Improve the matching of stencil coarsening

Conflicts:
	Grid/allocator/MemoryManager.h

.travis.yml

@@ -9,6 +9,11 @@ matrix:
     - os:        osx
       osx_image: xcode8.3
       compiler: clang
+      env: PREC=single
+    - os:        osx
+      osx_image: xcode8.3
+      compiler: clang
+      env: PREC=double
       
 before_install:
     - export GRIDDIR=`pwd`
@@ -50,7 +55,7 @@ script:
     - make -j4
     - make install
     - cd $CWD/build
-    - ../configure --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
+    - ../configure --enable-precision=$PREC --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
     - make -j4 
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - make check

Grid/DisableWarnings.h

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

Grid/GridStd.h

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

Grid/Grid_Eigen_Dense.h

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

Grid/Makefile.am

@@ -21,7 +21,6 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
-
 all: version-cache Version.h
 
 version-cache:
@@ -54,17 +53,6 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
-#extra_sources+=$(ZWILS_FERMION_FILES)
-extra_sources+=$(WILS_FERMION_FILES)
-extra_sources+=$(STAG_FERMION_FILES)
-if BUILD_GPARITY
-  extra_sources+=$(GP_FERMION_FILES)
-endif
-if BUILD_FERMION_REPS
-  extra_sources+=$(ADJ_FERMION_FILES)
-  extra_sources+=$(TWOIND_FERMION_FILES)
-endif
-
 lib_LIBRARIES = libGrid.a
 
 CCFILES += $(extra_sources)

Grid/algorithms/CoarsenedMatrix.h

@@ -31,7 +31,6 @@ 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);
 
@@ -50,24 +49,20 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
   Lattice<dotp> fine_inner_msk(fine);
 
   // Multiply could be fused with innerProduct
-  // Single block sum kernel could do both masks.
   fine_inner = localInnerProduct(fineX,fineY);
   mult(fine_inner_msk, fine_inner,FineMask);
   blockSum(CoarseInner,fine_inner_msk);
 }
 
-
 class Geometry {
 public:
   int npoint;
-  int base;
   std::vector<int> directions   ;
   std::vector<int> displacements;
-  std::vector<int> points_dagger;
 
   Geometry(int _d)  {
     
-    base = (_d==5) ? 1:0;
+    int base = (_d==5) ? 1:0;
 
     // make coarse grid stencil for 4d , not 5d
     if ( _d==5 ) _d=4;
@@ -75,50 +70,19 @@ 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;
+    std::cout <<GridLogMessage << "Geometry "<<std::endl;
+    for(int p=0;p<npoint;p++){
+      std::cout <<GridLogMessage << "point " <<p<<" dir "<<directions[p]<<" delta " <<displacements[p]<<std::endl;
+    }
   }
 };
   
@@ -140,8 +104,8 @@ public:
   Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
     CoarseGrid(_CoarseGrid),
     FineGrid(_FineGrid),
-    subspace(nbasis,_FineGrid),
-    checkerboard(_checkerboard)
+    checkerboard(_checkerboard),
+    subspace(nbasis,_FineGrid)
   {
   };
   
@@ -296,7 +260,7 @@ public:
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>
-class CoarsenedMatrix : public CheckerBoardedSparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
 public:
     
   typedef iVector<CComplex,nbasis >           siteVector;
@@ -306,59 +270,35 @@ 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; }
 
+  virtual std::vector<int> Directions(void)   { return geom.directions; };
+  virtual std::vector<int> Displacements(void){ return geom.displacements; };
   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;
@@ -372,6 +312,9 @@ public:
 
     int osites=Grid()->oSites();
 
+    autoView(st,Stencil,AcceleratorRead);
+    siteVector *CBp=Stencil.CommBuf();
+
     accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
       int ss = sss/nbasis;
       int b  = sss%nbasis;
@@ -380,14 +323,14 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom_v.npoint;point++){
+      for(int point=0;point<geom.npoint;point++){
 
-	SE=Stencil_v.GetEntry(ptype,point,ss);
+	SE=st.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]);
+	  nbr = coalescedRead(CBp[SE->_offset]);
 	}
 	acceleratorSynchronise();
 
@@ -396,7 +339,7 @@ public:
 	}
       }
       coalescedWrite(out_v[ss](b),res);
-      });
+    });
 
     for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   };
@@ -408,72 +351,12 @@ public:
       return M(in,out);
     } else {
       // corresponds to Galerkin coarsening
-      return MdagNonHermitian(in, out);
+      CoarseVector tmp(Grid());
+      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];
-
-    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];
-
-	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;
@@ -483,7 +366,6 @@ public:
   {
     conformable(_grid,in.Grid());
     conformable(_grid,out.Grid());
-    out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
     Vector<Aview> AcceleratorViewContainer;
@@ -492,7 +374,6 @@ 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;
@@ -506,12 +387,12 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      SE=Stencil_v.GetEntry(ptype,point,ss);
+      SE=Stencil.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]);
+	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
       }
       acceleratorSynchronise();
 
@@ -535,11 +416,23 @@ public:
       MdirCalc(in,out[p],p);
     }
   };
-  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
-
+  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp)
+  {
     this->MdirComms(in);
 
-    MdirCalc(in,out,geom.point(dir,disp));
+    int ndim = in.Grid()->Nd();
+
+    int point=-1;
+    for(int p=0;p<geom.npoint;p++){
+      if( (dir==geom.directions[p])&&(disp==geom.displacements[p])) point=p;
+    }
+    assert(point!=-1);// Must find
+
+    std::cout <<GridLogMessage << "Mdir point "<<point<<" dir "<<dir<<" disp "<<disp  <<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      std::cout <<GridLogMessage << "point " <<p<<" dir "<<geom.directions[p]<<" delta " <<geom.displacements[p]<<std::endl;
+    }
+    MdirCalc(in,out,point);
   };
 
   void Mdiag(const CoarseVector &in, CoarseVector &out)
@@ -548,286 +441,63 @@ public:
     MdirCalc(in, out, point); // No comms
   };
 
-  void Mooee(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerNo, InverseNo);
-  }
-
-  void MooeeInv(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerNo, InverseYes);
-  }
-
-  void MooeeDag(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerYes, InverseNo);
-  }
-
-  void MooeeInvDag(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerYes, InverseYes);
-  }
-
-  void Meooe(const CoarseVector &in, CoarseVector &out) {
-    if(in.Checkerboard() == Odd) {
-      DhopEO(in, out, DaggerNo);
-    } else {
-      DhopOE(in, out, DaggerNo);
-    }
-  }
-
-  void MeooeDag(const CoarseVector &in, CoarseVector &out) {
-    if(in.Checkerboard() == Odd) {
-      DhopEO(in, out, DaggerYes);
-    } else {
-      DhopOE(in, out, DaggerYes);
-    }
-  }
-
-  void Dhop(const CoarseVector &in, CoarseVector &out, int dag) {
-    conformable(in.Grid(), _grid); // verifies full grid
-    conformable(in.Grid(), out.Grid());
-
-    out.Checkerboard() = in.Checkerboard();
-
-    DhopInternal(Stencil, A, in, out, dag);
-  }
-
-  void DhopOE(const CoarseVector &in, CoarseVector &out, int dag) {
-    conformable(in.Grid(), _cbgrid);    // verifies half grid
-    conformable(in.Grid(), out.Grid()); // drops the cb check
-
-    assert(in.Checkerboard() == Even);
-    out.Checkerboard() = Odd;
-
-    DhopInternal(StencilEven, Aodd, in, out, dag);
-  }
-
-  void DhopEO(const CoarseVector &in, CoarseVector &out, int dag) {
-    conformable(in.Grid(), _cbgrid);    // verifies half grid
-    conformable(in.Grid(), out.Grid()); // drops the cb check
-
-    assert(in.Checkerboard() == Odd);
-    out.Checkerboard() = Even;
-
-    DhopInternal(StencilOdd, Aeven, in, out, dag);
-  }
-
-  void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) {
-    out.Checkerboard() = in.Checkerboard();
-    assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
-
-    CoarseMatrix *Aself = nullptr;
-    if(in.Grid()->_isCheckerBoarded) {
-      if(in.Checkerboard() == Odd) {
-        Aself = (inv) ? &AselfInvOdd : &Aodd[geom.npoint-1];
-        DselfInternal(StencilOdd, *Aself, in, out, dag);
-      } else {
-        Aself = (inv) ? &AselfInvEven : &Aeven[geom.npoint-1];
-        DselfInternal(StencilEven, *Aself, in, out, dag);
-      }
-    } else {
-      Aself = (inv) ? &AselfInv : &A[geom.npoint-1];
-      DselfInternal(Stencil, *Aself, in, out, dag);
-    }
-    assert(Aself != nullptr);
-  }
-
-  void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
-                       const CoarseVector &in, CoarseVector &out, int dag) {
-    int point = geom.npoint-1;
-    autoView( out_v, out, AcceleratorWrite);
-    autoView( in_v,  in,  AcceleratorRead);
-    autoView( st_v,  st,  AcceleratorRead);
-    autoView( a_v,   a,   AcceleratorRead);
-
-    const int Nsimd = CComplex::Nsimd();
-    typedef decltype(coalescedRead(in_v[0])) calcVector;
-    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
-
-    RealD* dag_factor_p = &dag_factor[0];
-
-    if(dag) {
-      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
-        int ss = sss/nbasis;
-        int b  = sss%nbasis;
-        calcComplex res = Zero();
-        calcVector nbr;
-        int ptype;
-        StencilEntry *SE;
-
-        SE=st_v.GetEntry(ptype,point,ss);
-
-        if(SE->_is_local) {
-          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-        } else {
-          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
-        }
-        acceleratorSynchronise();
-
-        for(int bb=0;bb<nbasis;bb++) {
-          res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(a_v[ss](b,bb))*nbr(bb);
-        }
-        coalescedWrite(out_v[ss](b),res);
-      });
-    } else {
-      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
-        int ss = sss/nbasis;
-        int b  = sss%nbasis;
-        calcComplex res = Zero();
-        calcVector nbr;
-        int ptype;
-        StencilEntry *SE;
-
-        SE=st_v.GetEntry(ptype,point,ss);
-
-        if(SE->_is_local) {
-          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-        } else {
-          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
-        }
-        acceleratorSynchronise();
-
-        for(int bb=0;bb<nbasis;bb++) {
-          res = res + coalescedRead(a_v[ss](b,bb))*nbr(bb);
-        }
-        coalescedWrite(out_v[ss](b),res);
-      });
-    }
-  }
-
-  void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
-                    const CoarseVector &in, CoarseVector &out, int dag) {
-    SimpleCompressor<siteVector> compressor;
-
-    st.HaloExchange(in,compressor);
-    autoView( in_v,  in,  AcceleratorRead);
-    autoView( out_v, out, AcceleratorWrite);
-    autoView( st_v , st,  AcceleratorRead);
-    typedef LatticeView<Cobj> Aview;
-
-    // determine in what order we need the points
-    int npoint = geom.npoint-1;
-    Vector<int> points(npoint, 0);
-    for(int p=0; p<npoint; p++)
-      points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
-
-    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];
-          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];
-          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) 	:
+ 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),
-    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)
+    A(geom.npoint,&CoarseGrid)
   {
-    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)
+  void Test(Aggregation<Fobj,CComplex,nbasis> &_Aggregates,GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop)
   {
-    fillFactor();
-  };
+    typedef Lattice<Fobj> FineField;
+    CoarseVector Cin(_grid);
+    CoarseVector Cout(_grid);
+    CoarseVector CFout(_grid);
 
-  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;
-    }
+    FineField Fin(FineGrid);
+    FineField Fout(FineGrid);
 
-    // GPU readable prefactor
-    thread_for(i, nbasis*nbasis, {
-      int j = i/nbasis;
-      int k = i%nbasis;
-      dag_factor[i] = dag_factor_eigen(j, k);
-    });
+
+    std::vector<int> seeds({1,2,3,4,5});
+    GridParallelRNG RNG(_grid);  RNG.SeedFixedIntegers(seeds);
+    gaussian(RNG,Cin);
+
+    _Aggregates.PromoteFromSubspace(Cin,Fin);
+    _Aggregates.ProjectToSubspace(Cin,Fin);
+
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    std::cout << GridLogMessage<< " Testing M  "<<std::endl;
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    // Coarse operator
+    this->M(Cin,Cout);
+    // Fine projected operator
+    _Aggregates.PromoteFromSubspace(Cin,Fin);
+    linop.Op(Fin,Fout);
+    _Aggregates.ProjectToSubspace(CFout,Fout);
+
+    CFout = CFout-Cout;
+    RealD diff = norm2(CFout);
+    std::cout << GridLogMessage<< " diff  "<<diff<<std::endl;
+    assert(diff<1.0e-5);
+
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    std::cout << GridLogMessage<< " Testing Mdag  "<<std::endl;
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    // Coarse operator
+    Mdag(Cin,Cout);
+    // Fine operator
+    linop.AdjOp(Fin,Fout);
+    _Aggregates.ProjectToSubspace(CFout,Fout);
+
+    CFout = CFout-Cout;
+    diff = norm2(CFout);
+    std::cout << GridLogMessage<< " diff  "<<diff<<std::endl; 
+    assert(diff<1.0e-5);
   }
 
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
@@ -836,8 +506,6 @@ public:
     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);
 
@@ -868,13 +536,22 @@ public:
 
     CoarseScalar InnerProd(Grid()); 
 
-    std::cout << GridLogMessage<< "CoarsenMatrix Orthog "<< std::endl;
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog " << std::endl;
     // Orthogonalise the subblocks over the basis
     blockOrthogonalise(InnerProd,Subspace.subspace);
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog done " << std::endl;
+
+    auto OpDirections    = linop.Directions();
+    auto OpDisplacements = linop.Displacements();
+
+    std::cout<<" Coarsening an operator with "<< OpDirections.size()<<" terms "<<std::endl;
+    for(int p=0;p<OpDirections.size();p++) {
+      assert(OpDirections[p]==geom.directions[p]);
+      assert(OpDisplacements[p]==geom.displacements[p]);
+    }
 
     // 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++)
     { 
@@ -907,7 +584,21 @@ public:
     evenmask = where(mod(bcb,2)==(Integer)0,one,zero);
     oddmask  = one-evenmask;
 
+    /*
+    {
+      phi=Subspace.subspace[0];
+      linop.OpDirAll(phi,Mphi_p);
+      for(int p=0;p<geom.npoint-1;p++){
+	int dir=geom.directions[p];
+	int disp=geom.displacements[p];
+	linop.OpDir(phi,Mphi,dir,disp);
+	Mphi=Mphi-Mphi_p[p];
+	std::cout << GridLogMessage <<" Direction mapping check " <<norm2(Mphi)<<std::endl;
+      }
+    }
+*/
     assert(self_stencil!=-1);
+    int lhermitian=hermitian;
 
     for(int i=0;i<nbasis;i++){
 
@@ -924,7 +615,7 @@ public:
 	int dir   = geom.directions[p];
 	int disp  = geom.displacements[p];
 
-	if ( (disp==-1) || (!hermitian ) ) {
+	if ( (disp==-1) || (!lhermitian ) ) {
 
 	  ////////////////////////////////////////////////////////////////////////
 	  // Pick out contributions coming from this cell and neighbour cell
@@ -942,7 +633,7 @@ 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) ) {
+	    if ( lhermitian && (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];
@@ -954,11 +645,11 @@ public:
 		}
 	      }
 	    }
-
 	  }
 	}
       }
 
+      std::cout << GridLogMessage<< "CoarsenMatrix Diag "<<std::endl;
       ///////////////////////////////////////////
       // Faster alternate self coupling.. use hermiticity to save 2x
       ///////////////////////////////////////////
@@ -990,57 +681,35 @@ public:
 
       }
     }
-    if(hermitian) {
-      std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
+
+    MemoryManager::PrintBytes();
+
+    // Auto self test
+    Test( Subspace,FineGrid,linop);
+
+#if 0
+    ///////////////////////////
+    // test code worth preserving in if block
+    ///////////////////////////
+    std::cout<<GridLogMessage<< " Computed matrix elements "<< self_stencil <<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      std::cout<<GridLogMessage<< "A["<<p<<"]" << std::endl;
+      std::cout<<GridLogMessage<< "\n"<<A[p] << std::endl;
     }
+    std::cout<<GridLogMessage<< " picking by block0 "<< self_stencil <<std::endl;
+
+    phi=Subspace.subspace[0];
+    std::vector<int> bc(FineGrid->_ndimension,0);
+    blockPick(Grid(),phi,tmp,bc);      // Pick out a block
+    linop.Op(tmp,Mphi);                // Apply big dop
+    blockProject(iProj,Mphi,Subspace.subspace); // project it and print it
+    std::cout<<GridLogMessage<< " Computed matrix elements from block zero only "<<std::endl;
+    std::cout<<GridLogMessage<< iProj <<std::endl;
+    std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
+#endif
 
