finer timers in Benchmark_IO

Jamie's fix

.travis.yml

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

Grid/DisableWarnings.h

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

Grid/GridStd.h

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

Grid/Makefile.am

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

Grid/algorithms/CoarsenedMatrix.h

@@ -31,6 +31,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_ALGORITHM_COARSENED_MATRIX_H
 #define  GRID_ALGORITHM_COARSENED_MATRIX_H
 
+#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
 
 NAMESPACE_BEGIN(Grid);
 
@@ -59,12 +60,14 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
 class Geometry {
 public:
   int npoint;
+  int base;
   std::vector<int> directions   ;
   std::vector<int> displacements;
+  std::vector<int> points_dagger;
 
   Geometry(int _d)  {
     
-    int base = (_d==5) ? 1:0;
+    base = (_d==5) ? 1:0;
 
     // make coarse grid stencil for 4d , not 5d
     if ( _d==5 ) _d=4;
@@ -72,16 +75,51 @@ public:
     npoint = 2*_d+1;
     directions.resize(npoint);
     displacements.resize(npoint);
+    points_dagger.resize(npoint);
     for(int d=0;d<_d;d++){
       directions[d   ] = d+base;
       directions[d+_d] = d+base;
       displacements[d  ] = +1;
       displacements[d+_d]= -1;
+      points_dagger[d   ] = d+_d;
+      points_dagger[d+_d] = d;
     }
     directions   [2*_d]=0;
     displacements[2*_d]=0;
+    points_dagger[2*_d]=2*_d;
   }
 
+  int point(int dir, int disp) {
+    assert(disp == -1 || disp == 0 || disp == 1);
+    assert(base+0 <= dir && dir < base+4);
+
+    // directions faster index = new indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  1  2  3  0  1  2  3  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  2  3  4  1  2  3  4  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+
+    // displacements faster index = old indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  0  1  1  2  2  3  3  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  1  2  2  3  3  4  4  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+
+    if(dir == 0 and disp == 0)
+      return 8;
+    else // New indexing
+      return (1 - disp) / 2 * 4 + dir - base;
+    // else // Old indexing
+    //   return (4 * (dir - base) + 1 - disp) / 2;
+  }
 };
   
 template<class Fobj,class CComplex,int nbasis>
@@ -258,7 +296,7 @@ public:
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>
-class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+class CoarsenedMatrix : public CheckerBoardedSparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
 public:
     
   typedef iVector<CComplex,nbasis >           siteVector;
@@ -268,33 +306,59 @@ public:
   typedef iMatrix<CComplex,nbasis >  Cobj;
   typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
   typedef Lattice<Fobj >        FineField;
+  typedef CoarseVector FermionField;
+
+  // enrich interface, use default implementation as in FermionOperator ///////
+  void Dminus(CoarseVector const& in, CoarseVector& out) { out = in; }
+  void DminusDag(CoarseVector const& in, CoarseVector& out) { out = in; }
+  void ImportPhysicalFermionSource(CoarseVector const& input, CoarseVector& imported) { imported = input; }
+  void ImportUnphysicalFermion(CoarseVector const& input, CoarseVector& imported) { imported = input; }
+  void ExportPhysicalFermionSolution(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
+  void ExportPhysicalFermionSource(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
 
   ////////////////////
   // Data members
   ////////////////////
   Geometry         geom;
   GridBase *       _grid; 
+  GridBase*        _cbgrid;
   int hermitian;
 
   CartesianStencil<siteVector,siteVector,int> Stencil; 
+  CartesianStencil<siteVector,siteVector,int> StencilEven;
+  CartesianStencil<siteVector,siteVector,int> StencilOdd;
 
   std::vector<CoarseMatrix> A;
+  std::vector<CoarseMatrix> Aeven;
+  std::vector<CoarseMatrix> Aodd;
+
+  CoarseMatrix AselfInv;
+  CoarseMatrix AselfInvEven;
+  CoarseMatrix AselfInvOdd;
+
+  Vector<RealD> dag_factor;
 
   ///////////////////////
   // Interface
   ///////////////////////
   GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know
+  GridBase * RedBlackGrid()     { return _cbgrid; };
+
+  int ConstEE() { return 0; }
 
   void M (const CoarseVector &in, CoarseVector &out)
   {
     conformable(_grid,in.Grid());
     conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
 
     SimpleCompressor<siteVector> compressor;
 
     Stencil.HaloExchange(in,compressor);
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+    auto& geom_v = geom;
     typedef LatticeView<Cobj> Aview;
       
     Vector<Aview> AcceleratorViewContainer;
@@ -316,14 +380,14 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom.npoint;point++){
+      for(int point=0;point<geom_v.npoint;point++){
 
-	SE=Stencil.GetEntry(ptype,point,ss);
+	SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
 	if(SE->_is_local) { 
 	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
 	} else {
-	  nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
+	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
 	}
 	acceleratorSynchronise();
 
@@ -344,12 +408,72 @@ public:
       return M(in,out);
     } else {
       // corresponds to Galerkin coarsening
-      CoarseVector tmp(Grid());
+      return MdagNonHermitian(in, out);
-      G5C(tmp, in); 
-      M(tmp, out);
-      G5C(out, out);
     }
   };
+
+  void MdagNonHermitian(const CoarseVector &in, CoarseVector &out)
+  {
+    conformable(_grid,in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    SimpleCompressor<siteVector> compressor;
+
+    Stencil.HaloExchange(in,compressor);
+    autoView( in_v , in, AcceleratorRead);
+    autoView( out_v , out, AcceleratorWrite);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+    auto& geom_v = geom;
+    typedef LatticeView<Cobj> Aview;
+
+    Vector<Aview> AcceleratorViewContainer;
+
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    Aview *Aview_p = & AcceleratorViewContainer[0];
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    int osites=Grid()->oSites();
+
+    Vector<int> points(geom.npoint, 0);
+    for(int p=0; p<geom.npoint; p++)
+      points[p] = geom.points_dagger[p];
+
+    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;
@@ -359,6 +483,7 @@ public:
   {
     conformable(_grid,in.Grid());
     conformable(_grid,out.Grid());
+    out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
     Vector<Aview> AcceleratorViewContainer;
@@ -367,6 +492,7 @@ public:
 
     autoView( out_v , out, AcceleratorWrite);
     autoView( in_v  , in, AcceleratorRead);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
 
     const int Nsimd = CComplex::Nsimd();
     typedef decltype(coalescedRead(in_v[0])) calcVector;
@@ -380,12 +506,12 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      SE=Stencil.GetEntry(ptype,point,ss);
+      SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
       if(SE->_is_local) { 
 	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
       } else {
-	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
+	nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
       }
       acceleratorSynchronise();
 
@@ -413,34 +539,7 @@ public:
 
     this->MdirComms(in);
 
-    int ndim = in.Grid()->Nd();
+    MdirCalc(in,out,geom.point(dir,disp));
-
-    //////////////
-    // 4D action like wilson
-    // 0+ => 0 
-    // 0- => 1
-    // 1+ => 2 
-    // 1- => 3
-    // etc..
-    //////////////
-    // 5D action like DWF
-    // 1+ => 0 
-    // 1- => 1
-    // 2+ => 2 
-    // 2- => 3
-    // etc..
-    auto point = [dir, disp, ndim](){
-      if(dir == 0 and disp == 0)
-	return 8;
-      else if ( ndim==4 ) { 
-	return (4 * dir + 1 - disp) / 2;
-      } else { 
-	return (4 * (dir-1) + 1 - disp) / 2;
-      }
-    }();
-
-    MdirCalc(in,out,point);
-
   };
 
   void Mdiag(const CoarseVector &in, CoarseVector &out)
@@ -449,23 +548,296 @@ 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) 	:
-
     _grid(&CoarseGrid),
+    _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
     Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-      A(geom.npoint,&CoarseGrid)
+    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    A(geom.npoint,&CoarseGrid),
+    Aeven(geom.npoint,_cbgrid),
+    Aodd(geom.npoint,_cbgrid),
+    AselfInv(&CoarseGrid),
+    AselfInvEven(_cbgrid),
+    AselfInvOdd(_cbgrid),
+    dag_factor(nbasis*nbasis)
   {
+    fillFactor();
   };
 
+  CoarsenedMatrix(GridCartesian &CoarseGrid, GridRedBlackCartesian &CoarseRBGrid, int hermitian_=0) 	:
+
+    _grid(&CoarseGrid),
+    _cbgrid(&CoarseRBGrid),
+    geom(CoarseGrid._ndimension),
+    hermitian(hermitian_),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    A(geom.npoint,&CoarseGrid),
+    Aeven(geom.npoint,&CoarseRBGrid),
+    Aodd(geom.npoint,&CoarseRBGrid),
+    AselfInv(&CoarseGrid),
+    AselfInvEven(&CoarseRBGrid),
+    AselfInvOdd(&CoarseRBGrid),
+    dag_factor(nbasis*nbasis)
+  {
+    fillFactor();
+  };
+
+  void fillFactor() {
+    Eigen::MatrixXd dag_factor_eigen = Eigen::MatrixXd::Ones(nbasis, nbasis);
+    if(!hermitian) {
+      const int nb = nbasis/2;
+      dag_factor_eigen.block(0,nb,nb,nb) *= -1.0;
+      dag_factor_eigen.block(nb,0,nb,nb) *= -1.0;
+    }
+
+    // GPU readable prefactor
+    thread_for(i, nbasis*nbasis, {
+      int j = i/nbasis;
+      int k = i%nbasis;
+      dag_factor[i] = dag_factor_eigen(j, k);
+    });
+  }
+
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
 		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
   {
     typedef Lattice<typename Fobj::tensor_reduced> FineComplexField;
     typedef typename Fobj::scalar_type scalar_type;
 
+    std::cout << GridLogMessage<< "CoarsenMatrix "<< std::endl;
+
     FineComplexField one(FineGrid); one=scalar_type(1.0,0.0);
     FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0);
 
@@ -496,11 +868,13 @@ public:
 
     CoarseScalar InnerProd(Grid()); 
 
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog "<< std::endl;
     // Orthogonalise the subblocks over the basis
     blockOrthogonalise(InnerProd,Subspace.subspace);
 
     // Compute the matrix elements of linop between this orthonormal
     // set of vectors.
+    std::cout << GridLogMessage<< "CoarsenMatrix masks "<< std::endl;
     int self_stencil=-1;
     for(int p=0;p<geom.npoint;p++)
     { 
@@ -539,7 +913,7 @@ public:
 
       phi=Subspace.subspace[i];
 
-      //      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl;
+      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl;
       linop.OpDirAll(phi,Mphi_p);
       linop.OpDiag  (phi,Mphi_p[geom.npoint-1]);
 
@@ -568,6 +942,18 @@ public:
 	    autoView( A_self  , A[self_stencil], AcceleratorWrite);
 
 	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
+	    if ( hermitian && (disp==-1) ) {
+	      for(int pp=0;pp<geom.npoint;pp++){// Find the opposite link and set <j|A|i> = <i|A|j>*
+		int dirp   = geom.directions[pp];
+		int dispp  = geom.displacements[pp];
+		if ( (dirp==dir) && (dispp==1) ){
+		  auto sft = conjugate(Cshift(oZProj,dir,1));
+		  autoView( sft_v    ,  sft  , AcceleratorWrite);
+		  autoView( A_pp     ,  A[pp], AcceleratorWrite);
+		  accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_pp[ss](i,j),sft_v(ss)); });
+		}
+	      }
+	    }
 
 	  }
 	}
@@ -606,28 +992,54 @@ public:
     }
     if(hermitian) {
       std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
-      ForceHermitian();
-    }
     }
 
-  void ForceHermitian(void) {
+    InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl;
-    CoarseMatrix Diff  (Grid());
+    FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl;
-    for(int p=0;p<geom.npoint;p++){
-      int dir   = geom.directions[p];
-      int disp  = geom.displacements[p];
-      if(disp==-1) {
-	// Find the opposite link
-	for(int pp=0;pp<geom.npoint;pp++){
-	  int dirp   = geom.directions[pp];
-	  int dispp  = geom.displacements[pp];
-	  if ( (dirp==dir) && (dispp==1) ){
-	    //	    Diff = adj(Cshift(A[p],dir,1)) - A[pp]; 
-	    //	    std::cout << GridLogMessage<<" Replacing stencil leg "<<pp<<" with leg "<<p<< " diff "<<norm2(Diff) <<std::endl;
-	    A[pp] = adj(Cshift(A[p],dir,1));
-	  }
   }
+
+  void InvertSelfStencilLink() {
+    std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl;
+    int localVolume = Grid()->lSites();
+
+    typedef typename Cobj::scalar_object scalar_object;
+
+    autoView(Aself_v,    A[geom.npoint-1], CpuRead);
+    autoView(AselfInv_v, AselfInv,         CpuWrite);
+    thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke
+      Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis);
+      Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis);
+
+      scalar_object selfLink    = Zero();
+      scalar_object selfLinkInv = Zero();
+
+      Coordinate lcoor;
+
+      Grid()->LocalIndexToLocalCoor(site, lcoor);
+      peekLocalSite(selfLink, Aself_v, lcoor);
+
+      for (int i = 0; i < nbasis; ++i)
+        for (int j = 0; j < nbasis; ++j)
+          selfLinkEigen(i, j) = static_cast<ComplexD>(TensorRemove(selfLink(i, j)));
+
+      selfLinkInvEigen = selfLinkEigen.inverse();
+
+      for(int i = 0; i < nbasis; ++i)
+        for(int j = 0; j < nbasis; ++j)
+          selfLinkInv(i, j) = selfLinkInvEigen(i, j);
+
+      pokeLocalSite(selfLinkInv, AselfInv_v, lcoor);
+    });
   }
+
+  void FillHalfCbs() {
+    std::cout << GridLogDebug << "CoarsenedMatrix::FillHalfCbs" << std::endl;
+    for(int p = 0; p < geom.npoint; ++p) {
+      pickCheckerboard(Even, Aeven[p], A[p]);
+      pickCheckerboard(Odd, Aodd[p], A[p]);
     }
+    pickCheckerboard(Even, AselfInvEven, AselfInv);
+    pickCheckerboard(Odd, AselfInvOdd, AselfInv);
   }
 };
 

Grid/allocator/AlignedAllocator.cc

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

Grid/allocator/AlignedAllocator.h

@@ -165,9 +165,18 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-//template<class T> using commAllocator = devAllocator<T>;
+#ifdef ACCELERATOR_CSHIFT
+// Cshift on device
+template<class T> using cshiftAllocator = devAllocator<T>;
+#else
+// Cshift on host
+template<class T> using cshiftAllocator = std::allocator<T>;
+#endif
+
 template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
+template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
 template<class T> using commVector = std::vector<T,devAllocator<T> >;
+template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryManager.h

@@ -34,8 +34,6 @@ NAMESPACE_BEGIN(Grid);
 
 // Move control to configure.ac and Config.h?
 
-#define ALLOCATION_CACHE
-#define GRID_ALLOC_ALIGN (2*1024*1024)
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
 /*Pinning pages is costly*/

Grid/allocator/MemoryManagerCache.cc

@@ -1,11 +1,12 @@
 #include <Grid/GridCore.h>
-
 #ifndef GRID_UVM
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
+//define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
 #define dprintf(...)
 
+
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
 ////////////////////////////////////////////////////////////
@@ -103,7 +104,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  //  dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -111,7 +112,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    //    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -125,7 +126,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  //  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   if(AccCache.state==AccDirty) {
@@ -136,7 +137,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    //    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -149,7 +150,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  //  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -164,7 +165,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  //  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -227,18 +228,24 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   // Find if present, otherwise get or force an empty
   ////////////////////////////////////////////////////////////////////////////
   if ( EntryPresent(CpuPtr)==0 ){
-    EvictVictims(bytes);
     EntryCreate(CpuPtr,bytes,mode,hint);
   }
 
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
-  
+  if (!AccCache.AccPtr) {
+    EvictVictims(bytes); 
+  } 
   assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
 
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
+    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
+		    (uint64_t)AccCache.CpuPtr,
+		    (uint64_t)CpuPtr,
+		    (uint64_t)AccCache.bytes,
+		    (uint64_t)bytes);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -285,21 +292,21 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    //    printf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    //    printf("Consistent entry into device accLock %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    //    printf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
@@ -361,13 +368,16 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   // Find if present, otherwise get or force an empty
   ////////////////////////////////////////////////////////////////////////////
   if ( EntryPresent(CpuPtr)==0 ){
-    EvictVictims(bytes);
     EntryCreate(CpuPtr,bytes,mode,transient);
   }
 
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
+  if (!AccCache.AccPtr) {
+     EvictVictims(bytes);
+  }
+
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
 

Grid/allocator/MemoryManagerShared.cc

@@ -1,7 +1,6 @@
 #include <Grid/GridCore.h>
 #ifdef GRID_UVM
 
-#warning "Grid is assuming unified virtual memory address space"
 NAMESPACE_BEGIN(Grid);
 /////////////////////////////////////////////////////////////////////////////////
 // View management is 1:1 address space mapping

Grid/cartesian/Cartesian_red_black.h

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

Grid/communicator/Communicator_base.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -108,6 +107,8 @@ public:
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
+  void GlobalMax(RealD &);
+  void GlobalMax(RealF &);
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
   void GlobalSum(RealD &);

Grid/communicator/Communicator_mpi3.cc

@@ -44,7 +44,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
 
-#if defined (TOFU) // FUGAKU, credits go to Issaku Kanamori
+#ifndef GRID_COMMS_THREADS
     nCommThreads=1;
     // wrong results here too
     // For now: comms-overlap leads to wrong results in Benchmark_wilson even on single node MPI runs
@@ -275,6 +275,16 @@ void CartesianCommunicator::GlobalXOR(uint64_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalMax(float &f)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSum(float &f){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
@@ -358,16 +368,19 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
+  int tag;
 
   if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
+    tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
   if ( gdest == MPI_UNDEFINED ) {
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
+    tag= dir+_processor*32;
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;

Grid/communicator/Communicator_none.cc

@@ -67,6 +67,8 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 
 CartesianCommunicator::~CartesianCommunicator(){}
 
+void CartesianCommunicator::GlobalMax(float &){}
+void CartesianCommunicator::GlobalMax(double &){}
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}

Grid/communicator/SharedMemory.h

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

Grid/communicator/SharedMemoryMPI.cc

@@ -7,6 +7,7 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christoph Lehner <christoph@lhnr.de>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -169,6 +170,23 @@ static inline int divides(int a,int b)
 }
 void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
 {
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
   ////////////////////////////////////////////////////////////////
   // Powers of 2,3,5 only in prime decomposition for now
   ////////////////////////////////////////////////////////////////
@@ -457,8 +475,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
-  if ( WorldRank == 0 ){
+  //  if ( WorldRank == 0 ){
-    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes 
+  if ( 1 ){
+    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -665,7 +684,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -715,7 +733,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
 {
 #ifdef GRID_CUDA
   cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
@@ -771,22 +789,15 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
 
-#ifdef GRID_IBM_SUMMIT
+#ifdef GRID_SHM_FORCE_MPI
-  // Hide the shared memory path between sockets 
+  // Hide the shared memory path between ranks
-  // if even number of nodes
+  {
-  if ( (ShmSize & 0x1)==0 ) {
-    int SocketSize = ShmSize/2;
-    int mySocket = ShmRank/SocketSize; 
     for(int r=0;r<size;r++){
-      int hisRank=ShmRanks[r];
+      if ( r!=rank ) {
-      if ( hisRank!= MPI_UNDEFINED ) {
-	int hisSocket=hisRank/SocketSize;
-	if ( hisSocket != mySocket ) {
 	ShmRanks[r] = MPI_UNDEFINED;
       }
     }
   }
-  }
 #endif
 
   SharedMemoryTest();

Grid/communicator/SharedMemoryNone.cc

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

Grid/cshift/Cshift_common.h

@@ -35,7 +35,7 @@ extern Vector<std::pair<int,int> > Cshift_table;
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -73,12 +73,19 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
      }
   }
   {
-    autoView(rhs_v , rhs, AcceleratorRead);
     auto buffer_p = & buffer[0];
     auto table = &Cshift_table[0];
+#ifdef ACCELERATOR_CSHIFT    
+    autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for(i,ent,{
+      buffer_p[table[i].first]=rhs_v[table[i].second];
+    });
+#endif
   }
 }
 
@@ -103,21 +110,36 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for2d(n,e1,b,e2,1,{
+    accelerator_for(nn,e1*e2,1,{
+	int n = nn%e1;
+	int b = nn/e1;
 	int o      =   n*n1;
 	int offset = b+n*e2;
 	
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
-  } else { 
+#else
-    autoView(rhs_v , rhs, AcceleratorRead);
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+	int o      =   n*n1;
+	int offset = b+n*e2;
 	
+	vobj temp =rhs_v[so+o+b];
+	extract<vobj>(temp,pointers,offset);
+      });
+#endif
+  } else { 
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-    accelerator_for2d(n,e1,b,e2,1,{
+#ifdef ACCELERATOR_CSHIFT    
+    autoView(rhs_v , rhs, AcceleratorRead);
+    accelerator_for(nn,e1*e2,1,{
+	int n = nn%e1;
+	int b = nn/e1;
 
 	Coordinate coor;
 
@@ -134,13 +156,33 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+
+	Coordinate coor;
+
+	int o=n*n1;
+	int oindex = o+b;
+
+       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
+
+	int ocb=1<<cb;
+	int offset = b+n*e2;
+
+	if ( ocb & cbmask ) {
+	  vobj temp =rhs_v[so+o+b];
+	  extract<vobj>(temp,pointers,offset);
+	}
+      });
+#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -182,12 +224,19 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
   }
   
   {
-    autoView( rhs_v, rhs, AcceleratorWrite);
     auto buffer_p = & buffer[0];
     auto table = &Cshift_table[0];
+#ifdef ACCELERATOR_CSHIFT    
+    autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
     });
+#else
+    autoView( rhs_v, rhs, CpuWrite);
+    thread_for(i,ent,{
+      rhs_v[table[i].first]=buffer_p[table[i].second];
+    });
+#endif
   }
 }
 
@@ -208,19 +257,30 @@ 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];
-    accelerator_for2d(n,e1,b,e2,1,{
+#ifdef ACCELERATOR_CSHIFT    
+    autoView( rhs_v , rhs, AcceleratorWrite);
+    accelerator_for(nn,e1*e2,1,{
+	int n = nn%e1;
+	int b = nn/e1;
 	int o      = n*_slice_stride;
 	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
+#else
+    autoView( rhs_v , rhs, CpuWrite);
+    thread_for2d(n,e1,b,e2,{
+	int o      = n*_slice_stride;
+	int offset = b+n*_slice_block;
+	merge(rhs_v[so+o+b],pointers,offset);
+    });
+#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
     // Test_cshift_red_black code.
-    //    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
+    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
     assert(0); // This will fail if hit on GPU
     autoView( rhs_v, rhs, CpuWrite);
@@ -280,12 +340,20 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
+    auto table = &Cshift_table[0];
+#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
-    auto table = &Cshift_table[0];
     accelerator_for(i,ent,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
   }
 }
 
@@ -324,12 +392,20 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
+    auto table = &Cshift_table[0];
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
-    auto table = &Cshift_table[0];
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
+#else
+    autoView( rhs_v, rhs, CpuRead);
+    autoView( lhs_v, lhs, CpuWrite);
+    thread_for(i,ent,{
+      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+    });
+#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

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

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_ET.h

@@ -342,19 +342,14 @@ 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));
@@ -456,20 +451,17 @@ GridTrinOpClass(TrinaryWhere,
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
-GRID_DEF_UNOP(adj, UnaryAdj);
+//GRID_DEF_UNOP(adj, UnaryAdj);
-GRID_DEF_UNOP(conjugate, UnaryConj);
+//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);
@@ -494,27 +486,27 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<decltype(expr.op.func(vecEval(0, expr.arg1)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
 {
-  Lattice<decltype(expr.op.func(vecEval(0, expr.arg1)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-  -> Lattice<decltype(expr.op.func(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 >
 {
-  Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
 				   vecEval(0, expr.arg2),
-				   vecEval(0, expr.arg3)))> 
+				   vecEval(0, expr.arg3)))>::type >
 {
-  Lattice<decltype(expr.op.func(vecEval(0, expr.arg1),
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
 				vecEval(0, expr.arg2),
-			        vecEval(0, expr.arg3)))>  ret(expr);
+			        vecEval(0, expr.arg3)))>::type >  ret(expr);
   return ret;
 }
 #define EXPRESSION_CLOSURE(function)					\

Grid/lattice/Lattice_basis.h

@@ -62,7 +62,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
     basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
-#if ( (!defined(GRID_SYCL)) && (!defined(GRID_CUDA)) && (!defined(GRID_HIP)) )
+#if ( (!defined(GRID_CUDA)) )
   int max_threads = thread_max();
   Vector < vobj > Bt(Nm * max_threads);
   thread_region
@@ -161,11 +161,13 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
   double * Qt_j = & Qt_jv[0];
   for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
+  auto basis_vp=& basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    auto B=coalescedRead(zz);
+    vobj zzz=Zero();
+    auto B=coalescedRead(zzz);
     for(int k=k0; k<k1; ++k){
-      B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
+      B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]);
     }
     coalescedWrite(result_v[ss], B);
   });

Grid/lattice/Lattice_reality.h

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

Grid/lattice/Lattice_reduction.h

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

Grid/lattice/Lattice_transfer.h

@@ -97,6 +97,20 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> &
   out = in;
 }
 
