Merge branch 'feature/ddhmc' of https://github.com/paboyle/Grid into feature/ddhmc

Print PCI device IDs on node 0

.gitignore

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

.travis.yml

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

Grid/DisableWarnings.h

@@ -34,10 +34,15 @@ directory
 
 #if defined __GNUC__ && __GNUC__>=6
 #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+#if defined __GNUC__ && __GNUC__>=6
+#pragma GCC diagnostic ignored "-Wpsabi"
 #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/GridQCDcore.h

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

Grid/Makefile.am

@@ -54,9 +54,11 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
-extra_sources+=$(ZWILS_FERMION_FILES)
 extra_sources+=$(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

Grid/algorithms/Algorithms.h

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

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);
+    auto& geom_v = geom;
     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<geom_v.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());
+      return MdagNonHermitian(in, out);
-      G5C(tmp, in); 
-      M(tmp, out);
-      G5C(out, 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);
+    auto& geom_v = geom;
+    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<geom_v.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();
+    MdirCalc(in,out,geom.point(dir,disp));
-
-    //////////////
-    // 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);
-
   };
 
   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) {
- CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	: 
+    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();
     }
+
+    InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl;
+    FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl;
   }
 
-  void ForceHermitian(void) {
+  void InvertSelfStencilLink() {
-    CoarseMatrix Diff  (Grid());
+    std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl;
-    for(int p=0;p<geom.npoint;p++){
+    int localVolume = Grid()->lSites();
-      int dir   = geom.directions[p];
+
-      int disp  = geom.displacements[p];
+    typedef typename Cobj::scalar_object scalar_object;
-      if(disp==-1) {
+
-	// Find the opposite link
+    autoView(Aself_v,    A[geom.npoint-1], CpuRead);
-	for(int pp=0;pp<geom.npoint;pp++){
+    autoView(AselfInv_v, AselfInv,         CpuWrite);
-	  int dirp   = geom.directions[pp];
+    thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke
-	  int dispp  = geom.displacements[pp];
+      Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis);
-	  if ( (dirp==dir) && (dispp==1) ){
+      Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis);
-	    //	    Diff = adj(Cshift(A[p],dir,1)) - A[pp]; 
+
-	    //	    std::cout << GridLogMessage<<" Replacing stencil leg "<<pp<<" with leg "<<p<< " diff "<<norm2(Diff) <<std::endl;
+      scalar_object selfLink    = Zero();
-	    A[pp] = adj(Cshift(A[p],dir,1));
+      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

@@ -223,9 +223,14 @@ class SchurOperatorBase :  public LinearOperatorBase<Field> {
     Mpc(in,tmp);
     MpcDag(tmp,out);
   }
+  virtual  void MpcMpcDag(const Field &in, Field &out) {
+    Field tmp(in.Grid());
+    tmp.Checkerboard() = in.Checkerboard();
+    MpcDag(in,tmp);
+    Mpc(tmp,out);
+  }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    out.Checkerboard() = in.Checkerboard();
+    HermOp(in,out);
-    MpcDagMpc(in,out);
     ComplexD dot= innerProduct(in,out); 
     n1=real(dot);
     n2=norm2(out);
@@ -276,6 +281,16 @@ template<class Matrix,class Field>
       axpy(out,-1.0,tmp,out);
     }
 };
+// Mpc MpcDag system presented as the HermOp
+template<class Matrix,class Field>
+class SchurDiagMooeeDagOperator :  public SchurDiagMooeeOperator<Matrix,Field> {
+ public:
+  virtual void HermOp(const Field &in, Field &out){
+    out.Checkerboard() = in.Checkerboard();
+    this->MpcMpcDag(in,out);
+  }
+  SchurDiagMooeeDagOperator (Matrix &Mat): SchurDiagMooeeOperator<Matrix,Field>(Mat){};
+};
 template<class Matrix,class Field>
   class SchurDiagOneOperator :  public SchurOperatorBase<Field> {
  protected:
@@ -530,6 +545,16 @@ 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> {

Grid/algorithms/approx/Chebyshev.h

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

Grid/algorithms/iterative/ConjugateGradient.h

@@ -102,7 +102,7 @@ public:
     // Check if guess is really REALLY good :)
     if (cp <= rsq) {
       TrueResidual = std::sqrt(a/ssq);
-      std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
+      std::cout << GridLogMessage << "ConjugateGradient guess is converged already "<<TrueResidual<< " tol "<< Tolerance<< std::endl;
       IterationsToComplete = 0;	
       return;
     }

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -48,19 +48,29 @@ NAMESPACE_BEGIN(Grid);
     Integer TotalInnerIterations; //Number of inner CG iterations
     Integer TotalOuterIterations; //Number of restarts
     Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
+    RealD TrueResidual;
 
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
-    
+
-    MixedPrecisionConjugateGradient(RealD tol, 
+    MixedPrecisionConjugateGradient(RealD Tol,
+				    Integer maxinnerit, 
+				    Integer maxouterit, 
+				    GridBase* _sp_grid, 
+				    LinearOperatorBase<FieldF> &_Linop_f, 
+				    LinearOperatorBase<FieldD> &_Linop_d) :
+      MixedPrecisionConjugateGradient(Tol, Tol, maxinnerit, maxouterit, _sp_grid, _Linop_f, _Linop_d) {};
+
+    MixedPrecisionConjugateGradient(RealD Tol,
+				    RealD InnerTol,
 				    Integer maxinnerit, 
 				    Integer maxouterit, 
 				    GridBase* _sp_grid, 
 				    LinearOperatorBase<FieldF> &_Linop_f, 
 				    LinearOperatorBase<FieldD> &_Linop_d) :
       Linop_f(_Linop_f), Linop_d(_Linop_d),
-      Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
+      Tolerance(Tol), InnerTolerance(InnerTol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
-      OuterLoopNormMult(100.), guesser(NULL){ };
+      OuterLoopNormMult(100.), guesser(NULL){ assert(InnerTol < 1.0e-1);};
 
     void useGuesser(LinearFunction<FieldF> &g){
       guesser = &g;
@@ -79,6 +89,11 @@ NAMESPACE_BEGIN(Grid);
     RealD stop = src_norm * Tolerance*Tolerance;
 
     GridBase* DoublePrecGrid = src_d_in.Grid();
+
+    //Generate precision change workspaces
+    precisionChangeWorkspace wk_dp_from_sp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace wk_sp_from_dp(SinglePrecGrid, DoublePrecGrid);
+
     FieldD tmp_d(DoublePrecGrid);
     tmp_d.Checkerboard() = cb;
     
@@ -119,7 +134,7 @@ NAMESPACE_BEGIN(Grid);
       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d);
+      precisionChange(src_f, src_d, wk_sp_from_dp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -137,7 +152,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f);
+      precisionChange(tmp_d, sol_f, wk_dp_from_sp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);
@@ -149,6 +164,7 @@ NAMESPACE_BEGIN(Grid);
     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
     CG_d(Linop_d, src_d_in, sol_d);
     TotalFinalStepIterations = CG_d.IterationsToComplete;
+    TrueResidual = CG_d.TrueResidual;
 
     TotalTimer.Stop();
     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -52,7 +52,7 @@ public:
   MultiShiftFunction shifts;
   std::vector<RealD> TrueResidualShift;
 
-  ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
+  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
     MaxIterations(maxit),
     shifts(_shifts)
   { 
@@ -182,6 +182,9 @@ public:
     for(int s=0;s<nshift;s++) {
       axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
     }
+
+    std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
+    
   
   ///////////////////////////////////////
   // Timers

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -0,0 +1,411 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+    
+//For safety the a final regular CG is applied to clean up if necessary
+
+//Linop to add shift to input linop, used in cleanup CG
+namespace ConjugateGradientMultiShiftMixedPrecSupport{
+template<typename Field>
+class ShiftedLinop: public LinearOperatorBase<Field>{
+public:
+  LinearOperatorBase<Field> &linop_base;
+  RealD shift;
+
+  ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
+
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
+  
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+
+  void HermOp(const Field &in, Field &out){
+    linop_base.HermOp(in, out);
+    axpy(out, shift, in, out);
+  }    
+
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+};
+};
+
+
+template<class FieldD, class FieldF,
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq
+				       ) : 
+    MaxIterations(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GridBase *DoublePrecGrid = src_d.Grid();
+    precisionChangeWorkspace wk_f_from_d(SinglePrecGrid, DoublePrecGrid);
+    precisionChangeWorkspace wk_d_from_f(DoublePrecGrid, SinglePrecGrid);
+
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    assert(psi_d.size()==nshift);
+    assert(mass.size()==nshift);
+    assert(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF r_f(SinglePrecGrid);
+    FieldF p_f(SinglePrecGrid);
+    FieldF tmp_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+    FieldF src_f(SinglePrecGrid);
+    precisionChange(src_f, src_d, wk_f_from_d);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      assert( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      ps_d[s] = src_d;
+    }
+    // r and p for primary
+    r_f=src_f; //residual maintained in single
+    p_f=src_f;
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+  
+    //MdagM+m[0]
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    axpy(mmp_f,mass[0],p_f,mmp_f);
+    RealD rn = norm2(p_f);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_f,b,mmp_f,r_f);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterations;k++){    
+      a = c /cp;
+
+      //Update double precision search direction by residual
+      PrecChangeTimer.Start();
+      precisionChange(r_d, r_f, wk_d_from_f);
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_d[s],a,ps_d[s],r_d);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChange(p_f, p_d, wk_f_from_d); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+
+      cp=c;
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f); 
+      d=real(innerProduct(p_f,mmp_f));    
+      MatrixTimer.Stop();  
+
+      AXPYTimer.Start();
+      axpy(mmp_f,mass[0],p_f,mmp_f);
+      AXPYTimer.Stop();
+      RealD rn = norm2(p_f);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+    
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update double precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
+	}
+      }
+
+      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
+      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
+      AXPYTimer.Stop();
+
+      c = c_f;
+
+      if(k % ReliableUpdateFreq == 0){
+	//Replace r with true residual
+	MatrixTimer.Start();  
+	Linop_d.HermOp(psi_d[0],mmp_d); 
+	MatrixTimer.Stop();  
+
+	AXPYTimer.Start();
+	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
+
+	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
+	AXPYTimer.Stop();
+
+	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
+	
+	PrecChangeTimer.Start();
+	precisionChange(r_f, r_d, wk_f_from_d);
+	PrecChangeTimer.Stop();
+	c = c_d;
+      }
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsq[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged ){
+
+	SolverTimer.Stop();
+	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
+	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
+      
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+
+   
+    }
+    // ugly hack
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    //  assert(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/Deflation.h

@@ -54,15 +54,23 @@ class DeflatedGuesser: public LinearFunction<Field> {
 private:
   const std::vector<Field> &evec;
   const std::vector<RealD> &eval;
+  const unsigned int       N;
 
 public:
 
-  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);

Grid/algorithms/iterative/SchurRedBlack.h

@@ -40,7 +40,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    *        (-MoeMee^{-1}   1 )   
    * L^{dag} = ( 1       Mee^{-dag} Moe^{dag} )
    *           ( 0       1                    )
-   * L^{-d}  = ( 1      -Mee^{-dag} Moe^{dag} )
+   * L^{-dag}= ( 1      -Mee^{-dag} Moe^{dag} )
    *           ( 0       1                    )
    *
    * U^-1 = (1   -Mee^{-1} Meo)
@@ -82,7 +82,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    * c) M_oo^-dag Doo^{dag} Doo Moo^-1 phi_0 = M_oo^-dag (D_oo)^dag L^{-1}  eta_o
    *                              eta_o'     = M_oo^-dag (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
    *                              psi_o = M_oo^-1 phi_o
-   * TODO: Deflation 
+   *
+   *
    */
 namespace Grid {
 
@@ -97,6 +98,7 @@ namespace Grid {
   protected:
     typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
     OperatorFunction<Field> & _HermitianRBSolver;
+    
     int CBfactorise;
     bool subGuess;
     bool useSolnAsInitGuess; // if true user-supplied solution vector is used as initial guess for solver
@@ -132,6 +134,31 @@ namespace Grid {
       (*this)(_Matrix,in,out,guess);
     }
 
+    void RedBlackSource(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &src_o) 
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      Field tmp(grid);
+      int nblock = in.size();
+      for(int b=0;b<nblock;b++){
+	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+      }
+    }
+    // James can write his own deflated guesser
+    // with optimised code for the inner products
+    //    RedBlackSolveSplitGrid();
+    //    RedBlackSolve(_Matrix,src_o,sol_o); 
+
+    void RedBlackSolution(Matrix &_Matrix, const std::vector<Field> &in, const std::vector<Field> &sol_o, std::vector<Field> &out)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      Field tmp(grid);
+      int nblock = in.size();
+      for(int b=0;b<nblock;b++) {
+	pickCheckerboard(Even,tmp,in[b]);
+	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]);
+      }
+    }
+
     template<class Guesser>
     void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess) 
     {
@@ -150,24 +177,29 @@ namespace Grid {
       ////////////////////////////////////////////////
       // Prepare RedBlack source
       ////////////////////////////////////////////////
-      for(int b=0;b<nblock;b++){
+      RedBlackSource(_Matrix,in,src_o);
-	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+	//      for(int b=0;b<nblock;b++){
-      }
+	//	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+	//      }
+      
       ////////////////////////////////////////////////
       // Make the guesses
       ////////////////////////////////////////////////
       if ( subGuess ) guess_save.resize(nblock,grid);
 
-      for(int b=0;b<nblock;b++){
+      
-        if(useSolnAsInitGuess) {
+      if(useSolnAsInitGuess) {
+        for(int b=0;b<nblock;b++){
           pickCheckerboard(Odd, sol_o[b], out[b]);
-        } else {
-          guess(src_o[b],sol_o[b]); 
         }
+      } else {
+        guess(src_o, sol_o); 
+      }
 
-	if ( subGuess ) { 
+	    if ( subGuess ) { 
-	  guess_save[b] = sol_o[b];
+        for(int b=0;b<nblock;b++){
-	}
+          guess_save[b] = sol_o[b];
+        }
       }
       //////////////////////////////////////////////////////////////
       // Call the block solver
@@ -189,13 +221,20 @@ namespace Grid {
 	/////////////////////////////////////////////////
 	// Check unprec residual if possible
 	/////////////////////////////////////////////////
-	if ( ! subGuess ) {
+	if ( ! subGuess ) {	  
-	  _Matrix.M(out[b],resid); 
+
+	  if ( this->adjoint() ) _Matrix.Mdag(out[b],resid); 
+	  else                   _Matrix.M(out[b],resid); 
+
 	  resid = resid-in[b];
 	  RealD ns = norm2(in[b]);
 	  RealD nr = norm2(resid);
 	
-	  std::cout<<GridLogMessage<< "SchurRedBlackBase solver true unprec resid["<<b<<"] "<<std::sqrt(nr/ns) << std::endl;
+	  std::cout<<GridLogMessage<< "SchurRedBlackBase adjoint "<< this->adjoint() << std::endl;
+	  if ( this->adjoint() ) 
+	    std::cout<<GridLogMessage<< "SchurRedBlackBase adjoint solver true unprec resid["<<b<<"] "<<std::sqrt(nr/ns) << std::endl;
+	  else                   
+	    std::cout<<GridLogMessage<< "SchurRedBlackBase solver true unprec resid["<<b<<"] "<<std::sqrt(nr/ns) << std::endl;
 	} else {
 	  std::cout<<GridLogMessage<< "SchurRedBlackBase Guess subtracted after solve["<<b<<"] " << std::endl;
 	}
@@ -249,12 +288,21 @@ namespace Grid {
 
       // Verify the unprec residual
       if ( ! subGuess ) {
-        _Matrix.M(out,resid); 
+
+	std::cout<<GridLogMessage<< "SchurRedBlackBase adjoint "<< this->adjoint() << std::endl;
+	
+	if ( this->adjoint() ) _Matrix.Mdag(out,resid); 
+	else                   _Matrix.M(out,resid); 
+
         resid = resid-in;
         RealD ns = norm2(in);
         RealD nr = norm2(resid);
 