-    InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl;
-    FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl;
   }
 
-  void InvertSelfStencilLink() {
-    std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl;
-    int localVolume = Grid()->lSites();
-
-    typedef typename Cobj::scalar_object scalar_object;
-
-    autoView(Aself_v,    A[geom.npoint-1], CpuRead);
-    autoView(AselfInv_v, AselfInv,         CpuWrite);
-    thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke
-      Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis);
-      Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis);
-
-      scalar_object selfLink    = Zero();
-      scalar_object selfLinkInv = Zero();
-
-      Coordinate lcoor;
-
-      Grid()->LocalIndexToLocalCoor(site, lcoor);
-      peekLocalSite(selfLink, Aself_v, lcoor);
-
-      for (int i = 0; i < nbasis; ++i)
-        for (int j = 0; j < nbasis; ++j)
-          selfLinkEigen(i, j) = static_cast<ComplexD>(TensorRemove(selfLink(i, j)));
-
-      selfLinkInvEigen = selfLinkEigen.inverse();
-
-      for(int i = 0; i < nbasis; ++i)
-        for(int j = 0; j < nbasis; ++j)
-          selfLinkInv(i, j) = selfLinkInvEigen(i, j);
-
-      pokeLocalSite(selfLinkInv, AselfInv_v, lcoor);
-    });
-  }
-
-  void FillHalfCbs() {
-    std::cout << GridLogDebug << "CoarsenedMatrix::FillHalfCbs" << std::endl;
-    for(int p = 0; p < geom.npoint; ++p) {
-      pickCheckerboard(Even, Aeven[p], A[p]);
-      pickCheckerboard(Odd, Aodd[p], A[p]);
-    }
-    pickCheckerboard(Even, AselfInvEven, AselfInv);
-    pickCheckerboard(Odd, AselfInvOdd, AselfInv);
-  }
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/LinearOperator.h

@@ -52,6 +52,9 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
+
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 };
 
 
@@ -76,6 +79,9 @@ class MdagMLinearOperator : public LinearOperatorBase<Field> {
 public:
   MdagMLinearOperator(Matrix &Mat): _Mat(Mat){};
 
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
+
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
     _Mat.Mdiag(in,out);
@@ -111,6 +117,8 @@ class ShiftedMdagMLinearOperator : public LinearOperatorBase<Field> {
   Matrix &_Mat;
   RealD _shift;
 public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   ShiftedMdagMLinearOperator(Matrix &Mat,RealD shift): _Mat(Mat), _shift(shift){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
@@ -151,6 +159,8 @@ template<class Matrix,class Field>
 class HermitianLinearOperator : public LinearOperatorBase<Field> {
   Matrix &_Mat;
 public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   HermitianLinearOperator(Matrix &Mat): _Mat(Mat){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
@@ -182,6 +192,8 @@ template<class Matrix,class Field>
 class NonHermitianLinearOperator : public LinearOperatorBase<Field> {
   Matrix &_Mat;
 public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   NonHermitianLinearOperator(Matrix &Mat): _Mat(Mat){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
@@ -255,6 +267,8 @@ template<class Matrix,class Field>
   class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
  public:
     Matrix &_Mat;
+    virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+    virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
     SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
     virtual  void Mpc      (const Field &in, Field &out) {
       Field tmp(in.Grid());
@@ -281,6 +295,8 @@ template<class Matrix,class Field>
  protected:
     Matrix &_Mat;
  public:
+    virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+    virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
     SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
     
     virtual void Mpc      (const Field &in, Field &out) {
@@ -307,6 +323,8 @@ template<class Matrix,class Field>
  protected:
     Matrix &_Mat;
  public:
+    virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+    virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
     SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
     
     virtual void Mpc      (const Field &in, Field &out) {
@@ -372,6 +390,8 @@ class NonHermitianSchurDiagMooeeOperator :  public NonHermitianSchurOperatorBase
 {
  public:
   Matrix& _Mat;
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
  NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
   virtual void Mpc(const Field& in, Field& out) {
     Field tmp(in.Grid());
@@ -405,6 +425,8 @@ class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Fi
   Matrix &_Mat;
   
  public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
   virtual void Mpc(const Field& in, Field& out) {
     Field tmp(in.Grid());
@@ -435,6 +457,8 @@ class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Fi
   Matrix& _Mat;
   
  public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
  NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
 
   virtual void Mpc(const Field& in, Field& out) {
@@ -475,6 +499,8 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   Field tmp;
   RealD mass;
  public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
   { 
     assert( _Mat.isTrivialEE() );

Grid/algorithms/SparseMatrix.h

@@ -48,6 +48,8 @@ public:
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -73,6 +75,8 @@ public:
   virtual  void MooeeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeInvDag (const Field &in, Field &out)=0;
 
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/AdefGeneric.h

@@ -1,4 +1,4 @@
-    /*************************************************************************************
+   /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_ITERATIVE_GENERIC_PCG
 #define GRID_ALGORITHMS_ITERATIVE_GENERIC_PCG
 
+NAMESPACE_BEGIN(Grid);
   /*
    * Compared to Tang-2009:  P=Pleft. P^T = PRight Q=MssInv. 
    * Script A = SolverMatrix 
@@ -50,53 +51,54 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    * Vout = x
    */
 
-// abstract base
-template<class Field, class CoarseField>
+
+template<class Field, class CoarseField, class Aggregates>
 class TwoLevelFlexiblePcg : public LinearFunction<Field>
 {
  public:
+
   int verbose;
+
   RealD   Tolerance;
   Integer MaxIterations;
-  const int mmax = 5;
-  GridBase *grid;
-  GridBase *coarsegrid;
+  const int mmax = 4;
+  GridBase *FineGrid;
+  GridBase *CoarseGrid;
 
-  LinearOperatorBase<Field>   *_Linop
-  OperatorFunction<Field>     *_Smoother,
-  LinearFunction<CoarseField> *_CoarseSolver;
-
-  // Need somthing that knows how to get from Coarse to fine and back again
+  LinearOperatorBase<Field>   &_Linop;
+  LinearFunction<Field>     &_Smoother;
+  LinearFunction<CoarseField> &_CoarseSolver;
+  Aggregates                  &_Aggregates;
   
   // more most opertor functions
   TwoLevelFlexiblePcg(RealD tol,
-		     Integer maxit,
-		     LinearOperatorBase<Field> *Linop,
-		     LinearOperatorBase<Field> *SmootherLinop,
-		     OperatorFunction<Field>   *Smoother,
-		     OperatorFunction<CoarseField>  CoarseLinop
-		     ) : 
-      Tolerance(tol), 
-      MaxIterations(maxit),
-      _Linop(Linop),
-      _PreconditionerLinop(PrecLinop),
-      _Preconditioner(Preconditioner)
+		      Integer maxit,
+		      LinearOperatorBase<Field> *Linop,
+		      LinearFunction<Field>   *Smoother,
+		      LinearFunction<CoarseField> *CoarseSolver,
+		      Aggregates *AggP
+		      ) : 
+  Tolerance(tol), 
+    MaxIterations(maxit),
+    _Linop(*Linop),
+    _Smoother(*Smoother),
+    _CoarseSolver(*CoarseSolver),
+    _Aggregates(*AggP)
   { 
+    CoarseGrid=_Aggregates.CoarseGrid;
+    FineGrid=_Aggregates.FineGrid;
     verbose=0;
   };
 
   // The Pcg routine is common to all, but the various matrices differ from derived 
   // implementation to derived implmentation
-  void operator() (const Field &src, Field &psi){
   void operator() (const Field &src, Field &psi){
 
     psi.Checkerboard() = src.Checkerboard();
-    grid             = src.Grid();
 
-    RealD f;
     RealD rtzp,rtz,a,d,b;
-    RealD rptzp;
-    RealD tn;
+    //    RealD rptzp;
+    //    RealD tn;
     RealD guess = norm2(psi);
     RealD ssq   = norm2(src);
     RealD rsq   = ssq*Tolerance*Tolerance;
@@ -104,15 +106,15 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
     /////////////////////////////
     // Set up history vectors
     /////////////////////////////
-    std::vector<Field> p  (mmax,grid);
-    std::vector<Field> mmp(mmax,grid);
+    std::vector<Field> p  (mmax,FineGrid);
+    std::vector<Field> mmp(mmax,FineGrid);
     std::vector<RealD> pAp(mmax);
 
-    Field x  (grid); x = psi;
-    Field z  (grid);
-    Field tmp(grid);
-    Field r  (grid);
-    Field mu (grid);
+    Field x  (FineGrid); x = psi;
+    Field z  (FineGrid);
+    Field tmp(FineGrid);
+    Field r  (FineGrid);
+    Field mu (FineGrid);
   
     //////////////////////////
     // x0 = Vstart -- possibly modify guess
@@ -121,13 +123,13 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
     Vstart(x,src);
 
     // r0 = b -A x0
-    HermOp(x,mmp); // Shouldn't this be something else?
+    _Linop.HermOp(x,mmp[0]); // Shouldn't this be something else?
     axpy (r, -1.0,mmp[0], src);    // Recomputes r=src-Ax0
 
     //////////////////////////////////
     // Compute z = M1 x
     //////////////////////////////////
-    M1(r,z,tmp,mp,SmootherMirs);
+    M1(r,z);
     rtzp =real(innerProduct(r,z));
 
     ///////////////////////////////////////
@@ -143,7 +145,7 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       int peri_kp = (k+1) % mmax;
 
       rtz=rtzp;
-      d= M3(p[peri_k],mp,mmp[peri_k],tmp);
+      d= M3(p[peri_k],mmp[peri_k]);
       a = rtz/d;
     
       // Memorise this
@@ -153,13 +155,13 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       RealD rn = axpy_norm(r,-a,mmp[peri_k],r);
 
       // Compute z = M x
-      M1(r,z,tmp,mp);
+      M1(r,z);
 
       rtzp =real(innerProduct(r,z));
 
       M2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
 
-      p[peri_kp]=p[peri_k];
+      p[peri_kp]=mu;
 
       // Standard search direction  p -> z + b p    ; b = 
       b = (rtzp)/rtz;
@@ -181,7 +183,7 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       // Stopping condition
       if ( rn <= rsq ) { 
 
-	HermOp(x,mmp); // Shouldn't this be something else?
+	_Linop.HermOp(x,mmp[0]); // Shouldn't this be something else?
 	axpy(tmp,-1.0,src,mmp[0]);
 	
 	RealD psinorm = sqrt(norm2(x));
@@ -190,7 +192,8 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
 	RealD true_residual = tmpnorm/srcnorm;
 	std::cout<<GridLogMessage<<"TwoLevelfPcg:   true residual is "<<true_residual<<std::endl;
 	std::cout<<GridLogMessage<<"TwoLevelfPcg: target residual was"<<Tolerance<<std::endl;
-	return k;
+
+	return;
       }
     }
     // Non-convergence
@@ -199,48 +202,40 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
 
  public:
 
-  virtual void M(Field & in,Field & out,Field & tmp) {
-
-  }
-
-  virtual void M1(Field & in, Field & out) {// the smoother
-
+  virtual void M1(Field & in, Field & out) 
+  {// the smoother
     // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
-    Field tmp(grid);
-    Field Min(grid);
+    Field tmp(FineGrid);
+    Field Min(FineGrid);
 
-    PcgM(in,Min); // Smoother call
+    CoarseField PleftProj(CoarseGrid);
+    CoarseField PleftMss_proj(CoarseGrid);
 
-    HermOp(Min,out);
+    _Smoother(in,Min); // Smoother call
+
+    _Linop.HermOp(Min,out);
     axpy(tmp,-1.0,out,in);          // tmp  = in - A Min
 
-    ProjectToSubspace(tmp,PleftProj);     
-    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
-    PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    _Aggregates.ProjectToSubspace(PleftProj,tmp);     
+    _CoarseSolver(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
+    _Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
     axpy(out,1.0,Min,tmp); // Min+tmp
   }
 
-  virtual void M2(const Field & in, Field & out) {
+  virtual void M2(const Field & in, Field & out) 
+  {
     out=in;
-    // Must override for Def2 only
-    //  case PcgDef2:
-    //    Pright(in,out);
-    //    break;
   }
 
-  virtual RealD M3(const Field & p, Field & mmp){
+  virtual RealD M3(const Field & p, Field & mmp)
+  {
     double d,dd;
-    HermOpAndNorm(p,mmp,d,dd);
+    _Linop.HermOpAndNorm(p,mmp,d,dd);
     return dd;
-    // Must override for Def1 only
-    //  case PcgDef1:
-    //    d=linop_d->Mprec(p,mmp,tmp,0,1);// Dag no
-    //      linop_d->Mprec(mmp,mp,tmp,1);// Dag yes
-    //    Pleft(mp,mmp);
-    //    d=real(linop_d->inner(p,mmp));
   }
 
-  virtual void VstartDef2(Field & xconst Field & src){
+  virtual void Vstart(Field & x,const Field & src)
+  {
     //case PcgDef2:
     //case PcgAdef2: 
     //case PcgAdef2f:
@@ -256,142 +251,79 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
     //                   = src_s - (A guess)_s - src_s  + (A guess)_s 
     //                   = 0 
     ///////////////////////////////////
-    Field r(grid);
-    Field mmp(grid);
+    Field r(FineGrid);
+    Field mmp(FineGrid);
+
+    CoarseField PleftProj(CoarseGrid);
+    CoarseField PleftMss_proj(CoarseGrid);
     
-    HermOp(x,mmp);
+    _Linop.HermOp(x,mmp);
     axpy (r, -1.0, mmp, src);        // r_{-1} = src - A x
-    ProjectToSubspace(r,PleftProj);     
-    ApplyInverseCG(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    PromoteFromSubspace(PleftMss_proj,mmp);  
+    _Aggregates.ProjectToSubspace(PleftProj,r);     
+    _CoarseSolver(PleftProj,PleftMss_proj); // Ass^{-1} r_s
+    _Aggregates.PromoteFromSubspace(PleftMss_proj,mmp);  
     x=x+mmp;
 
   }
 
-  virtual void Vstart(Field & x,const Field & src){
-    return;
-  }
-
   /////////////////////////////////////////////////////////////////////
   // Only Def1 has non-trivial Vout. Override in Def1
   /////////////////////////////////////////////////////////////////////
   virtual void   Vout  (Field & in, Field & out,Field & src){
     out = in;
-    //case PcgDef1:
-    //    //Qb + PT x
-    //    ProjectToSubspace(src,PleftProj);     
-    //    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    //    PromoteFromSubspace(PleftMss_proj,tmp);  
-    //    
-    //    Pright(in,out);
-    //    
-    //    linop_d->axpy(out,tmp,out,1.0);
-    //    break;
   }
 
   ////////////////////////////////////////////////////////////////////////////////////////////////
   // Pright and Pleft are common to all implementations
   ////////////////////////////////////////////////////////////////////////////////////////////////
-  virtual void Pright(Field & in,Field & out){
+  virtual void Pright(Field & in,Field & out)
+  {
     // P_R  = [ 1              0 ] 
     //        [ -Mss^-1 Msb    0 ] 
-    Field in_sbar(grid);
+    Field in_sbar(FineGrid);
 
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
+    CoarseField PleftProj(CoarseGrid);
+    CoarseField PleftMss_proj(CoarseGrid);
+
+    _Aggregates.ProjectToSubspace(PleftProj,in);     
+    _Aggregates.PromoteFromSubspace(PleftProj,out);  
     axpy(in_sbar,-1.0,out,in);       // in_sbar = in - in_s 
 
-    HermOp(in_sbar,out);
-    ProjectToSubspace(out,PleftProj);           // Mssbar in_sbar  (project)
+    _Linop.HermOp(in_sbar,out);
+    _Aggregates.ProjectToSubspace(PleftProj,out);           // Mssbar in_sbar  (project)
 
