
+    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
+	calcSpinor tmp;
+	multLink(tmp,Umu_v[sss],phi_v(sss),mu);
+	coalescedWrite(out_v[sss],tmp);
     });
   }
 					   
@@ -180,18 +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);
@@ -204,6 +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    
   }
 };
 
@@ -212,17 +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/WilsonKernels.h

@@ -49,9 +49,17 @@ public:
 
   INHERIT_IMPL_TYPES(Impl);
   typedef FermionOperator<Impl> Base;
-   
+  typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;   
 public:
 
+#ifdef GRID_SYCL
+#define SYCL_HACK
+#endif  
+#ifdef SYCL_HACK
+  static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
+			       int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
+#endif
+  
   static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,
 			 int interior=1,int exterior=1) ;

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

@@ -642,7 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -799,7 +799,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
-  LatticeInteger zz (UGrid);   zz=0.0;
+  PropagatorField zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
   for (int s=0;s<Ls;s++) {
 
@@ -826,8 +826,9 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
+  unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
   ////////////////////////////////////////////////
@@ -850,7 +851,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
 
-  LatticeInteger zz (UGrid);   zz=0.0;
+  PropagatorField  zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
 
   for(int s=0;s<Ls;s++){
@@ -880,17 +881,29 @@ 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 +920,7 @@ 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 +935,13 @@ 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 +1,363 @@
+/*************************************************************************************
+
+    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>
+CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
+                                                             GridCartesian& Fgrid,
+                                                             GridRedBlackCartesian& Hgrid,
+                                                             const RealD _mass,
+                                                             const RealD _csw_r,
+                                                             const RealD _csw_t,
+                                                             const RealD _cF,
+                                                             const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                                             const ImplParams& impl_p)
+  : WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
+  , csw_r(_csw_r)
+  , csw_t(_csw_t)
+  , cF(_cF)
+  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , Diagonal(&Fgrid),        Triangle(&Fgrid)
+  , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
+  , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
+  , DiagonalInv(&Fgrid),     TriangleInv(&Fgrid)
+  , DiagonalInvEven(&Hgrid), TriangleInvEven(&Hgrid)
+  , DiagonalInvOdd(&Hgrid),  TriangleInvOdd(&Hgrid)
+  , Tmp(&Fgrid)
+  , BoundaryMask(&Fgrid)
+  , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
+{
+  csw_r *= 0.5;
+  csw_t *= 0.5;
+  if (clover_anisotropy.isAnisotropic)
+    csw_r /= clover_anisotropy.xi_0;
+
+  ImportGauge(_Umu);
+  if (open_boundaries)
+    CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::Dhop(in, out, dag);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::DhopOE(in, out, dag);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::DhopEO(in, out, dag);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+  WilsonBase::DhopDir(in, out, dir, disp);
+  if(this->open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+  WilsonBase::DhopDirAll(in, out);
+  if(this->open_boundaries) {
+    for(auto& o : out) ApplyBoundaryMask(o);
+  }
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
+  out.Checkerboard() = in.Checkerboard();
+  WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
+  Mooee(in, Tmp);
+  axpy(out, 1.0, out, Tmp);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
+  out.Checkerboard() = in.Checkerboard();
+  WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
+  MooeeDag(in, Tmp);
+  axpy(out, 1.0, out, Tmp);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
+  WilsonBase::Meooe(in, out);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
+  WilsonBase::MeooeDag(in, out);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
+  if(in.Grid()->_isCheckerBoarded) {
+    if(in.Checkerboard() == Odd) {
+      MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
+    } else {
+      MooeeInternal(in, out, DiagonalEven, TriangleEven);
+    }
+  } else {
+    MooeeInternal(in, out, Diagonal, Triangle);
+  }
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
+  Mooee(in, out); // blocks are hermitian
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
+  if(in.Grid()->_isCheckerBoarded) {
+    if(in.Checkerboard() == Odd) {
+      MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
+    } else {
+      MooeeInternal(in, out, DiagonalInvEven, TriangleInvEven);
+    }
+  } else {
+    MooeeInternal(in, out, DiagonalInv, TriangleInv);
+  }
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
+  MooeeInv(in, out); // blocks are hermitian
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+  DhopDir(in, out, dir, disp);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+  DhopDirAll(in, out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+  assert(!open_boundaries); // TODO check for changes required for open bc
+
+  // 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*Helpers::Cmunu(U, lambda, mu, nu);   // checked
+      count++;
+    }
+
+    pokeLorentz(clover_force, U[mu] * force_mu, mu);
+  }
+  //clover_force *= csw;
+  force += clover_force;
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+  assert(0);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+  assert(0);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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>
+void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
+  // NOTE: parts copied from original implementation
+
+  // Import gauge into base class
+  double t0 = usecond();
+  WilsonBase::ImportGauge(_Umu); // NOTE: called here and in wilson constructor -> performed twice, but can't avoid that
+
+  // Initialize temporary variables
+  double t1 = usecond();
+  conformable(_Umu.Grid(), this->GaugeGrid());
+  GridBase* grid = _Umu.Grid();
+  typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
+  CloverField TmpOriginal(grid);
+
+  // Compute the field strength terms mu>nu
+  double t2 = usecond();
+  WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
+  WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
+  WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
+
+  // Compute the Clover Operator acting on Colour and Spin
+  // multiply here by the clover coefficients for the anisotropy
+  double t3 = usecond();
+  TmpOriginal  = Helpers::fillCloverYZ(Bx) * csw_r;
+  TmpOriginal += Helpers::fillCloverXZ(By) * csw_r;
+  TmpOriginal += Helpers::fillCloverXY(Bz) * csw_r;
+  TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
+  TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
+  TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
+  TmpOriginal += this->diag_mass;
+
+  // Convert the data layout of the clover term
+  double t4 = usecond();
+  CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
+
+  // Possible modify the boundary values
+  double t5 = usecond();
+  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+
+  // Invert the clover term in the improved layout
+  double t6 = usecond();
+  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+
+  // Fill the remaining clover fields
+  double t7 = 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 t8 = usecond();
+#if 0
+  std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
+            << ", allocations = "               << (t2 - t1) / 1e6
+            << ", field strength = "            << (t3 - t2) / 1e6
+            << ", fill clover = "               << (t4 - t3) / 1e6
+            << ", convert = "                   << (t5 - t4) / 1e6
+            << ", boundaries = "                << (t6 - t5) / 1e6
+            << ", inversions = "                << (t7 - t6) / 1e6
+            << ", pick cbs = "                  << (t8 - t7) / 1e6
+            << ", total = "                     << (t8 - t0) / 1e6
+            << std::endl;
+#endif
+}
+
+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=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);				\
@@ -140,13 +158,35 @@ 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>
+template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
 					  DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
 					  SiteSpinor *buf, int sF, int sU, 
@@ -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;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  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,local,perm, ptype;
 
   StencilEntry *SE;
@@ -215,13 +241,13 @@ 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);
   }
 }
 
 
 template <class Impl>
-template <int Naik>
+template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st, 
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 					     SiteSpinor *buf, int sF, int sU, 
@@ -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);  
@@ -294,13 +305,13 @@ 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);
   }
 }
 
 
 template <class Impl>
-template <int Naik>
+template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 					     SiteSpinor *buf, int sF, int sU, 
@@ -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);			\
+    int perm= SE->_permute;						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
   } else {							\
-      chi_p = &in[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  } else {							\
-    chi_p = &buf[SE->_offset];					\
-  }								\
-  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>
@@ -78,18 +71,20 @@ StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
 // Int, Ext, Int+Ext cases for comms overlap
 ////////////////////////////////////////////////////////////////////////////////////
 template <class Impl>
-template <int Naik>
+template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGeneric(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;
+  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);
   }
 };
 
@@ -126,17 +121,20 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
   // Only contributions from interior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-template <int Naik>
+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);
   }
 };
 
@@ -174,18 +172,21 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
   // Only contributions from exterior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-template <int Naik>
+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;
@@ -212,10 +213,11 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
     GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
     }
     if ( nmu ) {
+      auto _out = coalescedRead(out[sF],lane);
       if ( dag ) {
-	out[sF] = out[sF] - Uchi;
+	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
@@ -33,6 +34,45 @@
 
