Improve the matching of stencil coarsening

Conflicts:
	Grid/allocator/MemoryManager.h

.travis.yml

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

Grid/DisableWarnings.h

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

Grid/GridStd.h

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

Grid/Grid_Eigen_Dense.h

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

Grid/Makefile.am

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

Grid/algorithms/CoarsenedMatrix.h

@@ -31,7 +31,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_ALGORITHM_COARSENED_MATRIX_H
 #define  GRID_ALGORITHM_COARSENED_MATRIX_H
 
-#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
 
 NAMESPACE_BEGIN(Grid);
 
@@ -50,24 +49,20 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
   Lattice<dotp> fine_inner_msk(fine);
 
   // Multiply could be fused with innerProduct
-  // Single block sum kernel could do both masks.
   fine_inner = localInnerProduct(fineX,fineY);
   mult(fine_inner_msk, fine_inner,FineMask);
   blockSum(CoarseInner,fine_inner_msk);
 }
 
-
 class Geometry {
 public:
   int npoint;
-  int base;
   std::vector<int> directions   ;
   std::vector<int> displacements;
-  std::vector<int> points_dagger;
 
   Geometry(int _d)  {
     
-    base = (_d==5) ? 1:0;
+    int base = (_d==5) ? 1:0;
 
     // make coarse grid stencil for 4d , not 5d
     if ( _d==5 ) _d=4;
@@ -75,50 +70,19 @@ public:
     npoint = 2*_d+1;
     directions.resize(npoint);
     displacements.resize(npoint);
-    points_dagger.resize(npoint);
     for(int d=0;d<_d;d++){
       directions[d   ] = d+base;
       directions[d+_d] = d+base;
       displacements[d  ] = +1;
       displacements[d+_d]= -1;
-      points_dagger[d   ] = d+_d;
-      points_dagger[d+_d] = d;
     }
     directions   [2*_d]=0;
     displacements[2*_d]=0;
-    points_dagger[2*_d]=2*_d;
-  }
 
-  int point(int dir, int disp) {
-    assert(disp == -1 || disp == 0 || disp == 1);
-    assert(base+0 <= dir && dir < base+4);
-
-    // directions faster index = new indexing
-    // 4d (base = 0):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   0  1  2  3  0  1  2  3  0
-    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
-    // 5d (base = 1):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   1  2  3  4  1  2  3  4  0
-    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
-
-    // displacements faster index = old indexing
-    // 4d (base = 0):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   0  0  1  1  2  2  3  3  0
-    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
-    // 5d (base = 1):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   1  1  2  2  3  3  4  4  0
-    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
-
-    if(dir == 0 and disp == 0)
-      return 8;
-    else // New indexing
-      return (1 - disp) / 2 * 4 + dir - base;
-    // else // Old indexing
-    //   return (4 * (dir - base) + 1 - disp) / 2;
+    std::cout <<GridLogMessage << "Geometry "<<std::endl;
+    for(int p=0;p<npoint;p++){
+      std::cout <<GridLogMessage << "point " <<p<<" dir "<<directions[p]<<" delta " <<displacements[p]<<std::endl;
+    }
   }
 };
   
@@ -140,8 +104,8 @@ public:
   Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
     CoarseGrid(_CoarseGrid),
     FineGrid(_FineGrid),
-    subspace(nbasis,_FineGrid),
-    checkerboard(_checkerboard)
+    checkerboard(_checkerboard),
+    subspace(nbasis,_FineGrid)
   {
   };
   
@@ -296,7 +260,7 @@ public:
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>
-class CoarsenedMatrix : public CheckerBoardedSparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
 public:
     
   typedef iVector<CComplex,nbasis >           siteVector;
@@ -306,59 +270,35 @@ public:
   typedef iMatrix<CComplex,nbasis >  Cobj;
   typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
   typedef Lattice<Fobj >        FineField;
-  typedef CoarseVector FermionField;
-
-  // enrich interface, use default implementation as in FermionOperator ///////
-  void Dminus(CoarseVector const& in, CoarseVector& out) { out = in; }
-  void DminusDag(CoarseVector const& in, CoarseVector& out) { out = in; }
-  void ImportPhysicalFermionSource(CoarseVector const& input, CoarseVector& imported) { imported = input; }
-  void ImportUnphysicalFermion(CoarseVector const& input, CoarseVector& imported) { imported = input; }
-  void ExportPhysicalFermionSolution(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
-  void ExportPhysicalFermionSource(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
 
   ////////////////////
   // Data members
   ////////////////////
   Geometry         geom;
   GridBase *       _grid; 
-  GridBase*        _cbgrid;
   int hermitian;
 
   CartesianStencil<siteVector,siteVector,int> Stencil; 
-  CartesianStencil<siteVector,siteVector,int> StencilEven;
-  CartesianStencil<siteVector,siteVector,int> StencilOdd;
 
   std::vector<CoarseMatrix> A;
-  std::vector<CoarseMatrix> Aeven;
-  std::vector<CoarseMatrix> Aodd;
-
-  CoarseMatrix AselfInv;
-  CoarseMatrix AselfInvEven;
-  CoarseMatrix AselfInvOdd;
-
-  Vector<RealD> dag_factor;
-
+    
   ///////////////////////
   // Interface
   ///////////////////////
   GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know
-  GridBase * RedBlackGrid()     { return _cbgrid; };
-
-  int ConstEE() { return 0; }
 
+  virtual std::vector<int> Directions(void)   { return geom.directions; };
+  virtual std::vector<int> Displacements(void){ return geom.displacements; };
   void M (const CoarseVector &in, CoarseVector &out)
   {
     conformable(_grid,in.Grid());
     conformable(in.Grid(),out.Grid());
-    out.Checkerboard() = in.Checkerboard();
 
     SimpleCompressor<siteVector> compressor;
 
     Stencil.HaloExchange(in,compressor);
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
-    autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
     typedef LatticeView<Cobj> Aview;
       
     Vector<Aview> AcceleratorViewContainer;
@@ -372,6 +312,9 @@ public:
 
     int osites=Grid()->oSites();
 
+    autoView(st,Stencil,AcceleratorRead);
+    siteVector *CBp=Stencil.CommBuf();
+
     accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
       int ss = sss/nbasis;
       int b  = sss%nbasis;
@@ -380,14 +323,14 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom_v.npoint;point++){
+      for(int point=0;point<geom.npoint;point++){
 
-	SE=Stencil_v.GetEntry(ptype,point,ss);
+	SE=st.GetEntry(ptype,point,ss);
 	  
 	if(SE->_is_local) { 
 	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
 	} else {
-	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
+	  nbr = coalescedRead(CBp[SE->_offset]);
 	}
 	acceleratorSynchronise();
 
@@ -396,7 +339,7 @@ public:
 	}
       }
       coalescedWrite(out_v[ss](b),res);
-      });
+    });
 
     for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   };
@@ -408,72 +351,12 @@ public:
       return M(in,out);
     } else {
       // corresponds to Galerkin coarsening
-      return MdagNonHermitian(in, out);
+      CoarseVector tmp(Grid());
+      G5C(tmp, in); 
+      M(tmp, out);
+      G5C(out, out);
     }
   };
-
-  void MdagNonHermitian(const CoarseVector &in, CoarseVector &out)
-  {
-    conformable(_grid,in.Grid());
-    conformable(in.Grid(),out.Grid());
-    out.Checkerboard() = in.Checkerboard();
-
-    SimpleCompressor<siteVector> compressor;
-
-    Stencil.HaloExchange(in,compressor);
-    autoView( in_v , in, AcceleratorRead);
-    autoView( out_v , out, AcceleratorWrite);
-    autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
-    typedef LatticeView<Cobj> Aview;
-
-    Vector<Aview> AcceleratorViewContainer;
-
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
-    Aview *Aview_p = & AcceleratorViewContainer[0];
-
-    const int Nsimd = CComplex::Nsimd();
-    typedef decltype(coalescedRead(in_v[0])) calcVector;
-    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
-
-    int osites=Grid()->oSites();
-
-    Vector<int> points(geom.npoint, 0);
-    for(int p=0; p<geom.npoint; p++)
-      points[p] = geom.points_dagger[p];
-
-    RealD* dag_factor_p = &dag_factor[0];
-
-    accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
-      int ss = sss/nbasis;
-      int b  = sss%nbasis;
-      calcComplex res = Zero();
-      calcVector nbr;
-      int ptype;
-      StencilEntry *SE;
-
-      for(int p=0;p<geom_v.npoint;p++){
-        int point = points[p];
-
-	SE=Stencil_v.GetEntry(ptype,point,ss);
-
-	if(SE->_is_local) {
-	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-	} else {
-	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
-	}
-	acceleratorSynchronise();
-
-	for(int bb=0;bb<nbasis;bb++) {
-	  res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
-	}
-      }
-      coalescedWrite(out_v[ss](b),res);
-      });
-
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
-  }
-
   void MdirComms(const CoarseVector &in)
   {
     SimpleCompressor<siteVector> compressor;
@@ -483,7 +366,6 @@ public:
   {
     conformable(_grid,in.Grid());
     conformable(_grid,out.Grid());
-    out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
     Vector<Aview> AcceleratorViewContainer;
@@ -492,7 +374,6 @@ public:
 
     autoView( out_v , out, AcceleratorWrite);
     autoView( in_v  , in, AcceleratorRead);
-    autoView( Stencil_v  , Stencil, AcceleratorRead);
 
     const int Nsimd = CComplex::Nsimd();
     typedef decltype(coalescedRead(in_v[0])) calcVector;
@@ -506,12 +387,12 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      SE=Stencil_v.GetEntry(ptype,point,ss);
+      SE=Stencil.GetEntry(ptype,point,ss);
 	  
       if(SE->_is_local) { 
 	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
       } else {
-	nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
+	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
       }
       acceleratorSynchronise();
 
@@ -535,11 +416,23 @@ public:
       MdirCalc(in,out[p],p);
     }
   };
-  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
-
+  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp)
+  {
     this->MdirComms(in);
 
-    MdirCalc(in,out,geom.point(dir,disp));
+    int ndim = in.Grid()->Nd();
+
+    int point=-1;
+    for(int p=0;p<geom.npoint;p++){
+      if( (dir==geom.directions[p])&&(disp==geom.displacements[p])) point=p;
+    }
+    assert(point!=-1);// Must find
+
+    std::cout <<GridLogMessage << "Mdir point "<<point<<" dir "<<dir<<" disp "<<disp  <<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      std::cout <<GridLogMessage << "point " <<p<<" dir "<<geom.directions[p]<<" delta " <<geom.displacements[p]<<std::endl;
+    }
+    MdirCalc(in,out,point);
   };
 
   void Mdiag(const CoarseVector &in, CoarseVector &out)
@@ -548,286 +441,63 @@ public:
     MdirCalc(in, out, point); // No comms
   };
 
-  void Mooee(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerNo, InverseNo);
-  }
-
-  void MooeeInv(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerNo, InverseYes);
-  }
-
-  void MooeeDag(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerYes, InverseNo);
-  }
-
-  void MooeeInvDag(const CoarseVector &in, CoarseVector &out) {
-    MooeeInternal(in, out, DaggerYes, InverseYes);
-  }
-
-  void Meooe(const CoarseVector &in, CoarseVector &out) {
-    if(in.Checkerboard() == Odd) {
-      DhopEO(in, out, DaggerNo);
-    } else {
-      DhopOE(in, out, DaggerNo);
-    }
-  }
-
-  void MeooeDag(const CoarseVector &in, CoarseVector &out) {
-    if(in.Checkerboard() == Odd) {
-      DhopEO(in, out, DaggerYes);
-    } else {
-      DhopOE(in, out, DaggerYes);
-    }
-  }
-
-  void Dhop(const CoarseVector &in, CoarseVector &out, int dag) {
-    conformable(in.Grid(), _grid); // verifies full grid
-    conformable(in.Grid(), out.Grid());
-
-    out.Checkerboard() = in.Checkerboard();
-
-    DhopInternal(Stencil, A, in, out, dag);
-  }
-
-  void DhopOE(const CoarseVector &in, CoarseVector &out, int dag) {
-    conformable(in.Grid(), _cbgrid);    // verifies half grid
-    conformable(in.Grid(), out.Grid()); // drops the cb check
-
-    assert(in.Checkerboard() == Even);
-    out.Checkerboard() = Odd;
-
-    DhopInternal(StencilEven, Aodd, in, out, dag);
-  }
-
-  void DhopEO(const CoarseVector &in, CoarseVector &out, int dag) {
-    conformable(in.Grid(), _cbgrid);    // verifies half grid
-    conformable(in.Grid(), out.Grid()); // drops the cb check
-
-    assert(in.Checkerboard() == Odd);
-    out.Checkerboard() = Even;
-
-    DhopInternal(StencilOdd, Aeven, in, out, dag);
-  }
-
-  void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) {
-    out.Checkerboard() = in.Checkerboard();
-    assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
-
-    CoarseMatrix *Aself = nullptr;
-    if(in.Grid()->_isCheckerBoarded) {
-      if(in.Checkerboard() == Odd) {
-        Aself = (inv) ? &AselfInvOdd : &Aodd[geom.npoint-1];
-        DselfInternal(StencilOdd, *Aself, in, out, dag);
-      } else {
-        Aself = (inv) ? &AselfInvEven : &Aeven[geom.npoint-1];
-        DselfInternal(StencilEven, *Aself, in, out, dag);
-      }
-    } else {
-      Aself = (inv) ? &AselfInv : &A[geom.npoint-1];
-      DselfInternal(Stencil, *Aself, in, out, dag);
-    }
-    assert(Aself != nullptr);
-  }
-
-  void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
-                       const CoarseVector &in, CoarseVector &out, int dag) {
-    int point = geom.npoint-1;
-    autoView( out_v, out, AcceleratorWrite);
-    autoView( in_v,  in,  AcceleratorRead);
-    autoView( st_v,  st,  AcceleratorRead);
-    autoView( a_v,   a,   AcceleratorRead);
-
-    const int Nsimd = CComplex::Nsimd();
-    typedef decltype(coalescedRead(in_v[0])) calcVector;
-    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
-
-    RealD* dag_factor_p = &dag_factor[0];
-
-    if(dag) {
-      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
-        int ss = sss/nbasis;
-        int b  = sss%nbasis;
-        calcComplex res = Zero();
-        calcVector nbr;
-        int ptype;
-        StencilEntry *SE;
-
-        SE=st_v.GetEntry(ptype,point,ss);
-
-        if(SE->_is_local) {
-          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-        } else {
-          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
-        }
-        acceleratorSynchronise();
-
-        for(int bb=0;bb<nbasis;bb++) {
-          res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(a_v[ss](b,bb))*nbr(bb);
-        }
-        coalescedWrite(out_v[ss](b),res);
-      });
-    } else {
-      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
-        int ss = sss/nbasis;
-        int b  = sss%nbasis;
-        calcComplex res = Zero();
-        calcVector nbr;
-        int ptype;
-        StencilEntry *SE;
-
-        SE=st_v.GetEntry(ptype,point,ss);
-
-        if(SE->_is_local) {
-          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-        } else {
-          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
-        }
-        acceleratorSynchronise();
-
-        for(int bb=0;bb<nbasis;bb++) {
-          res = res + coalescedRead(a_v[ss](b,bb))*nbr(bb);
-        }
-        coalescedWrite(out_v[ss](b),res);
-      });
-    }
-  }
-
-  void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
-                    const CoarseVector &in, CoarseVector &out, int dag) {
-    SimpleCompressor<siteVector> compressor;
-
-    st.HaloExchange(in,compressor);
-    autoView( in_v,  in,  AcceleratorRead);
-    autoView( out_v, out, AcceleratorWrite);
-    autoView( st_v , st,  AcceleratorRead);
-    typedef LatticeView<Cobj> Aview;
-
-    // determine in what order we need the points
-    int npoint = geom.npoint-1;
-    Vector<int> points(npoint, 0);
-    for(int p=0; p<npoint; p++)
-      points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
-
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead));
-    Aview *Aview_p = & AcceleratorViewContainer[0];
-
-    const int Nsimd = CComplex::Nsimd();
-    typedef decltype(coalescedRead(in_v[0])) calcVector;
-    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
-
-    RealD* dag_factor_p = &dag_factor[0];
-
-    if(dag) {
-      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
-        int ss = sss/nbasis;
-        int b  = sss%nbasis;
-        calcComplex res = Zero();
-        calcVector nbr;
-        int ptype;
-        StencilEntry *SE;
-
-        for(int p=0;p<npoint;p++){
-          int point = points[p];
-          SE=st_v.GetEntry(ptype,point,ss);
-
-          if(SE->_is_local) {
-            nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-          } else {
-            nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
-          }
-          acceleratorSynchronise();
-
-          for(int bb=0;bb<nbasis;bb++) {
-            res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
-          }
-        }
-        coalescedWrite(out_v[ss](b),res);
-      });
-    } else {
-      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
-        int ss = sss/nbasis;
-        int b  = sss%nbasis;
-        calcComplex res = Zero();
-        calcVector nbr;
-        int ptype;
-        StencilEntry *SE;
-
-        for(int p=0;p<npoint;p++){
-          int point = points[p];
-          SE=st_v.GetEntry(ptype,point,ss);
-
-          if(SE->_is_local) {
-            nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
-          } else {
-            nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
-          }
-          acceleratorSynchronise();
-
-          for(int bb=0;bb<nbasis;bb++) {
-            res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
-          }
-        }
-        coalescedWrite(out_v[ss](b),res);
-      });
-    }
-
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose();
-  }
   