+template<typename T>
+accelerator_inline EnableIf<isGridFundamental<T>> convertType(T & out, const T & in) {
+  out = in;
+}
+
+// This would allow for conversions between GridFundamental types, but is not strictly needed as yet
+/*template<typename T1, typename T2>
+accelerator_inline typename std::enable_if<isGridFundamental<T1>::value && isGridFundamental<T2>::value>::type
+// Or to make this very broad, conversions between anything that's not a GridTensor could be allowed
+//accelerator_inline typename std::enable_if<!isGridTensor<T1>::value && !isGridTensor<T2>::value>::type
+convertType(T1 & out, const T2 & in) {
+  out = in;
+}*/
+
 #ifdef GRID_SIMT
 accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
   ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
@@ -117,18 +131,18 @@ accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
   Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
 }
 
-template<typename T1,typename T2,int N>
+template<typename T1,typename T2>
-  accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in);
+accelerator_inline void convertType(iScalar<T1> & out, const iScalar<T2> & in) {
-template<typename T1,typename T2,int N>
+  convertType(out._internal,in._internal);
-  accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in);
+}
 
-template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
+template<typename T1,typename T2>
-accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
+accelerator_inline NotEnableIf<isGridScalar<T1>> convertType(T1 & out, const iScalar<T2> & in) {
   convertType(out,in._internal);
 }
 
 template<typename T1,typename T2>
-accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
+accelerator_inline NotEnableIf<isGridScalar<T2>> convertType(iScalar<T1> & out, const T2 & in) {
   convertType(out._internal,in);
 }
 
@@ -145,11 +159,6 @@ accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & i
     convertType(out._internal[i],in._internal[i]);
 }
 
-template<typename T, typename std::enable_if<isGridFundamental<T>::value, T>::type* = nullptr>
-accelerator_inline void convertType(T & out, const T & in) {
-  out = in;
-}
-
 template<typename T1,typename T2>
 accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
   autoView( out_v , out,AcceleratorWrite);

Grid/lattice/Lattice_view.h

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

Grid/lattice/Lattice_where.h

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

Grid/parallelIO/BinaryIO.h

@@ -271,7 +271,7 @@ class BinaryIO {
 			      uint32_t &scidac_csumb)
   {
     grid->Barrier();
-    GridStopWatch timer; 
+    GridStopWatch timer, insideTimer; 
     GridStopWatch bstimer;
     
     nersc_csum=0;
@@ -363,7 +363,10 @@ class BinaryIO {
 	std::cout<< GridLogMessage<<"IOobject: MPI read I/O "<< file<< std::endl;
 	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);    assert(ierr==0);
 	ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);    assert(ierr==0);
-	ierr=MPI_File_read_all(fh, &iodata[0], 1, localArray, &status);    assert(ierr==0);
+  insideTimer.Start();
+	ierr=MPI_File_read_all(fh, &iodata[0], 1, localArray, &status);
+  insideTimer.Stop();
+  assert(ierr==0);
 	MPI_File_close(&fh);
 	MPI_Type_free(&fileArray);
 	MPI_Type_free(&localArray);
@@ -438,7 +441,9 @@ class BinaryIO {
         assert(ierr == 0);
 
         std::cout << GridLogDebug << "MPI write I/O write all " << file << std::endl;
+        insideTimer.Start();
         ierr = MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);
+        insideTimer.Stop();
         assert(ierr == 0);
 
         MPI_Offset os;
@@ -518,6 +523,11 @@ class BinaryIO {
     uint64_t bytes = sizeof(fobj)*iodata.size()*nrank;
     std::cout<< lastPerf.size <<"bytes in "<< timer.Elapsed() <<" "
 	     << lastPerf.mbytesPerSecond <<" MB/s "<<std::endl;
+    std::cout << GridLogMessage << "IOobject: pure MPI IO call " 
+              << lastPerf.size <<" bytes in " 
+              << insideTimer.Elapsed() << " "
+              << lastPerf.size/1024./1024./(insideTimer.useconds()/1.0e6)
+              <<" MB/s "<<std::endl;
 
     std::cout<<GridLogMessage<<"IOobject: endian and checksum overhead "<<bstimer.Elapsed()  <<std::endl;
 

Grid/parallelIO/IldgIO.h

@@ -619,12 +619,12 @@ class IldgWriter : public ScidacWriter {
   // Don't require scidac records EXCEPT checksum
   // Use Grid MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
+  template <class stats = PeriodicGaugeStatistics>
-  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
+  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef Lattice<vLorentzColourMatrixD> GaugeField;
-    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -636,6 +636,9 @@ class IldgWriter : public ScidacWriter {
 
     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
 
+    stats Stats;
+    Stats(Umu,header);
+    
     std::string format = header.floating_point;
     header.ensemble_id    = description;
     header.ensemble_label = description;
@@ -705,10 +708,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
+  template <class stats = PeriodicGaugeStatistics>
-  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
+  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
 
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef Lattice<vLorentzColourMatrixD > GaugeField;
     typedef typename GaugeField::vector_object  vobj;
     typedef typename vobj::scalar_object sobj;
 
@@ -921,7 +924,8 @@ class IldgReader : public GridLimeReader {
 
     if ( found_FieldMetaData || found_usqcdInfo ) {
       FieldMetaData checker;
-      GaugeStatistics(Umu,checker);
+      stats Stats;
+      Stats(Umu,checker);
       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;

Grid/parallelIO/MetaData.h

@@ -128,7 +128,7 @@ inline void MachineCharacteristics(FieldMetaData &header)
   std::time_t t = std::time(nullptr);
   std::tm tm_ = *std::localtime(&t);
   std::ostringstream oss; 
-  //      oss << std::put_time(&tm_, "%c %Z");
+  oss << std::put_time(&tm_, "%c %Z");
   header.creation_date = oss.str();
   header.archive_date  = header.creation_date;
 
@@ -176,29 +176,18 @@ template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMet
   GridMetaData(grid,header); 
   MachineCharacteristics(header);
 }
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
+template<class Impl>
+class GaugeStatistics
 {
-  // How to convert data precision etc...
+public:
-  header.link_trace=WilsonLoops<PeriodicGimplF>::linkTrace(data);
+  void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
-  header.plaquette =WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
-}
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
   {
-  // How to convert data precision etc...
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
-  header.link_trace=WilsonLoops<PeriodicGimplD>::linkTrace(data);
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(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);
   }
+};
+typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
+typedef GaugeStatistics<ConjugateGimplD> ConjugateGaugeStatistics;
 template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
 {
   GridBase *grid = field.Grid();
@@ -206,7 +195,6 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
   header.floating_point = format;
   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
   GridMetaData(grid,header); 
-  GaugeStatistics(field,header);
   MachineCharacteristics(header);
 }
 

Grid/parallelIO/NerscIO.h

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

Grid/parallelIO/OpenQcdIO.h

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

Grid/parallelIO/OpenQcdIOChromaReference.h

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

Grid/qcd/QCD.h

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

Grid/qcd/action/ActionBase.h

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

Grid/qcd/action/fermion/Fermion.h

@@ -291,12 +291,6 @@ typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
 typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
 typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
 
-#ifndef GRID_CUDA
-typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR;
-typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF;
-typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD;
-#endif
-
 NAMESPACE_END(Grid);
 
 ////////////////////

Grid/qcd/action/fermion/FermionOperatorImpl.h

@@ -153,8 +153,8 @@ public:
   typedef typename Impl::StencilImpl             StencilImpl;		\
   typedef typename Impl::ImplParams               ImplParams;	        \
   typedef typename Impl::StencilImpl::View_type  StencilView;		\
-  typedef typename ViewMap<FermionField>::Type      FermionFieldView;	\
+  typedef const typename ViewMap<FermionField>::Type      FermionFieldView;	\
-  typedef typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
+  typedef const typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
 
 #define INHERIT_IMPL_TYPES(Base)		\
   INHERIT_GIMPL_TYPES(Base)			\
@@ -183,7 +183,8 @@ NAMESPACE_CHECK(ImplStaggered);
 /////////////////////////////////////////////////////////////////////////////
 // Single flavour one component spinors with colour index. 5d vec
 /////////////////////////////////////////////////////////////////////////////
-#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
+// Deprecate Vec5d
-NAMESPACE_CHECK(ImplStaggered5dVec);  
+//#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
+//NAMESPACE_CHECK(ImplStaggered5dVec);  
 
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

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

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

Grid/qcd/action/fermion/StaggeredImpl.h

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

Grid/qcd/action/fermion/WilsonCloverFermion.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -61,7 +61,7 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) {
+  accelerator_inline int CommDatumSize(void) const {
     return sizeof(SiteHalfCommSpinor);
   }
 
@@ -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) {
+  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
     _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){
+				   Integer type,Integer o) const {
     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) {    
+				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
     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)
+					   Integer j,Integer k, Integer m,Integer type) const
   {
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
@@ -117,7 +117,7 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) { return false; }
+  accelerator_inline bool DecompressionStep(void) const { 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) {
+  accelerator_inline int CommDatumSize(void) const {
     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) {
+  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
     _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){
+		       Integer type,Integer o) const {
     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){
+  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
     SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
     precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
   }
@@ -186,7 +186,7 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
 			       SiteHalfSpinor *out1,
 			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type){
+			       Integer j,Integer k, Integer m,Integer type) const {
     SiteHalfSpinor temp1, temp2,temp3,temp4;
     SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
     SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
@@ -200,7 +200,7 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) { return true; }
+  accelerator_inline bool DecompressionStep(void) const { return true; }
 
 };
 

Grid/qcd/action/fermion/WilsonImpl.h

@@ -72,7 +72,7 @@ public:
   typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
   typedef WilsonImplParams ImplParams;
   typedef WilsonStencil<SiteSpinor, SiteHalfSpinor,ImplParams> StencilImpl;
-  typedef typename StencilImpl::View_type StencilView;
+  typedef const typename StencilImpl::View_type StencilView;
     
   ImplParams Params;
 
@@ -106,11 +106,15 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
+    const int Nsimd = SiteHalfSpinor::Nsimd();
     autoView( out_v, out, AcceleratorWrite);
     autoView( phi_v, phi, AcceleratorRead);
     autoView( Umu_v, Umu, AcceleratorRead);
-    accelerator_for(sss,out.Grid()->oSites(),1,{
+    typedef decltype(coalescedRead(out_v[0]))   calcSpinor;
-	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
+    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
+	calcSpinor tmp;
+	multLink(tmp,Umu_v[sss],phi_v(sss),mu);
+	coalescedWrite(out_v[sss],tmp);
     });
   }
 					   
@@ -180,18 +184,22 @@ public:
       mat = TraceIndex<SpinIndex>(P); 
     }
       
-    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
+    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds)
+    {
       for (int mu = 0; mu < Nd; mu++)
       mat[mu] = PeekIndex<LorentzIndex>(Uds, mu);
     }
 
-
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu)
-  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
+  {
-      
+#undef USE_OLD_INSERT_FORCE    
     int Ls=Btilde.Grid()->_fdimensions[0];
+    autoView( mat_v , mat, AcceleratorWrite);
+#ifdef USE_OLD_INSERT_FORCE    
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
     {
+      const int Nsimd = SiteSpinor::Nsimd();
       autoView( tmp_v , tmp, AcceleratorWrite);
       autoView( Btilde_v , Btilde, AcceleratorRead);
       autoView( Atilde_v , Atilde, AcceleratorRead);
@@ -204,6 +212,29 @@ public:
 	});
     }
     PokeIndex<LorentzIndex>(mat,tmp,mu);
+#else
+    {
+      const int Nsimd = SiteSpinor::Nsimd();
+      autoView( Btilde_v , Btilde, AcceleratorRead);
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,mat.Grid()->oSites(),Nsimd,{
+	  int sU=sss;
+  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
+  	  ColorMatrixType sum;
+	  zeroit(sum);  
+	  for(int s=0;s<Ls;s++){
+	    int sF = s+Ls*sU;
+  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
+  	      auto bb = coalescedRead(Btilde_v[sF]()(spn) ); //color vector
+  	      auto aa = coalescedRead(Atilde_v[sF]()(spn) );
+	      auto op = outerProduct(bb,aa);
+  	      sum = sum + op;
+	    }
+	  }
+  	  coalescedWrite(mat_v[sU](mu)(), sum);
+      });
+    }
+#endif    
   }
 };
 

Grid/qcd/action/fermion/WilsonKernels.h

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

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

@@ -642,7 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -826,7 +826,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
@@ -880,11 +880,23 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
+  Integer sign = 1; // sign flip for vector/tadpole
   if ( curr_type == Current::Axial ) {
     for(int s=0;s<Ls/2;s++){
       G_s[s] = -1.0;
     }
   }
+  else if ( curr_type == Current::Tadpole ) {
+    auto b=this->_b;
+    auto c=this->_c;
+    if ( b == 1 && c == 0 ) {
+      sign = -1;    
+    }
+    else {
+      std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
+      assert(b==1 && c==0);
+    }
+  }
 
   for(int s=0;s<Ls;s++){
 
@@ -907,7 +919,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     tmp    = Cshift(tmp,mu,1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu);
-    tmp    = G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
+    tmp    = sign*G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
     tmp    = where((lcoor>=tmin),tmp,zz); // Mask the time 
     L_Q    = where((lcoor<=tmax),tmp,zz); // Position of current complicated
 

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

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

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

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

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

@@ -35,39 +35,32 @@ NAMESPACE_BEGIN(Grid);
 #define GENERIC_STENCIL_LEG(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
+    int perm= SE->_permute;						\
-      chi_p = χ						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
-      permute(chi,  in[SE->_offset], ptype);			\
   } else {							\
-      chi_p = &in[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  } else {							\
+  acceleratorSynchronise();					\
-    chi_p = &buf[SE->_offset];					\
+  multLink(Uchi, U[sU], chi, Dir);			
-  }								\
-  multLink(Uchi, U[sU], *chi_p, Dir);			
 
 #define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
+    int perm= SE->_permute;						\
-      chi_p = χ						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
-      permute(chi,  in[SE->_offset], ptype);			\
-    } else {							\
-      chi_p = &in[SE->_offset];					\
-    }								\
   } else if ( st.same_node[Dir] ) {				\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);                 \
   }								\
   if (SE->_is_local || st.same_node[Dir] ) {			\
-    multLink(Uchi, U[sU], *chi_p, Dir);				\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 #define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
-    multLink(Uchi, U[sU], *chi_p, Dir);				\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 template <class Impl>
@@ -84,12 +77,14 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
 					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
-  const SiteSpinor *chi_p;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
-  SiteSpinor chi;
+  calcSpinor chi;
-  SiteSpinor Uchi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //
@@ -118,7 +113,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
     if ( dag ) { 
       Uchi = - Uchi;
     } 
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -130,13 +125,16 @@ template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 						SiteSpinor *buf, int sF, int sU, 
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
+						const FermionFieldView &in, FermionFieldView &out,int dag)
-  const SiteSpinor *chi_p;
+{
-  SiteSpinor chi;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
-  SiteSpinor Uchi;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //    int sF=LLs*sU+s;
@@ -165,7 +163,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
     if ( dag ) {
       Uchi = - Uchi;
     }
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -178,14 +176,17 @@ template <int Naik> accelerator_inline
 void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 						SiteSpinor *buf, int sF, int sU,
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
+						const FermionFieldView &in, FermionFieldView &out,int dag)
-  const SiteSpinor *chi_p;
+{
-  //  SiteSpinor chi;
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
-  SiteSpinor Uchi;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int nmu=0;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   //  for(int s=0;s<LLs;s++){
   //    int sF=LLs*sU+s;
@@ -212,10 +213,11 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
     GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
     }
     if ( nmu ) {
+      auto _out = coalescedRead(out[sF],lane);
       if ( dag ) {
-	out[sF] = out[sF] - Uchi;
+	coalescedWrite(out[sF], _out-Uchi,lane);
       } else { 
-	out[sF] = out[sF] + Uchi;
+	coalescedWrite(out[sF], _out+Uchi,lane);
       }
     }
   }
@@ -261,6 +263,8 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo,
   GridBase *FGrid=in.Grid();  
   GridBase *UGrid=U.Grid();  
   typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
   autoView( UUU_v , UUU, AcceleratorRead);
   autoView( U_v   ,   U, AcceleratorRead);
   autoView( in_v  ,  in, AcceleratorRead);
@@ -301,6 +305,8 @@ void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo,
   GridBase *FGrid=in.Grid();  
   GridBase *UGrid=U.Grid();  
   typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
   autoView( UUU_v ,   U, AcceleratorRead);
   autoView( U_v   ,   U, AcceleratorRead);
   autoView( in_v  ,  in, AcceleratorRead);

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

@@ -92,20 +92,16 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   int lvol = _Umu.Grid()->lSites();
   int DimRep = Impl::Dimension;
 
-  Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-  Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-
-  Coordinate lcoor;
-  typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
-
   {
     autoView(CTv,CloverTerm,CpuRead);
     autoView(CTIv,CloverTermInv,CpuWrite);
-    for (int site = 0; site < lvol; site++) {
+    thread_for(site, lvol, {
+      Coordinate lcoor;
       grid->LocalIndexToLocalCoor(site, lcoor);
-      EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
       peekLocalSite(Qx, CTv, lcoor);
-      Qxinv = Zero();
       //if (csw!=0){
       for (int j = 0; j < Ns; j++)
 	for (int k = 0; k < Ns; k++)
@@ -126,21 +122,21 @@ 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
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
 
-  pickCheckerboard(Even, CloverTermDagEven, closure(adj(CloverTerm)));
+  pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
-  pickCheckerboard(Odd, CloverTermDagOdd, closure(adj(CloverTerm)));
+  pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
 
   pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
   pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
 
-  pickCheckerboard(Even, CloverTermInvDagEven, closure(adj(CloverTermInv)));
+  pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
-  pickCheckerboard(Odd, CloverTermInvDagOdd, closure(adj(CloverTermInv)));
+  pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
 }
 
 template <class Impl>

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

@@ -397,6 +397,7 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
 template <class Impl>
 void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -408,6 +409,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 template <class Impl>
 void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -420,6 +422,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
 template <class Impl>
 void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 

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

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

@@ -25,6 +25,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
+
+// GCC 10 messes up SVE instruction scheduling using -O3, but
+// -O3 -fno-schedule-insns -fno-schedule-insns2 does wonders
+// performance now is better than armclang 20.2
+
 #ifdef KERNEL_DAG
 #define DIR0_PROJ    XP_PROJ
 #define DIR1_PROJ    YP_PROJ
@@ -110,6 +115,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     }                                                       \
     RECON;								                    \
 
+/*
+NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
+    though I expected that it would improve on performance
+*/
+
 #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
   base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
   PREFETCH1_CHIMU(base);						            \
@@ -132,60 +142,48 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
   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;								\
+      }									\
+  base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
+  PREFETCH_CHIMU(base);						\
   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);							\
     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;								\
     nmu++;								\
   }
 
 #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
   nmu=0;								\
+  { ZERO_PSI;}								\
   base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
   if((!local)&&(!st.same_node[Dir]) ) {					\
     LOAD_CHI(base);							\
     MULT_2SPIN_1(Dir);					                    \
-    PREFETCH_CHIMU(base);                                   \
-    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    if (s == 0) {                                           \
-      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
-    }                                                       \
     RECON;								\
     nmu++;								\
   }
@@ -193,6 +191,8 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 #define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
 
 #endif
+
+
 {
   int nmu;
   int local,perm, ptype;
@@ -209,7 +209,6 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
     //    int sUn=lo.Reorder(ssn);
     int sUn=ssn;
-    LOCK_GAUGE(0);
 #else
     int sU =ssU;
     int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
@@ -295,6 +294,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       std::cout << "----------------------------------------------------" << std::endl;
 #endif
 
+      // DC ZVA test
+      // { uint64_t basestore = (uint64_t)&out[ss];
+      //   PREFETCH_RESULT_L2_STORE(basestore); }
+
+
       ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJ,DIR5_RECON);
 
 #ifdef SHOW
@@ -308,6 +312,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       std::cout << "----------------------------------------------------" << std::endl;
 #endif
 
+      // DC ZVA test
+      //{ uint64_t basestore = (uint64_t)&out[ss];
+      //  PREFETCH_RESULT_L2_STORE(basestore); }
+
+
       ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJ,DIR6_RECON);
 
 #ifdef SHOW
@@ -321,6 +330,11 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       std::cout << "----------------------------------------------------" << std::endl;
 #endif
 
+      // DC ZVA test
+      //{ uint64_t basestore = (uint64_t)&out[ss];
+      //  PREFETCH_RESULT_L2_STORE(basestore); }
+
+
       ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJ,DIR7_RECON);
 
 #ifdef SHOW
@@ -341,6 +355,7 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
       base = (uint64_t) &out[ss];
       basep= st.GetPFInfo(nent,plocal); ent++;
       basep = (uint64_t) &out[ssn];
+      //PREFETCH_RESULT_L1_STORE(base);
       RESULT(base,basep);
 
 #ifdef SHOW

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

@@ -76,7 +76,24 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 #define REGISTER
 
-#define LOAD_CHIMU \
+#ifdef GRID_SIMT
+#define LOAD_CHIMU(ptype)		\
+  {const SiteSpinor & ref (in[offset]);	\
+    Chimu_00=coalescedReadPermute<ptype>(ref()(0)(0),perm,lane);	\
+    Chimu_01=coalescedReadPermute<ptype>(ref()(0)(1),perm,lane);		\
+    Chimu_02=coalescedReadPermute<ptype>(ref()(0)(2),perm,lane);		\
+    Chimu_10=coalescedReadPermute<ptype>(ref()(1)(0),perm,lane);		\
+    Chimu_11=coalescedReadPermute<ptype>(ref()(1)(1),perm,lane);		\
+    Chimu_12=coalescedReadPermute<ptype>(ref()(1)(2),perm,lane);		\
+    Chimu_20=coalescedReadPermute<ptype>(ref()(2)(0),perm,lane);		\
+    Chimu_21=coalescedReadPermute<ptype>(ref()(2)(1),perm,lane);		\
+    Chimu_22=coalescedReadPermute<ptype>(ref()(2)(2),perm,lane);		\
+    Chimu_30=coalescedReadPermute<ptype>(ref()(3)(0),perm,lane);		\
+    Chimu_31=coalescedReadPermute<ptype>(ref()(3)(1),perm,lane);		\
+    Chimu_32=coalescedReadPermute<ptype>(ref()(3)(2),perm,lane);	}
+#define PERMUTE_DIR(dir) ;
+#else
+#define LOAD_CHIMU(ptype)		\
   {const SiteSpinor & ref (in[offset]);	\
     Chimu_00=ref()(0)(0);\
     Chimu_01=ref()(0)(1);\
@@ -91,24 +108,24 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     Chimu_31=ref()(3)(1);\
     Chimu_32=ref()(3)(2);}
 
-#define LOAD_CHI\
+#define PERMUTE_DIR(dir)			\
-  {const SiteHalfSpinor &ref(buf[offset]);	\
+  permute##dir(Chi_00,Chi_00);	\
-    Chi_00 = ref()(0)(0);\
+      permute##dir(Chi_01,Chi_01);\
-    Chi_01 = ref()(0)(1);\
+      permute##dir(Chi_02,Chi_02);\
-    Chi_02 = ref()(0)(2);\
+      permute##dir(Chi_10,Chi_10);	\
-    Chi_10 = ref()(1)(0);\
+      permute##dir(Chi_11,Chi_11);\
-    Chi_11 = ref()(1)(1);\
+      permute##dir(Chi_12,Chi_12);
-    Chi_12 = ref()(1)(2);}
+
+#endif
 
-// To splat or not to splat depends on the implementation
 #define MULT_2SPIN(A)\
   {auto & ref(U[sU](A));						\
-   Impl::loadLinkElement(U_00,ref()(0,0));	\
+    U_00=coalescedRead(ref()(0,0),lane);				\
-   Impl::loadLinkElement(U_10,ref()(1,0));	\
+    U_10=coalescedRead(ref()(1,0),lane);				\
-   Impl::loadLinkElement(U_20,ref()(2,0));	\
+    U_20=coalescedRead(ref()(2,0),lane);				\
-   Impl::loadLinkElement(U_01,ref()(0,1));	\
+    U_01=coalescedRead(ref()(0,1),lane);				\
-   Impl::loadLinkElement(U_11,ref()(1,1));	\
+    U_11=coalescedRead(ref()(1,1),lane);				\
-   Impl::loadLinkElement(U_21,ref()(2,1));	\
+    U_21=coalescedRead(ref()(2,1),lane);				\
     UChi_00 = U_00*Chi_00;						\
     UChi_10 = U_00*Chi_10;						\
     UChi_01 = U_10*Chi_00;						\
@@ -121,9 +138,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     UChi_11+= U_11*Chi_11;						\
     UChi_02+= U_21*Chi_01;						\
     UChi_12+= U_21*Chi_11;						\
-    Impl::loadLinkElement(U_00,ref()(0,2));	\
+    U_00=coalescedRead(ref()(0,2),lane);				\
-    Impl::loadLinkElement(U_10,ref()(1,2));	\
+    U_10=coalescedRead(ref()(1,2),lane);				\
-    Impl::loadLinkElement(U_20,ref()(2,2));	\
+    U_20=coalescedRead(ref()(2,2),lane);				\
     UChi_00+= U_00*Chi_02;						\
     UChi_10+= U_00*Chi_12;						\
     UChi_01+= U_10*Chi_02;						\
@@ -131,14 +148,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     UChi_02+= U_20*Chi_02;						\
     UChi_12+= U_20*Chi_12;}
 
-
+#define LOAD_CHI				\
-#define PERMUTE_DIR(dir)			\
+  {const SiteHalfSpinor &ref(buf[offset]);	\
-      permute##dir(Chi_00,Chi_00);\
+    Chi_00 = coalescedRead(ref()(0)(0),lane);	\
-      permute##dir(Chi_01,Chi_01);\
+    Chi_01 = coalescedRead(ref()(0)(1),lane);	\
-      permute##dir(Chi_02,Chi_02);\
+    Chi_02 = coalescedRead(ref()(0)(2),lane);	\
-      permute##dir(Chi_10,Chi_10);\
+    Chi_10 = coalescedRead(ref()(1)(0),lane);	\
-      permute##dir(Chi_11,Chi_11);\
+    Chi_11 = coalescedRead(ref()(1)(1),lane);	\
-      permute##dir(Chi_12,Chi_12);
+    Chi_12 = coalescedRead(ref()(1)(2),lane);}
 