-        std::cout<<GridLogMessage << "SchurRedBlackBase solver true unprec resid "<< std::sqrt(nr/ns) << std::endl;
+	  if ( this->adjoint() ) 
+	    std::cout<<GridLogMessage<< "SchurRedBlackBase adjoint solver true unprec resid "<<std::sqrt(nr/ns) << std::endl;
+	  else                   
+	    std::cout<<GridLogMessage<< "SchurRedBlackBase solver true unprec resid "<<std::sqrt(nr/ns) << std::endl;
+
       } else {
         std::cout << GridLogMessage << "SchurRedBlackBase Guess subtracted after solve." << std::endl;
       }
@@ -263,6 +311,7 @@ namespace Grid {
     /////////////////////////////////////////////////////////////
     // Override in derived. 
     /////////////////////////////////////////////////////////////
+    virtual bool adjoint(void) { return false; }
     virtual void RedBlackSource  (Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)                =0;
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)          =0;
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)                           =0;
@@ -616,6 +665,127 @@ namespace Grid {
         this->_HermitianRBSolver(_OpEO, src_o, sol_o); 
       }
   };
+
+  /*
+   * Red black Schur decomposition
+   *
+   *  M = (Mee Meo) =  (1             0 )   (Mee   0               )  (1 Mee^{-1} Meo)
+   *      (Moe Moo)    (Moe Mee^-1    1 )   (0   Moo-Moe Mee^-1 Meo)  (0   1         )
+   *                =         L                     D                     U
+   *
+   * L^-1 = (1              0 )
+   *        (-MoeMee^{-1}   1 )   
+   * L^{dag} = ( 1       Mee^{-dag} Moe^{dag} )
+   *           ( 0       1                    )
+   *
+   * U^-1 = (1   -Mee^{-1} Meo)
+   *        (0    1           )
+   * U^{dag} = ( 1                 0)
+   *           (Meo^dag Mee^{-dag} 1)
+   * U^{-dag} = (  1                 0)
+   *            (-Meo^dag Mee^{-dag} 1)
+   *
+   *
+   ***********************
+   *     M^dag psi = eta
+   ***********************
+   *
+   * Really for Mobius: (Wilson - easier to just use gamma 5 hermiticity)
+   *
+   *    Mdag psi     =         Udag  Ddag  Ldag psi = eta
+   *
+   * U^{-dag} = (  1                 0)
+   *            (-Meo^dag Mee^{-dag} 1)
+   *
+   *
+   * i)                D^dag phi =  (U^{-dag}  eta)
+   *                        eta'_e = eta_e
+   *                        eta'_o = (eta_o - Meo^dag Mee^{-dag} eta_e)
+   * 
+   *      phi_o = D_oo^-dag eta'_o = D_oo^-dag (eta_o - Meo^dag Mee^{-dag} eta_e)
+   *
+   *      phi_e = D_ee^-dag eta'_e = D_ee^-dag eta_e
+   * 
+   * Solve: 
+   *
+   *      D_oo D_oo^dag phi_o = D_oo (eta_o - Meo^dag Mee^{-dag} eta_e)
+   *
+   * ii) 
+   *      phi = L^dag psi => psi = L^-dag phi. 
+   *
+   * L^{-dag} = ( 1      -Mee^{-dag} Moe^{dag} )
+   *            ( 0       1                    )
+   *
+   *   => sol_e = M_ee^-dag * ( src_e - Moe^dag phi_o )...
+   *   => sol_o = phi_o
+   */
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Site diagonal has Mooee on it, but solve the Adjoint system
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagMooeeDagSolve : public SchurRedBlackBase<Field> {
+  public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
+
+    virtual bool adjoint(void) { return true; }
+    SchurRedBlackDiagMooeeDagSolve(OperatorFunction<Field> &HermitianRBSolver,
+				   const bool initSubGuess = false,
+				   const bool _solnAsInitGuess = false)  
+      : SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
+
+    //////////////////////////////////////////////////////
+    // Override RedBlack specialisation
+    //////////////////////////////////////////////////////
+    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd ,src_o,src);
+      /////////////////////////////////////////////////////
+      // src_o = (source_o - Moe^dag MeeInvDag source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInvDag(src_e,tmp);  assert(  tmp.Checkerboard() ==Even);
+      _Matrix.MeooeDag   (tmp,Mtmp);   assert( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+
+      // get the right Mpc
+      SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
+      _HermOpEO.Mpc(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);
+    }
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
+    {
+      SchurDiagMooeeDagOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
+    };
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
+    {
+      SchurDiagMooeeDagOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);
+    }
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field  sol_e(grid);
+      Field  tmp(grid);
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-dag * ( src_e - Moe^dag phi_o )...
+      // sol_o = phi_o
+      ///////////////////////////////////////////////////
+      _Matrix.MeooeDag(sol_o,tmp);      assert(tmp.Checkerboard()==Even);
+      tmp = src_e-tmp;                  assert(tmp.Checkerboard()==Even);
+      _Matrix.MooeeInvDag(tmp,sol_e);   assert(sol_e.Checkerboard()==Even);
+      
+      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+    }
+  };
+
 }
 
 #endif

Grid/allocator/AlignedAllocator.h

@@ -173,7 +173,8 @@ template<class T> using cshiftAllocator = devAllocator<T>;
 template<class T> using cshiftAllocator = std::allocator<T>;
 #endif
 
-template<class T> using Vector     = std::vector<T,uvmAllocator<T> >;           
+template<class T> using Vector        = std::vector<T,uvmAllocator<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> >;
 

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 : ------------------------------------ "<<std::endl;
-  std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
+  std::cout << " MemoryManager : PrintBytes "<<std::endl;
-  std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<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
 
@@ -181,13 +225,13 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 #ifdef ALLOCATION_CACHE
   bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
   int cache = type + small;
-  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);  
+  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
 #endif
 }
 
-void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) 
+void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
   assert(ncache>0);
 #ifdef GRID_OMP
@@ -211,6 +255,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 
   if ( entries[v].valid ) {
     ret = entries[v].address;
+    cacheBytes -= entries[v].bytes;
     entries[v].valid = 0;
     entries[v].address = NULL;
     entries[v].bytes = 0;
@@ -219,6 +264,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
   entries[v].address=ptr;
   entries[v].bytes  =bytes;
   entries[v].valid  =1;
+  cacheBytes += bytes;
 
   return ret;
 }
@@ -228,13 +274,13 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 #ifdef ALLOCATION_CACHE
   bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
   int cache = type+small;
-  return Lookup(bytes,Entries[cache],Ncache[cache]);
+  return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
 #endif
 }
 
-void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) 
+void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
   assert(ncache>0);
 #ifdef GRID_OMP
@@ -243,6 +289,7 @@ 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*/
@@ -84,14 +82,15 @@ 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 *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) ;
+  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
   static void PrintBytes(void);
  public:

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;

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

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 &);

Grid/communicator/Communicator_mpi3.cc

@@ -275,6 +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);
@@ -360,7 +370,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( gfrom ==MPI_UNDEFINED) {
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
     tag= dir+from*32;
     ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
     assert(ierr==0);
@@ -368,12 +378,18 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
     off_node_bytes+=bytes;
   }
 
-  if ( gdest == MPI_UNDEFINED ) {
+  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
     tag= dir+_processor*32;
     ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
+  } else {
+    // TODO : make a OMP loop on CPU, call threaded bcopy
+    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+    assert(shm!=NULL);
+    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+    acceleratorCopySynchronise(); // MPI prob slower
   }
 
   if ( CommunicatorPolicy == CommunicatorPolicySequential ) {

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 &){}

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
@@ -34,6 +35,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
+#endif
+#ifdef GRID_SYCl
+
 #endif
 
 NAMESPACE_BEGIN(Grid); 
@@ -69,6 +73,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   WorldNodes = WorldSize/WorldShmSize;
   assert( (WorldNodes * WorldShmSize) == WorldSize );
 
+
   // FIXME: Check all WorldShmSize are the same ?
 
   /////////////////////////////////////////////////////////////////////
@@ -169,6 +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
   ////////////////////////////////////////////////////////////////
@@ -428,7 +450,47 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#if defined(GRID_CUDA) ||defined(GRID_HIP)
+#if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
+
+//if defined(GRID_SYCL)
+#if 0
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  void * ShmCommBuf ; 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the pointer array for shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Each MPI rank should allocate our own buffer
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  ShmCommBuf = acceleratorAllocDevice(bytes);
+
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+
+  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+
+  SharedMemoryZero(ShmCommBuf,bytes);
+
+  assert(WorldShmSize == 1);
+  for(int r=0;r<WorldShmSize;r++){
+    WorldShmCommBufs[r] = ShmCommBuf;
+  }
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+}
+#endif
+
+#if defined(GRID_CUDA) ||defined(GRID_HIP) ||defined(GRID_SYCL)  
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -452,18 +514,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   ShmCommBuf = acceleratorAllocDevice(bytes);
-
   if (ShmCommBuf == (void *)NULL ) {
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
-  //  if ( WorldRank == 0 ){
+  if ( WorldRank == 0 ){
-  if ( 1 ){
     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
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -473,6 +533,29 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
     //////////////////////////////////////////////////
+    void * thisBuf = ShmCommBuf;
+    if(!Stencil_force_mpi) {
+#ifdef GRID_SYCL_LEVEL_ZERO_IPC
+    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+      
+    ze_ipc_mem_handle_t ihandle;
+    clone_mem_t handle;
+
+    if ( r==WorldShmRank ) { 
+      auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
+      if ( err != ZE_RESULT_SUCCESS ) {
+	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	exit(EXIT_FAILURE);
+      } else {
+	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+      }
+      memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
+      handle.pid = getpid();
+    }
+#endif
 #ifdef GRID_CUDA
     cudaIpcMemHandle_t handle;
     if ( r==WorldShmRank ) { 
@@ -493,6 +576,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       }
     }
 #endif
+
     //////////////////////////////////////////////////
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
@@ -508,7 +592,35 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     ///////////////////////////////////////////////////////////////
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
-    void * thisBuf = ShmCommBuf;
+
+#ifdef GRID_SYCL_LEVEL_ZERO_IPC
+    if ( r!=WorldShmRank ) {
+      thisBuf = nullptr;
+      std::cout<<"mapping seeking remote pid/fd "
+	       <<handle.pid<<"/"
+	       <<handle.fd<<std::endl;
+
+      int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
+      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
+      //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
+      int myfd  = syscall(438,pidfd,handle.fd,0);
+
+      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
+      
+      memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
+
+      auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
+      if ( err != ZE_RESULT_SUCCESS ) {
+	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	exit(EXIT_FAILURE);
+      } else {
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl;
+      }
+      assert(thisBuf!=nullptr);
+    }
+#endif
 #ifdef GRID_CUDA
     if ( r!=WorldShmRank ) { 
       auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
@@ -530,6 +642,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     ///////////////////////////////////////////////////////////////
     // Save a copy of the device buffers
     ///////////////////////////////////////////////////////////////
+    }
     WorldShmCommBufs[r] = thisBuf;
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
@@ -539,6 +652,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
+#endif
+
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
@@ -709,16 +824,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 /////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
-#ifdef GRID_CUDA
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  cudaMemset(dest,0,bytes);
+  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
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
+  acceleratorCopyToDevice(src,dest,bytes);
 #else   
   bcopy(src,dest,bytes);
 #endif
@@ -782,7 +897,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
 #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_common.h

@@ -110,9 +110,11 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
-#ifdef ACCELERATOR_CSHIFT    
+#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;
 	
@@ -135,7 +137,9 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
 #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;
 
 	Coordinate coor;
 
@@ -257,7 +261,9 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
     int _slice_block = rhs.Grid()->_slice_block[dimension];
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v , rhs, AcceleratorWrite);
-    accelerator_for2d(n,e1,b,e2,1,{
+    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);
@@ -274,7 +280,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 
     // 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);

Grid/cshift/Cshift_mpi.h

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

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_arith.h

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

Grid/lattice/Lattice_basis.h

@@ -62,7 +62,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
     basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
-#if ( (!defined(GRID_SYCL)) && (!defined(GRID_CUDA)) )
+#if ( (!defined(GRID_CUDA)) )
   int max_threads = thread_max();
   Vector < vobj > Bt(Nm * max_threads);
   thread_region
@@ -125,7 +125,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 
 	for(int k=k0; k<k1; ++k){
 	  auto tmp = coalescedRead(Bp[ss*nrot+j]);
-	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss]));
+	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_vp[k][sss]));
 	}
       });
 
@@ -134,7 +134,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	int jj  =j0+j;
 	int ss =sj/nrot;
 	int sss=ss+s;
-	coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j]));
+	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
       });
   }
 #endif
@@ -164,7 +164,8 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
   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_vp[k][ss]);
     }

Grid/lattice/Lattice_crc.h

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

Grid/lattice/Lattice_reduction.h

@@ -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_rng.h

@@ -32,8 +32,9 @@
 #include <random>
 
 #ifdef RNG_SITMO
-#include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
+#include <Grid/random/sitmo_prng_engine.hpp>
 #endif 
+#include <Grid/random/gaussian.h>
 
 #if defined(RNG_SITMO)
 #define RNG_FAST_DISCARD
@@ -142,7 +143,7 @@ public:
 
   std::vector<RngEngine>                             _generators;
   std::vector<std::uniform_real_distribution<RealD> > _uniform;
-  std::vector<std::normal_distribution<RealD> >       _gaussian;
+  std::vector<Grid::gaussian_distribution<RealD> >    _gaussian;
   std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
   std::vector<std::uniform_int_distribution<uint32_t> > _uid;
 
@@ -243,7 +244,7 @@ public:
   GridSerialRNG() : GridRNGbase() {
     _generators.resize(1);
     _uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
+    _gaussian.resize(1,gaussian_distribution<RealD>(0.0,1.0) );
     _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(1,std::uniform_int_distribution<uint32_t>() );
   }
@@ -357,7 +358,7 @@ public:
 
     _generators.resize(_vol);
     _uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
+    _gaussian.resize(_vol,gaussian_distribution<RealD>(0.0,1.0) );
     _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
   }

Grid/lattice/Lattice_transfer.h

@@ -97,6 +97,20 @@ 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 +131,18 @@ 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>
+template<typename T1,typename T2>
-  accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in);
+accelerator_inline void convertType(iScalar<T1> & out, const iScalar<T2> & in) {
-template<typename T1,typename T2,int N>
+  convertType(out._internal,in._internal);
-  accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in);
+}
 
-template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
+template<typename T1,typename T2>
-accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
+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 +159,6 @@ 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);
@@ -355,16 +364,22 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   autoView( coarseData_ , coarseData, AcceleratorWrite);
   autoView( fineData_   , fineData, AcceleratorRead);
 
+  auto coarseData_p = &coarseData_[0];
+  auto fineData_p = &fineData_[0];
+  
   Coordinate fine_rdimensions = fine->_rdimensions;
   Coordinate coarse_rdimensions = coarse->_rdimensions;
+
+  vobj zz = Zero();
   
   accelerator_for(sc,coarse->oSites(),1,{
 
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
-      coarseData_[sc]=Zero();
 
+      vobj cd = zz;
+      
       for(int sb=0;sb<blockVol;sb++){
 
 	int sf;
@@ -374,9 +389,11 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
 	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
 
-	coarseData_[sc]=coarseData_[sc]+fineData_[sf];
+	cd=cd+fineData_p[sf];
       }
 
+      coarseData_p[sc] = cd;
+
     });
   return;
 }
@@ -768,7 +785,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
 
 
 template<class vobj>
-void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
+void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 {
   typedef typename vobj::scalar_object sobj;
 
@@ -993,54 +1010,96 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//Convert a Lattice from one precision to another
+//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
-template<class VobjOut, class VobjIn>
+class precisionChangeWorkspace{
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+  std::pair<Integer,Integer>* fmap_device; //device pointer
-{
+public:
-  assert(out.Grid()->Nd() == in.Grid()->Nd());
+  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid){
-  for(int d=0;d<out.Grid()->Nd();d++){
+    //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
-    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
+    assert(out_grid->Nd() == in_grid->Nd());
-  }
+    for(int d=0;d<out_grid->Nd();d++){
-  out.Checkerboard() = in.Checkerboard();
+      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
-  GridBase *in_grid=in.Grid();
-  GridBase *out_grid = out.Grid();
-
-  typedef typename VobjOut::scalar_object SobjOut;
-  typedef typename VobjIn::scalar_object SobjIn;
-
-  int ndim = out.Grid()->Nd();
-  int out_nsimd = out_grid->Nsimd();
-    
-  std::vector<Coordinate > out_icoor(out_nsimd);
-      
-  for(int lane=0; lane < out_nsimd; lane++){
-    out_icoor[lane].resize(ndim);
-    out_grid->iCoorFromIindex(out_icoor[lane], lane);
-  }
-        
-  std::vector<SobjOut> in_slex_conv(in_grid->lSites());
-  unvectorizeToLexOrdArray(in_slex_conv, in);
-    
-  autoView( out_v , out, CpuWrite);
-  thread_for(out_oidx,out_grid->oSites(),{
-    Coordinate out_ocoor(ndim);
-    out_grid->oCoorFromOindex(out_ocoor, out_oidx);
-
-    ExtractPointerArray<SobjOut> ptrs(out_nsimd);      
-
-    Coordinate lcoor(out_grid->Nd());
-      
-    for(int lane=0; lane < out_nsimd; lane++){
-      for(int mu=0;mu<ndim;mu++)
-	lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
-	
-      int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
-      ptrs[lane] = &in_slex_conv[llex];
     }
-    merge(out_v[out_oidx], ptrs, 0);
+    int Nsimd_out = out_grid->Nsimd();
-  });
+
+    std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
+    for(int lane=0; lane < out_grid->Nsimd(); lane++)
+      out_grid->iCoorFromIindex(out_icorrs[lane], lane);
+  
+    std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
+    thread_for(out_oidx,out_grid->oSites(),{
+	Coordinate out_ocorr; 
+	out_grid->oCoorFromOindex(out_ocorr, out_oidx);
+      
+	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
+	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
+	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
+	
+	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
+	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
+	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
+	  int in_oidx = 0, in_lane = 0;
+	  for(int d=0;d<in_grid->_ndimension;d++){
+	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
+	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
+	  }
+	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
+	}
+      });
+
+    //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
+    size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
+    fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
+    acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes); 
+  }
+
+  //Prevent moving or copying
+  precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
+  precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
+  
+  std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
+
+  ~precisionChangeWorkspace(){
+    acceleratorFreeDevice(fmap_device);
+  }
+};
+
+
+//Convert a lattice of one precision to another. The input workspace contains the mapping data.
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
+  static_assert( std::is_same<typename VobjOut::DoublePrecision, typename VobjIn::DoublePrecision>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+
+  out.Checkerboard() = in.Checkerboard();
+  constexpr int Nsimd_out = VobjOut::Nsimd();
+
+  std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
+
+  //Do the copy/precision change
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( in_v , in, AcceleratorRead);
+
+  accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
+      std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
+      for(int out_lane=0; out_lane < Nsimd_out; out_lane++){      
+	int in_oidx = fmap_osite[out_lane].first;
+	int in_lane = fmap_osite[out_lane].second;
+	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
+      }
+    });
 }
 
+//Convert a Lattice from one precision to another
+//Generate the workspace in place; if multiple calls with the same mapping are performed, consider pregenerating the workspace and reusing
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
+  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
+  precisionChange(out, in, workspace);
+}
+
+
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////