-  CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
+ CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	: 
+
     _grid(&CoarseGrid),
-    _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
     Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
-    A(geom.npoint,&CoarseGrid),
-    Aeven(geom.npoint,_cbgrid),
-    Aodd(geom.npoint,_cbgrid),
-    AselfInv(&CoarseGrid),
-    AselfInvEven(_cbgrid),
-    AselfInvOdd(_cbgrid),
-    dag_factor(nbasis*nbasis)
+    A(geom.npoint,&CoarseGrid)
   {
-    fillFactor();
   };
 
-  CoarsenedMatrix(GridCartesian &CoarseGrid, GridRedBlackCartesian &CoarseRBGrid, int hermitian_=0) 	:
-
-    _grid(&CoarseGrid),
-    _cbgrid(&CoarseRBGrid),
-    geom(CoarseGrid._ndimension),
-    hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
-    A(geom.npoint,&CoarseGrid),
-    Aeven(geom.npoint,&CoarseRBGrid),
-    Aodd(geom.npoint,&CoarseRBGrid),
-    AselfInv(&CoarseGrid),
-    AselfInvEven(&CoarseRBGrid),
-    AselfInvOdd(&CoarseRBGrid),
-    dag_factor(nbasis*nbasis)
+  void Test(Aggregation<Fobj,CComplex,nbasis> &_Aggregates,GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop)
   {
-    fillFactor();
-  };
+    typedef Lattice<Fobj> FineField;
+    CoarseVector Cin(_grid);
+    CoarseVector Cout(_grid);
+    CoarseVector CFout(_grid);
 
-  void fillFactor() {
-    Eigen::MatrixXd dag_factor_eigen = Eigen::MatrixXd::Ones(nbasis, nbasis);
-    if(!hermitian) {
-      const int nb = nbasis/2;
-      dag_factor_eigen.block(0,nb,nb,nb) *= -1.0;
-      dag_factor_eigen.block(nb,0,nb,nb) *= -1.0;
-    }
+    FineField Fin(FineGrid);
+    FineField Fout(FineGrid);
 
-    // GPU readable prefactor
-    thread_for(i, nbasis*nbasis, {
-      int j = i/nbasis;
-      int k = i%nbasis;
-      dag_factor[i] = dag_factor_eigen(j, k);
-    });
+
+    std::vector<int> seeds({1,2,3,4,5});
+    GridParallelRNG RNG(_grid);  RNG.SeedFixedIntegers(seeds);
+    gaussian(RNG,Cin);
+
+    _Aggregates.PromoteFromSubspace(Cin,Fin);
+    _Aggregates.ProjectToSubspace(Cin,Fin);
+
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    std::cout << GridLogMessage<< " Testing M  "<<std::endl;
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    // Coarse operator
+    this->M(Cin,Cout);
+    // Fine projected operator
+    _Aggregates.PromoteFromSubspace(Cin,Fin);
+    linop.Op(Fin,Fout);
+    _Aggregates.ProjectToSubspace(CFout,Fout);
+
+    CFout = CFout-Cout;
+    RealD diff = norm2(CFout);
+    std::cout << GridLogMessage<< " diff  "<<diff<<std::endl;
+    assert(diff<1.0e-5);
+
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    std::cout << GridLogMessage<< " Testing Mdag  "<<std::endl;
+    std::cout << GridLogMessage<< "************  "<<std::endl;
+    // Coarse operator
+    Mdag(Cin,Cout);
+    // Fine operator
+    linop.AdjOp(Fin,Fout);
+    _Aggregates.ProjectToSubspace(CFout,Fout);
+
+    CFout = CFout-Cout;
+    diff = norm2(CFout);
+    std::cout << GridLogMessage<< " diff  "<<diff<<std::endl; 
+    assert(diff<1.0e-5);
   }
 
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
@@ -836,8 +506,6 @@ public:
     typedef Lattice<typename Fobj::tensor_reduced> FineComplexField;
     typedef typename Fobj::scalar_type scalar_type;
 
-    std::cout << GridLogMessage<< "CoarsenMatrix "<< std::endl;
-
     FineComplexField one(FineGrid); one=scalar_type(1.0,0.0);
     FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0);
 
@@ -868,13 +536,22 @@ public:
 
     CoarseScalar InnerProd(Grid()); 
 
-    std::cout << GridLogMessage<< "CoarsenMatrix Orthog "<< std::endl;
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog " << std::endl;
     // Orthogonalise the subblocks over the basis
     blockOrthogonalise(InnerProd,Subspace.subspace);
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog done " << std::endl;
+
+    auto OpDirections    = linop.Directions();
+    auto OpDisplacements = linop.Displacements();
+
+    std::cout<<" Coarsening an operator with "<< OpDirections.size()<<" terms "<<std::endl;
+    for(int p=0;p<OpDirections.size();p++) {
+      assert(OpDirections[p]==geom.directions[p]);
+      assert(OpDisplacements[p]==geom.displacements[p]);
+    }
 
     // Compute the matrix elements of linop between this orthonormal
     // set of vectors.
-    std::cout << GridLogMessage<< "CoarsenMatrix masks "<< std::endl;
     int self_stencil=-1;
     for(int p=0;p<geom.npoint;p++)
     { 
@@ -907,7 +584,21 @@ public:
     evenmask = where(mod(bcb,2)==(Integer)0,one,zero);
     oddmask  = one-evenmask;
 
+    /*
+    {
+      phi=Subspace.subspace[0];
+      linop.OpDirAll(phi,Mphi_p);
+      for(int p=0;p<geom.npoint-1;p++){
+	int dir=geom.directions[p];
+	int disp=geom.displacements[p];
+	linop.OpDir(phi,Mphi,dir,disp);
+	Mphi=Mphi-Mphi_p[p];
+	std::cout << GridLogMessage <<" Direction mapping check " <<norm2(Mphi)<<std::endl;
+      }
+    }
+*/
     assert(self_stencil!=-1);
+    int lhermitian=hermitian;
 
     for(int i=0;i<nbasis;i++){
 
@@ -924,7 +615,7 @@ public:
 	int dir   = geom.directions[p];
 	int disp  = geom.displacements[p];
 
-	if ( (disp==-1) || (!hermitian ) ) {
+	if ( (disp==-1) || (!lhermitian ) ) {
 
 	  ////////////////////////////////////////////////////////////////////////
 	  // Pick out contributions coming from this cell and neighbour cell
@@ -942,7 +633,7 @@ public:
 	    autoView( A_self  , A[self_stencil], AcceleratorWrite);
 
 	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
-	    if ( hermitian && (disp==-1) ) {
+	    if ( lhermitian && (disp==-1) ) {
 	      for(int pp=0;pp<geom.npoint;pp++){// Find the opposite link and set <j|A|i> = <i|A|j>*
 		int dirp   = geom.directions[pp];
 		int dispp  = geom.displacements[pp];
@@ -954,11 +645,11 @@ public:
 		}
 	      }
 	    }
-
 	  }
 	}
       }
 
+      std::cout << GridLogMessage<< "CoarsenMatrix Diag "<<std::endl;
       ///////////////////////////////////////////
       // Faster alternate self coupling.. use hermiticity to save 2x
       ///////////////////////////////////////////
@@ -990,57 +681,35 @@ public:
 
       }
     }
-    if(hermitian) {
-      std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
+
+    MemoryManager::PrintBytes();
+
+    // Auto self test
+    Test( Subspace,FineGrid,linop);
+
+#if 0
+    ///////////////////////////
+    // test code worth preserving in if block
+    ///////////////////////////
+    std::cout<<GridLogMessage<< " Computed matrix elements "<< self_stencil <<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      std::cout<<GridLogMessage<< "A["<<p<<"]" << std::endl;
+      std::cout<<GridLogMessage<< "\n"<<A[p] << std::endl;
     }
+    std::cout<<GridLogMessage<< " picking by block0 "<< self_stencil <<std::endl;
+
+    phi=Subspace.subspace[0];
+    std::vector<int> bc(FineGrid->_ndimension,0);
+    blockPick(Grid(),phi,tmp,bc);      // Pick out a block
+    linop.Op(tmp,Mphi);                // Apply big dop
+    blockProject(iProj,Mphi,Subspace.subspace); // project it and print it
+    std::cout<<GridLogMessage<< " Computed matrix elements from block zero only "<<std::endl;
+    std::cout<<GridLogMessage<< iProj <<std::endl;
+    std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
+#endif
 
-    InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl;
-    FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl;
   }
 
-  void InvertSelfStencilLink() {
-    std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl;
-    int localVolume = Grid()->lSites();
-
-    typedef typename Cobj::scalar_object scalar_object;
-
-    autoView(Aself_v,    A[geom.npoint-1], CpuRead);
-    autoView(AselfInv_v, AselfInv,         CpuWrite);
-    thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke
-      Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis);
-      Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis);
-
-      scalar_object selfLink    = Zero();
-      scalar_object selfLinkInv = Zero();
-
-      Coordinate lcoor;
-
-      Grid()->LocalIndexToLocalCoor(site, lcoor);
-      peekLocalSite(selfLink, Aself_v, lcoor);
-
-      for (int i = 0; i < nbasis; ++i)
-        for (int j = 0; j < nbasis; ++j)
-          selfLinkEigen(i, j) = static_cast<ComplexD>(TensorRemove(selfLink(i, j)));
-
-      selfLinkInvEigen = selfLinkEigen.inverse();
-
-      for(int i = 0; i < nbasis; ++i)
-        for(int j = 0; j < nbasis; ++j)
-          selfLinkInv(i, j) = selfLinkInvEigen(i, j);
-
-      pokeLocalSite(selfLinkInv, AselfInv_v, lcoor);
-    });
-  }
-
-  void FillHalfCbs() {
-    std::cout << GridLogDebug << "CoarsenedMatrix::FillHalfCbs" << std::endl;
-    for(int p = 0; p < geom.npoint; ++p) {
-      pickCheckerboard(Even, Aeven[p], A[p]);
-      pickCheckerboard(Odd, Aodd[p], A[p]);
-    }
-    pickCheckerboard(Even, AselfInvEven, AselfInv);
-    pickCheckerboard(Odd, AselfInvOdd, AselfInv);
-  }
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/LinearOperator.h

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

Grid/algorithms/SparseMatrix.h

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

Grid/algorithms/iterative/AdefGeneric.h

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

Grid/algorithms/iterative/Deflation.h

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

Grid/algorithms/iterative/PrecConjugateResidual.h

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

Grid/allocator/AlignedAllocator.cc

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

Grid/allocator/AlignedAllocator.h

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

Grid/allocator/MemoryManager.cc

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

Grid/allocator/MemoryManager.h

@@ -34,6 +34,8 @@ NAMESPACE_BEGIN(Grid);
 
 // Move control to configure.ac and Config.h?
 
+#define ALLOCATION_CACHE
+#define GRID_ALLOC_ALIGN (2*1024*1024)
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
 /*Pinning pages is costly*/
@@ -91,8 +93,8 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);

Grid/allocator/MemoryManagerCache.cc

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

Grid/allocator/MemoryManagerShared.cc

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

Grid/communicator/Communicator_base.h

@@ -138,6 +138,21 @@ public:
 		      int recv_from_rank,
 		      int bytes);
   
+  void SendRecvPacket(void *xmit,
+		      void *recv,
+		      int xmit_to_rank,
+		      int recv_from_rank,
+		      int bytes);
+  
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+			   void *xmit,
+			   int xmit_to_rank,
+			   void *recv,
+			   int recv_from_rank,
+			   int bytes);
+  
+  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,
 			       void *recv,

Grid/communicator/Communicator_mpi3.cc

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

Grid/communicator/Communicator_none.cc

@@ -77,6 +77,15 @@ void CartesianCommunicator::GlobalSumVector(uint64_t *,int N){}
 void CartesianCommunicator::GlobalXOR(uint32_t &){}
 void CartesianCommunicator::GlobalXOR(uint64_t &){}
 
+void CartesianCommunicator::SendRecvPacket(void *xmit,
+					   void *recv,
+					   int xmit_to_rank,
+					   int recv_from_rank,
+					   int bytes)
+{
+  assert(0);
+}
+
 
 // Basic Halo comms primitive -- should never call in single node
 void CartesianCommunicator::SendToRecvFrom(void *xmit,
@@ -87,6 +96,20 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes)
+{
+  assert(0);
+}
+
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  assert(0);
+}
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   bcopy(in,out,bytes*words);
@@ -114,6 +137,10 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int recv_from_rank,
 						     int bytes, int dir)
 {
+  std::vector<CommsRequest_t> list;
+  // Discard the "dir"
+  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  SendToRecvFromComplete(list);
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
@@ -123,10 +150,13 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int recv_from_rank,
 							 int bytes, int dir)
 {
+  // Discard the "dir"
+  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
   return 2.0*bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {
+  SendToRecvFromComplete(waitall);
 }
 
 void CartesianCommunicator::StencilBarrier(void){};

Grid/communicator/SharedMemory.h

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

Grid/communicator/SharedMemoryMPI.cc

@@ -32,9 +32,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
 #endif
-#ifdef GRID_HIP
-#include <hip/hip_runtime_api.h>
-#endif
 
 NAMESPACE_BEGIN(Grid); 
 #define header "SharedMemoryMpi: "
@@ -428,7 +425,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#if defined(GRID_CUDA) ||defined(GRID_HIP)
+#ifdef GRID_CUDA
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -451,16 +448,21 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
-
-  if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+#ifndef GRID_MPI3_SHM_NONE
+  auto err =  cudaMalloc(&ShmCommBuf, bytes);
+#else
+  auto err =  cudaMallocManaged(&ShmCommBuf, bytes);
+#endif
+  if ( err !=  cudaSuccess) {
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
     exit(EXIT_FAILURE);  
   }
-  //  if ( WorldRank == 0 ){
-  if ( 1 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+  if ( WorldRank == 0 ){
+    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
 
@@ -473,26 +475,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
     //////////////////////////////////////////////////
-#ifdef GRID_CUDA
     cudaIpcMemHandle_t handle;
+    
     if ( r==WorldShmRank ) { 
-      auto err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
-#endif
-#ifdef GRID_HIP
-    hipIpcMemHandle_t handle;    
-    if ( r==WorldShmRank ) { 
-      auto err = hipIpcGetMemHandle(&handle,ShmCommBuf);
-      if ( err !=  hipSuccess) {
-	std::cerr << " SharedMemoryMPI.cc hipIpcGetMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
-	exit(EXIT_FAILURE);
-      }
-    }
-#endif
     //////////////////////////////////////////////////
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
@@ -509,24 +500,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
     void * thisBuf = ShmCommBuf;
-#ifdef GRID_CUDA
     if ( r!=WorldShmRank ) { 
-      auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
-#endif
-#ifdef GRID_HIP
-    if ( r!=WorldShmRank ) { 
-      auto err = hipIpcOpenMemHandle(&thisBuf,handle,hipIpcMemLazyEnablePeerAccess);
-      if ( err !=  hipSuccess) {
-	std::cerr << " SharedMemoryMPI.cc hipIpcOpenMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
-	exit(EXIT_FAILURE);
-      }
-    }
-#endif
     ///////////////////////////////////////////////////////////////
     // Save a copy of the device buffers
     ///////////////////////////////////////////////////////////////
@@ -666,6 +646,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
+      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -715,7 +696,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
 {
 #ifdef GRID_CUDA
   cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
@@ -771,12 +752,19 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
 