 //      hspin(0)=fspin(0)+timesI(fspin(3));
 //      hspin(1)=fspin(1)+timesI(fspin(2));
@@ -353,13 +370,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   result_31-= UChi_11;	\
   result_32-= UChi_12;
 
-#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
+#define HAND_STENCIL_LEGB(PROJ,PERM,DIR,RECON)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
   offset = SE->_offset;				\
   local  = SE->_is_local;			\
   perm   = SE->_permute;			\
   if ( local ) {				\
-    LOAD_CHIMU;					\
+    LOAD_CHIMU(PERM);				\
     PROJ;					\
     if ( perm) {				\
       PERMUTE_DIR(PERM);			\
@@ -367,6 +384,37 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   } else {					\
     LOAD_CHI;					\
   }						\
+  acceleratorSynchronise();			\
+  MULT_2SPIN(DIR);				\
+  RECON;					
+
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
+  SE=&st_p[DIR+8*ss];				\
+  ptype=st_perm[DIR];				\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU(PERM);				\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else {					\
+    LOAD_CHI;					\
+  }						\
+  acceleratorSynchronise();			\
+  MULT_2SPIN(DIR);				\
+  RECON;					
+
+#define HAND_STENCIL_LEGA(PROJ,PERM,DIR,RECON)				\
+  SE=&st_p[DIR+8*ss];							\
+  ptype=st_perm[DIR];							\
+ /*SE=st.GetEntry(ptype,DIR,ss);*/					\
+  offset = SE->_offset;				\
+  perm   = SE->_permute;			\
+  LOAD_CHIMU(PERM);				\
+  PROJ;						\
   MULT_2SPIN(DIR);				\
   RECON;					
 
@@ -376,7 +424,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   local  = SE->_is_local;			\
   perm   = SE->_permute;			\
   if ( local ) {				\
-    LOAD_CHIMU;					\
+    LOAD_CHIMU(PERM);				\
     PROJ;					\
     if ( perm) {				\
       PERMUTE_DIR(PERM);			\
@@ -384,10 +432,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   } else if ( st.same_node[DIR] ) {		\
     LOAD_CHI;					\
   }						\
+  acceleratorSynchronise();			\
   if (local || st.same_node[DIR] ) {		\
     MULT_2SPIN(DIR);				\
     RECON;					\
-  }
+  }						\
+  acceleratorSynchronise();			
 
 #define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
@@ -397,44 +447,44 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     MULT_2SPIN(DIR);				\
     RECON;					\
     nmu++;					\
-  }
+  }						\
+  acceleratorSynchronise();			
 
 #define HAND_RESULT(ss)				\
   {						\
     SiteSpinor & ref (out[ss]);			\
-    vstream(ref()(0)(0),result_00);		\
+    coalescedWrite(ref()(0)(0),result_00,lane);		\
-    vstream(ref()(0)(1),result_01);		\
+    coalescedWrite(ref()(0)(1),result_01,lane);		\
-    vstream(ref()(0)(2),result_02);		\
+    coalescedWrite(ref()(0)(2),result_02,lane);		\
-    vstream(ref()(1)(0),result_10);		\
+    coalescedWrite(ref()(1)(0),result_10,lane);		\
-    vstream(ref()(1)(1),result_11);		\
+    coalescedWrite(ref()(1)(1),result_11,lane);		\
-    vstream(ref()(1)(2),result_12);		\
+    coalescedWrite(ref()(1)(2),result_12,lane);		\
-    vstream(ref()(2)(0),result_20);		\
+    coalescedWrite(ref()(2)(0),result_20,lane);		\
-    vstream(ref()(2)(1),result_21);		\
+    coalescedWrite(ref()(2)(1),result_21,lane);		\
-    vstream(ref()(2)(2),result_22);		\
+    coalescedWrite(ref()(2)(2),result_22,lane);		\
-    vstream(ref()(3)(0),result_30);		\
+    coalescedWrite(ref()(3)(0),result_30,lane);		\
-    vstream(ref()(3)(1),result_31);		\
+    coalescedWrite(ref()(3)(1),result_31,lane);		\
-    vstream(ref()(3)(2),result_32);		\
+    coalescedWrite(ref()(3)(2),result_32,lane);		\
   }
 
 #define HAND_RESULT_EXT(ss)				\
-  if (nmu){					\
+  {							\
     SiteSpinor & ref (out[ss]);				\
-    ref()(0)(0)+=result_00;		\
+    coalescedWrite(ref()(0)(0),coalescedRead(ref()(0)(0))+result_00,lane);	\
-    ref()(0)(1)+=result_01;		\
+    coalescedWrite(ref()(0)(1),coalescedRead(ref()(0)(1))+result_01,lane);	\
-    ref()(0)(2)+=result_02;		\
+    coalescedWrite(ref()(0)(2),coalescedRead(ref()(0)(2))+result_02,lane);	\
-    ref()(1)(0)+=result_10;		\
+    coalescedWrite(ref()(1)(0),coalescedRead(ref()(1)(0))+result_10,lane);	\
-    ref()(1)(1)+=result_11;		\
+    coalescedWrite(ref()(1)(1),coalescedRead(ref()(1)(1))+result_11,lane);	\
-    ref()(1)(2)+=result_12;		\
+    coalescedWrite(ref()(1)(2),coalescedRead(ref()(1)(2))+result_12,lane);	\
-    ref()(2)(0)+=result_20;		\
+    coalescedWrite(ref()(2)(0),coalescedRead(ref()(2)(0))+result_20,lane);	\
-    ref()(2)(1)+=result_21;		\
+    coalescedWrite(ref()(2)(1),coalescedRead(ref()(2)(1))+result_21,lane);	\
-    ref()(2)(2)+=result_22;		\
+    coalescedWrite(ref()(2)(2),coalescedRead(ref()(2)(2))+result_22,lane);	\
-    ref()(3)(0)+=result_30;		\
+    coalescedWrite(ref()(3)(0),coalescedRead(ref()(3)(0))+result_30,lane);	\
-    ref()(3)(1)+=result_31;		\
+    coalescedWrite(ref()(3)(1),coalescedRead(ref()(3)(1))+result_31,lane);	\
-    ref()(3)(2)+=result_32;		\
+    coalescedWrite(ref()(3)(2),coalescedRead(ref()(3)(2))+result_32,lane);	\
   }
 
-
+#define HAND_DECLARATIONS(Simd)			\
-#define HAND_DECLARATIONS(a)			\
   Simd result_00;				\
   Simd result_01;				\
   Simd result_02;				\
@@ -467,18 +517,18 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   Simd U_21;
 
 #define ZERO_RESULT							\
-  result_00=Zero();				\
+  zeroit(result_00);							\
-  result_01=Zero();				\
+  zeroit(result_01);							\
-  result_02=Zero();				\
+  zeroit(result_02);							\
-  result_10=Zero();				\
+  zeroit(result_10);							\
-  result_11=Zero();				\
+  zeroit(result_11);							\
-  result_12=Zero();				\
+  zeroit(result_12);							\
-  result_20=Zero();				\
+  zeroit(result_20);							\
-  result_21=Zero();				\
+  zeroit(result_21);							\
-  result_22=Zero();				\
+  zeroit(result_22);							\
-  result_30=Zero();				\
+  zeroit(result_30);							\
-  result_31=Zero();				\
+  zeroit(result_31);							\
-  result_32=Zero();			
+  zeroit(result_32);			
 
 #define Chimu_00 Chi_00
 #define Chimu_01 Chi_01
@@ -495,15 +545,53 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+
+#ifdef SYCL_HACK
 template<class Impl> accelerator_inline void 
-WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
-				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+				      int ss,int sU,const SiteSpinor *in, SiteSpinor *out)
 {
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef iSinglet<Simd> vCplx;
+  //  typedef decltype( coalescedRead( vCplx()()() )) Simt;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+  HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
+  HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+#endif
+
+template<class Impl> accelerator_inline void 
+WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  auto st_p = st._entries_p;						
+  auto st_perm = st._permute_type;					
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
+
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
   int offset,local,perm, ptype;
   StencilEntry *SE;
@@ -523,10 +611,16 @@ template<class Impl>  accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
+  auto st_p = st._entries_p;						
+  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
   int offset,local,perm, ptype;
@@ -546,11 +640,17 @@ template<class Impl>  accelerator_inline void
 WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
+  auto st_p = st._entries_p;						
+  auto st_perm = st._permute_type;					
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
   int offset,local,perm, ptype;
   StencilEntry *SE;
@@ -570,10 +670,16 @@ template<class Impl> accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
+  auto st_p = st._entries_p;						
+  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
   int offset,local,perm, ptype;
@@ -593,11 +699,17 @@ template<class Impl>  accelerator_inline void
 WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
+  auto st_p = st._entries_p;						
+  auto st_perm = st._permute_type;					
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
   int offset, ptype;
   StencilEntry *SE;
@@ -618,10 +730,16 @@ template<class Impl>  accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
+  auto st_p = st._entries_p;						
+  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
+  typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
 
-  HAND_DECLARATIONS(ignore);
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+
+  HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
   int offset, ptype;
@@ -682,3 +800,4 @@ NAMESPACE_END(Grid);
 #undef HAND_RESULT
 #undef HAND_RESULT_INT
 #undef HAND_RESULT_EXT
+#undef HAND_DECLARATIONS

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

@@ -416,6 +416,20 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
+#define KERNEL_CALL_TMP(A) \
+  const uint64_t    NN = Nsite*Ls;					\
+  auto U_p = & U_v[0];							\
+  auto in_p = & in_v[0];						\
+  auto out_p = & out_v[0];						\
+  auto st_p = st_v._entries_p;						\
+  auto st_perm = st_v._permute_type;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_perm,st_p,U_p,buf,sF,sU,in_p,out_p);	\
+    });									\
+  accelerator_barrier();
+
 #define KERNEL_CALLNB(A)						\
   const uint64_t    NN = Nsite*Ls;					\
   accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
@@ -445,20 +459,24 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifndef GRID_CUDA
+#ifdef SYCL_HACK     
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
+#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
+#endif     
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteInt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteInt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
    }
@@ -476,20 +494,20 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDag); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDag);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
    }

Grid/qcd/action/gauge/Gauge.cc

@@ -0,0 +1,38 @@
+/*************************************************************************************
+
+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

@@ -96,7 +96,7 @@ public:
   ///////////////////////////////////////////////////////////
   // Move these to another class
   // HMC auxiliary functions
-  static inline void generate_momenta(Field &P, GridParallelRNG &pRNG) 
+  static inline void generate_momenta(Field &P, GridSerialRNG & sRNG, GridParallelRNG &pRNG) 
   {
     // Zbigniew Srocinsky thesis:
     //
@@ -154,6 +154,10 @@ 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 Cshift(adj(Link), mu, -1);
+    return PeriodicBC::CovShiftIdentityBackward(Link, mu);
   }
   static inline GaugeLinkField
   CovShiftIdentityForward(const GaugeLinkField &Link, int mu) {
-    return Link;
+    return PeriodicBC::CovShiftIdentityForward(Link,mu);
   }
   static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu) {
-    return Cshift(Link, mu, 1);
+    return PeriodicBC::ShiftStaple(Link,mu);
   }
 
   static inline bool isPeriodicGaugeField(void) { return true; }
@@ -74,7 +74,13 @@ public:
 
 // Composition with smeared link, bc's etc.. probably need multiple inheritance
 // Variable precision "S" and variable Nc
-template <class GimplTypes> class ConjugateGaugeImpl : public GimplTypes {
+class ConjugateGaugeImplBase {
+protected:
+  static std::vector<int> _conjDirs;
+};
+
+  template <class GimplTypes> class ConjugateGaugeImpl : public GimplTypes, ConjugateGaugeImplBase {
+private:
 public:
   INHERIT_GIMPL_TYPES(GimplTypes);
 
@@ -84,47 +90,56 @@ public:
   ////////////////////////////////////////////////////////////////////////////////////////////////////////////
   template <class covariant>
   static Lattice<covariant> CovShiftForward(const GaugeLinkField &Link, int mu,
-                                            const Lattice<covariant> &field) {
+                                            const Lattice<covariant> &field)
+  {
+    assert(_conjDirs.size() == Nd);
+    if(_conjDirs[mu]) 
       return ConjugateBC::CovShiftForward(Link, mu, field);
+    else
+      return PeriodicBC::CovShiftForward(Link, mu, field);
   }
 
   template <class covariant>
   static Lattice<covariant> CovShiftBackward(const GaugeLinkField &Link, int mu,
-                                             const Lattice<covariant> &field) {
+                                             const Lattice<covariant> &field)
+  {
+    assert(_conjDirs.size() == Nd);
+    if(_conjDirs[mu]) 
       return ConjugateBC::CovShiftBackward(Link, mu, field);
+    else 
+      return PeriodicBC::CovShiftBackward(Link, mu, field);
   }
 
   static inline GaugeLinkField
-  CovShiftIdentityBackward(const GaugeLinkField &Link, int mu) {
+  CovShiftIdentityBackward(const GaugeLinkField &Link, int mu)
-    GridBase *grid = Link.Grid();
+  {
-    int Lmu = grid->GlobalDimensions()[mu] - 1;
+    assert(_conjDirs.size() == Nd);
-
+    if(_conjDirs[mu]) 
-    Lattice<iScalar<vInteger>> coor(grid);
+      return ConjugateBC::CovShiftIdentityBackward(Link, mu);
-    LatticeCoordinate(coor, mu);
+    else 
-
+      return PeriodicBC::CovShiftIdentityBackward(Link, 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) {
+  CovShiftIdentityForward(const GaugeLinkField &Link, int mu)
-    return Link;
+  {
+    assert(_conjDirs.size() == Nd);
+    if(_conjDirs[mu]) 
+      return ConjugateBC::CovShiftIdentityForward(Link,mu);
+    else
+      return PeriodicBC::CovShiftIdentityForward(Link,mu);
   }
 
-  static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu) {
+  static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu)
-    GridBase *grid = Link.Grid();
+  {
-    int Lmu = grid->GlobalDimensions()[mu] - 1;
+    assert(_conjDirs.size() == Nd);
-
+    if(_conjDirs[mu]) 
-    Lattice<iScalar<vInteger>> coor(grid);
+      return ConjugateBC::ShiftStaple(Link,mu);
-    LatticeCoordinate(coor, mu);
+    else     
-
+      return PeriodicBC::ShiftStaple(Link,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; }
 };
 

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

@@ -49,7 +49,7 @@ public:
 
   virtual std::string action_name(){return "PlaqPlusRectangleAction";}
       
-  virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {}; // noop as no pseudoferms
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {}; // noop as no pseudoferms
       
   virtual std::string LogParameters(){
     std::stringstream sstream;

Grid/qcd/action/gauge/WilsonGaugeAction.h

@@ -54,8 +54,7 @@ public:
     return sstream.str();
   }
 
-  virtual void refresh(const GaugeField &U,
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG){};  // noop as no pseudoferms
-                       GridParallelRNG &pRNG){};  // noop as no pseudoferms
 
   virtual RealD S(const GaugeField &U) {
     RealD plaq = WilsonLoops<Gimpl>::avgPlaquette(U);

Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h

@@ -124,7 +124,7 @@ NAMESPACE_BEGIN(Grid);
       //
       // As a check of rational require \Phi^dag M_{EOFA} \Phi == eta^dag M^-1/2^dag M M^-1/2 eta = eta^dag eta
       //
-      virtual void refresh(const GaugeField& U, GridParallelRNG& pRNG)
+      virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
       {
         Lop.ImportGauge(U);
         Rop.ImportGauge(U);

Grid/qcd/action/pseudofermion/OneFlavourEvenOddRational.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
 Grid physics library, www.github.com/paboyle/Grid
@@ -43,8 +42,7 @@ NAMESPACE_BEGIN(Grid);
 //
 
 template <class Impl>
-class OneFlavourEvenOddRationalPseudoFermionAction
+class OneFlavourEvenOddRationalPseudoFermionAction : public Action<typename Impl::GaugeField> {
-  : public Action<typename Impl::GaugeField> {
 public:
   INHERIT_IMPL_TYPES(Impl);
 
@@ -103,7 +101,7 @@ public:
     return sstream.str();
   }
   
-  virtual void refresh(const GaugeField &U, GridParallelRNG &pRNG) {
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
     // P(phi) = e^{- phi^dag (MpcdagMpc)^-1/2 phi}
     //        = e^{- phi^dag (MpcdagMpc)^-1/4 (MpcdagMpc)^-1/4 phi}
     // Phi = MpcdagMpc^{1/4} eta
@@ -156,7 +154,10 @@ public:
 
     msCG(Mpc, PhiOdd, Y);
 
-    if ( (rand()%param.BoundsCheckFreq)==0 ) { 
+    auto grid = FermOp.FermionGrid();
+    auto r=rand();
+    grid->Broadcast(0,r);
+    if ( (r%param.BoundsCheckFreq)==0 ) { 
       FermionField gauss(FermOp.FermionRedBlackGrid());
       gauss = PhiOdd;
       HighBoundCheck(Mpc,gauss,param.hi);

Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h

@@ -101,7 +101,7 @@ NAMESPACE_BEGIN(Grid);
       }
       
       
-      virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {
+      virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
 
 	// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi       
 	//
@@ -170,7 +170,10 @@ NAMESPACE_BEGIN(Grid);
 	msCG_M(MdagM,X,Y);
 
 	// Randomly apply rational bounds checks.
-	if ( (rand()%param.BoundsCheckFreq)==0 ) { 
+	auto grid = NumOp.FermionGrid();
+        auto r=rand();
+        grid->Broadcast(0,r);
+        if ( (r%param.BoundsCheckFreq)==0 ) { 
 	  FermionField gauss(NumOp.FermionRedBlackGrid());
 	  gauss = PhiOdd;
 	  HighBoundCheck(MdagM,gauss,param.hi);

Grid/qcd/action/pseudofermion/OneFlavourRational.h

@@ -98,7 +98,7 @@ NAMESPACE_BEGIN(Grid);
 
 
       
-      virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {
+      virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
 
 	
 	// P(phi) = e^{- phi^dag (MdagM)^-1/2 phi}
@@ -142,7 +142,10 @@ NAMESPACE_BEGIN(Grid);
 
 	msCG(MdagMOp,Phi,Y);
 
-	if ( (rand()%param.BoundsCheckFreq)==0 ) { 
+	auto grid = FermOp.FermionGrid();
+        auto r=rand();
+        grid->Broadcast(0,r);
+        if ( (r%param.BoundsCheckFreq)==0 ) { 
 	  FermionField gauss(FermOp.FermionGrid());
 	  gauss = Phi;
 	  HighBoundCheck(MdagMOp,gauss,param.hi);

Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h

@@ -95,7 +95,7 @@ NAMESPACE_BEGIN(Grid);
       }
       
 
-      virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {
+      virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
 
 	// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi       
 	//
@@ -156,7 +156,10 @@ NAMESPACE_BEGIN(Grid);
 	msCG_M(MdagM,X,Y);
 
 	// Randomly apply rational bounds checks.
-	if ( (rand()%param.BoundsCheckFreq)==0 ) { 
+        auto grid = NumOp.FermionGrid();
+        auto r=rand();
+        grid->Broadcast(0,r);
+        if ( (r%param.BoundsCheckFreq)==0 ) { 	
 	  FermionField gauss(NumOp.FermionGrid());
 	  gauss = Phi;
 	  HighBoundCheck(MdagM,gauss,param.hi);

Grid/qcd/action/pseudofermion/TwoFlavour.h

@@ -73,7 +73,7 @@ public:
   //////////////////////////////////////////////////////////////////////////////////////
   // Push the gauge field in to the dops. Assume any BC's and smearing already applied
   //////////////////////////////////////////////////////////////////////////////////////
-  virtual void refresh(const GaugeField &U, GridParallelRNG &pRNG) {
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
     // P(phi) = e^{- phi^dag (MdagM)^-1 phi}
     // Phi = Mdag eta
     // P(eta) = e^{- eta^dag eta}

Grid/qcd/action/pseudofermion/TwoFlavourEvenOdd.h

@@ -77,7 +77,7 @@ public:
   //////////////////////////////////////////////////////////////////////////////////////
   // Push the gauge field in to the dops. Assume any BC's and smearing already applied
   //////////////////////////////////////////////////////////////////////////////////////
-  virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
     
     // P(phi) = e^{- phi^dag (MpcdagMpc)^-1 phi}
     // Phi = McpDag eta 

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -84,7 +84,7 @@ NAMESPACE_BEGIN(Grid);
       } 
 
       
-      virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {
+      virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
 
         // P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
         //

Grid/qcd/action/pseudofermion/TwoFlavourRatio.h

@@ -64,7 +64,7 @@ public:
     return sstream.str();
   }  
       
-  virtual void refresh(const GaugeField &U, GridParallelRNG& pRNG) {
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
 
     // P(phi) = e^{- phi^dag V (MdagM)^-1 Vdag phi}
     //

Grid/qcd/action/scalar/ScalarAction.h

@@ -55,7 +55,7 @@ public:
   }
   virtual std::string action_name() {return "ScalarAction";}
 
-  virtual void refresh(const Field &U, GridParallelRNG &pRNG) {}  // noop as no pseudoferms
+  virtual void refresh(const Field &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {}  // noop as no pseudoferms
 
   virtual RealD S(const Field &p) {
     return (mass_square * 0.5 + Nd) * ScalarObs<Impl>::sumphisquared(p) +

Grid/qcd/action/scalar/ScalarImpl.h

@@ -27,7 +27,7 @@ public:
   typedef Field              FermionField;
   typedef Field              PropagatorField;
     
-  static inline void generate_momenta(Field& P, GridParallelRNG& pRNG){
+  static inline void generate_momenta(Field& P, GridSerialRNG &sRNG, GridParallelRNG& pRNG){
     RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR); // CPS/UKQCD momentum rescaling
     gaussian(pRNG, P);
     P *= scale; 
@@ -55,6 +55,10 @@ public:
     U = 1.0;
   }
 
+  static inline void Project(Field &U) {
+    return;
+  }
+    
   static void MomentumSpacePropagator(Field &out, RealD m)
   {
     GridBase           *grid = out.Grid();
@@ -147,7 +151,7 @@ public:
       out = one / out;
     }
 
-    static inline void generate_momenta(Field &P, GridParallelRNG &pRNG)
+    static inline void generate_momenta(Field &P, GridSerialRNG & sRNG, GridParallelRNG &pRNG)
     {
       RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR); // CPS/UKQCD momentum rescaling
 #ifndef USE_FFT_ACCELERATION
@@ -234,6 +238,10 @@ public:
 #endif //USE_FFT_ACCELERATION
   }
 
+  static inline void Project(Field &U) {
+    return;
+  }
+
   static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
     Group::GaussianFundamentalLieAlgebraMatrix(pRNG, U);
   }

Grid/qcd/action/scalar/ScalarInteractionAction.h

@@ -77,7 +77,7 @@ public:
 
   virtual std::string action_name() { return "ScalarAction"; }
 
-  virtual void refresh(const Field &U, GridParallelRNG &pRNG) {}
+  virtual void refresh(const Field &U, GridSerialRNG & sRNG, GridParallelRNG &pRNG) {}
 
   virtual RealD S(const Field &p)
   {

Grid/qcd/hmc/GenericHMCrunner.h

@@ -159,6 +159,13 @@ private:
       Resources.GetCheckPointer()->CheckpointRestore(Parameters.StartTrajectory, U,
 						     Resources.GetSerialRNG(),
 						     Resources.GetParallelRNG());
+    } else {
+      // others
+      std::cout << GridLogError << "Unrecognized StartingType\n";
+      std::cout
+	<< GridLogError
+	<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+      exit(1);
     }
 
     Smearing.set_Field(U);

Grid/qcd/hmc/HMC.h

@@ -139,7 +139,7 @@ private:
   // Evolution
   /////////////////////////////////////////////////////////
   RealD evolve_hmc_step(Field &U) {
-    TheIntegrator.refresh(U, pRNG);  // set U and initialize P and phi's
+    TheIntegrator.refresh(U, sRNG, pRNG);  // set U and initialize P and phi's
 
     RealD H0 = TheIntegrator.S(U);  // initial state action
 

Grid/qcd/hmc/checkpointers/BaseCheckpointer.h

@@ -74,7 +74,7 @@ public:
       conf_file = os.str();
     }
   } 
-
+  virtual ~BaseHmcCheckpointer(){};
   void check_filename(const std::string &filename){
     std::ifstream f(filename.c_str());
     if(!f.good()){
@@ -82,7 +82,6 @@ public:
       abort();
     };
   }
-
   virtual void initialize(const CheckpointerParameters &Params) = 0;
 
   virtual void CheckpointRestore(int traj, typename Impl::Field &U,

Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h

@@ -45,6 +45,7 @@ private:
 
 public:
   INHERIT_GIMPL_TYPES(Implementation);
+  typedef GaugeStatistics<Implementation> GaugeStats;
 
   ILDGHmcCheckpointer(const CheckpointerParameters &Params_) { initialize(Params_); }
 
@@ -78,7 +79,7 @@ public:
       BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
       IldgWriter _IldgWriter(grid->IsBoss());
       _IldgWriter.open(config);
-      _IldgWriter.writeConfiguration(U, traj, config, config);
+      _IldgWriter.writeConfiguration<GaugeStats>(U, traj, config, config);
       _IldgWriter.close();
 
       std::cout << GridLogMessage << "Written ILDG Configuration on " << config
@@ -105,7 +106,7 @@ public:
     FieldMetaData header;
     IldgReader _IldgReader;
     _IldgReader.open(config);
-    _IldgReader.readConfiguration(U,header);  // format from the header
+    _IldgReader.readConfiguration<GaugeStats>(U,header);  // format from the header
     _IldgReader.close();
 
     std::cout << GridLogMessage << "Read ILDG Configuration from " << config

Grid/qcd/hmc/checkpointers/NerscCheckpointer.h

@@ -43,6 +43,7 @@ private:
 
 public:
   INHERIT_GIMPL_TYPES(Gimpl);  // only for gauge configurations
+  typedef GaugeStatistics<Gimpl> GaugeStats;
   