-    ApplyInverse     (PleftProj,PleftMss_proj); // Mss^{-1} Mssbar 
-    PromoteFromSubspace(PleftMss_proj,out);     // 
+    _CoarseSolver(PleftProj,PleftMss_proj); // Mss^{-1} Mssbar 
+    _Aggregates.PromoteFromSubspace(PleftMss_proj,out);     // 
 
     axpy(out,-1.0,out,in_sbar);     // in_sbar - Mss^{-1} Mssbar in_sbar
   }
-  virtual void Pleft (Field & in,Field & out){
+  virtual void Pleft (Field & in,Field & out)
+  {
     // P_L  = [ 1  -Mbs Mss^-1] 
     //        [ 0   0         ] 
-    Field in_sbar(grid);
-    Field    tmp2(grid);
-    Field    Mtmp(grid);
+    Field in_sbar(FineGrid);
+    Field    tmp2(FineGrid);
+    Field    Mtmp(FineGrid);
 
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
+    CoarseField PleftProj(CoarseGrid);
+    CoarseField PleftMss_proj(CoarseGrid);
+
+    _Aggregates.ProjectToSubspace(PleftProj,in);     
+    _Aggregates.PromoteFromSubspace(PleftProj,out);  
     axpy(in_sbar,-1.0,out,in);      // in_sbar = in - in_s
 
-    ApplyInverse(PleftProj,PleftMss_proj); // Mss^{-1} in_s
-    PromoteFromSubspace(PleftMss_proj,out);
+    _CoarseSolver(PleftProj,PleftMss_proj); // Mss^{-1} in_s
+    _Aggregates.PromoteFromSubspace(PleftMss_proj,out);
 
-    HermOp(out,Mtmp);
+    _Linop.HermOp(out,Mtmp);
 
-    ProjectToSubspace(Mtmp,PleftProj);      // Msbar s Mss^{-1}
-    PromoteFromSubspace(PleftProj,tmp2);
+    _Aggregates.ProjectToSubspace(PleftProj,Mtmp);      // Msbar s Mss^{-1}
+    _Aggregates.PromoteFromSubspace(PleftProj,tmp2);
 
     axpy(out,-1.0,tmp2,Mtmp);
     axpy(out,-1.0,out,in_sbar);     // in_sbar - Msbars Mss^{-1} in_s
   }
-}
+};
+NAMESPACE_END(Grid);
 
-template<class Field>
-class TwoLevelFlexiblePcgADef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp){
-
-  } 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp){
-
-  }
-  virtual void M2(Field & in, Field & out){
-
-  }
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp){
-
-  }
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp){
-
-  }
-}
-/*
-template<class Field>
-class TwoLevelFlexiblePcgAD : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef1 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-  virtual void   Vout  (Field & in, Field & out,Field & src,Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgV11: public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-*/
 #endif

Grid/algorithms/iterative/Deflation.h

@@ -60,6 +60,8 @@ public:
   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
 
   virtual void operator()(const Field &src,Field &guess) {
+    RealD t=-usecond();
+    
     guess = Zero();
     assert(evec.size()==eval.size());
     auto N = evec.size();
@@ -68,6 +70,8 @@ public:
       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
     }
     guess.Checkerboard() = src.Checkerboard();
+    t+=usecond();
+    std::cout<<GridLogMessage<<"\t\t\t" << "Deflated guess took "<< t/1000.0<< "ms" <<std::endl;
   }
 };
 

Grid/algorithms/iterative/PrecConjugateResidual.h

@@ -59,7 +59,7 @@ public:
     GridBase *grid = src.Grid();
     Field r(grid),  p(grid), Ap(grid), Ar(grid), z(grid);
       
-    psi=zero;
+    psi=Zero();
     r  = src;
     Preconditioner(r,p);
 

Grid/allocator/AlignedAllocator.cc

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

Grid/allocator/AlignedAllocator.h

@@ -53,7 +53,11 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
+#ifdef GRID_UVM
+    _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
+#else 
     _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
+#endif
     assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
@@ -62,7 +66,11 @@ public:
   { 
     size_type bytes = __n * sizeof(_Tp);
     profilerFree(bytes);
+#ifdef GRID_UVM
+    MemoryManager::SharedFree((void *)__p,bytes);
+#else
     MemoryManager::CpuFree((void *)__p,bytes);
+#endif
   }
 
   // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
@@ -165,18 +173,9 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-#ifdef ACCELERATOR_CSHIFT
-// Cshift on device
-template<class T> using cshiftAllocator = devAllocator<T>;
-#else
-// Cshift on host
-template<class T> using cshiftAllocator = std::allocator<T>;
-#endif
-
-template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
-template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
+//template<class T> using commAllocator = devAllocator<T>;
+template<class T> using Vector     = std::vector<T,uvmAllocator<T> >;           
 template<class T> using commVector = std::vector<T,devAllocator<T> >;
-template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryManager.cc

@@ -9,11 +9,13 @@ NAMESPACE_BEGIN(Grid);
 #define AccSmall (3)
 #define Shared   (4)
 #define SharedSmall (5)
+uint64_t total_cache;
 uint64_t total_shared;
 uint64_t total_device;
 uint64_t total_host;;
 void MemoryManager::PrintBytes(void)
 {
+  std::cout << " MemoryManager : "<<total_cache <<" cache       bytes "<<std::endl;
   std::cout << " MemoryManager : "<<total_shared<<" shared      bytes "<<std::endl;
   std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
   std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<std::endl;
@@ -35,6 +37,8 @@ void *MemoryManager::AcceleratorAllocate(size_t bytes)
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocDevice(bytes);
     total_device+=bytes;
+  } else {
+    //    std::cout <<"AcceleratorAllocate: cache hit Device pointer "<<std::hex<<ptr<<std::dec<<" "<<bytes<<std::endl;
   }
   return ptr;
 }
@@ -53,8 +57,10 @@ void *MemoryManager::SharedAllocate(size_t bytes)
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocShared(bytes);
     total_shared+=bytes;
-    //    std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
+    //    std::cout <<"SharedAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
     //    PrintBytes();
+  } else {
+    //    std::cout <<"SharedAllocate: cache hit Shared pointer "<<std::hex<<ptr<<std::dec<<" "<<bytes<<std::endl;
   }
   return ptr;
 }
@@ -74,6 +80,9 @@ void *MemoryManager::CpuAllocate(size_t bytes)
   if ( ptr == (void *) NULL ) {
     ptr = (void *) acceleratorAllocShared(bytes);
     total_host+=bytes;
+    //    std::cout <<"CpuAllocate: allocated Cpu pointer "<<std::hex<<ptr<<std::dec<<std::endl;
+  } else {
+    //    std::cout <<"CpufAllocate: cache hit Cpu pointer "<<std::hex<<ptr<<std::dec<<" "<<bytes<<std::endl;
   }
   return ptr;
 }
@@ -120,7 +129,7 @@ void MemoryManager::Init(void)
   str= getenv("GRID_ALLOC_NCACHE_LARGE");
   if ( str ) {
     Nc = atoi(str);
-    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+    if ( (Nc>=0) && (Nc <= NallocCacheMax)) {
       Ncache[Cpu]=Nc;
       Ncache[Acc]=Nc;
       Ncache[Shared]=Nc;
@@ -130,7 +139,7 @@ void MemoryManager::Init(void)
   str= getenv("GRID_ALLOC_NCACHE_SMALL");
   if ( str ) {
     Nc = atoi(str);
-    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+    if ( (Nc>=0) && (Nc <= NallocCacheMax)) {
       Ncache[CpuSmall]=Nc;
       Ncache[AccSmall]=Nc;
       Ncache[SharedSmall]=Nc;
@@ -211,6 +220,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 
   if ( entries[v].valid ) {
     ret = entries[v].address;
+    total_cache-=entries[v].bytes;
     entries[v].valid = 0;
     entries[v].address = NULL;
     entries[v].bytes = 0;
@@ -219,6 +229,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
   entries[v].address=ptr;
   entries[v].bytes  =bytes;
   entries[v].valid  =1;
+  total_cache+=entries[v].bytes;
 
   return ret;
 }
@@ -243,6 +254,7 @@ 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;
+      total_cache-=bytes;
       return entries[e].address;
     }
   }

Grid/allocator/MemoryManager.h

@@ -34,6 +34,8 @@ 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*/
@@ -91,8 +93,8 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);

Grid/allocator/MemoryManagerCache.cc

@@ -1,12 +1,11 @@
 #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
 ////////////////////////////////////////////////////////////
@@ -104,7 +103,7 @@ 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);
@@ -112,7 +111,7 @@ 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);
@@ -126,7 +125,7 @@ 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) {
@@ -137,7 +136,7 @@ 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);
@@ -150,7 +149,7 @@ 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;
@@ -165,7 +164,7 @@ 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++;
@@ -228,24 +227,18 @@ 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);
   }
@@ -292,21 +285,21 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    //    printf("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++;
-    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
+    //    printf("Consistent entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+    //    printf("AccDirty entry into device accLock %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
@@ -368,16 +361,13 @@ 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
 

Grid/allocator/MemoryManagerShared.cc

@@ -1,6 +1,7 @@
 #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,const Coordinate &rdim,const Coordinate &chk_dim_msk)
+accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex, Coordinate &rdim, Coordinate &chk_dim_msk)
 {
   int nd=rdim.size();
   Coordinate coor(nd);

Grid/communicator/Communicator_base.h

@@ -1,3 +1,4 @@
+
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -107,8 +108,6 @@ public:
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
-  void GlobalMax(RealD &);
-  void GlobalMax(RealF &);
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
   void GlobalSum(RealD &);
@@ -139,6 +138,21 @@ public:
 		      int recv_from_rank,
 		      int bytes);
   
+  void SendRecvPacket(void *xmit,
+		      void *recv,
+		      int xmit_to_rank,
+		      int recv_from_rank,
+		      int bytes);
+  
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+			   void *xmit,
+			   int xmit_to_rank,
+			   void *recv,
+			   int recv_from_rank,
+			   int bytes);
+  
+  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,
 			       void *recv,

Grid/communicator/Communicator_mpi3.cc

@@ -44,7 +44,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
 
-#ifndef GRID_COMMS_THREADS
+#if defined (TOFU) // FUGAKU, credits go to Issaku Kanamori
     nCommThreads=1;
     // wrong results here too
     // For now: comms-overlap leads to wrong results in Benchmark_wilson even on single node MPI runs
@@ -275,16 +275,6 @@ 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);
@@ -368,19 +358,16 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
-  int tag;
 
   if ( gfrom ==MPI_UNDEFINED) {
-    tag= dir+from*32;
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
   if ( gdest == MPI_UNDEFINED ) {
-    tag= dir+_processor*32;
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;

Grid/communicator/Communicator_none.cc

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

Grid/communicator/SharedMemory.h

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

Grid/communicator/SharedMemoryMPI.cc

@@ -32,9 +32,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
 #endif
-#ifdef GRID_HIP
-#include <hip/hip_runtime_api.h>
-#endif
 
 NAMESPACE_BEGIN(Grid); 
 #define header "SharedMemoryMpi: "
@@ -428,7 +425,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#if defined(GRID_CUDA) ||defined(GRID_HIP)
+#ifdef GRID_CUDA
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -451,16 +448,21 @@ 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;
+#ifndef GRID_MPI3_SHM_NONE
+  auto err =  cudaMalloc(&ShmCommBuf, bytes);
+#else
+  auto err =  cudaMallocManaged(&ShmCommBuf, bytes);
+#endif
+  if ( err !=  cudaSuccess) {
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
     exit(EXIT_FAILURE);  
   }
-  //  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;
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+  if ( WorldRank == 0 ){
+    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
 
@@ -473,26 +475,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
     //////////////////////////////////////////////////
-#ifdef GRID_CUDA
     cudaIpcMemHandle_t handle;
+    
     if ( r==WorldShmRank ) { 
-      auto err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
-#endif
-#ifdef GRID_HIP
-    hipIpcMemHandle_t handle;    
-    if ( r==WorldShmRank ) { 
-      auto err = hipIpcGetMemHandle(&handle,ShmCommBuf);
-      if ( err !=  hipSuccess) {
-	std::cerr << " SharedMemoryMPI.cc hipIpcGetMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
-	exit(EXIT_FAILURE);
-      }
-    }
-#endif
     //////////////////////////////////////////////////
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
@@ -509,24 +500,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
     void * thisBuf = ShmCommBuf;
-#ifdef GRID_CUDA
     if ( r!=WorldShmRank ) { 
-      auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
-#endif
-#ifdef GRID_HIP
-    if ( r!=WorldShmRank ) { 
-      auto err = hipIpcOpenMemHandle(&thisBuf,handle,hipIpcMemLazyEnablePeerAccess);
-      if ( err !=  hipSuccess) {
-	std::cerr << " SharedMemoryMPI.cc hipIpcOpenMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
-	exit(EXIT_FAILURE);
-      }
-    }
-#endif
     ///////////////////////////////////////////////////////////////
     // Save a copy of the device buffers
     ///////////////////////////////////////////////////////////////
@@ -666,6 +646,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
+      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -715,7 +696,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
 {
 #ifdef GRID_CUDA
   cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
@@ -771,12 +752,19 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
 
-#ifdef GRID_SHM_FORCE_MPI
-  // Hide the shared memory path between ranks
-  {
+#ifdef GRID_IBM_SUMMIT
+  // Hide the shared memory path between sockets 
+  // if even number of nodes
+  if ( (ShmSize & 0x1)==0 ) {
+    int SocketSize = ShmSize/2;
+    int mySocket = ShmRank/SocketSize; 
     for(int r=0;r<size;r++){
-      if ( r!=rank ) {
-	ShmRanks[r] = MPI_UNDEFINED;
+      int hisRank=ShmRanks[r];
+      if ( hisRank!= MPI_UNDEFINED ) {
+	int hisSocket=hisRank/SocketSize;
+	if ( hisSocket != mySocket ) {
+	  ShmRanks[r] = MPI_UNDEFINED;
+	}
       }
     }
   }

Grid/communicator/SharedMemoryNone.cc

@@ -29,7 +29,6 @@ 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)
@@ -56,38 +55,6 @@ 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 ; 
@@ -116,15 +83,7 @@ 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

@@ -35,7 +35,7 @@ extern Vector<std::pair<int,int> > Cshift_table;
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -73,19 +73,12 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
      }
   }
   {
+    autoView(rhs_v , rhs, AcceleratorRead);
     auto buffer_p = & buffer[0];
     auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for(i,ent,{
-      buffer_p[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
   }
 }
 
@@ -110,7 +103,6 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
-#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for2d(n,e1,b,e2,1,{
 	int o      =   n*n1;
@@ -119,22 +111,12 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-	int o      =   n*n1;
-	int offset = b+n*e2;
-	
-	vobj temp =rhs_v[so+o+b];
-	extract<vobj>(temp,pointers,offset);
-      });
-#endif
   } else { 
+    autoView(rhs_v , rhs, AcceleratorRead);
+
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for2d(n,e1,b,e2,1,{
 
 	Coordinate coor;
@@ -152,33 +134,13 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-
-	Coordinate coor;
-
-	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
-	int offset = b+n*e2;
-
-	if ( ocb & cbmask ) {
-	  vobj temp =rhs_v[so+o+b];
-	  extract<vobj>(temp,pointers,offset);
-	}
-      });
-#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -220,19 +182,12 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   }
   
   {
+    autoView( rhs_v, rhs, AcceleratorWrite);
     auto buffer_p = & buffer[0];
     auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
     });
-#else
-    autoView( rhs_v, rhs, CpuWrite);
-    thread_for(i,ent,{
-      rhs_v[table[i].first]=buffer_p[table[i].second];
-    });
-#endif
   }
 }
 
@@ -253,23 +208,14 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
+    autoView( rhs_v , rhs, AcceleratorWrite);
     int _slice_stride = rhs.Grid()->_slice_stride[dimension];
     int _slice_block = rhs.Grid()->_slice_block[dimension];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v , rhs, AcceleratorWrite);
     accelerator_for2d(n,e1,b,e2,1,{
 	int o      = n*_slice_stride;
 	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
-#else
-    autoView( rhs_v , rhs, CpuWrite);
-    thread_for2d(n,e1,b,e2,{
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
-	merge(rhs_v[so+o+b],pointers,offset);
-    });
-#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
@@ -334,20 +280,12 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
+    auto table = &Cshift_table[0];
     accelerator_for(i,ent,vobj::Nsimd(),{
       coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    autoView(lhs_v , lhs, CpuWrite);
-    thread_for(i,ent,{
-      lhs_v[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
   }
 }
 