 NAMESPACE_BEGIN(Grid);
 
+template<class Impl>
+WilsonCloverFermion<Impl>::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)
@@ -67,10 +107,13 @@ void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 template <class Impl>
 void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
+  double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
+  double t1 = usecond();
   GridBase *grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
 
+  double t2 = usecond();
   // Compute the field strength terms mu>nu
   WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
   WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
@@ -79,33 +122,32 @@ 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  = 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;
   CloverTerm += diag_mass;
 
+  double t4 = usecond();
   int lvol = _Umu.Grid()->lSites();
   int DimRep = Impl::Dimension;
 
-  Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-  Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-
-  Coordinate lcoor;
-  typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
-
+  double t5 = usecond();
   {
     autoView(CTv,CloverTerm,CpuRead);
     autoView(CTIv,CloverTermInv,CpuWrite);
-    for (int site = 0; site < lvol; site++) {
+    thread_for(site, lvol, {
+      Coordinate lcoor;
       grid->LocalIndexToLocalCoor(site, lcoor);
-      EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      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);
-      Qxinv = Zero();
       //if (csw!=0){
       for (int j = 0; j < Ns; j++)
 	for (int k = 0; k < Ns; k++)
@@ -126,21 +168,36 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
       //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
       //  }
       pokeLocalSite(Qxinv, CTIv, lcoor);
-    }
+    });
   }
 
+  double t6 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
 
-  pickCheckerboard(Even, CloverTermDagEven, closure(adj(CloverTerm)));
-  pickCheckerboard(Odd, CloverTermDagOdd, closure(adj(CloverTerm)));
+  pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
+  pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
 
   pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
   pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
 
-  pickCheckerboard(Even, CloverTermInvDagEven, closure(adj(CloverTermInv)));
-  pickCheckerboard(Odd, CloverTermInvDagOdd, closure(adj(CloverTermInv)));
+  pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
+  pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
+  double t7 = usecond();
+
+#if 0
+  std::cout << GridLogMessage << "WilsonCloverFermion::ImportGauge timings:"
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
+            << ", allocations = "               << (t2 - t1) / 1e6
+            << ", field strength = "            << (t3 - t2) / 1e6
+            << ", fill clover = "               << (t4 - t3) / 1e6
+            << ", misc = "                      << (t5 - t4) / 1e6
+            << ", inversions = "                << (t6 - t5) / 1e6
+            << ", pick cbs = "                  << (t7 - t6) / 1e6
+            << ", total = "                     << (t7 - t0) / 1e6
+            << std::endl;
+#endif
 }
 
 template <class Impl>
@@ -171,7 +228,7 @@ template <class Impl>
 void WilsonCloverFermion<Impl>::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)
@@ -186,12 +243,12 @@ 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
@@ -209,18 +266,98 @@ 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>
+void WilsonCloverFermion<Impl>::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*Helpers::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>

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

@@ -75,6 +75,91 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
   StencilOdd.BuildSurfaceList(1,vol4);
 }
 
+template<class Impl>
+void WilsonFermion<Impl>::Report(void)
+{
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+  RealD volume = 1;
+  Coordinate latt = _grid->GlobalDimensions();
+  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of DhopEO Calls   : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion TotalTime   /Calls        : " << DhopTotalTime   / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime    /Calls        : " << DhopCommTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion FaceTime    /Calls        : " << DhopFaceTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime1/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime2/Calls        : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
+
+    // Average the compute time
+    _grid->GlobalSum(DhopComputeTime);
+    DhopComputeTime/=NP;
+    RealD mflops = 1320*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+
+    RealD Fullmflops = 1320*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+
+    // how to count flops here?
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call               ? : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node      ? : " << mflops/NP << std::endl;
+
+    // how to count flops here?
+    RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)        ? : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full) ? : " << Fullmflops/NP << std::endl;  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil"    <<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+  }
+  if ( DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil     Reporti()"    <<std::endl;  Stencil.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilEven Reporti()"<<std::endl;  StencilEven.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd  Reporti()" <<std::endl;  StencilOdd.Reporti(DhopCalls);
+  }
+}
+
+template<class Impl>
+void WilsonFermion<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0; // ok
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopComputeTime2= 0;
+  DhopFaceTime    = 0;
+  DhopTotalTime   = 0;
+
+  DerivCalls       = 0; // ok
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+  Stencil.ZeroCountersi();
+  StencilEven.ZeroCountersi();
+  StencilOdd.ZeroCountersi();
+}
+
+
 template <class Impl>
 void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
@@ -234,6 +319,7 @@ template <class Impl>
 void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
                                         GaugeField &mat, const FermionField &A,
                                         const FermionField &B, int dag) {
+  DerivCalls++;
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -242,8 +328,11 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
   FermionField Atilde(B.Grid());
   Atilde = A;
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B, compressor);
+  DerivCommTime+=usecond();
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma (1+g)<->(1-g) if dag
@@ -251,6 +340,7 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     int gamma = mu;
     if (!dag) gamma += Nd;
 
+    DerivDhopComputeTime -= usecond();
     int Ls=1;
     Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, B.Grid()->oSites(), B, Btilde, mu, gamma);
 
@@ -258,7 +348,9 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     // spin trace outer product
     //////////////////////////////////////////////////
     Impl::InsertForce4D(mat, Btilde, Atilde, mu);
+    DerivDhopComputeTime += usecond();
   }
+  DerivComputeTime += usecond();
 }
 
 template <class Impl>
@@ -305,6 +397,7 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
 template <class Impl>
 void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -316,6 +409,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 template <class Impl>
 void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -328,6 +422,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
 template <class Impl>
 void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -392,13 +487,14 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
+  DhopTotalTime-=usecond();
 #ifdef GRID_OMP
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
 #endif
     DhopInternalSerial(st,lo,U,in,out,dag);
-
+  DhopTotalTime+=usecond();
 }
 
 template <class Impl>
@@ -417,38 +513,53 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
   /////////////////////////////
   std::vector<std::vector<CommsRequest_t> > requests;
   st.Prepare();
+  DhopFaceTime-=usecond();
   st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
+
+  DhopCommTime -=usecond();
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);
+  DhopFaceTime+=usecond();
 
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt;
+  DhopComputeTime-=usecond();
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   } else {
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   }
+  DhopComputeTime+=usecond();
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
+  DhopCommTime   +=usecond();
+
+  DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
+
+  DhopComputeTime2-=usecond();
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   } else {
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   }
+  DhopComputeTime2+=usecond();
 };
 
 
@@ -460,14 +571,18 @@ void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
 {
   assert((dag == DaggerNo) || (dag == DaggerYes));
   Compressor compressor(dag);
+  DhopCommTime-=usecond();
   st.HaloExchange(in, compressor);
+  DhopCommTime+=usecond();
 
+  DhopComputeTime-=usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   } else {
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   }
+  DhopComputeTime+=usecond();
 };
 /*Change ends */
 

Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h

@@ -0,0 +1,450 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+
+
+    Source file: ./lib/qcd/action/fermion/WilsonKernelsAsmA64FX.h
+
+    Copyright (C) 2020
+
+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 */
+#pragma once
+
+//#if defined(A64FXASM)
+#if defined(A64FX)
+
+// safety include
+#include <arm_sve.h>
+
+// undefine everything related to kernels
+#include <simd/Fujitsu_A64FX_undef.h>
+
+
+    ///////////////////////////////////////////////////////////
+    // If we are A64FX specialise the single precision routine
+    ///////////////////////////////////////////////////////////
+#if defined(DSLASHINTRIN)
+//#pragma message ("A64FX Dslash: intrin")
+#include <simd/Fujitsu_A64FX_intrin_single.h>
+#else
+#pragma message ("A64FX Dslash: asm")
+#include <simd/Fujitsu_A64FX_asm_single.h>
+#endif
+
+/// Switch off the 5d vectorised code optimisations
+#undef DWFVEC5D
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, undag Kernel, single
+/////////////////////////////////////////////////////////////////
+#undef KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplF>::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<ZWilsonImplF>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#define INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplF>::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<ZWilsonImplF>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#define EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplF>::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<ZWilsonImplF>::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>
+
+
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, dag Kernel, single
+/////////////////////////////////////////////////////////////////
+#define KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplF>::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<ZWilsonImplF>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#define INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplF>::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<ZWilsonImplF>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#define EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplF>::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<ZWilsonImplF>::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>
+
+
+
+// undefine
+#include <simd/Fujitsu_A64FX_undef.h>
+
+///////////////////////////////////////////////////////////
+// If we are A64FX specialise the double precision routine
+///////////////////////////////////////////////////////////
+
+#if defined(DSLASHINTRIN)
+#include <simd/Fujitsu_A64FX_intrin_double.h>
+#else
+#include <simd/Fujitsu_A64FX_asm_double.h>
+#endif
+
+// former KNL
+//#define MAYBEPERM(A,perm) if (perm) { A ; }
+//#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
+//#define COMPLEX_SIGNS(isigns) vComplexD *isigns = &signsD[0];
+
+
+#define INTERIOR_AND_EXTERIOR
+#undef  INTERIOR
+#undef  EXTERIOR
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, undag Kernel, double
+/////////////////////////////////////////////////////////////////
+#undef KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplD>::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<ZWilsonImplD>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#define INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplD>::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<ZWilsonImplD>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#define EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplD>::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<ZWilsonImplD>::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>
+
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, dag Kernel, double
+/////////////////////////////////////////////////////////////////
+#define KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplD>::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<ZWilsonImplD>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#define INTERIOR
+#undef EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplD>::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<ZWilsonImplD>::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>
+
+
+#undef INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#define EXTERIOR
+
+#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
+template<> void
+WilsonKernels<WilsonImplD>::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<ZWilsonImplD>::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>
+
+
+
+
+// undefs
+#include <simd/Fujitsu_A64FX_undef.h>
+
+#endif //A64FXASM

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/WilsonKernelsAsmBodyA64FX.h

@@ -0,0 +1,395 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: WilsonKernelsAsmBodyA64FX.h
+
+    Copyright (C) 2020
+
+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 */
+
+// GCC 10 messes up SVE instruction scheduling using -O3, but
+// -O3 -fno-schedule-insns -fno-schedule-insns2 does wonders
+// performance now is better than armclang 20.2
+
+#ifdef KERNEL_DAG
+#define DIR0_PROJ    XP_PROJ
+#define DIR1_PROJ    YP_PROJ
+#define DIR2_PROJ    ZP_PROJ
+#define DIR3_PROJ    TP_PROJ
+#define DIR4_PROJ    XM_PROJ
+#define DIR5_PROJ    YM_PROJ
+#define DIR6_PROJ    ZM_PROJ
+#define DIR7_PROJ    TM_PROJ
+#define DIR0_RECON   XP_RECON
+#define DIR1_RECON   YP_RECON_ACCUM
+#define DIR2_RECON   ZP_RECON_ACCUM
+#define DIR3_RECON   TP_RECON_ACCUM
+#define DIR4_RECON   XM_RECON_ACCUM
+#define DIR5_RECON   YM_RECON_ACCUM
+#define DIR6_RECON   ZM_RECON_ACCUM
+#define DIR7_RECON   TM_RECON_ACCUM
+#else
+#define DIR0_PROJ    XM_PROJ
+#define DIR1_PROJ    YM_PROJ
+#define DIR2_PROJ    ZM_PROJ
+#define DIR3_PROJ    TM_PROJ
+#define DIR4_PROJ    XP_PROJ
+#define DIR5_PROJ    YP_PROJ
+#define DIR6_PROJ    ZP_PROJ
+#define DIR7_PROJ    TP_PROJ
+#define DIR0_RECON   XM_RECON
+#define DIR1_RECON   YM_RECON_ACCUM
+#define DIR2_RECON   ZM_RECON_ACCUM
+#define DIR3_RECON   TM_RECON_ACCUM
+#define DIR4_RECON   XP_RECON_ACCUM
+#define DIR5_RECON   YP_RECON_ACCUM
+#define DIR6_RECON   ZP_RECON_ACCUM
+#define DIR7_RECON   TP_RECON_ACCUM
+#endif
+
+//using namespace std;
+
+#undef SHOW
+//#define SHOW
+
+#undef WHERE
+
+#ifdef INTERIOR_AND_EXTERIOR
+#define WHERE "INT_AND_EXT"
+#endif
+
+#ifdef INTERIOR
+#define WHERE "INT"
+#endif
+
+#ifdef EXTERIOR
+#define WHERE "EXT"
+#endif
+
+//#pragma message("here")
+
+
+
+////////////////////////////////////////////////////////////////////////////////
+// Comms then compute kernel
+////////////////////////////////////////////////////////////////////////////////
+#ifdef INTERIOR_AND_EXTERIOR
+
+#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+      basep = st.GetPFInfo(nent,plocal); nent++;			\
+      if ( local ) {							            \
+    LOAD_CHIMU(base);                                       \
+    LOAD_TABLE(PERMUTE_DIR);                                \
+    PROJ;							                        \
+    MAYBEPERM(PERMUTE_DIR,perm);					        \
+      } else {								                \
+	  LOAD_CHI(base);							                \
+      }									                    \
+      base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
+    MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    PREFETCH_CHIMU_L2(basep);                               \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+    MULT_2SPIN_2;					                        \
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+
+/*
+NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
+    though I expected that it would improve on performance
+*/
+
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
+  PREFETCH1_CHIMU(base);						            \
+  ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
+
+#define RESULT(base,basep) SAVE_RESULT(base,basep);
+
+#endif
+
+////////////////////////////////////////////////////////////////////////////////
+// Pre comms kernel -- prefetch like normal because it is mostly right
+////////////////////////////////////////////////////////////////////////////////
+#ifdef INTERIOR
+
+#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+      basep = st.GetPFInfo(nent,plocal); nent++;			\
+      if ( local ) {							\
+  LOAD_CHIMU(base);                                       \
+  LOAD_TABLE(PERMUTE_DIR);                                \
+  PROJ;							                        \
+  MAYBEPERM(PERMUTE_DIR,perm);					        \
+      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}			\
+      if ( local || st.same_node[Dir] ) {				\
+  MULT_2SPIN_1(Dir);					                    \
+  MULT_2SPIN_2;					                        \
+  RECON;								\
+      }									\
+  base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
+  PREFETCH_CHIMU(base);						\
+  PREFETCH_CHIMU_L2(basep);                               \
+
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
+  PREFETCH1_CHIMU(base);						\
+  { ZERO_PSI; }								\
+  ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
+
+#define RESULT(base,basep) SAVE_RESULT(base,basep);
+
+#endif
+
+////////////////////////////////////////////////////////////////////////////////
+// Post comms kernel
+////////////////////////////////////////////////////////////////////////////////
+#ifdef EXTERIOR
+
+#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
+  if((!local)&&(!st.same_node[Dir]) ) {					\
+    LOAD_CHI(base);							\
+    MULT_2SPIN_1(Dir);					                    \
+    MULT_2SPIN_2;					                        \
+    RECON;								\
+    nmu++;								\
+  }
+
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+  nmu=0;								\
+  { ZERO_PSI;}								\
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
+  if((!local)&&(!st.same_node[Dir]) ) {					\
+    LOAD_CHI(base);							\
+    MULT_2SPIN_1(Dir);					                    \
+    MULT_2SPIN_2;					                        \
+    RECON;								\
+    nmu++;								\
+  }
+
+#define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
+
+#endif
+
+
+{
+  int nmu;
+  int local,perm, ptype;
+  uint64_t base;
+  uint64_t basep;
+  const uint64_t plocal =(uint64_t) & in[0];
+
+  MASK_REGS;
+  int nmax=U.oSites();
+  for(int site=0;site<Ns;site++) {
+#ifndef EXTERIOR
+    //    int sU =lo.Reorder(ssU);
+    int sU =ssU;
+    int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
+    //    int sUn=lo.Reorder(ssn);
+    int sUn=ssn;
+#else
+    int sU =ssU;
+    int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
+    int sUn=ssn;
+#endif
+    for(int s=0;s<Ls;s++) {
+      ss =sU*Ls+s;
+      ssn=sUn*Ls+s;
+      int  ent=ss*8;// 2*Ndim
+      int nent=ssn*8;
+
+      uint64_t delta_base, delta_base_p;
+
+   ASM_LEG_XP(Xp,Yp,PERMUTE_DIR3,DIR0_PROJ,DIR0_RECON);
+
+#ifdef SHOW
+      float rescale = 64. * 12.;
+      std::cout << "=================================================================" << std::endl;
+      std::cout << "ss = " << ss << "   ssn = " << ssn << std::endl;
+      std::cout << "sU = " << sU << "   ssU = " << ssU << std::endl;
+      std::cout << " " << std::endl;
+
+
+      std::cout << "Dir = " << Xp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Xp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Yp,Zp,PERMUTE_DIR2,DIR1_PROJ,DIR1_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Yp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Yp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Zp,Tp,PERMUTE_DIR1,DIR2_PROJ,DIR2_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Zp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Zp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Tp,Xm,PERMUTE_DIR0,DIR3_PROJ,DIR3_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Tp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Tp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Xm,Ym,PERMUTE_DIR3,DIR4_PROJ,DIR4_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Xm << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Xm] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      // DC ZVA test
+      // { uint64_t basestore = (uint64_t)&out[ss];
+      //   PREFETCH_RESULT_L2_STORE(basestore); }
+
+
+      ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJ,DIR5_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Ym << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Ym] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      // DC ZVA test
+      //{ uint64_t basestore = (uint64_t)&out[ss];
+      //  PREFETCH_RESULT_L2_STORE(basestore); }
+
+
+      ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJ,DIR6_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Zm << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Zm] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      // DC ZVA test
+      //{ uint64_t basestore = (uint64_t)&out[ss];
+      //  PREFETCH_RESULT_L2_STORE(basestore); }
+
+
+      ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJ,DIR7_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Tm << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Tm] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+#ifdef EXTERIOR
+      if (nmu==0) break;
+      //      if (nmu!=0) std::cout << "EXT "<<sU<<std::endl;
+#endif
+      base = (uint64_t) &out[ss];
+      basep= st.GetPFInfo(nent,plocal); ent++;
+      basep = (uint64_t) &out[ssn];
+      //PREFETCH_RESULT_L1_STORE(base);
+      RESULT(base,basep);
+
+#ifdef SHOW
+      std::cout << "Dir = FINAL        " <<  WHERE<< std::endl;;
+
+      base_ss = base;
+      std::cout << "base              = " << (base - (uint64_t) &out[0])/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+    }
+    ssU++;
+    UNLOCK_GAUGE(0);
+  }
+}
+
+#undef DIR0_PROJ
+#undef DIR1_PROJ
+#undef DIR2_PROJ
+#undef DIR3_PROJ
+#undef DIR4_PROJ
+#undef DIR5_PROJ
+#undef DIR6_PROJ
+#undef DIR7_PROJ
+#undef DIR0_RECON
+#undef DIR1_RECON
+#undef DIR2_RECON
+#undef DIR3_RECON
+#undef DIR4_RECON
+#undef DIR5_RECON
+#undef DIR6_RECON
+#undef DIR7_RECON
+#undef ASM_LEG
+#undef ASM_LEG_XP
+#undef RESULT