   NerscHmcCheckpointer(const CheckpointerParameters &Params_) { initialize(Params_); }
 
@@ -60,7 +61,7 @@ public:
       int precision32 = 1;
       int tworow = 0;
       NerscIO::writeRNGState(sRNG, pRNG, rng);
-      NerscIO::writeConfiguration(U, config, tworow, precision32);
+      NerscIO::writeConfiguration<GaugeStats>(U, config, tworow, precision32);
     }
   };
 
@@ -74,7 +75,7 @@ public:
 
     FieldMetaData header;
     NerscIO::readRNGState(sRNG, pRNG, header, rng);
-    NerscIO::readConfiguration(U, header, config);
+    NerscIO::readConfiguration<GaugeStats>(U, header, config);
   };
 };
 

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,6 +33,7 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
+#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -78,8 +79,19 @@ protected:
   RepresentationPolicy Representations;
   IntegratorParameters Params;
 
+  //Filters allow the user to manipulate the conjugate momentum, for example to freeze links in DDHMC
+  //It is applied whenever the momentum is updated / refreshed
+  //The default filter does nothing
+  MomentumFilterBase<MomentaField> const* MomFilter;
+
   const ActionSet<Field, RepresentationPolicy> as;
 
+  //Get a pointer to a shared static instance of the "do-nothing" momentum filter to serve as a default
+  static MomentumFilterBase<MomentaField> const* getDefaultMomFilter(){ 
+    static MomentumFilterNone<MomentaField> filter;
+    return &filter;
+  }
+
   void update_P(Field& U, int level, double ep) 
   {
     t_P[level] += ep;
@@ -135,6 +147,8 @@ protected:
 
     // Force from the other representations
     as[level].apply(update_P_hireps, Representations, Mom, U, ep);
+
+    MomFilter->applyFilter(Mom);
   }
 
   void update_U(Field& U, double ep) 
@@ -174,12 +188,24 @@ public:
     t_P.resize(levels, 0.0);
     t_U = 0.0;
     // initialization of smearer delegated outside of Integrator
+
+    //Default the momentum filter to "do-nothing"
+    MomFilter = getDefaultMomFilter();
   };
 
   virtual ~Integrator() {}
 
   virtual std::string integrator_name() = 0;
   
+  //Set the momentum filter allowing for manipulation of the conjugate momentum
+  void setMomentumFilter(const MomentumFilterBase<MomentaField> &filter){
+    MomFilter = &filter;
+  }
+
+  //Access the conjugate momentum
+  const MomentaField & getMomentum() const{ return P; }
+  
+
   void print_parameters()
   {
     std::cout << GridLogMessage << "[Integrator] Name : "<< integrator_name() << std::endl;
@@ -210,10 +236,9 @@ public:
   // over the representations
   struct _refresh {
     template <class FieldType, class Repr>
-    void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep,
+    void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, GridSerialRNG & sRNG, GridParallelRNG& pRNG) {
-                    GridParallelRNG& pRNG) {
       for (int a = 0; a < repr_set.size(); ++a){
-        repr_set.at(a)->refresh(Rep.U, pRNG);
+        repr_set.at(a)->refresh(Rep.U, sRNG, pRNG);
       
 	std::cout << GridLogDebug << "Hirep refreshing pseudofermions" << std::endl;
       }
@@ -221,12 +246,12 @@ public:
   } refresh_hireps{};
 
   // Initialization of momenta and actions
-  void refresh(Field& U, GridParallelRNG& pRNG) 
+  void refresh(Field& U,  GridSerialRNG & sRNG, GridParallelRNG& pRNG) 
   {
     assert(P.Grid() == U.Grid());
     std::cout << GridLogIntegrator << "Integrator refresh\n";
 
-    FieldImplementation::generate_momenta(P, pRNG);
+    FieldImplementation::generate_momenta(P, sRNG, pRNG);
 
     // Update the smeared fields, can be implemented as observer
     // necessary to keep the fields updated even after a reject
@@ -243,12 +268,14 @@ public:
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
-        as[level].actions.at(actionID)->refresh(Us, pRNG);
+        as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
       }
 
       // Refresh the higher representation actions
-      as[level].apply(refresh_hireps, Representations, pRNG);
+      as[level].apply(refresh_hireps, Representations, sRNG, pRNG);
     }
+
+    MomFilter->applyFilter(P);
   }
 
   // to be used by the actionlevel class to iterate
@@ -313,6 +340,8 @@ public:
       std::cout << GridLogIntegrator << " times[" << level << "]= " << t_P[level] << " " << t_U << std::endl;
     }
 
+    FieldImplementation::Project(U);
+
     // and that we indeed got to the end of the trajectory
     assert(fabs(t_U - Params.trajL) < 1.0e-6);
 

Grid/qcd/hmc/integrators/MomentumFilter.h

@@ -0,0 +1,94 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/MomentumFilter.h
+
+Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+//--------------------------------------------------------------------
+#ifndef MOMENTUM_FILTER
+#define MOMENTUM_FILTER
+
+NAMESPACE_BEGIN(Grid);
+
+//These filter objects allow the user to manipulate the conjugate momentum as part of the update / refresh
+
+template<typename MomentaField>
+struct MomentumFilterBase{
+  virtual void applyFilter(MomentaField &P) const;
+};
+
+//Do nothing
+template<typename MomentaField>
+struct MomentumFilterNone: public MomentumFilterBase<MomentaField>{
+  void applyFilter(MomentaField &P) const override{}
+};
+
+//Multiply each site/direction by a Lorentz vector complex number field
+//Can be used to implement a mask, zeroing out sites
+template<typename MomentaField>
+struct MomentumFilterApplyPhase: public MomentumFilterBase<MomentaField>{
+  typedef typename MomentaField::vector_type vector_type; //SIMD-vectorized complex type
+  typedef typename MomentaField::scalar_type scalar_type; //scalar complex type
+  typedef iVector<iScalar<iScalar<vector_type> >, Nd > LorentzScalarType; //complex phase for each site/direction
+  typedef Lattice<LorentzScalarType> LatticeLorentzScalarType;
+  
+  LatticeLorentzScalarType phase;
+ 
+  MomentumFilterApplyPhase(const LatticeLorentzScalarType _phase): phase(_phase){}
+
+  //Default to uniform field of (1,0)
+  MomentumFilterApplyPhase(GridBase* _grid): phase(_grid){
+    LorentzScalarType one;
+    for(int mu=0;mu<Nd;mu++)
+      one(mu)()() = scalar_type(1.);
+    
+    phase = one;
+  }
+
+  void applyFilter(MomentaField &P) const override{
+    conformable(P,phase);
+    autoView( P_v , P, AcceleratorWrite);
+    autoView( phase_v , phase, AcceleratorRead);
+
+    accelerator_for(ss,P_v.size(),MomentaField::vector_type::Nsimd(),{
+    	auto site_mom = P_v(ss);
+    	auto site_phase = phase_v(ss);
+	for(int mu=0;mu<Nd;mu++)
+	  site_mom(mu) = site_mom(mu) * site_phase(mu);
+    	coalescedWrite(P_v[ss], site_mom);
+      });
+    
+  }
+
+
+};
+
+
+
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/qcd/modules/Modules.h

@@ -99,7 +99,7 @@ public:
   virtual Prod* getPtr() = 0;
 
   // add a getReference? 
-  
+  virtual ~HMCModuleBase(){};
   virtual void print_parameters(){};  // default to nothing
 };
 

Grid/qcd/smearing/StoutSmearing.h

@@ -85,21 +85,18 @@ public:
 
     std::cout << GridLogDebug << "Stout smearing started\n";
 
-    // Smear the configurations
+    // C contains the staples multiplied by some rho
+    u_smr = U ; // set the smeared field to the current gauge field
     SmearBase->smear(C, U);
 
     for (int mu = 0; mu < Nd; mu++) {
-      if( mu == OrthogDim )
+      if( mu == OrthogDim ) continue ;
-        tmp = 1.0;  // Don't smear in the orthogonal direction
+      // u_smr = exp(iQ_mu)*U_mu apart from Orthogdim
-      else {
-        tmp = peekLorentz(C, mu);
       Umu = peekLorentz(U, mu);
-        iq_mu = Ta(
+      tmp = peekLorentz(C, mu);
-                   tmp *
+      iq_mu = Ta( tmp * adj(Umu));  
-                   adj(Umu));  // iq_mu = Ta(Omega_mu) to match the signs with the paper
       exponentiate_iQ(tmp, iq_mu);
-      }
+      pokeLorentz(u_smr, tmp * Umu, mu);
-      pokeLorentz(u_smr, tmp * Umu, mu);  // u_smr = exp(iQ_mu)*U_mu
     }
     std::cout << GridLogDebug << "Stout smearing completed\n";
   };

Grid/qcd/spin/TwoSpinor.h

@@ -128,7 +128,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spProjTm (iVector<vtype,Nhs> &hspin,const iVector<vtype,Ns> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   hspin(0)=fspin(0)-fspin(2);
   hspin(1)=fspin(1)-fspin(3);
 }
@@ -138,40 +137,50 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
  *  0 0 -1  0
  *  0 0  0 -1
  */
-
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spProj5p (iVector<vtype,Nhs> &hspin,const iVector<vtype,Ns> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   hspin(0)=fspin(0);
   hspin(1)=fspin(1);
 }
 
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spProj5m (iVector<vtype,Nhs> &hspin,const iVector<vtype,Ns> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   hspin(0)=fspin(2);
   hspin(1)=fspin(3);
 }
   
-//  template<class vtype> accelerator_inline void fspProj5p (iVector<vtype,Ns> &rfspin,const iVector<vtype,Ns> &fspin)
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spProj5p (iVector<vtype,Ns> &rfspin,const iVector<vtype,Ns> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   rfspin(0)=fspin(0);
   rfspin(1)=fspin(1);
   rfspin(2)=Zero();
   rfspin(3)=Zero();
 }
-//  template<class vtype> accelerator_inline void fspProj5m (iVector<vtype,Ns> &rfspin,const iVector<vtype,Ns> &fspin)
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spProj5m (iVector<vtype,Ns> &rfspin,const iVector<vtype,Ns> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   rfspin(0)=Zero();
   rfspin(1)=Zero();
   rfspin(2)=fspin(2);
   rfspin(3)=fspin(3);
 }
 
+template<class vtype,int N,IfCoarsened<iVector<vtype,N> > = 0> accelerator_inline void spProj5p (iVector<vtype,N> &rfspin,const iVector<vtype,N> &fspin)
+{
+  const int hN = N>>1;
+  for(int s=0;s<hN;s++){
+    rfspin(s)=fspin(s);
+    rfspin(s+hN)=Zero();
+  }
+}
+template<class vtype,int N,IfCoarsened<iVector<vtype,N> > = 0> accelerator_inline void spProj5m (iVector<vtype,N> &rfspin,const iVector<vtype,N> &fspin)
+{
+  const int hN = N>>1;
+  for(int s=0;s<hN;s++){
+    rfspin(s)=Zero();
+    rfspin(s+hN)=fspin(s+hN);
+  }
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Reconstruction routines to move back again to four spin
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -183,7 +192,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
  */
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconXp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=timesMinusI(hspin(1));
@@ -191,7 +199,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconXm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=timesI(hspin(1));
@@ -199,7 +206,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconXp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)-=timesI(hspin(1));
@@ -207,7 +213,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconXm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)+=timesI(hspin(1));
@@ -221,7 +226,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
 
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconYp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)= hspin(1);
@@ -229,7 +233,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconYm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=-hspin(1);
@@ -237,7 +240,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconYp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)+=hspin(1);
@@ -245,7 +247,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconYm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)-=hspin(1);
@@ -260,7 +261,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
  */
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconZp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=timesMinusI(hspin(0));
@@ -268,7 +268,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconZm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=     timesI(hspin(0));
@@ -276,7 +275,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconZp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)-=timesI(hspin(0));
@@ -284,7 +282,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconZm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)+=timesI(hspin(0));
@@ -298,7 +295,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
  */
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconTp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=hspin(0);
@@ -306,7 +302,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spReconTm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0);
   fspin(1)=hspin(1);
   fspin(2)=-hspin(0);
@@ -314,7 +309,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconTp (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)+=hspin(0);
@@ -322,7 +316,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumReconTm (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0);
   fspin(1)+=hspin(1);
   fspin(2)-=hspin(0);
@@ -336,7 +329,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
  */
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spRecon5p (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=hspin(0)+hspin(0); // add is lower latency than mul
   fspin(1)=hspin(1)+hspin(1); // probably no measurable diffence though
   fspin(2)=Zero();
@@ -344,7 +336,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void spRecon5m (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)=Zero();
   fspin(1)=Zero();
   fspin(2)=hspin(0)+hspin(0);
@@ -352,7 +343,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void s
 }
 template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void accumRecon5p (iVector<vtype,Ns> &fspin,const iVector<vtype,Nhs> &hspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
   fspin(0)+=hspin(0)+hspin(0);
   fspin(1)+=hspin(1)+hspin(1);
 }
@@ -372,7 +362,6 @@ template<class vtype,IfSpinor<iVector<vtype,Ns> > = 0> accelerator_inline void a
 //////////
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjXp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjXp(hspin._internal[i],fspin._internal[i]);
   }
@@ -426,26 +415,21 @@ template<class rtype,class vtype,int N> accelerator_inline void accumReconXp (iM
     }}
 }
 
-
-
 ////////
 // Xm
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjXm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjXm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjXm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjXm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjXm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjXm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -455,19 +439,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProjXm (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spReconXm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconXm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconXm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconXm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconXm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconXm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -476,45 +457,37 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconXm (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconXm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconXm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconXm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconXm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconXm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconXm(hspin._internal[i][j],fspin._internal[i][j]);
     }}
 }
 
-
-
 ////////
 // Yp
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjYp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjYp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjYp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjYp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjYp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjYp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -524,19 +497,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProjYp (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spReconYp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconYp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconYp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconYp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconYp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconYp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -545,66 +515,55 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconYp (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconYp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconYp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconYp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconYp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconYp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconYp(hspin._internal[i][j],fspin._internal[i][j]);
     }}
 }
 
-
 ////////
 // Ym
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjYm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjYm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjYm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjYm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjYm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjYm(hspin._internal[i][j],fspin._internal[i][j]);
     }}
 }
 
-
 template<class rtype,class vtype> accelerator_inline void spReconYm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconYm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconYm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,const iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconYm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconYm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconYm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -613,19 +572,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconYm (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconYm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconYm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconYm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconYm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconYm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconYm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -638,19 +594,16 @@ template<class rtype,class vtype,int N> accelerator_inline void accumReconYm (iM
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjZp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjZp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjZp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjZp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjZp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjZp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -660,19 +613,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProjZp (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spReconZp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconZp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconZp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconZp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconZp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconZp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -681,19 +631,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconZp (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconZp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconZp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconZp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconZp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconZp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconZp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -706,19 +653,16 @@ template<class rtype,class vtype,int N> accelerator_inline void accumReconZp (iM
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjZm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjZm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjZm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjZm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjZm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjZm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -728,19 +672,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProjZm (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spReconZm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconZm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconZm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconZm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconZm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconZm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -749,19 +690,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconZm (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconZm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconZm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconZm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconZm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconZm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconZm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -774,19 +712,16 @@ template<class rtype,class vtype,int N> accelerator_inline void accumReconZm (iM
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjTp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjTp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjTp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjTp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjTp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjTp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -796,19 +731,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProjTp (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spReconTp (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconTp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconTp (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconTp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconTp (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconTp(hspin._internal[i][j],fspin._internal[i][j]);
@@ -817,44 +749,37 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconTp (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconTp (iScalar<rtype> &hspin, iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconTp(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconTp (iVector<rtype,N> &hspin, const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconTp(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconTp (iMatrix<rtype,N> &hspin, const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconTp(hspin._internal[i][j],fspin._internal[i][j]);
     }}
 }
 
-
 ////////
 // Tm
 ////////
 template<class rtype,class vtype> accelerator_inline void spProjTm (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProjTm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProjTm (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProjTm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spProjTm (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProjTm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -864,19 +789,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProjTm (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spReconTm (iScalar<rtype> &hspin, const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spReconTm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spReconTm (iVector<rtype,N> &hspin, const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spReconTm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spReconTm (iMatrix<rtype,N> &hspin, const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spReconTm(hspin._internal[i][j],fspin._internal[i][j]);
@@ -885,44 +807,37 @@ template<class rtype,class vtype,int N> accelerator_inline void spReconTm (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumReconTm (iScalar<rtype> &hspin, const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumReconTm(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumReconTm (iVector<rtype,N> &hspin, const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumReconTm(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumReconTm (iMatrix<rtype,N> &hspin, const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumReconTm(hspin._internal[i][j],fspin._internal[i][j]);
     }}
 }
 
-
 ////////
 // 5p
 ////////
-template<class rtype,class vtype> accelerator_inline void spProj5p (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
+template<class rtype,class vtype,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5p (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProj5p(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProj5p (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProj5p(hspin._internal[i],fspin._internal[i]);
   }
 }
-template<class rtype,class vtype,int N> accelerator_inline void spProj5p (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
+template<class rtype,class vtype,int N,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5p (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProj5p(hspin._internal[i][j],fspin._internal[i][j]);
@@ -931,19 +846,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProj5p (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spRecon5p (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spRecon5p(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spRecon5p (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spRecon5p(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spRecon5p (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spRecon5p(hspin._internal[i][j],fspin._internal[i][j]);
@@ -952,19 +864,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spRecon5p (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumRecon5p (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumRecon5p(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumRecon5p (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumRecon5p(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumRecon5p (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumRecon5p(hspin._internal[i][j],fspin._internal[i][j]);
@@ -972,24 +881,18 @@ template<class rtype,class vtype,int N> accelerator_inline void accumRecon5p (iM
 }
 
 // four spinor projectors for chiral proj
-//  template<class vtype> accelerator_inline void fspProj5p (iScalar<vtype> &hspin,const iScalar<vtype> &fspin)
+template<class vtype,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5p (iScalar<vtype> &hspin,const iScalar<vtype> &fspin)
-template<class vtype> accelerator_inline void spProj5p (iScalar<vtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProj5p(hspin._internal,fspin._internal);
 }
-//  template<class vtype,int N> accelerator_inline void fspProj5p (iVector<vtype,N> &hspin,iVector<vtype,N> &fspin)
+template<class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5p (iVector<vtype,N> &hspin,const iVector<vtype,N> &fspin)
-template<class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProj5p (iVector<vtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProj5p(hspin._internal[i],fspin._internal[i]);
   }
 }
-//  template<class vtype,int N> accelerator_inline void fspProj5p (iMatrix<vtype,N> &hspin,iMatrix<vtype,N> &fspin)
+template<class vtype,int N,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5p (iMatrix<vtype,N> &hspin,const iMatrix<vtype,N> &fspin)
-template<class vtype,int N> accelerator_inline void spProj5p (iMatrix<vtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProj5p(hspin._internal[i][j],fspin._internal[i][j]);
@@ -1001,17 +904,17 @@ template<class vtype,int N> accelerator_inline void spProj5p (iMatrix<vtype,N> &
 // 5m
 ////////
 
-template<class rtype,class vtype> accelerator_inline void spProj5m (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
+template<class rtype,class vtype,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5m (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
   spProj5m(hspin._internal,fspin._internal);
 }
-template<class rtype,class vtype,int N,IfNotSpinor<iVector<rtype,N> > = 0> accelerator_inline void spProj5m (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
+template<class rtype,class vtype,int N,IfNotSpinor<iVector<rtype,N> > = 0,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5m (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
   for(int i=0;i<N;i++) {
     spProj5m(hspin._internal[i],fspin._internal[i]);
   }
 }
-template<class rtype,class vtype,int N> accelerator_inline void spProj5m (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
+template<class rtype,class vtype,int N,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5m (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
@@ -1021,19 +924,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spProj5m (iMatri
 
 template<class rtype,class vtype> accelerator_inline void spRecon5m (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spRecon5m(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spRecon5m (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spRecon5m(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void spRecon5m (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spRecon5m(hspin._internal[i][j],fspin._internal[i][j]);
@@ -1042,19 +942,16 @@ template<class rtype,class vtype,int N> accelerator_inline void spRecon5m (iMatr
 
 template<class rtype,class vtype> accelerator_inline void accumRecon5m (iScalar<rtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   accumRecon5m(hspin._internal,fspin._internal);
 }
 template<class rtype,class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void accumRecon5m (iVector<rtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     accumRecon5m(hspin._internal[i],fspin._internal[i]);
   }
 }
 template<class rtype,class vtype,int N> accelerator_inline void accumRecon5m (iMatrix<rtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       accumRecon5m(hspin._internal[i][j],fspin._internal[i][j]);
@@ -1063,24 +960,18 @@ template<class rtype,class vtype,int N> accelerator_inline void accumRecon5m (iM
 
 
 // four spinor projectors for chiral proj
-//  template<class vtype> accelerator_inline void fspProj5m (iScalar<vtype> &hspin,const iScalar<vtype> &fspin)
+template<class vtype,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5m (iScalar<vtype> &hspin,const iScalar<vtype> &fspin)
-template<class vtype> accelerator_inline void spProj5m (iScalar<vtype> &hspin,const iScalar<vtype> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iScalar<vtype>,SpinorIndex>::notvalue,iScalar<vtype> >::type *temp;
   spProj5m(hspin._internal,fspin._internal);
 }
-//  template<class vtype,int N> accelerator_inline void fspProj5m (iVector<vtype,N> &hspin,iVector<vtype,N> &fspin)
+template<class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5m (iVector<vtype,N> &hspin,const iVector<vtype,N> &fspin)
-template<class vtype,int N,IfNotSpinor<iVector<vtype,N> > = 0> accelerator_inline void spProj5m (iVector<vtype,N> &hspin,const iVector<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iVector<vtype,N>,SpinorIndex>::notvalue,iVector<vtype,N> >::type *temp;
   for(int i=0;i<N;i++) {
     spProj5m(hspin._internal[i],fspin._internal[i]);
   }
 }
-//  template<class vtype,int N> accelerator_inline void fspProj5m (iMatrix<vtype,N> &hspin,iMatrix<vtype,N> &fspin)
+template<class vtype,int N,IfNotCoarsened<iScalar<vtype> > = 0> accelerator_inline void spProj5m (iMatrix<vtype,N> &hspin,const iMatrix<vtype,N> &fspin)
-template<class vtype,int N> accelerator_inline void spProj5m (iMatrix<vtype,N> &hspin,const iMatrix<vtype,N> &fspin)
 {
-  //typename std::enable_if<matchGridTensorIndex<iMatrix<vtype,N>,SpinorIndex>::notvalue,iMatrix<vtype,N> >::type *temp;
   for(int i=0;i<N;i++){ 
     for(int j=0;j<N;j++){
       spProj5m(hspin._internal[i][j],fspin._internal[i][j]);

Grid/qcd/utils/CovariantCshift.h

@@ -53,6 +53,24 @@ namespace PeriodicBC {
     return Cshift(tmp,mu,-1);// moves towards positive mu
   }
 
+  template<class gauge> Lattice<gauge>
+  CovShiftIdentityBackward(const Lattice<gauge> &Link, int mu) 
+  {
+    return Cshift(adj(Link), mu, -1);
+  }
+
+  template<class gauge> Lattice<gauge>
+  CovShiftIdentityForward(const Lattice<gauge> &Link, int mu)
+  {
+    return Link;
+  }
+
+  template<class gauge> Lattice<gauge>
+  ShiftStaple(const Lattice<gauge> &Link, int mu)
+  {
+    return Cshift(Link, mu, 1);
+  }
+  
   template<class gauge,class Expr,typename std::enable_if<is_lattice_expr<Expr>::value,void>::type * = nullptr>
     auto  CovShiftForward(const Lattice<gauge> &Link, 
 			  int mu,
@@ -70,6 +88,7 @@ namespace PeriodicBC {
     return CovShiftBackward(Link,mu,arg);
   }
 
+
 }
 
 
@@ -139,6 +158,38 @@ namespace ConjugateBC {
     //    std::cout<<"Gparity::CovCshiftBackward mu="<<mu<<std::endl;
     return Cshift(tmp,mu,-1);// moves towards positive mu
   }
+  template<class gauge> Lattice<gauge>
+  CovShiftIdentityBackward(const Lattice<gauge> &Link, int mu) {
+    GridBase *grid = Link.Grid();
+    int Lmu = grid->GlobalDimensions()[mu] - 1;
+
+    Lattice<iScalar<vInteger>> coor(grid);
+    LatticeCoordinate(coor, mu);
+
+    Lattice<gauge> tmp(grid);
+    tmp = adj(Link);
+    tmp = where(coor == Lmu, conjugate(tmp), tmp);
+    return Cshift(tmp, mu, -1); // moves towards positive mu
+  }
+  template<class gauge> Lattice<gauge>
+  CovShiftIdentityForward(const Lattice<gauge> &Link, int mu) {
+    return Link;
+  }
+
+  template<class gauge> Lattice<gauge>
+  ShiftStaple(const Lattice<gauge> &Link, int mu)
+  {
+    GridBase *grid = Link.Grid();
+    int Lmu = grid->GlobalDimensions()[mu] - 1;
+
+    Lattice<iScalar<vInteger>> coor(grid);
+    LatticeCoordinate(coor, mu);
+
+    Lattice<gauge> tmp(grid);
+    tmp = Cshift(Link, mu, 1);
+    tmp = where(coor == Lmu, conjugate(tmp), tmp);
+    return tmp;
+  }
 
   template<class gauge,class Expr,typename std::enable_if<is_lattice_expr<Expr>::value,void>::type * = nullptr>
     auto  CovShiftForward(const Lattice<gauge> &Link, 