Grid/lattice/Lattice_view.h

@@ -67,9 +67,14 @@ 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; };
   accelerator_inline uint64_t size(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);
+  autoView(iftrue_v,iftrue,CpuRead);
-  std::vector<scalar_object> truevals (Nsimd);
+  autoView(iffalse_v,iffalse,CpuRead);
-  std::vector<scalar_object> falsevals(Nsimd);
+  autoView(predicate_v,predicate,CpuRead);
-
+  autoView(ret_v,ret,CpuWrite);
-  parallel_for(int ss=0;ss<iftrue._grid->oSites(); ss++){
+  Integer NN= grid->oSites();
-
+  thread_for(ss,NN,{
-    extract(iftrue._odata[ss]   ,truevals);
+    Integer mask;
-    extract(iffalse._odata[ss]  ,falsevals);
+    scalar_object trueval;
-    extract<vInteger,Integer>(TensorRemove(predicate._odata[ss]),mask);
+    scalar_object falseval;
-
+    for(int l=0;l<Nsimd;l++){
-    for(int s=0;s<Nsimd;s++){
+      trueval =extractLane(l,iftrue_v[ss]);
-      if (mask[s]) falsevals[s]=truevals[s];
+      falseval=extractLane(l,iffalse_v[ss]);
+      mask    =extractLane(l,predicate_v[ss]);
+      if (mask) falseval=trueval;
+      insertLane(l,ret_v[ss],falseval);
     }
-
+  });
-    merge(ret._odata[ss],falsevals);
-  }
 }
 
 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.cc

@@ -69,6 +69,7 @@ GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(0);
@@ -79,6 +80,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogPerformance.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
+  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
     if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
@@ -87,7 +89,8 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -182,6 +182,7 @@ extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogHMC;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/IldgIO.h

@@ -123,7 +123,7 @@ assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
  ////////////////////////////////////////////////////////////
  // Helper to fill out metadata
  ////////////////////////////////////////////////////////////
- template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
+template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
 					  FieldMetaData &header,
 					  scidacRecord & _scidacRecord,
 					  scidacFile   & _scidacFile) 
@@ -619,12 +619,12 @@ class IldgWriter : public ScidacWriter {
   // Don't require scidac records EXCEPT checksum
   // Use Grid MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
+  template <class stats = PeriodicGaugeStatistics>
-  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
+  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef Lattice<vLorentzColourMatrixD> GaugeField;
-    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -636,6 +636,9 @@ 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 +708,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
+  template <class stats = PeriodicGaugeStatistics>
-  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
+  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 +924,8 @@ 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...
+public:
-  header.link_trace=WilsonLoops<PeriodicGimplF>::linkTrace(data);
+  void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
-  header.plaquette =WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
+  {
-}
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
-{
+  }
-  // How to convert data precision etc...
+};
-  header.link_trace=WilsonLoops<PeriodicGimplD>::linkTrace(data);
+typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
-  header.plaquette =WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
+typedef GaugeStatistics<ConjugateGimplD> ConjugateGaugeStatistics;
-}
-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);
-}
 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

@@ -39,6 +39,10 @@ using namespace Grid;
 ////////////////////////////////////////////////////////////////////////////////
 class NerscIO : public BinaryIO { 
 public:
+  typedef Lattice<vLorentzColourMatrixD> GaugeField;
+
+  // Enable/disable exiting if the plaquette in the header does not match the value computed (default true)
+  static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; }
 
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
@@ -129,12 +133,12 @@ public:
   // Now the meat: the object readers
   /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-  template<class vsimd>
+  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+  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 +157,23 @@ 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 +181,7 @@ 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;
@@ -196,22 +200,29 @@ public:
       std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
       exit(0);
     }
-    assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
+    if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
     assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
     assert(nersc_csum == header.checksum );
       
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
-  template<class vsimd>
+  // Preferred interface
-  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+  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 vLorentzColourMatrixD vobj;
-
-    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
@@ -219,8 +230,8 @@ public:
     // Following should become arguments
     ///////////////////////////////////////////
     header.sequence_number = 1;
-    header.ensemble_id     = "UKQCD";
+    header.ensemble_id     = std::string("UKQCD");
-    header.ensemble_label  = "DWF";
+    header.ensemble_label  = ens_label;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -229,28 +240,28 @@ public:
 
     GridMetaData(grid,header);
     assert(header.nd==4);
-    GaugeStatistics(Umu,header);
+    GaugeStats Stats; Stats(Umu,header);
     MachineCharacteristics(header);
 
-	uint64_t offset;
+    uint64_t offset;
 
     // Sod it -- always write 3x3 double
     header.floating_point = std::string("IEEE64BIG");
     header.data_type      = std::string("4D_SU3_GAUGE_3x3");
     GaugeSimpleUnmunger<fobj3D,sobj> munge;
-	if ( grid->IsBoss() ) { 
+    if ( grid->IsBoss() ) { 
-	  truncate(file);
+      truncate(file);
-    offset = writeHeader(header,file);
+      offset = writeHeader(header,file);
-	}
+    }
-	grid->Broadcast(0,(void *)&offset,sizeof(offset));
+    grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
 					      nersc_csum,scidac_csuma,scidac_csumb);
     header.checksum = nersc_csum;
-	if ( grid->IsBoss() ) { 
+    if ( grid->IsBoss() ) { 
-    writeHeader(header,file);
+      writeHeader(header,file);
-	}
+    }
 
     std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
 	     <<std::hex<<header.checksum

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

@@ -63,6 +63,7 @@ static constexpr int Ngp=2; // gparity index range
 #define ColourIndex  (2)
 #define SpinIndex    (1)
 #define LorentzIndex (0)
+#define GparityFlavourIndex (0)
 
 // Also should make these a named enum type
 static constexpr int DaggerNo=0;
@@ -80,6 +81,15 @@ 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> > ;
+
+const int GparityFlavourTensorIndex = 3; //TensorLevel counts from the bottom!
+
 // ChrisK very keen to add extra space for Gparity doubling.
 //
 // Also add domain wall index, in a way where Wilson operator 
@@ -94,6 +104,7 @@ template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iSca
 template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
 template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
 template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
+template<typename vtype> using iLorentzVector             = iVector<iScalar<iScalar<vtype> >, Nd > ;
 template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
 template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >;
 template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >;
@@ -103,8 +114,10 @@ template<typename vtype> using iHalfSpinColourVector      = iScalar<iVector<iVec
     template<typename vtype> using iSpinColourSpinColourMatrix  = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
 
 
+template<typename vtype> using iGparityFlavourVector                = iVector<iScalar<iScalar<vtype> >, Ngp>;
 template<typename vtype> using iGparitySpinColourVector       = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
 template<typename vtype> using iGparityHalfSpinColourVector   = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
+template<typename vtype> using iGparityFlavourMatrix = iMatrix<iScalar<iScalar<vtype> >, Ngp>;
 
 // Spin matrix
 typedef iSpinMatrix<Complex  >          SpinMatrix;
@@ -151,7 +164,16 @@ typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
 
-// LorentzColour
+// LorentzVector
+typedef iLorentzVector<Complex  > LorentzVector;
+typedef iLorentzVector<ComplexF > LorentzVectorF;
+typedef iLorentzVector<ComplexD > LorentzVectorD;
+
+typedef iLorentzVector<vComplex > vLorentzVector;
+typedef iLorentzVector<vComplexF> vLorentzVectorF;
+typedef iLorentzVector<vComplexD> vLorentzVectorD;
+
+// LorentzColourMatrix
 typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix;
 typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
 typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
@@ -169,6 +191,16 @@ typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
 typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
 
+//G-parity flavour matrix
+typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
+typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
+
+typedef iGparityFlavourMatrix<vComplex> vGparityFlavourMatrix;
+typedef iGparityFlavourMatrix<vComplexF> vGparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<vComplexD> vGparityFlavourMatrixD;
+
+
 // Spin vector
 typedef iSpinVector<Complex >           SpinVector;
 typedef iSpinVector<ComplexF>           SpinVectorF;
@@ -213,6 +245,16 @@ typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
 typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
 typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
 typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
+
+//G-parity flavour vector
+typedef iGparityFlavourVector<Complex >         GparityFlavourVector;
+typedef iGparityFlavourVector<ComplexF>         GparityFlavourVectorF;
+typedef iGparityFlavourVector<ComplexD>         GparityFlavourVectorD;
+
+typedef iGparityFlavourVector<vComplex >         vGparityFlavourVector;
+typedef iGparityFlavourVector<vComplexF>         vGparityFlavourVectorF;
+typedef iGparityFlavourVector<vComplexD>         vGparityFlavourVectorD;
+
     
 // singlets
 typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type.
@@ -256,6 +298,10 @@ typedef Lattice<vLorentzColourMatrix>  LatticeLorentzColourMatrix;
 typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
 typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD;
 
+typedef Lattice<vLorentzVector>  LatticeLorentzVector;
+typedef Lattice<vLorentzVectorF> LatticeLorentzVectorF;
+typedef Lattice<vLorentzVectorD> LatticeLorentzVectorD;
+
 // DoubleStored gauge field
 typedef Lattice<vDoubleStoredColourMatrix>  LatticeDoubleStoredColourMatrix;
 typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;

Grid/qcd/action/Action.h

@@ -30,8 +30,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_QCD_ACTION_H
+#pragma once
-#define GRID_QCD_ACTION_H
 
 ////////////////////////////////////////////
 // Abstract base interface
@@ -51,4 +50,4 @@ NAMESPACE_CHECK(Fermion);
 #include <Grid/qcd/action/pseudofermion/PseudoFermion.h>
 NAMESPACE_CHECK(PseudoFermion);
 
-#endif
+

Grid/qcd/action/ActionBase.h

@@ -40,8 +40,31 @@ class Action
 
 public:
   bool is_smeared = false;
+  RealD deriv_norm_sum;
+  RealD deriv_max_sum;
+  int   deriv_num;
+  RealD deriv_us;
+  RealD S_us;
+  RealD refresh_us;
+  void  reset_timer(void)        {
+    deriv_us = S_us = refresh_us = 0.0;
+    deriv_num=0;
+    deriv_norm_sum = deriv_max_sum=0.0;
+  }
+  void  deriv_log(RealD nrm, RealD max) { deriv_max_sum+=max; deriv_norm_sum+=nrm; deriv_num++;}
+  RealD deriv_max_average(void)         { return deriv_max_sum/deriv_num; };
+  RealD deriv_norm_average(void)        { return deriv_norm_sum/deriv_num; };
+  RealD deriv_timer(void)        { return deriv_us; };
+  RealD S_timer(void)            { return deriv_us; };
+  RealD refresh_timer(void)      { return deriv_us; };
+  void deriv_timer_start(void)   { deriv_us-=usecond(); }
+  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
+  void refresh_timer_start(void) { refresh_us-=usecond(); }
+  void refresh_timer_stop(void)  { refresh_us+=usecond(); }
+  void S_timer_start(void)       { S_us-=usecond(); }
+  void S_timer_stop(void)        { S_us+=usecond(); }
   // Heatbath?
-  virtual void refresh(const GaugeField& U, 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/ActionCore.h

@@ -58,6 +58,8 @@ NAMESPACE_CHECK(Scalar);
 ////////////////////////////////////////////
 // Utility functions
 ////////////////////////////////////////////
+#include <Grid/qcd/action/domains/Domains.h>
+
 #include <Grid/qcd/utils/Metric.h>
 NAMESPACE_CHECK(Metric);
 #include <Grid/qcd/utils/CovariantLaplacian.h>

Grid/qcd/action/ActionParams.h

@@ -36,28 +36,34 @@ NAMESPACE_BEGIN(Grid);
 
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
-  Coordinate twists;
+  Coordinate twists; //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
-  GparityWilsonImplParams() : twists(Nd, 0) {};
+                     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+  bool locally_periodic;
+  GparityWilsonImplParams() : twists(Nd, 0), locally_periodic(false) {};
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
+  bool locally_periodic;
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
+      locally_periodic = false;
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
+    locally_periodic = false;
   }
 };
 
 struct StaggeredImplParams {
-  StaggeredImplParams()  {};
+  bool locally_periodic;
+  StaggeredImplParams() : locally_periodic(false) {};
 };
   
-  struct OneFlavourRationalParams : Serializable {
+struct OneFlavourRationalParams : Serializable {
     GRID_SERIALIZABLE_CLASS_MEMBERS(OneFlavourRationalParams, 
 				    RealD, lo, 
 				    RealD, hi, 
@@ -85,6 +91,50 @@ struct StaggeredImplParams {
         precision(_precision),
         BoundsCheckFreq(_BoundsCheckFreq){};
   };
+
+
+  /*Action parameters for the generalized rational action
+    The approximation is for (M^dag M)^{1/inv_pow}
+    where inv_pow is the denominator of the fractional power.
+    Default inv_pow=2 for square root, making this equivalent to 
+    the OneFlavourRational action
+  */
+    struct RationalActionParams : Serializable {
+    GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams, 
+				    int, inv_pow, 
+				    RealD, lo, //low eigenvalue bound of rational approx
+				    RealD, hi, //high eigenvalue bound of rational approx
+				    int,   MaxIter,  //maximum iterations in msCG
+				    RealD, action_tolerance,  //msCG tolerance in action evaluation
+				    int,   action_degree, //rational approx tolerance in action evaluation
+				    RealD, md_tolerance,  //msCG tolerance in MD integration
+				    int,   md_degree, //rational approx tolerance in MD integration
+				    int,   precision, //precision of floating point arithmetic
+				    int,   BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
+  // constructor 
+  RationalActionParams(int _inv_pow = 2,
+		       RealD _lo      = 0.0, 
+		       RealD _hi      = 1.0, 
+		       int _maxit     = 1000,
+		       RealD _action_tolerance      = 1.0e-8, 
+		       int _action_degree    = 10,
+		       RealD _md_tolerance      = 1.0e-8, 
+		       int _md_degree    = 10,
+		       int _precision = 64,
+		       int _BoundsCheckFreq=20)
+    : inv_pow(_inv_pow), 
+      lo(_lo),
+      hi(_hi),
+      MaxIter(_maxit),
+      action_tolerance(_action_tolerance),
+      action_degree(_action_degree),
+      md_tolerance(_md_tolerance),
+      md_degree(_md_degree),
+      precision(_precision),
+      BoundsCheckFreq(_BoundsCheckFreq){};
+  };
+
+
   
 NAMESPACE_END(Grid);
 

Grid/qcd/action/domains/DDHMCFilter.h

@@ -2,14 +2,12 @@
 
 Grid physics library, www.github.com/paboyle/Grid
 
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+Source file: ./lib/qcd/hmc/DDHMC.h
 
-Copyright (C) 2015, 2020
+Copyright (C) 2021
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: Christopher Kelly
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -28,24 +26,27 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
-/*  END LEGAL */
+			   /*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
 
 NAMESPACE_BEGIN(Grid);
+////////////////////////////////////////////////////
+// DDHMC filter with sub-block size B[mu]
+////////////////////////////////////////////////////
 
-#include "impl.h"
+template<typename MomentaField>
-template class WilsonKernels<IMPLEMENTATION>;
+struct DDHMCFilter: public MomentumFilterBase<MomentaField>
+{
+  Coordinate Block;
+  int Width;
+  
+  DDHMCFilter(const Coordinate &_Block): Block(_Block) {}
+
+  void applyFilter(MomentaField &P) const override
+  {
+    DomainDecomposition Domains(Block);
+    Domains.ProjectDDHMC(P);
+  }
+};
 
 NAMESPACE_END(Grid);
+

Grid/qcd/action/domains/DirichletFilter.h

@@ -0,0 +1,98 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/momentum/DirichletFilter.h
+
+Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+
+////////////////////////////////////////////////////
+// Dirichlet filter with sub-block size B[mu]
+////////////////////////////////////////////////////
+#pragma once 
+
+#include <Grid/qcd/action/domains/DomainDecomposition.h>
+
+NAMESPACE_BEGIN(Grid);
+
+
+template<typename MomentaField>
+struct DirichletFilter: public MomentumFilterBase<MomentaField>
+{
+  Coordinate Block;
+  
+  DirichletFilter(const Coordinate &_Block): Block(_Block) {}
+
+  // Edge detect using domain projectors
+  void applyFilter (MomentaField &U) const override
+  {
+    DomainDecomposition Domains(Block);
+    GridBase *grid = U.Grid();
+    LatticeInteger  coor(grid);
+    LatticeInteger  face(grid);
+    LatticeInteger  one(grid);   one = 1;
+    LatticeInteger  zero(grid); zero = 0;
+    LatticeInteger  omega(grid);
+    LatticeInteger  omegabar(grid);
+    LatticeInteger  tmp(grid);
+
+    omega=one;    Domains.ProjectDomain(omega,0);
+    omegabar=one; Domains.ProjectDomain(omegabar,1);
+    
+    LatticeInteger nface(grid); nface=Zero();
+    
+    MomentaField projected(grid); projected=Zero();
+    typedef decltype(PeekIndex<LorentzIndex>(U,0)) MomentaLinkField;
+    MomentaLinkField  Umu(grid);
+    MomentaLinkField   zz(grid); zz=Zero();
+
+    int dims = grid->Nd();
+    Coordinate Global=grid->GlobalDimensions();
+    assert(dims==Nd);
+
+    for(int mu=0;mu<Nd;mu++){
+
+      if ( Block[mu]!=0 ) {
+
+	Umu = PeekIndex<LorentzIndex>(U,mu);
+
+	// Upper face 
+ 	tmp = Cshift(omegabar,mu,1);
+	tmp = tmp + omega;
+	face = where(tmp == Integer(2),one,zero );
+
+ 	tmp = Cshift(omega,mu,1);
+	tmp = tmp + omegabar;
+	face = where(tmp == Integer(2),one,face );
+
+	Umu = where(face,zz,Umu);
+
+	PokeIndex<LorentzIndex>(U, Umu, mu);
+      }
+    }
+  }
+  
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/domains/DomainDecomposition.h