@@ -386,20 +324,12 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
+    auto table = &Cshift_table[0];
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
-#else
-    autoView( rhs_v, rhs, CpuRead);
-    autoView( lhs_v, lhs, CpuWrite);
-    thread_for(i,ent,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

@@ -101,8 +101,7 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
 }
-#define ACCELERATOR_CSHIFT_NO_COPY
-#ifdef ACCELERATOR_CSHIFT_NO_COPY
+
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
   typedef typename vobj::vector_type vector_type;
@@ -122,9 +121,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  cshiftVector<vobj> send_buf(buffer_size);
-  cshiftVector<vobj> recv_buf(buffer_size);
-    
+  commVector<vobj> send_buf(buffer_size);
+  commVector<vobj> recv_buf(buffer_size);
+
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
 
@@ -139,7 +138,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
     } else {
 
-      int words = buffer_size;
+      int words = send_buf.size();
       if (cbmask != 0x3) words=words>>1;
 
       int bytes = words * sizeof(vobj);
@@ -151,14 +150,12 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
-      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
-
       grid->Barrier();
 
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
@@ -198,15 +195,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd);
-  std::vector<cshiftVector<scalar_object> >  recv_buf_extract(Nsimd);
-  scalar_object *  recv_buf_extract_mpi;
-  scalar_object *  send_buf_extract_mpi;
- 
-  for(int s=0;s<Nsimd;s++){
-    send_buf_extract[s].resize(buffer_size);
-    recv_buf_extract[s].resize(buffer_size);
-  }
+  std::vector<commVector<scalar_object> >   send_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
+  std::vector<commVector<scalar_object> >   recv_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
 
   int bytes = buffer_size*sizeof(scalar_object);
 
@@ -252,204 +242,11 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	grid->Barrier();
-
-	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
-	recv_buf_extract_mpi = &recv_buf_extract[i][0];
-	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
+	grid->SendToRecvFrom((void *)&send_buf_extract[nbr_lane][0],
 			     xmit_to_rank,
-			     (void *)recv_buf_extract_mpi,
+			     (void *)&recv_buf_extract[i][0],
 			     recv_from_rank,
 			     bytes);
-
-	grid->Barrier();
-
-	rpointers[i] = &recv_buf_extract[i][0];
-      } else { 
-	rpointers[i] = &send_buf_extract[nbr_lane][0];
-      }
-
-    }
-    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-  }
-
-}
-#else 
-template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  GridBase *grid=rhs.Grid();
-  Lattice<vobj> temp(rhs.Grid());
-
-  int fd              = rhs.Grid()->_fdimensions[dimension];
-  int rd              = rhs.Grid()->_rdimensions[dimension];
-  int pd              = rhs.Grid()->_processors[dimension];
-  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  assert(simd_layout==1);
-  assert(comm_dim==1);
-  assert(shift>=0);
-  assert(shift<fd);
-  
-  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  cshiftVector<vobj> send_buf_v(buffer_size);
-  cshiftVector<vobj> recv_buf_v(buffer_size);
-  vobj *send_buf;
-  vobj *recv_buf;
-  {
-    grid->ShmBufferFreeAll();
-    size_t bytes = buffer_size*sizeof(vobj);
-    send_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-    recv_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-  }
-    
-  int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  for(int x=0;x<rd;x++){       
-
-    int sx        =  (x+sshift)%rd;
-    int comm_proc = ((x+sshift)/rd)%pd;
-    
-    if (comm_proc==0) {
-
-      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-
-    } else {
-
-      int words = buffer_size;
-      if (cbmask != 0x3) words=words>>1;
-
-      int bytes = words * sizeof(vobj);
-
-      Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
-
-      //      int rank           = grid->_processor;
-      int recv_from_rank;
-      int xmit_to_rank;
-      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-
-
-      grid->Barrier();
-
-      acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
-      grid->SendToRecvFrom((void *)&send_buf[0],
-			   xmit_to_rank,
-			   (void *)&recv_buf[0],
-			   recv_from_rank,
-			   bytes);
-      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
-
-      grid->Barrier();
-
-      Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
-    }
-  }
-}
-
-template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  GridBase *grid=rhs.Grid();
-  const int Nsimd = grid->Nsimd();
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-   
-  int fd = grid->_fdimensions[dimension];
-  int rd = grid->_rdimensions[dimension];
-  int ld = grid->_ldimensions[dimension];
-  int pd = grid->_processors[dimension];
-  int simd_layout     = grid->_simd_layout[dimension];
-  int comm_dim        = grid->_processors[dimension] >1 ;
-
-  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
-  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
-  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
-
-  assert(comm_dim==1);
-  assert(simd_layout==2);
-  assert(shift>=0);
-  assert(shift<fd);
-
-  int permute_type=grid->PermuteType(dimension);
-
-  ///////////////////////////////////////////////
-  // Simd direction uses an extract/merge pair
-  ///////////////////////////////////////////////
-  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
-  //  int words = sizeof(vobj)/sizeof(vector_type);
-
-  std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd);
-  std::vector<cshiftVector<scalar_object> >  recv_buf_extract(Nsimd);
-  scalar_object *  recv_buf_extract_mpi;
-  scalar_object *  send_buf_extract_mpi;
-  {
-    size_t bytes = sizeof(scalar_object)*buffer_size;
-    grid->ShmBufferFreeAll();
-    send_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-    recv_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-  }
-  for(int s=0;s<Nsimd;s++){
-    send_buf_extract[s].resize(buffer_size);
-    recv_buf_extract[s].resize(buffer_size);
-  }
-
-  int bytes = buffer_size*sizeof(scalar_object);
-
-  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
-  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
-
-  ///////////////////////////////////////////
-  // Work out what to send where
-  ///////////////////////////////////////////
-  int cb    = (cbmask==0x2)? Odd : Even;
-  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  // loop over outer coord planes orthog to dim
-  for(int x=0;x<rd;x++){       
-
-    // FIXME call local permute copy if none are offnode.
-    for(int i=0;i<Nsimd;i++){       
-      pointers[i] = &send_buf_extract[i][0];
-    }
-    int sx   = (x+sshift)%rd;
-    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
-
-    for(int i=0;i<Nsimd;i++){
-      
-      int inner_bit = (Nsimd>>(permute_type+1));
-      int ic= (i&inner_bit)? 1:0;
-
-      int my_coor          = rd*ic + x;
-      int nbr_coor         = my_coor+sshift;
-      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-
-      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
-      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
-      int nbr_lane = (i&(~inner_bit));
-
-      int recv_from_rank;
-      int xmit_to_rank;
-
-      if (nbr_ic) nbr_lane|=inner_bit;
-
-      assert (sx == nbr_ox);
-
-      if(nbr_proc){
-	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
-
-	grid->Barrier();
-
-	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
-	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
-			     xmit_to_rank,
-			     (void *)recv_buf_extract_mpi,
-			     recv_from_rank,
-			     bytes);
-	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
-
 	grid->Barrier();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -461,7 +258,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   }
 
 }
-#endif
+
 NAMESPACE_END(Grid); 
 
 #endif

Grid/lattice/Lattice_ET.h

@@ -342,14 +342,19 @@ inline void ExpressionViewClose(LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
 
 GridUnopClass(UnarySub, -a);
 GridUnopClass(UnaryNot, Not(a));
+GridUnopClass(UnaryAdj, adj(a));
+GridUnopClass(UnaryConj, conjugate(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
+GridUnopClass(UnaryToReal, toReal(a));
+GridUnopClass(UnaryToComplex, toComplex(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
 GridUnopClass(UnarySqrt, sqrt(a));
+GridUnopClass(UnaryRsqrt, rsqrt(a));
 GridUnopClass(UnarySin, sin(a));
 GridUnopClass(UnaryCos, cos(a));
 GridUnopClass(UnaryAsin, asin(a));
@@ -451,17 +456,20 @@ GridTrinOpClass(TrinaryWhere,
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
-//GRID_DEF_UNOP(adj, UnaryAdj);
-//GRID_DEF_UNOP(conjugate, UnaryConj);
+GRID_DEF_UNOP(adj, UnaryAdj);
+GRID_DEF_UNOP(conjugate, UnaryConj);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
+GRID_DEF_UNOP(toReal, UnaryToReal);
+GRID_DEF_UNOP(toComplex, UnaryToComplex);
 GRID_DEF_UNOP(timesI, UnaryTimesI);
 GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
 GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the
                                // abs-fabs-dabs-labs thing
 GRID_DEF_UNOP(sqrt, UnarySqrt);
+GRID_DEF_UNOP(rsqrt, UnaryRsqrt);
 GRID_DEF_UNOP(sin, UnarySin);
 GRID_DEF_UNOP(cos, UnaryCos);
 GRID_DEF_UNOP(asin, UnaryAsin);
@@ -486,27 +494,27 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
+  -> Lattice<decltype(expr.op.func(vecEval(0, expr.arg1)))> 
 {
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > ret(expr);
+  Lattice<decltype(expr.op.func(vecEval(0, expr.arg1)))> ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type >
+  -> Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))> 
 {
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
+  Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))> ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+  -> Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),
 				   vecEval(0, expr.arg2),
-				   vecEval(0, expr.arg3)))>::type >
+				   vecEval(0, expr.arg3)))> 
 {
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+  Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),
 				vecEval(0, expr.arg2),
-			        vecEval(0, expr.arg3)))>::type >  ret(expr);
+			        vecEval(0, expr.arg3)))>  ret(expr);
   return ret;
 }
 #define EXPRESSION_CLOSURE(function)					\

Grid/lattice/Lattice_arith.h

@@ -60,9 +60,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   autoView( lhs_v , lhs, AcceleratorRead);
   autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     auto rhs_t=rhs_v(ss);
-    auto tmp  =ret_v(ss);
     mac(&tmp,&lhs_t,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
   });
@@ -124,7 +124,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
+    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     mac(&tmp,&lhs_t,&rhs);
     coalescedWrite(ret_v[ss],tmp);
@@ -182,7 +182,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
+    decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
     mac(&tmp,&lhs,&rhs_t);
     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_CUDA)) )
+#if ( (!defined(GRID_SYCL)) && (!defined(GRID_CUDA)) && (!defined(GRID_HIP)) )
   int max_threads = thread_max();
   Vector < vobj > Bt(Nm * max_threads);
   thread_region
@@ -161,13 +161,11 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
   double * Qt_j = & Qt_jv[0];
   for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
-  auto basis_vp=& basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    vobj zzz=Zero();
-    auto B=coalescedRead(zzz);
+    auto B=coalescedRead(zz);
     for(int k=k0; k<k1; ++k){
-      B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]);
+      B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
     }
     coalescedWrite(result_v[ss], B);
   });

Grid/lattice/Lattice_reality.h

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

Grid/lattice/Lattice_reduction.h

@@ -96,34 +96,8 @@ 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)
 {
@@ -167,32 +141,6 @@ 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)

Grid/lattice/Lattice_reduction_gpu.h

@@ -2,13 +2,12 @@ NAMESPACE_BEGIN(Grid);
 
 #ifdef GRID_HIP
 extern hipDeviceProp_t *gpu_props;
-#define WARP_SIZE 64
 #endif
 #ifdef GRID_CUDA
 extern cudaDeviceProp *gpu_props;
-#define WARP_SIZE 32
 #endif
 
+#define WARP_SIZE 32
 __device__ unsigned int retirementCount = 0;
 
 template <class Iterator>
@@ -65,7 +64,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
   
   // cannot use overloaded operators for sobj as they are not volatile-qualified
   memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
-  acceleratorSynchronise();
+  __syncwarp();
   
   const Iterator VEC = WARP_SIZE;
   const Iterator vid = tid & (VEC-1);
@@ -79,9 +78,9 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
       beta += temp;
       memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
     }
-    acceleratorSynchronise();
+    __syncwarp();
   }
-  acceleratorSynchroniseAll();
+  __syncthreads();
   
   if (threadIdx.x == 0) {
     beta  = Zero();
@@ -91,7 +90,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
     }
     memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
   }
-  acceleratorSynchroniseAll();
+  __syncthreads();
 }
 
 

Grid/lattice/Lattice_transfer.h

@@ -127,11 +127,6 @@ accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
   convertType(out,in._internal);
 }
 
-template<typename T1, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
-accelerator_inline void convertType(T1 & out, const iScalar<T1> & in) {
-  convertType(out,in._internal);
-}
-
 template<typename T1,typename T2>
 accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
   convertType(out._internal,in);

Grid/lattice/Lattice_view.h

@@ -52,6 +52,7 @@ public:
 // This will be safe to call from accelerator_for and is trivially copy constructible
 // The copy constructor for this will need to be used by device lambda functions
 /////////////////////////////////////////////////////////////////////////////////////////
+#undef LATTICE_BOUNDS_CHECK
 template<class vobj> 
 class LatticeView : public LatticeAccelerator<vobj>
 {
@@ -61,19 +62,36 @@ public:
   void * cpu_ptr;
 #ifdef GRID_SIMT
   accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { 
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
     return coalescedRead(this->_odata[i]); 
   }
 #else 
-  accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
+  accelerator_inline const vobj & operator()(size_t i) const {
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
+    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]; };
+  accelerator_inline const vobj & operator[](size_t i) const { 
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
 #endif
+    return this->_odata[i]; 
+  };
+  accelerator_inline vobj       & operator[](size_t i)       { 
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
+    return this->_odata[i]; 
+  };
 
   accelerator_inline uint64_t begin(void) const { return 0;};
   accelerator_inline uint64_t end(void)   const { return this->_odata_size; };

Grid/log/Log.h

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

Grid/parallelIO/BinaryIO.cc

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

Grid/parallelIO/BinaryIO.h

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

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 stats = PeriodicGaugeStatistics>
-  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
+  template <class vsimd>
+  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<vLorentzColourMatrixD> GaugeField;
-    typedef vLorentzColourMatrixD vobj;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -636,9 +636,6 @@ 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;
@@ -708,10 +705,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
+  template <class vsimd>
+  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
 
-    typedef Lattice<vLorentzColourMatrixD > GaugeField;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
     typedef typename GaugeField::vector_object  vobj;
     typedef typename vobj::scalar_object sobj;
 
@@ -924,8 +921,7 @@ class IldgReader : public GridLimeReader {
 
     if ( found_FieldMetaData || found_usqcdInfo ) {
       FieldMetaData checker;
-      stats Stats;
-      Stats(Umu,checker);
+      GaugeStatistics(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

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

Grid/parallelIO/NerscIO.h

@@ -40,8 +40,6 @@ using namespace Grid;
 class NerscIO : public BinaryIO { 
 public:
 
-  typedef Lattice<vLorentzColourMatrixD> GaugeField;
-
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -131,12 +129,12 @@ public:
   // Now the meat: the object readers
   /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void readConfiguration(GaugeField &Umu,
+  template<class vsimd>
+  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
 				       FieldMetaData& header,
-				       std::string file,
-				       GaugeStats GaugeStatisticsCalculator=GaugeStats())
+				       std::string file)
   {
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
     GridBase *grid = Umu.Grid();
     uint64_t offset = readHeader(file,Umu.Grid(),header);
@@ -155,23 +153,23 @@ 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<vLorentzColourMatrixD, LorentzColour2x3F> 
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3D> 
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, 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<vLorentzColourMatrixD,LorentzColourMatrixF>
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixD>
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
@@ -179,7 +177,7 @@ public:
       assert(0);
     }
 
-    GaugeStats Stats; Stats(Umu,clone);
+    GaugeStatistics(Umu,clone);
 
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl;
@@ -205,13 +203,15 @@ public:
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
+  template<class vsimd>
+  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32)
   {
-    typedef vLorentzColourMatrixD vobj;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+
+    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
@@ -229,7 +229,7 @@ public:
 
     GridMetaData(grid,header);
     assert(header.nd==4);
-    GaugeStats Stats; Stats(Umu,header);
+    GaugeStatistics(Umu,header);
     MachineCharacteristics(header);
 
 	uint64_t offset;
@@ -238,19 +238,19 @@ public:
     header.floating_point = std::string("IEEE64BIG");
     header.data_type      = std::string("4D_SU3_GAUGE_3x3");
     GaugeSimpleUnmunger<fobj3D,sobj> munge;
-    if ( grid->IsBoss() ) { 
-      truncate(file);
-      offset = writeHeader(header,file);
-    }
-    grid->Broadcast(0,(void *)&offset,sizeof(offset));
+	if ( grid->IsBoss() ) { 
+	  truncate(file);
+    offset = writeHeader(header,file);
+	}
+	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() ) { 
-      writeHeader(header,file);
-    }
+	if ( grid->IsBoss() ) { 
+    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;
 