-#ifdef GRID_SHM_FORCE_MPI
-  // Hide the shared memory path between ranks
-  {
+#ifdef GRID_IBM_SUMMIT
+  // Hide the shared memory path between sockets 
+  // if even number of nodes
+  if ( (ShmSize & 0x1)==0 ) {
+    int SocketSize = ShmSize/2;
+    int mySocket = ShmRank/SocketSize; 
     for(int r=0;r<size;r++){
-      if ( r!=rank ) {
-	ShmRanks[r] = MPI_UNDEFINED;
+      int hisRank=ShmRanks[r];
+      if ( hisRank!= MPI_UNDEFINED ) {
+	int hisSocket=hisRank/SocketSize;
+	if ( hisSocket != mySocket ) {
+	  ShmRanks[r] = MPI_UNDEFINED;
+	}
       }
     }
   }

Grid/communicator/SharedMemoryNone.cc

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

Grid/cshift/Cshift_common.h

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

Grid/cshift/Cshift_mpi.h

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

Grid/lattice/Lattice_ET.h

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

Grid/lattice/Lattice_arith.h

@@ -60,9 +60,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   autoView( lhs_v , lhs, AcceleratorRead);
   autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     auto rhs_t=rhs_v(ss);
-    auto tmp  =ret_v(ss);
     mac(&tmp,&lhs_t,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
   });
@@ -124,7 +124,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
+    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     mac(&tmp,&lhs_t,&rhs);
     coalescedWrite(ret_v[ss],tmp);
@@ -182,7 +182,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
+    decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
     mac(&tmp,&lhs,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);

Grid/lattice/Lattice_basis.h

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

Grid/lattice/Lattice_reality.h

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

Grid/lattice/Lattice_reduction_gpu.h

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

Grid/lattice/Lattice_transfer.h

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

Grid/lattice/Lattice_view.h

@@ -52,6 +52,7 @@ public:
 // This will be safe to call from accelerator_for and is trivially copy constructible
 // The copy constructor for this will need to be used by device lambda functions
 /////////////////////////////////////////////////////////////////////////////////////////
+#undef LATTICE_BOUNDS_CHECK
 template<class vobj> 
 class LatticeView : public LatticeAccelerator<vobj>
 {
@@ -61,14 +62,36 @@ public:
   void * cpu_ptr;
 #ifdef GRID_SIMT
   accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { 
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
     return coalescedRead(this->_odata[i]); 
   }
 #else 
-  accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
+  accelerator_inline const vobj & operator()(size_t i) const {
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
+    return this->_odata[i];
+  }
 #endif
 
-  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
+  accelerator_inline const vobj & operator[](size_t i) const { 
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
+    return this->_odata[i]; 
+  };
+  accelerator_inline vobj       & operator[](size_t i)       { 
+#ifdef LATTICE_BOUNDS_CHECK
+    assert(i<this->_odata_size);
+    assert(i>=0);
+#endif
+    return this->_odata[i]; 
+  };
 
   accelerator_inline uint64_t begin(void) const { return 0;};
   accelerator_inline uint64_t end(void)   const { return this->_odata_size; };

Grid/log/Log.h

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

Grid/parallelIO/BinaryIO.cc

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

Grid/parallelIO/BinaryIO.h

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

Grid/parallelIO/IldgIO.h

@@ -123,7 +123,7 @@ assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
  ////////////////////////////////////////////////////////////
  // Helper to fill out metadata
  ////////////////////////////////////////////////////////////
-template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
+ template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
 					  FieldMetaData &header,
 					  scidacRecord & _scidacRecord,
 					  scidacFile   & _scidacFile) 
@@ -619,12 +619,12 @@ class IldgWriter : public ScidacWriter {
   // Don't require scidac records EXCEPT checksum
   // Use Grid MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
+  template <class vsimd>
+  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<vLorentzColourMatrixD> GaugeField;
-    typedef vLorentzColourMatrixD vobj;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -636,9 +636,6 @@ class IldgWriter : public ScidacWriter {
 
     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
 
-    stats Stats;
-    Stats(Umu,header);
-    
     std::string format = header.floating_point;
     header.ensemble_id    = description;
     header.ensemble_label = description;
@@ -708,10 +705,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
+  template <class vsimd>
+  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
 
-    typedef Lattice<vLorentzColourMatrixD > GaugeField;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
     typedef typename GaugeField::vector_object  vobj;
     typedef typename vobj::scalar_object sobj;
 
@@ -924,8 +921,7 @@ class IldgReader : public GridLimeReader {
 
     if ( found_FieldMetaData || found_usqcdInfo ) {
       FieldMetaData checker;
-      stats Stats;
-      Stats(Umu,checker);
+      GaugeStatistics(Umu,checker);
       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;

Grid/parallelIO/MetaData.h

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

Grid/parallelIO/NerscIO.h

@@ -40,8 +40,6 @@ using namespace Grid;
 class NerscIO : public BinaryIO { 
 public:
 
-  typedef Lattice<vLorentzColourMatrixD> GaugeField;
-
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -131,12 +129,12 @@ public:
   // Now the meat: the object readers
   /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void readConfiguration(GaugeField &Umu,
+  template<class vsimd>
+  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
 				       FieldMetaData& header,
-				       std::string file,
-				       GaugeStats GaugeStatisticsCalculator=GaugeStats())
+				       std::string file)
   {
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
     GridBase *grid = Umu.Grid();
     uint64_t offset = readHeader(file,Umu.Grid(),header);
@@ -155,23 +153,23 @@ public:
     // munger is a function of <floating point, Real, data_type>
     if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3D> 
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
     } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixD>
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
@@ -179,7 +177,7 @@ public:
       assert(0);
     }
 
-    GaugeStats Stats; Stats(Umu,clone);
+    GaugeStatistics(Umu,clone);
 
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl;
@@ -205,13 +203,15 @@ public:
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
+  template<class vsimd>
+  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32)
   {
-    typedef vLorentzColourMatrixD vobj;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+
+    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
@@ -229,7 +229,7 @@ public:
 
     GridMetaData(grid,header);
     assert(header.nd==4);
-    GaugeStats Stats; Stats(Umu,header);
+    GaugeStatistics(Umu,header);
     MachineCharacteristics(header);
 
 	uint64_t offset;
@@ -238,19 +238,19 @@ public:
     header.floating_point = std::string("IEEE64BIG");
     header.data_type      = std::string("4D_SU3_GAUGE_3x3");
     GaugeSimpleUnmunger<fobj3D,sobj> munge;
-    if ( grid->IsBoss() ) { 
-      truncate(file);
-      offset = writeHeader(header,file);
-    }
-    grid->Broadcast(0,(void *)&offset,sizeof(offset));
+	if ( grid->IsBoss() ) { 
+	  truncate(file);
+    offset = writeHeader(header,file);
+	}
+	grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
 					      nersc_csum,scidac_csuma,scidac_csumb);
     header.checksum = nersc_csum;
-    if ( grid->IsBoss() ) { 
-      writeHeader(header,file);
-    }
+	if ( grid->IsBoss() ) { 
+    writeHeader(header,file);
+	}
 
     std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
 	     <<std::hex<<header.checksum

Grid/parallelIO/OpenQcdIO.h

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

Grid/parallelIO/OpenQcdIOChromaReference.h

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

Grid/qcd/QCD.h

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

Grid/qcd/action/fermion/FermionOperator.h

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

Grid/qcd/action/fermion/GparityWilsonImpl.h

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

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

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

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

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

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

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

Grid/qcd/action/fermion/StaggeredKernels.h

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

Grid/qcd/action/fermion/WilsonFermion.h

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

Grid/qcd/action/fermion/WilsonFermion5D.h

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

Grid/qcd/action/fermion/g5HermitianLinop.h

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

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

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

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

@@ -146,7 +146,7 @@ NAMESPACE_BEGIN(Grid);
 
 
 template <class Impl>
-template <int Naik> accelerator_inline
+template <int Naik>
 void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
 					  DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
 					  SiteSpinor *buf, int sF, int sU, 
@@ -221,7 +221,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
 
 
 template <class Impl>
-template <int Naik> accelerator_inline
+template <int Naik>
 void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st, 
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 					     SiteSpinor *buf, int sF, int sU, 
@@ -300,7 +300,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
 
 
 template <class Impl>
-template <int Naik> accelerator_inline
+template <int Naik>
 void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 					     SiteSpinor *buf, int sF, int sU, 

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

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

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

@@ -92,16 +92,20 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   int lvol = _Umu.Grid()->lSites();
   int DimRep = Impl::Dimension;
 
+  Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+  Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+
+  Coordinate lcoor;
+  typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
+
   {
     autoView(CTv,CloverTerm,CpuRead);
     autoView(CTIv,CloverTermInv,CpuWrite);
-    thread_for(site, lvol, {
-      Coordinate lcoor;
+    for (int site = 0; site < lvol; site++) {
       grid->LocalIndexToLocalCoor(site, lcoor);
-      Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-      Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-      typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
+      EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
       peekLocalSite(Qx, CTv, lcoor);
+      Qxinv = Zero();
       //if (csw!=0){
       for (int j = 0; j < Ns; j++)
 	for (int k = 0; k < Ns; k++)
@@ -122,7 +126,7 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
       //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
       //  }
       pokeLocalSite(Qxinv, CTIv, lcoor);
-    });
+    }
   }
 
   // Separate the even and odd parts

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

@@ -38,6 +38,9 @@ Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 // undefine everything related to kernels
 #include <simd/Fujitsu_A64FX_undef.h>
 
+// enable A64FX body
+#define WILSONKERNELSASMBODYA64FX
+//#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
 
     ///////////////////////////////////////////////////////////
     // If we are A64FX specialise the single precision routine
@@ -60,89 +63,119 @@ Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 #define INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #define EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 
 /////////////////////////////////////////////////////////////////
@@ -152,89 +185,119 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
 #define INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #define EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 
 // undefine
@@ -267,89 +330,119 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFie
 #define INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #define EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 /////////////////////////////////////////////////////////////////
 // XYZT vectorised, dag Kernel, double
@@ -358,93 +451,124 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
 #define INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #define INTERIOR
 #undef EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 #undef INTERIOR_AND_EXTERIOR
 #undef INTERIOR
 #define EXTERIOR
-
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
-#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
 template<> void
 WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
-
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
 
 
 
 // undefs
+#undef WILSONKERNELSASMBODYA64FX
 #include <simd/Fujitsu_A64FX_undef.h>
 
 #endif //A64FXASM

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

@@ -25,11 +25,6 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-
-// GCC 10 messes up SVE instruction scheduling using -O3, but
-// -O3 -fno-schedule-insns -fno-schedule-insns2 does wonders
-// performance now is better than armclang 20.2
-
 #ifdef KERNEL_DAG
 #define DIR0_PROJ    XP_PROJ
 #define DIR1_PROJ    YP_PROJ
@@ -102,7 +97,7 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     PROJ;							                        \
     MAYBEPERM(PERMUTE_DIR,perm);					        \
       } else {								                \
-	  LOAD_CHI(base);							                \
+	LOAD_CHI(base);							                \
       }									                    \
       base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
     MULT_2SPIN_1(Dir);					                    \
@@ -115,11 +110,6 @@ Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
     }                                                       \
     RECON;								                    \
 
-/*
-NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
-    though I expected that it would improve on performance
-*/
-
 #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
   base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
   PREFETCH1_CHIMU(base);						            \
@@ -136,63 +126,73 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
 
 #define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
       basep = st.GetPFInfo(nent,plocal); nent++;			\
-      if ( local ) {							\
-  LOAD_CHIMU(base);                                       \
-  LOAD_TABLE(PERMUTE_DIR);                                \
-  PROJ;							                        \
-  MAYBEPERM(PERMUTE_DIR,perm);					        \
-      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}			\
-      if ( local || st.same_node[Dir] ) {				\
-  MULT_2SPIN_1(Dir);					                    \
-  MULT_2SPIN_2;					                        \
-  RECON;								\
-      }									\
-  base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
-  PREFETCH_CHIMU(base);						\
-  PREFETCH_CHIMU_L2(basep);                               \
+      if ( local ) {							            \
+    LOAD_CHIMU(base);                                       \
+    LOAD_TABLE(PERMUTE_DIR);                                \
+    PROJ;							                        \
+    MAYBEPERM(PERMUTE_DIR,perm);					        \
+      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}	    \
+      base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
+      if ( local || st.same_node[Dir] ) {				    \
+    MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+    MULT_2SPIN_2;					                        \
+    if (s == 0) {                                           \
+       if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    PREFETCH_CHIMU_L2(basep);                               \
+      } else { PREFETCH_CHIMU(base); }								                    \
 
 #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
   base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
   PREFETCH1_CHIMU(base);						\
-  { ZERO_PSI; }								\
   ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
 
 #define RESULT(base,basep) SAVE_RESULT(base,basep);
 
 #endif
-
 ////////////////////////////////////////////////////////////////////////////////
 // Post comms kernel
 ////////////////////////////////////////////////////////////////////////////////
 #ifdef EXTERIOR
 
+
 #define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
-  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
-  if((!local)&&(!st.same_node[Dir]) ) {					\
-    LOAD_CHI(base);							\
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
+  if((!local)&&(!st.same_node[Dir]) ) {					    \
+    LOAD_CHI(base);							                \
     MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    RECON;								\
-    nmu++;								\
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    nmu++;								                    \
   }
 
-#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
-  nmu=0;								\
-  { ZERO_PSI;}								\
-  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;		\
-  if((!local)&&(!st.same_node[Dir]) ) {					\
-    LOAD_CHI(base);							\
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
+  nmu=0;								                    \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;\
+  if((!local)&&(!st.same_node[Dir]) ) {					    \
+    LOAD_CHI(base);							                \
     MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
     MULT_2SPIN_2;					                        \
-    RECON;								\
-    nmu++;								\
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    nmu++;								                    \
   }
 
 #define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
 
 #endif
-
-
 {
   int nmu;
   int local,perm, ptype;
@@ -209,6 +209,7 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
     int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
     //    int sUn=lo.Reorder(ssn);
     int sUn=ssn;
+    LOCK_GAUGE(0);
 #else
     int sU =ssU;
     int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
@@ -294,11 +295,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       std::cout << "----------------------------------------------------" << std::endl;
 #endif
 
-      // DC ZVA test
-      // { uint64_t basestore = (uint64_t)&out[ss];
-      //   PREFETCH_RESULT_L2_STORE(basestore); }
-
-
       ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJ,DIR5_RECON);
 
 #ifdef SHOW
@@ -312,11 +308,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       std::cout << "----------------------------------------------------" << std::endl;
 #endif
 
-      // DC ZVA test
-      //{ uint64_t basestore = (uint64_t)&out[ss];
-      //  PREFETCH_RESULT_L2_STORE(basestore); }
-
-
       ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJ,DIR6_RECON);
 
 #ifdef SHOW
@@ -330,11 +321,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       std::cout << "----------------------------------------------------" << std::endl;
 #endif
 
-      // DC ZVA test
-      //{ uint64_t basestore = (uint64_t)&out[ss];
-      //  PREFETCH_RESULT_L2_STORE(basestore); }
-
-
       ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJ,DIR7_RECON);
 
 #ifdef SHOW
@@ -355,7 +341,6 @@ NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
       base = (uint64_t) &out[ss];
       basep= st.GetPFInfo(nent,plocal); ent++;
       basep = (uint64_t) &out[ssn];
-      //PREFETCH_RESULT_L1_STORE(base);
       RESULT(base,basep);
 
 #ifdef SHOW

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

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

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

@@ -495,7 +495,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl> accelerator_inline void 
+template<class Impl> void 
 WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
@@ -519,7 +519,7 @@ WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,Site
   HAND_RESULT(ss);
 }
 
-template<class Impl>  accelerator_inline
+template<class Impl>
 void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
@@ -542,7 +542,7 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView
   HAND_RESULT(ss);
 }
 
-template<class Impl>  accelerator_inline void 
+template<class Impl> void 
 WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
@@ -566,7 +566,7 @@ WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,Si
   HAND_RESULT(ss);
 }
 
-template<class Impl> accelerator_inline
+template<class Impl>
 void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
@@ -589,7 +589,7 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldVi
   HAND_RESULT(ss);
 }
 
-template<class Impl>  accelerator_inline void 
+template<class Impl> void 
 WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
@@ -614,7 +614,7 @@ WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,Si
   HAND_RESULT_EXT(ss);
 }
 
-template<class Impl>  accelerator_inline
+template<class Impl>
 void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {

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

@@ -114,7 +114,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   ////////////////////////////////////////////////////////////////////
   // All legs kernels ; comms then compute
   ////////////////////////////////////////////////////////////////////
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,
 					     SiteHalfSpinor *buf, int sF,
 					     int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -140,7 +140,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldV
   coalescedWrite(out[sF],result,lane);
 };
 
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U,
 					  SiteHalfSpinor *buf, int sF,
 					  int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -169,7 +169,7 @@ void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView
   ////////////////////////////////////////////////////////////////////
   // Interior kernels
   ////////////////////////////////////////////////////////////////////
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFieldView &U,
 						SiteHalfSpinor *buf, int sF,
 						int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -197,7 +197,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFi
   coalescedWrite(out[sF], result,lane);
 };
 