@@ -0,0 +1,187 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/domains/DomainDecomposition.h
+
+Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+
+////////////////////////////////////////////////////
+// Dirichlet filter with sub-block size B[mu]
+////////////////////////////////////////////////////
+#pragma once 
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+struct DomainDecomposition
+{
+  Coordinate Block;
+  static constexpr RealD factor = 0.6;
+
+  DomainDecomposition(const Coordinate &_Block): Block(_Block){ assert(Block.size()==Nd);};
+  
+  template<class Field>
+  void ProjectDomain(Field &f,Integer domain)
+  {
+    GridBase *grid = f.Grid();
+    int dims = grid->Nd();
+    int isDWF= (dims==Nd+1);
+    assert((dims==Nd)||(dims==Nd+1));
+
+    Field   zz(grid);  zz = Zero();
+    LatticeInteger coor(grid);
+    LatticeInteger domaincoor(grid);
+    LatticeInteger mask(grid); mask = Integer(1);
+    LatticeInteger zi(grid);     zi = Integer(0);
+    for(int d=0;d<Nd;d++){
+      Integer B= Block[d];
+      if ( B ) {
+	LatticeCoordinate(coor,d+isDWF);
+	domaincoor = mod(coor,B);
+	mask = where(domaincoor==Integer(0),zi,mask);
+	mask = where(domaincoor==Integer(B-1),zi,mask);
+      }
+    }
+    if ( !domain )
+      f = where(mask==Integer(1),f,zz);
+    else 
+      f = where(mask==Integer(0),f,zz);
+  };
+  template<class GaugeField>
+  void ProjectDDHMC(GaugeField &U)
+  {
+    GridBase *grid = U.Grid();
+    Coordinate Global=grid->GlobalDimensions();
+    GaugeField zzz(grid); zzz = Zero();
+    LatticeInteger coor(grid); 
+
+    GaugeField Uorg(grid); Uorg = U;
+    
+    auto zzz_mu = PeekIndex<LorentzIndex>(zzz,0);
+    ////////////////////////////////////////////////////
+    // Zero BDY layers
+    ////////////////////////////////////////////////////
+    for(int mu=0;mu<Nd;mu++) {
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+	LatticeCoordinate(coor,mu);
+
+
+	////////////////////////////////
+	// OmegaBar - zero all links contained in slice B-1,0 and
+	// mu links connecting to Omega
+	////////////////////////////////
+
+	U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+
+	auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	U_mu = where(mod(coor,B1)==Integer(B1-2),zzz_mu,U_mu); 
+	PokeIndex<LorentzIndex>(U, U_mu, mu);
+
+      }
+    }
+   
+    ////////////////////////////////////////////
+    // Omega interior slow the evolution
+    // Tricky as we need to take the smallest of values imposed by each cut
+    // Do them in order or largest to smallest and smallest writes last
+    ////////////////////////////////////////////
+    RealD f= factor;
+#if 0    
+    for(int mu=0;mu<Nd;mu++) {
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+
+	auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	auto Uorg_mu= PeekIndex<LorentzIndex>(Uorg,mu);
+	// In the plane
+	U = where(mod(coor,B1)==Integer(B1-5),Uorg*f,U); 
+	U = where(mod(coor,B1)==Integer(4)   ,Uorg*f,U); 
+
+	// Perp links
+       	U_mu = where(mod(coor,B1)==Integer(B1-6),Uorg_mu*f,U_mu);
+	U_mu = where(mod(coor,B1)==Integer(4)   ,Uorg_mu*f,U_mu);
+
+	PokeIndex<LorentzIndex>(U, U_mu, mu);
+      }
+    }
+#endif
+    for(int mu=0;mu<Nd;mu++) {
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+
+	auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	auto Uorg_mu= PeekIndex<LorentzIndex>(Uorg,mu);
+	// In the plane
+	U = where(mod(coor,B1)==Integer(B1-4),Uorg*f*f,U); 
+	U = where(mod(coor,B1)==Integer(3)   ,Uorg*f*f,U); 
+
+	// Perp links
+       	U_mu = where(mod(coor,B1)==Integer(B1-5),Uorg_mu*f*f,U_mu);
+	U_mu = where(mod(coor,B1)==Integer(3)   ,Uorg_mu*f*f,U_mu);
+
+	PokeIndex<LorentzIndex>(U, U_mu, mu);
+      }
+    }
+    for(int mu=0;mu<Nd;mu++) {
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+
+	auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	auto Uorg_mu= PeekIndex<LorentzIndex>(Uorg,mu);
+	// In the plane
+	U = where(mod(coor,B1)==Integer(B1-3),Uorg*f*f*f,U); 
+	U = where(mod(coor,B1)==Integer(2)   ,Uorg*f*f*f,U); 
+
+	// Perp links
+       	U_mu = where(mod(coor,B1)==Integer(B1-4),Uorg_mu*f*f*f,U_mu);
+	U_mu = where(mod(coor,B1)==Integer(2)   ,Uorg_mu*f*f*f,U_mu);
+
+	PokeIndex<LorentzIndex>(U, U_mu, mu);
+      }
+    }
+    for(int mu=0;mu<Nd;mu++) {
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+
+	auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	auto Uorg_mu= PeekIndex<LorentzIndex>(Uorg,mu);
+	// In the plane
+	U = where(mod(coor,B1)==Integer(B1-2),zzz,U); 
+	U = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+
+	// Perp links
+	U_mu = where(mod(coor,B1)==Integer(B1-3),Uorg_mu*f*f*f*f,U_mu);
+	U_mu = where(mod(coor,B1)==Integer(1)   ,Uorg_mu*f*f*f*f,U_mu);
+
+	PokeIndex<LorentzIndex>(U, U_mu, mu);
+      }
+    }
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/domains/Domains.h

@@ -2,14 +2,11 @@
 
 Grid physics library, www.github.com/paboyle/Grid
 
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+Source file: ./lib/qcd/action/momentum/Domains.h
 
-Copyright (C) 2015, 2020
+Copyright (C) 2021
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -28,24 +25,15 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
-/*  END LEGAL */
+			   /*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
 
-#ifndef AVX512
+////////////////////////////////////////////////////
-#ifndef QPX
+// Dirichlet filter with sub-block size B[mu]
-#ifndef A64FX
+////////////////////////////////////////////////////
-#ifndef A64FXFIXEDSIZE
+#pragma once 
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
 
-NAMESPACE_BEGIN(Grid);
+#include <Grid/qcd/action/domains/DomainDecomposition.h>
+#include <Grid/qcd/action/domains/MomentumFilter.h>
+#include <Grid/qcd/action/domains/DirichletFilter.h>
+#include <Grid/qcd/action/domains/DDHMCFilter.h>
 
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/domains/MomentumFilter.h

@@ -0,0 +1,91 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/MomentumFilter.h
+
+Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+//--------------------------------------------------------------------
+#pragma once 
+
+NAMESPACE_BEGIN(Grid);
+
+//These filter objects allow the user to manipulate the conjugate momentum as part of the update / refresh
+
+template<typename MomentaField>
+struct MomentumFilterBase{
+  virtual void applyFilter(MomentaField &P) const = 0;
+};
+
+//Do nothing
+template<typename MomentaField>
+struct MomentumFilterNone: public MomentumFilterBase<MomentaField>{
+  void applyFilter(MomentaField &P) const override{}
+};
+
+//Multiply each site/direction by a Lorentz vector complex number field
+//Can be used to implement a mask, zeroing out sites
+template<typename MomentaField>
+struct MomentumFilterApplyPhase: public MomentumFilterBase<MomentaField>{
+  typedef typename MomentaField::vector_type vector_type; //SIMD-vectorized complex type
+  typedef typename MomentaField::scalar_type scalar_type; //scalar complex type
+  typedef iVector<iScalar<iScalar<vector_type> >, Nd > LorentzScalarType; //complex phase for each site/direction
+  typedef Lattice<LorentzScalarType> LatticeLorentzScalarType;
+  
+  LatticeLorentzScalarType phase;
+ 
+  MomentumFilterApplyPhase(const LatticeLorentzScalarType _phase): phase(_phase){}
+
+  //Default to uniform field of (1,0)
+  MomentumFilterApplyPhase(GridBase* _grid): phase(_grid){
+    LorentzScalarType one;
+    for(int mu=0;mu<Nd;mu++)
+      one(mu)()() = scalar_type(1.);
+    
+    phase = one;
+  }
+
+  void applyFilter(MomentaField &P) const override{
+    conformable(P,phase);
+    autoView( P_v , P, AcceleratorWrite);
+    autoView( phase_v , phase, AcceleratorRead);
+
+    accelerator_for(ss,P_v.size(),MomentaField::vector_type::Nsimd(),{
+    	auto site_mom = P_v(ss);
+    	auto site_phase = phase_v(ss);
+	for(int mu=0;mu<Nd;mu++)
+	  site_mom(mu) = site_mom(mu) * site_phase(mu);
+    	coalescedWrite(P_v[ss], site_mom);
+      });
+    
+  }
+
+
+};
+
+
+
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -60,6 +60,8 @@ public:
   ///////////////////////////////////////////////////////////////
   virtual void Dminus(const FermionField &psi, FermionField &chi);
   virtual void DminusDag(const FermionField &psi, FermionField &chi);
+  virtual void ImportFourDimPseudoFermion(const FermionField &input,FermionField &imported);
+  virtual void ExportFourDimPseudoFermion(const FermionField &solution,FermionField &exported);
   virtual void ExportPhysicalFermionSolution(const FermionField &solution5d,FermionField &exported4d);
   virtual void ExportPhysicalFermionSource(const FermionField &solution5d, FermionField &exported4d);
   virtual void ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d);

Grid/qcd/action/fermion/DirichletFermionOperator.h

@@ -0,0 +1,185 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/DirichletFermionOperator.h
+
+    Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+////////////////////////////////////////////////////////////////
+// Wrap a fermion operator in Dirichlet BC's at node boundary
+////////////////////////////////////////////////////////////////
+    
+template<class Impl>
+class DirichletFermionOperator : public FermionOperator<Impl>
+{
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+
+  // Data members
+  int CommsMode;
+  Coordinate Block;
+  DirichletFilter<GaugeField> Filter;
+  FermionOperator<Impl> & FermOp;
+  
+  // Constructor / bespoke
+  DirichletFermionOperator(FermionOperator<Impl> & _FermOp, Coordinate &_Block)
+    : FermOp(_FermOp), Block(_Block), Filter(Block)
+  {
+    // Save what the comms mode should be under normal BCs
+    CommsMode = WilsonKernelsStatic::Comms;
+    assert((CommsMode == WilsonKernelsStatic::CommsAndCompute)
+         ||(CommsMode == WilsonKernelsStatic::CommsThenCompute));
+
+    // Check the block size divides local lattice
+    GridBase *grid = FermOp.GaugeGrid();
+
+    int blocks_per_rank = 1;
+    Coordinate LocalDims = grid->LocalDimensions();
+    Coordinate GlobalDims= grid->GlobalDimensions();
+    assert(Block.size()==LocalDims.size());
+
+    for(int d=0;d<LocalDims.size();d++){
+      if (Block[d]&&(Block[d]<=GlobalDims[d])){
+	int r = LocalDims[d] % Block[d];
+	assert(r == 0);
+	blocks_per_rank *= (LocalDims[d] / Block[d]);
+      }
+    }
+    // Even blocks per node required // could be relaxed but inefficient use of hardware as idle nodes in boundary operator R
+    assert( blocks_per_rank != 0);
+
+    // Possible checks that SIMD lanes are used with full occupancy???
+  };
+  virtual ~DirichletFermionOperator(void) = default;
+
+  void DirichletOn(void)   {
+    assert(WilsonKernelsStatic::Comms!= WilsonKernelsStatic::CommsDirichlet);
+    //    WilsonKernelsStatic::Comms = WilsonKernelsStatic::CommsDirichlet;
+  }
+  void DirichletOff(void)  {
+    //    assert(WilsonKernelsStatic::Comms== WilsonKernelsStatic::CommsDirichlet);
+    //    WilsonKernelsStatic::Comms = CommsMode;
+  }
+
+  // Implement the full interface
+  virtual FermionField &tmp(void) { return FermOp.tmp(); };
+
+  virtual GridBase *FermionGrid(void)         { return FermOp.FermionGrid(); }
+  virtual GridBase *FermionRedBlackGrid(void) { return FermOp.FermionRedBlackGrid(); }
+  virtual GridBase *GaugeGrid(void)           { return FermOp.GaugeGrid(); }
+  virtual GridBase *GaugeRedBlackGrid(void)   { return FermOp.GaugeRedBlackGrid(); }
+  
+  // override multiply
+  virtual void  M    (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.M(in,out);    DirichletOff();  };
+  virtual void  Mdag (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.Mdag(in,out); DirichletOff();  };
+
+  // half checkerboard operaions
+  virtual void   Meooe       (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.Meooe(in,out);    DirichletOff(); };  
+  virtual void   MeooeDag    (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.MeooeDag(in,out); DirichletOff(); };
+  virtual void   Mooee       (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.Mooee(in,out);    DirichletOff(); };
+  virtual void   MooeeDag    (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.MooeeDag(in,out); DirichletOff(); };
+  virtual void   MooeeInv    (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.MooeeInv(in,out); DirichletOff(); };
+  virtual void   MooeeInvDag (const FermionField &in, FermionField &out) { DirichletOn(); FermOp.MooeeInvDag(in,out); DirichletOff(); };
+
+  // non-hermitian hopping term; half cb or both
+  virtual void Dhop  (const FermionField &in, FermionField &out,int dag) { DirichletOn(); FermOp.Dhop(in,out,dag);    DirichletOff(); };
+  virtual void DhopOE(const FermionField &in, FermionField &out,int dag) { DirichletOn(); FermOp.DhopOE(in,out,dag);  DirichletOff(); };
+  virtual void DhopEO(const FermionField &in, FermionField &out,int dag) { DirichletOn(); FermOp.DhopEO(in,out,dag);  DirichletOff(); };
+  virtual void DhopDir(const FermionField &in, FermionField &out,int dir,int disp) { DirichletOn(); FermOp.DhopDir(in,out,dir,disp);  DirichletOff(); };
+
+  // force terms; five routines; default to Dhop on diagonal
+  virtual void MDeriv  (GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.MDeriv(mat,U,V,dag);};
+  virtual void MoeDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.MoeDeriv(mat,U,V,dag);};
+  virtual void MeoDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.MeoDeriv(mat,U,V,dag);};
+  virtual void MooDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.MooDeriv(mat,U,V,dag);};
+  virtual void MeeDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.MeeDeriv(mat,U,V,dag);};
+
+  virtual void DhopDeriv  (GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.DhopDeriv(mat,U,V,dag);};
+  virtual void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.DhopDerivEO(mat,U,V,dag);};
+  virtual void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){FermOp.DhopDerivOE(mat,U,V,dag);};
+
+  virtual void  Mdiag  (const FermionField &in, FermionField &out) { Mooee(in,out);};
+  virtual void  Mdir   (const FermionField &in, FermionField &out,int dir,int disp){FermOp.Mdir(in,out,dir,disp);};
+  virtual void  MdirAll(const FermionField &in, std::vector<FermionField> &out)    {FermOp.MdirAll(in,out);};
+
+  ///////////////////////////////////////////////
+  // Updates gauge field during HMC
+  ///////////////////////////////////////////////
+  DoubledGaugeField &GetDoubledGaugeField(void){ return FermOp.GetDoubledGaugeField(); };
+  DoubledGaugeField &GetDoubledGaugeFieldE(void){ return FermOp.GetDoubledGaugeFieldE(); };
+  DoubledGaugeField &GetDoubledGaugeFieldO(void){ return FermOp.GetDoubledGaugeFieldO(); };
+  virtual void ImportGauge(const GaugeField & _U)
+  {
+    GaugeField U = _U;
+    // Filter gauge field to apply Dirichlet
+    Filter.applyFilter(U);
+    FermOp.ImportGauge(U);
+  }
+  ///////////////////////////////////////////////
+  // Physical field import/export
+  ///////////////////////////////////////////////
+  virtual void Dminus(const FermionField &psi, FermionField &chi)    { FermOp.Dminus(psi,chi); }
+  virtual void DminusDag(const FermionField &psi, FermionField &chi) { FermOp.DminusDag(psi,chi); }
+  virtual void ImportFourDimPseudoFermion(const FermionField &input,FermionField &imported)   { FermOp.ImportFourDimPseudoFermion(input,imported);}
+  virtual void ExportFourDimPseudoFermion(const FermionField &solution,FermionField &exported){ FermOp.ExportFourDimPseudoFermion(solution,exported);}
+  virtual void ImportPhysicalFermionSource(const FermionField &input,FermionField &imported)  { FermOp.ImportPhysicalFermionSource(input,imported);}
+  virtual void ImportUnphysicalFermion(const FermionField &input,FermionField &imported)      { FermOp.ImportUnphysicalFermion(input,imported);}
+  virtual void ExportPhysicalFermionSolution(const FermionField &solution,FermionField &exported) {FermOp.ExportPhysicalFermionSolution(solution,exported);}
+  virtual void ExportPhysicalFermionSource(const FermionField &solution,FermionField &exported)   {FermOp.ExportPhysicalFermionSource(solution,exported);}
+  //////////////////////////////////////////////////////////////////////
+  // Should never be used
+  //////////////////////////////////////////////////////////////////////
+  virtual void MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
+  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary,std::vector<double> twist) {assert(0);}
+  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass) { assert(0);}
+  virtual void ContractConservedCurrent(PropagatorField &q_in_1,
+					PropagatorField &q_in_2,
+					PropagatorField &q_out,
+					PropagatorField &phys_src,
+					Current curr_type,
+					unsigned int mu)
+  {assert(0);};
+  virtual void SeqConservedCurrent(PropagatorField &q_in, 
+				   PropagatorField &q_out,
+				   PropagatorField &phys_src,
+				   Current curr_type,
+				   unsigned int mu,
+				   unsigned int tmin, 
+				   unsigned int tmax,
+				   ComplexField &lattice_cmplx)
+  {assert(0);};
+      // Only reimplemented in Wilson5D 
+      // Default to just a zero correlation function
+  virtual void ContractJ5q(FermionField &q_in   ,ComplexField &J5q) { J5q=Zero(); };
+  virtual void ContractJ5q(PropagatorField &q_in,ComplexField &J5q) { J5q=Zero(); };
+  
+};
+
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/fermion/Fermion.h