-    PeriodicGaugeStatistics Stats; Stats(Umu, clone);
+    GaugeStatistics(Umu, clone);
 
     RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
 

Grid/parallelIO/OpenQcdIOChromaReference.h

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

Grid/qcd/QCD.h

@@ -47,7 +47,7 @@ static constexpr int Ym = 5;
 static constexpr int Zm = 6;
 static constexpr int Tm = 7;
 
-static constexpr int Nc=Config_Nc;
+static constexpr int Nc=3;
 static constexpr int Ns=4;
 static constexpr int Nd=4;
 static constexpr int Nhs=2; // half spinor
@@ -77,13 +77,13 @@ const int SpinorIndex = 2;
 template<typename T> struct isSpinor {
   static constexpr bool value = (SpinorIndex==T::TensorLevel);
 };
+const int CoarseIndex = 4;
+template<typename T> struct isCoarsened {
+  static constexpr bool value = (CoarseIndex<=T::TensorLevel);
+};
 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> > ;
 

Grid/qcd/action/fermion/FermionOperator.h

@@ -89,7 +89,8 @@ public:
   virtual void  Mdiag  (const FermionField &in, FermionField &out) { Mooee(in,out);};   // Same as Mooee applied to both CB's
   virtual void  Mdir   (const FermionField &in, FermionField &out,int dir,int disp)=0;   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
   virtual void  MdirAll(const FermionField &in, std::vector<FermionField> &out)=0;   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
-
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 
   virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -88,7 +88,7 @@ public:
 					  const _Spinor &chi, 
 					  int mu, 
 					  StencilEntry *SE,
-					  const StencilView &St) 
+					  StencilView &St) 
   {
     int direction = St._directions[mu];
     int distance  = St._distances[mu];
@@ -97,30 +97,42 @@ 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));
 
-    auto UU0=coalescedRead(U(0)(mu));
-    auto UU1=coalescedRead(U(1)(mu));
-    
-    //Decide whether we do a G-parity flavor twist
-    //Note: this assumes (but does not check) that sl==1 || sl==2 i.e. max 2 SIMD lanes in G-parity dir
-    //It also assumes (but does not check) that abs(distance) == 1
-    int permute_lane = (sl==1) 
-    || ((distance== 1)&&(icoor[direction]==1))
-    || ((distance==-1)&&(icoor[direction]==0));
+      if ( permute_lane ) { 
+	tmp(0) = chi(1);
+	tmp(1) = chi(0);
+      } else {
+	tmp(0) = chi(0);
+	tmp(1) = chi(1);
+      }
 
-    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
+      auto UU0=coalescedRead(U(0)(mu));
+      auto UU1=coalescedRead(U(1)(mu));
 
-    //Apply the links
-    int f_upper = permute_lane ? 1 : 0;
-    int f_lower = !f_upper;
+      mult(&phi(0),&UU0,&tmp(0));
+      mult(&phi(1),&UU1,&tmp(1));
 
-    mult(&phi(0),&UU0,&chi(f_upper));
-    mult(&phi(1),&UU1,&chi(f_lower));
+    } else {
+
+      auto UU0=coalescedRead(U(0)(mu));
+      auto UU1=coalescedRead(U(1)(mu));
+
+      mult(&phi(0),&UU0,&chi(0));
+      mult(&phi(1),&UU1,&chi(1));
+
+    }
 
 #else
     typedef _Spinor vobj;

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -44,6 +44,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
   typedef StaggeredKernels<Impl> Kernels;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -49,6 +49,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
   typedef StaggeredKernels<Impl> Kernels;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 

Grid/qcd/action/fermion/MADWF.h

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

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -47,6 +47,9 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   ////////////////////////////////////////
   // Performance monitoring
   ////////////////////////////////////////

Grid/qcd/action/fermion/StaggeredKernels.h

@@ -56,80 +56,55 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
 		 DoubledGaugeField &U,
 		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
   
-  void DhopDirKernel(StencilImpl &st,
-		     const DoubledGaugeFieldView &U,
-		     const DoubledGaugeFieldView &UUU, SiteSpinor * buf,
-		     int sF, int sU,
-		     const FermionFieldView &in,
-		     const FermionFieldView &out, int dir,int disp);
+  void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+		     int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
  protected:    
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Generic Nc kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   template<int Naik> 
-   static accelerator_inline
-   void DhopSiteGeneric(const StencilView &st, 
-			const DoubledGaugeFieldView &U,
-			const DoubledGaugeFieldView &UUU, 
+   template<int Naik> accelerator_inline
+   void DhopSiteGeneric(StencilView &st, 
+			DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
-			const FermionFieldView &in,
-			const FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteGenericInt(const StencilView &st, 
-			   const DoubledGaugeFieldView &U,
-			   const DoubledGaugeFieldView &UUU, 
+			const FermionFieldView &in, FermionFieldView &out,int dag);
+   template<int Naik> accelerator_inline
+   void DhopSiteGenericInt(StencilView &st, 
+			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			   SiteSpinor * buf, int LLs, int sU, 
-			   const FermionFieldView &in,
-			   const FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteGenericExt(const StencilView &st, 
-			   const DoubledGaugeFieldView &U,
-			   const DoubledGaugeFieldView &UUU,
- 			   SiteSpinor * buf, int LLs, int sU, 
-			   const FermionFieldView &in,
-			   const FermionFieldView &out,int dag);
+			   const FermionFieldView &in, FermionFieldView &out,int dag);
+   template<int Naik> accelerator_inline
+   void DhopSiteGenericExt(StencilView &st, 
+			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
+			   SiteSpinor * buf, int LLs, int sU, 
+			   const FermionFieldView &in, FermionFieldView &out,int dag);
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteHand(const StencilView &st, 
-		     const DoubledGaugeFieldView &U,
-		     const DoubledGaugeFieldView &UUU, 
- 		     SiteSpinor * buf, int LLs, int sU, 
-		     const FermionFieldView &in,
-		     const FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteHandInt(const StencilView &st, 
-			const DoubledGaugeFieldView &U,
-			const DoubledGaugeFieldView &UUU, 
- 			SiteSpinor * buf, int LLs, int sU, 
-			const FermionFieldView &in,
-			const FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteHandExt(const StencilView &st, 
-			const DoubledGaugeFieldView &U,
-			const DoubledGaugeFieldView &UUU, 
- 			SiteSpinor * buf, int LLs, int sU, 
-			const FermionFieldView &in,
-			const FermionFieldView &out,int dag);
+   template<int Naik> accelerator_inline
+   void DhopSiteHand(StencilView &st, 
+		     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
+		     SiteSpinor * buf, int LLs, int sU, 
+		     const FermionFieldView &in, FermionFieldView &out,int dag);
+   template<int Naik> accelerator_inline
+   void DhopSiteHandInt(StencilView &st, 
+			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
+			SiteSpinor * buf, int LLs, int sU, 
+			const FermionFieldView &in, FermionFieldView &out,int dag);
+   template<int Naik> accelerator_inline
+   void DhopSiteHandExt(StencilView &st, 
+			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
+			SiteSpinor * buf, int LLs, int sU, 
+			const FermionFieldView &in, FermionFieldView &out,int dag);
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Asm Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   
-   void DhopSiteAsm(const StencilView &st, 
-		    const DoubledGaugeFieldView &U,
-		    const DoubledGaugeFieldView &UUU, 
- 		    SiteSpinor * buf, int LLs, int sU, 
-		    const FermionFieldView &in,
-		    const FermionFieldView &out,int dag);
+   void DhopSiteAsm(StencilView &st, 
+		    DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
+		    SiteSpinor * buf, int LLs, int sU, 
+		    const FermionFieldView &in, FermionFieldView &out,int dag);
   
 public:
 

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -61,7 +61,7 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) const {
+  accelerator_inline int CommDatumSize(void) {
     return sizeof(SiteHalfCommSpinor);
   }
 
@@ -69,7 +69,7 @@ public:
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
   template<class _SiteHalfSpinor, class _SiteSpinor>
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
+  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
     _SiteHalfSpinor tmp;
     projector::Proj(tmp,in,mu,dag);
     vstream(buf[o],tmp);
@@ -81,7 +81,7 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
 				   const SiteHalfSpinor * __restrict__ vp0,
 				   const SiteHalfSpinor * __restrict__ vp1,
-				   Integer type,Integer o) const {
+				   Integer type,Integer o){
     SiteHalfSpinor tmp1;
     SiteHalfSpinor tmp2;
     exchange(tmp1,tmp2,vp0[o],vp1[o],type);
@@ -93,7 +93,7 @@ public:
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
   accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
-				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
+				     SiteHalfSpinor * __restrict__ in, Integer o) {    
     assert(0);
   }
 
@@ -103,7 +103,7 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
 					   SiteHalfSpinor * __restrict__ out1,
 					   const SiteSpinor * __restrict__ in,
-					   Integer j,Integer k, Integer m,Integer type) const
+					   Integer j,Integer k, Integer m,Integer type)
   {
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
@@ -117,7 +117,7 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return false; }
+  accelerator_inline bool DecompressionStep(void) { return false; }
 
 };
 
@@ -142,7 +142,7 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) const {
+  accelerator_inline int CommDatumSize(void) {
     return sizeof(SiteHalfCommSpinor);
   }
 
@@ -150,7 +150,7 @@ public:
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
   template<class _SiteHalfSpinor, class _SiteSpinor>
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
+  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
     _SiteHalfSpinor hsp;
     SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
     projector::Proj(hsp,in,mu,dag);
@@ -163,7 +163,7 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
                        SiteHalfSpinor *vp0,
                        SiteHalfSpinor *vp1,
-		       Integer type,Integer o) const {
+		       Integer type,Integer o){
     SiteHalfSpinor vt0,vt1;
     SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
     SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
@@ -175,7 +175,7 @@ public:
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
+  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o){
     SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
     precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
   }
@@ -186,7 +186,7 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
 			       SiteHalfSpinor *out1,
 			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type) const {
+			       Integer j,Integer k, Integer m,Integer type){
     SiteHalfSpinor temp1, temp2,temp3,temp4;
     SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
     SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
@@ -200,7 +200,7 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return true; }
+  accelerator_inline bool DecompressionStep(void) { return true; }
 
 };
 

Grid/qcd/action/fermion/WilsonFermion.h

@@ -63,6 +63,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
   typedef WilsonKernels<Impl> Kernels;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -72,6 +72,9 @@ public:
   typedef WilsonKernels<Impl> Kernels;
   PmuStat stat;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 

Grid/qcd/action/fermion/WilsonImpl.h

@@ -95,7 +95,7 @@ public:
 					  const _Spinor &chi,
 					  int mu,
 					  StencilEntry *SE,
-					  const StencilView &St) 
+					  StencilView &St) 
   {
     multLink(phi,U,chi,mu);
   }
@@ -106,15 +106,11 @@ 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);
-    typedef decltype(coalescedRead(out_v[0]))   calcSpinor;
-    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
-	calcSpinor tmp;
-	multLink(tmp,Umu_v[sss],phi_v(sss),mu);
-	coalescedWrite(out_v[sss],tmp);
+    accelerator_for(sss,out.Grid()->oSites(),1,{
+	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
     });
   }
 					   

Grid/qcd/action/fermion/WilsonKernels.h

@@ -49,7 +49,6 @@ public:
 
   INHERIT_IMPL_TYPES(Impl);
   typedef FermionOperator<Impl> Base;
-  typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;
    
 public:
 
@@ -69,87 +68,73 @@ public:
 
 private:
 
-  static accelerator_inline void DhopDirK(const StencilView &st, const DoubledGaugeFieldView &U,
-					  SiteHalfSpinor * buf, int sF, int sU,
-					  const FermionFieldView &in,const FermionFieldView &out, int dirdisp, int gamma);
+  static accelerator_inline void DhopDirK(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor * buf,
+				   int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dirdisp, int gamma);
 
-  static accelerator_inline void DhopDirXp(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirYp(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirZp(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirTp(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirXm(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirYm(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirZm(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
-  static accelerator_inline void DhopDirTm(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,
-					   const FermionFieldView &in, const FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirXp(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirYp(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirZp(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirTp(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirXm(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirYm(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirZm(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
+  static accelerator_inline void DhopDirTm(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,int sF,int sU,const FermionFieldView &in,FermionFieldView &out,int dirdisp);
       
   // Specialised variants
-  static accelerator void GenericDhopSite(const StencilView &st,
-					  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					  int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-       
-  static accelerator void GenericDhopSiteDag(const StencilView &st, const  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					     int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-   
-  static accelerator void GenericDhopSiteInt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					     int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-       
-  static accelerator void GenericDhopSiteDagInt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-						int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-   
-  static accelerator void GenericDhopSiteExt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					     int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-       
-  static accelerator void GenericDhopSiteDagExt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-						int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
+  static accelerator void GenericDhopSite(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					  int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
+      
+  static accelerator void GenericDhopSiteDag(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+						    int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
+  
+  static accelerator void GenericDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+						    int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
+      
+  static accelerator void GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+						int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
+  
+  static accelerator void GenericDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					     int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
+      
+  static accelerator void GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+						       int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
 
-// Keep Hand unrolled 
-  static accelerator void HandDhopSiteSycl(StencilVector st_perm, StencilEntry *st_p,  SiteDoubledGaugeField *U, SiteHalfSpinor * buf,
-					   int sF, int sU, const SiteSpinor *in, SiteSpinor *out);
+  static void AsmDhopSite(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+			  int sF, int sU, int Ls, int Nsite, const FermionFieldView &in,FermionFieldView &out);
+  
+  static void AsmDhopSiteDag(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+			     int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out);
+  
+  static void AsmDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+			     int sF, int sU, int Ls, int Nsite, const FermionFieldView &in,FermionFieldView &out);
+  
+  static void AsmDhopSiteDagInt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+				int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out);
+  
+  static void AsmDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+			     int sF, int sU, int Ls, int Nsite, const FermionFieldView &in,FermionFieldView &out);
+  
+  static void AsmDhopSiteDagExt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+				int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out);
 
-  static accelerator void HandDhopSite(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-				       int sF, int sU, const FermionFieldView &in,const FermionFieldView &out);
-   
-  static accelerator void HandDhopSiteDag(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					  int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-   
-  static accelerator void HandDhopSiteInt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					  int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
+// Keep Hand unrolled temporarily  
+  static accelerator void HandDhopSite(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+				       int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
   
-  static accelerator void HandDhopSiteDagInt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					     int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
+  static accelerator void HandDhopSiteDag(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					  int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
   
-  static accelerator void HandDhopSiteExt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					  int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-   
-  static accelerator void HandDhopSiteDagExt(const StencilView &st, const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-					     int sF, int sU, const FermionFieldView &in, const FermionFieldView &out);
-  //AVX 512 ASM
-  static void AsmDhopSite(const StencilView &st,  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-			  int sF, int sU, int Ls, int Nsite, const FermionFieldView &in,const FermionFieldView &out);
+  static accelerator void HandDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					  int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
   
-  static void AsmDhopSiteDag(const StencilView &st,  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-			     int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, const FermionFieldView &out);
+  static accelerator void HandDhopSiteDagInt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					     int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
   
-  static void AsmDhopSiteInt(const StencilView &st,  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-			     int sF, int sU, int Ls, int Nsite, const FermionFieldView &in,const FermionFieldView &out);
+  static accelerator void HandDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					  int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
   
-  static void AsmDhopSiteDagInt(const StencilView &st,  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-				int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, const FermionFieldView &out);
-  
-  static void AsmDhopSiteExt(const StencilView &st,  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-			     int sF, int sU, int Ls, int Nsite, const FermionFieldView &in,const FermionFieldView &out);
-  
-  static void AsmDhopSiteDagExt(const StencilView &st,  const DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
-				int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, const FermionFieldView &out);
-
+  static accelerator void HandDhopSiteDagExt(StencilView &st,  DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
+					     int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
  public:
  WilsonKernels(const ImplParams &p = ImplParams()) : Base(p){};
 };

Grid/qcd/action/fermion/g5HermitianLinop.h

@@ -79,6 +79,8 @@ public:
     _Mat.M(in,tmp);
     G5R5(out,tmp);
   }
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
 };
 
 
@@ -127,6 +129,8 @@ public:
     _Mat.M(in,tmp);
     out=g5*tmp;
   }
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
 };
 