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U,
 							 SiteHalfSpinor *buf, int sF,
 							 int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -227,7 +227,7 @@ void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeField
 ////////////////////////////////////////////////////////////////////
 // Exterior kernels
 ////////////////////////////////////////////////////////////////////
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFieldView &U,
 						SiteHalfSpinor *buf, int sF,
 						int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -258,7 +258,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFi
   }
 };
 
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U,
 					     SiteHalfSpinor *buf, int sF,
 					     int sU, const FermionFieldView &in, FermionFieldView &out)
@@ -290,7 +290,7 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeField
 };
 
 #define DhopDirMacro(Dir,spProj,spRecon)	\
-  template <class Impl> accelerator_inline				\
+  template <class Impl>							\
   void WilsonKernels<Impl>::DhopDir##Dir(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF, \
 					 int sU, const FermionFieldView &in, FermionFieldView &out, int dir) \
   {									\
@@ -318,7 +318,7 @@ DhopDirMacro(Ym,spProjYm,spReconYm);
 DhopDirMacro(Zm,spProjZm,spReconZm);
 DhopDirMacro(Tm,spProjTm,spReconTm);
 
-template <class Impl> accelerator_inline
+template <class Impl>
 void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF,
 				    int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int gamma)
 {

Grid/qcd/action/gauge/Gauge.cc

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

Grid/qcd/action/gauge/GaugeImplTypes.h

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

Grid/qcd/action/gauge/GaugeImplementations.h

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

Grid/qcd/action/scalar/ScalarImpl.h

@@ -54,10 +54,6 @@ public:
   static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
     U = 1.0;
   }
-
-  static inline void Project(Field &U) {
-    return;
-  }
     
   static void MomentumSpacePropagator(Field &out, RealD m)
   {
@@ -238,10 +234,6 @@ 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/hmc/GenericHMCrunner.h

@@ -159,13 +159,6 @@ 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

@@ -95,7 +95,7 @@ private:
 
   typedef typename IntegratorType::Field Field;
   typedef std::vector< HmcObservable<Field> * > ObsListType;
-
+  
   //pass these from the resource manager
   GridSerialRNG &sRNG;   
   GridParallelRNG &pRNG; 

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()){
@@ -82,6 +82,7 @@ public:
       abort();
     };
   }
+
   virtual void initialize(const CheckpointerParameters &Params) = 0;
 
   virtual void CheckpointRestore(int traj, typename Impl::Field &U,

Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h

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

Grid/qcd/hmc/checkpointers/NerscCheckpointer.h

@@ -43,8 +43,7 @@ private:
 
 public:
   INHERIT_GIMPL_TYPES(Gimpl);  // only for gauge configurations
-  typedef GaugeStatistics<Gimpl> GaugeStats;
-  
+
   NerscHmcCheckpointer(const CheckpointerParameters &Params_) { initialize(Params_); }
 
   void initialize(const CheckpointerParameters &Params_) {
@@ -61,7 +60,7 @@ public:
       int precision32 = 1;
       int tworow = 0;
       NerscIO::writeRNGState(sRNG, pRNG, rng);
-      NerscIO::writeConfiguration<GaugeStats>(U, config, tworow, precision32);
+      NerscIO::writeConfiguration(U, config, tworow, precision32);
     }
   };
 
@@ -75,7 +74,7 @@ public:
 
     FieldMetaData header;
     NerscIO::readRNGState(sRNG, pRNG, header, rng);
-    NerscIO::readConfiguration<GaugeStats>(U, header, config);
+    NerscIO::readConfiguration(U, header, config);
   };
 };
 

Grid/qcd/hmc/integrators/Integrator.h

@@ -313,8 +313,6 @@ 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/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/utils/BaryonUtils.h

@@ -51,7 +51,7 @@ public:
 
   private: 
   template <class mobj, class robj>
-  static void BaryonSite(const mobj &D1,
+  static void baryon_site(const mobj &D1,
 				 const mobj &D2,
 				 const mobj &D3,
 				 const Gamma GammaA_left,
@@ -61,18 +61,8 @@ public:
 				 const int parity,
 				 const bool * wick_contractions,
   				 robj &result);
-  template <class mobj, class robj>
-  static void BaryonSiteMatrix(const mobj &D1,
-         const mobj &D2,
-         const mobj &D3,
-         const Gamma GammaA_left,
-         const Gamma GammaB_left,
-         const Gamma GammaA_right,
-         const Gamma GammaB_right,
-         const bool * wick_contractions,
-           robj &result);
   public:
-  static void WickContractions(std::string qi, 
+  static void Wick_Contractions(std::string qi, 
                  std::string qf, 
                  bool* wick_contractions);
   static void ContractBaryons(const PropagatorField &q1_left,
@@ -85,17 +75,8 @@ public:
 				 const bool* wick_contractions,
 				 const int parity,
 				 ComplexField &baryon_corr);
-  static void ContractBaryonsMatrix(const PropagatorField &q1_left,
-         const PropagatorField &q2_left,
-         const PropagatorField &q3_left,
-         const Gamma GammaA_left,
-         const Gamma GammaB_left,
-         const Gamma GammaA_right,
-         const Gamma GammaB_right,
-         const bool* wick_contractions,
-         SpinMatrixField &baryon_corr);
   template <class mobj, class robj>
-  static void ContractBaryonsSliced(const mobj &D1,
+  static void ContractBaryons_Sliced(const mobj &D1,
 				 const mobj &D2,
 				 const mobj &D3,
 				 const Gamma GammaA_left,
@@ -106,20 +87,9 @@ public:
 				 const int parity,
 				 const int nt,
 				 robj &result);
-  template <class mobj, class robj>
-  static void ContractBaryonsSlicedMatrix(const mobj &D1,
-         const mobj &D2,
-         const mobj &D3,
-         const Gamma GammaA_left,
-         const Gamma GammaB_left,
-         const Gamma GammaA_right,
-         const Gamma GammaB_right,
-         const bool* wick_contractions,
-         const int nt,
-         robj &result);
   private:
   template <class mobj, class mobj2, class robj>
-  static void BaryonGamma3ptGroup1Site(
+  static void Baryon_Gamma_3pt_Group1_Site(
            const mobj &Dq1_ti,
            const mobj2 &Dq2_spec,
            const mobj2 &Dq3_spec,
@@ -131,7 +101,7 @@ public:
            robj &result);
 
   template <class mobj, class mobj2, class robj>
-  static void BaryonGamma3ptGroup2Site(
+  static void Baryon_Gamma_3pt_Group2_Site(
            const mobj2 &Dq1_spec,
            const mobj &Dq2_ti,
            const mobj2 &Dq3_spec,
@@ -143,7 +113,7 @@ public:
            robj &result);
 
   template <class mobj, class mobj2, class robj>
-  static void BaryonGamma3ptGroup3Site(
+  static void Baryon_Gamma_3pt_Group3_Site(
            const mobj2 &Dq1_spec,
            const mobj2 &Dq2_spec,
            const mobj &Dq3_ti,
@@ -155,7 +125,7 @@ public:
            robj &result);
   public:
   template <class mobj>
-  static void BaryonGamma3pt(
+  static void Baryon_Gamma_3pt(
            const PropagatorField &q_ti,
            const mobj &Dq_spec1,
            const mobj &Dq_spec2,
@@ -168,7 +138,7 @@ public:
            SpinMatrixField &stn_corr);
   private: 
   template <class mobj, class mobj2, class robj>
-  static void SigmaToNucleonQ1EyeSite(const mobj &Dq_loop,
+  static void Sigma_to_Nucleon_Q1_Eye_site(const mobj &Dq_loop,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
 						 const mobj &Ds_ti,
@@ -177,7 +147,7 @@ public:
 		                 		 const Gamma GammaB_nucl,
 						 robj &result);
   template <class mobj, class mobj2, class robj>
-  static void SigmaToNucleonQ1NonEyeSite(const mobj &Du_ti,
+  static void Sigma_to_Nucleon_Q1_NonEye_site(const mobj &Du_ti,
 						 const mobj &Du_tf,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
@@ -189,7 +159,7 @@ public:
 
 
   template <class mobj, class mobj2, class robj>
-  static void SigmaToNucleonQ2EyeSite(const mobj &Dq_loop,
+  static void Sigma_to_Nucleon_Q2_Eye_site(const mobj &Dq_loop,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
 						 const mobj &Ds_ti,
@@ -198,7 +168,7 @@ public:
 		                 		 const Gamma GammaB_nucl,
 						 robj &result);
   template <class mobj, class mobj2, class robj>
-  static void SigmaToNucleonQ2NonEyeSite(const mobj &Du_ti,
+  static void Sigma_to_Nucleon_Q2_NonEye_site(const mobj &Du_ti,
 						 const mobj &Du_tf,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
@@ -209,7 +179,7 @@ public:
 						 robj &result);
   public:
   template <class mobj>
-  static void SigmaToNucleonEye(const PropagatorField &qq_loop,
+  static void Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
 				 const mobj &Du_spec,
 				 const PropagatorField &qd_tf,
 				 const PropagatorField &qs_ti,
@@ -219,7 +189,7 @@ public:
 		                 const std::string op,
 				 SpinMatrixField &stn_corr);
   template <class mobj>
-  static void SigmaToNucleonNonEye(const PropagatorField &qq_ti,
+  static void Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
 				 const PropagatorField &qq_tf,
 				 const mobj &Du_spec,
 				 const PropagatorField &qd_tf,
@@ -247,7 +217,7 @@ const Real BaryonUtils<FImpl>::epsilon_sgn[6] = {1.,1.,1.,-1.,-1.,-1.};
 //This is the old version
 template <class FImpl>
 template <class mobj, class robj>
-void BaryonUtils<FImpl>::BaryonSite(const mobj &D1,
+void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
                 const mobj &D2,
                 const mobj &D3,
                          const Gamma GammaA_i,
@@ -359,132 +329,12 @@ void BaryonUtils<FImpl>::BaryonSite(const mobj &D1,
     }}
 }
 
-//New version without parity projection or trace
-template <class FImpl>
-template <class mobj, class robj>
-void BaryonUtils<FImpl>::BaryonSiteMatrix(const mobj &D1,
-                const mobj &D2,
-                const mobj &D3,
-                         const Gamma GammaA_i,
-                         const Gamma GammaB_i,
-                         const Gamma GammaA_f,
-                         const Gamma GammaB_f,
-                const bool * wick_contraction,
-                robj &result)
-{
-
-    auto D1_GAi =  D1 * GammaA_i;
-    auto GAf_D1_GAi = GammaA_f * D1_GAi;
-    auto GBf_D1_GAi = GammaB_f * D1_GAi;
-
-    auto D2_GBi = D2 * GammaB_i;
-    auto GBf_D2_GBi = GammaB_f * D2_GBi;
-    auto GAf_D2_GBi = GammaA_f * D2_GBi;
-
-    auto GBf_D3 = GammaB_f * D3;
-    auto GAf_D3 = GammaA_f * D3;
-
-    for (int ie_f=0; ie_f < 6 ; ie_f++){
-        int a_f = epsilon[ie_f][0]; //a
-        int b_f = epsilon[ie_f][1]; //b
-        int c_f = epsilon[ie_f][2]; //c
-    for (int ie_i=0; ie_i < 6 ; ie_i++){
-        int a_i = epsilon[ie_i][0]; //a'
-        int b_i = epsilon[ie_i][1]; //b'
-        int c_i = epsilon[ie_i][2]; //c'
-
-        Real ee = epsilon_sgn[ie_f] * epsilon_sgn[ie_i];
-        //This is the \delta_{456}^{123} part
-        if (wick_contraction[0]){
-            for (int rho_i=0; rho_i<Ns; rho_i++){
-            for (int rho_f=0; rho_f<Ns; rho_f++){
-                auto GAf_D1_GAi_rr_cc = GAf_D1_GAi()(rho_f,rho_i)(c_f,c_i);
-                for (int alpha_f=0; alpha_f<Ns; alpha_f++){
-                for (int beta_i=0; beta_i<Ns; beta_i++){
-                    result()(rho_f,rho_i)() += ee  * GAf_D1_GAi_rr_cc
-                                        * D2_GBi    ()(alpha_f,beta_i)(a_f,a_i)
-                                        * GBf_D3    ()(alpha_f,beta_i)(b_f,b_i);
-                }}
-            }}
-        }   
-        //This is the \delta_{456}^{231} part
-        if (wick_contraction[1]){
-            for (int rho_i=0; rho_i<Ns; rho_i++){
-            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
-                auto D1_GAi_ar_ac = D1_GAi()(alpha_f,rho_i)(a_f,c_i);
-                for (int beta_i=0; beta_i<Ns; beta_i++){
-                  auto GBf_D2_GBi_ab_ba = GBf_D2_GBi ()(alpha_f,beta_i)(b_f,a_i);
-                for (int rho_f=0; rho_f<Ns; rho_f++){
-                    result()(rho_f,rho_i)() += ee  * D1_GAi_ar_ac
-                                        * GBf_D2_GBi_ab_ba
-                                        * GAf_D3        ()(rho_f,beta_i)(c_f,b_i);
-                }}
-            }}
-        }   
-        //This is the \delta_{456}^{312} part
-        if (wick_contraction[2]){
-            for (int rho_i=0; rho_i<Ns; rho_i++){
-            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
-                auto GBf_D1_GAi_ar_bc = GBf_D1_GAi()(alpha_f,rho_i)(b_f,c_i);
-                for (int beta_i=0; beta_i<Ns; beta_i++){
-                  auto D3_ab_ab = D3 ()(alpha_f,beta_i)(a_f,b_i);
-                for (int rho_f=0; rho_f<Ns; rho_f++){
-                    result()(rho_f,rho_i)() += ee  * GBf_D1_GAi_ar_bc
-                                        * GAf_D2_GBi    ()(rho_f,beta_i)(c_f,a_i)
-                                        * D3_ab_ab;
-                }}
-            }}
-        }   
-        //This is the \delta_{456}^{132} part
-        if (wick_contraction[3]){
-            for (int rho_i=0; rho_i<Ns; rho_i++){
-            for (int rho_f=0; rho_f<Ns; rho_f++){
-                auto GAf_D1_GAi_rr_cc = GAf_D1_GAi()(rho_f,rho_i)(c_f,c_i);
-                for (int alpha_f=0; alpha_f<Ns; alpha_f++){
-                for (int beta_i=0; beta_i<Ns; beta_i++){
-                    result()(rho_f,rho_i)() -= ee  * GAf_D1_GAi_rr_cc
-                                        * GBf_D2_GBi    ()(alpha_f,beta_i)(b_f,a_i)
-                                        * D3            ()(alpha_f,beta_i)(a_f,b_i);
-                }}
-            }}
-        }   
-        //This is the \delta_{456}^{321} part
-        if (wick_contraction[4]){
-            for (int rho_i=0; rho_i<Ns; rho_i++){
-            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
-                auto GBf_D1_GAi_ar_bc = GBf_D1_GAi()(alpha_f,rho_i)(b_f,c_i);
-                for (int beta_i=0; beta_i<Ns; beta_i++){
-                  auto D2_GBi_ab_aa = D2_GBi()(alpha_f,beta_i)(a_f,a_i);
-                for (int rho_f=0; rho_f<Ns; rho_f++){
-                    result()(rho_f,rho_i)() -= ee  * GBf_D1_GAi_ar_bc
-                                        * D2_GBi_ab_aa
-                                        * GAf_D3    ()(rho_f,beta_i)(c_f,b_i);
-                }}
-            }}
-        }   
-        //This is the \delta_{456}^{213} part
-        if (wick_contraction[5]){
-            for (int rho_i=0; rho_i<Ns; rho_i++){
-            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
-                auto D1_GAi_ar_ac = D1_GAi()(alpha_f,rho_i)(a_f,c_i);
-                for (int beta_i=0; beta_i<Ns; beta_i++){
-                  auto GBf_D3_ab_bb = GBf_D3()(alpha_f,beta_i)(b_f,b_i);
-                for (int rho_f=0; rho_f<Ns; rho_f++){
-                    result()(rho_f,rho_i)() -= ee  * D1_GAi_ar_ac
-                                        * GAf_D2_GBi    ()(rho_f,beta_i)(c_f,a_i)
-                                        * GBf_D3_ab_bb;
-                }}
-            }}
-        }
-    }}
-}
-
 /* Computes which wick contractions should be performed for a    *
  * baryon 2pt function given the initial and finals state quark  *
  * flavours.                                                     *
  * The array wick_contractions must be of length 6               */
 template<class FImpl>
-void BaryonUtils<FImpl>::WickContractions(std::string qi, std::string qf, bool* wick_contractions) {
+void BaryonUtils<FImpl>::Wick_Contractions(std::string qi, std::string qf, bool* wick_contractions) {
     const int epsilon[6][3] = {{0,1,2},{1,2,0},{2,0,1},{0,2,1},{2,1,0},{1,0,2}};
     for (int ie=0; ie < 6 ; ie++) {
         wick_contractions[ie] = (qi.size() == 3 && qf.size() == 3
@@ -514,6 +364,11 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
 
   assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
   assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
+
+  std::cout << "GammaA (left) " << (GammaA_left.g) <<  std::endl;
+  std::cout << "GammaB (left) " << (GammaB_left.g) <<  std::endl;
+  std::cout << "GammaA (right) " << (GammaA_right.g) <<  std::endl;
+  std::cout << "GammaB (right) " << (GammaB_right.g) <<  std::endl;
  
   assert(parity==1 || parity == -1 && "Parity must be +1 or -1");
 