@@ -101,6 +101,12 @@ NAMESPACE_CHECK(WilsonTM5);
 #include <Grid/qcd/action/fermion/PauliVillarsInverters.h>
 #include <Grid/qcd/action/fermion/Reconstruct5Dprop.h>
 #include <Grid/qcd/action/fermion/MADWF.h>
+////////////////////////////////////////////////////////////////////
+// DDHMC related 
+////////////////////////////////////////////////////////////////////
+#include <Grid/qcd/action/fermion/DirichletFermionOperator.h>
+#include <Grid/qcd/action/fermion/SchurFactoredFermionOperator.h>
+
 NAMESPACE_CHECK(DWFutils);
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -115,9 +121,9 @@ typedef WilsonFermion<WilsonImplR> WilsonFermionR;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
-typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
+//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
-typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
+//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
-typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
+//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
 
 typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
@@ -158,41 +164,41 @@ typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
 
-typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
+//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
-typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
+//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
-typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
+//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
 
 typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
+//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
-typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
+//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
-typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
+//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
 
 typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
+//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
-typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
+//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
-typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
+//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
 
 typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
+//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
-typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
+//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
-typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
+//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
 
 typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
+//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
-typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
+//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
-typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
+//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
 
 // Ls vectorised
 typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
@@ -235,49 +241,49 @@ typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 
-typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
+//typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
-typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
+//typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
-typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
+//typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
 
 typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
+//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
-typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
+//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
-typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
+//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
 
 typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
+//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
-typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
+//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
-typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
+//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
 
 typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
+//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
-typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
+//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
-typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
+//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
 
 typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
+//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
-typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
+//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
-typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
+//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
 
 typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 
-typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
+//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
-typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
+//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
-typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
+//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
 
 typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
@@ -291,12 +297,6 @@ 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/FermionCore.h

@@ -25,8 +25,7 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef  GRID_QCD_FERMION_CORE_H
+#pragma once
-#define  GRID_QCD_FERMION_CORE_H
 
 #include <Grid/GridCore.h>
 #include <Grid/GridQCDcore.h>
@@ -45,4 +44,3 @@ NAMESPACE_CHECK(FermionOperator);
 #include <Grid/qcd/action/fermion/StaggeredKernels.h>        //used by all wilson type fermions
 NAMESPACE_CHECK(Kernels);
 
-#endif

Grid/qcd/action/fermion/FermionOperator.h

@@ -140,6 +140,9 @@ public:
   // Updates gauge field during HMC
   ///////////////////////////////////////////////
   virtual void ImportGauge(const GaugeField & _U)=0;
+  virtual DoubledGaugeField &GetDoubledGaugeField(void)  =0;
+  virtual DoubledGaugeField &GetDoubledGaugeFieldE(void)  =0;
+  virtual DoubledGaugeField &GetDoubledGaugeFieldO(void)  =0;
 
   //////////////////////////////////////////////////////////////////////
   // Conserved currents, either contract at sink or insert sequentially.
@@ -171,6 +174,16 @@ public:
       ///////////////////////////////////////////////
       virtual void Dminus(const FermionField &psi, FermionField &chi)    { chi=psi; }
       virtual void DminusDag(const FermionField &psi, FermionField &chi) { chi=psi; }
+
+      virtual void ImportFourDimPseudoFermion(const FermionField &input,FermionField &imported)
+      {
+	imported = input;
+      };
+      virtual void ExportFourDimPseudoFermion(const FermionField &solution,FermionField &exported)
+      {
+	exported=solution;
+      };
+
       virtual void ImportPhysicalFermionSource(const FermionField &input,FermionField &imported)
       {
 	imported = input;

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 const typename ViewMap<FermionField>::Type      FermionFieldView;	\
-  typedef typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
+  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> 
+// Deprecate Vec5d
-NAMESPACE_CHECK(ImplStaggered5dVec);  
+//#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
+//NAMESPACE_CHECK(ImplStaggered5dVec);  
 
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -30,6 +30,18 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+/*
+  Policy implementation for G-parity boundary conditions
+
+  Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular
+  field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first
+  with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field
+  to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to
+  either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor.
+
+  Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
+  mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+ */
 template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal>
 class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > {
 public:
@@ -97,42 +109,30 @@ 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] ) {
-      
-      int permute_lane = (sl==1) 
-    	|| ((distance== 1)&&(icoor[direction]==1))
-	|| ((distance==-1)&&(icoor[direction]==0));
 
-      if ( permute_lane ) { 
+    auto UU0=coalescedRead(U(0)(mu));
-	tmp(0) = chi(1);
+    auto UU1=coalescedRead(U(1)(mu));
-	tmp(1) = chi(0);
+    
-      } else {
+    //Decide whether we do a G-parity flavor twist
-	tmp(0) = chi(0);
+    //Note: this assumes (but does not check) that sl==1 || sl==2 i.e. max 2 SIMD lanes in G-parity dir
-	tmp(1) = chi(1);
+    //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));
 
-      auto UU0=coalescedRead(U(0)(mu));
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction
-      auto UU1=coalescedRead(U(1)(mu));
 
-      mult(&phi(0),&UU0,&tmp(0));
+    //Apply the links
-      mult(&phi(1),&UU1,&tmp(1));
+    int f_upper = permute_lane ? 1 : 0;
+    int f_lower = !f_upper;
 
-    } else {
+    mult(&phi(0),&UU0,&chi(f_upper));
-
+    mult(&phi(1),&UU1,&chi(f_lower));
-      auto UU0=coalescedRead(U(0)(mu));
-      auto UU1=coalescedRead(U(1)(mu));
-
-      mult(&phi(0),&UU0,&chi(0));
-      mult(&phi(1),&UU1,&chi(1));
-
-    }
 
 #else
     typedef _Spinor vobj;
@@ -151,10 +151,10 @@ public:
     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code
     assert((sl == 1) || (sl == 2));
 
-    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
+    //If this site is an global boundary site, perform the G-parity flavor twist
-
+    if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) {
       if ( sl == 2 ) {
-       
+	//Only do the twist for lanes on the edge of the physical node
 	ExtractBuffer<sobj> vals(Nsimd);
 
 	extract(chi,vals);
@@ -209,6 +209,19 @@ public:
     reg = memory;
   }
 
+
+  //Poke 'poke_f0' onto flavor 0 and 'poke_f1' onto flavor 1 in direction mu of the doubled gauge field Uds
+  inline void pokeGparityDoubledGaugeField(DoubledGaugeField &Uds, const GaugeLinkField &poke_f0, const GaugeLinkField &poke_f1, const int mu){
+    autoView(poke_f0_v, poke_f0, CpuRead);
+    autoView(poke_f1_v, poke_f1, CpuRead);
+    autoView(Uds_v, Uds, CpuWrite);
+    thread_foreach(ss,poke_f0_v,{
+	Uds_v[ss](0)(mu) = poke_f0_v[ss]();
+	Uds_v[ss](1)(mu) = poke_f1_v[ss]();
+      });
+  }
+    
+
   inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
   {
     conformable(Uds.Grid(),GaugeGrid);
@@ -219,14 +232,19 @@ public:
     GaugeLinkField Uconj(GaugeGrid);
    
     Lattice<iScalar<vInteger> > coor(GaugeGrid);
-        
+
-    for(int mu=0;mu<Nd;mu++){
+    //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
-          
+    //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs        
-      LatticeCoordinate(coor,mu);
+    for(int mu=0;mu<Nd-1;mu++){
+
+      if( Params.twists[mu] ){
+	LatticeCoordinate(coor,mu);
+      }
           
       U     = PeekIndex<LorentzIndex>(Umu,mu);
       Uconj = conjugate(U);
      
+      // Implement the isospin rotation sign on the boundary between f=1 and f=0
       // This phase could come from a simple bc 1,1,-1,1 ..
       int neglink = GaugeGrid->GlobalDimensions()[mu]-1;
       if ( Params.twists[mu] ) { 
@@ -241,7 +259,7 @@ public:
 	thread_foreach(ss,U_v,{
 	    Uds_v[ss](0)(mu) = U_v[ss]();
 	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
-	  });
+	});
       }
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
@@ -272,6 +290,38 @@ public:
         });
       }
     }
+
+    { //periodic / antiperiodic temporal BCs
+      int mu = Nd-1;
+      int L   = GaugeGrid->GlobalDimensions()[mu];
+      int Lmu = L - 1;
+
+      LatticeCoordinate(coor, mu);
+
+      U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links
+      
+      GaugeLinkField *Upoke = &U;
+
+      if(Params.twists[mu]){ //antiperiodic
+	Utmp =  where(coor == Lmu, -U, U);
+	Upoke = &Utmp;
+      }
+    
+      Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links      
+      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu);
+
+      //Get the barrel-shifted field
+      Utmp = adj(Cshift(U, mu, -1)); //is a forward shift!
+      Upoke = &Utmp;
+
+      if(Params.twists[mu]){
+	U = where(coor == 0, -Utmp, Utmp);  //boundary phase
+	Upoke = &U;
+      }
+      
+      Uconj = conjugate(*Upoke);
+      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4);
+    }
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
@@ -310,28 +360,48 @@ public:
   inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
     assert(0);
   }
-  
+ 
   inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
-
+    int Ls=Btilde.Grid()->_fdimensions[0];
-    int Ls = Btilde.Grid()->_fdimensions[0];
+    
-        
-    GaugeLinkField tmp(mat.Grid());
-    tmp = Zero();
     {
-      autoView( tmp_v , tmp, CpuWrite);
+      GridBase *GaugeGrid = mat.Grid();
-      autoView( Atilde_v , Atilde, CpuRead);
+      Lattice<iScalar<vInteger> > coor(GaugeGrid);
-      autoView( Btilde_v , Btilde, CpuRead);
+
-      thread_for(ss,tmp.Grid()->oSites(),{
+      if( Params.twists[mu] ){
-	  for (int s = 0; s < Ls; s++) {
+	LatticeCoordinate(coor,mu);
-	    int sF = s + Ls * ss;
+      }
-	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+
-	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+      autoView( mat_v , mat, AcceleratorWrite);
-	  }
+      autoView( Btilde_v , Btilde, AcceleratorRead);
-	});
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{	  
+  	  int sU=sss;
+  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
+  	  ColorMatrixType sum;
+  	  zeroit(sum);
+  	  for(int s=0;s<Ls;s++){
+  	    int sF = s+Ls*sU;
+  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
+	      //Flavor 0
+  	      auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector
+  	      auto aa = coalescedRead(Atilde_v[sF](0)(spn) );
+  	      sum = sum + outerProduct(bb,aa);
+
+  	      //Flavor 1
+  	      bb = coalescedRead(Btilde_v[sF](1)(spn) );
+  	      aa = coalescedRead(Atilde_v[sF](1)(spn) );
+  	      sum = sum + conjugate(outerProduct(bb,aa));
+  	    }
+  	  }	    
+  	  coalescedWrite(mat_v[sU](mu)(), sum);
+  	});
     }
-    PokeIndex<LorentzIndex>(mat, tmp, mu);
-    return;
   }
+
+
+  
+
   
 };
 
@@ -339,8 +409,8 @@ 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<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
-typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
+//typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
-typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
+//typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -141,8 +141,11 @@ public:
   void ImportGauge(const GaugeField &_Uthin, const GaugeField &_Ufat);
   void ImportGaugeSimple(const GaugeField &_UUU    ,const GaugeField &_U);
   void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
-  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  virtual DoubledGaugeField &GetDoubledGaugeField(void)  override { return Umu; };
-  DoubledGaugeField &GetUUU(void) { return UUUmu; };
+  virtual DoubledGaugeField &GetDoubledGaugeFieldE(void) override { return UmuEven; };
+  virtual DoubledGaugeField &GetDoubledGaugeFieldO(void) override { return UmuOdd; };
+  virtual DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  virtual DoubledGaugeField &GetUUU(void) { return UUUmu; };
   void CopyGaugeCheckerboards(void);
 
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -160,17 +160,20 @@ public:
 			       RealD _c1=1.0, RealD _c2=1.0,RealD _u0=1.0,
 			     const ImplParams &p= ImplParams());
     
-    // DoubleStore gauge field in operator
+  // DoubleStore gauge field in operator
-    void ImportGauge      (const GaugeField &_Uthin ) { assert(0); }
+  void ImportGauge      (const GaugeField &_Uthin ) { assert(0); }
   void ImportGauge(const GaugeField &_Uthin,const GaugeField &_Ufat);
-    void ImportGaugeSimple(const GaugeField &_UUU,const GaugeField &_U);
+  void ImportGaugeSimple(const GaugeField &_UUU,const GaugeField &_U);
-    void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
+  void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
-    // Give a reference; can be used to do an assignment or copy back out after import
+  // Give a reference; can be used to do an assignment or copy back out after import
-    // if Carleton wants to cache them and not use the ImportSimple
+  // if Carleton wants to cache them and not use the ImportSimple
-    DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  virtual DoubledGaugeField &GetDoubledGaugeField(void)  override { return Umu; };
-    DoubledGaugeField &GetUUU(void) { return UUUmu; };
+  virtual DoubledGaugeField &GetDoubledGaugeFieldE(void) override { return UmuEven; };
-    void CopyGaugeCheckerboards(void);
+  virtual DoubledGaugeField &GetDoubledGaugeFieldO(void) override { return UmuOdd; };
-    
+  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  DoubledGaugeField &GetUUU(void) { return UUUmu; };
+  void CopyGaugeCheckerboards(void);
+  
   ///////////////////////////////////////////////////////////////
   // Data members require to support the functionality
   ///////////////////////////////////////////////////////////////

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); 
+    GridBase *src_grid = src.Grid();
-    std::cout << GridLogMessage << " src4 " <<norm2(src4)<<std::endl;
+
+    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/NaiveStaggeredFermion.h

@@ -135,6 +135,9 @@ public:
 
   // DoubleStore impl dependent
   void ImportGauge      (const GaugeField &_U );
+  DoubledGaugeField &GetDoubledGaugeField(void){ return Umu; };
+  DoubledGaugeField &GetDoubledGaugeFieldE(void){ return UmuEven; };
+  DoubledGaugeField &GetDoubledGaugeFieldO(void){ return UmuOdd; };
   DoubledGaugeField &GetU(void)   { return Umu ; } ;
   void CopyGaugeCheckerboards(void);
 