Grid/qcd/utils/LinalgUtils.h

@@ -188,7 +188,6 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
   z.Checkerboard() = x.Checkerboard();
   conformable(x,z);
   int Ls = grid->_rdimensions[0];
-  Gamma G5(Gamma::Algebra::Gamma5);
   autoView( x_v, x, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
   uint64_t nloop = grid->oSites()/Ls;
@@ -196,7 +195,13 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
     uint64_t ss = sss*Ls;
     for(int s=0;s<Ls;s++){
       int sp = Ls-1-s;
-      coalescedWrite(z_v[ss+sp],G5*x_v(ss+s));
+      auto tmp = x_v(ss+s);
+      decltype(tmp) tmp_p;
+      decltype(tmp) tmp_m;
+      spProj5p(tmp_p,tmp);
+      spProj5m(tmp_m,tmp);
+      // Use of spProj5m, 5p captures the coarse space too
+      coalescedWrite(z_v[ss+sp],tmp_p - tmp_m);
     }
   });
 }
@@ -208,10 +213,20 @@ void G5C(Lattice<vobj> &z, const Lattice<vobj> &x)
   z.Checkerboard() = x.Checkerboard();
   conformable(x, z);
 
-  Gamma G5(Gamma::Algebra::Gamma5);
+  autoView( x_v, x, AcceleratorRead);
-  z = G5 * x;
+  autoView( z_v, z, AcceleratorWrite);
+  uint64_t nloop = grid->oSites();
+  accelerator_for(ss,nloop,vobj::Nsimd(),{
+    auto tmp = x_v(ss);
+    decltype(tmp) tmp_p;
+    decltype(tmp) tmp_m;
+    spProj5p(tmp_p,tmp);
+    spProj5m(tmp_m,tmp);
+    coalescedWrite(z_v[ss],tmp_p - tmp_m);
+  });
 }
 
+/*
 template<class CComplex, int nbasis>
 void G5C(Lattice<iVector<CComplex, nbasis>> &z, const Lattice<iVector<CComplex, nbasis>> &x)
 {
@@ -234,6 +249,7 @@ void G5C(Lattice<iVector<CComplex, nbasis>> &z, const Lattice<iVector<CComplex,
     }
   });
 }
+*/
 
 NAMESPACE_END(Grid);
 

Grid/qcd/utils/Metric.h

@@ -93,13 +93,13 @@ public:
   GeneralisedMomenta(GridBase* grid, Metric<MomentaField>& M): M(M), Mom(grid), AuxMom(grid), AuxField(grid){}
 
   // Correct
-  void MomentaDistribution(GridParallelRNG& pRNG){
+  void MomentaDistribution(GridSerialRNG & sRNG, GridParallelRNG& pRNG){
     // Generate a distribution for
     // P^dag G P
     // where G = M^-1
 
     // Generate gaussian momenta
-    Implementation::generate_momenta(Mom, pRNG);
+    Implementation::generate_momenta(Mom, sRNG, pRNG);
     // Modify the distribution with the metric
     M.MSquareRoot(Mom);
 
@@ -107,8 +107,8 @@ public:
       // Auxiliary momenta
       // do nothing if trivial, so hide in the metric
       MomentaField AuxMomTemp(Mom.Grid());
-      Implementation::generate_momenta(AuxMom, pRNG);
+      Implementation::generate_momenta(AuxMom, sRNG, pRNG);
-      Implementation::generate_momenta(AuxField, pRNG);
+      Implementation::generate_momenta(AuxField, sRNG, pRNG);
       // Modify the distribution with the metric
       // Aux^dag M Aux
       M.MInvSquareRoot(AuxMom);  // AuxMom = M^{-1/2} AuxMomTemp

Grid/qcd/utils/SUn.h

@@ -449,7 +449,8 @@ public:
     LatticeReal alpha(grid);
 
     //    std::cout<<GridLogMessage<<"xi "<<xi <<std::endl;
-    alpha = toReal(2.0 * xi);
+    xi = 2.0 *xi;
+    alpha = toReal(xi);
 
     do {
       // A. Generate two uniformly distributed pseudo-random numbers R and R',
@@ -734,7 +735,6 @@ public:
     }
   }
 
-
   template <typename GaugeField>
   static void HotConfiguration(GridParallelRNG &pRNG, GaugeField &out) {
     typedef typename GaugeField::vector_type vector_type;
@@ -799,6 +799,88 @@ public:
   }
 };
 
+template<int N>
+LatticeComplexD Determinant(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  LatticeComplexD ret(grid);
+
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us()()(i,j);
+      }}
+    ComplexD det = EigenU.determinant();
+    pokeLocalSite(det,ret_v,lcoor);
+  });
+  return ret;
+}
+template<int N>
+static void ProjectSUn(Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
+{
+  Umu      = ProjectOnGroup(Umu);
+  auto det = Determinant(Umu);
+
+  det = conjugate(det);
+
+  for(int i=0;i<N;i++){
+    auto element = PeekIndex<ColourIndex>(Umu,N-1,i);
+    element = element * det;
+    PokeIndex<ColourIndex>(Umu,element,Nc-1,i);
+  }
+}
+template<int N>
+static void ProjectSUn(Lattice<iVector<iScalar<iMatrix<vComplexD, N> >,Nd> > &U)
+{
+  GridBase *grid=U.Grid();
+  // Reunitarise
+  for(int mu=0;mu<Nd;mu++){
+    auto Umu = PeekIndex<LorentzIndex>(U,mu);
+    Umu      = ProjectOnGroup(Umu);
+    ProjectSUn(Umu);
+    PokeIndex<LorentzIndex>(U,Umu,mu);
+  }
+}
+// Explicit specialisation for SU(3).
+// Explicit specialisation for SU(3).
+static void
+ProjectSU3 (Lattice<iScalar<iScalar<iMatrix<vComplexD, 3> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  const int x=0;
+  const int y=1;
+  const int z=2;
+  // Reunitarise
+  Umu = ProjectOnGroup(Umu);
+  autoView(Umu_v,Umu,CpuWrite);
+  thread_for(ss,grid->oSites(),{
+      auto cm = Umu_v[ss];
+      cm()()(2,x) = adj(cm()()(0,y)*cm()()(1,z)-cm()()(0,z)*cm()()(1,y)); //x= yz-zy
+      cm()()(2,y) = adj(cm()()(0,z)*cm()()(1,x)-cm()()(0,x)*cm()()(1,z)); //y= zx-xz
+      cm()()(2,z) = adj(cm()()(0,x)*cm()()(1,y)-cm()()(0,y)*cm()()(1,x)); //z= xy-yx
+      Umu_v[ss]=cm;
+  });
+}
+static void ProjectSU3(Lattice<iVector<iScalar<iMatrix<vComplexD, 3> >,Nd> > &U)
+{
+  GridBase *grid=U.Grid();
+  // Reunitarise
+  for(int mu=0;mu<Nd;mu++){
+    auto Umu = PeekIndex<LorentzIndex>(U,mu);
+    Umu      = ProjectOnGroup(Umu);
+    ProjectSU3(Umu);
+    PokeIndex<LorentzIndex>(U,Umu,mu);
+  }
+}
+
 typedef SU<2> SU2;
 typedef SU<3> SU3;
 typedef SU<4> SU4;

Grid/serialisation/JSON_IO.cc

@@ -26,7 +26,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#ifndef __NVCC__
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
 
 NAMESPACE_BEGIN(Grid);
 

Grid/simd/Fujitsu_A64FX_asm_double.h

@@ -1,779 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: Fujitsu_A64FX_asm_double.h
-
-    Copyright (C) 2020
-
-Author: Nils Meyer <nils.meyer@ur.de>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#define LOAD_CHIMU(base)               LOAD_CHIMU_INTERLEAVED_A64FXd(base)  
-#define PREFETCH_CHIMU_L1(A)           PREFETCH_CHIMU_L1_INTERNAL_A64FXd(A)  
-#define PREFETCH_GAUGE_L1(A)           PREFETCH_GAUGE_L1_INTERNAL_A64FXd(A)  
-#define PREFETCH_CHIMU_L2(A)           PREFETCH_CHIMU_L2_INTERNAL_A64FXd(A)  
-#define PREFETCH_GAUGE_L2(A)           PREFETCH_GAUGE_L2_INTERNAL_A64FXd(A)  
-#define PF_GAUGE(A)  
-#define PREFETCH_RESULT_L2_STORE(A)    PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXd(A)  
-#define PREFETCH_RESULT_L1_STORE(A)    PREFETCH_RESULT_L1_STORE_INTERNAL_A64FXd(A)  
-#define PREFETCH1_CHIMU(A)             PREFETCH_CHIMU_L1(A)  
-#define PREFETCH_CHIMU(A)              PREFETCH_CHIMU_L1(A)  
-#define LOCK_GAUGE(A)  
-#define UNLOCK_GAUGE(A)  
-#define MASK_REGS                      DECLARATIONS_A64FXd  
-#define SAVE_RESULT(A,B)               RESULT_A64FXd(A); PREFETCH_RESULT_L2_STORE(B)  
-#define MULT_2SPIN_1(Dir)              MULT_2SPIN_1_A64FXd(Dir)  
-#define MULT_2SPIN_2                   MULT_2SPIN_2_A64FXd  
-#define LOAD_CHI(base)                 LOAD_CHI_A64FXd(base)  
-#define ADD_RESULT(base,basep)         LOAD_CHIMU(base); ADD_RESULT_INTERNAL_A64FXd; RESULT_A64FXd(base)  
-#define XP_PROJ                        XP_PROJ_A64FXd  
-#define YP_PROJ                        YP_PROJ_A64FXd  
-#define ZP_PROJ                        ZP_PROJ_A64FXd  
-#define TP_PROJ                        TP_PROJ_A64FXd  
-#define XM_PROJ                        XM_PROJ_A64FXd  
-#define YM_PROJ                        YM_PROJ_A64FXd  
-#define ZM_PROJ                        ZM_PROJ_A64FXd  
-#define TM_PROJ                        TM_PROJ_A64FXd  
-#define XP_RECON                       XP_RECON_A64FXd  
-#define XM_RECON                       XM_RECON_A64FXd  
-#define XM_RECON_ACCUM                 XM_RECON_ACCUM_A64FXd  
-#define YM_RECON_ACCUM                 YM_RECON_ACCUM_A64FXd  
-#define ZM_RECON_ACCUM                 ZM_RECON_ACCUM_A64FXd  
-#define TM_RECON_ACCUM                 TM_RECON_ACCUM_A64FXd  
-#define XP_RECON_ACCUM                 XP_RECON_ACCUM_A64FXd  
-#define YP_RECON_ACCUM                 YP_RECON_ACCUM_A64FXd  
-#define ZP_RECON_ACCUM                 ZP_RECON_ACCUM_A64FXd  
-#define TP_RECON_ACCUM                 TP_RECON_ACCUM_A64FXd  
-#define PERMUTE_DIR0                   0  
-#define PERMUTE_DIR1                   1  
-#define PERMUTE_DIR2                   2  
-#define PERMUTE_DIR3                   3  
-#define PERMUTE                        PERMUTE_A64FXd;  
-#define LOAD_TABLE(Dir)                if (Dir == 0) { LOAD_TABLE0; } else if (Dir == 1) { LOAD_TABLE1; } else if (Dir == 2) { LOAD_TABLE2; }  
-#define MAYBEPERM(Dir,perm)            if (Dir != 3) { if (perm) { PERMUTE; } }  
-// DECLARATIONS
-#define DECLARATIONS_A64FXd  \
-    const uint64_t lut[4][8] = { \
-        {4, 5, 6, 7, 0, 1, 2, 3}, \
-        {2, 3, 0, 1, 6, 7, 4, 5}, \
-        {1, 0, 3, 2, 5, 4, 7, 6}, \
-        {0, 1, 2, 4, 5, 6, 7, 8} };\
-asm ( \
-    "fmov z31.d , 0 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// RESULT
-#define RESULT_A64FXd(base)  \
-{ \
-asm ( \
-    "str z0, [%[storeptr], -6, mul vl] \n\t" \
-    "str z1, [%[storeptr], -5, mul vl] \n\t" \
-    "str z2, [%[storeptr], -4, mul vl] \n\t" \
-    "str z3, [%[storeptr], -3, mul vl] \n\t" \
-    "str z4, [%[storeptr], -2, mul vl] \n\t" \
-    "str z5, [%[storeptr], -1, mul vl] \n\t" \
-    "str z6, [%[storeptr], 0, mul vl] \n\t" \
-    "str z7, [%[storeptr], 1, mul vl] \n\t" \
-    "str z8, [%[storeptr], 2, mul vl] \n\t" \
-    "str z9, [%[storeptr], 3, mul vl] \n\t" \
-    "str z10, [%[storeptr], 4, mul vl] \n\t" \
-    "str z11, [%[storeptr], 5, mul vl] \n\t" \
-    :  \
-    : [storeptr] "r" (base + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_CHIMU_L2 (prefetch to L2)
-#define PREFETCH_CHIMU_L2_INTERNAL_A64FXd(base)  \
-{ \
-asm ( \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_CHIMU_L1 (prefetch to L1)
-#define PREFETCH_CHIMU_L1_INTERNAL_A64FXd(base)  \
-{ \
-asm ( \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_GAUGE_L2 (prefetch to L2)
-#define PREFETCH_GAUGE_L2_INTERNAL_A64FXd(A)  \
-{ \
-    const auto & ref(U[sUn](A)); uint64_t baseU = (uint64_t)&ref + 3 * 3 * 64; \
-asm ( \
-    "prfd PLDL2STRM, p5, [%[fetchptr], -4, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 12, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 16, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 20, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 24, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 28, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_GAUGE_L1 (prefetch to L1)
-#define PREFETCH_GAUGE_L1_INTERNAL_A64FXd(A)  \
-{ \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
-asm ( \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHI
-#define LOAD_CHI_A64FXd(base)  \
-{ \
-asm ( \
-    "ldr z12, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], 2, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], 5, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHIMU
-#define LOAD_CHIMU_INTERLEAVED_A64FXd(base)  \
-{ \
-asm ( \
-    "ptrue p5.d \n\t" \
-    "ldr z12, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z21, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z18, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z22, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z19, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z23, [%[fetchptr], 5, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z20, [%[fetchptr], 2, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHIMU_0213
-#define LOAD_CHIMU_0213_A64FXd  \
-{ \
-    const SiteSpinor & ref(in[offset]); \
-asm ( \
-    "ptrue p5.d \n\t" \
-    "ldr z12, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z18, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z19, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z20, [%[fetchptr], 2, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z21, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z22, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z23, [%[fetchptr], 5, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (&ref[2][0]) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHIMU_0312
-#define LOAD_CHIMU_0312_A64FXd  \
-{ \
-    const SiteSpinor & ref(in[offset]); \
-asm ( \
-    "ptrue p5.d \n\t" \
-    "ldr z12, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z21, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z22, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z23, [%[fetchptr], 5, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z18, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z19, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z20, [%[fetchptr], 2, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (&ref[2][0]) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_TABLE0
-#define LOAD_TABLE0  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (0) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_TABLE1
-#define LOAD_TABLE1  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (1) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_TABLE2
-#define LOAD_TABLE2  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (2) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_TABLE3
-#define LOAD_TABLE3  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (3) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// PERMUTE
-#define PERMUTE_A64FXd  \
-asm ( \
-    "tbl z12.d, { z12.d }, z30.d \n\t"  \
-    "tbl z13.d, { z13.d }, z30.d \n\t"  \
-    "tbl z14.d, { z14.d }, z30.d \n\t"  \
-    "tbl z15.d, { z15.d }, z30.d \n\t"  \
-    "tbl z16.d, { z16.d }, z30.d \n\t"  \
-    "tbl z17.d, { z17.d }, z30.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_GAUGE
-#define LOAD_GAUGE  \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
-{ \
-asm ( \
-    "ptrue p5.d \n\t" \
-    "ldr z24, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z25, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z26, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z27, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z28, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z29, [%[fetchptr], 1, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// MULT_2SPIN
-#define MULT_2SPIN_1_A64FXd(A)  \
-{ \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
-asm ( \
-    "ldr z24, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z25, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z26, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z27, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z28, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z29, [%[fetchptr], 1, mul vl] \n\t" \
-    "movprfx z18.d, p5/m, z31.d \n\t" \
-    "fcmla z18.d, p5/m, z24.d, z12.d, 0 \n\t" \
-    "movprfx z21.d, p5/m, z31.d \n\t" \
-    "fcmla z21.d, p5/m, z24.d, z15.d, 0 \n\t" \
-    "movprfx z19.d, p5/m, z31.d \n\t" \
-    "fcmla z19.d, p5/m, z25.d, z12.d, 0 \n\t" \
-    "movprfx z22.d, p5/m, z31.d \n\t" \
-    "fcmla z22.d, p5/m, z25.d, z15.d, 0 \n\t" \
-    "movprfx z20.d, p5/m, z31.d \n\t" \
-    "fcmla z20.d, p5/m, z26.d, z12.d, 0 \n\t" \
-    "movprfx z23.d, p5/m, z31.d \n\t" \
-    "fcmla z23.d, p5/m, z26.d, z15.d, 0 \n\t" \
-    "fcmla z18.d, p5/m, z24.d, z12.d, 90 \n\t" \
-    "fcmla z21.d, p5/m, z24.d, z15.d, 90 \n\t" \
-    "fcmla z19.d, p5/m, z25.d, z12.d, 90 \n\t" \
-    "fcmla z22.d, p5/m, z25.d, z15.d, 90 \n\t" \
-    "fcmla z20.d, p5/m, z26.d, z12.d, 90 \n\t" \
-    "fcmla z23.d, p5/m, z26.d, z15.d, 90 \n\t" \
-    "ldr z24, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z25, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z26, [%[fetchptr], 2, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// MULT_2SPIN_BACKEND
-#define MULT_2SPIN_2_A64FXd  \
-{ \
-asm ( \
-    "fcmla z18.d, p5/m, z27.d, z13.d, 0 \n\t" \
-    "fcmla z21.d, p5/m, z27.d, z16.d, 0 \n\t" \
-    "fcmla z19.d, p5/m, z28.d, z13.d, 0 \n\t" \
-    "fcmla z22.d, p5/m, z28.d, z16.d, 0 \n\t" \
-    "fcmla z20.d, p5/m, z29.d, z13.d, 0 \n\t" \
-    "fcmla z23.d, p5/m, z29.d, z16.d, 0 \n\t" \
-    "fcmla z18.d, p5/m, z27.d, z13.d, 90 \n\t" \
-    "fcmla z21.d, p5/m, z27.d, z16.d, 90 \n\t" \
-    "fcmla z19.d, p5/m, z28.d, z13.d, 90 \n\t" \
-    "fcmla z22.d, p5/m, z28.d, z16.d, 90 \n\t" \
-    "fcmla z20.d, p5/m, z29.d, z13.d, 90 \n\t" \
-    "fcmla z23.d, p5/m, z29.d, z16.d, 90 \n\t" \
-    "fcmla z18.d, p5/m, z24.d, z14.d, 0 \n\t" \
-    "fcmla z21.d, p5/m, z24.d, z17.d, 0 \n\t" \
-    "fcmla z19.d, p5/m, z25.d, z14.d, 0 \n\t" \
-    "fcmla z22.d, p5/m, z25.d, z17.d, 0 \n\t" \
-    "fcmla z20.d, p5/m, z26.d, z14.d, 0 \n\t" \
-    "fcmla z23.d, p5/m, z26.d, z17.d, 0 \n\t" \
-    "fcmla z18.d, p5/m, z24.d, z14.d, 90 \n\t" \
-    "fcmla z21.d, p5/m, z24.d, z17.d, 90 \n\t" \
-    "fcmla z19.d, p5/m, z25.d, z14.d, 90 \n\t" \
-    "fcmla z22.d, p5/m, z25.d, z17.d, 90 \n\t" \
-    "fcmla z20.d, p5/m, z26.d, z14.d, 90 \n\t" \
-    "fcmla z23.d, p5/m, z26.d, z17.d, 90 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XP_PROJ
-#define XP_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fcadd z12.d, p5/m, z12.d, z21.d, 90 \n\t" \
-    "fcadd z13.d, p5/m, z13.d, z22.d, 90 \n\t" \
-    "fcadd z14.d, p5/m, z14.d, z23.d, 90 \n\t" \
-    "fcadd z15.d, p5/m, z15.d, z18.d, 90 \n\t" \
-    "fcadd z16.d, p5/m, z16.d, z19.d, 90 \n\t" \
-    "fcadd z17.d, p5/m, z17.d, z20.d, 90 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XP_RECON
-#define XP_RECON_A64FXd  \
-asm ( \
-    "movprfx z6.d, p5/m, z31.d \n\t" \
-    "fcadd z6.d, p5/m, z6.d, z21.d, 270 \n\t" \
-    "movprfx z7.d, p5/m, z31.d \n\t" \
-    "fcadd z7.d, p5/m, z7.d, z22.d, 270 \n\t" \
-    "movprfx z8.d, p5/m, z31.d \n\t" \
-    "fcadd z8.d, p5/m, z8.d, z23.d, 270 \n\t" \
-    "movprfx z9.d, p5/m, z31.d \n\t" \
-    "fcadd z9.d, p5/m, z9.d, z18.d, 270 \n\t" \
-    "movprfx z10.d, p5/m, z31.d \n\t" \
-    "fcadd z10.d, p5/m, z10.d, z19.d, 270 \n\t" \
-    "movprfx z11.d, p5/m, z31.d \n\t" \
-    "fcadd z11.d, p5/m, z11.d, z20.d, 270 \n\t" \
-    "mov z0.d, p5/m, z18.d \n\t" \
-    "mov z1.d, p5/m, z19.d \n\t" \
-    "mov z2.d, p5/m, z20.d \n\t" \
-    "mov z3.d, p5/m, z21.d \n\t" \
-    "mov z4.d, p5/m, z22.d \n\t" \
-    "mov z5.d, p5/m, z23.d \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// XP_RECON_ACCUM
-#define XP_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fcadd z9.d, p5/m, z9.d, z18.d, 270 \n\t" \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fcadd z10.d, p5/m, z10.d, z19.d, 270 \n\t" \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fcadd z11.d, p5/m, z11.d, z20.d, 270 \n\t" \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fcadd z6.d, p5/m, z6.d, z21.d, 270 \n\t" \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fcadd z7.d, p5/m, z7.d, z22.d, 270 \n\t" \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fcadd z8.d, p5/m, z8.d, z23.d, 270 \n\t" \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YP_PROJ
-#define YP_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fsub z12.d, p5/m, z12.d, z21.d \n\t" \
-    "fsub z13.d, p5/m, z13.d, z22.d \n\t" \
-    "fsub z14.d, p5/m, z14.d, z23.d \n\t" \
-    "fadd z15.d, p5/m, z15.d, z18.d \n\t"  \
-    "fadd z16.d, p5/m, z16.d, z19.d \n\t"  \
-    "fadd z17.d, p5/m, z17.d, z20.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// ZP_PROJ
-#define ZP_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fcadd z12.d, p5/m, z12.d, z18.d, 90 \n\t" \
-    "fcadd z13.d, p5/m, z13.d, z19.d, 90 \n\t" \
-    "fcadd z14.d, p5/m, z14.d, z20.d, 90 \n\t" \
-    "fcadd z15.d, p5/m, z15.d, z21.d, 270 \n\t" \
-    "fcadd z16.d, p5/m, z16.d, z22.d, 270 \n\t" \
-    "fcadd z17.d, p5/m, z17.d, z23.d, 270 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// TP_PROJ
-#define TP_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fadd z12.d, p5/m, z12.d, z18.d \n\t"  \
-    "fadd z13.d, p5/m, z13.d, z19.d \n\t"  \
-    "fadd z14.d, p5/m, z14.d, z20.d \n\t"  \
-    "fadd z15.d, p5/m, z15.d, z21.d \n\t"  \
-    "fadd z16.d, p5/m, z16.d, z22.d \n\t"  \
-    "fadd z17.d, p5/m, z17.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XM_PROJ
-#define XM_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fcadd z12.d, p5/m, z12.d, z21.d, 270 \n\t" \
-    "fcadd z13.d, p5/m, z13.d, z22.d, 270 \n\t" \
-    "fcadd z14.d, p5/m, z14.d, z23.d, 270 \n\t" \
-    "fcadd z15.d, p5/m, z15.d, z18.d, 270 \n\t" \
-    "fcadd z16.d, p5/m, z16.d, z19.d, 270 \n\t" \
-    "fcadd z17.d, p5/m, z17.d, z20.d, 270 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XM_RECON
-#define XM_RECON_A64FXd  \
-asm ( \
-    "movprfx z6.d, p5/m, z31.d \n\t" \
-    "fcadd z6.d, p5/m, z6.d, z21.d, 90 \n\t" \
-    "movprfx z7.d, p5/m, z31.d \n\t" \
-    "fcadd z7.d, p5/m, z7.d, z22.d, 90 \n\t" \
-    "movprfx z8.d, p5/m, z31.d \n\t" \
-    "fcadd z8.d, p5/m, z8.d, z23.d, 90 \n\t" \
-    "movprfx z9.d, p5/m, z31.d \n\t" \
-    "fcadd z9.d, p5/m, z9.d, z18.d, 90 \n\t" \
-    "movprfx z10.d, p5/m, z31.d \n\t" \
-    "fcadd z10.d, p5/m, z10.d, z19.d, 90 \n\t" \
-    "movprfx z11.d, p5/m, z31.d \n\t" \
-    "fcadd z11.d, p5/m, z11.d, z20.d, 90 \n\t" \
-    "mov z0.d, p5/m, z18.d \n\t" \
-    "mov z1.d, p5/m, z19.d \n\t" \
-    "mov z2.d, p5/m, z20.d \n\t" \
-    "mov z3.d, p5/m, z21.d \n\t" \
-    "mov z4.d, p5/m, z22.d \n\t" \
-    "mov z5.d, p5/m, z23.d \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YM_PROJ
-#define YM_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fadd z12.d, p5/m, z12.d, z21.d \n\t"  \
-    "fadd z13.d, p5/m, z13.d, z22.d \n\t"  \
-    "fadd z14.d, p5/m, z14.d, z23.d \n\t"  \
-    "fsub z15.d, p5/m, z15.d, z18.d \n\t" \
-    "fsub z16.d, p5/m, z16.d, z19.d \n\t" \
-    "fsub z17.d, p5/m, z17.d, z20.d \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// ZM_PROJ
-#define ZM_PROJ_A64FXd  \
-{ \
-asm ( \
-    "fcadd z12.d, p5/m, z12.d, z18.d, 270 \n\t" \
-    "fcadd z13.d, p5/m, z13.d, z19.d, 270 \n\t" \
-    "fcadd z14.d, p5/m, z14.d, z20.d, 270 \n\t" \
-    "fcadd z15.d, p5/m, z15.d, z21.d, 90 \n\t" \
-    "fcadd z16.d, p5/m, z16.d, z22.d, 90 \n\t" \
-    "fcadd z17.d, p5/m, z17.d, z23.d, 90 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// TM_PROJ
-#define TM_PROJ_A64FXd  \
-{ \
-asm ( \
-    "ptrue p5.d \n\t" \
-    "fsub z12.d, p5/m, z12.d, z18.d \n\t" \
-    "fsub z13.d, p5/m, z13.d, z19.d \n\t" \
-    "fsub z14.d, p5/m, z14.d, z20.d \n\t" \
-    "fsub z15.d, p5/m, z15.d, z21.d \n\t" \
-    "fsub z16.d, p5/m, z16.d, z22.d \n\t" \
-    "fsub z17.d, p5/m, z17.d, z23.d \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XM_RECON_ACCUM
-#define XM_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fcadd z9.d, p5/m, z9.d, z18.d, 90 \n\t" \
-    "fcadd z10.d, p5/m, z10.d, z19.d, 90 \n\t" \
-    "fcadd z11.d, p5/m, z11.d, z20.d, 90 \n\t" \
-    "fcadd z6.d, p5/m, z6.d, z21.d, 90 \n\t" \
-    "fcadd z7.d, p5/m, z7.d, z22.d, 90 \n\t" \
-    "fcadd z8.d, p5/m, z8.d, z23.d, 90 \n\t" \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YP_RECON_ACCUM
-#define YP_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fsub z9.d, p5/m, z9.d, z18.d \n\t" \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fsub z10.d, p5/m, z10.d, z19.d \n\t" \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fsub z11.d, p5/m, z11.d, z20.d \n\t" \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fadd z6.d, p5/m, z6.d, z21.d \n\t"  \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fadd z7.d, p5/m, z7.d, z22.d \n\t"  \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    "fadd z8.d, p5/m, z8.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YM_RECON_ACCUM
-#define YM_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fadd z9.d, p5/m, z9.d, z18.d \n\t"  \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fadd z10.d, p5/m, z10.d, z19.d \n\t"  \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fadd z11.d, p5/m, z11.d, z20.d \n\t"  \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fsub z6.d, p5/m, z6.d, z21.d \n\t" \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fsub z7.d, p5/m, z7.d, z22.d \n\t" \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    "fsub z8.d, p5/m, z8.d, z23.d \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// ZP_RECON_ACCUM
-#define ZP_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fcadd z6.d, p5/m, z6.d, z18.d, 270 \n\t" \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fcadd z7.d, p5/m, z7.d, z19.d, 270 \n\t" \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fcadd z8.d, p5/m, z8.d, z20.d, 270 \n\t" \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fcadd z9.d, p5/m, z9.d, z21.d, 90 \n\t" \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fcadd z10.d, p5/m, z10.d, z22.d, 90 \n\t" \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fcadd z11.d, p5/m, z11.d, z23.d, 90 \n\t" \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// ZM_RECON_ACCUM
-#define ZM_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fcadd z6.d, p5/m, z6.d, z18.d, 90 \n\t" \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fcadd z7.d, p5/m, z7.d, z19.d, 90 \n\t" \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fcadd z8.d, p5/m, z8.d, z20.d, 90 \n\t" \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fcadd z9.d, p5/m, z9.d, z21.d, 270 \n\t" \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fcadd z10.d, p5/m, z10.d, z22.d, 270 \n\t" \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fcadd z11.d, p5/m, z11.d, z23.d, 270 \n\t" \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// TP_RECON_ACCUM
-#define TP_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fadd z6.d, p5/m, z6.d, z18.d \n\t"  \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fadd z7.d, p5/m, z7.d, z19.d \n\t"  \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fadd z8.d, p5/m, z8.d, z20.d \n\t"  \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fadd z9.d, p5/m, z9.d, z21.d \n\t"  \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fadd z10.d, p5/m, z10.d, z22.d \n\t"  \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    "fadd z11.d, p5/m, z11.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// TM_RECON_ACCUM
-#define TM_RECON_ACCUM_A64FXd  \
-asm ( \
-    "fadd z0.d, p5/m, z0.d, z18.d \n\t"  \
-    "fsub z6.d, p5/m, z6.d, z18.d \n\t" \
-    "fadd z1.d, p5/m, z1.d, z19.d \n\t"  \
-    "fsub z7.d, p5/m, z7.d, z19.d \n\t" \
-    "fadd z2.d, p5/m, z2.d, z20.d \n\t"  \
-    "fsub z8.d, p5/m, z8.d, z20.d \n\t" \
-    "fadd z3.d, p5/m, z3.d, z21.d \n\t"  \
-    "fsub z9.d, p5/m, z9.d, z21.d \n\t" \
-    "fadd z4.d, p5/m, z4.d, z22.d \n\t"  \
-    "fsub z10.d, p5/m, z10.d, z22.d \n\t" \
-    "fadd z5.d, p5/m, z5.d, z23.d \n\t"  \
-    "fsub z11.d, p5/m, z11.d, z23.d \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// ZERO_PSI
-#define ZERO_PSI_A64FXd  \
-asm ( \
-    "ptrue p5.d \n\t" \
-    "fmov z0.d , 0 \n\t" \
-    "fmov z1.d , 0 \n\t" \
-    "fmov z2.d , 0 \n\t" \
-    "fmov z3.d , 0 \n\t" \
-    "fmov z4.d , 0 \n\t" \
-    "fmov z5.d , 0 \n\t" \
-    "fmov z6.d , 0 \n\t" \
-    "fmov z7.d , 0 \n\t" \
-    "fmov z8.d , 0 \n\t" \
-    "fmov z9.d , 0 \n\t" \
-    "fmov z10.d , 0 \n\t" \
-    "fmov z11.d , 0 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// PREFETCH_RESULT_L2_STORE (prefetch store to L2)
-#define PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXd(base)  \
-{ \
-asm ( \
-    "prfd PSTL2STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PSTL2STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PSTL2STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_RESULT_L1_STORE (prefetch store to L1)
-#define PREFETCH_RESULT_L1_STORE_INTERNAL_A64FXd(base)  \
-{ \
-asm ( \
-    "prfd PSTL1STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PSTL1STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PSTL1STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// ADD_RESULT_INTERNAL
-#define ADD_RESULT_INTERNAL_A64FXd  \
-asm ( \
-    "fadd z0.d, p5/m, z0.d, z12.d \n\t"  \
-    "fadd z1.d, p5/m, z1.d, z13.d \n\t"  \
-    "fadd z2.d, p5/m, z2.d, z14.d \n\t"  \
-    "fadd z3.d, p5/m, z3.d, z15.d \n\t"  \
-    "fadd z4.d, p5/m, z4.d, z16.d \n\t"  \
-    "fadd z5.d, p5/m, z5.d, z17.d \n\t"  \
-    "fadd z6.d, p5/m, z6.d, z18.d \n\t"  \
-    "fadd z7.d, p5/m, z7.d, z19.d \n\t"  \
-    "fadd z8.d, p5/m, z8.d, z20.d \n\t"  \
-    "fadd z9.d, p5/m, z9.d, z21.d \n\t"  \
-    "fadd z10.d, p5/m, z10.d, z22.d \n\t"  \
-    "fadd z11.d, p5/m, z11.d, z23.d \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-