@@ -542,62 +397,13 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
     auto D2 = v2[ss];
     auto D3 = v3[ss];
     vobj result=Zero();
-    BaryonSite(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contractions,result);
+    baryon_site(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contractions,result);
     vbaryon_corr[ss] = result; 
   }  );//end loop over lattice sites
 
   t += usecond();
 
-  std::cout << GridLogDebug << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s " << std::endl;
-}
-
-template<class FImpl>
-void BaryonUtils<FImpl>::ContractBaryonsMatrix(const PropagatorField &q1_left,
-             const PropagatorField &q2_left,
-             const PropagatorField &q3_left,
-                         const Gamma GammaA_left,
-                         const Gamma GammaB_left,
-                         const Gamma GammaA_right,
-                         const Gamma GammaB_right,
-             const bool* wick_contractions,
-             SpinMatrixField &baryon_corr)
-{
-
-  assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
-  assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
- 
-  GridBase *grid = q1_left.Grid();
-  
-  autoView(vbaryon_corr, baryon_corr,CpuWrite);
-  autoView( v1 , q1_left, CpuRead);
-  autoView( v2 , q2_left, CpuRead);
-  autoView( v3 , q3_left, CpuRead);
-
-  // Real bytes =0.;
-  // bytes += grid->oSites() * (432.*sizeof(vComplex) + 126.*sizeof(int) + 36.*sizeof(Real));
-  // for (int ie=0; ie < 6 ; ie++){
-  //   if(ie==0 or ie==3){
-  //      bytes += grid->oSites() * (4.*sizeof(int) + 4752.*sizeof(vComplex)) * wick_contractions[ie];
-  //   }
-  //   else{
-  //      bytes += grid->oSites() * (64.*sizeof(int) + 5184.*sizeof(vComplex)) * wick_contractions[ie];
-  //   }
-  // }
-  // Real t=0.;
-  // t =-usecond();
-
-  accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
-    auto D1 = v1[ss];
-    auto D2 = v2[ss];
-    auto D3 = v3[ss];
-    sobj result=Zero();
-    BaryonSiteMatrix(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,wick_contractions,result);
-    vbaryon_corr[ss] = result; 
-  }  );//end loop over lattice sites
-
-  // t += usecond();
-
-  // std::cout << GridLogDebug << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s " << std::endl;
+  std::cout << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s " << std::endl;
 
 }
 
@@ -608,7 +414,7 @@ void BaryonUtils<FImpl>::ContractBaryonsMatrix(const PropagatorField &q1_left,
  * Wick_Contractions function above                               */
 template <class FImpl>
 template <class mobj, class robj>
-void BaryonUtils<FImpl>::ContractBaryonsSliced(const mobj &D1,
+void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
 						 const mobj &D2,
 						 const mobj &D3,
 				                 const Gamma GammaA_left,
@@ -623,33 +429,16 @@ void BaryonUtils<FImpl>::ContractBaryonsSliced(const mobj &D1,
 
   assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
   assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
+
+  std::cout << "GammaA (left) " << (GammaA_left.g) <<  std::endl;
+  std::cout << "GammaB (left) " << (GammaB_left.g) <<  std::endl;
+  std::cout << "GammaA (right) " << (GammaA_right.g) <<  std::endl;
+  std::cout << "GammaB (right) " << (GammaB_right.g) <<  std::endl;
  
   assert(parity==1 || parity == -1 && "Parity must be +1 or -1");
 
   for (int t=0; t<nt; t++) {
-    BaryonSite(D1[t],D2[t],D3[t],GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contractions,result[t]);
-  }
-}
-
-template <class FImpl>
-template <class mobj, class robj>
-void BaryonUtils<FImpl>::ContractBaryonsSlicedMatrix(const mobj &D1,
-             const mobj &D2,
-             const mobj &D3,
-                         const Gamma GammaA_left,
-                         const Gamma GammaB_left,
-                         const Gamma GammaA_right,
-                         const Gamma GammaB_right,
-             const bool* wick_contractions,
-             const int nt,
-             robj &result)
-{
-
-  assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
-  assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
-
-  for (int t=0; t<nt; t++) {
-    BaryonSiteMatrix(D1[t],D2[t],D3[t],GammaA_left,GammaB_left,GammaA_right,GammaB_right,wick_contractions,result[t]);
+    baryon_site(D1[t],D2[t],D3[t],GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contractions,result[t]);
   }
 }
 
@@ -665,7 +454,7 @@ void BaryonUtils<FImpl>::ContractBaryonsSlicedMatrix(const mobj &D1,
  * Dq4_tf is a quark line from t_f to t_J */
 template<class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::BaryonGamma3ptGroup1Site(
+void BaryonUtils<FImpl>::Baryon_Gamma_3pt_Group1_Site(
                         const mobj &Dq1_ti,
                         const mobj2 &Dq2_spec,
                         const mobj2 &Dq3_spec,
@@ -757,7 +546,7 @@ void BaryonUtils<FImpl>::BaryonGamma3ptGroup1Site(
  * Dq4_tf is a quark line from t_f to t_J */
 template<class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::BaryonGamma3ptGroup2Site(
+void BaryonUtils<FImpl>::Baryon_Gamma_3pt_Group2_Site(
                         const mobj2 &Dq1_spec,
                         const mobj &Dq2_ti,
                         const mobj2 &Dq3_spec,
@@ -847,7 +636,7 @@ void BaryonUtils<FImpl>::BaryonGamma3ptGroup2Site(
  * Dq4_tf is a quark line from t_f to t_J */
 template<class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::BaryonGamma3ptGroup3Site(
+void BaryonUtils<FImpl>::Baryon_Gamma_3pt_Group3_Site(
                         const mobj2 &Dq1_spec,
                         const mobj2 &Dq2_spec,
                         const mobj &Dq3_ti,
@@ -939,7 +728,7 @@ void BaryonUtils<FImpl>::BaryonGamma3ptGroup3Site(
  * https://aportelli.github.io/Hadrons-doc/#/mcontraction        */
 template<class FImpl>
 template <class mobj>
-void BaryonUtils<FImpl>::BaryonGamma3pt(
+void BaryonUtils<FImpl>::Baryon_Gamma_3pt(
                         const PropagatorField &q_ti,
                         const mobj &Dq_spec1,
                         const mobj &Dq_spec2,
@@ -962,7 +751,7 @@ void BaryonUtils<FImpl>::BaryonGamma3pt(
             auto Dq_ti = vq_ti[ss];
             auto Dq_tf = vq_tf[ss];
             sobj result=Zero();
-            BaryonGamma3ptGroup1Site(Dq_ti,Dq_spec1,Dq_spec2,Dq_tf,GammaJ,GammaBi,GammaBf,wick_contraction,result);
+            Baryon_Gamma_3pt_Group1_Site(Dq_ti,Dq_spec1,Dq_spec2,Dq_tf,GammaJ,GammaBi,GammaBf,wick_contraction,result);
             vcorr[ss] += result; 
         });//end loop over lattice sites
     } else if (group == 2) {
@@ -970,7 +759,7 @@ void BaryonUtils<FImpl>::BaryonGamma3pt(
             auto Dq_ti = vq_ti[ss];
             auto Dq_tf = vq_tf[ss];
             sobj result=Zero();
-            BaryonGamma3ptGroup2Site(Dq_spec1,Dq_ti,Dq_spec2,Dq_tf,GammaJ,GammaBi,GammaBf,wick_contraction,result);
+            Baryon_Gamma_3pt_Group2_Site(Dq_spec1,Dq_ti,Dq_spec2,Dq_tf,GammaJ,GammaBi,GammaBf,wick_contraction,result);
             vcorr[ss] += result; 
         });//end loop over lattice sites
     } else if (group == 3) {
@@ -978,7 +767,7 @@ void BaryonUtils<FImpl>::BaryonGamma3pt(
             auto Dq_ti = vq_ti[ss];
             auto Dq_tf = vq_tf[ss];
             sobj result=Zero();
-            BaryonGamma3ptGroup3Site(Dq_spec1,Dq_spec2,Dq_ti,Dq_tf,GammaJ,GammaBi,GammaBf,wick_contraction,result);
+            Baryon_Gamma_3pt_Group3_Site(Dq_spec1,Dq_spec2,Dq_ti,Dq_tf,GammaJ,GammaBi,GammaBf,wick_contraction,result);
 
             vcorr[ss] += result; 
         });//end loop over lattice sites
@@ -998,7 +787,7 @@ void BaryonUtils<FImpl>::BaryonGamma3pt(
  * Ds_ti is a quark line from t_i to t_H */
 template <class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::SigmaToNucleonQ1EyeSite(const mobj &Dq_loop,
+void BaryonUtils<FImpl>::Sigma_to_Nucleon_Q1_Eye_site(const mobj &Dq_loop,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
 						 const mobj &Ds_ti,
@@ -1049,7 +838,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonQ1EyeSite(const mobj &Dq_loop,
  * Ds_ti is a quark line from t_i to t_H */
 template <class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::SigmaToNucleonQ1NonEyeSite(const mobj &Du_ti,
+void BaryonUtils<FImpl>::Sigma_to_Nucleon_Q1_NonEye_site(const mobj &Du_ti,
 						 const mobj &Du_tf,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
@@ -1108,7 +897,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonQ1NonEyeSite(const mobj &Du_ti,
  * Ds_ti is a quark line from t_i to t_H */
 template <class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::SigmaToNucleonQ2EyeSite(const mobj &Dq_loop,
+void BaryonUtils<FImpl>::Sigma_to_Nucleon_Q2_Eye_site(const mobj &Dq_loop,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
 						 const mobj &Ds_ti,
@@ -1159,7 +948,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonQ2EyeSite(const mobj &Dq_loop,
  * Ds_ti is a quark line from t_i to t_H */
 template <class FImpl>
 template <class mobj, class mobj2, class robj>
-void BaryonUtils<FImpl>::SigmaToNucleonQ2NonEyeSite(const mobj &Du_ti,
+void BaryonUtils<FImpl>::Sigma_to_Nucleon_Q2_NonEye_site(const mobj &Du_ti,
 						 const mobj &Du_tf,
 						 const mobj2 &Du_spec,
 						 const mobj &Dd_tf,
@@ -1213,7 +1002,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonQ2NonEyeSite(const mobj &Du_ti,
 
 template<class FImpl>
 template <class mobj>
-void BaryonUtils<FImpl>::SigmaToNucleonEye(const PropagatorField &qq_loop,
+void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
 						 const mobj &Du_spec,
 						 const PropagatorField &qd_tf,
 						 const PropagatorField &qs_ti,
@@ -1240,9 +1029,9 @@ void BaryonUtils<FImpl>::SigmaToNucleonEye(const PropagatorField &qq_loop,
     auto Ds_ti = vs_ti[ss];
     sobj result=Zero();
     if(op == "Q1"){
-      SigmaToNucleonQ1EyeSite(Dq_loop,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
+      Sigma_to_Nucleon_Q1_Eye_site(Dq_loop,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
     } else if(op == "Q2"){
-      SigmaToNucleonQ2EyeSite(Dq_loop,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
+      Sigma_to_Nucleon_Q2_Eye_site(Dq_loop,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
     } else {
       assert(0 && "Weak Operator not correctly specified");
     }
@@ -1252,7 +1041,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonEye(const PropagatorField &qq_loop,
 
 template<class FImpl>
 template <class mobj>
-void BaryonUtils<FImpl>::SigmaToNucleonNonEye(const PropagatorField &qq_ti,
+void BaryonUtils<FImpl>::Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
 						 const PropagatorField &qq_tf,
 						 const mobj &Du_spec,
 						 const PropagatorField &qd_tf,
@@ -1282,9 +1071,9 @@ void BaryonUtils<FImpl>::SigmaToNucleonNonEye(const PropagatorField &qq_ti,
     auto Ds_ti = vs_ti[ss];
     sobj result=Zero();
     if(op == "Q1"){
-      SigmaToNucleonQ1NonEyeSite(Dq_ti,Dq_tf,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
+      Sigma_to_Nucleon_Q1_NonEye_site(Dq_ti,Dq_tf,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
     } else if(op == "Q2"){
-      SigmaToNucleonQ2NonEyeSite(Dq_ti,Dq_tf,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
+      Sigma_to_Nucleon_Q2_NonEye_site(Dq_ti,Dq_tf,Du_spec,Dd_tf,Ds_ti,Gamma_H,GammaB_sigma,GammaB_nucl,result);
     } else {
       assert(0 && "Weak Operator not correctly specified");
     }

Grid/qcd/utils/CovariantCshift.h

@@ -53,24 +53,6 @@ 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,
@@ -88,7 +70,6 @@ namespace PeriodicBC {
     return CovShiftBackward(Link,mu,arg);
   }
 
-
 }
 
 
@@ -158,38 +139,6 @@ 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/SUn.h

@@ -449,8 +449,7 @@ public:
     LatticeReal alpha(grid);
 
     //    std::cout<<GridLogMessage<<"xi "<<xi <<std::endl;
-    xi = 2.0 *xi;
-    alpha = toReal(xi);
+    alpha = toReal(2.0 * xi);
 
     do {
       // A. Generate two uniformly distributed pseudo-random numbers R and R',
@@ -735,6 +734,7 @@ public:
     }
   }
 
+
   template <typename GaugeField>
   static void HotConfiguration(GridParallelRNG &pRNG, GaugeField &out) {
     typedef typename GaugeField::vector_type vector_type;
@@ -799,88 +799,6 @@ 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>
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef __NVCC__
 
 NAMESPACE_BEGIN(Grid);
 

Grid/simd/Fujitsu_A64FX_asm_double.h

@@ -0,0 +1,779 @@
+/*************************************************************************************
+
+    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

@@ -0,0 +1,779 @@
+/*************************************************************************************
+
+    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

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

Grid/simd/Fujitsu_A64FX_intrin_single.h

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

Grid/simd/Fujitsu_A64FX_undef.h

@@ -46,7 +46,6 @@ 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

@@ -38,22 +38,9 @@ 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 {
 
-
-
-typedef struct Half2_t { half x; half y; } Half2;
-
 #define COALESCE_GRANULARITY ( GEN_SIMD_WIDTH )
 
 template<class pair>
@@ -138,14 +125,14 @@ inline accelerator GpuVector<N,datum> operator/(const GpuVector<N,datum> l,const
 }
 
 constexpr int NSIMD_RealH    = COALESCE_GRANULARITY / sizeof(half);
-constexpr int NSIMD_ComplexH = COALESCE_GRANULARITY / sizeof(Half2);
+constexpr int NSIMD_ComplexH = COALESCE_GRANULARITY / sizeof(half2);
 constexpr int NSIMD_RealF    = COALESCE_GRANULARITY / sizeof(float);
 constexpr int NSIMD_ComplexF = COALESCE_GRANULARITY / sizeof(float2);
 constexpr int NSIMD_RealD    = COALESCE_GRANULARITY / sizeof(double);
 constexpr int NSIMD_ComplexD = COALESCE_GRANULARITY / sizeof(double2);
 constexpr int NSIMD_Integer  = COALESCE_GRANULARITY / sizeof(Integer);
 
-typedef GpuComplex<Half2  > GpuComplexH;
+typedef GpuComplex<half2  > GpuComplexH;
 typedef GpuComplex<float2 > GpuComplexF;
 typedef GpuComplex<double2> GpuComplexD;
 