Grid/qcd/action/fermion/SchurFactoredFermionOperator.h

@@ -0,0 +1,534 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/SchurFactoredFermionOperator.h
+
+    Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include <Grid/qcd/utils/MixedPrecisionOperatorFunction.h>
+#include <Grid/qcd/action/domains/Domains.h>
+
+NAMESPACE_BEGIN(Grid);
+
+  ////////////////////////////////////////////////////////
+  // Some explanation of class structure for domain decomposition:
+  //
+  // Need a dirichlet operator for two flavour determinant - acts on both Omega and OmegaBar.
+  //
+  // Possible gain if the global sums and CG are run independently?? Could measure this.
+  //
+  // Types of operations
+  //
+  // 1) assemble local det dOmega det dOmegaBar pseudofermion
+  //
+  // - DirichletFermionOperator - can either do a global solve, or independent/per cell coefficients.
+  //
+  // 2) assemble dOmegaInverse and dOmegaBarInverse in R
+  //
+  // - DirichletFermionOperator - can also be used to 
+  //                                       - need two or more cells per node. Options
+  //                                       - a) solve one cell at a time, no new code, CopyRegion and reduced /split Grids
+  //                                       - b) solve multiple cells in parallel. predicated dslash implementation
+  //
+  //                                       - b) has more parallelism, experience with block solver suggest might not be aalgorithmically inefficient
+  //                                         a) has more cache friendly and easier code.
+  //                                         b) is easy to implement in a "trial" or inefficient code with projection.
+  //
+  // 3)  Additional functionality for domain operations
+  //
+  // - SchurFactoredFermionOperator  - Need a DDHMC utility - whether used in two flavour or one flavour 
+  //
+  // - dBoundary - needs non-dirichlet operator
+  // - Contains one Dirichlet Op, and one non-Dirichlet op. Implements dBoundary etc...
+  // - The Dirichlet ops can be passed to dOmega(Bar) solvers etc...
+  //
+  ////////////////////////////////////////////////////////
+
+
+template<class ImplD,class ImplF>
+class SchurFactoredFermionOperator : public ImplD
+{
+  INHERIT_IMPL_TYPES(ImplD);
+  
+  typedef typename ImplF::FermionField FermionFieldF;
+  typedef typename ImplD::FermionField FermionFieldD;
+
+  typedef SchurDiagMooeeOperator<FermionOperator<ImplD>,FermionFieldD> LinearOperatorD;
+  typedef SchurDiagMooeeOperator<FermionOperator<ImplF>,FermionFieldF> LinearOperatorF;
+  typedef SchurDiagMooeeDagOperator<FermionOperator<ImplD>,FermionFieldD> LinearOperatorDagD;
+  typedef SchurDiagMooeeDagOperator<FermionOperator<ImplF>,FermionFieldF> LinearOperatorDagF;
+
+  typedef MixedPrecisionConjugateGradientOperatorFunction<FermionOperator<ImplD>,
+							  FermionOperator<ImplF>,
+							  LinearOperatorD,
+							  LinearOperatorF> MxPCG;
+
+  typedef MixedPrecisionConjugateGradientOperatorFunction<FermionOperator<ImplD>,
+							  FermionOperator<ImplF>,
+							  LinearOperatorDagD,
+							  LinearOperatorDagF> MxDagPCG;
+public:
+
+  GridBase *FermionGrid(void) { return PeriodicFermOpD.FermionGrid(); };
+  GridBase *GaugeGrid(void)   { return PeriodicFermOpD.GaugeGrid(); };
+  
+  FermionOperator<ImplD> & DirichletFermOpD;
+  FermionOperator<ImplF> & DirichletFermOpF;
+  FermionOperator<ImplD> & PeriodicFermOpD; 
+  FermionOperator<ImplF> & PeriodicFermOpF; 
+
+  LinearOperatorD DirichletLinOpD;
+  LinearOperatorF DirichletLinOpF;
+  LinearOperatorD PeriodicLinOpD;
+  LinearOperatorF PeriodicLinOpF;
+
+  LinearOperatorDagD DirichletLinOpDagD;
+  LinearOperatorDagF DirichletLinOpDagF;
+  LinearOperatorDagD PeriodicLinOpDagD;
+  LinearOperatorDagF PeriodicLinOpDagF;
+
+  // Can tinker with these in the pseudofermion for force vs. action solves
+  Integer maxinnerit;
+  Integer maxouterit;
+  RealD tol;
+  RealD tolinner;
+  
+  Coordinate Block;
+
+  DomainDecomposition Domains;
+
+  SchurFactoredFermionOperator(FermionOperator<ImplD>  & _PeriodicFermOpD,
+			       FermionOperator<ImplF>  & _PeriodicFermOpF,
+			       FermionOperator<ImplD>  & _DirichletFermOpD,
+			       FermionOperator<ImplF>  & _DirichletFermOpF,
+			       Coordinate &_Block)
+    : Block(_Block), Domains(Block),
+
+      PeriodicFermOpD(_PeriodicFermOpD),
+      PeriodicFermOpF(_PeriodicFermOpF),
+      DirichletFermOpD(_DirichletFermOpD),
+      DirichletFermOpF(_DirichletFermOpF),
+      DirichletLinOpD(DirichletFermOpD),
+      DirichletLinOpF(DirichletFermOpF),
+      PeriodicLinOpD(PeriodicFermOpD),
+      PeriodicLinOpF(PeriodicFermOpF),
+      DirichletLinOpDagD(DirichletFermOpD),
+      DirichletLinOpDagF(DirichletFermOpF),
+      PeriodicLinOpDagD(PeriodicFermOpD),
+      PeriodicLinOpDagF(PeriodicFermOpF)
+  {
+    tol=1.0e-10;
+    tolinner=1.0e-6;
+    maxinnerit=1000;
+    maxouterit=10;
+    assert(PeriodicFermOpD.FermionGrid() == DirichletFermOpD.FermionGrid());
+    assert(PeriodicFermOpF.FermionGrid() == DirichletFermOpF.FermionGrid());
+  };
+
+  enum Domain { Omega=0, OmegaBar=1 };
+
+  void ImportGauge(const GaugeField &Umu)
+  {
+    // Single precision will update in the mixed prec CG
+    PeriodicFermOpD.ImportGauge(Umu);
+    GaugeField dUmu(Umu.Grid());
+    dUmu=Umu;
+    //    DirchletBCs(dUmu);
+    DirichletFilter<GaugeField> Filter(Block);
+    Filter.applyFilter(dUmu);
+    DirichletFermOpD.ImportGauge(dUmu);
+  }
+
+/*
+  void ProjectBoundaryBothDomains (FermionField &f,int sgn)
+  {
+    assert((sgn==1)||(sgn==-1));
+    Real rsgn = sgn;
+
+    Gamma::Algebra Gmu [] = {
+      Gamma::Algebra::GammaX,
+      Gamma::Algebra::GammaY,
+      Gamma::Algebra::GammaZ,
+      Gamma::Algebra::GammaT
+    };
+
+    GridBase *grid = f.Grid();
+    LatticeInteger  coor(grid);
+    LatticeInteger  face(grid);
+    LatticeInteger  one(grid); one = 1;
+    LatticeInteger  zero(grid); zero = 0;
+    LatticeInteger nface(grid); nface=Zero();
+    
+    FermionField projected(grid); projected=Zero();
+    FermionField sp_proj  (grid);
+
+    int dims = grid->Nd();
+    int isDWF= (dims==Nd+1);
+    assert((dims==Nd)||(dims==Nd+1));
+    Coordinate Global=grid->GlobalDimensions();
+
+    for(int mu=0;mu<Nd;mu++){
+
+      if ( Block[mu] <= Global[mu+isDWF] ) {
+	// need to worry about DWF 5th dim first
+	LatticeCoordinate(coor,mu+isDWF); 
+      
+	face = where(mod(coor,Block[mu]) == Integer(0),one,zero );
+	nface = nface + face;
+
+	Gamma G(Gmu[mu]);
+	// Lower face receives (1-gamma)/2 in normal forward hopping term
+	sp_proj  = 0.5*(f-G*f*rsgn);
+	projected= where(face,sp_proj,projected);
+	//projected= where(face,f,projected);
+      
+	face = where(mod(coor,Block[mu]) == Integer(Block[mu]-1) ,one,zero );
+	nface = nface + face;
+
+	// Upper face receives (1+gamma)/2 in normal backward hopping term
+	sp_proj = 0.5*(f+G*f*rsgn);
+	projected= where(face,sp_proj,projected);
+	//projected= where(face,f,projected);
+      }
+      
+    }
+    // Initial Zero() where nface==0.
+    // Keep the spin projected faces where nface==1
+    // Full spinor where nface>=2
+    projected = where(nface>Integer(1),f,projected);
+    f=projected;
+  }
+*/
+  void ProjectBoundaryBothDomains (FermionField &f,int sgn)
+  {
+    assert((sgn==1)||(sgn==-1));
+    Real rsgn = sgn;
+
+    Gamma::Algebra Gmu [] = {
+      Gamma::Algebra::GammaX,
+      Gamma::Algebra::GammaY,
+      Gamma::Algebra::GammaZ,
+      Gamma::Algebra::GammaT
+    };
+
+    GridBase *grid = f.Grid();
+    LatticeInteger  coor(grid);
+    LatticeInteger  face(grid);
+    LatticeInteger  one(grid);   one = 1;
+    LatticeInteger  zero(grid); zero = 0;
+    LatticeInteger  omega(grid);
+    LatticeInteger  omegabar(grid);
+    LatticeInteger  tmp(grid);
+
+    omega=one;    Domains.ProjectDomain(omega,0);
+    omegabar=one; Domains.ProjectDomain(omegabar,1);
+    
+    LatticeInteger nface(grid); nface=Zero();
+    
+    FermionField projected(grid); projected=Zero();
+    FermionField sp_proj  (grid);
+
+    int dims = grid->Nd();
+    int isDWF= (dims==Nd+1);
+    assert((dims==Nd)||(dims==Nd+1));
+    Coordinate Global=grid->GlobalDimensions();
+
+    for(int mmu=0;mmu<Nd;mmu++){
+      Gamma G(Gmu[mmu]);
+
+      // need to worry about DWF 5th dim first
+      int mu = mmu+isDWF;
+      if ( Block[mmu] && (Block[mmu] <= Global[mu]) ) {
+
+	// Lower face receives (1-gamma)/2 in normal forward hopping term
+ 	tmp = Cshift(omegabar,mu,-1);
+	tmp = tmp + omega;
+	face = where(tmp == Integer(2),one,zero );
+
+ 	tmp = Cshift(omega,mu,-1);
+	tmp = tmp + omegabar;
+	face = where(tmp == Integer(2),one,face );
+
+	nface = nface + face;
+
+	sp_proj  = 0.5*(f-G*f*rsgn);
+	projected= where(face,sp_proj,projected);
+
+	// Upper face receives (1+gamma)/2 in normal backward hopping term
+ 	tmp = Cshift(omegabar,mu,1);
+	tmp = tmp + omega;
+	face = where(tmp == Integer(2),one,zero );
+
+ 	tmp = Cshift(omega,mu,1);
+	tmp = tmp + omegabar;
+	face = where(tmp == Integer(2),one,face );
+
+	nface = nface + face;
+
+	sp_proj = 0.5*(f+G*f*rsgn);
+	projected= where(face,sp_proj,projected);
+      }
+      
+    }
+    // Initial Zero() where nface==0.
+    // Keep the spin projected faces where nface==1
+    // Full spinor where nface>=2
+    projected = where(nface>Integer(1),f,projected);
+    f=projected;
+  }
+
+  void ProjectDomain(FermionField &f,int domain)
+  {
+/*
+    GridBase *grid = f.Grid();
+    int dims = grid->Nd();
+    int isDWF= (dims==Nd+1);
+    assert((dims==Nd)||(dims==Nd+1));
+
+    FermionField zz(grid); zz=Zero();
+    LatticeInteger coor(grid);
+    LatticeInteger domaincb(grid); domaincb=Zero();
+    for(int d=0;d<Nd;d++){
+      LatticeCoordinate(coor,d+isDWF);
+      domaincb = domaincb + div(coor,Block[d]);
+    }
+    f = where(mod(domaincb,2)==Integer(domain),f,zz);
+*/
+    Domains.ProjectDomain(f,domain);
+
+  };
+  void ProjectOmegaBar   (FermionField &f) {ProjectDomain(f,OmegaBar);}
+  void ProjectOmega      (FermionField &f) {ProjectDomain(f,Omega);}
+  // See my notes(!).
+  // Notation: Following Luscher, we introduce projectors $\hPdb$ with both spinor and space structure
+  // projecting all spinor elements in $\Omega$ connected by $\Ddb$ to $\bar{\Omega}$,
+  void ProjectBoundaryBar(FermionField &f)
+  {
+    ProjectBoundaryBothDomains(f,1);
+    ProjectOmega(f);
+  }
+  // and $\hPd$ projecting all spinor elements in $\bar{\Omega}$ connected by $\Dd$ to $\Omega$.
+  void ProjectBoundary   (FermionField &f)
+  {
+    ProjectBoundaryBothDomains(f,1);
+    ProjectOmegaBar(f);
+    //    DumpSliceNorm("ProjectBoundary",f,f.Grid()->Nd()-1);
+  };
+
+  void dBoundary    (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmegaBar(tmp);
+    PeriodicFermOpD.M(tmp,out);
+    ProjectOmega(out);
+  };
+  void dBoundaryDag (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmega(tmp);
+    PeriodicFermOpD.Mdag(tmp,out);
+    ProjectOmegaBar(out);
+  };
+  void dBoundaryBar (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmega(tmp);
+    PeriodicFermOpD.M(tmp,out);
+    ProjectOmegaBar(out);
+  };
+  void dBoundaryBarDag (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmegaBar(tmp);
+    PeriodicFermOpD.Mdag(tmp,out);
+    ProjectOmega(out);
+  };
+  void dOmega       (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmega(tmp);
+    DirichletFermOpD.M(tmp,out);
+    ProjectOmega(out);
+  };
+  void dOmegaBar    (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmegaBar(tmp);
+    DirichletFermOpD.M(tmp,out);
+    ProjectOmegaBar(out);
+  };
+  void dOmegaDag       (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmega(tmp);
+    DirichletFermOpD.Mdag(tmp,out);
+    ProjectOmega(out);
+  };
+  void dOmegaBarDag    (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmegaBar(tmp);
+    DirichletFermOpD.Mdag(tmp,out);
+    ProjectOmegaBar(out);
+  };
+  void dOmegaInv   (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmega(tmp);
+    dOmegaInvAndOmegaBarInv(tmp,out); // Inefficient warning
+    ProjectOmega(out);
+  };
+  void dOmegaBarInv(FermionField &in,FermionField &out)
+  {    
+    FermionField tmp(in);
+    ProjectOmegaBar(tmp);
+    dOmegaInvAndOmegaBarInv(tmp,out);
+    ProjectOmegaBar(out);
+  };
+  void dOmegaDagInv   (FermionField &in,FermionField &out)
+  {
+    FermionField tmp(in);
+    ProjectOmega(tmp);
+    dOmegaDagInvAndOmegaBarDagInv(tmp,out);
+    ProjectOmega(out);
+  };
+  void dOmegaBarDagInv(FermionField &in,FermionField &out)
+  {    
+    FermionField tmp(in);
+    ProjectOmegaBar(tmp);
+    dOmegaDagInvAndOmegaBarDagInv(tmp,out);
+    ProjectOmegaBar(out);
+  };
+  void dOmegaInvAndOmegaBarInv(FermionField &in,FermionField &out)
+  {
+    MxPCG OmegaSolver(tol,
+		      tolinner,
+		      maxinnerit,
+		      maxouterit,
+		      DirichletFermOpF.FermionRedBlackGrid(),
+		      DirichletFermOpF,
+		      DirichletFermOpD,
+		      DirichletLinOpF,
+		      DirichletLinOpD);
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(OmegaSolver);
+    PrecSolve(DirichletFermOpD,in,out);
+  };
+  void dOmegaDagInvAndOmegaBarDagInv(FermionField &in,FermionField &out)
+  {
+    MxDagPCG OmegaDagSolver(tol,
+			    tolinner,
+			    maxinnerit,
+			    maxouterit,
+			    DirichletFermOpF.FermionRedBlackGrid(),
+			    DirichletFermOpF,
+			    DirichletFermOpD,
+			    DirichletLinOpDagF,
+			    DirichletLinOpDagD);
+    SchurRedBlackDiagMooeeDagSolve<FermionField> PrecSolve(OmegaDagSolver);
+    PrecSolve(DirichletFermOpD,in,out);
+  };
+
+  // Rdag = Pdbar - DdbarDag DomegabarDagInv  DdDag DomegaDagInv Pdbar 
+  void RDag(FermionField &in,FermionField &out)
+  {
+    FermionField tmp1(PeriodicFermOpD.FermionGrid());
+    FermionField tmp2(PeriodicFermOpD.FermionGrid());
+    out = in;
+    ProjectBoundaryBar(out);
+    dOmegaDagInv(out,tmp1);   
+    dBoundaryDag(tmp1,tmp2);   
+    dOmegaBarDagInv(tmp2,tmp1);
+    dBoundaryBarDag(tmp1,tmp2); 
+    out = out - tmp2;
+  };
+
+  // R = Pdbar - Pdbar DomegaInv Dd DomegabarInv Ddbar
+  void R(FermionField &in,FermionField &out)
+  {
+    FermionField tmp1(PeriodicFermOpD.FermionGrid());
+    FermionField tmp2(PeriodicFermOpD.FermionGrid());
+    out = in;
+    ProjectBoundaryBar(out);
+    dBoundaryBar(out,tmp1); 
+    dOmegaBarInv(tmp1,tmp2);
+    dBoundary(tmp2,tmp1);   
+    dOmegaInv(tmp1,tmp2);   
+    out = in - tmp2 ;       
+    ProjectBoundaryBar(out);
+    //    DumpSliceNorm("R",out,out.Grid()->Nd()-1);
+  };
+  
+  // R = Pdbar - Pdbar Dinv Ddbar 
+  void RInv(FermionField &in,FermionField &out)
+  {
+    FermionField tmp1(PeriodicFermOpD.FermionGrid());
+    dBoundaryBar(in,out);
+    Dinverse(out,tmp1);  
+    out =in -tmp1; 
+    ProjectBoundaryBar(out);
+  };
+  // R = Pdbar - DdbarDag DinvDag Pdbar 
+  void RDagInv(FermionField &in,FermionField &out)
+  {
+    FermionField tmp(PeriodicFermOpD.FermionGrid());
+    FermionField Pin(PeriodicFermOpD.FermionGrid());
+    Pin = in; ProjectBoundaryBar(Pin);
+    DinverseDag(Pin,out);  
+    dBoundaryBarDag(out,tmp);
+    out =Pin -tmp; 
+  };
+  // Non-dirichlet inverter using red-black preconditioning
+  void Dinverse(FermionField &in,FermionField &out)
+  {
+    MxPCG DSolver(tol,
+		  tolinner,
+		  maxinnerit,
+		  maxouterit,
+		  PeriodicFermOpF.FermionRedBlackGrid(),
+		  PeriodicFermOpF,
+		  PeriodicFermOpD,
+		  PeriodicLinOpF,
+		  PeriodicLinOpD);
+    SchurRedBlackDiagMooeeSolve<FermionField> Solve(DSolver);
+    Solve(PeriodicFermOpD,in,out);
+  }
+  void DinverseDag(FermionField &in,FermionField &out)
+  {
+    MxDagPCG DdagSolver(tol,
+			tolinner,
+			maxinnerit,
+			maxouterit,
+			PeriodicFermOpF.FermionRedBlackGrid(),
+			PeriodicFermOpF,
+			PeriodicFermOpD,
+			PeriodicLinOpDagF,
+			PeriodicLinOpDagD);
+    SchurRedBlackDiagMooeeDagSolve<FermionField> Solve(DdagSolver);
+    Solve(PeriodicFermOpD,in,out);
+  }
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/fermion/StaggeredImpl.h

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

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -245,7 +245,7 @@ public:
     return out;
   }
 
-private:
+protected:
   // here fixing the 4 dimensions, make it more general?
 