Grid/simd/Fujitsu_A64FX_asm_single.h

@@ -1,779 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: Fujitsu_A64FX_asm_single.h
-
-    Copyright (C) 2020
-
-Author: Nils Meyer <nils.meyer@ur.de>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#define LOAD_CHIMU(base)               LOAD_CHIMU_INTERLEAVED_A64FXf(base)  
-#define PREFETCH_CHIMU_L1(A)           PREFETCH_CHIMU_L1_INTERNAL_A64FXf(A)  
-#define PREFETCH_GAUGE_L1(A)           PREFETCH_GAUGE_L1_INTERNAL_A64FXf(A)  
-#define PREFETCH_CHIMU_L2(A)           PREFETCH_CHIMU_L2_INTERNAL_A64FXf(A)  
-#define PREFETCH_GAUGE_L2(A)           PREFETCH_GAUGE_L2_INTERNAL_A64FXf(A)  
-#define PF_GAUGE(A)  
-#define PREFETCH_RESULT_L2_STORE(A)    PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXf(A)  
-#define PREFETCH_RESULT_L1_STORE(A)    PREFETCH_RESULT_L1_STORE_INTERNAL_A64FXf(A)  
-#define PREFETCH1_CHIMU(A)             PREFETCH_CHIMU_L1(A)  
-#define PREFETCH_CHIMU(A)              PREFETCH_CHIMU_L1(A)  
-#define LOCK_GAUGE(A)  
-#define UNLOCK_GAUGE(A)  
-#define MASK_REGS                      DECLARATIONS_A64FXf  
-#define SAVE_RESULT(A,B)               RESULT_A64FXf(A); PREFETCH_RESULT_L2_STORE(B)  
-#define MULT_2SPIN_1(Dir)              MULT_2SPIN_1_A64FXf(Dir)  
-#define MULT_2SPIN_2                   MULT_2SPIN_2_A64FXf  
-#define LOAD_CHI(base)                 LOAD_CHI_A64FXf(base)  
-#define ADD_RESULT(base,basep)         LOAD_CHIMU(base); ADD_RESULT_INTERNAL_A64FXf; RESULT_A64FXf(base)  
-#define XP_PROJ                        XP_PROJ_A64FXf  
-#define YP_PROJ                        YP_PROJ_A64FXf  
-#define ZP_PROJ                        ZP_PROJ_A64FXf  
-#define TP_PROJ                        TP_PROJ_A64FXf  
-#define XM_PROJ                        XM_PROJ_A64FXf  
-#define YM_PROJ                        YM_PROJ_A64FXf  
-#define ZM_PROJ                        ZM_PROJ_A64FXf  
-#define TM_PROJ                        TM_PROJ_A64FXf  
-#define XP_RECON                       XP_RECON_A64FXf  
-#define XM_RECON                       XM_RECON_A64FXf  
-#define XM_RECON_ACCUM                 XM_RECON_ACCUM_A64FXf  
-#define YM_RECON_ACCUM                 YM_RECON_ACCUM_A64FXf  
-#define ZM_RECON_ACCUM                 ZM_RECON_ACCUM_A64FXf  
-#define TM_RECON_ACCUM                 TM_RECON_ACCUM_A64FXf  
-#define XP_RECON_ACCUM                 XP_RECON_ACCUM_A64FXf  
-#define YP_RECON_ACCUM                 YP_RECON_ACCUM_A64FXf  
-#define ZP_RECON_ACCUM                 ZP_RECON_ACCUM_A64FXf  
-#define TP_RECON_ACCUM                 TP_RECON_ACCUM_A64FXf  
-#define PERMUTE_DIR0                   0  
-#define PERMUTE_DIR1                   1  
-#define PERMUTE_DIR2                   2  
-#define PERMUTE_DIR3                   3  
-#define PERMUTE                        PERMUTE_A64FXf;  
-#define LOAD_TABLE(Dir)                if (Dir == 0) { LOAD_TABLE0; } else if (Dir == 1) { LOAD_TABLE1 } else if (Dir == 2) { LOAD_TABLE2; } else if (Dir == 3) { LOAD_TABLE3; }  
-#define MAYBEPERM(A,perm)              if (perm) { PERMUTE; }  
-// DECLARATIONS
-#define DECLARATIONS_A64FXf  \
-    const uint32_t lut[4][16] = { \
-        {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7}, \
-        {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11}, \
-        {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13}, \
-        {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14} }; \
-asm ( \
-    "fmov z31.s , 0 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// RESULT
-#define RESULT_A64FXf(base)  \
-{ \
-asm ( \
-    "str z0, [%[storeptr], -6, mul vl] \n\t" \
-    "str z1, [%[storeptr], -5, mul vl] \n\t" \
-    "str z2, [%[storeptr], -4, mul vl] \n\t" \
-    "str z3, [%[storeptr], -3, mul vl] \n\t" \
-    "str z4, [%[storeptr], -2, mul vl] \n\t" \
-    "str z5, [%[storeptr], -1, mul vl] \n\t" \
-    "str z6, [%[storeptr], 0, mul vl] \n\t" \
-    "str z7, [%[storeptr], 1, mul vl] \n\t" \
-    "str z8, [%[storeptr], 2, mul vl] \n\t" \
-    "str z9, [%[storeptr], 3, mul vl] \n\t" \
-    "str z10, [%[storeptr], 4, mul vl] \n\t" \
-    "str z11, [%[storeptr], 5, mul vl] \n\t" \
-    :  \
-    : [storeptr] "r" (base + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_CHIMU_L2 (prefetch to L2)
-#define PREFETCH_CHIMU_L2_INTERNAL_A64FXf(base)  \
-{ \
-asm ( \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_CHIMU_L1 (prefetch to L1)
-#define PREFETCH_CHIMU_L1_INTERNAL_A64FXf(base)  \
-{ \
-asm ( \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_GAUGE_L2 (prefetch to L2)
-#define PREFETCH_GAUGE_L2_INTERNAL_A64FXf(A)  \
-{ \
-    const auto & ref(U[sUn](A)); uint64_t baseU = (uint64_t)&ref + 3 * 3 * 64; \
-asm ( \
-    "prfd PLDL2STRM, p5, [%[fetchptr], -4, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 12, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 16, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 20, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 24, mul vl] \n\t" \
-    "prfd PLDL2STRM, p5, [%[fetchptr], 28, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_GAUGE_L1 (prefetch to L1)
-#define PREFETCH_GAUGE_L1_INTERNAL_A64FXf(A)  \
-{ \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
-asm ( \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PLDL1STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHI
-#define LOAD_CHI_A64FXf(base)  \
-{ \
-asm ( \
-    "ldr z12, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], 2, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], 5, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHIMU
-#define LOAD_CHIMU_INTERLEAVED_A64FXf(base)  \
-{ \
-asm ( \
-    "ptrue p5.s \n\t" \
-    "ldr z12, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z21, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z18, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z22, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z19, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z23, [%[fetchptr], 5, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z20, [%[fetchptr], 2, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHIMU_0213
-#define LOAD_CHIMU_0213_A64FXf  \
-{ \
-    const SiteSpinor & ref(in[offset]); \
-asm ( \
-    "ptrue p5.s \n\t" \
-    "ldr z12, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z18, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z19, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z20, [%[fetchptr], 2, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z21, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z22, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z23, [%[fetchptr], 5, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (&ref[2][0]) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_CHIMU_0312
-#define LOAD_CHIMU_0312_A64FXf  \
-{ \
-    const SiteSpinor & ref(in[offset]); \
-asm ( \
-    "ptrue p5.s \n\t" \
-    "ldr z12, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z21, [%[fetchptr], 3, mul vl] \n\t" \
-    "ldr z13, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z22, [%[fetchptr], 4, mul vl] \n\t" \
-    "ldr z14, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z23, [%[fetchptr], 5, mul vl] \n\t" \
-    "ldr z15, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z18, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z16, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z19, [%[fetchptr], 1, mul vl] \n\t" \
-    "ldr z17, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z20, [%[fetchptr], 2, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (&ref[2][0]) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// LOAD_TABLE0
-#define LOAD_TABLE0  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (0) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_TABLE1
-#define LOAD_TABLE1  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (1) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_TABLE2
-#define LOAD_TABLE2  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (2) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_TABLE3
-#define LOAD_TABLE3  \
-asm ( \
-    "ldr z30, [%[tableptr], %[index], mul vl] \n\t" \
-    :  \
-    : [tableptr] "r" (&lut[0]),[index] "i" (3) \
-    : "memory","cc","p5","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// PERMUTE
-#define PERMUTE_A64FXf  \
-asm ( \
-    "tbl z12.s, { z12.s }, z30.s \n\t"  \
-    "tbl z13.s, { z13.s }, z30.s \n\t"  \
-    "tbl z14.s, { z14.s }, z30.s \n\t"  \
-    "tbl z15.s, { z15.s }, z30.s \n\t"  \
-    "tbl z16.s, { z16.s }, z30.s \n\t"  \
-    "tbl z17.s, { z17.s }, z30.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// LOAD_GAUGE
-#define LOAD_GAUGE  \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
-{ \
-asm ( \
-    "ptrue p5.s \n\t" \
-    "ldr z24, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z25, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z26, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z27, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z28, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z29, [%[fetchptr], 1, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// MULT_2SPIN
-#define MULT_2SPIN_1_A64FXf(A)  \
-{ \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
-asm ( \
-    "ldr z24, [%[fetchptr], -6, mul vl] \n\t" \
-    "ldr z25, [%[fetchptr], -3, mul vl] \n\t" \
-    "ldr z26, [%[fetchptr], 0, mul vl] \n\t" \
-    "ldr z27, [%[fetchptr], -5, mul vl] \n\t" \
-    "ldr z28, [%[fetchptr], -2, mul vl] \n\t" \
-    "ldr z29, [%[fetchptr], 1, mul vl] \n\t" \
-    "movprfx z18.s, p5/m, z31.s \n\t" \
-    "fcmla z18.s, p5/m, z24.s, z12.s, 0 \n\t" \
-    "movprfx z21.s, p5/m, z31.s \n\t" \
-    "fcmla z21.s, p5/m, z24.s, z15.s, 0 \n\t" \
-    "movprfx z19.s, p5/m, z31.s \n\t" \
-    "fcmla z19.s, p5/m, z25.s, z12.s, 0 \n\t" \
-    "movprfx z22.s, p5/m, z31.s \n\t" \
-    "fcmla z22.s, p5/m, z25.s, z15.s, 0 \n\t" \
-    "movprfx z20.s, p5/m, z31.s \n\t" \
-    "fcmla z20.s, p5/m, z26.s, z12.s, 0 \n\t" \
-    "movprfx z23.s, p5/m, z31.s \n\t" \
-    "fcmla z23.s, p5/m, z26.s, z15.s, 0 \n\t" \
-    "fcmla z18.s, p5/m, z24.s, z12.s, 90 \n\t" \
-    "fcmla z21.s, p5/m, z24.s, z15.s, 90 \n\t" \
-    "fcmla z19.s, p5/m, z25.s, z12.s, 90 \n\t" \
-    "fcmla z22.s, p5/m, z25.s, z15.s, 90 \n\t" \
-    "fcmla z20.s, p5/m, z26.s, z12.s, 90 \n\t" \
-    "fcmla z23.s, p5/m, z26.s, z15.s, 90 \n\t" \
-    "ldr z24, [%[fetchptr], -4, mul vl] \n\t" \
-    "ldr z25, [%[fetchptr], -1, mul vl] \n\t" \
-    "ldr z26, [%[fetchptr], 2, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (baseU + 2 * 3 * 64) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// MULT_2SPIN_BACKEND
-#define MULT_2SPIN_2_A64FXf  \
-{ \
-asm ( \
-    "fcmla z18.s, p5/m, z27.s, z13.s, 0 \n\t" \
-    "fcmla z21.s, p5/m, z27.s, z16.s, 0 \n\t" \
-    "fcmla z19.s, p5/m, z28.s, z13.s, 0 \n\t" \
-    "fcmla z22.s, p5/m, z28.s, z16.s, 0 \n\t" \
-    "fcmla z20.s, p5/m, z29.s, z13.s, 0 \n\t" \
-    "fcmla z23.s, p5/m, z29.s, z16.s, 0 \n\t" \
-    "fcmla z18.s, p5/m, z27.s, z13.s, 90 \n\t" \
-    "fcmla z21.s, p5/m, z27.s, z16.s, 90 \n\t" \
-    "fcmla z19.s, p5/m, z28.s, z13.s, 90 \n\t" \
-    "fcmla z22.s, p5/m, z28.s, z16.s, 90 \n\t" \
-    "fcmla z20.s, p5/m, z29.s, z13.s, 90 \n\t" \
-    "fcmla z23.s, p5/m, z29.s, z16.s, 90 \n\t" \
-    "fcmla z18.s, p5/m, z24.s, z14.s, 0 \n\t" \
-    "fcmla z21.s, p5/m, z24.s, z17.s, 0 \n\t" \
-    "fcmla z19.s, p5/m, z25.s, z14.s, 0 \n\t" \
-    "fcmla z22.s, p5/m, z25.s, z17.s, 0 \n\t" \
-    "fcmla z20.s, p5/m, z26.s, z14.s, 0 \n\t" \
-    "fcmla z23.s, p5/m, z26.s, z17.s, 0 \n\t" \
-    "fcmla z18.s, p5/m, z24.s, z14.s, 90 \n\t" \
-    "fcmla z21.s, p5/m, z24.s, z17.s, 90 \n\t" \
-    "fcmla z19.s, p5/m, z25.s, z14.s, 90 \n\t" \
-    "fcmla z22.s, p5/m, z25.s, z17.s, 90 \n\t" \
-    "fcmla z20.s, p5/m, z26.s, z14.s, 90 \n\t" \
-    "fcmla z23.s, p5/m, z26.s, z17.s, 90 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XP_PROJ
-#define XP_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fcadd z12.s, p5/m, z12.s, z21.s, 90 \n\t" \
-    "fcadd z13.s, p5/m, z13.s, z22.s, 90 \n\t" \
-    "fcadd z14.s, p5/m, z14.s, z23.s, 90 \n\t" \
-    "fcadd z15.s, p5/m, z15.s, z18.s, 90 \n\t" \
-    "fcadd z16.s, p5/m, z16.s, z19.s, 90 \n\t" \
-    "fcadd z17.s, p5/m, z17.s, z20.s, 90 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XP_RECON
-#define XP_RECON_A64FXf  \
-asm ( \
-    "movprfx z6.s, p5/m, z31.s \n\t" \
-    "fcadd z6.s, p5/m, z6.s, z21.s, 270 \n\t" \
-    "movprfx z7.s, p5/m, z31.s \n\t" \
-    "fcadd z7.s, p5/m, z7.s, z22.s, 270 \n\t" \
-    "movprfx z8.s, p5/m, z31.s \n\t" \
-    "fcadd z8.s, p5/m, z8.s, z23.s, 270 \n\t" \
-    "movprfx z9.s, p5/m, z31.s \n\t" \
-    "fcadd z9.s, p5/m, z9.s, z18.s, 270 \n\t" \
-    "movprfx z10.s, p5/m, z31.s \n\t" \
-    "fcadd z10.s, p5/m, z10.s, z19.s, 270 \n\t" \
-    "movprfx z11.s, p5/m, z31.s \n\t" \
-    "fcadd z11.s, p5/m, z11.s, z20.s, 270 \n\t" \
-    "mov z0.s, p5/m, z18.s \n\t" \
-    "mov z1.s, p5/m, z19.s \n\t" \
-    "mov z2.s, p5/m, z20.s \n\t" \
-    "mov z3.s, p5/m, z21.s \n\t" \
-    "mov z4.s, p5/m, z22.s \n\t" \
-    "mov z5.s, p5/m, z23.s \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// XP_RECON_ACCUM
-#define XP_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fcadd z9.s, p5/m, z9.s, z18.s, 270 \n\t" \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fcadd z10.s, p5/m, z10.s, z19.s, 270 \n\t" \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fcadd z11.s, p5/m, z11.s, z20.s, 270 \n\t" \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fcadd z6.s, p5/m, z6.s, z21.s, 270 \n\t" \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fcadd z7.s, p5/m, z7.s, z22.s, 270 \n\t" \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fcadd z8.s, p5/m, z8.s, z23.s, 270 \n\t" \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YP_PROJ
-#define YP_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fsub z12.s, p5/m, z12.s, z21.s \n\t" \
-    "fsub z13.s, p5/m, z13.s, z22.s \n\t" \
-    "fsub z14.s, p5/m, z14.s, z23.s \n\t" \
-    "fadd z15.s, p5/m, z15.s, z18.s \n\t"  \
-    "fadd z16.s, p5/m, z16.s, z19.s \n\t"  \
-    "fadd z17.s, p5/m, z17.s, z20.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// ZP_PROJ
-#define ZP_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fcadd z12.s, p5/m, z12.s, z18.s, 90 \n\t" \
-    "fcadd z13.s, p5/m, z13.s, z19.s, 90 \n\t" \
-    "fcadd z14.s, p5/m, z14.s, z20.s, 90 \n\t" \
-    "fcadd z15.s, p5/m, z15.s, z21.s, 270 \n\t" \
-    "fcadd z16.s, p5/m, z16.s, z22.s, 270 \n\t" \
-    "fcadd z17.s, p5/m, z17.s, z23.s, 270 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// TP_PROJ
-#define TP_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fadd z12.s, p5/m, z12.s, z18.s \n\t"  \
-    "fadd z13.s, p5/m, z13.s, z19.s \n\t"  \
-    "fadd z14.s, p5/m, z14.s, z20.s \n\t"  \
-    "fadd z15.s, p5/m, z15.s, z21.s \n\t"  \
-    "fadd z16.s, p5/m, z16.s, z22.s \n\t"  \
-    "fadd z17.s, p5/m, z17.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XM_PROJ
-#define XM_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fcadd z12.s, p5/m, z12.s, z21.s, 270 \n\t" \
-    "fcadd z13.s, p5/m, z13.s, z22.s, 270 \n\t" \
-    "fcadd z14.s, p5/m, z14.s, z23.s, 270 \n\t" \
-    "fcadd z15.s, p5/m, z15.s, z18.s, 270 \n\t" \
-    "fcadd z16.s, p5/m, z16.s, z19.s, 270 \n\t" \
-    "fcadd z17.s, p5/m, z17.s, z20.s, 270 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XM_RECON
-#define XM_RECON_A64FXf  \
-asm ( \
-    "movprfx z6.s, p5/m, z31.s \n\t" \
-    "fcadd z6.s, p5/m, z6.s, z21.s, 90 \n\t" \
-    "movprfx z7.s, p5/m, z31.s \n\t" \
-    "fcadd z7.s, p5/m, z7.s, z22.s, 90 \n\t" \
-    "movprfx z8.s, p5/m, z31.s \n\t" \
-    "fcadd z8.s, p5/m, z8.s, z23.s, 90 \n\t" \
-    "movprfx z9.s, p5/m, z31.s \n\t" \
-    "fcadd z9.s, p5/m, z9.s, z18.s, 90 \n\t" \
-    "movprfx z10.s, p5/m, z31.s \n\t" \
-    "fcadd z10.s, p5/m, z10.s, z19.s, 90 \n\t" \
-    "movprfx z11.s, p5/m, z31.s \n\t" \
-    "fcadd z11.s, p5/m, z11.s, z20.s, 90 \n\t" \
-    "mov z0.s, p5/m, z18.s \n\t" \
-    "mov z1.s, p5/m, z19.s \n\t" \
-    "mov z2.s, p5/m, z20.s \n\t" \
-    "mov z3.s, p5/m, z21.s \n\t" \
-    "mov z4.s, p5/m, z22.s \n\t" \
-    "mov z5.s, p5/m, z23.s \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YM_PROJ
-#define YM_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fadd z12.s, p5/m, z12.s, z21.s \n\t"  \
-    "fadd z13.s, p5/m, z13.s, z22.s \n\t"  \
-    "fadd z14.s, p5/m, z14.s, z23.s \n\t"  \
-    "fsub z15.s, p5/m, z15.s, z18.s \n\t" \
-    "fsub z16.s, p5/m, z16.s, z19.s \n\t" \
-    "fsub z17.s, p5/m, z17.s, z20.s \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// ZM_PROJ
-#define ZM_PROJ_A64FXf  \
-{ \
-asm ( \
-    "fcadd z12.s, p5/m, z12.s, z18.s, 270 \n\t" \
-    "fcadd z13.s, p5/m, z13.s, z19.s, 270 \n\t" \
-    "fcadd z14.s, p5/m, z14.s, z20.s, 270 \n\t" \
-    "fcadd z15.s, p5/m, z15.s, z21.s, 90 \n\t" \
-    "fcadd z16.s, p5/m, z16.s, z22.s, 90 \n\t" \
-    "fcadd z17.s, p5/m, z17.s, z23.s, 90 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// TM_PROJ
-#define TM_PROJ_A64FXf  \
-{ \
-asm ( \
-    "ptrue p5.s \n\t" \
-    "fsub z12.s, p5/m, z12.s, z18.s \n\t" \
-    "fsub z13.s, p5/m, z13.s, z19.s \n\t" \
-    "fsub z14.s, p5/m, z14.s, z20.s \n\t" \
-    "fsub z15.s, p5/m, z15.s, z21.s \n\t" \
-    "fsub z16.s, p5/m, z16.s, z22.s \n\t" \
-    "fsub z17.s, p5/m, z17.s, z23.s \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); \
-}
-// XM_RECON_ACCUM
-#define XM_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fcadd z9.s, p5/m, z9.s, z18.s, 90 \n\t" \
-    "fcadd z10.s, p5/m, z10.s, z19.s, 90 \n\t" \
-    "fcadd z11.s, p5/m, z11.s, z20.s, 90 \n\t" \
-    "fcadd z6.s, p5/m, z6.s, z21.s, 90 \n\t" \
-    "fcadd z7.s, p5/m, z7.s, z22.s, 90 \n\t" \
-    "fcadd z8.s, p5/m, z8.s, z23.s, 90 \n\t" \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YP_RECON_ACCUM
-#define YP_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fsub z9.s, p5/m, z9.s, z18.s \n\t" \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fsub z10.s, p5/m, z10.s, z19.s \n\t" \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fsub z11.s, p5/m, z11.s, z20.s \n\t" \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fadd z6.s, p5/m, z6.s, z21.s \n\t"  \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fadd z7.s, p5/m, z7.s, z22.s \n\t"  \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    "fadd z8.s, p5/m, z8.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// YM_RECON_ACCUM
-#define YM_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fadd z9.s, p5/m, z9.s, z18.s \n\t"  \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fadd z10.s, p5/m, z10.s, z19.s \n\t"  \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fadd z11.s, p5/m, z11.s, z20.s \n\t"  \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fsub z6.s, p5/m, z6.s, z21.s \n\t" \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fsub z7.s, p5/m, z7.s, z22.s \n\t" \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    "fsub z8.s, p5/m, z8.s, z23.s \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// ZP_RECON_ACCUM
-#define ZP_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fcadd z6.s, p5/m, z6.s, z18.s, 270 \n\t" \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fcadd z7.s, p5/m, z7.s, z19.s, 270 \n\t" \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fcadd z8.s, p5/m, z8.s, z20.s, 270 \n\t" \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fcadd z9.s, p5/m, z9.s, z21.s, 90 \n\t" \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fcadd z10.s, p5/m, z10.s, z22.s, 90 \n\t" \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fcadd z11.s, p5/m, z11.s, z23.s, 90 \n\t" \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// ZM_RECON_ACCUM
-#define ZM_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fcadd z6.s, p5/m, z6.s, z18.s, 90 \n\t" \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fcadd z7.s, p5/m, z7.s, z19.s, 90 \n\t" \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fcadd z8.s, p5/m, z8.s, z20.s, 90 \n\t" \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fcadd z9.s, p5/m, z9.s, z21.s, 270 \n\t" \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fcadd z10.s, p5/m, z10.s, z22.s, 270 \n\t" \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fcadd z11.s, p5/m, z11.s, z23.s, 270 \n\t" \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// TP_RECON_ACCUM
-#define TP_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fadd z6.s, p5/m, z6.s, z18.s \n\t"  \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fadd z7.s, p5/m, z7.s, z19.s \n\t"  \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fadd z8.s, p5/m, z8.s, z20.s \n\t"  \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fadd z9.s, p5/m, z9.s, z21.s \n\t"  \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fadd z10.s, p5/m, z10.s, z22.s \n\t"  \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    "fadd z11.s, p5/m, z11.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// TM_RECON_ACCUM
-#define TM_RECON_ACCUM_A64FXf  \
-asm ( \
-    "fadd z0.s, p5/m, z0.s, z18.s \n\t"  \
-    "fsub z6.s, p5/m, z6.s, z18.s \n\t" \
-    "fadd z1.s, p5/m, z1.s, z19.s \n\t"  \
-    "fsub z7.s, p5/m, z7.s, z19.s \n\t" \
-    "fadd z2.s, p5/m, z2.s, z20.s \n\t"  \
-    "fsub z8.s, p5/m, z8.s, z20.s \n\t" \
-    "fadd z3.s, p5/m, z3.s, z21.s \n\t"  \
-    "fsub z9.s, p5/m, z9.s, z21.s \n\t" \
-    "fadd z4.s, p5/m, z4.s, z22.s \n\t"  \
-    "fsub z10.s, p5/m, z10.s, z22.s \n\t" \
-    "fadd z5.s, p5/m, z5.s, z23.s \n\t"  \
-    "fsub z11.s, p5/m, z11.s, z23.s \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// ZERO_PSI
-#define ZERO_PSI_A64FXf  \
-asm ( \
-    "ptrue p5.s \n\t" \
-    "fmov z0.s , 0 \n\t" \
-    "fmov z1.s , 0 \n\t" \
-    "fmov z2.s , 0 \n\t" \
-    "fmov z3.s , 0 \n\t" \
-    "fmov z4.s , 0 \n\t" \
-    "fmov z5.s , 0 \n\t" \
-    "fmov z6.s , 0 \n\t" \
-    "fmov z7.s , 0 \n\t" \
-    "fmov z8.s , 0 \n\t" \
-    "fmov z9.s , 0 \n\t" \
-    "fmov z10.s , 0 \n\t" \
-    "fmov z11.s , 0 \n\t" \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-
-// PREFETCH_RESULT_L2_STORE (prefetch store to L2)
-#define PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXf(base)  \
-{ \
-asm ( \
-    "prfd PSTL2STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PSTL2STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PSTL2STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// PREFETCH_RESULT_L1_STORE (prefetch store to L1)
-#define PREFETCH_RESULT_L1_STORE_INTERNAL_A64FXf(base)  \
-{ \
-asm ( \
-    "prfd PSTL1STRM, p5, [%[fetchptr], 0, mul vl] \n\t" \
-    "prfd PSTL1STRM, p5, [%[fetchptr], 4, mul vl] \n\t" \
-    "prfd PSTL1STRM, p5, [%[fetchptr], 8, mul vl] \n\t" \
-    :  \
-    : [fetchptr] "r" (base) \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31","memory" \
-); \
-}
-// ADD_RESULT_INTERNAL
-#define ADD_RESULT_INTERNAL_A64FXf  \
-asm ( \
-    "fadd z0.s, p5/m, z0.s, z12.s \n\t"  \
-    "fadd z1.s, p5/m, z1.s, z13.s \n\t"  \
-    "fadd z2.s, p5/m, z2.s, z14.s \n\t"  \
-    "fadd z3.s, p5/m, z3.s, z15.s \n\t"  \
-    "fadd z4.s, p5/m, z4.s, z16.s \n\t"  \
-    "fadd z5.s, p5/m, z5.s, z17.s \n\t"  \
-    "fadd z6.s, p5/m, z6.s, z18.s \n\t"  \
-    "fadd z7.s, p5/m, z7.s, z19.s \n\t"  \
-    "fadd z8.s, p5/m, z8.s, z20.s \n\t"  \
-    "fadd z9.s, p5/m, z9.s, z21.s \n\t"  \
-    "fadd z10.s, p5/m, z10.s, z22.s \n\t"  \
-    "fadd z11.s, p5/m, z11.s, z23.s \n\t"  \
-    :  \
-    :  \
-    : "p5","cc","z0","z1","z2","z3","z4","z5","z6","z7","z8","z9","z10","z11","z12","z13","z14","z15","z16","z17","z18","z19","z20","z21","z22","z23","z24","z25","z26","z27","z28","z29","z30","z31" \
-); 
-