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

@@ -646,7 +646,7 @@ NAMESPACE_BEGIN(Grid);
   HAND_RESULT_EXT(ss,F)
 
 #define HAND_SPECIALISE_GPARITY(IMPL)					\
-  template<> void						\
+  template<> accelerator_inline void						\
   WilsonKernels<IMPL>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
 				    int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
@@ -662,7 +662,7 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> void						\
+  template<> accelerator_inline void						\
   WilsonKernels<IMPL>::HandDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
 				       int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
@@ -678,7 +678,7 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> void						\
+  template<> accelerator_inline void						\
   WilsonKernels<IMPL>::HandDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
 				       int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
@@ -694,7 +694,7 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> void						\
+  template<> accelerator_inline void						\
   WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
@@ -710,7 +710,7 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> void							\
+  template<> accelerator_inline void							\
   WilsonKernels<IMPL>::HandDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
 				       int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
@@ -727,7 +727,7 @@ NAMESPACE_BEGIN(Grid);
     nmu = 0;								\
     HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
-  template<> void						\
+  template<> accelerator_inline void						\
   WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\

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

@@ -76,7 +76,24 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 #define REGISTER
 
-#define LOAD_CHIMU \
+#ifdef GRID_SIMT
+#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);	}
+#define PERMUTE_DIR(dir) ;
+#else
+#define LOAD_CHIMU(Ptype)		\
   {const SiteSpinor & ref (in[offset]);	\
     Chimu_00=ref()(0)(0);\
     Chimu_01=ref()(0)(1);\
@@ -91,24 +108,24 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     Chimu_31=ref()(3)(1);\
     Chimu_32=ref()(3)(2);}
 
-#define LOAD_CHI\
-  {const SiteHalfSpinor &ref(buf[offset]);	\
-    Chi_00 = ref()(0)(0);\
-    Chi_01 = ref()(0)(1);\
-    Chi_02 = ref()(0)(2);\
-    Chi_10 = ref()(1)(0);\
-    Chi_11 = ref()(1)(1);\
-    Chi_12 = ref()(1)(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);
+
+#endif
 
-// To splat or not to splat depends on the implementation
 #define MULT_2SPIN(A)\
   {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));	\
+    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);				\
     UChi_00 = U_00*Chi_00;						\
     UChi_10 = U_00*Chi_10;						\
     UChi_01 = U_10*Chi_00;						\
@@ -121,9 +138,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     UChi_11+= U_11*Chi_11;						\
     UChi_02+= U_21*Chi_01;						\
     UChi_12+= U_21*Chi_11;						\
-    Impl::loadLinkElement(U_00,ref()(0,2));	\
-    Impl::loadLinkElement(U_10,ref()(1,2));	\
-    Impl::loadLinkElement(U_20,ref()(2,2));	\
+    U_00=coalescedRead(ref()(0,2),lane);				\
+    U_10=coalescedRead(ref()(1,2),lane);				\
+    U_20=coalescedRead(ref()(2,2),lane);				\
     UChi_00+= U_00*Chi_02;						\
     UChi_10+= U_00*Chi_12;						\
     UChi_01+= U_10*Chi_02;						\
@@ -131,14 +148,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     UChi_02+= U_20*Chi_02;						\
     UChi_12+= U_20*Chi_12;}
 
-
-#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);
+#define LOAD_CHI				\
+  {const SiteHalfSpinor &ref(buf[offset]);	\
+    Chi_00 = coalescedRead(ref()(0)(0),lane);	\
+    Chi_01 = coalescedRead(ref()(0)(1),lane);	\
+    Chi_02 = coalescedRead(ref()(0)(2),lane);	\
+    Chi_10 = coalescedRead(ref()(1)(0),lane);	\
+    Chi_11 = coalescedRead(ref()(1)(1),lane);	\
+    Chi_12 = coalescedRead(ref()(1)(2),lane);}
 
 //      hspin(0)=fspin(0)+timesI(fspin(3));
 //      hspin(1)=fspin(1)+timesI(fspin(2));
@@ -353,13 +370,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   result_31-= UChi_11;	\
   result_32-= UChi_12;
 