   RealD csw_r;                                               // Clover coefficient - spatial

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,const SiteSpinor &in) const {
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
+    typedef decltype(coalescedRead(buf)) sobj;
-    _SiteHalfSpinor tmp;
+    sobj sp;
-    projector::Proj(tmp,in,mu,dag);
+    auto sin = coalescedRead(in);
-    vstream(buf[o],tmp);
+    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,const SiteSpinor &in) const {
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
+    SiteHalfSpinor hsp;
-    _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;
+  void ZeroCountersi(void)  {  }
-  double timer1;
+  void Reporti(int calls)  {  }
-  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;
-  }
 
   std::vector<int> surface_list;
 
@@ -287,9 +303,11 @@ public:
 		int npoints,
 		int checkerboard,
 		const std::vector<int> &directions,
-		const std::vector<int> &distances,Parameters p)  
+		const std::vector<int> &distances,
-    : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
+		bool locally_periodic,
-  { 
+		Parameters p)  
+    : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,locally_periodic,p)
+  {
     ZeroCountersi();
     surface_list.resize(0);
     this->same_node.resize(npoints);
@@ -321,26 +339,18 @@ 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 +362,9 @@ 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 +400,6 @@ 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

@@ -146,8 +146,11 @@ public:
   void DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
                     const FermionField &in, FermionField &out, int dag);
 
+  void DhopInternalDirichletComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+				  const FermionField &in, FermionField &out, int dag);
+  
   void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
-                    const FermionField &in, FermionField &out, int dag);
+				   const FermionField &in, FermionField &out, int dag);
 
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
@@ -157,7 +160,10 @@ public:
 
   // DoubleStore impl dependent
   void ImportGauge(const GaugeField &_Umu);
-
+  DoubledGaugeField &GetDoubledGaugeField(void){ return Umu; };
+  DoubledGaugeField &GetDoubledGaugeFieldE(void){ return UmuEven; };
+  DoubledGaugeField &GetDoubledGaugeFieldO(void){ return UmuOdd; };
+  
   ///////////////////////////////////////////////////////////////
   // Data members require to support the functionality
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -165,7 +165,14 @@ public:
 			       const FermionField &in, 
 			       FermionField &out,
 			       int dag);
-    
+
+  void DhopInternalDirichletComms(StencilImpl & st,
+				  LebesgueOrder &lo,
+				  DoubledGaugeField &U,
+				  const FermionField &in, 
+				  FermionField &out,
+				  int dag);
+  
   // Constructors
   WilsonFermion5D(GaugeField &_Umu,
 		  GridCartesian         &FiveDimGrid,
@@ -174,19 +181,11 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
     
-  // Constructors
-  /*
-    WilsonFermion5D(int simd, 
-    GaugeField &_Umu,
-    GridCartesian         &FiveDimGrid,
-    GridRedBlackCartesian &FiveDimRedBlackGrid,
-    GridCartesian         &FourDimGrid,
-    double _M5,const ImplParams &p= ImplParams());
-  */
-    
   // DoubleStore
   void ImportGauge(const GaugeField &_Umu);
-    
+  DoubledGaugeField &GetDoubledGaugeField(void){ return Umu; };
+  DoubledGaugeField &GetDoubledGaugeFieldE(void){ return UmuEven; };
+  DoubledGaugeField &GetDoubledGaugeFieldO(void){ return UmuOdd; };
   ///////////////////////////////////////////////////////////////
   // Data members require to support the functionality
   ///////////////////////////////////////////////////////////////

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,{
+    typedef decltype(coalescedRead(out_v[0]))   calcSpinor;
-	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
+    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<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
-typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
+//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
-typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
+//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<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
-typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
+//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
-typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
+//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

@@ -39,7 +39,7 @@ NAMESPACE_BEGIN(Grid);
 class WilsonKernelsStatic { 
 public:
   enum { OptGeneric, OptHandUnroll, OptInlineAsm };
-  enum { CommsAndCompute, CommsThenCompute };
+  enum { CommsAndCompute, CommsThenCompute, CommsDirichlet };
   static int Opt;  
   static int Comms;
 };
@@ -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

@@ -112,7 +112,6 @@ void CayleyFermion5D<Impl>::ImportUnphysicalFermion(const FermionField &input4d,
   axpby_ssp_pminus(tmp, 0., tmp, 1., tmp, Ls-1, Ls-1);
   imported5d=tmp;
 }
-
 template<class Impl>  
 void CayleyFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
 {
@@ -127,6 +126,37 @@ void CayleyFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &inpu
   axpby_ssp_pminus(tmp, 0., tmp, 1., tmp, Ls-1, Ls-1);
   Dminus(tmp,imported5d);
 }
+////////////////////////////////////////////////////
+// Added for fourD pseudofermion det estimation
+////////////////////////////////////////////////////
+template<class Impl>  
+void CayleyFermion5D<Impl>::ImportFourDimPseudoFermion(const FermionField &input4d,FermionField &imported5d)
+{
+  int Ls = this->Ls;
+  FermionField tmp(this->FermionGrid());
+  conformable(imported5d.Grid(),this->FermionGrid());
+  conformable(input4d.Grid()   ,this->GaugeGrid());
+  tmp = Zero();
+  InsertSlice(input4d, tmp, 0   , 0);
+  InsertSlice(input4d, tmp, Ls-1, 0);
+  axpby_ssp_pminus(tmp, 0., tmp, 1., tmp, 0, 0);
+  axpby_ssp_pplus (tmp, 0., tmp, 1., tmp, Ls-1, Ls-1);
+  imported5d=tmp;
+}
+template<class Impl>  
+void CayleyFermion5D<Impl>::ExportFourDimPseudoFermion(const FermionField &solution5d,FermionField &exported4d)
+{
+  int Ls = this->Ls;
+  FermionField tmp(this->FermionGrid());
+  tmp = solution5d;
+  conformable(solution5d.Grid(),this->FermionGrid());
+  conformable(exported4d.Grid(),this->GaugeGrid());
+  axpby_ssp_pminus(tmp, 0., solution5d, 1., solution5d, 0, 0);
+  axpby_ssp_pplus (tmp, 1., tmp       , 1., solution5d, 0, Ls-1);
+  ExtractSlice(exported4d, tmp, 0, 0);
+}
+
+// Dminus
 template<class Impl>  
 void CayleyFermion5D<Impl>::Dminus(const FermionField &psi, FermionField &chi)
 {
@@ -642,7 +672,7 @@ 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,
@@ -826,7 +856,7 @@ 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;
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
@@ -880,17 +910,29 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
+  RealD sign = 1; // 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;    
+    }
+    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 sr = Ls-1-s;
-    int srp= (sr+1)%Ls;
+    //    int srp= (sr+1)%Ls;
 
     // Mobius parameters
     auto b=this->bs[s];
@@ -907,7 +949,7 @@ 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
 

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

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

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

@@ -32,25 +32,50 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-#define LOAD_CHI(b)		\
+#ifdef GRID_SIMT
+
+#define LOAD_CHI(ptype,b)			\
+  const SiteSpinor & ref (b[offset]);				\
+  Chi_0=coalescedReadPermute<ptype>(ref()()(0),perm,lane);	\
+  Chi_1=coalescedReadPermute<ptype>(ref()()(1),perm,lane);	\
+  Chi_2=coalescedReadPermute<ptype>(ref()()(2),perm,lane);
+
+#define LOAD_CHI_COMMS(b)		\
   const SiteSpinor & ref (b[offset]);	\
-    Chi_0=ref()()(0);\
+  Chi_0=coalescedRead(ref()()(0),lane);	\
-    Chi_1=ref()()(1);\
+  Chi_1=coalescedRead(ref()()(1),lane);	\
-    Chi_2=ref()()(2);
+  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));      \
+    U_00=coalescedRead(ref()(0,0),lane);				\
-   Impl::loadLinkElement(U_10,ref()(1,0));      \
+    U_10=coalescedRead(ref()(1,0),lane);				\
-   Impl::loadLinkElement(U_20,ref()(2,0));      \
+    U_20=coalescedRead(ref()(2,0),lane);				\
-   Impl::loadLinkElement(U_01,ref()(0,1));      \
+    U_01=coalescedRead(ref()(0,1),lane);				\
-   Impl::loadLinkElement(U_11,ref()(1,1));      \
+    U_11=coalescedRead(ref()(1,1),lane);				\
-   Impl::loadLinkElement(U_21,ref()(2,1));      \
+    U_21=coalescedRead(ref()(2,1),lane);				\
-   Impl::loadLinkElement(U_02,ref()(0,2));     \
+    U_02=coalescedRead(ref()(0,2),lane);				\
-   Impl::loadLinkElement(U_12,ref()(1,2));     \
+    U_12=coalescedRead(ref()(1,2),lane);				\
-   Impl::loadLinkElement(U_22,ref()(2,2));     \
+    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));      \
+    U_00=coalescedRead(ref()(0,0),lane);				\
-   Impl::loadLinkElement(U_10,ref()(1,0));      \
+    U_10=coalescedRead(ref()(1,0),lane);				\
-   Impl::loadLinkElement(U_20,ref()(2,0));      \
+    U_20=coalescedRead(ref()(2,0),lane);				\
-   Impl::loadLinkElement(U_01,ref()(0,1));      \
+    U_01=coalescedRead(ref()(0,1),lane);				\
-   Impl::loadLinkElement(U_11,ref()(1,1));      \
+    U_11=coalescedRead(ref()(1,1),lane);				\
-   Impl::loadLinkElement(U_21,ref()(2,1));      \
+    U_21=coalescedRead(ref()(2,1),lane);				\
-   Impl::loadLinkElement(U_02,ref()(0,2));     \
+    U_02=coalescedRead(ref()(0,2),lane);				\
-   Impl::loadLinkElement(U_12,ref()(1,2));     \
+    U_12=coalescedRead(ref()(1,2),lane);				\
-   Impl::loadLinkElement(U_22,ref()(2,2));     \
+    U_22=coalescedRead(ref()(2,2),lane);				\
     UChi ## _0 += U_00*Chi_0;	       \
     UChi ## _1 += U_10*Chi_0;\
     UChi ## _2 += U_20*Chi_0;\
@@ -83,24 +108,18 @@ NAMESPACE_BEGIN(Grid);
     UChi ## _2 += U_22*Chi_2;
 
 
-#define PERMUTE_DIR(dir)			\
-  permute##dir(Chi_0,Chi_0);			\
-  permute##dir(Chi_1,Chi_1);			\
-  permute##dir(Chi_2,Chi_2);
-
-
 #define HAND_STENCIL_LEG_BASE(Dir,Perm,skew)	\
   SE=st.GetEntry(ptype,Dir+skew,sF);	\
   offset = SE->_offset;			\
   local  = SE->_is_local;		\
   perm   = SE->_permute;		\
   if ( local ) {						\
-    LOAD_CHI(in);					\
+    LOAD_CHI(Perm,in);						\
     if ( perm) {						\
       PERMUTE_DIR(Perm);					\
     }								\
   } else {							\
-    LOAD_CHI(buf);						\
+    LOAD_CHI_COMMS(buf);					\
   }								
 
 #define HAND_STENCIL_LEG_BEGIN(Dir,Perm,skew,even)		\
@@ -116,19 +135,18 @@ NAMESPACE_BEGIN(Grid);
   }
 
 
-
 #define HAND_STENCIL_LEG_INT(U,Dir,Perm,skew,even)	\
   SE=st.GetEntry(ptype,Dir+skew,sF);			\
   offset = SE->_offset;					\
   local  = SE->_is_local;				\
   perm   = SE->_permute;				\
   if ( local ) {					\
-    LOAD_CHI(in);				\
+    LOAD_CHI(Perm,in);					\
     if ( perm) {					\
       PERMUTE_DIR(Perm);				\
     }							\
   } else if ( st.same_node[Dir] ) {			\
-    LOAD_CHI(buf);					\
+    LOAD_CHI_COMMS(buf);				\
   }							\
   if (local || st.same_node[Dir] ) {		\
     MULT_ADD(U,Dir,even);				\
@@ -140,10 +158,32 @@ NAMESPACE_BEGIN(Grid);
   local  = SE->_is_local;				\
   if ((!local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
-    { LOAD_CHI(buf);	  }					\
+    { LOAD_CHI_COMMS(buf);	  }				\
     { MULT_ADD(U,Dir,even); }					\
   }								
 
+#define HAND_DECLARATIONS(Simd) \
+  Simd even_0;			\
+  Simd even_1;			\
+  Simd even_2;			\
+  Simd odd_0;			\
+  Simd odd_1;			\
+  Simd odd_2;		        \
+		      		\
+  Simd Chi_0;			\
+  Simd Chi_1;			\
+  Simd Chi_2;			\
+				\
+  Simd U_00;			\
+  Simd U_10;			\
+  Simd U_20;			\
+  Simd U_01;			\
+  Simd U_11;			\
+  Simd U_21;			\
+  Simd U_02;			\
+  Simd U_12;			\
+  Simd U_22;			
+  
 
 template <class Impl>
 template <int Naik> accelerator_inline
@@ -155,28 +195,14 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
 
-  Simd Chi_0;    // two spinor; 6 regs
+  const int Nsimd = SiteHalfSpinor::Nsimd();
-  Simd Chi_1;
+  const int lane=acceleratorSIMTlane(Nsimd);
-  Simd Chi_2;
+  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;
@@ -215,7 +241,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
       result()()(1) = even_1 + odd_1;
       result()()(2) = even_2 + odd_2;
     }
-    vstream(out[sF],result);
+    coalescedWrite(out[sF],result);
   }
 }
 
@@ -230,28 +256,13 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
+  const int Nsimd = SiteHalfSpinor::Nsimd();
-  Simd even_1;
+  const int lane=acceleratorSIMTlane(Nsimd);
-  Simd even_2;
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
-  Simd odd_0; // 12 regs on knc
+  HAND_DECLARATIONS(Simt);
-  Simd odd_1;
-  Simd odd_2;
 
-  Simd Chi_0;    // two spinor; 6 regs
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
-  Simd Chi_1;
+  calcSiteSpinor result;
-  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;
   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();
+    zeroit(even_0);    zeroit(even_1);    zeroit(even_2);
-     odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
+    zeroit(odd_0);    zeroit(odd_1);    zeroit(odd_2);
 
     skew = 0;
     HAND_STENCIL_LEG_INT(U,Xp,3,skew,even);  
@@ -294,7 +305,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
       result()()(1) = even_1 + odd_1;
       result()()(2) = even_2 + odd_2;
     }
-    vstream(out[sF],result);
+    coalescedWrite(out[sF],result);
   }
 }
 
@@ -309,28 +320,13 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
+  const int Nsimd = SiteHalfSpinor::Nsimd();
-  Simd even_1;
+  const int lane=acceleratorSIMTlane(Nsimd);
-  Simd even_2;
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
-  Simd odd_0; // 12 regs on knc
+  HAND_DECLARATIONS(Simt);
-  Simd odd_1;
-  Simd odd_2;
 
-  Simd Chi_0;    // two spinor; 6 regs
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
-  Simd Chi_1;
+  calcSiteSpinor result;
-  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;
   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();
+    zeroit(even_0);    zeroit(even_1);    zeroit(even_2);
-     odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
+    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) {						\
+    int perm= SE->_permute;						\
-      chi_p = χ						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
-      permute(chi,  in[SE->_offset], ptype);			\
-    } else {							\
-      chi_p = &in[SE->_offset];					\
-    }								\
   } else {							\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  multLink(Uchi, U[sU], *chi_p, Dir);			
+  acceleratorSynchronise();					\
+  multLink(Uchi, U[sU], chi, Dir);			
 
 #define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
+    int perm= SE->_permute;						\
-      chi_p = χ						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
-      permute(chi,  in[SE->_offset], ptype);			\
-    } else {							\
-      chi_p = &in[SE->_offset];					\
-    }								\
   } 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];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
-    multLink(Uchi, U[sU], *chi_p, Dir);				\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 template <class Impl>
@@ -84,12 +77,14 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
 					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
-  const SiteSpinor *chi_p;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
-  SiteSpinor chi;
+  calcSpinor chi;
-  SiteSpinor Uchi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //
@@ -118,7 +113,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
     if ( dag ) { 
       Uchi = - Uchi;
     } 
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -130,13 +125,16 @@ template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 						SiteSpinor *buf, int sF, int sU, 
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
+						const FermionFieldView &in, FermionFieldView &out,int dag)
-  const SiteSpinor *chi_p;
+{
-  SiteSpinor chi;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
-  SiteSpinor Uchi;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //    int sF=LLs*sU+s;
@@ -165,7 +163,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
     if ( dag ) {
       Uchi = - Uchi;
     }
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -178,14 +176,17 @@ template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 						SiteSpinor *buf, int sF, int sU,
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
+						const FermionFieldView &in, FermionFieldView &out,int dag)
-  const SiteSpinor *chi_p;
+{
-  //  SiteSpinor chi;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
-  SiteSpinor Uchi;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int nmu=0;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //    int sF=LLs*sU+s;
@@ -211,11 +212,12 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
     GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
     }
-    if ( nmu ) { 
+    if ( nmu ) {
-      if ( dag ) { 
+      auto _out = coalescedRead(out[sF],lane);
-	out[sF] = out[sF] - Uchi;
+      if ( dag ) {
+	coalescedWrite(out[sF], _out-Uchi,lane);
       } else { 
-	out[sF] = out[sF] + Uchi;
+	coalescedWrite(out[sF], _out+Uchi,lane);
       }
     }
   }
@@ -261,6 +263,8 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo,
   GridBase *FGrid=in.Grid();  
   GridBase *UGrid=U.Grid();  
   typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
   autoView( UUU_v , UUU, AcceleratorRead);
   autoView( U_v   ,   U, AcceleratorRead);
   autoView( in_v  ,  in, AcceleratorRead);
@@ -301,6 +305,8 @@ void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo,
   GridBase *FGrid=in.Grid();  
   GridBase *UGrid=U.Grid();  
   typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
   autoView( UUU_v ,   U, AcceleratorRead);
   autoView( U_v   ,   U, AcceleratorRead);
   autoView( in_v  ,  in, AcceleratorRead);

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

@@ -92,20 +92,16 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   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();
-
   {
     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 SiteCloverType::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++)
@@ -126,7 +122,7 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
       //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
       //  }
       pokeLocalSite(Qxinv, CTIv, lcoor);
-    }
+    });
   }
 