 NAMESPACE_END(Grid);

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

@@ -642,7 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-#if (!defined(GRID_HIP))
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -799,7 +799,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
-  PropagatorField zz (UGrid);   zz=0.0;
+  LatticeInteger zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
   for (int s=0;s<Ls;s++) {
 
@@ -826,7 +826,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
-#if (!defined(GRID_HIP))
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
@@ -850,7 +850,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
 
-  PropagatorField  zz (UGrid);   zz=0.0;
+  LatticeInteger zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
 
   for(int s=0;s<Ls;s++){

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

@@ -618,13 +618,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteAsm(const StencilView &st,
-					 const DoubledGaugeFieldView &U,
-					 const DoubledGaugeFieldView &UUU,
- 					 SiteSpinor *buf, int sF,
-					 int sU,
-					 const FermionFieldView &in,
-					 const FermionFieldView &out,int dag) 
+void StaggeredKernels<Impl>::DhopSiteAsm(StencilView &st,
+					 DoubledGaugeFieldView &U,
+					 DoubledGaugeFieldView &UUU,
+					 SiteSpinor *buf, int sF,
+					 int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   assert(0);
 };
@@ -685,13 +683,11 @@ void StaggeredKernels<Impl>::DhopSiteAsm(const StencilView &st,
 
   // This is the single precision 5th direction vectorised kernel
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(const StencilView &st,
-								    const DoubledGaugeFieldView &U,
-								    const DoubledGaugeFieldView &UUU,
- 								    SiteSpinor *buf, int sF,
-								    int sU,
-								    const FermionFieldView &in,
-								    const FermionFieldView &out,int dag) 
+template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
+								    DoubledGaugeFieldView &U,
+								    DoubledGaugeFieldView &UUU,
+								    SiteSpinor *buf, int sF,
+								    int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -742,13 +738,11 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(const Stenci
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(const StencilView &st, 
-								    const DoubledGaugeFieldView &U,
-								    const DoubledGaugeFieldView &UUU,
- 								    SiteSpinor *buf, int sF,
-								    int sU,
-								    const FermionFieldView &in,
-								    const FermionFieldView &out, int dag) 
+template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView &st, 
+								    DoubledGaugeFieldView &U,
+								    DoubledGaugeFieldView &UUU,
+								    SiteSpinor *buf, int sF,
+								    int sU, const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -830,13 +824,11 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(const Stenci
   // This is the single precision 5th direction vectorised kernel
 
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(const StencilView &st, 
-							       const DoubledGaugeFieldView &U,
-							       const DoubledGaugeFieldView &UUU,
- 							       SiteSpinor *buf, int sF,
-							       int sU,
-							       const FermionFieldView &in,
-							       const FermionFieldView &out,int dag) 
+template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st, 
+							       DoubledGaugeFieldView &U,
+							       DoubledGaugeFieldView &UUU,
+							       SiteSpinor *buf, int sF,
+							       int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -901,13 +893,11 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(const StencilView
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(const StencilView &st, 
-							       const DoubledGaugeFieldView &U,
-							       const DoubledGaugeFieldView &UUU,
- 							       SiteSpinor *buf, int sF,
-							       int sU,
-							       const FermionFieldView &in,
-							       const FermionFieldView &out,int dag) 
+template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilView &st, 
+							       DoubledGaugeFieldView &U,
+							       DoubledGaugeFieldView &UUU,
+							       SiteSpinor *buf, int sF,
+							       int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;

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

@@ -146,13 +146,11 @@ NAMESPACE_BEGIN(Grid);
 
 
 template <class Impl>
-template <int Naik> accelerator_inline
-void StaggeredKernels<Impl>::DhopSiteHand(const StencilView &st,
-					  const DoubledGaugeFieldView &U,
-					  const DoubledGaugeFieldView &UUU,
- 					  SiteSpinor *buf, int sF, int sU, 
-					  const FermionFieldView &in,
-					  const FermionFieldView &out,int dag) 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
+					  DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
+					  SiteSpinor *buf, int sF, int sU, 
+					  const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -223,13 +221,11 @@ void StaggeredKernels<Impl>::DhopSiteHand(const StencilView &st,
 
 
 template <class Impl>
-template <int Naik> accelerator_inline
-void StaggeredKernels<Impl>::DhopSiteHandInt(const StencilView &st, 
-					     const DoubledGaugeFieldView &U,
-					     const DoubledGaugeFieldView &UUU,
- 					     SiteSpinor *buf, int sF, int sU, 
-					     const FermionFieldView &in,
-					     const FermionFieldView &out,int dag) 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st, 
+					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
+					     SiteSpinor *buf, int sF, int sU, 
+					     const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -304,13 +300,11 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(const StencilView &st,
 
 
 template <class Impl>
-template <int Naik> accelerator_inline
-void StaggeredKernels<Impl>::DhopSiteHandExt(const StencilView &st,
-					     const DoubledGaugeFieldView &U,
-					     const DoubledGaugeFieldView &UUU,
- 					     SiteSpinor *buf, int sF, int sU, 
-					     const FermionFieldView &in,
-					     const FermionFieldView &out,int dag) 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
+					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
+					     SiteSpinor *buf, int sF, int sU, 
+					     const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;

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

@@ -78,11 +78,11 @@ StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
 // Int, Ext, Int+Ext cases for comms overlap
 ////////////////////////////////////////////////////////////////////////////////////
 template <class Impl>
-template <int Naik> accelerator_inline
-void StaggeredKernels<Impl>::DhopSiteGeneric(const StencilView &st, 
-					     const DoubledGaugeFieldView &U, const DoubledGaugeFieldView &UUU,
- 					     SiteSpinor *buf, int sF, int sU, 
-					     const FermionFieldView &in, const FermionFieldView &out, int dag) 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st, 
+					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
+					     SiteSpinor *buf, int sF, int sU, 
+					     const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
   const SiteSpinor *chi_p;
   SiteSpinor chi;
@@ -126,12 +126,11 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(const StencilView &st,
   // Only contributions from interior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-template <int Naik> accelerator_inline
-void StaggeredKernels<Impl>::DhopSiteGenericInt(const StencilView &st, 
-						const DoubledGaugeFieldView &U, const DoubledGaugeFieldView &UUU,
- 						SiteSpinor *buf, int sF, int sU, 
-						const FermionFieldView &in, const FermionFieldView &out,int dag)
-{
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st, 
+						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
+						SiteSpinor *buf, int sF, int sU, 
+						const FermionFieldView &in, FermionFieldView &out,int dag) {
   const SiteSpinor *chi_p;
   SiteSpinor chi;
   SiteSpinor Uchi;
@@ -175,14 +174,11 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(const StencilView &st,
   // Only contributions from exterior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-template <int Naik> accelerator_inline
-void StaggeredKernels<Impl>::DhopSiteGenericExt(const StencilView &st, 
-						const DoubledGaugeFieldView &U,
-						const DoubledGaugeFieldView &UUU,
- 						SiteSpinor *buf, int sF, int sU,
-						const FermionFieldView &in,
-						const FermionFieldView &out,int dag)
-{
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st, 
+						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
+						SiteSpinor *buf, int sF, int sU,
+						const FermionFieldView &in, FermionFieldView &out,int dag) {
   const SiteSpinor *chi_p;
   //  SiteSpinor chi;
   SiteSpinor Uchi;
@@ -228,14 +224,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(const StencilView &st,
 ////////////////////////////////////////////////////////////////////////////////////
 // Driving / wrapping routine to select right kernel
 ////////////////////////////////////////////////////////////////////////////////////
-template <class Impl> 
-void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st,
-					   const DoubledGaugeFieldView &U,
-					   const DoubledGaugeFieldView &UUU,
-					   SiteSpinor * buf,
-					   int sF, int sU,
-					   const FermionFieldView &in,
-					   const FermionFieldView &out, int dir,int disp)
+template <class Impl>
+void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+					   int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp)
 {
   // Disp should be either +1,-1,+3,-3
   // What about "dag" ?
@@ -262,9 +253,8 @@ void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st,
       ThisKernel::A(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag);		\
   });
 
-template <class Impl> 
-void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st,
-					  LebesgueOrder &lo, 
+template <class Impl>
+void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
 					  DoubledGaugeField &U, DoubledGaugeField &UUU, 
 					  const FermionField &in, FermionField &out, int dag, int interior,int exterior)
 {
@@ -303,7 +293,7 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st,
   }
   assert(0 && " Kernel optimisation case not covered ");
 }
-template <class Impl> 
+template <class Impl>
 void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
 				       DoubledGaugeField &U,
 				       const FermionField &in, FermionField &out, int dag, int interior,int exterior)

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

@@ -92,16 +92,20 @@ 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);
-    thread_for(site, lvol, {
-      Coordinate lcoor;
+    for (int site = 0; site < lvol; site++) {
       grid->LocalIndexToLocalCoor(site, lcoor);
-      Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-      Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-      typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
+      EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
       peekLocalSite(Qx, CTv, lcoor);
+      Qxinv = Zero();
       //if (csw!=0){
       for (int j = 0; j < Ns; j++)
 	for (int k = 0; k < Ns; k++)
@@ -122,7 +126,7 @@ 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/WilsonKernelsAsmA64FX.h

@@ -38,6 +38,9 @@ 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
@@ -60,89 +63,119 @@ 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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
 
 /////////////////////////////////////////////////////////////////
@@ -152,89 +185,119 @@ 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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
@@ -267,89 +330,119 @@ 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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
@@ -358,93 +451,124 @@ 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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#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<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/WilsonKernelsAsmBodyA64FX.h

@@ -25,11 +25,6 @@ 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
@@ -102,7 +97,7 @@ 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);					                    \
@@ -115,11 +110,6 @@ 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);						            \
@@ -136,63 +126,73 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
 
 #define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
       basep = st.GetPFInfo(nent,plocal); nent++;			\
-      if ( local ) {							\
-  LOAD_CHIMU(base);                                       \
-  LOAD_TABLE(PERMUTE_DIR);                                \
-  PROJ;							                        \
-  MAYBEPERM(PERMUTE_DIR,perm);					        \
-      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}			\
-      if ( local || st.same_node[Dir] ) {				\
-  MULT_2SPIN_1(Dir);					                    \
-  MULT_2SPIN_2;					                        \
-  RECON;								\
-      }									\
-  base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
-  PREFETCH_CHIMU(base);						\
-  PREFETCH_CHIMU_L2(basep);                               \
+      if ( local ) {							            \
+    LOAD_CHIMU(base);                                       \
+    LOAD_TABLE(PERMUTE_DIR);                                \
+    PROJ;							                        \
+    MAYBEPERM(PERMUTE_DIR,perm);					        \
+      }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] ) {				    \
+    MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+    MULT_2SPIN_2;					                        \
+    if (s == 0) {                                           \
+       if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    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++;		\
-  if((!local)&&(!st.same_node[Dir]) ) {					\
-    LOAD_CHI(base);							\
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
+  if((!local)&&(!st.same_node[Dir]) ) {					    \
+    LOAD_CHI(base);							                \
     MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    RECON;								\
-    nmu++;								\
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    nmu++;								                    \
   }
 
-#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
-  nmu=0;								\
-  { ZERO_PSI;}								\
-  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
-  if((!local)&&(!st.same_node[Dir]) ) {					\
-    LOAD_CHI(base);							\
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
+  nmu=0;								                    \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;\
+  if((!local)&&(!st.same_node[Dir]) ) {					    \
+    LOAD_CHI(base);							                \
     MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    RECON;								\
-    nmu++;								\
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    nmu++;								                    \
   }
 
 #define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
 
 #endif
-
-
 {
   int nmu;
   int local,perm, ptype;
@@ -209,6 +209,7 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
     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;
@@ -294,11 +295,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       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
@@ -312,11 +308,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       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
@@ -330,11 +321,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       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
@@ -355,7 +341,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       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/WilsonKernelsAsmImplementation.h

@@ -38,46 +38,46 @@ NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////
 // Default to no assembler implementation
-// Will specialise to AVX512 if available
+// Will specialise to 
 ///////////////////////////////////////////////////////////
 template<class Impl> void 
-WilsonKernels<Impl >::AsmDhopSite(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-				  int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, const FermionFieldView &out)
+WilsonKernels<Impl >::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+				  int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 {
   assert(0);
 }
 
 template<class Impl> void 
-WilsonKernels<Impl >::AsmDhopSiteDag(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-				     int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, const FermionFieldView &out)
+WilsonKernels<Impl >::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+				     int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 {
   assert(0);
 }
 
 template<class Impl> void 
-WilsonKernels<Impl >::AsmDhopSiteInt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-				     int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, const FermionFieldView &out)
+WilsonKernels<Impl >::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+				     int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 {
   assert(0);
 }
 
 template<class Impl> void 
-WilsonKernels<Impl >::AsmDhopSiteDagInt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-					int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, const FermionFieldView &out)
+WilsonKernels<Impl >::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+					int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 {
   assert(0);
 }
 
 template<class Impl> void 
-WilsonKernels<Impl >::AsmDhopSiteExt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-				     int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, const FermionFieldView &out)
+WilsonKernels<Impl >::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+				     int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 {
   assert(0);
 }
 
 template<class Impl> void 
-WilsonKernels<Impl >::AsmDhopSiteDagExt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-					int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, const FermionFieldView &out)
+WilsonKernels<Impl >::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+					int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 {
   assert(0);
 }

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

@@ -646,14 +646,9 @@ NAMESPACE_BEGIN(Grid);
   HAND_RESULT_EXT(ss,F)
 
 #define HAND_SPECIALISE_GPARITY(IMPL)					\
-  template<> accelerator_inline void					\
-  WilsonKernels<IMPL>::HandDhopSiteSycl(StencilVector st_perm, StencilEntry *st_p, \
-					SiteDoubledGaugeField *U, SiteHalfSpinor * buf, \
-					int sF, int sU, const SiteSpinor *in, SiteSpinor *out) {} \
-  									\
-  template<> accelerator_inline void						\
-  WilsonKernels<IMPL>::HandDhopSite(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
-				    int ss,int sU,const FermionFieldView &in, const FermionFieldView &out) \
+  template<> void						\
+  WilsonKernels<IMPL>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
+				    int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
     typedef IMPL Impl;							\
     typedef typename Simd::scalar_type S;				\
@@ -667,9 +662,9 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> accelerator_inline void						\
-  WilsonKernels<IMPL>::HandDhopSiteDag(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
-				       int ss,int sU,const FermionFieldView &in, const FermionFieldView &out) \
+  template<> void						\
+  WilsonKernels<IMPL>::HandDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
+				       int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
     typedef IMPL Impl;							\
     typedef typename Simd::scalar_type S;				\
@@ -683,9 +678,9 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> accelerator_inline void						\
-  WilsonKernels<IMPL>::HandDhopSiteInt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
-				       int ss,int sU,const FermionFieldView &in, const FermionFieldView &out) \
+  template<> void						\
+  WilsonKernels<IMPL>::HandDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
+				       int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
     typedef IMPL Impl;							\
     typedef typename Simd::scalar_type S;				\
@@ -699,9 +694,9 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> accelerator_inline void						\
-  WilsonKernels<IMPL>::HandDhopSiteDagInt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
-					  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out) \
+  template<> void						\
+  WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
     typedef IMPL Impl;							\
     typedef typename Simd::scalar_type S;				\
@@ -715,9 +710,9 @@ NAMESPACE_BEGIN(Grid);
     HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
 									\
-  template<> accelerator_inline void							\
-  WilsonKernels<IMPL>::HandDhopSiteExt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
-				       int ss,int sU,const FermionFieldView &in, const FermionFieldView &out) \
+  template<> void							\
+  WilsonKernels<IMPL>::HandDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf, \
+				       int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
     typedef IMPL Impl;							\
     typedef typename Simd::scalar_type S;				\
@@ -732,9 +727,9 @@ NAMESPACE_BEGIN(Grid);
     nmu = 0;								\
     HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
   }									\
-  template<> accelerator_inline void						\
-  WilsonKernels<IMPL>::HandDhopSiteDagExt(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
-					  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out) \
+  template<> void						\
+  WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
   {									\
     typedef IMPL Impl;							\
     typedef typename Simd::scalar_type S;				\

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

@@ -495,9 +495,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl> accelerator_inline void 
-WilsonKernels<Impl>::HandDhopSite(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
-				  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out)
+template<class Impl> void 
+WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
@@ -519,9 +519,9 @@ WilsonKernels<Impl>::HandDhopSite(const StencilView &st, const DoubledGaugeField
   HAND_RESULT(ss);
 }
 