-#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
+#define HAND_STENCIL_LEGB(PROJ,PERM,DIR,RECON)	\
+  {int ptype;					\
    SE=st.GetEntry(ptype,DIR,ss);		\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
+   auto offset = SE->_offset;			\
+   auto local  = SE->_is_local;			\
+   auto perm   = SE->_permute;			\
    if ( local ) {				\
-    LOAD_CHIMU;					\
+     LOAD_CHIMU(PERM);				\
      PROJ;					\
      if ( perm) {				\
        PERMUTE_DIR(PERM);			\
@@ -367,16 +385,48 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
    } else {					\
      LOAD_CHI;					\
    }						\
+   acceleratorSynchronise();			\
    MULT_2SPIN(DIR);				\
-  RECON;					
+   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];							\
+    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)	\
+  { int ptype;						\
   SE=st.GetEntry(ptype,DIR,ss);				\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
+  auto offset = SE->_offset;					\
+  auto local  = SE->_is_local;					\
+  auto perm   = SE->_permute;					\
   if ( local ) {						\
-    LOAD_CHIMU;					\
+    LOAD_CHIMU(PERM);						\
     PROJ;							\
     if ( perm) {						\
       PERMUTE_DIR(PERM);					\
@@ -384,57 +434,60 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   } 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)	\
+  { int ptype;						\
   SE=st.GetEntry(ptype,DIR,ss);				\
-  offset = SE->_offset;				\
+  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]);				\
-    vstream(ref()(0)(0),result_00);		\
-    vstream(ref()(0)(1),result_01);		\
-    vstream(ref()(0)(2),result_02);		\
-    vstream(ref()(1)(0),result_10);		\
-    vstream(ref()(1)(1),result_11);		\
-    vstream(ref()(1)(2),result_12);		\
-    vstream(ref()(2)(0),result_20);		\
-    vstream(ref()(2)(1),result_21);		\
-    vstream(ref()(2)(2),result_22);		\
-    vstream(ref()(3)(0),result_30);		\
-    vstream(ref()(3)(1),result_31);		\
-    vstream(ref()(3)(2),result_32);		\
+    coalescedWrite(ref()(0)(0),result_00,lane);		\
+    coalescedWrite(ref()(0)(1),result_01,lane);		\
+    coalescedWrite(ref()(0)(2),result_02,lane);		\
+    coalescedWrite(ref()(1)(0),result_10,lane);		\
+    coalescedWrite(ref()(1)(1),result_11,lane);		\
+    coalescedWrite(ref()(1)(2),result_12,lane);		\
+    coalescedWrite(ref()(2)(0),result_20,lane);		\
+    coalescedWrite(ref()(2)(1),result_21,lane);		\
+    coalescedWrite(ref()(2)(2),result_22,lane);		\
+    coalescedWrite(ref()(3)(0),result_30,lane);		\
+    coalescedWrite(ref()(3)(1),result_31,lane);		\
+    coalescedWrite(ref()(3)(2),result_32,lane);		\
   }
 
 #define HAND_RESULT_EXT(ss)				\
-  if (nmu){					\
+  {							\
     SiteSpinor & ref (out[ss]);				\
-    ref()(0)(0)+=result_00;		\
-    ref()(0)(1)+=result_01;		\
-    ref()(0)(2)+=result_02;		\
-    ref()(1)(0)+=result_10;		\
-    ref()(1)(1)+=result_11;		\
-    ref()(1)(2)+=result_12;		\
-    ref()(2)(0)+=result_20;		\
-    ref()(2)(1)+=result_21;		\
-    ref()(2)(2)+=result_22;		\
-    ref()(3)(0)+=result_30;		\
-    ref()(3)(1)+=result_31;		\
-    ref()(3)(2)+=result_32;		\
+    coalescedWrite(ref()(0)(0),coalescedRead(ref()(0)(0))+result_00,lane);	\
+    coalescedWrite(ref()(0)(1),coalescedRead(ref()(0)(1))+result_01,lane);	\
+    coalescedWrite(ref()(0)(2),coalescedRead(ref()(0)(2))+result_02,lane);	\
+    coalescedWrite(ref()(1)(0),coalescedRead(ref()(1)(0))+result_10,lane);	\
+    coalescedWrite(ref()(1)(1),coalescedRead(ref()(1)(1))+result_11,lane);	\
+    coalescedWrite(ref()(1)(2),coalescedRead(ref()(1)(2))+result_12,lane);	\
+    coalescedWrite(ref()(2)(0),coalescedRead(ref()(2)(0))+result_20,lane);	\
+    coalescedWrite(ref()(2)(1),coalescedRead(ref()(2)(1))+result_21,lane);	\
+    coalescedWrite(ref()(2)(2),coalescedRead(ref()(2)(2))+result_22,lane);	\
+    coalescedWrite(ref()(3)(0),coalescedRead(ref()(3)(0))+result_30,lane);	\
+    coalescedWrite(ref()(3)(1),coalescedRead(ref()(3)(1))+result_31,lane);	\
+    coalescedWrite(ref()(3)(2),coalescedRead(ref()(3)(2))+result_32,lane);	\
   }
 
-
-#define HAND_DECLARATIONS(a)			\
+#define HAND_DECLARATIONS(Simd)			\
   Simd result_00;				\
   Simd result_01;				\
   Simd result_02;				\
@@ -467,18 +520,18 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   Simd U_21;
 
 #define ZERO_RESULT							\
-  result_00=Zero();				\
-  result_01=Zero();				\
-  result_02=Zero();				\
-  result_10=Zero();				\
-  result_11=Zero();				\
-  result_12=Zero();				\
-  result_20=Zero();				\
-  result_21=Zero();				\
-  result_22=Zero();				\
-  result_30=Zero();				\
-  result_31=Zero();				\
-  result_32=Zero();			
+  zeroit(result_00);							\
+  zeroit(result_01);							\
+  zeroit(result_02);							\
+  zeroit(result_10);							\
+  zeroit(result_11);							\
+  zeroit(result_12);							\
+  zeroit(result_20);							\
+  zeroit(result_21);							\
+  zeroit(result_22);							\
+  zeroit(result_30);							\
+  zeroit(result_31);							\
+  zeroit(result_32);			
 
 #define Chimu_00 Chi_00
 #define Chimu_01 Chi_01
@@ -495,19 +548,54 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl> void 
-WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
-				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+
+#ifdef SYCL_HACK
+template<class Impl> accelerator_inline void 
+WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
+				      int ss,int sU,const SiteSpinor *in, SiteSpinor *out)
 {
 // 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;
+  typedef iSinglet<Simd> vCplx;
+  //  typedef decltype( coalescedRead( vCplx()()() )) Simt;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  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);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+#endif
 
+template<class Impl> accelerator_inline void 
+WilsonKernels<Impl>::HandDhopSite(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;					
+// 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;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
+
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
+
+  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);
@@ -519,18 +607,22 @@ WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,Site
   HAND_RESULT(ss);
 }
 
-template<class Impl>
+template<class Impl>  accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDag(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;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  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);
@@ -542,17 +634,22 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView
   HAND_RESULT(ss);
 }
 
-template<class Impl> void 
+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;					
 // 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;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
   ZERO_RESULT;
   HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
@@ -566,17 +663,22 @@ WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,Si
   HAND_RESULT(ss);
 }
 
-template<class Impl>
+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;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  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);
@@ -589,17 +691,23 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldVi
   HAND_RESULT(ss);
 }
 
-template<class Impl> void 
+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;					
 // 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;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
-  int offset, ptype;
+  HAND_DECLARATIONS(Simt);
+
+  //  int offset, ptype;
   StencilEntry *SE;
   int nmu=0;
   ZERO_RESULT;
@@ -614,17 +722,23 @@ WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,Si
   HAND_RESULT_EXT(ss);
 }
 
-template<class Impl>
+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;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  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);
@@ -682,3 +796,4 @@ NAMESPACE_END(Grid);
 #undef HAND_RESULT
 #undef HAND_RESULT_INT
 #undef HAND_RESULT_EXT
+#undef HAND_DECLARATIONS