@@ -160,11 +147,13 @@ typedef GpuVector<NSIMD_Integer,  Integer     > GpuVectorI;
 accelerator_inline float half2float(half h)
 {
   float f;
-#if defined(GRID_CUDA) || defined(GRID_HIP)
+#ifdef GRID_SIMT
   f = __half2float(h);
 #else 
+  //f = __half2float(h);
+  __half_raw hr(h);
   Grid_half hh; 
-  hh.x = h.x;
+  hh.x = hr.x;
   f=  sfw_half_to_float(hh);
 #endif
   return f;
@@ -172,11 +161,13 @@ accelerator_inline float half2float(half h)
 accelerator_inline half float2half(float f)
 {
   half h;
-#if defined(GRID_CUDA) || defined(GRID_HIP)
+#ifdef GRID_SIMT
   h = __float2half(f);
 #else
   Grid_half hh = sfw_float_to_half(f);
-  h.x = hh.x;
+  __half_raw hr;  
+  hr.x = hh.x;
+  h = __half(hr);
 #endif
   return h;
 }
@@ -532,7 +523,7 @@ namespace Optimization {
     ////////////////////////////////////////////////////////////////////////////////////
     // Single / Half
     ////////////////////////////////////////////////////////////////////////////////////
-     static accelerator_inline GpuVectorCH StoH (GpuVectorCF a,GpuVectorCF b) {
+    static accelerator_inline GpuVectorCH StoH (GpuVectorCF a,GpuVectorCF b) {
       int N = GpuVectorCF::N;
       GpuVectorCH h;
       for(int i=0;i<N;i++) {

Grid/simd/Grid_vector_unops.h

@@ -125,6 +125,14 @@ 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/stencil/Stencil.h

@@ -269,7 +269,7 @@ public:
   std::vector<Vector<std::pair<int,int> > > face_table ;
   Vector<int> surface_list;
 
-  stencilVector<StencilEntry>  _entries; // Resident in managed memory
+  Vector<StencilEntry>  _entries; // Resident in managed memory
   std::vector<Packet> Packets;
   std::vector<Merge> Mergers;
   std::vector<Merge> MergersSHM;

Grid/tensors/Tensor_Ta.h

@@ -95,18 +95,14 @@ accelerator_inline iMatrix<vtype,N> ProjectOnGroup(const iMatrix<vtype,N> &arg)
   vtype nrm;
   vtype inner;
   for(int c1=0;c1<N;c1++){
-
-    // Normalises row c1
     zeroit(inner);	
     for(int c2=0;c2<N;c2++)
       inner += innerProduct(ret._internal[c1][c2],ret._internal[c1][c2]);
 
-    nrm = sqrt(inner);
-    nrm = 1.0/nrm;
+    nrm = rsqrt(inner);
     for(int c2=0;c2<N;c2++)
       ret._internal[c1][c2]*= nrm;
       
-    // Remove c1 from rows c1+1...N-1
     for (int b=c1+1; b<N; ++b){
       decltype(ret._internal[b][b]*ret._internal[b][b]) pr;
       zeroit(pr);
@@ -117,19 +113,7 @@ accelerator_inline iMatrix<vtype,N> ProjectOnGroup(const iMatrix<vtype,N> &arg)
 	ret._internal[b][c] -= pr * ret._internal[c1][c];
       }
     }
-  }
-
-  // Normalise last row
-  {
-    int c1 = N-1;
-    zeroit(inner);	
-    for(int c2=0;c2<N;c2++)
-      inner += innerProduct(ret._internal[c1][c2],ret._internal[c1][c2]);
-
-    nrm = sqrt(inner);
-    nrm = 1.0/nrm;
-    for(int c2=0;c2<N;c2++)
-      ret._internal[c1][c2]*= nrm;
+	  
   }
   // assuming the determinant is ok
   return ret;

Grid/tensors/Tensor_unary.h

@@ -84,6 +84,7 @@ NAMESPACE_BEGIN(Grid);
   }
 
 UNARY(sqrt);
+UNARY(rsqrt);
 UNARY(sin);
 UNARY(cos);
 UNARY(asin);

Grid/threads/Accelerator.cc

@@ -21,26 +21,22 @@ void acceleratorInit(void)
 #define ENV_RANK_SLURM         "SLURM_PROCID"
 #define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
 #define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
+  // We extract the local rank initialization using an environment variable
+  if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL) {
+    printf("OPENMPI detected\n");
+    rank = atoi(localRankStr);		
+  } else if ((localRankStr = getenv(ENV_LOCAL_RANK_MVAPICH)) != NULL) {
+    printf("MVAPICH detected\n");
+    rank = atoi(localRankStr);		
+  } else if ((localRankStr = getenv(ENV_LOCAL_RANK_SLURM)) != NULL) {
+    printf("SLURM detected\n");
+    rank = atoi(localRankStr);		
+  } else { 
+    printf("MPI version is unknown - bad things may happen\n");
+  }
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_SLURM  )) != NULL) { world_rank = atoi(localRankStr);}
-  // We extract the local rank initialization using an environment variable
-  if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL) {
-    if (!world_rank)
-      printf("OPENMPI detected\n");
-    rank = atoi(localRankStr);		
-  } else if ((localRankStr = getenv(ENV_LOCAL_RANK_MVAPICH)) != NULL) {
-    if (!world_rank)
-      printf("MVAPICH detected\n");
-    rank = atoi(localRankStr);		
-  } else if ((localRankStr = getenv(ENV_LOCAL_RANK_SLURM)) != NULL) {
-    if (!world_rank)
-      printf("SLURM detected\n");
-    rank = atoi(localRankStr);		
-  } else { 
-    if (!world_rank)
-      printf("MPI version is unknown - bad things may happen\n");
-  }
 
   size_t totalDeviceMem=0;
   for (int i = 0; i < nDevices; i++) {
@@ -52,14 +48,13 @@ void acceleratorInit(void)
     prop = gpu_props[i];
     totalDeviceMem = prop.totalGlobalMem;
     if ( world_rank == 0) {
-#ifndef GRID_DEFAULT_GPU
+#ifndef GRID_IBM_SUMMIT
       if ( i==rank ) {
 	printf("AcceleratorCudaInit[%d]: ========================\n",rank);
 	printf("AcceleratorCudaInit[%d]: Device Number    : %d\n", rank,i);
 	printf("AcceleratorCudaInit[%d]: ========================\n",rank);
 	printf("AcceleratorCudaInit[%d]: Device identifier: %s\n",rank, prop.name);
 
-
 	GPU_PROP_FMT(totalGlobalMem,"%lld");
 	GPU_PROP(managedMemory);
 	GPU_PROP(isMultiGpuBoard);
@@ -77,17 +72,11 @@ void acceleratorInit(void)
 #undef GPU_PROP_FMT    
 #undef GPU_PROP
 
-#ifdef GRID_DEFAULT_GPU
+#ifdef GRID_IBM_SUMMIT
   // IBM Jsrun makes cuda Device numbering screwy and not match rank
-  if ( world_rank == 0 ) {
-    printf("AcceleratorCudaInit: using default device \n");
-    printf("AcceleratorCudaInit: assume user either uses a) IBM jsrun, or \n");
-    printf("AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
-    printf("AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no \n");
-  }
+  if ( world_rank == 0 )  printf("AcceleratorCudaInit: IBM Summit or similar - use default device\n");
 #else
   printf("AcceleratorCudaInit: rank %d setting device to node rank %d\n",world_rank,rank);
-  printf("AcceleratorCudaInit: Configure options --enable-select-gpu=yes \n");
   cudaSetDevice(rank);
 #endif
   if ( world_rank == 0 )  printf("AcceleratorCudaInit: ================================================\n");
@@ -120,24 +109,20 @@ void acceleratorInit(void)
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
 
-  printf("world_rank %d has %d devices\n",world_rank,nDevices);
-  size_t totalDeviceMem=0;
   for (int i = 0; i < nDevices; i++) {
 
 #define GPU_PROP_FMT(canMapHostMemory,FMT)     printf("AcceleratorHipInit:   " #canMapHostMemory ": " FMT" \n",prop.canMapHostMemory);
 #define GPU_PROP(canMapHostMemory)             GPU_PROP_FMT(canMapHostMemory,"%d");
     
     hipGetDeviceProperties(&gpu_props[i], i);
-    hipDeviceProp_t prop; 
-    prop = gpu_props[i];
-    totalDeviceMem = prop.totalGlobalMem;
     if ( world_rank == 0) {
+      hipDeviceProp_t prop; 
+      prop = gpu_props[i];
       printf("AcceleratorHipInit: ========================\n");
       printf("AcceleratorHipInit: Device Number    : %d\n", i);
       printf("AcceleratorHipInit: ========================\n");
       printf("AcceleratorHipInit: Device identifier: %s\n", prop.name);
 
-      GPU_PROP_FMT(totalGlobalMem,"%lu");
       //      GPU_PROP(managedMemory);
       GPU_PROP(isMultiGpuBoard);
       GPU_PROP(warpSize);
@@ -146,21 +131,13 @@ void acceleratorInit(void)
       //      GPU_PROP(singleToDoublePrecisionPerfRatio);
     }
   }
-  MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
 #undef GPU_PROP_FMT    
 #undef GPU_PROP
-
-#ifdef GRID_DEFAULT_GPU
-  if ( world_rank == 0 ) {
-    printf("AcceleratorHipInit: using default device \n");
-    printf("AcceleratorHipInit: assume user either uses a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
-    printf("AcceleratorHipInit: Configure options --enable-summit, --enable-select-gpu=no \n");
-  }
+#ifdef GRID_IBM_SUMMIT
+  // IBM Jsrun makes cuda Device numbering screwy and not match rank
+  if ( world_rank == 0 )  printf("AcceleratorHipInit: IBM Summit or similar - NOT setting device to node rank\n");
 #else
-  if ( world_rank == 0 ) {
-    printf("AcceleratorHipInit: rank %d setting device to node rank %d\n",world_rank,rank);
-    printf("AcceleratorHipInit: Configure options --enable-select-gpu=yes \n");
-  }
+  if ( world_rank == 0 )  printf("AcceleratorHipInit: setting device to node rank\n");
   hipSetDevice(rank);
 #endif
   if ( world_rank == 0 )  printf("AcceleratorHipInit: ================================================\n");

Grid/threads/Accelerator.h

@@ -151,6 +151,9 @@ inline void *acceleratorAllocShared(size_t bytes)
     ptr = (void *) NULL;
     printf(" cudaMallocManaged failed for %d %s \n",bytes,cudaGetErrorString(err));
   }
+  //  size_t free,total;
+  //  cudaMemGetInfo(&free,&total);
+  //  std::cout << "Malloc managed "<<bytes<<" "<<free<<"/"<<total<<std::endl;
   return ptr;
 };
 inline void *acceleratorAllocDevice(size_t bytes)
@@ -161,23 +164,23 @@ inline void *acceleratorAllocDevice(size_t bytes)
     ptr = (void *) NULL;
     printf(" cudaMalloc failed for %d %s \n",bytes,cudaGetErrorString(err));
   }
+  //  size_t free,total;
+  //  cudaMemGetInfo(&free,&total);
+  //  std::cout << "Malloc device "<<bytes<<" "<<free<<"/"<<total<<std::endl;
   return ptr;
 };
 inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
-inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
 inline int  acceleratorIsCommunicable(void *ptr)
 {
-  //  int uvm=0;
-  //  auto 
-  //  cuerr = cuPointerGetAttribute( &uvm, CU_POINTER_ATTRIBUTE_IS_MANAGED, (CUdeviceptr) ptr);
-  //  assert(cuerr == cudaSuccess );
-  //  if(uvm) return 0;
-  //  else    return 1;
-    return 1;
+  int uvm;
+  auto 
+  cuerr = cuPointerGetAttribute( &uvm, CU_POINTER_ATTRIBUTE_IS_MANAGED, (CUdeviceptr) ptr);
+  assert(cuerr == cudaSuccess );
+  if(uvm) return 0;
+  else    return 1;
 }
 
 #endif
@@ -232,10 +235,8 @@ inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*t
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
-inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
 inline int  acceleratorIsCommunicable(void *ptr)
 {
 #if 0
@@ -312,13 +313,17 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 
 inline void *acceleratorAllocShared(size_t bytes)
 {
+#if 0
   void *ptr=NULL;
   auto err = hipMallocManaged((void **)&ptr,bytes);
   if( err != hipSuccess ) {
     ptr = (void *) NULL;
-    printf(" hipMallocManaged failed for %ld %s \n",bytes,hipGetErrorString(err));
+    printf(" hipMallocManaged failed for %d %s \n",bytes,hipGetErrorString(err));
   }
   return ptr;
+#else
+  return malloc(bytes);
+#endif
 };
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
 
@@ -328,17 +333,15 @@ inline void *acceleratorAllocDevice(size_t bytes)
   auto err = hipMalloc((void **)&ptr,bytes);
   if( err != hipSuccess ) {
     ptr = (void *) NULL;
-    printf(" hipMalloc failed for %ld %s \n",bytes,hipGetErrorString(err));
+    printf(" hipMalloc failed for %d %s \n",bytes,hipGetErrorString(err));
   }
   return ptr;
 };
 
-inline void acceleratorFreeShared(void *ptr){ hipFree(ptr);};
+inline void acceleratorFreeShared(void *ptr){ free(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ hipFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
-inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(base,value,bytes);}
 
 #endif
 
@@ -361,7 +364,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 //////////////////////////////////////////////
 // CPU Target - No accelerator just thread instead
 //////////////////////////////////////////////
-
+#define GRID_ALLOC_ALIGN (2*1024*1024) // 2MB aligned 
 #if ( (!defined(GRID_SYCL)) && (!defined(GRID_CUDA)) && (!defined(GRID_HIP)) )
 
 #undef GRID_SIMT
@@ -376,10 +379,8 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
-inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { memset(base,value,bytes);}
 #ifdef HAVE_MM_MALLOC_H
 inline void *acceleratorAllocShared(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
 inline void *acceleratorAllocDevice(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
@@ -402,8 +403,6 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
 inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 #endif
 
-
-
 ///////////////////////////////////////////////////
 // Synchronise across local threads for divergence resynch
 ///////////////////////////////////////////////////

Grid/util/Init.cc

@@ -473,13 +473,11 @@ void Grid_init(int *argc,char ***argv)
     LebesgueOrder::UseLebesgueOrder=1;
   }
   CartesianCommunicator::nCommThreads = 1;
-#ifdef GRID_COMMS_THREADS  
   if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-threads") ){
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--comms-threads");
     GridCmdOptionInt(arg,CartesianCommunicator::nCommThreads);
     assert(CartesianCommunicator::nCommThreads > 0);
   }
-#endif  
   if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");
     GridCmdOptionIntVector(arg,LebesgueOrder::Block);

README

@@ -111,10 +111,11 @@ Now you can execute the `configure` script to generate makefiles (here from a bu
 
 ``` bash
 mkdir build; cd build
-../configure --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
+../configure --enable-precision=double --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
 ```
 
-where `--enable-simd=` set the SIMD type, `--enable-
+where `--enable-precision=` set the default precision,
+`--enable-simd=` set the SIMD type, `--enable-
 comms=`, and `<path>` should be replaced by the prefix path where you want to
 install Grid. Other options are detailed in the next section, you can also use `configure
 --help` to display them. Like with any other program using GNU autotool, the
@@ -145,8 +146,8 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-numa`: enable NUMA first touch optimisation
 - `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
-- `--enable-precision={single|double}`: set the default precision (default: `double`). **Deprecated option**
-- `--enable-comms=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
+- `--enable-precision={single|double}`: set the default precision (default: `double`).
+- `--enable-precision=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.
 - `--enable-chroma`: enable Chroma regression tests.
@@ -200,7 +201,8 @@ Alternatively, some CPU codenames can be directly used:
 The following configuration is recommended for the Intel Knights Landing platform:
 
 ``` bash
-../configure --enable-simd=KNL        \
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
              --enable-comms=mpi-auto  \
              --enable-mkl             \
              CXX=icpc MPICXX=mpiicpc
@@ -210,7 +212,8 @@ The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
 If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
-../configure --enable-simd=KNL        \
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
              --enable-comms=mpi       \
              --enable-mkl             \
              CXX=CC CC=cc
@@ -229,7 +232,8 @@ for interior communication. This is the mpi3 communications implementation.
 We recommend four ranks per node for best performance, but optimum is local volume dependent.
 