-template<class Impl>  accelerator_inline
-void WilsonKernels<Impl>::HandDhopSiteDag(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-					  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out)
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -542,9 +542,9 @@ void WilsonKernels<Impl>::HandDhopSiteDag(const StencilView &st,const DoubledGau
   HAND_RESULT(ss);
 }
 
-template<class Impl>  accelerator_inline void 
-WilsonKernels<Impl>::HandDhopSiteInt(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
-					  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out)
+template<class Impl> void 
+WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
@@ -566,9 +566,9 @@ WilsonKernels<Impl>::HandDhopSiteInt(const StencilView &st,const DoubledGaugeFie
   HAND_RESULT(ss);
 }
 
-template<class Impl> accelerator_inline
-void WilsonKernels<Impl>::HandDhopSiteDagInt(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-						  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out)
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -589,9 +589,9 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(const StencilView &st,const Doubled
   HAND_RESULT(ss);
 }
 
-template<class Impl>  accelerator_inline void 
-WilsonKernels<Impl>::HandDhopSiteExt(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
-					  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out)
+template<class Impl> void 
+WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
@@ -614,9 +614,9 @@ WilsonKernels<Impl>::HandDhopSiteExt(const StencilView &st,const DoubledGaugeFie
   HAND_RESULT_EXT(ss);
 }
 
-template<class Impl>  accelerator_inline
-void WilsonKernels<Impl>::HandDhopSiteDagExt(const StencilView &st,const DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
-						  int ss,int sU,const FermionFieldView &in, const FermionFieldView &out)
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -682,4 +682,3 @@ NAMESPACE_END(Grid);
 #undef HAND_RESULT
 #undef HAND_RESULT_INT
 #undef HAND_RESULT_EXT
-#undef HAND_DECLARATIONS

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

@@ -1,598 +0,0 @@
-   /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/qcd/action/fermion/WilsonKernelsHand.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-
-#pragma once
-
-#include <Grid/qcd/action/fermion/FermionCore.h>
-
-
-#undef LOAD_CHIMU  
-#undef LOAD_CHI 
-#undef MULT_2SPIN
-#undef PERMUTE_DIR
-#undef XP_PROJ  
-#undef YP_PROJ  
-#undef ZP_PROJ  
-#undef TP_PROJ  
-#undef XM_PROJ  
-#undef YM_PROJ  
-#undef ZM_PROJ  
-#undef TM_PROJ  
-#undef XP_RECON 
-#undef XP_RECON_ACCUM 
-#undef XM_RECON 
-#undef XM_RECON_ACCUM 
-#undef YP_RECON_ACCUM 
-#undef YM_RECON_ACCUM 
-#undef ZP_RECON_ACCUM 
-#undef ZM_RECON_ACCUM 
-#undef TP_RECON_ACCUM 
-#undef TM_RECON_ACCUM 
-#undef ZERO_RESULT				 
-#undef Chimu_00
-#undef Chimu_01
-#undef Chimu_02
-#undef Chimu_10
-#undef Chimu_11
-#undef Chimu_12
-#undef Chimu_20
-#undef Chimu_21
-#undef Chimu_22
-#undef Chimu_30
-#undef Chimu_31
-#undef Chimu_32
-#undef HAND_STENCIL_LEG
-#undef HAND_STENCIL_LEG_INT
-#undef HAND_STENCIL_LEG_EXT
-#undef HAND_RESULT
-#undef HAND_RESULT_INT
-#undef HAND_RESULT_EXT
-
-#define REGISTER
-
-#ifdef GRID_SIMT
-#define LOAD_CHIMU(ptype)		\
-  {const SiteSpinor & ref (in[offset]);	\
-    Chimu_00=coalescedReadPermute<ptype>(ref()(0)(0),perm);	\
-    Chimu_01=coalescedReadPermute<ptype>(ref()(0)(1),perm);	\
-    Chimu_02=coalescedReadPermute<ptype>(ref()(0)(2),perm);	\
-    Chimu_10=coalescedReadPermute<ptype>(ref()(1)(0),perm);	\
-    Chimu_11=coalescedReadPermute<ptype>(ref()(1)(1),perm);	\
-    Chimu_12=coalescedReadPermute<ptype>(ref()(1)(2),perm);	\
-    Chimu_20=coalescedReadPermute<ptype>(ref()(2)(0),perm);	\
-    Chimu_21=coalescedReadPermute<ptype>(ref()(2)(1),perm);	\
-    Chimu_22=coalescedReadPermute<ptype>(ref()(2)(2),perm);	\
-    Chimu_30=coalescedReadPermute<ptype>(ref()(3)(0),perm);	\
-    Chimu_31=coalescedReadPermute<ptype>(ref()(3)(1),perm);	\
-    Chimu_32=coalescedReadPermute<ptype>(ref()(3)(2),perm);	}
-
-#define PERMUTE_DIR(dir) ;
-#else
-#define LOAD_CHIMU(ptype)		\
-  {const SiteSpinor & ref (in[offset]);	\
-    Chimu_00=coalescedRead(ref()(0)(0));	\
-    Chimu_01=coalescedRead(ref()(0)(1));	\
-    Chimu_02=coalescedRead(ref()(0)(2));	\
-    Chimu_10=coalescedRead(ref()(1)(0));	\
-    Chimu_11=coalescedRead(ref()(1)(1));	\
-    Chimu_12=coalescedRead(ref()(1)(2));	\
-    Chimu_20=coalescedRead(ref()(2)(0));	\
-    Chimu_21=coalescedRead(ref()(2)(1));	\
-    Chimu_22=coalescedRead(ref()(2)(2));	\
-    Chimu_30=coalescedRead(ref()(3)(0));	\
-    Chimu_31=coalescedRead(ref()(3)(1));	\
-    Chimu_32=coalescedRead(ref()(3)(2));	}
-
-#define PERMUTE_DIR(dir)			\
-  permute##dir(Chi_00,Chi_00);	\
-      permute##dir(Chi_01,Chi_01);\
-      permute##dir(Chi_02,Chi_02);\
-      permute##dir(Chi_10,Chi_10);	\
-      permute##dir(Chi_11,Chi_11);\
-      permute##dir(Chi_12,Chi_12);
-#endif
-
-#define MULT_2SPIN(A)\
-  {auto & ref(U[sU](A));					\
-  U_00=coalescedRead(ref()(0,0));				\
-  U_10=coalescedRead(ref()(1,0));					\
-  U_20=coalescedRead(ref()(2,0));					\
-  U_01=coalescedRead(ref()(0,1));					\
-  U_11=coalescedRead(ref()(1,1));					\
-  U_21=coalescedRead(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;					\
-    U_00=coalescedRead(ref()(0,2));				\
-    U_10=coalescedRead(ref()(1,2));				\
-    U_20=coalescedRead(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 LOAD_CHI				\
-  {const SiteHalfSpinor &ref(buf[offset]);	\
-    Chi_00 = coalescedRead(ref()(0)(0));	\
-    Chi_01 = coalescedRead(ref()(0)(1));	\
-    Chi_02 = coalescedRead(ref()(0)(2));	\
-    Chi_10 = coalescedRead(ref()(1)(0));	\
-    Chi_11 = coalescedRead(ref()(1)(1));	\
-    Chi_12 = coalescedRead(ref()(1)(2));}
-
-//      hspin(0)=fspin(0)+timesI(fspin(3));
-//      hspin(1)=fspin(1)+timesI(fspin(2));
-#define XP_PROJ \
-    Chi_00 = Chimu_00+timesI(Chimu_30);\
-    Chi_01 = Chimu_01+timesI(Chimu_31);\
-    Chi_02 = Chimu_02+timesI(Chimu_32);\
-    Chi_10 = Chimu_10+timesI(Chimu_20);\
-    Chi_11 = Chimu_11+timesI(Chimu_21);\
-    Chi_12 = Chimu_12+timesI(Chimu_22);
-
-#define YP_PROJ \
-    Chi_00 = Chimu_00-Chimu_30;\
-    Chi_01 = Chimu_01-Chimu_31;\
-    Chi_02 = Chimu_02-Chimu_32;\
-    Chi_10 = Chimu_10+Chimu_20;\
-    Chi_11 = Chimu_11+Chimu_21;\
-    Chi_12 = Chimu_12+Chimu_22;
-
-#define ZP_PROJ \
-  Chi_00 = Chimu_00+timesI(Chimu_20);		\
-  Chi_01 = Chimu_01+timesI(Chimu_21);		\
-  Chi_02 = Chimu_02+timesI(Chimu_22);		\
-  Chi_10 = Chimu_10-timesI(Chimu_30);		\
-  Chi_11 = Chimu_11-timesI(Chimu_31);		\
-  Chi_12 = Chimu_12-timesI(Chimu_32);
-
-#define TP_PROJ \
-  Chi_00 = Chimu_00+Chimu_20;		\
-  Chi_01 = Chimu_01+Chimu_21;		\
-  Chi_02 = Chimu_02+Chimu_22;		\
-  Chi_10 = Chimu_10+Chimu_30;		\
-  Chi_11 = Chimu_11+Chimu_31;		\
-  Chi_12 = Chimu_12+Chimu_32;
-
-
-//      hspin(0)=fspin(0)-timesI(fspin(3));
-//      hspin(1)=fspin(1)-timesI(fspin(2));
-#define XM_PROJ \
-    Chi_00 = Chimu_00-timesI(Chimu_30);\
-    Chi_01 = Chimu_01-timesI(Chimu_31);\
-    Chi_02 = Chimu_02-timesI(Chimu_32);\
-    Chi_10 = Chimu_10-timesI(Chimu_20);\
-    Chi_11 = Chimu_11-timesI(Chimu_21);\
-    Chi_12 = Chimu_12-timesI(Chimu_22);
-
-#define YM_PROJ \
-    Chi_00 = Chimu_00+Chimu_30;\
-    Chi_01 = Chimu_01+Chimu_31;\
-    Chi_02 = Chimu_02+Chimu_32;\
-    Chi_10 = Chimu_10-Chimu_20;\
-    Chi_11 = Chimu_11-Chimu_21;\
-    Chi_12 = Chimu_12-Chimu_22;
-
-#define ZM_PROJ \
-  Chi_00 = Chimu_00-timesI(Chimu_20);		\
-  Chi_01 = Chimu_01-timesI(Chimu_21);		\
-  Chi_02 = Chimu_02-timesI(Chimu_22);		\
-  Chi_10 = Chimu_10+timesI(Chimu_30);		\
-  Chi_11 = Chimu_11+timesI(Chimu_31);		\
-  Chi_12 = Chimu_12+timesI(Chimu_32);
-
-#define TM_PROJ \
-  Chi_00 = Chimu_00-Chimu_20;		\
-  Chi_01 = Chimu_01-Chimu_21;		\
-  Chi_02 = Chimu_02-Chimu_22;		\
-  Chi_10 = Chimu_10-Chimu_30;		\
-  Chi_11 = Chimu_11-Chimu_31;		\
-  Chi_12 = Chimu_12-Chimu_32;
-
-//      fspin(0)=hspin(0);
-//      fspin(1)=hspin(1);
-//      fspin(2)=timesMinusI(hspin(1));
-//      fspin(3)=timesMinusI(hspin(0));
-#define XP_RECON\
-  result_00 = UChi_00;\
-  result_01 = UChi_01;\
-  result_02 = UChi_02;\
-  result_10 = UChi_10;\
-  result_11 = UChi_11;\
-  result_12 = UChi_12;\
-  result_20 = timesMinusI(UChi_10);\
-  result_21 = timesMinusI(UChi_11);\
-  result_22 = timesMinusI(UChi_12);\
-  result_30 = timesMinusI(UChi_00);\
-  result_31 = timesMinusI(UChi_01);\
-  result_32 = timesMinusI(UChi_02);
-
-#define XP_RECON_ACCUM\
-  result_00+=UChi_00;\
-  result_01+=UChi_01;\
-  result_02+=UChi_02;\
-  result_10+=UChi_10;\
-  result_11+=UChi_11;\
-  result_12+=UChi_12;\
-  result_20-=timesI(UChi_10);\
-  result_21-=timesI(UChi_11);\
-  result_22-=timesI(UChi_12);\
-  result_30-=timesI(UChi_00);\
-  result_31-=timesI(UChi_01);\
-  result_32-=timesI(UChi_02);
-
-#define XM_RECON\
-  result_00 = UChi_00;\
-  result_01 = UChi_01;\
-  result_02 = UChi_02;\
-  result_10 = UChi_10;\
-  result_11 = UChi_11;\
-  result_12 = UChi_12;\
-  result_20 = timesI(UChi_10);\
-  result_21 = timesI(UChi_11);\
-  result_22 = timesI(UChi_12);\
-  result_30 = timesI(UChi_00);\
-  result_31 = timesI(UChi_01);\
-  result_32 = timesI(UChi_02);
-
-#define XM_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20+= timesI(UChi_10);\
-  result_21+= timesI(UChi_11);\
-  result_22+= timesI(UChi_12);\
-  result_30+= timesI(UChi_00);\
-  result_31+= timesI(UChi_01);\
-  result_32+= timesI(UChi_02);
-
-#define YP_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20+= UChi_10;\
-  result_21+= UChi_11;\
-  result_22+= UChi_12;\
-  result_30-= UChi_00;\
-  result_31-= UChi_01;\
-  result_32-= UChi_02;
-
-#define YM_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20-= UChi_10;\
-  result_21-= UChi_11;\
-  result_22-= UChi_12;\
-  result_30+= UChi_00;\
-  result_31+= UChi_01;\
-  result_32+= UChi_02;
-
-#define ZP_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20-= timesI(UChi_00);			\
-  result_21-= timesI(UChi_01);			\
-  result_22-= timesI(UChi_02);			\
-  result_30+= timesI(UChi_10);			\
-  result_31+= timesI(UChi_11);			\
-  result_32+= timesI(UChi_12);
-
-#define ZM_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20+= timesI(UChi_00);			\
-  result_21+= timesI(UChi_01);			\
-  result_22+= timesI(UChi_02);			\
-  result_30-= timesI(UChi_10);			\
-  result_31-= timesI(UChi_11);			\
-  result_32-= timesI(UChi_12);
-
-#define TP_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20+= UChi_00;			\
-  result_21+= UChi_01;			\
-  result_22+= UChi_02;			\
-  result_30+= UChi_10;			\
-  result_31+= UChi_11;			\
-  result_32+= UChi_12;
-
-#define TM_RECON_ACCUM\
-  result_00+= UChi_00;\
-  result_01+= UChi_01;\
-  result_02+= UChi_02;\
-  result_10+= UChi_10;\
-  result_11+= UChi_11;\
-  result_12+= UChi_12;\
-  result_20-= UChi_00;	\
-  result_21-= UChi_01;	\
-  result_22-= UChi_02;	\
-  result_30-= UChi_10;	\
-  result_31-= UChi_11;	\
-  result_32-= UChi_12;
-
-#define HAND_STENCIL_LEGA(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;					\
-  }						\
-  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;			\
-  LOAD_CHIMU(PERM);				\
-  PROJ;						\
-  MULT_2SPIN(DIR);				\
-  RECON;					
-
-
-#define HAND_STENCIL_LEG_INT(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;					\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else if ( st.same_node[DIR] ) {		\
-    LOAD_CHI;					\
-  }						\
-  if (local || st.same_node[DIR] ) {		\
-    MULT_2SPIN(DIR);				\
-    RECON;					\
-  }
-
-#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {	\
-    LOAD_CHI;					\
-    MULT_2SPIN(DIR);				\
-    RECON;					\
-    nmu++;					\
-  }
-
-#define HAND_RESULT(ss)				\
-  {						\
-    SiteSpinor & ref (out[ss]);			\
-    coalescedWrite(ref()(0)(0),result_00);		\
-    coalescedWrite(ref()(0)(1),result_01);		\
-    coalescedWrite(ref()(0)(2),result_02);		\
-    coalescedWrite(ref()(1)(0),result_10);		\
-    coalescedWrite(ref()(1)(1),result_11);		\
-    coalescedWrite(ref()(1)(2),result_12);		\
-    coalescedWrite(ref()(2)(0),result_20);		\
-    coalescedWrite(ref()(2)(1),result_21);		\
-    coalescedWrite(ref()(2)(2),result_22);		\
-    coalescedWrite(ref()(3)(0),result_30);		\
-    coalescedWrite(ref()(3)(1),result_31);		\
-    coalescedWrite(ref()(3)(2),result_32);		\
-  }
-
-#define HAND_RESULT_EXT(ss)			\
-  if (nmu){					\
-    SiteSpinor & ref (out[ss]);		\
-    ref()(0)(0)+=result_00;		\
-    ref()(0)(1)+=result_01;		\
-    ref()(0)(2)+=result_02;		\
-    ref()(1)(0)+=result_10;		\
-    ref()(1)(1)+=result_11;		\
-    ref()(1)(2)+=result_12;		\
-    ref()(2)(0)+=result_20;		\
-    ref()(2)(1)+=result_21;		\
-    ref()(2)(2)+=result_22;		\
-    ref()(3)(0)+=result_30;		\
-    ref()(3)(1)+=result_31;		\
-    ref()(3)(2)+=result_32;		\
-  }
-
-#define HAND_DECLARATIONS(Simd)			\
-  Simd result_00;				\
-  Simd result_01;				\
-  Simd result_02;				\
-  Simd result_10;				\
-  Simd result_11;				\
-  Simd result_12;				\
-  Simd result_20;				\
-  Simd result_21;				\
-  Simd result_22;				\
-  Simd result_30;				\
-  Simd result_31;				\
-  Simd result_32;				\
-  Simd Chi_00;					\
-  Simd Chi_01;					\
-  Simd Chi_02;					\
-  Simd Chi_10;					\
-  Simd Chi_11;					\
-  Simd Chi_12;					\
-  Simd UChi_00;					\
-  Simd UChi_01;					\
-  Simd UChi_02;					\
-  Simd UChi_10;					\
-  Simd UChi_11;					\
-  Simd UChi_12;					\
-  Simd U_00;					\
-  Simd U_10;					\
-  Simd U_20;					\
-  Simd U_01;					\
-  Simd U_11;					\
-  Simd U_21;
-
-#define ZERO_RESULT				\
-  result_00=Zero();				\
-  result_01=Zero();				\
-  result_02=Zero();				\
-  result_10=Zero();				\
-  result_11=Zero();				\
-  result_12=Zero();				\
-  result_20=Zero();				\
-  result_21=Zero();				\
-  result_22=Zero();				\
-  result_30=Zero();				\
-  result_31=Zero();				\
-  result_32=Zero();			
-
-#define Chimu_00 Chi_00
-#define Chimu_01 Chi_01
-#define Chimu_02 Chi_02
-#define Chimu_10 Chi_10
-#define Chimu_11 Chi_11
-#define Chimu_12 Chi_12
-#define Chimu_20 UChi_00
-#define Chimu_21 UChi_01
-#define Chimu_22 UChi_02
-#define Chimu_30 UChi_10
-#define Chimu_31 UChi_11
-#define Chimu_32 UChi_12
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Impl> accelerator_inline void 
-WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
-				      int ss,int sU,const SiteSpinor *in, SiteSpinor *out)
-{
-// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-  typedef typename Simd::scalar_type S;
-  typedef typename Simd::vector_type V;
-  typedef iSinglet<Simd> vCplx;
-  //  typedef decltype( coalescedRead( vCplx()()() )) Simt;
-  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
-
-  HAND_DECLARATIONS(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);
-}
-
-////////////// Wilson ; uses this implementation /////////////////////
-
-NAMESPACE_END(Grid);
-#undef LOAD_CHIMU  
-#undef LOAD_CHI 
-#undef MULT_2SPIN
-#undef PERMUTE_DIR
-#undef XP_PROJ  
-#undef YP_PROJ  
-#undef ZP_PROJ  
-#undef TP_PROJ  
-#undef XM_PROJ  
-#undef YM_PROJ  
-#undef ZM_PROJ  
-#undef TM_PROJ  
-#undef XP_RECON 
-#undef XP_RECON_ACCUM 
-#undef XM_RECON 
-#undef XM_RECON_ACCUM 
-#undef YP_RECON_ACCUM 
-#undef YM_RECON_ACCUM 
-#undef ZP_RECON_ACCUM 
-#undef ZM_RECON_ACCUM 
-#undef TP_RECON_ACCUM 
-#undef TM_RECON_ACCUM 
-#undef ZERO_RESULT				 
-#undef Chimu_00
-#undef Chimu_01
-#undef Chimu_02
-#undef Chimu_10
-#undef Chimu_11
-#undef Chimu_12
-#undef Chimu_20
-#undef Chimu_21
-#undef Chimu_22
-#undef Chimu_30
-#undef Chimu_31
-#undef Chimu_32
-#undef HAND_STENCIL_LEG
-#undef HAND_STENCIL_LEG_INT
-#undef HAND_STENCIL_LEG_EXT
-#undef HAND_RESULT
-#undef HAND_RESULT_INT
-#undef HAND_RESULT_EXT
-#undef HAND_DECLARATIONS