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

@@ -0,0 +1,943 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonKernelsHand.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+#pragma once
+
+#include <Grid/qcd/action/fermion/FermionCore.h>
+
+
+#undef LOAD_CHIMU
+#undef LOAD_CHI
+#undef MULT_2SPIN
+#undef PERMUTE_DIR
+#undef XP_PROJ
+#undef YP_PROJ
+#undef ZP_PROJ
+#undef TP_PROJ
+#undef XM_PROJ
+#undef YM_PROJ
+#undef ZM_PROJ
+#undef TM_PROJ
+#undef XP_RECON
+#undef XP_RECON_ACCUM
+#undef XM_RECON
+#undef XM_RECON_ACCUM
+#undef YP_RECON_ACCUM
+#undef YM_RECON_ACCUM
+#undef ZP_RECON_ACCUM
+#undef ZM_RECON_ACCUM
+#undef TP_RECON_ACCUM
+#undef TM_RECON_ACCUM
+#undef ZERO_RESULT
+#undef Chimu_00
+#undef Chimu_01
+#undef Chimu_02
+#undef Chimu_10
+#undef Chimu_11
+#undef Chimu_12
+#undef Chimu_20
+#undef Chimu_21
+#undef Chimu_22
+#undef Chimu_30
+#undef Chimu_31
+#undef Chimu_32
+#undef HAND_STENCIL_LEG
+#undef HAND_STENCIL_LEG_INT
+#undef HAND_STENCIL_LEG_EXT
+#undef HAND_RESULT
+#undef HAND_RESULT_INT
+#undef HAND_RESULT_EXT
+
+#define REGISTER
+
+#define LOAD_CHIMU \
+  {const SiteSpinor & ref (in[offset]);	\
+    Chimu_00=ref()(0)(0);\
+    Chimu_01=ref()(0)(1);\
+    Chimu_02=ref()(0)(2);\
+    Chimu_10=ref()(1)(0);\
+    Chimu_11=ref()(1)(1);\
+    Chimu_12=ref()(1)(2);\
+    Chimu_20=ref()(2)(0);\
+    Chimu_21=ref()(2)(1);\
+    Chimu_22=ref()(2)(2);\
+    Chimu_30=ref()(3)(0);\
+    Chimu_31=ref()(3)(1);\
+    Chimu_32=ref()(3)(2);\
+    std::cout << std::endl << "DEBUG -- LOAD_CHIMU" << std::endl; \
+    std::cout << "Chimu_00 -- " <<  Chimu_00 << std::endl; \
+    std::cout << "Chimu_01 -- " <<  Chimu_01 << std::endl; \
+    std::cout << "Chimu_02 -- " <<  Chimu_02 << std::endl; \
+    std::cout << "Chimu_10 -- " <<  Chimu_10 << std::endl; \
+    std::cout << "Chimu_11 -- " <<  Chimu_11 << std::endl; \
+    std::cout << "Chimu_12 -- " <<  Chimu_12 << std::endl; \
+    std::cout << "Chimu_20 -- " <<  Chimu_20 << std::endl; \
+    std::cout << "Chimu_21 -- " <<  Chimu_21 << std::endl; \
+    std::cout << "Chimu_22 -- " <<  Chimu_22 << std::endl; \
+    std::cout << "Chimu_30 -- " <<  Chimu_30 << std::endl; \
+    std::cout << "Chimu_31 -- " <<  Chimu_31 << std::endl; \
+    std::cout << "Chimu_32 -- " <<  Chimu_32 << std::endl; \
+}
+
+#define LOAD_CHI\
+  {const SiteHalfSpinor &ref(buf[offset]);	\
+    Chi_00 = ref()(0)(0);\
+    Chi_01 = ref()(0)(1);\
+    Chi_02 = ref()(0)(2);\
+    Chi_10 = ref()(1)(0);\
+    Chi_11 = ref()(1)(1);\
+    Chi_12 = ref()(1)(2);\
+    std::cout << std::endl << "DEBUG -- LOAD_CHI" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl; \
+  }
+
+// To splat or not to splat depends on the implementation
+#define MULT_2SPIN(A)\
+  {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));	\
+    UChi_00 = U_00*Chi_00;\
+    UChi_10 = U_00*Chi_10;\
+    UChi_01 = U_10*Chi_00;\
+    UChi_11 = U_10*Chi_10;\
+    UChi_02 = U_20*Chi_00;\
+    UChi_12 = U_20*Chi_10;\
+    UChi_00+= U_01*Chi_01;\
+    UChi_10+= U_01*Chi_11;\
+    UChi_01+= U_11*Chi_01;\
+    UChi_11+= U_11*Chi_11;\
+    UChi_02+= U_21*Chi_01;\
+    UChi_12+= U_21*Chi_11;\
+    Impl::loadLinkElement(U_00,ref()(0,2));	\
+    Impl::loadLinkElement(U_10,ref()(1,2));	\
+    Impl::loadLinkElement(U_20,ref()(2,2));	\
+    UChi_00+= U_00*Chi_02;\
+    UChi_10+= U_00*Chi_12;\
+    UChi_01+= U_10*Chi_02;\
+    UChi_11+= U_10*Chi_12;\
+    UChi_02+= U_20*Chi_02;\
+    UChi_12+= U_20*Chi_12;\
+    std::cout << std::endl << "DEBUG -- MULT_2SPIN" << std::endl; \
+    std::cout << "UChi_00 -- " <<  UChi_00 << std::endl; \
+    std::cout << "UChi_01 -- " <<  UChi_01 << std::endl; \
+    std::cout << "UChi_02 -- " <<  UChi_02 << std::endl; \
+    std::cout << "UChi_10 -- " <<  UChi_10 << std::endl; \
+    std::cout << "UChi_11 -- " <<  UChi_11 << std::endl; \
+    std::cout << "UChi_12 -- " <<  UChi_12 << std::endl; \
+    }
+
+
+#define PERMUTE_DIR(dir)			\
+std::cout << std::endl << "DEBUG -- PERM PRE" << std::endl; \
+std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+std::cout << "Chi_12 -- " <<  Chi_12 << std::endl; \
+      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);\
+  std::cout << std::endl << "DEBUG -- PERM POST" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+//      hspin(0)=fspin(0)+timesI(fspin(3));
+//      hspin(1)=fspin(1)+timesI(fspin(2));
+#define XP_PROJ \
+    Chi_00 = Chimu_00+timesI(Chimu_30);\
+    Chi_01 = Chimu_01+timesI(Chimu_31);\
+    Chi_02 = Chimu_02+timesI(Chimu_32);\
+    Chi_10 = Chimu_10+timesI(Chimu_20);\
+    Chi_11 = Chimu_11+timesI(Chimu_21);\
+    Chi_12 = Chimu_12+timesI(Chimu_22);\
+    std::cout << std::endl << "DEBUG -- XP_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define YP_PROJ \
+    Chi_00 = Chimu_00-Chimu_30;\
+    Chi_01 = Chimu_01-Chimu_31;\
+    Chi_02 = Chimu_02-Chimu_32;\
+    Chi_10 = Chimu_10+Chimu_20;\
+    Chi_11 = Chimu_11+Chimu_21;\
+    Chi_12 = Chimu_12+Chimu_22;\
+    std::cout << std::endl << "DEBUG -- YP_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define ZP_PROJ \
+  Chi_00 = Chimu_00+timesI(Chimu_20);		\
+  Chi_01 = Chimu_01+timesI(Chimu_21);		\
+  Chi_02 = Chimu_02+timesI(Chimu_22);		\
+  Chi_10 = Chimu_10-timesI(Chimu_30);		\
+  Chi_11 = Chimu_11-timesI(Chimu_31);		\
+  Chi_12 = Chimu_12-timesI(Chimu_32);\
+  std::cout << std::endl << "DEBUG -- ZP_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define TP_PROJ \
+  Chi_00 = Chimu_00+Chimu_20;		\
+  Chi_01 = Chimu_01+Chimu_21;		\
+  Chi_02 = Chimu_02+Chimu_22;		\
+  Chi_10 = Chimu_10+Chimu_30;		\
+  Chi_11 = Chimu_11+Chimu_31;		\
+  Chi_12 = Chimu_12+Chimu_32;\
+  std::cout << std::endl << "DEBUG -- TP_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+
+//      hspin(0)=fspin(0)-timesI(fspin(3));
+//      hspin(1)=fspin(1)-timesI(fspin(2));
+#define XM_PROJ \
+    Chi_00 = Chimu_00-timesI(Chimu_30);\
+    Chi_01 = Chimu_01-timesI(Chimu_31);\
+    Chi_02 = Chimu_02-timesI(Chimu_32);\
+    Chi_10 = Chimu_10-timesI(Chimu_20);\
+    Chi_11 = Chimu_11-timesI(Chimu_21);\
+    Chi_12 = Chimu_12-timesI(Chimu_22);\
+    std::cout << std::endl << "DEBUG -- XM_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define YM_PROJ \
+    Chi_00 = Chimu_00+Chimu_30;\
+    Chi_01 = Chimu_01+Chimu_31;\
+    Chi_02 = Chimu_02+Chimu_32;\
+    Chi_10 = Chimu_10-Chimu_20;\
+    Chi_11 = Chimu_11-Chimu_21;\
+    Chi_12 = Chimu_12-Chimu_22;\
+    std::cout << std::endl << "DEBUG -- YM_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define ZM_PROJ \
+  Chi_00 = Chimu_00-timesI(Chimu_20);		\
+  Chi_01 = Chimu_01-timesI(Chimu_21);		\
+  Chi_02 = Chimu_02-timesI(Chimu_22);		\
+  Chi_10 = Chimu_10+timesI(Chimu_30);		\
+  Chi_11 = Chimu_11+timesI(Chimu_31);		\
+  Chi_12 = Chimu_12+timesI(Chimu_32);\
+  std::cout << std::endl << "DEBUG -- ZM_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define TM_PROJ \
+  Chi_00 = Chimu_00-Chimu_20;		\
+  Chi_01 = Chimu_01-Chimu_21;		\
+  Chi_02 = Chimu_02-Chimu_22;		\
+  Chi_10 = Chimu_10-Chimu_30;		\
+  Chi_11 = Chimu_11-Chimu_31;		\
+  Chi_12 = Chimu_12-Chimu_32;\
+  std::cout << std::endl << "DEBUG -- TM_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+//      fspin(0)=hspin(0);
+//      fspin(1)=hspin(1);
+//      fspin(2)=timesMinusI(hspin(1));
+//      fspin(3)=timesMinusI(hspin(0));
+#define XP_RECON\
+  result_00 = UChi_00;\
+  result_01 = UChi_01;\
+  result_02 = UChi_02;\
+  result_10 = UChi_10;\
+  result_11 = UChi_11;\
+  result_12 = UChi_12;\
+  result_20 = timesMinusI(UChi_10);\
+  result_21 = timesMinusI(UChi_11);\
+  result_22 = timesMinusI(UChi_12);\
+  result_30 = timesMinusI(UChi_00);\
+  result_31 = timesMinusI(UChi_01);\
+  result_32 = timesMinusI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XP_RECON" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define XP_RECON_ACCUM\
+  result_00+=UChi_00;\
+  result_01+=UChi_01;\
+  result_02+=UChi_02;\
+  result_10+=UChi_10;\
+  result_11+=UChi_11;\
+  result_12+=UChi_12;\
+  result_20-=timesI(UChi_10);\
+  result_21-=timesI(UChi_11);\
+  result_22-=timesI(UChi_12);\
+  result_30-=timesI(UChi_00);\
+  result_31-=timesI(UChi_01);\
+  result_32-=timesI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define XM_RECON\
+  result_00 = UChi_00;\
+  result_01 = UChi_01;\
+  result_02 = UChi_02;\
+  result_10 = UChi_10;\
+  result_11 = UChi_11;\
+  result_12 = UChi_12;\
+  result_20 = timesI(UChi_10);\
+  result_21 = timesI(UChi_11);\
+  result_22 = timesI(UChi_12);\
+  result_30 = timesI(UChi_00);\
+  result_31 = timesI(UChi_01);\
+  result_32 = timesI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XM_RECON" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define XM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= timesI(UChi_10);\