Grid/simd/Fujitsu_A64FX_intrin_double.h

@@ -38,10 +38,11 @@ Author: Nils Meyer <nils.meyer@ur.de>
 #define LOCK_GAUGE(A)  
 #define UNLOCK_GAUGE(A)  
 #define MASK_REGS                      DECLARATIONS_A64FXd  
-#define SAVE_RESULT(A,B)               RESULT_A64FXd(A); PREFETCH_RESULT_L2_STORE(B)  
+#define SAVE_RESULT(A,B)               RESULT_A64FXd(A);  
 #define MULT_2SPIN_1(Dir)              MULT_2SPIN_1_A64FXd(Dir)  
 #define MULT_2SPIN_2                   MULT_2SPIN_2_A64FXd  
 #define LOAD_CHI(base)                 LOAD_CHI_A64FXd(base)  
+#define ZERO_PSI                       ZERO_PSI_A64FXd  
 #define ADD_RESULT(base,basep)         LOAD_CHIMU(base); ADD_RESULT_INTERNAL_A64FXd; RESULT_A64FXd(base)  
 #define XP_PROJ                        XP_PROJ_A64FXd  
 #define YP_PROJ                        YP_PROJ_A64FXd  
@@ -70,6 +71,7 @@ Author: Nils Meyer <nils.meyer@ur.de>
 #define MAYBEPERM(Dir,perm)            if (Dir != 3) { if (perm) { PERMUTE; } }  
 // DECLARATIONS
 #define DECLARATIONS_A64FXd  \
+    uint64_t baseU; \
     const uint64_t lut[4][8] = { \
         {4, 5, 6, 7, 0, 1, 2, 3}, \
         {2, 3, 0, 1, 6, 7, 4, 5}, \
@@ -126,114 +128,114 @@ Author: Nils Meyer <nils.meyer@ur.de>
 // RESULT
 #define RESULT_A64FXd(base)  \
 { \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + -6 * 64), result_00);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(-6), result_00);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + -5 * 64), result_01);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(-5), result_01);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + -4 * 64), result_02);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(-4), result_02);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + -3 * 64), result_10);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(-3), result_10);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + -2 * 64), result_11);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(-2), result_11);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + -1 * 64), result_12);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(-1), result_12);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + 0 * 64), result_20);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(0), result_20);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + 1 * 64), result_21);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(1), result_21);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + 2 * 64), result_22);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(2), result_22);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + 3 * 64), result_30);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(3), result_30);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + 4 * 64), result_31);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(4), result_31);  \
-    svst1(pg1, (float64_t*)(base + 2 * 3 * 64 + 5 * 64), result_32);  \
+    svst1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64),(int64_t)(5), result_32);  \
 }
 // PREFETCH_CHIMU_L2 (prefetch to L2)
 #define PREFETCH_CHIMU_L2_INTERNAL_A64FXd(base)  \
 { \
-    svprfd(pg1, (int64_t*)(base + 0), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(0), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(base + 256), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(4), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(base + 512), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(8), SV_PLDL2STRM); \
 }
 // PREFETCH_CHIMU_L1 (prefetch to L1)
 #define PREFETCH_CHIMU_L1_INTERNAL_A64FXd(base)  \
 { \
-    svprfd(pg1, (int64_t*)(base + 0), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(0), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(base + 256), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(4), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(base + 512), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(8), SV_PLDL1STRM); \
 }
 // PREFETCH_GAUGE_L2 (prefetch to L2)
 #define PREFETCH_GAUGE_L2_INTERNAL_A64FXd(A)  \
 { \
-    const auto & ref(U[sUn](A)); uint64_t baseU = (uint64_t)&ref + 3 * 3 * 64; \
+    const auto & ref(U[sUn](A)); baseU = (uint64_t)&ref + 3 * 3 * 64; \
-    svprfd(pg1, (int64_t*)(baseU + -256), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(-4), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 0), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(0), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 256), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(4), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 512), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(8), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 768), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(12), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1024), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(16), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1280), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(20), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1536), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(24), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1792), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(28), SV_PLDL2STRM); \
 }
 // PREFETCH_GAUGE_L1 (prefetch to L1)
 #define PREFETCH_GAUGE_L1_INTERNAL_A64FXd(A)  \
 { \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
+    const auto & ref(U[sU](A)); baseU = (uint64_t)&ref; \
-    svprfd(pg1, (int64_t*)(baseU + 0), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(0), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 256), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(4), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 512), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(8), SV_PLDL1STRM); \
 }
 // LOAD_CHI
 #define LOAD_CHI_A64FXd(base)  \
 { \
-    Chi_00 = svld1(pg1, (float64_t*)(base + 0 * 64));  \
+    Chi_00 = svld1_vnum(pg1, (float64_t*)(base), (int64_t)(0));  \
-    Chi_01 = svld1(pg1, (float64_t*)(base + 1 * 64));  \
+    Chi_01 = svld1_vnum(pg1, (float64_t*)(base), (int64_t)(1));  \
-    Chi_02 = svld1(pg1, (float64_t*)(base + 2 * 64));  \
+    Chi_02 = svld1_vnum(pg1, (float64_t*)(base), (int64_t)(2));  \
-    Chi_10 = svld1(pg1, (float64_t*)(base + 3 * 64));  \
+    Chi_10 = svld1_vnum(pg1, (float64_t*)(base), (int64_t)(3));  \
-    Chi_11 = svld1(pg1, (float64_t*)(base + 4 * 64));  \
+    Chi_11 = svld1_vnum(pg1, (float64_t*)(base), (int64_t)(4));  \
-    Chi_12 = svld1(pg1, (float64_t*)(base + 5 * 64));  \
+    Chi_12 = svld1_vnum(pg1, (float64_t*)(base), (int64_t)(5));  \
 }
 // LOAD_CHIMU
 #define LOAD_CHIMU_INTERLEAVED_A64FXd(base)  \
 { \
-    Chimu_00 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -6 * 64));  \
+    Chimu_00 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-6));  \
-    Chimu_30 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 3 * 64));  \
+    Chimu_30 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(3));  \
-    Chimu_10 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -3 * 64));  \
+    Chimu_10 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-3));  \
-    Chimu_20 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 0 * 64));  \
+    Chimu_20 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(0));  \
-    Chimu_01 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -5 * 64));  \
+    Chimu_01 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-5));  \
-    Chimu_31 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 4 * 64));  \
+    Chimu_31 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(4));  \
-    Chimu_11 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -2 * 64));  \
+    Chimu_11 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-2));  \
-    Chimu_21 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 1 * 64));  \
+    Chimu_21 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(1));  \
-    Chimu_02 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -4 * 64));  \
+    Chimu_02 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-4));  \
-    Chimu_32 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 5 * 64));  \
+    Chimu_32 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(5));  \
-    Chimu_12 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -1 * 64));  \
+    Chimu_12 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-1));  \
-    Chimu_22 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 2 * 64));  \
+    Chimu_22 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(2));  \
 }
 // LOAD_CHIMU_0213
 #define LOAD_CHIMU_0213_A64FXd  \
 { \
     const SiteSpinor & ref(in[offset]); \
-    Chimu_00 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -6 * 64));  \
+    Chimu_00 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-6));  \
-    Chimu_20 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 0 * 64));  \
+    Chimu_20 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(0));  \
-    Chimu_01 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -5 * 64));  \
+    Chimu_01 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-5));  \
-    Chimu_21 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 1 * 64));  \
+    Chimu_21 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(1));  \
-    Chimu_02 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -4 * 64));  \
+    Chimu_02 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-4));  \
-    Chimu_22 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 2 * 64));  \
+    Chimu_22 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(2));  \
-    Chimu_10 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -3 * 64));  \
+    Chimu_10 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-3));  \
-    Chimu_30 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 3 * 64));  \
+    Chimu_30 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(3));  \
-    Chimu_11 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -2 * 64));  \
+    Chimu_11 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-2));  \
-    Chimu_31 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 4 * 64));  \
+    Chimu_31 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(4));  \
-    Chimu_12 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -1 * 64));  \
+    Chimu_12 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-1));  \
-    Chimu_32 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 5 * 64));  \
+    Chimu_32 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(5));  \
 }
 // LOAD_CHIMU_0312
 #define LOAD_CHIMU_0312_A64FXd  \
 { \
     const SiteSpinor & ref(in[offset]); \
-    Chimu_00 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -6 * 64));  \
+    Chimu_00 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-6));  \
-    Chimu_30 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 3 * 64));  \
+    Chimu_30 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(3));  \
-    Chimu_01 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -5 * 64));  \
+    Chimu_01 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-5));  \
-    Chimu_31 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 4 * 64));  \
+    Chimu_31 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(4));  \
-    Chimu_02 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -4 * 64));  \
+    Chimu_02 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-4));  \
-    Chimu_32 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 5 * 64));  \
+    Chimu_32 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(5));  \
-    Chimu_10 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -3 * 64));  \
+    Chimu_10 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-3));  \
-    Chimu_20 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 0 * 64));  \
+    Chimu_20 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(0));  \
-    Chimu_11 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -2 * 64));  \
+    Chimu_11 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-2));  \
-    Chimu_21 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 1 * 64));  \
+    Chimu_21 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(1));  \
-    Chimu_12 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + -1 * 64));  \
+    Chimu_12 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(-1));  \
-    Chimu_22 = svld1(pg1, (float64_t*)(base + 2 * 3 * 64 + 2 * 64));  \
+    Chimu_22 = svld1_vnum(pg1, (float64_t*)(base + 2 * 3 * 64), (int64_t)(2));  \
 }
 // LOAD_TABLE0
 #define LOAD_TABLE0  \
@@ -261,26 +263,26 @@ Author: Nils Meyer <nils.meyer@ur.de>
     Chi_12 = svtbl(Chi_12, table0);    
 
 // LOAD_GAUGE
-#define LOAD_GAUGE  \
+#define LOAD_GAUGE(A)  \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
 { \
-    U_00 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -6 * 64));  \
+    const auto & ref(U[sU](A)); baseU = (uint64_t)&ref; \
-    U_10 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -3 * 64));  \
+    U_00 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-6));  \
-    U_20 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + 0 * 64));  \
+    U_10 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-3));  \
-    U_01 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -5 * 64));  \
+    U_20 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(0));  \
-    U_11 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -2 * 64));  \
+    U_01 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-5));  \
-    U_21 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + 1 * 64));  \
+    U_11 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-2));  \
+    U_21 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(1));  \
 }
 // MULT_2SPIN
 #define MULT_2SPIN_1_A64FXd(A)  \
 { \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
+    const auto & ref(U[sU](A)); baseU = (uint64_t)&ref; \
-    U_00 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -6 * 64));  \
+    U_00 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-6));  \
-    U_10 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -3 * 64));  \
+    U_10 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-3));  \
-    U_20 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + 0 * 64));  \
+    U_20 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(0));  \
-    U_01 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -5 * 64));  \
+    U_01 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-5));  \
-    U_11 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -2 * 64));  \
+    U_11 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-2));  \
-    U_21 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + 1 * 64));  \
+    U_21 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(1));  \
     UChi_00 = svcmla_x(pg1, zero0, U_00, Chi_00, 0); \
     UChi_10 = svcmla_x(pg1, zero0, U_00, Chi_10, 0); \
     UChi_01 = svcmla_x(pg1, zero0, U_10, Chi_00, 0); \
@@ -293,9 +295,9 @@ Author: Nils Meyer <nils.meyer@ur.de>
     UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_10, 90); \
     UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_00, 90); \
     UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_10, 90); \
-    U_00 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -4 * 64));  \
+    U_00 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-4));  \
-    U_10 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -1 * 64));  \
+    U_10 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(-1));  \
-    U_20 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + 2 * 64));  \
+    U_20 = svld1_vnum(pg1, (float64_t*)(baseU + 2 * 3 * 64), (int64_t)(2));  \
 }
 // MULT_2SPIN_BACKEND
 #define MULT_2SPIN_2_A64FXd  \
@@ -570,12 +572,12 @@ Author: Nils Meyer <nils.meyer@ur.de>
     result_31 = svdup_f64(0.); \
     result_32 = svdup_f64(0.); 
 
-// PREFETCH_RESULT_L2_STORE (prefetch store to L2)
+// PREFETCH_RESULT_L2_STORE (uses DC ZVA for cache line zeroing)
 #define PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXd(base)  \
 { \
-    svprfd(pg1, (int64_t*)(base + 0), SV_PSTL2STRM); \
+    asm( "dc zva, %[fetchptr] \n\t" : : [fetchptr] "r" (base + 256 * 0) : "memory" ); \
-    svprfd(pg1, (int64_t*)(base + 256), SV_PSTL2STRM); \
+    asm( "dc zva, %[fetchptr] \n\t" : : [fetchptr] "r" (base + 256 * 1) : "memory" ); \
-    svprfd(pg1, (int64_t*)(base + 512), SV_PSTL2STRM); \
+    asm( "dc zva, %[fetchptr] \n\t" : : [fetchptr] "r" (base + 256 * 2) : "memory" ); \
 }
 // PREFETCH_RESULT_L1_STORE (prefetch store to L1)
 #define PREFETCH_RESULT_L1_STORE_INTERNAL_A64FXd(base)  \