 ``` bash
-../configure --enable-simd=KNL        \
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
              --enable-comms=mpi3-auto \
              --enable-mkl             \
              CC=icpc MPICXX=mpiicpc 
@@ -240,7 +244,8 @@ We recommend four ranks per node for best performance, but optimum is local volu
 The following configuration is recommended for the Intel Haswell platform:
 
 ``` bash
-../configure --enable-simd=AVX2       \
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
              --enable-comms=mpi3-auto \
              --enable-mkl             \
              CXX=icpc MPICXX=mpiicpc
@@ -257,7 +262,8 @@ where `<path>` is the UNIX prefix where GMP and MPFR are installed.
 If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
-../configure --enable-simd=AVX2       \
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
              --enable-comms=mpi3      \
              --enable-mkl             \
              CXX=CC CC=cc
@@ -274,7 +280,8 @@ This is the default.
 The following configuration is recommended for the Intel Skylake platform:
 
 ``` bash
-../configure --enable-simd=AVX512     \
+../configure --enable-precision=double\
+             --enable-simd=AVX512     \
              --enable-comms=mpi3      \
              --enable-mkl             \
              CXX=mpiicpc
@@ -291,7 +298,8 @@ where `<path>` is the UNIX prefix where GMP and MPFR are installed.
 If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
-../configure --enable-simd=AVX512     \
+../configure --enable-precision=double\
+             --enable-simd=AVX512     \
              --enable-comms=mpi3      \
              --enable-mkl             \
              CXX=CC CC=cc
@@ -322,7 +330,8 @@ and 8 threads per rank.
 The following configuration is recommended for the AMD EPYC platform.
 
 ``` bash
-../configure --enable-simd=AVX2       \
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
              --enable-comms=mpi3 \
              CXX=mpicxx 
 ```

README.md

@@ -115,10 +115,11 @@ Now you can execute the `configure` script to generate makefiles (here from a bu
 
 ``` bash
 mkdir build; cd build
-../configure --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
+../configure --enable-precision=double --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
 ```
 
-where `--enable-simd=` set the SIMD type, `--enable-
+where `--enable-precision=` set the default precision,
+`--enable-simd=` set the SIMD type, `--enable-
 comms=`, and `<path>` should be replaced by the prefix path where you want to
 install Grid. Other options are detailed in the next section, you can also use `configure
 --help` to display them. Like with any other program using GNU autotool, the
@@ -149,8 +150,8 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-numa`: enable NUMA first touch optimisation
 - `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
-- `--enable-precision={single|double}`: set the default precision (default: `double`). **Deprecated option**
-- `--enable-comms=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
+- `--enable-precision={single|double}`: set the default precision (default: `double`).
+- `--enable-precision=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.
 - `--enable-chroma`: enable Chroma regression tests.
@@ -204,7 +205,8 @@ Alternatively, some CPU codenames can be directly used:
 The following configuration is recommended for the Intel Knights Landing platform:
 
 ``` bash
-../configure --enable-simd=KNL        \
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
              --enable-comms=mpi-auto  \
              --enable-mkl             \
              CXX=icpc MPICXX=mpiicpc
@@ -214,7 +216,8 @@ The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
 If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
-../configure --enable-simd=KNL        \
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
              --enable-comms=mpi       \
              --enable-mkl             \
              CXX=CC CC=cc
@@ -233,7 +236,8 @@ for interior communication. This is the mpi3 communications implementation.
 We recommend four ranks per node for best performance, but optimum is local volume dependent.
 
 ``` bash
-../configure --enable-simd=KNL        \
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
              --enable-comms=mpi3-auto \
              --enable-mkl             \
              CC=icpc MPICXX=mpiicpc 
@@ -244,7 +248,8 @@ We recommend four ranks per node for best performance, but optimum is local volu
 The following configuration is recommended for the Intel Haswell platform:
 
 ``` bash
-../configure --enable-simd=AVX2       \
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
              --enable-comms=mpi3-auto \
              --enable-mkl             \
              CXX=icpc MPICXX=mpiicpc
@@ -261,7 +266,8 @@ where `<path>` is the UNIX prefix where GMP and MPFR are installed.
 If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
-../configure --enable-simd=AVX2       \
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
              --enable-comms=mpi3      \
              --enable-mkl             \
              CXX=CC CC=cc
@@ -278,7 +284,8 @@ This is the default.
 The following configuration is recommended for the Intel Skylake platform:
 
 ``` bash
-../configure --enable-simd=AVX512     \
+../configure --enable-precision=double\
+             --enable-simd=AVX512     \
              --enable-comms=mpi3      \
              --enable-mkl             \
              CXX=mpiicpc
@@ -295,7 +302,8 @@ where `<path>` is the UNIX prefix where GMP and MPFR are installed.
 If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
-../configure --enable-simd=AVX512     \
+../configure --enable-precision=double\
+             --enable-simd=AVX512     \
              --enable-comms=mpi3      \
              --enable-mkl             \
              CXX=CC CC=cc
@@ -326,7 +334,8 @@ and 8 threads per rank.
 The following configuration is recommended for the AMD EPYC platform.
 
 ``` bash
-../configure --enable-simd=AVX2       \
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
              --enable-comms=mpi3 \
              CXX=mpicxx 
 ```

SVE_README.txt

@@ -12,31 +12,31 @@ module load mpi/openmpi-aarch64
 
 scl enable gcc-toolset-10 bash
 
-../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=g++ CC=gcc CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN"
+../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=g++ CC=gcc CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN"
 
 * gcc 10.1 prebuild w/ MPI, QPACE4 interactive login
 
 scl enable gcc-toolset-10 bash
 module load mpi/openmpi-aarch64
 
-../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=mpi-auto --enable-shm=shmget --enable-openmp CXX=mpicxx CC=mpicc CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN"
+../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=mpi-auto --enable-shm=shmget --enable-openmp CXX=mpicxx CC=mpicc CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN"
 
 ------------------------------------------------------------------------------
 
 * armclang 20.2 (qp4)
 
-../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DA64FX -DARMCLANGCOMPAT -DA64FXASM -DDSLASHINTRIN"
+../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DA64FX -DARMCLANGCOMPAT -DA64FXASM -DDSLASHINTRIN"
 
 ------------------------------------------------------------------------------
 
 * gcc 10.0.1 VLA (merlin)
 
-../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=g++-10.0.1 CC=gcc-10.0.1 CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FX -DA64FXASM -DDSLASHINTRIN" LDFLAGS=-static GRID_LDFLAGS=-static MPI_CXXLDFLAGS=-static
+../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=g++-10.0.1 CC=gcc-10.0.1 CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FX -DA64FXASM -DDSLASHINTRIN" LDFLAGS=-static GRID_LDFLAGS=-static MPI_CXXLDFLAGS=-static
 
 
 * gcc 10.0.1 fixed-size ACLE (merlin)
 
-../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=g++-10.0.1 CC=gcc-10.0.1 CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN"
+../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=g++-10.0.1 CC=gcc-10.0.1 CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN"
 
 
 * gcc 10.0.1 fixed-size ACLE (fjt) w/ MPI
@@ -46,34 +46,34 @@ export OMPI_CXX=g++-10.0.1
 export MPICH_CC=gcc-10.0.1
 export MPICH_CXX=g++-10.0.1
 
-$ ../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=mpi3 --enable-openmp CXX=mpiFCC CC=mpifcc CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN -DTOFU -I/opt/FJSVxtclanga/tcsds-1.2.25/include/mpi/fujitsu -lrt" LDFLAGS="-L/opt/FJSVxtclanga/tcsds-1.2.25/lib64 -lrt"
+$ ../configure --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=mpi3 --enable-openmp CXX=mpiFCC CC=mpifcc CXXFLAGS="-std=c++11 -march=armv8-a+sve -msve-vector-bits=512 -fno-gcse -DA64FXFIXEDSIZE -DA64FXASM -DDSLASHINTRIN -DTOFU -I/opt/FJSVxtclanga/tcsds-1.2.25/include/mpi/fujitsu -lrt" LDFLAGS="-L/opt/FJSVxtclanga/tcsds-1.2.25/lib64 -lrt"
 
 --------------------------------------------------------
 
 * armclang 20.0 VLA (merlin)
 
-../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -fno-unroll-loops -mllvm -vectorizer-min-trip-count=2 -march=armv8-a+sve -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN" LDFLAGS=-static GRID_LDFLAGS=-static MPI_CXXLDFLAGS=-static
+../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -fno-unroll-loops -mllvm -vectorizer-min-trip-count=2 -march=armv8-a+sve -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN" LDFLAGS=-static GRID_LDFLAGS=-static MPI_CXXLDFLAGS=-static
 
 TODO check ARMCLANGCOMPAT
 
 
 * armclang 20.1 VLA (merlin)
 
-../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN" LDFLAGS=-static GRID_LDFLAGS=-static MPI_CXXLDFLAGS=-static
+../configure --with-lime=/home/men04359/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN" LDFLAGS=-static GRID_LDFLAGS=-static MPI_CXXLDFLAGS=-static
 
 TODO check ARMCLANGCOMPAT
 
 
 * armclang 20.1 VLA (fjt cluster)
 
-../configure --with-lime=$HOME/local --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU"
+../configure --with-lime=$HOME/local --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU"
 
 TODO check ARMCLANGCOMPAT
 
 
 * armclang 20.1 VLA w/MPI (fjt cluster)
 
-../configure --with-lime=$HOME/local --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=mpi3 --enable-openmp CXX=mpiFCC CC=mpifcc CXXFLAGS="-std=c++11 -mcpu=a64fx -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU -I/opt/FJSVxtclanga/tcsds-1.2.25/include/mpi/fujitsu -lrt" LDFLAGS="-L/opt/FJSVxtclanga/tcsds-1.2.25/lib64"
+../configure --with-lime=$HOME/local --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=mpi3 --enable-openmp CXX=mpiFCC CC=mpifcc CXXFLAGS="-std=c++11 -mcpu=a64fx -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU -I/opt/FJSVxtclanga/tcsds-1.2.25/include/mpi/fujitsu -lrt" LDFLAGS="-L/opt/FJSVxtclanga/tcsds-1.2.25/lib64"
 
 No ARMCLANGCOMPAT -> still correct ?
 
@@ -81,9 +81,9 @@ No ARMCLANGCOMPAT -> still correct ?
 
 * Fujitsu fcc
 
-../configure --with-lime=$HOME/grid-a64fx/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=none --enable-openmp --with-mpfr=/home/users/gre/gre-1/grid-a64fx/mpfr-build/install CXX=FCC CC=fcc CXXFLAGS="-Nclang -Kfast -DA64FX -DA64FXASM -DDSLASHINTRIN"
+../configure --with-lime=$HOME/grid-a64fx/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp --with-mpfr=/home/users/gre/gre-1/grid-a64fx/mpfr-build/install CXX=FCC CC=fcc CXXFLAGS="-Nclang -Kfast -DA64FX -DA64FXASM -DDSLASHINTRIN"
 
 
 * Fujitsu fcc w/ MPI
 
-../configure --with-lime=$HOME/grid-a64fx/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-comms=mpi --enable-openmp --with-mpfr=/home/users/gre/gre-1/grid-a64fx/mpfr-build/install CXX=mpiFCC CC=mpifcc CXXFLAGS="-Nclang -Kfast -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU"
+../configure --with-lime=$HOME/grid-a64fx/lime/c-lime --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=mpi --enable-openmp --with-mpfr=/home/users/gre/gre-1/grid-a64fx/mpfr-build/install CXX=mpiFCC CC=mpifcc CXXFLAGS="-Nclang -Kfast -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU"

benchmarks/Benchmark_IO.cc

@@ -1,18 +1,10 @@
+
 #include "Benchmark_IO.hpp"
 
-#ifndef BENCH_IO_LMIN
-#define BENCH_IO_LMIN 8
-#endif
-
 #ifndef BENCH_IO_LMAX
-#define BENCH_IO_LMAX 32
+#define BENCH_IO_LMAX 40
 #endif
 
-#ifndef BENCH_IO_NPASS
-#define BENCH_IO_NPASS 10
-#endif
-
-#ifdef HAVE_LIME
 using namespace Grid;
 
 std::string filestem(const int l)
@@ -20,182 +12,37 @@ std::string filestem(const int l)
   return "iobench_l" + std::to_string(l);
 }
 
-int vol(const int i)
-{
-  return BENCH_IO_LMIN + 2*i;
-}
-
-int volInd(const int l)
-{
-  return (l - BENCH_IO_LMIN)/2;
-}
-
-template <typename Mat>
-void stats(Mat &mean, Mat &stdDev, const std::vector<Mat> &data)
-{
-  auto            nr = data[0].rows(), nc = data[0].cols();
-  Eigen::MatrixXd sqSum(nr, nc);
-  double          n = static_cast<double>(data.size());
-
-  assert(n > 1.);
-  mean  = Mat::Zero(nr, nc);
-  sqSum = Mat::Zero(nr, nc);
-  for (auto &d: data)
-  {
-    mean  += d;
-    sqSum += d.cwiseProduct(d);
-  }
-  stdDev = ((sqSum - mean.cwiseProduct(mean)/n)/(n - 1.)).cwiseSqrt();
-  mean  /= n;
-}
-
-#define grid_printf(...) \
-{\
-  char _buf[1024];\
-  sprintf(_buf, __VA_ARGS__);\
-  MSG << _buf;\
-}
-
-enum {sRead = 0, sWrite = 1, gRead = 2, gWrite = 3};
-
 int main (int argc, char ** argv)
 {
+#ifdef HAVE_LIME
   Grid_init(&argc,&argv);
 
-  int64_t                      threads = GridThread::GetThreads();
-  auto                         mpi     = GridDefaultMpi();
-  unsigned int                 nVol    = (BENCH_IO_LMAX - BENCH_IO_LMIN)/2 + 1;
-  unsigned int                 nRelVol = (BENCH_IO_LMAX - 24)/2 + 1;
-  std::vector<Eigen::MatrixXd> perf(BENCH_IO_NPASS, Eigen::MatrixXd::Zero(nVol, 4));
-  std::vector<Eigen::VectorXd> avPerf(BENCH_IO_NPASS, Eigen::VectorXd::Zero(4));
-  std::vector<int>             latt;
-
+  int64_t threads = GridThread::GetThreads();
   MSG << "Grid is setup to use " << threads << " threads" << std::endl;
-  MSG << "MPI partition " << mpi << std::endl;
-  for (unsigned int i = 0; i < BENCH_IO_NPASS; ++i)
+  MSG << SEP << std::endl;
+  MSG << "Benchmark Lime write" << std::endl;
+  MSG << SEP << std::endl;
+  for (int l = 4; l <= BENCH_IO_LMAX; l += 2)
   {
-    MSG << BIGSEP << std::endl;
-    MSG << "Pass " << i + 1 << "/" << BENCH_IO_NPASS << std::endl;
-    MSG << BIGSEP << std::endl;
-    MSG << SEP << std::endl;
-    MSG << "Benchmark std write" << std::endl;
-    MSG << SEP << std::endl;
-    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
-    {
-      latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
+    auto             mpi  = GridDefaultMpi();
+    std::vector<int> latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
 
-      MSG << "-- Local volume " << l << "^4" << std::endl;
-      writeBenchmark<LatticeFermion>(latt, filestem(l), stdWrite<LatticeFermion>);
-      perf[i](volInd(l), sWrite) = BinaryIO::lastPerf.mbytesPerSecond;
-    }
-
-    MSG << SEP << std::endl;
-    MSG << "Benchmark std read" << std::endl;
-    MSG << SEP << std::endl;
-    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
-    {
-      latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
-
-      MSG << "-- Local volume " << l << "^4" << std::endl;
-      readBenchmark<LatticeFermion>(latt, filestem(l), stdRead<LatticeFermion>);
-      perf[i](volInd(l), sRead) = BinaryIO::lastPerf.mbytesPerSecond;
-    }
-
-  #ifdef HAVE_LIME
-    MSG << SEP << std::endl;
-    MSG << "Benchmark Grid C-Lime write" << std::endl;
-    MSG << SEP << std::endl;
-    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
-    {
-      latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
-
-      MSG << "-- Local volume " << l << "^4" << std::endl;
-      writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>);
-      perf[i](volInd(l), gWrite) = BinaryIO::lastPerf.mbytesPerSecond;
-    }
-
-    MSG << SEP << std::endl;
-    MSG << "Benchmark Grid C-Lime read" << std::endl;
-    MSG << SEP << std::endl;
-    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
-    {
-      latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
-
-      MSG << "-- Local volume " << l << "^4" << std::endl;
-      readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>);
-      perf[i](volInd(l), gRead) = BinaryIO::lastPerf.mbytesPerSecond;
-    }
-#endif
-    avPerf[i].fill(0.);
-    for (int f = 0; f < 4; ++f)
-    for (int l = 24; l <= BENCH_IO_LMAX; l += 2)
-    {
-      avPerf[i](f) += perf[i](volInd(l), f);
-    }
-    avPerf[i] /= nRelVol;
+    std::cout << "-- Local volume " << l << "^4" << std::endl;
+    writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>);
   }
 