+  result_21+= timesI(UChi_11);\
+  result_22+= timesI(UChi_12);\
+  result_30+= timesI(UChi_00);\
+  result_31+= timesI(UChi_01);\
+  result_32+= timesI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define YP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= UChi_10;\
+  result_21+= UChi_11;\
+  result_22+= UChi_12;\
+  result_30-= UChi_00;\
+  result_31-= UChi_01;\
+  result_32-= UChi_02;\
+  std::cout << std::endl << "DEBUG -- YP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define YM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= UChi_10;\
+  result_21-= UChi_11;\
+  result_22-= UChi_12;\
+  result_30+= UChi_00;\
+  result_31+= UChi_01;\
+  result_32+= UChi_02;\
+  std::cout << std::endl << "DEBUG -- YM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define ZP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= timesI(UChi_00);			\
+  result_21-= timesI(UChi_01);			\
+  result_22-= timesI(UChi_02);			\
+  result_30+= timesI(UChi_10);			\
+  result_31+= timesI(UChi_11);			\
+  result_32+= timesI(UChi_12);\
+  std::cout << std::endl << "DEBUG -- ZP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define ZM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= timesI(UChi_00);			\
+  result_21+= timesI(UChi_01);			\
+  result_22+= timesI(UChi_02);			\
+  result_30-= timesI(UChi_10);			\
+  result_31-= timesI(UChi_11);			\
+  result_32-= timesI(UChi_12);\
+  std::cout << std::endl << "DEBUG -- ZM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define TP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= UChi_00;			\
+  result_21+= UChi_01;			\
+  result_22+= UChi_02;			\
+  result_30+= UChi_10;			\
+  result_31+= UChi_11;			\
+  result_32+= UChi_12;\
+  std::cout << std::endl << "DEBUG -- TP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define TM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= UChi_00;	\
+  result_21-= UChi_01;	\
+  result_22-= UChi_02;	\
+  result_30-= UChi_10;	\
+  result_31-= UChi_11;	\
+  result_32-= UChi_12;\
+  std::cout << std::endl << "DEBUG -- TM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU;					\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else {					\
+    LOAD_CHI;					\
+  }						\
+  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;					\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else if ( st.same_node[DIR] ) {		\
+    LOAD_CHI;					\
+  }						\
+  if (local || st.same_node[DIR] ) {		\
+    MULT_2SPIN(DIR);				\
+    RECON;					\
+  }
+
+#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++;					\
+  }
+
+#define HAND_RESULT(ss)				\
+  {						\
+    SiteSpinor & ref (out[ss]);		\
+    vstream(ref()(0)(0),result_00);		\
+    vstream(ref()(0)(1),result_01);		\
+    vstream(ref()(0)(2),result_02);		\
+    vstream(ref()(1)(0),result_10);		\
+    vstream(ref()(1)(1),result_11);		\
+    vstream(ref()(1)(2),result_12);		\
+    vstream(ref()(2)(0),result_20);		\
+    vstream(ref()(2)(1),result_21);		\
+    vstream(ref()(2)(2),result_22);		\
+    vstream(ref()(3)(0),result_30);		\
+    vstream(ref()(3)(1),result_31);		\
+    vstream(ref()(3)(2),result_32);		\
+    std::cout << std::endl << "DEBUG -- RESULT" << std::endl; \
+    std::cout << "result_00 -- " <<  result_00 << std::endl; \
+    std::cout << "result_01 -- " <<  result_01 << std::endl; \
+    std::cout << "result_02 -- " <<  result_02 << std::endl; \
+    std::cout << "result_10 -- " <<  result_10 << std::endl; \
+    std::cout << "result_11 -- " <<  result_11 << std::endl; \
+    std::cout << "result_12 -- " <<  result_12 << std::endl; \
+    std::cout << "result_20 -- " <<  result_20 << std::endl; \
+    std::cout << "result_21 -- " <<  result_21 << std::endl; \
+    std::cout << "result_22 -- " <<  result_22 << std::endl; \
+    std::cout << "result_30 -- " <<  result_30 << std::endl; \
+    std::cout << "result_31 -- " <<  result_31 << std::endl; \
+    std::cout << "result_32 -- " <<  result_32 << std::endl;\
+  }
+
+#define HAND_RESULT_EXT(ss)			\
+  if (nmu){					\
+    SiteSpinor & ref (out[ss]);		\
+    ref()(0)(0)+=result_00;		\
+    ref()(0)(1)+=result_01;		\
+    ref()(0)(2)+=result_02;		\
+    ref()(1)(0)+=result_10;		\
+    ref()(1)(1)+=result_11;		\
+    ref()(1)(2)+=result_12;		\
+    ref()(2)(0)+=result_20;		\
+    ref()(2)(1)+=result_21;		\
+    ref()(2)(2)+=result_22;		\
+    ref()(3)(0)+=result_30;		\
+    ref()(3)(1)+=result_31;		\
+    ref()(3)(2)+=result_32;		\
+    std::cout << std::endl << "DEBUG -- RESULT EXT" << std::endl; \
+    std::cout << "result_00 -- " <<  result_00 << std::endl; \
+    std::cout << "result_01 -- " <<  result_01 << std::endl; \
+    std::cout << "result_02 -- " <<  result_02 << std::endl; \
+    std::cout << "result_10 -- " <<  result_10 << std::endl; \
+    std::cout << "result_11 -- " <<  result_11 << std::endl; \
+    std::cout << "result_12 -- " <<  result_12 << std::endl; \
+    std::cout << "result_20 -- " <<  result_20 << std::endl; \
+    std::cout << "result_21 -- " <<  result_21 << std::endl; \
+    std::cout << "result_22 -- " <<  result_22 << std::endl; \
+    std::cout << "result_30 -- " <<  result_30 << std::endl; \
+    std::cout << "result_31 -- " <<  result_31 << std::endl; \
+    std::cout << "result_32 -- " <<  result_32 << std::endl;\
+  }
+
+
+#define HAND_DECLARATIONS(a)			\
+  Simd result_00;				\
+  Simd result_01;				\
+  Simd result_02;				\
+  Simd result_10;				\
+  Simd result_11;				\
+  Simd result_12;				\
+  Simd result_20;				\
+  Simd result_21;				\
+  Simd result_22;				\
+  Simd result_30;				\
+  Simd result_31;				\
+  Simd result_32;				\
+  Simd Chi_00;					\
+  Simd Chi_01;					\
+  Simd Chi_02;					\
+  Simd Chi_10;					\
+  Simd Chi_11;					\
+  Simd Chi_12;					\
+  Simd UChi_00;					\
+  Simd UChi_01;					\
+  Simd UChi_02;					\
+  Simd UChi_10;					\
+  Simd UChi_11;					\
+  Simd UChi_12;					\
+  Simd U_00;					\
+  Simd U_10;					\
+  Simd U_20;					\
+  Simd U_01;					\
+  Simd U_11;					\
+  Simd U_21;\
+  Simd debugreg;\
+  svbool_t pg1;        \
+  pg1 = svptrue_b64();        \
+
+#define ZERO_RESULT				\
+  result_00=Zero();				\
+  result_01=Zero();				\
+  result_02=Zero();				\
+  result_10=Zero();				\
+  result_11=Zero();				\
+  result_12=Zero();				\
+  result_20=Zero();				\
+  result_21=Zero();				\
+  result_22=Zero();				\
+  result_30=Zero();				\
+  result_31=Zero();				\
+  result_32=Zero();
+
+#define Chimu_00 Chi_00
+#define Chimu_01 Chi_01
+#define Chimu_02 Chi_02
+#define Chimu_10 Chi_10
+#define Chimu_11 Chi_11
+#define Chimu_12 Chi_12
+#define Chimu_20 UChi_00
+#define Chimu_21 UChi_01
+#define Chimu_22 UChi_02
+#define Chimu_30 UChi_10
+#define Chimu_31 UChi_11
+#define Chimu_32 UChi_12
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Impl> void
+WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+// 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;
+
+  HAND_DECLARATIONS(ignore);
+
+  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);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  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);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl> void
+WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+// 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;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  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);
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl> void
+WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+// 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;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset, ptype;
+  StencilEntry *SE;
+  int nmu=0;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT_EXT(ss);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset, ptype;
+  int nmu=0;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT_EXT(ss);
+}
+
+////////////// Wilson ; uses this implementation /////////////////////
+
+NAMESPACE_END(Grid);
+#undef LOAD_CHIMU
+#undef LOAD_CHI
+#undef MULT_2SPIN
+#undef PERMUTE_DIR
+#undef XP_PROJ
+#undef YP_PROJ
+#undef ZP_PROJ
+#undef TP_PROJ
+#undef XM_PROJ
+#undef YM_PROJ
+#undef ZM_PROJ
+#undef TM_PROJ
+#undef XP_RECON
+#undef XP_RECON_ACCUM
+#undef XM_RECON
+#undef XM_RECON_ACCUM
+#undef YP_RECON_ACCUM
+#undef YM_RECON_ACCUM
+#undef ZP_RECON_ACCUM
+#undef ZM_RECON_ACCUM
+#undef TP_RECON_ACCUM
+#undef TM_RECON_ACCUM
+#undef ZERO_RESULT
+#undef Chimu_00
+#undef Chimu_01
+#undef Chimu_02
+#undef Chimu_10
+#undef Chimu_11
+#undef Chimu_12
+#undef Chimu_20
+#undef Chimu_21
+#undef Chimu_22
+#undef Chimu_30
+#undef Chimu_31
+#undef Chimu_32
+#undef HAND_STENCIL_LEG
+#undef HAND_STENCIL_LEG_INT
+#undef HAND_STENCIL_LEG_EXT
+#undef HAND_RESULT
+#undef HAND_RESULT_INT
+#undef HAND_RESULT_EXT