   // Separate the even and odd parts

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

@@ -51,9 +51,9 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   _FiveDimRedBlackGrid(&FiveDimRedBlackGrid),
   _FourDimGrid        (&FourDimGrid),
   _FourDimRedBlackGrid(&FourDimRedBlackGrid),
-  Stencil    (_FiveDimGrid,npoint,Even,directions,displacements,p),
+  Stencil    (_FiveDimGrid,npoint,Even,directions,displacements,p.locally_periodic,p),
-  StencilEven(_FiveDimRedBlackGrid,npoint,Even,directions,displacements,p), // source is Even
+  StencilEven(_FiveDimRedBlackGrid,npoint,Even,directions,displacements,p.locally_periodic,p), // source is Even
-  StencilOdd (_FiveDimRedBlackGrid,npoint,Odd ,directions,displacements,p), // source is Odd
+  StencilOdd (_FiveDimRedBlackGrid,npoint,Odd ,directions,displacements,p.locally_periodic,p), // source is Odd
   M5(_M5),
   Umu(_FourDimGrid),
   UmuEven(_FourDimRedBlackGrid),
@@ -361,10 +361,21 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
                                          const FermionField &in, FermionField &out,int dag)
 {
   DhopTotalTime-=usecond();
-  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
+
+  assert(  (WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute)
+	 ||(WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute)
+         ||(WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsDirichlet) );
+
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) {
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
-  else 
+  }
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute ) {
     DhopInternalSerialComms(st,lo,U,in,out,dag);
+  }
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsDirichlet ) {
+    DhopInternalDirichletComms(st,lo,U,in,out,dag);
+  }
+  
   DhopTotalTime+=usecond();
 }
 
@@ -431,6 +442,30 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   DhopComputeTime2+=usecond();
 }
 
+template<class Impl>
+void WilsonFermion5D<Impl>::DhopInternalDirichletComms(StencilImpl & st, LebesgueOrder &lo,
+						       DoubledGaugeField & U,
+						       const FermionField &in, FermionField &out,int dag)
+{
+  Compressor compressor(dag);
+
+  int LLs = in.Grid()->_rdimensions[0];
+  int len =  U.Grid()->oSites();
+      
+  /////////////////////////////
+  // do the compute interior
+  /////////////////////////////
+  int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
+  DhopComputeTime-=usecond();
+  if (dag == DaggerYes) {
+    Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
+  } else {
+    Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
+  }
+  accelerator_barrier();
+  DhopComputeTime+=usecond();
+}
+
 
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,

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

@@ -47,9 +47,9 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
     Kernels(p),
     _grid(&Fgrid),
     _cbgrid(&Hgrid),
-    Stencil(&Fgrid, npoint, Even, directions, displacements,p),
+    Stencil(&Fgrid, npoint, Even, directions, displacements,p.locally_periodic,p),
-    StencilEven(&Hgrid, npoint, Even, directions,displacements,p),  // source is Even
+    StencilEven(&Hgrid, npoint, Even, directions,displacements,p.locally_periodic,p),  // source is Even
-    StencilOdd(&Hgrid, npoint, Odd, directions,displacements,p),  // source is Odd
+    StencilOdd(&Hgrid, npoint, Odd, directions,displacements,p.locally_periodic,p),  // source is Odd
     mass(_mass),
     Lebesgue(_grid),
     LebesgueEvenOdd(_cbgrid),
@@ -397,6 +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());
 
@@ -408,6 +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
 
@@ -420,6 +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
 
@@ -485,12 +488,21 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        FermionField &out, int dag)
 {
   DhopTotalTime-=usecond();
-#ifdef GRID_OMP
+
-  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
+  assert(  (WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute)
+	 ||(WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute)
+         ||(WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsDirichlet) );
+
+
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) {
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
-  else
+  }
-#endif
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute ) {
     DhopInternalSerial(st,lo,U,in,out,dag);
+  }
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsDirichlet ) {
+    DhopInternalDirichletComms(st,lo,U,in,out,dag);
+  }
   DhopTotalTime+=usecond();
 }
 
@@ -559,6 +571,29 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
   DhopComputeTime2+=usecond();
 };
 
+template <class Impl>
+void WilsonFermion<Impl>::DhopInternalDirichletComms(StencilImpl &st, LebesgueOrder &lo,
+						     DoubledGaugeField &U,
+						     const FermionField &in,
+						     FermionField &out, int dag)
+{
+  assert((dag == DaggerNo) || (dag == DaggerYes));
+
+  Compressor compressor(dag);
+  int len =  U.Grid()->oSites();
+
+  /////////////////////////////
+  // 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();
+};
 
 template <class Impl>
 void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,

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

@@ -38,9 +38,6 @@ Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 // undefine everything related to kernels
 #include <simd/Fujitsu_A64FX_undef.h>
 
-// enable A64FX body
-#define WILSONKERNELSASMBODYA64FX
-//#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
 
     ///////////////////////////////////////////////////////////
     // If we are A64FX specialise the single precision routine
@@ -63,119 +60,89 @@ Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+//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>
 
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+//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>
 
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+//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>
 
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 
 /////////////////////////////////////////////////////////////////
@@ -185,119 +152,89 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+//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>
 
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+//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>
 
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+//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>
 
-template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 
 // undefine
@@ -330,119 +267,89 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFie
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+// 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>
 
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+// 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>
 
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+// 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>
 
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 /////////////////////////////////////////////////////////////////
 // XYZT vectorised, dag Kernel, double
@@ -451,124 +358,93 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+// 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>
 
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+// 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>
 
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 #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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
+#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)
-#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
-template<> void
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+// template<> void
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+// WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-#if defined (WILSONKERNELSASMBODYA64FX)
+// 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+// #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
+
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+// #pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
-#endif
+// 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>
 
-template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#if defined (WILSONKERNELSASMBODYA64FX)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-#else
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-#endif
 
 
 
 // undefs
-#undef WILSONKERNELSASMBODYA64FX
 #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 
+//template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
+//
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
+//
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
+//
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
+//
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-
+//
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-				    
+//
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-				    
+//
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
-							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							 int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#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 
+//template<> void
-WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
-template<> void 
+//template<> void
-WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+//WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+//							    int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
-#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+//#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
 #endif  // VEC 5D
 

Grid/qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h

@@ -25,6 +25,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     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
@@ -97,7 +102,7 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     PROJ;							                        \
     MAYBEPERM(PERMUTE_DIR,perm);					        \
       } else {								                \
-	LOAD_CHI(base);							                \
+	  LOAD_CHI(base);							                \
       }									                    \
       base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
     MULT_2SPIN_1(Dir);					                    \
@@ -110,6 +115,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     }                                                       \
     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);						            \
@@ -126,73 +136,63 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 
 #define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
       basep = st.GetPFInfo(nent,plocal); nent++;			\
-      if ( local ) {							            \
+      if ( local ) {							\
-    LOAD_CHIMU(base);                                       \
+  LOAD_CHIMU(base);                                       \
-    LOAD_TABLE(PERMUTE_DIR);                                \
+  LOAD_TABLE(PERMUTE_DIR);                                \
-    PROJ;							                        \
+  PROJ;							                        \
-    MAYBEPERM(PERMUTE_DIR,perm);					        \
+  MAYBEPERM(PERMUTE_DIR,perm);					        \
-      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}	    \
+      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}			\
-      base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
+      if ( local || st.same_node[Dir] ) {				\
-      if ( local || st.same_node[Dir] ) {				    \
+  MULT_2SPIN_1(Dir);					                    \
-    MULT_2SPIN_1(Dir);					                    \
+  MULT_2SPIN_2;					                        \
-    PREFETCH_CHIMU(base);                                   \
+  RECON;								\
-    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+      }									\
-    MULT_2SPIN_2;					                        \
+  base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
-    if (s == 0) {                                           \
+  PREFETCH_CHIMU(base);						\
-       if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+  PREFETCH_CHIMU_L2(basep);                               \
-    }                                                       \
-    RECON;								                    \
-    PREFETCH_CHIMU_L2(basep);                               \
-      } else { PREFETCH_CHIMU(base); }								                    \
 
 #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++; \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
-  if((!local)&&(!st.same_node[Dir]) ) {					    \
+  if((!local)&&(!st.same_node[Dir]) ) {					\
-    LOAD_CHI(base);							                \
+    LOAD_CHI(base);							\
     MULT_2SPIN_1(Dir);					                    \
-    PREFETCH_CHIMU(base);                                   \
-    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    if (s == 0) {                                           \
+    RECON;								\
-      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    nmu++;								\
-    }                                                       \
-    RECON;								                    \
-    nmu++;								                    \
   }
 
-#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
-  nmu=0;								                    \
+  nmu=0;								\
-  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;\
+  { ZERO_PSI;}								\
-  if((!local)&&(!st.same_node[Dir]) ) {					    \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
-    LOAD_CHI(base);							                \
+  if((!local)&&(!st.same_node[Dir]) ) {					\
+    LOAD_CHI(base);							\
     MULT_2SPIN_1(Dir);					                    \
-    PREFETCH_CHIMU(base);                                   \
-    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    if (s == 0) {                                           \
+    RECON;								\
-      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    nmu++;								\
-    }                                                       \
-    RECON;								                    \
-    nmu++;								                    \
   }
 
 #define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
 
 #endif
+
+
 {
   int nmu;
   int local,perm, ptype;
@@ -209,7 +209,6 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
     //    int sUn=lo.Reorder(ssn);
     int sUn=ssn;
-    LOCK_GAUGE(0);
 #else
     int sU =ssU;
     int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
@@ -295,6 +294,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       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
@@ -308,6 +312,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       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
@@ -321,6 +330,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       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
@@ -341,6 +355,7 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       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

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);\
@@ -91,55 +108,55 @@ 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);}
-
-// 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;}
-
-
 #define PERMUTE_DIR(dir)			\
-      permute##dir(Chi_00,Chi_00);\
+  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_10,Chi_10);	\
       permute##dir(Chi_11,Chi_11);\
       permute##dir(Chi_12,Chi_12);
 
+#endif
+
+#define MULT_2SPIN(A)\
+  {auto & ref(U[sU](A));						\
+    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;						\
+    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;						\
+    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;						\
+    UChi_11+= U_10*Chi_12;						\
+    UChi_02+= U_20*Chi_02;						\
+    UChi_12+= U_20*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));
 #define XP_PROJ \
@@ -353,13 +370,13 @@ 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)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
   offset = SE->_offset;				\
   local  = SE->_is_local;			\
   perm   = SE->_permute;			\
   if ( local ) {				\
-    LOAD_CHIMU;					\
+    LOAD_CHIMU(PERM);				\
     PROJ;					\
     if ( perm) {				\
       PERMUTE_DIR(PERM);			\
@@ -367,6 +384,37 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   } else {					\
     LOAD_CHI;					\
   }						\
+  acceleratorSynchronise();			\
+  MULT_2SPIN(DIR);				\
+  RECON;					
+
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
+  SE=&st_p[DIR+8*ss];				\
+  ptype=st_perm[DIR];				\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU(PERM);				\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else {					\
+    LOAD_CHI;					\
+  }						\
+  acceleratorSynchronise();			\
+  MULT_2SPIN(DIR);				\
+  RECON;					
+
+#define HAND_STENCIL_LEGA(PROJ,PERM,DIR,RECON)				\
+  SE=&st_p[DIR+8*ss];							\
+  ptype=st_perm[DIR];							\
+ /*SE=st.GetEntry(ptype,DIR,ss);*/					\
+  offset = SE->_offset;				\
+  perm   = SE->_permute;			\
+  LOAD_CHIMU(PERM);				\
+  PROJ;						\
   MULT_2SPIN(DIR);				\
   RECON;					
 
@@ -376,7 +424,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   local  = SE->_is_local;			\
   perm   = SE->_permute;			\
   if ( local ) {				\
-    LOAD_CHIMU;					\
+    LOAD_CHIMU(PERM);				\
     PROJ;					\
     if ( perm) {				\
       PERMUTE_DIR(PERM);			\
@@ -384,10 +432,12 @@ 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)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
@@ -397,44 +447,44 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     MULT_2SPIN(DIR);				\
     RECON;					\
     nmu++;					\
-  }
+  }						\
+  acceleratorSynchronise();			
 
 #define HAND_RESULT(ss)				\
   {						\
-    SiteSpinor & ref (out[ss]);		\
+    SiteSpinor & ref (out[ss]);			\
-    vstream(ref()(0)(0),result_00);		\
+    coalescedWrite(ref()(0)(0),result_00,lane);		\
-    vstream(ref()(0)(1),result_01);		\
+    coalescedWrite(ref()(0)(1),result_01,lane);		\
-    vstream(ref()(0)(2),result_02);		\
+    coalescedWrite(ref()(0)(2),result_02,lane);		\
-    vstream(ref()(1)(0),result_10);		\
+    coalescedWrite(ref()(1)(0),result_10,lane);		\
-    vstream(ref()(1)(1),result_11);		\
+    coalescedWrite(ref()(1)(1),result_11,lane);		\
-    vstream(ref()(1)(2),result_12);		\
+    coalescedWrite(ref()(1)(2),result_12,lane);		\
-    vstream(ref()(2)(0),result_20);		\
+    coalescedWrite(ref()(2)(0),result_20,lane);		\
-    vstream(ref()(2)(1),result_21);		\
+    coalescedWrite(ref()(2)(1),result_21,lane);		\
-    vstream(ref()(2)(2),result_22);		\
+    coalescedWrite(ref()(2)(2),result_22,lane);		\
-    vstream(ref()(3)(0),result_30);		\
+    coalescedWrite(ref()(3)(0),result_30,lane);		\
-    vstream(ref()(3)(1),result_31);		\
+    coalescedWrite(ref()(3)(1),result_31,lane);		\
-    vstream(ref()(3)(2),result_32);		\
+    coalescedWrite(ref()(3)(2),result_32,lane);		\
   }
 
-#define HAND_RESULT_EXT(ss)			\
+#define HAND_RESULT_EXT(ss)				\
-  if (nmu){					\
+  {							\
-    SiteSpinor & ref (out[ss]);		\
+    SiteSpinor & ref (out[ss]);				\
-    ref()(0)(0)+=result_00;		\
+    coalescedWrite(ref()(0)(0),coalescedRead(ref()(0)(0))+result_00,lane);	\
-    ref()(0)(1)+=result_01;		\
+    coalescedWrite(ref()(0)(1),coalescedRead(ref()(0)(1))+result_01,lane);	\
-    ref()(0)(2)+=result_02;		\
+    coalescedWrite(ref()(0)(2),coalescedRead(ref()(0)(2))+result_02,lane);	\
-    ref()(1)(0)+=result_10;		\
+    coalescedWrite(ref()(1)(0),coalescedRead(ref()(1)(0))+result_10,lane);	\
-    ref()(1)(1)+=result_11;		\
+    coalescedWrite(ref()(1)(1),coalescedRead(ref()(1)(1))+result_11,lane);	\
-    ref()(1)(2)+=result_12;		\
+    coalescedWrite(ref()(1)(2),coalescedRead(ref()(1)(2))+result_12,lane);	\
-    ref()(2)(0)+=result_20;		\
+    coalescedWrite(ref()(2)(0),coalescedRead(ref()(2)(0))+result_20,lane);	\
-    ref()(2)(1)+=result_21;		\
+    coalescedWrite(ref()(2)(1),coalescedRead(ref()(2)(1))+result_21,lane);	\
-    ref()(2)(2)+=result_22;		\
+    coalescedWrite(ref()(2)(2),coalescedRead(ref()(2)(2))+result_22,lane);	\
-    ref()(3)(0)+=result_30;		\
+    coalescedWrite(ref()(3)(0),coalescedRead(ref()(3)(0))+result_30,lane);	\
-    ref()(3)(1)+=result_31;		\
+    coalescedWrite(ref()(3)(1),coalescedRead(ref()(3)(1))+result_31,lane);	\
-    ref()(3)(2)+=result_32;		\
+    coalescedWrite(ref()(3)(2),coalescedRead(ref()(3)(2))+result_32,lane);	\
   }
 
-
+#define HAND_DECLARATIONS(Simd)			\
-#define HAND_DECLARATIONS(a)			\
   Simd result_00;				\
   Simd result_01;				\
   Simd result_02;				\
@@ -466,19 +516,19 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   Simd U_11;					\
   Simd U_21;
 
-#define ZERO_RESULT				\
+#define ZERO_RESULT							\
-  result_00=Zero();				\
+  zeroit(result_00);							\
-  result_01=Zero();				\
+  zeroit(result_01);							\
-  result_02=Zero();				\
+  zeroit(result_02);							\
-  result_10=Zero();				\
+  zeroit(result_10);							\
-  result_11=Zero();				\
+  zeroit(result_11);							\
-  result_12=Zero();				\
+  zeroit(result_12);							\
-  result_20=Zero();				\
+  zeroit(result_20);							\
-  result_21=Zero();				\
+  zeroit(result_21);							\
-  result_22=Zero();				\
+  zeroit(result_22);							\
-  result_30=Zero();				\
+  zeroit(result_30);							\
-  result_31=Zero();				\
+  zeroit(result_31);							\
-  result_32=Zero();			
+  zeroit(result_32);			
 
 #define Chimu_00 Chi_00
 #define Chimu_01 Chi_01
@@ -495,15 +545,53 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+
+#ifdef SYCL_HACK
 template<class Impl> accelerator_inline void 
-WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
-				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+				      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);
 
   int offset,local,perm, ptype;
   StencilEntry *SE;
@@ -523,10 +611,16 @@ 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;
@@ -546,11 +640,17 @@ 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;
@@ -570,10 +670,16 @@ 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;
@@ -593,11 +699,17 @@ 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);
+
+  HAND_DECLARATIONS(Simt);
 
   int offset, ptype;
   StencilEntry *SE;
@@ -618,10 +730,16 @@ 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;
@@ -682,3 +800,4 @@ NAMESPACE_END(Grid);
 #undef HAND_RESULT
 #undef HAND_RESULT_INT
 #undef HAND_RESULT_EXT
+#undef HAND_DECLARATIONS

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

@@ -416,7 +416,21 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
-#define KERNEL_CALLNB(A) \
+#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(), {				\
       int sF = ss;							\
@@ -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
    }

Grid/qcd/action/fermion/instantiation/GparityWilsonImplDF/CayleyFermion5DInstantiationGparityWilsonImplDF.cc

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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