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

@@ -114,10 +114,10 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   ////////////////////////////////////////////////////////////////////
   // All legs kernels ; comms then compute
   ////////////////////////////////////////////////////////////////////
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::GenericDhopSiteDag(const StencilView &st, const DoubledGaugeFieldView &U,
- 					     SiteHalfSpinor *buf, int sF,
-					     int sU, const FermionFieldView &in, const FermionFieldView &out)
+template <class Impl>
+void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,
+					     SiteHalfSpinor *buf, int sF,
+					     int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0]))   calcHalfSpinor;
   typedef decltype(coalescedRead(in[0])) calcSpinor;
@@ -140,10 +140,10 @@ void WilsonKernels<Impl>::GenericDhopSiteDag(const StencilView &st, const Double
   coalescedWrite(out[sF],result,lane);
 };
 
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::GenericDhopSite(const StencilView &st, const DoubledGaugeFieldView &U,
- 					  SiteHalfSpinor *buf, int sF,
-					  int sU, const FermionFieldView &in, const FermionFieldView &out)
+template <class Impl>
+void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U,
+					  SiteHalfSpinor *buf, int sF,
+					  int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -169,10 +169,10 @@ void WilsonKernels<Impl>::GenericDhopSite(const StencilView &st, const DoubledGa
   ////////////////////////////////////////////////////////////////////
   // Interior kernels
   ////////////////////////////////////////////////////////////////////
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::GenericDhopSiteDagInt(const StencilView &st, const DoubledGaugeFieldView &U,
- 						SiteHalfSpinor *buf, int sF,
-						int sU, const FermionFieldView &in, const FermionFieldView &out)
+template <class Impl>
+void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFieldView &U,
+						SiteHalfSpinor *buf, int sF,
+						int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -197,10 +197,10 @@ void WilsonKernels<Impl>::GenericDhopSiteDagInt(const StencilView &st, const Dou
   coalescedWrite(out[sF], result,lane);
 };
 
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::GenericDhopSiteInt(const StencilView &st, const DoubledGaugeFieldView &U,
-					     SiteHalfSpinor *buf, int sF,
-					     int sU, const FermionFieldView &in, const FermionFieldView &out)
+template <class Impl>
+void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U,
+							 SiteHalfSpinor *buf, int sF,
+							 int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -227,10 +227,10 @@ void WilsonKernels<Impl>::GenericDhopSiteInt(const StencilView &st, const Double
 ////////////////////////////////////////////////////////////////////
 // Exterior kernels
 ////////////////////////////////////////////////////////////////////
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::GenericDhopSiteDagExt(const StencilView &st, const DoubledGaugeFieldView &U,
- 						SiteHalfSpinor *buf, int sF,
-						int sU, const FermionFieldView &in, const FermionFieldView &out)
+template <class Impl>
+void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFieldView &U,
+						SiteHalfSpinor *buf, int sF,
+						int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -258,10 +258,10 @@ void WilsonKernels<Impl>::GenericDhopSiteDagExt(const StencilView &st, const Dou
   }
 };
 
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::GenericDhopSiteExt(const StencilView &st, const DoubledGaugeFieldView &U,
- 					     SiteHalfSpinor *buf, int sF,
-					     int sU, const FermionFieldView &in, const FermionFieldView &out)
+template <class Impl>
+void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U,
+					     SiteHalfSpinor *buf, int sF,
+					     int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -290,9 +290,9 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(const StencilView &st, const Double
 };
 
 #define DhopDirMacro(Dir,spProj,spRecon)	\
-  template <class Impl> accelerator_inline				\
-  void WilsonKernels<Impl>::DhopDir##Dir(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF, \
-					 int sU, const FermionFieldView &in, const FermionFieldView &out, int dir) \
+  template <class Impl>							\
+  void WilsonKernels<Impl>::DhopDir##Dir(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF, \
+					 int sU, const FermionFieldView &in, FermionFieldView &out, int dir) \
   {									\
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;		\
   typedef decltype(coalescedRead(in[0]))  calcSpinor;			\
@@ -302,8 +302,8 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(const StencilView &st, const Double
   StencilEntry *SE;							\
   int ptype;								\
   const int Nsimd = SiteHalfSpinor::Nsimd();				\
-  const int lane=acceleratorSIMTlane(Nsimd);				\
-  									\
+  const int lane=acceleratorSIMTlane(Nsimd);					\
+									\
   SE = st.GetEntry(ptype, dir, sF);					\
   GENERIC_DHOPDIR_LEG_BODY(Dir,spProj,spRecon);				\
   coalescedWrite(out[sF], result,lane);					\
@@ -318,9 +318,9 @@ DhopDirMacro(Ym,spProjYm,spReconYm);
 DhopDirMacro(Zm,spProjZm,spReconZm);
 DhopDirMacro(Tm,spProjTm,spReconTm);
 
-template <class Impl> accelerator_inline
-void WilsonKernels<Impl>::DhopDirK(const StencilView &st, const DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF,
-				   int sU, const FermionFieldView &in, const FermionFieldView &out, int dir, int gamma)
+template <class Impl>
+void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF,
+				    int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int gamma)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -345,8 +345,8 @@ void WilsonKernels<Impl>::DhopDirK(const StencilView &st, const DoubledGaugeFiel
 }
 
 template <class Impl>
-void WilsonKernels<Impl>::DhopDirAll(StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor *buf, int Ls,
-				     int Nsite, const FermionField &in, std::vector<FermionField> &out)
+void WilsonKernels<Impl>::DhopDirAll( StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor *buf, int Ls,
+				      int Nsite, const FermionField &in, std::vector<FermionField> &out)
 {
    autoView(U_v  ,U,AcceleratorRead);
    autoView(in_v ,in,AcceleratorRead);
@@ -424,20 +424,6 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
   });
 
-#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_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
 #define ASM_CALL(A)							\
@@ -460,8 +446,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
 #ifndef GRID_CUDA
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
-     //     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
    } else if( interior ) {

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

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

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

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

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

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

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

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

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

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

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

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

Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc

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

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

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

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

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

Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc

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

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

@@ -32,7 +32,6 @@ directory
 #include <Grid/qcd/action/fermion/FermionCore.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementationSycl.h>
 
 #ifndef AVX512
 #ifndef QPX

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc

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

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

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

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

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc

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

Grid/qcd/action/gauge/Gauge.cc

@@ -1,38 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/gauge/Gauge.cc
-
-Copyright (C) 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/qcd/action/fermion/FermionCore.h>
-
-NAMESPACE_BEGIN(Grid);
-
-std::vector<int> ConjugateGaugeImplBase::_conjDirs;
-
-NAMESPACE_END(Grid);
-

Grid/qcd/action/gauge/GaugeImplTypes.h

@@ -154,10 +154,6 @@ public:
     return Hsum.real();
   }
 
-  static inline void Project(Field &U) {
-    ProjectSUn(U);
-  }
-
   static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
     SU<Nc>::HotConfiguration(pRNG, U);
   }

Grid/qcd/action/gauge/GaugeImplementations.h

@@ -59,14 +59,14 @@ public:
   }
   static inline GaugeLinkField
   CovShiftIdentityBackward(const GaugeLinkField &Link, int mu) {
-    return PeriodicBC::CovShiftIdentityBackward(Link, mu);
+    return Cshift(adj(Link), mu, -1);
   }
   static inline GaugeLinkField
   CovShiftIdentityForward(const GaugeLinkField &Link, int mu) {
-    return PeriodicBC::CovShiftIdentityForward(Link,mu);
+    return Link;
   }
   static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu) {
-    return PeriodicBC::ShiftStaple(Link,mu);
+    return Cshift(Link, mu, 1);
   }
 
   static inline bool isPeriodicGaugeField(void) { return true; }
@@ -74,13 +74,7 @@ public:
 
 // Composition with smeared link, bc's etc.. probably need multiple inheritance
 // Variable precision "S" and variable Nc
-class ConjugateGaugeImplBase {
-protected:
-  static std::vector<int> _conjDirs;
-};
-
-  template <class GimplTypes> class ConjugateGaugeImpl : public GimplTypes, ConjugateGaugeImplBase {
-private:
+template <class GimplTypes> class ConjugateGaugeImpl : public GimplTypes {
 public:
   INHERIT_GIMPL_TYPES(GimplTypes);
 
@@ -90,56 +84,47 @@ public:
   ////////////////////////////////////////////////////////////////////////////////////////////////////////////
   template <class covariant>
   static Lattice<covariant> CovShiftForward(const GaugeLinkField &Link, int mu,
-                                            const Lattice<covariant> &field)
-  {
-    assert(_conjDirs.size() == Nd);
-    if(_conjDirs[mu]) 
-      return ConjugateBC::CovShiftForward(Link, mu, field);
-    else
-      return PeriodicBC::CovShiftForward(Link, mu, field);
+                                            const Lattice<covariant> &field) {
+    return ConjugateBC::CovShiftForward(Link, mu, field);
   }
 
   template <class covariant>
   static Lattice<covariant> CovShiftBackward(const GaugeLinkField &Link, int mu,
-                                             const Lattice<covariant> &field)
-  {
-    assert(_conjDirs.size() == Nd);
-    if(_conjDirs[mu]) 
-      return ConjugateBC::CovShiftBackward(Link, mu, field);
-    else 
-      return PeriodicBC::CovShiftBackward(Link, mu, field);
+                                             const Lattice<covariant> &field) {
+    return ConjugateBC::CovShiftBackward(Link, mu, field);
   }
 
   static inline GaugeLinkField
-  CovShiftIdentityBackward(const GaugeLinkField &Link, int mu)
-  {
-    assert(_conjDirs.size() == Nd);
-    if(_conjDirs[mu]) 
-      return ConjugateBC::CovShiftIdentityBackward(Link, mu);
-    else 
-      return PeriodicBC::CovShiftIdentityBackward(Link, mu);
+  CovShiftIdentityBackward(const GaugeLinkField &Link, int mu) {
+    GridBase *grid = Link.Grid();
+    int Lmu = grid->GlobalDimensions()[mu] - 1;
+
+    Lattice<iScalar<vInteger>> coor(grid);
+    LatticeCoordinate(coor, mu);
+
+    GaugeLinkField tmp(grid);
+    tmp = adj(Link);
+    tmp = where(coor == Lmu, conjugate(tmp), tmp);
+    return Cshift(tmp, mu, -1); // moves towards positive mu
   }
   static inline GaugeLinkField
-  CovShiftIdentityForward(const GaugeLinkField &Link, int mu)
-  {
-    assert(_conjDirs.size() == Nd);
-    if(_conjDirs[mu]) 
-      return ConjugateBC::CovShiftIdentityForward(Link,mu);
-    else
-      return PeriodicBC::CovShiftIdentityForward(Link,mu);
+  CovShiftIdentityForward(const GaugeLinkField &Link, int mu) {
+    return Link;
   }
 
-  static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu)
-  {
-    assert(_conjDirs.size() == Nd);
-    if(_conjDirs[mu]) 
-      return ConjugateBC::ShiftStaple(Link,mu);
-    else     
-      return PeriodicBC::ShiftStaple(Link,mu);
+  static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu) {
+    GridBase *grid = Link.Grid();
+    int Lmu = grid->GlobalDimensions()[mu] - 1;
+
+    Lattice<iScalar<vInteger>> coor(grid);
+    LatticeCoordinate(coor, mu);
+
+    GaugeLinkField tmp(grid);
+    tmp = Cshift(Link, mu, 1);
+    tmp = where(coor == Lmu, conjugate(tmp), tmp);
+    return tmp;
   }
 
-  static inline void       setDirections(std::vector<int> &conjDirs) { _conjDirs=conjDirs; }
-  static inline std::vector<int> getDirections(void) { return _conjDirs; }
   static inline bool isPeriodicGaugeField(void) { return false; }
 };