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

@@ -114,7 +114,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   ////////////////////////////////////////////////////////////////////
   // All legs kernels ; comms then compute
   ////////////////////////////////////////////////////////////////////
-template <class Impl>
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,
 					     SiteHalfSpinor *buf, int sF,
 					     int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -140,7 +140,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldV
   coalescedWrite(out[sF],result,lane);
 };
 
-template <class Impl>
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U,
 					  SiteHalfSpinor *buf, int sF,
 					  int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -169,7 +169,7 @@ void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView
   ////////////////////////////////////////////////////////////////////
   // Interior kernels
   ////////////////////////////////////////////////////////////////////
-template <class Impl>
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFieldView &U,
 						SiteHalfSpinor *buf, int sF,
 						int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -197,7 +197,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFi
   coalescedWrite(out[sF], result,lane);
 };
 
-template <class Impl>
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U,
 							 SiteHalfSpinor *buf, int sF,
 							 int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -227,7 +227,7 @@ void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeField
 ////////////////////////////////////////////////////////////////////
 // Exterior kernels
 ////////////////////////////////////////////////////////////////////
-template <class Impl>
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFieldView &U,
 						SiteHalfSpinor *buf, int sF,
 						int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -258,7 +258,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFi
   }
 };
 
-template <class Impl>
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U,
 					     SiteHalfSpinor *buf, int sF,
 					     int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -290,7 +290,7 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeField
 };
 
 #define DhopDirMacro(Dir,spProj,spRecon)	\
-  template <class Impl>							\
+  template <class Impl> accelerator_inline				\
   void WilsonKernels<Impl>::DhopDir##Dir(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF, \
 					 int sU, const FermionFieldView &in, FermionFieldView &out, int dir) \
   {									\
@@ -318,7 +318,7 @@ DhopDirMacro(Ym,spProjYm,spReconYm);
 DhopDirMacro(Zm,spProjZm,spReconZm);
 DhopDirMacro(Tm,spProjTm,spReconTm);
 
-template <class Impl> 
+template <class Impl> accelerator_inline
 void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF,
 				    int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int gamma)
 {
@@ -416,6 +416,20 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
+#define KERNEL_CALL_TMP(A) \
+  const uint64_t    NN = Nsite*Ls;					\
+  auto U_p = & U_v[0];							\
+  auto in_p = & in_v[0];						\
+  auto out_p = & out_v[0];						\
+  auto st_p = st_v._entries_p;						\
+  auto st_perm = st_v._permute_type;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_perm,st_p,U_p,buf,sF,sU,in_p,out_p);	\
+    });									\
+  accelerator_barrier();
+
 #define KERNEL_CALLNB(A)						\
   const uint64_t    NN = Nsite*Ls;					\
   accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
@@ -445,20 +459,24 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifndef GRID_CUDA
+#ifdef SYCL_HACK     
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
+#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
+#endif     
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteInt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteInt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
    }
@@ -476,20 +494,20 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDag); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDag);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
    }
@@ -501,4 +519,3 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #undef ASM_CALL
 
 NAMESPACE_END(Grid);
-

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

@@ -0,0 +1,41 @@
+/*************************************************************************************
+
+    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>
+
+    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>
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class CompactWilsonCloverFermion<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