Grid/simd/Fujitsu_A64FX_intrin_single.h

@@ -38,10 +38,11 @@ Author: Nils Meyer <nils.meyer@ur.de>
 #define LOCK_GAUGE(A)  
 #define UNLOCK_GAUGE(A)  
 #define MASK_REGS                      DECLARATIONS_A64FXf  
-#define SAVE_RESULT(A,B)               RESULT_A64FXf(A); PREFETCH_RESULT_L2_STORE(B)  
+#define SAVE_RESULT(A,B)               RESULT_A64FXf(A);  
 #define MULT_2SPIN_1(Dir)              MULT_2SPIN_1_A64FXf(Dir)  
 #define MULT_2SPIN_2                   MULT_2SPIN_2_A64FXf  
 #define LOAD_CHI(base)                 LOAD_CHI_A64FXf(base)  
+#define ZERO_PSI                       ZERO_PSI_A64FXf  
 #define ADD_RESULT(base,basep)         LOAD_CHIMU(base); ADD_RESULT_INTERNAL_A64FXf; RESULT_A64FXf(base)  
 #define XP_PROJ                        XP_PROJ_A64FXf  
 #define YP_PROJ                        YP_PROJ_A64FXf  
@@ -70,6 +71,7 @@ Author: Nils Meyer <nils.meyer@ur.de>
 #define MAYBEPERM(A,perm)              if (perm) { PERMUTE; }  
 // DECLARATIONS
 #define DECLARATIONS_A64FXf  \
+    uint64_t baseU; \
     const uint32_t lut[4][16] = { \
         {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7}, \
         {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11}, \
@@ -126,114 +128,114 @@ Author: Nils Meyer <nils.meyer@ur.de>
 // RESULT
 #define RESULT_A64FXf(base)  \
 { \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + -6 * 64), result_00);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(-6), result_00);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + -5 * 64), result_01);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(-5), result_01);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + -4 * 64), result_02);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(-4), result_02);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + -3 * 64), result_10);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(-3), result_10);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + -2 * 64), result_11);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(-2), result_11);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + -1 * 64), result_12);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(-1), result_12);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + 0 * 64), result_20);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(0), result_20);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + 1 * 64), result_21);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(1), result_21);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + 2 * 64), result_22);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(2), result_22);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + 3 * 64), result_30);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(3), result_30);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + 4 * 64), result_31);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(4), result_31);  \
-    svst1(pg1, (float32_t*)(base + 2 * 3 * 64 + 5 * 64), result_32);  \
+    svst1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64),(int64_t)(5), result_32);  \
 }
 // PREFETCH_CHIMU_L2 (prefetch to L2)
 #define PREFETCH_CHIMU_L2_INTERNAL_A64FXf(base)  \
 { \
-    svprfd(pg1, (int64_t*)(base + 0), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(0), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(base + 256), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(4), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(base + 512), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(8), SV_PLDL2STRM); \
 }
 // PREFETCH_CHIMU_L1 (prefetch to L1)
 #define PREFETCH_CHIMU_L1_INTERNAL_A64FXf(base)  \
 { \
-    svprfd(pg1, (int64_t*)(base + 0), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(0), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(base + 256), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(4), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(base + 512), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(base), (int64_t)(8), SV_PLDL1STRM); \
 }
 // PREFETCH_GAUGE_L2 (prefetch to L2)
 #define PREFETCH_GAUGE_L2_INTERNAL_A64FXf(A)  \
 { \
-    const auto & ref(U[sUn](A)); uint64_t baseU = (uint64_t)&ref + 3 * 3 * 64; \
+    const auto & ref(U[sUn](A)); baseU = (uint64_t)&ref + 3 * 3 * 64; \
-    svprfd(pg1, (int64_t*)(baseU + -256), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(-4), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 0), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(0), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 256), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(4), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 512), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(8), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 768), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(12), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1024), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(16), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1280), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(20), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1536), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(24), SV_PLDL2STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 1792), SV_PLDL2STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(28), SV_PLDL2STRM); \
 }
 // PREFETCH_GAUGE_L1 (prefetch to L1)
 #define PREFETCH_GAUGE_L1_INTERNAL_A64FXf(A)  \
 { \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
+    const auto & ref(U[sU](A)); baseU = (uint64_t)&ref; \
-    svprfd(pg1, (int64_t*)(baseU + 0), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(0), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 256), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(4), SV_PLDL1STRM); \
-    svprfd(pg1, (int64_t*)(baseU + 512), SV_PLDL1STRM); \
+    svprfd_vnum(pg1, (void*)(baseU), (int64_t)(8), SV_PLDL1STRM); \
 }
 // LOAD_CHI
 #define LOAD_CHI_A64FXf(base)  \
 { \
-    Chi_00 = svld1(pg1, (float32_t*)(base + 0 * 64));  \
+    Chi_00 = svld1_vnum(pg1, (float32_t*)(base), (int64_t)(0));  \
-    Chi_01 = svld1(pg1, (float32_t*)(base + 1 * 64));  \
+    Chi_01 = svld1_vnum(pg1, (float32_t*)(base), (int64_t)(1));  \
-    Chi_02 = svld1(pg1, (float32_t*)(base + 2 * 64));  \
+    Chi_02 = svld1_vnum(pg1, (float32_t*)(base), (int64_t)(2));  \
-    Chi_10 = svld1(pg1, (float32_t*)(base + 3 * 64));  \
+    Chi_10 = svld1_vnum(pg1, (float32_t*)(base), (int64_t)(3));  \
-    Chi_11 = svld1(pg1, (float32_t*)(base + 4 * 64));  \
+    Chi_11 = svld1_vnum(pg1, (float32_t*)(base), (int64_t)(4));  \
-    Chi_12 = svld1(pg1, (float32_t*)(base + 5 * 64));  \
+    Chi_12 = svld1_vnum(pg1, (float32_t*)(base), (int64_t)(5));  \
 }
 // LOAD_CHIMU
 #define LOAD_CHIMU_INTERLEAVED_A64FXf(base)  \
 { \
-    Chimu_00 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -6 * 64));  \
+    Chimu_00 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-6));  \
-    Chimu_30 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 3 * 64));  \
+    Chimu_30 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(3));  \
-    Chimu_10 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -3 * 64));  \
+    Chimu_10 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-3));  \
-    Chimu_20 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 0 * 64));  \
+    Chimu_20 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(0));  \
-    Chimu_01 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -5 * 64));  \
+    Chimu_01 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-5));  \
-    Chimu_31 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 4 * 64));  \
+    Chimu_31 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(4));  \
-    Chimu_11 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -2 * 64));  \
+    Chimu_11 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-2));  \
-    Chimu_21 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 1 * 64));  \
+    Chimu_21 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(1));  \
-    Chimu_02 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -4 * 64));  \
+    Chimu_02 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-4));  \
-    Chimu_32 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 5 * 64));  \
+    Chimu_32 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(5));  \
-    Chimu_12 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -1 * 64));  \
+    Chimu_12 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-1));  \
-    Chimu_22 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 2 * 64));  \
+    Chimu_22 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(2));  \
 }
 // LOAD_CHIMU_0213
 #define LOAD_CHIMU_0213_A64FXf  \
 { \
     const SiteSpinor & ref(in[offset]); \
-    Chimu_00 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -6 * 64));  \
+    Chimu_00 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-6));  \
-    Chimu_20 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 0 * 64));  \
+    Chimu_20 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(0));  \
-    Chimu_01 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -5 * 64));  \
+    Chimu_01 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-5));  \
-    Chimu_21 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 1 * 64));  \
+    Chimu_21 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(1));  \
-    Chimu_02 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -4 * 64));  \
+    Chimu_02 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-4));  \
-    Chimu_22 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 2 * 64));  \
+    Chimu_22 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(2));  \
-    Chimu_10 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -3 * 64));  \
+    Chimu_10 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-3));  \
-    Chimu_30 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 3 * 64));  \
+    Chimu_30 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(3));  \
-    Chimu_11 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -2 * 64));  \
+    Chimu_11 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-2));  \
-    Chimu_31 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 4 * 64));  \
+    Chimu_31 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(4));  \
-    Chimu_12 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -1 * 64));  \
+    Chimu_12 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-1));  \
-    Chimu_32 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 5 * 64));  \
+    Chimu_32 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(5));  \
 }
 // LOAD_CHIMU_0312
 #define LOAD_CHIMU_0312_A64FXf  \
 { \
     const SiteSpinor & ref(in[offset]); \
-    Chimu_00 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -6 * 64));  \
+    Chimu_00 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-6));  \
-    Chimu_30 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 3 * 64));  \
+    Chimu_30 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(3));  \
-    Chimu_01 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -5 * 64));  \
+    Chimu_01 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-5));  \
-    Chimu_31 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 4 * 64));  \
+    Chimu_31 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(4));  \
-    Chimu_02 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -4 * 64));  \
+    Chimu_02 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-4));  \
-    Chimu_32 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 5 * 64));  \
+    Chimu_32 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(5));  \
-    Chimu_10 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -3 * 64));  \
+    Chimu_10 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-3));  \
-    Chimu_20 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 0 * 64));  \
+    Chimu_20 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(0));  \
-    Chimu_11 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -2 * 64));  \
+    Chimu_11 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-2));  \
-    Chimu_21 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 1 * 64));  \
+    Chimu_21 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(1));  \
-    Chimu_12 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + -1 * 64));  \
+    Chimu_12 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(-1));  \
-    Chimu_22 = svld1(pg1, (float32_t*)(base + 2 * 3 * 64 + 2 * 64));  \
+    Chimu_22 = svld1_vnum(pg1, (float32_t*)(base + 2 * 3 * 64), (int64_t)(2));  \
 }
 // LOAD_TABLE0
 #define LOAD_TABLE0  \
@@ -261,26 +263,26 @@ Author: Nils Meyer <nils.meyer@ur.de>
     Chi_12 = svtbl(Chi_12, table0);    
 
 // LOAD_GAUGE
-#define LOAD_GAUGE  \
+#define LOAD_GAUGE(A)  \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
 { \
-    U_00 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -6 * 64));  \
+    const auto & ref(U[sU](A)); baseU = (uint64_t)&ref; \
-    U_10 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -3 * 64));  \
+    U_00 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-6));  \
-    U_20 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + 0 * 64));  \
+    U_10 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-3));  \
-    U_01 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -5 * 64));  \
+    U_20 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(0));  \
-    U_11 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -2 * 64));  \
+    U_01 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-5));  \
-    U_21 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + 1 * 64));  \
+    U_11 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-2));  \
+    U_21 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(1));  \
 }
 // MULT_2SPIN
 #define MULT_2SPIN_1_A64FXf(A)  \
 { \
-    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
+    const auto & ref(U[sU](A)); baseU = (uint64_t)&ref; \
-    U_00 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -6 * 64));  \
+    U_00 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-6));  \
-    U_10 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -3 * 64));  \
+    U_10 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-3));  \
-    U_20 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + 0 * 64));  \
+    U_20 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(0));  \
-    U_01 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -5 * 64));  \
+    U_01 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-5));  \
-    U_11 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -2 * 64));  \
+    U_11 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-2));  \
-    U_21 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + 1 * 64));  \
+    U_21 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(1));  \
     UChi_00 = svcmla_x(pg1, zero0, U_00, Chi_00, 0); \
     UChi_10 = svcmla_x(pg1, zero0, U_00, Chi_10, 0); \
     UChi_01 = svcmla_x(pg1, zero0, U_10, Chi_00, 0); \
@@ -293,9 +295,9 @@ Author: Nils Meyer <nils.meyer@ur.de>
     UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_10, 90); \
     UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_00, 90); \
     UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_10, 90); \
-    U_00 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -4 * 64));  \
+    U_00 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-4));  \
-    U_10 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -1 * 64));  \
+    U_10 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(-1));  \
-    U_20 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + 2 * 64));  \
+    U_20 = svld1_vnum(pg1, (float32_t*)(baseU + 2 * 3 * 64), (int64_t)(2));  \
 }
 // MULT_2SPIN_BACKEND
 #define MULT_2SPIN_2_A64FXf  \
@@ -570,12 +572,12 @@ Author: Nils Meyer <nils.meyer@ur.de>
     result_31 = svdup_f32(0.); \
     result_32 = svdup_f32(0.); 
 
-// PREFETCH_RESULT_L2_STORE (prefetch store to L2)
+// PREFETCH_RESULT_L2_STORE (uses DC ZVA for cache line zeroing)
 #define PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXf(base)  \
 { \
-    svprfd(pg1, (int64_t*)(base + 0), SV_PSTL2STRM); \
+    asm( "dc zva, %[fetchptr] \n\t" : : [fetchptr] "r" (base + 256 * 0) : "memory" ); \
-    svprfd(pg1, (int64_t*)(base + 256), SV_PSTL2STRM); \
+    asm( "dc zva, %[fetchptr] \n\t" : : [fetchptr] "r" (base + 256 * 1) : "memory" ); \
-    svprfd(pg1, (int64_t*)(base + 512), SV_PSTL2STRM); \
+    asm( "dc zva, %[fetchptr] \n\t" : : [fetchptr] "r" (base + 256 * 2) : "memory" ); \
 }
 // PREFETCH_RESULT_L1_STORE (prefetch store to L1)
 #define PREFETCH_RESULT_L1_STORE_INTERNAL_A64FXf(base)  \

Grid/simd/Fujitsu_A64FX_undef.h

@@ -46,6 +46,7 @@ Author: Nils Meyer <nils.meyer@ur.de>
 #undef MULT_2SPIN_2
 #undef MAYBEPERM
 #undef LOAD_CHI
+#undef ZERO_PSI
 #undef XP_PROJ
 #undef YP_PROJ
 #undef ZP_PROJ

Grid/simd/Grid_gpu_vec.h

@@ -38,12 +38,20 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_HIP
 #include <hip/hip_fp16.h>
 #endif
+#ifdef GRID_SYCL
+namespace Grid {
+  typedef struct { uint16_t x;} half;
+  typedef struct { half   x; half   y;} half2;
+  typedef struct { float  x; float  y;} float2;
+  typedef struct { double x; double y;} double2;
+}
+#endif
+
 
 namespace Grid {
 
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+
-typedef struct { uint16_t x;} half;
+
-#endif
 typedef struct Half2_t { half x; half y; } Half2;
 
 #define COALESCE_GRANULARITY ( GEN_SIMD_WIDTH )
@@ -52,11 +60,26 @@ template<class pair>
 class GpuComplex {
 public:
   pair z;
-  typedef decltype(z.x) real;
+  typedef decltype(z.x) Real;
 public: 
   accelerator_inline GpuComplex() = default;
-  accelerator_inline GpuComplex(real re,real im) { z.x=re; z.y=im; };
+  accelerator_inline GpuComplex(Real re,Real im) { z.x=re; z.y=im; };
   accelerator_inline GpuComplex(const GpuComplex &zz) { z = zz.z;};
+  accelerator_inline Real real(void) const { return z.x; };
+  accelerator_inline Real imag(void) const { return z.y; };
+  accelerator_inline GpuComplex &operator=(const Zero &zz) { z.x = 0; z.y=0; return *this; };
+  accelerator_inline GpuComplex &operator*=(const GpuComplex &r) {
+    *this = (*this) * r;
+    return *this;
+  }
+  accelerator_inline GpuComplex &operator+=(const GpuComplex &r) {
+    *this = (*this) + r;
+    return *this;
+  }
+  accelerator_inline GpuComplex &operator-=(const GpuComplex &r) {
+    *this = (*this) - r;
+    return *this;
+  }
   friend accelerator_inline  GpuComplex operator+(const GpuComplex &lhs,const GpuComplex &rhs) { 
     GpuComplex r ; 
     r.z.x = lhs.z.x + rhs.z.x; 
@@ -149,6 +172,11 @@ typedef GpuVector<NSIMD_RealD,    double      > GpuVectorRD;
 typedef GpuVector<NSIMD_ComplexD, GpuComplexD > GpuVectorCD;
 typedef GpuVector<NSIMD_Integer,  Integer     > GpuVectorI;
 
+accelerator_inline GpuComplexF timesI(const GpuComplexF &r)     { return(GpuComplexF(-r.imag(),r.real()));}
+accelerator_inline GpuComplexD timesI(const GpuComplexD &r)     { return(GpuComplexD(-r.imag(),r.real()));}
+accelerator_inline GpuComplexF timesMinusI(const GpuComplexF &r){ return(GpuComplexF(r.imag(),-r.real()));}
+accelerator_inline GpuComplexD timesMinusI(const GpuComplexD &r){ return(GpuComplexD(r.imag(),-r.real()));}
+
 accelerator_inline float half2float(half h)
 {
   float f;
@@ -156,7 +184,7 @@ accelerator_inline float half2float(half h)
   f = __half2float(h);
 #else 
   Grid_half hh; 
-  hh.x = hr.x;
+  hh.x = h.x;
   f=  sfw_half_to_float(hh);
 #endif
   return f;

Grid/simd/Grid_vector_types.h

@@ -208,8 +208,8 @@ struct RealPart<complex<T> > {
 //////////////////////////////////////
 // type alias used to simplify the syntax of std::enable_if
 template <typename T> using Invoke = typename T::type;
-template <typename Condition, typename ReturnType> using EnableIf    = Invoke<std::enable_if<Condition::value, ReturnType> >;
+template <typename Condition, typename ReturnType = void> using EnableIf    = Invoke<std::enable_if<Condition::value, ReturnType> >;
-template <typename Condition, typename ReturnType> using NotEnableIf = Invoke<std::enable_if<!Condition::value, ReturnType> >;
+template <typename Condition, typename ReturnType = void> using NotEnableIf = Invoke<std::enable_if<!Condition::value, ReturnType> >;
 
 ////////////////////////////////////////////////////////
 // Check for complexity with type traits

Grid/simd/Grid_vector_unops.h

@@ -125,14 +125,6 @@ accelerator_inline Grid_simd<S, V> sqrt(const Grid_simd<S, V> &r) {
   return SimdApply(SqrtRealFunctor<S>(), r);
 }
 template <class S, class V>
-accelerator_inline Grid_simd<S, V> rsqrt(const Grid_simd<S, V> &r) {
-  return SimdApply(RSqrtRealFunctor<S>(), r);
-}
-template <class Scalar>
-accelerator_inline Scalar rsqrt(const Scalar &r) {
-  return (RSqrtRealFunctor<Scalar>(), r);
-}
-template <class S, class V>
 accelerator_inline Grid_simd<S, V> cos(const Grid_simd<S, V> &r) {
   return SimdApply(CosRealFunctor<S>(), r);
 }

Grid/simd/Simd.h

@@ -148,10 +148,14 @@ accelerator_inline void sub (ComplexF * __restrict__ y,const ComplexF * __restri
 accelerator_inline void add (ComplexF * __restrict__ y,const ComplexF * __restrict__ l,const ComplexF *__restrict__ r){ *y = (*l) + (*r); }
   
 //conjugate already supported for complex
-accelerator_inline ComplexF timesI(const ComplexF &r)     { return(r*ComplexF(0.0,1.0));}
+accelerator_inline ComplexF timesI(const ComplexF &r)     { return(ComplexF(-r.imag(),r.real()));}
-accelerator_inline ComplexD timesI(const ComplexD &r)     { return(r*ComplexD(0.0,1.0));}
+accelerator_inline ComplexD timesI(const ComplexD &r)     { return(ComplexD(-r.imag(),r.real()));}
-accelerator_inline ComplexF timesMinusI(const ComplexF &r){ return(r*ComplexF(0.0,-1.0));}
+accelerator_inline ComplexF timesMinusI(const ComplexF &r){ return(ComplexF(r.imag(),-r.real()));}
-accelerator_inline ComplexD timesMinusI(const ComplexD &r){ return(r*ComplexD(0.0,-1.0));}
+accelerator_inline ComplexD timesMinusI(const ComplexD &r){ return(ComplexD(r.imag(),-r.real()));}
+//accelerator_inline ComplexF timesI(const ComplexF &r)     { return(r*ComplexF(0.0,1.0));}
+//accelerator_inline ComplexD timesI(const ComplexD &r)     { return(r*ComplexD(0.0,1.0));}
+//accelerator_inline ComplexF timesMinusI(const ComplexF &r){ return(r*ComplexF(0.0,-1.0));}
+//accelerator_inline ComplexD timesMinusI(const ComplexD &r){ return(r*ComplexD(0.0,-1.0));}
 
 // define projections to real and imaginay parts
 accelerator_inline ComplexF projReal(const ComplexF &r){return( ComplexF(r.real(), 0.0));}

Grid/stencil/SimpleCompressor.h

@@ -7,20 +7,20 @@ template<class vobj>
 class SimpleCompressor {
 public:
   void Point(int) {};
-  accelerator_inline int  CommDatumSize(void) { return sizeof(vobj); }
+  accelerator_inline int  CommDatumSize(void) const { return sizeof(vobj); }
-  accelerator_inline bool DecompressionStep(void) { return false; }
+  accelerator_inline bool DecompressionStep(void) const { return false; }
-  template<class cobj> accelerator_inline void Compress(cobj *buf,int o,const cobj &in) { buf[o]=in; }
+  template<class cobj> accelerator_inline void Compress(cobj *buf,int o,const cobj &in) const { buf[o]=in; }
-  accelerator_inline void Exchange(vobj *mp,vobj *vp0,vobj *vp1,Integer type,Integer o){
+  accelerator_inline void Exchange(vobj *mp,vobj *vp0,vobj *vp1,Integer type,Integer o) const {
     exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
   }
-  accelerator_inline void Decompress(vobj *out,vobj *in, int o){ assert(0); }
+  accelerator_inline void Decompress(vobj *out,vobj *in, int o) const { assert(0); }
   accelerator_inline void CompressExchange(vobj *out0,vobj *out1,const vobj *in,
-			       int j,int k, int m,int type){
+			       int j,int k, int m,int type) const {
     exchange(out0[j],out1[j],in[k],in[m],type);
   }
   // For cshift. Cshift should drop compressor coupling altogether 
   // because I had to decouple the code from the Stencil anyway
-  accelerator_inline vobj operator() (const vobj &arg) {
+  accelerator_inline vobj operator() (const vobj &arg) const {
     return arg;
   }
 };

Grid/stencil/Stencil.h

@@ -147,16 +147,16 @@ class CartesianStencilAccelerator {
   cobj* u_recv_buf_p;
   cobj* u_send_buf_p;
 
-  accelerator_inline cobj *CommBuf(void) { return u_recv_buf_p; }
+  accelerator_inline cobj *CommBuf(void) const { return u_recv_buf_p; }
 
-  accelerator_inline int GetNodeLocal(int osite,int point) {
+  accelerator_inline int GetNodeLocal(int osite,int point) const {
     return this->_entries_p[point+this->_npoints*osite]._is_local;
   }
-  accelerator_inline StencilEntry * GetEntry(int &ptype,int point,int osite) {
+  accelerator_inline StencilEntry * GetEntry(int &ptype,int point,int osite) const {
     ptype = this->_permute_type[point]; return & this->_entries_p[point+this->_npoints*osite];
   }
 
-  accelerator_inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) {
+  accelerator_inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) const {
     uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
     local = this->_entries_p[ent]._is_local;
     perm  = this->_entries_p[ent]._permute;
@@ -168,14 +168,14 @@ class CartesianStencilAccelerator {
     }
   }
 
-  accelerator_inline uint64_t GetPFInfo(int ent,uint64_t base) {
+  accelerator_inline uint64_t GetPFInfo(int ent,uint64_t base) const {
     uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
     int local = this->_entries_p[ent]._is_local;
     if (local) return  base + this->_entries_p[ent]._byte_offset;
     else       return cbase + this->_entries_p[ent]._byte_offset;
   }
 
-  accelerator_inline void iCoorFromIindex(Coordinate &coor,int lane)
+  accelerator_inline void iCoorFromIindex(Coordinate &coor,int lane) const
   {
     Lexicographic::CoorFromIndex(coor,lane,this->_simd_layout);
   }
@@ -221,7 +221,7 @@ public:
   typedef typename cobj::vector_type vector_type;
   typedef typename cobj::scalar_type scalar_type;
   typedef typename cobj::scalar_object scalar_object;
-  typedef CartesianStencilView<vobj,cobj,Parameters> View_type;
+  typedef const CartesianStencilView<vobj,cobj,Parameters> View_type;
   typedef typename View_type::StencilVector StencilVector;
   ///////////////////////////////////////////
   // Helper structs
@@ -269,7 +269,7 @@ public:
   std::vector<Vector<std::pair<int,int> > > face_table ;
   Vector<int> surface_list;
 
-  Vector<StencilEntry>  _entries; // Resident in managed memory
+  stencilVector<StencilEntry>  _entries; // Resident in managed memory
   std::vector<Packet> Packets;
   std::vector<Merge> Mergers;
   std::vector<Merge> MergersSHM;

Grid/tensors/Tensor_SIMT.h

@@ -64,6 +64,71 @@ void coalescedWriteNonTemporal(vobj & __restrict__ vec,const vobj & __restrict__
 }
 #else
 
+
+//#ifndef GRID_SYCL
+#if 1
+// Use the scalar as our own complex on GPU ... thrust::complex or std::complex
+template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
+typename vsimd::scalar_type
+coalescedRead(const vsimd & __restrict__ vec,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
+{
+  typedef typename vsimd::scalar_type S;
+  S * __restrict__ p=(S *)&vec;
+  return p[lane];
+}
+template<int ptype,class vsimd,IfSimd<vsimd> = 0> accelerator_inline
+typename vsimd::scalar_type
+coalescedReadPermute(const vsimd & __restrict__ vec,int doperm,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
+{
+  typedef typename vsimd::scalar_type S;
+
+  S * __restrict__ p=(S *)&vec;
+  int mask = vsimd::Nsimd() >> (ptype + 1);
+  int plane= doperm ? lane ^ mask : lane;
+  return p[plane];
+}
+template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
+void coalescedWrite(vsimd & __restrict__ vec,
+		    const typename vsimd::scalar_type & __restrict__ extracted,
+		    int lane=acceleratorSIMTlane(vsimd::Nsimd()))
+{
+  typedef typename vsimd::scalar_type S;
+  S * __restrict__ p=(S *)&vec;
+  p[lane]=extracted;
+}
+#else
+// For SyCL have option to use GpuComplex from inside the vector type in SIMT loops
+// Faster for some reason
+template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
+typename vsimd::vector_type::datum
+coalescedRead(const vsimd & __restrict__ vec,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
+{
+  typedef typename vsimd::vector_type::datum S;
+  S * __restrict__ p=(S *)&vec;
+  return p[lane];
+}
+template<int ptype,class vsimd,IfSimd<vsimd> = 0> accelerator_inline
+typename vsimd::vector_type::datum
+coalescedReadPermute(const vsimd & __restrict__ vec,int doperm,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
+{
+  typedef typename vsimd::vector_type::datum S;
+
+  S * __restrict__ p=(S *)&vec;
+  int mask = vsimd::Nsimd() >> (ptype + 1);
+  int plane= doperm ? lane ^ mask : lane;
+  return p[plane];
+}
+template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
+void coalescedWrite(vsimd & __restrict__ vec,
+		    const typename vsimd::vector_type::datum & __restrict__ extracted,
+		    int lane=acceleratorSIMTlane(vsimd::Nsimd()))
+{
+  typedef typename vsimd::vector_type::datum S;
+  S * __restrict__ p=(S *)&vec;
+  p[lane]=extracted;
+}
+#endif
+
 //////////////////////////////////////////
 // Extract and insert slices on the GPU
 //////////////////////////////////////////