-  Eigen::MatrixXd mean(nVol, 4), stdDev(nVol, 4), rob(nVol, 4);
-  Eigen::VectorXd avMean(4), avStdDev(4), avRob(4);
-  double          n = BENCH_IO_NPASS;
-
-  stats(mean, stdDev, perf);
-  stats(avMean, avStdDev, avPerf);
-  rob.fill(100.);
-  rob -= 100.*stdDev.cwiseQuotient(mean.cwiseAbs());
-  avRob.fill(100.);
-  avRob -= 100.*avStdDev.cwiseQuotient(avMean.cwiseAbs());
-
-  MSG << BIGSEP << std::endl;
-  MSG << "SUMMARY" << std::endl;
-  MSG << BIGSEP << std::endl;
-  MSG << "Summary of individual results (all results in MB/s)." << std::endl;
-  MSG << "Every second colum gives the standard deviation of the previous column." << std::endl;
-  MSG << std::endl;
-  grid_printf("%4s %12s %12s %12s %12s %12s %12s %12s %12s\n",
-              "L", "std read", "std dev", "std write", "std dev",
-              "Grid read", "std dev", "Grid write", "std dev");
-  for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
+  MSG << "Benchmark Lime read" << std::endl;
+  MSG << SEP << std::endl;
+  for (int l = 4; l <= BENCH_IO_LMAX; l += 2)
   {
-    grid_printf("%4d %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n",
-                l, mean(volInd(l), sRead), stdDev(volInd(l), sRead),
-                mean(volInd(l), sWrite), stdDev(volInd(l), sWrite),
-                mean(volInd(l), gRead), stdDev(volInd(l), gRead),
-                mean(volInd(l), gWrite), stdDev(volInd(l), gWrite));
+    auto             mpi  = GridDefaultMpi();
+    std::vector<int> latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
+
+    std::cout << "-- Local volume " << l << "^4" << std::endl;
+    readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>);
   }
-  MSG << std::endl;
-  MSG << "Robustness of individual results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
-  MSG << std::endl;
-  grid_printf("%4s %12s %12s %12s %12s\n",
-              "L", "std read", "std write", "Grid read", "Grid write");
-  for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
-  {
-    grid_printf("%4d %12.1f %12.1f %12.1f %12.1f\n",
-                l, rob(volInd(l), sRead), rob(volInd(l), sWrite),
-                rob(volInd(l), gRead), rob(volInd(l), gWrite));
-  }
-  MSG << std::endl;
-  MSG << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX << "^4 (all results in MB/s)." << std::endl;
-  MSG << "Every second colum gives the standard deviation of the previous column." << std::endl;
-  MSG << std::endl;
-  grid_printf("%12s %12s %12s %12s %12s %12s %12s %12s\n",
-              "std read", "std dev", "std write", "std dev",
-              "Grid read", "std dev", "Grid write", "std dev");
-  grid_printf("%12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n",
-              avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite),
-              avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
-  MSG << std::endl;
-  MSG << "Robustness of volume-averaged results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
-  MSG << std::endl;
-  grid_printf("%12s %12s %12s %12s\n",
-              "std read", "std write", "Grid read", "Grid write");
-  grid_printf("%12.1f %12.1f %12.1f %12.1f\n",
-              avRob(sRead), avRob(sWrite), avRob(gRead), avRob(gWrite));
 
   Grid_finalize();
-
+#endif
   return EXIT_SUCCESS;
 }
-#else
-int main(int argc,char ** argv){}
-#endif

benchmarks/Benchmark_IO.hpp

@@ -2,12 +2,10 @@
 #define Benchmark_IO_hpp_
 
 #include <Grid/Grid.h>
+#ifdef HAVE_LIME
 #define MSG std::cout << GridLogMessage
 #define SEP \
-"-----------------------------------------------------------------------------"
-#define BIGSEP \
 "============================================================================="
-#ifdef HAVE_LIME
 
 namespace Grid {
 
@@ -16,152 +14,13 @@ using WriterFn = std::function<void(const std::string, Field &)> ;
 template <typename Field>
 using ReaderFn = std::function<void(Field &, const std::string)>;
 
-// AP 06/10/2020: Standard C version in case one is suspicious of the C++ API
-// 
-// template <typename Field>
-// void stdWrite(const std::string filestem, Field &vec)
-// {
-//   std::string   rankStr = std::to_string(vec.Grid()->ThisRank());
-//   std::FILE     *file = std::fopen((filestem + "." + rankStr + ".bin").c_str(), "wb");
-//   size_t        size;
-//   uint32_t      crc;
-//   GridStopWatch ioWatch, crcWatch;
-
-//   size = vec.Grid()->lSites()*sizeof(typename Field::scalar_object);
-//   autoView(vec_v, vec, CpuRead);
-//   crcWatch.Start();
-//   crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
-//   std::fwrite(&crc, sizeof(uint32_t), 1, file);
-//   crcWatch.Stop();
-//   MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
-//   ioWatch.Start();
-//   std::fwrite(vec_v.cpu_ptr, sizeof(typename Field::scalar_object), vec.Grid()->lSites(), file);
-//   ioWatch.Stop();
-//   std::fclose(file);
-//   size *= vec.Grid()->ProcessorCount();
-//   auto &p = BinaryIO::lastPerf;
-//   p.size            = size;
-//   p.time            = ioWatch.useconds();
-//   p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
-//   MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed() 
-//       << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
-//   MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
-// }
-//
-// template <typename Field>
-// void stdRead(Field &vec, const std::string filestem)
-// {
-//   std::string   rankStr = std::to_string(vec.Grid()->ThisRank());
-//   std::FILE     *file = std::fopen((filestem + "." + rankStr + ".bin").c_str(), "rb");
-//   size_t        size;
-//   uint32_t      crcRead, crcData;
-//   GridStopWatch ioWatch, crcWatch;
-
-//   size = vec.Grid()->lSites()*sizeof(typename Field::scalar_object);
-//   crcWatch.Start();
-//   std::fread(&crcRead, sizeof(uint32_t), 1, file);
-//   crcWatch.Stop();
-//   {
-//     autoView(vec_v, vec, CpuWrite);
-//     ioWatch.Start();
-//     std::fread(vec_v.cpu_ptr, sizeof(typename Field::scalar_object), vec.Grid()->lSites(), file);
-//     ioWatch.Stop();
-//     std::fclose(file);
-//   }
-//   {
-//     autoView(vec_v, vec, CpuRead);
-//     crcWatch.Start();
-//     crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
-//     crcWatch.Stop();
-//   }
-//   MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
-//   assert(crcData == crcRead);
-//   size *= vec.Grid()->ProcessorCount();
-//   auto &p = BinaryIO::lastPerf;
-//   p.size            = size;
-//   p.time            = ioWatch.useconds();
-//   p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
-//   MSG << "Std I/O read: Read " <<  p.size << " bytes in " << ioWatch.Elapsed() 
-//       << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
-//   MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
-// }
-
-template <typename Field>
-void stdWrite(const std::string filestem, Field &vec)
-{
-  std::string   rankStr = std::to_string(vec.Grid()->ThisRank());
-  std::ofstream file(filestem + "." + rankStr + ".bin", std::ios::out | std::ios::binary);
-  size_t        size, sizec;
-  uint32_t      crc;
-  GridStopWatch ioWatch, crcWatch;
-
-  size  = vec.Grid()->lSites()*sizeof(typename Field::scalar_object);
-  sizec = size/sizeof(char); // just in case of...
-  autoView(vec_v, vec, CpuRead);
-  crcWatch.Start();
-  crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
-  file.write(reinterpret_cast<char *>(&crc), sizeof(uint32_t)/sizeof(char));
-  crcWatch.Stop();
-  MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
-  ioWatch.Start();
-  file.write(reinterpret_cast<char *>(vec_v.cpu_ptr), sizec);
-  file.flush();
-  ioWatch.Stop();
-  size *= vec.Grid()->ProcessorCount();
-  auto &p = BinaryIO::lastPerf;
-  p.size            = size;
-  p.time            = ioWatch.useconds();
-  p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
-  MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed() 
-      << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
-  MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
-}
-
-template <typename Field>
-void stdRead(Field &vec, const std::string filestem)
-{
-  std::string   rankStr = std::to_string(vec.Grid()->ThisRank());
-  std::ifstream file(filestem + "." + rankStr + ".bin", std::ios::in | std::ios::binary);
-  size_t        size, sizec;
-  uint32_t      crcRead, crcData;
-  GridStopWatch ioWatch, crcWatch;
-
-  size  = vec.Grid()->lSites()*sizeof(typename Field::scalar_object);
-  sizec = size/sizeof(char); // just in case of...
-  crcWatch.Start();
-  file.read(reinterpret_cast<char *>(&crcRead), sizeof(uint32_t)/sizeof(char));
-  crcWatch.Stop();
-  {
-    autoView(vec_v, vec, CpuWrite);
-    ioWatch.Start();
-    file.read(reinterpret_cast<char *>(vec_v.cpu_ptr), sizec);
-    ioWatch.Stop();
-  }
-  {
-    autoView(vec_v, vec, CpuRead);
-    crcWatch.Start();
-    crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
-    crcWatch.Stop();
-  }
-  MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
-  assert(crcData == crcRead);
-  size *= vec.Grid()->ProcessorCount();
-  auto &p = BinaryIO::lastPerf;
-  p.size            = size;
-  p.time            = ioWatch.useconds();
-  p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
-  MSG << "Std I/O read: Read " <<  p.size << " bytes in " << ioWatch.Elapsed() 
-      << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
-  MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
-}
-
 template <typename Field>
 void limeWrite(const std::string filestem, Field &vec)
 {
   emptyUserRecord   record;
   ScidacWriter binWriter(vec.Grid()->IsBoss());
 
-  binWriter.open(filestem + ".lime.bin");
+  binWriter.open(filestem + ".bin");
   binWriter.writeScidacFieldRecord(vec, record);
   binWriter.close();
 }
@@ -172,7 +31,7 @@ void limeRead(Field &vec, const std::string filestem)
   emptyUserRecord   record;
   ScidacReader binReader;
 
-  binReader.open(filestem + ".lime.bin");
+  binReader.open(filestem + ".bin");
   binReader.readScidacFieldRecord(vec, record);
   binReader.close();
 }
@@ -214,18 +73,12 @@ void writeBenchmark(const Coordinate &latt, const std::string filename,
   auto                           simd = GridDefaultSimd(latt.size(), Field::vector_type::Nsimd());
   std::shared_ptr<GridCartesian> gBasePt(SpaceTimeGrid::makeFourDimGrid(latt, simd, mpi));
   std::shared_ptr<GridBase>      gPt;
-  std::random_device             rd;
 
   makeGrid(gPt, gBasePt, Ls, rb);
 
-  GridBase         *g = gPt.get();
-  GridParallelRNG  rng(g);
-  Field            vec(g);
-
-  rng.SeedFixedIntegers({static_cast<int>(rd()), static_cast<int>(rd()),
-                         static_cast<int>(rd()), static_cast<int>(rd()),
-                         static_cast<int>(rd()), static_cast<int>(rd()),
-                         static_cast<int>(rd()), static_cast<int>(rd())});
+  GridBase                       *g = gPt.get();
+  GridParallelRNG                rng(g);
+  Field                          vec(g);
 
   random(rng, vec);
   write(filename, vec);
@@ -243,8 +96,8 @@ void readBenchmark(const Coordinate &latt, const std::string filename,
 
   makeGrid(gPt, gBasePt, Ls, rb);
 
-  GridBase *g = gPt.get();
-  Field    vec(g);
+  GridBase                       *g = gPt.get();
+  Field                          vec(g);
 
   read(vec, filename);
 }

benchmarks/Benchmark_IO_vs_dir.cc

@@ -1,9 +1,14 @@
 #include "Benchmark_IO.hpp"
-#ifdef HAVE_LIME
+
+#define MSG std::cout << GridLogMessage
+#define SEP \
+"============================================================================="
+
 using namespace Grid;
 
 int main (int argc, char ** argv)
 {
+#ifdef HAVE_LIME
   std::vector<std::string> dir;
   unsigned int             Ls;
   bool                     rb;
@@ -29,74 +34,46 @@ int main (int argc, char ** argv)
   }
   Grid_init(&argc,&argv);
 
+
   int64_t threads = GridThread::GetThreads();
-  auto    mpi     = GridDefaultMpi();
-
   MSG << "Grid is setup to use " << threads << " threads" << std::endl;
-  MSG << "MPI partition " << mpi << std::endl;
-
   MSG << SEP << std::endl;
-  MSG << "Benchmark Grid std write" << std::endl;
+  MSG << "Benchmark double precision Lime write" << std::endl;
   MSG << SEP << std::endl;
   for (auto &d: dir)
   {
     MSG << "-- Directory " << d << std::endl;
-    writeBenchmark<LatticeFermion>(GridDefaultLatt(), d + "/ioBench", 
-                                   stdWrite<LatticeFermion>, Ls, rb);
+    writeBenchmark<LatticeFermion>(GridDefaultLatt(), d + "/ioBench", limeWrite<LatticeFermion>, Ls, rb);
   }
+
   MSG << SEP << std::endl;
-  MSG << "Benchmark Grid std read" << std::endl;
+  MSG << "Benchmark double precision Lime read" << std::endl;
   MSG << SEP << std::endl;
   for (auto &d: dir)
   {
     MSG << "-- Directory " << d << std::endl;
-    readBenchmark<LatticeFermion>(GridDefaultLatt(), d + "/ioBench", 
-                                  stdRead<LatticeFermion>, Ls, rb);
+    readBenchmark<LatticeFermion>(GridDefaultLatt(), d + "/ioBench", limeRead<LatticeFermion>, Ls, rb);
   }
 
-#ifdef HAVE_LIME
   MSG << SEP << std::endl;
-  MSG << "Benchmark Grid C-Lime write" << std::endl;
+  MSG << "Benchmark single precision Lime write" << std::endl;
   MSG << SEP << std::endl;
   for (auto &d: dir)
   {
     MSG << "-- Directory " << d << std::endl;
-    writeBenchmark<LatticeFermion>(GridDefaultLatt(), d + "/ioBench", 
-                                   limeWrite<LatticeFermion>, Ls, rb);
+    writeBenchmark<LatticeFermionF>(GridDefaultLatt(), d + "/ioBench", limeWrite<LatticeFermionF>, Ls, rb);
   }
+
   MSG << SEP << std::endl;
-  MSG << "Benchmark Grid C-Lime read" << std::endl;
+  MSG << "Benchmark single precision Lime read" << std::endl;
   MSG << SEP << std::endl;
   for (auto &d: dir)
   {
     MSG << "-- Directory " << d << std::endl;
-    readBenchmark<LatticeFermion>(GridDefaultLatt(), d + "/ioBench", 
-                                  limeRead<LatticeFermion>, Ls, rb);
+    readBenchmark<LatticeFermionF>(GridDefaultLatt(), d + "/ioBench", limeRead<LatticeFermionF>, Ls, rb);
   }
-#endif
-
-  // MSG << SEP << std::endl;
-  // MSG << "Benchmark single precision Lime write" << std::endl;
-  // MSG << SEP << std::endl;
-  // for (auto &d: dir)
-  // {
-  //   MSG << "-- Directory " << d << std::endl;
-  //   writeBenchmark<LatticeFermionF>(GridDefaultLatt(), d + "/ioBench", limeWrite<LatticeFermionF>, Ls, rb);
-  // }
-
-  // MSG << SEP << std::endl;
-  // MSG << "Benchmark single precision Lime read" << std::endl;
-  // MSG << SEP << std::endl;
-  // for (auto &d: dir)
-  // {
-  //   MSG << "-- Directory " << d << std::endl;
-  //   readBenchmark<LatticeFermionF>(GridDefaultLatt(), d + "/ioBench", limeRead<LatticeFermionF>, Ls, rb);
-  // }
 
   Grid_finalize();
-
+#endif
   return EXIT_SUCCESS;
 }
-#else
-int main(int argc,char ** argv){}
-#endif

benchmarks/Benchmark_ITT.cc

@@ -1,4 +1,4 @@
-/*************************************************************************************
+    /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -62,7 +62,7 @@ struct time_statistics{
 
 void comms_header(){
   std::cout <<GridLogMessage << " L  "<<"\t"<<" Ls  "<<"\t"
-            <<"bytes\t MB/s uni (err/min/max) \t\t MB/s bidi (err/min/max)"<<std::endl;
+            <<std::setw(11)<<"bytes"<<"MB/s uni (err/min/max)"<<"\t\t"<<"MB/s bidi (err/min/max)"<<std::endl;
 };
 
 Gamma::Algebra Gmu [] = {
@@ -125,7 +125,7 @@ public:
 	      lat*mpi_layout[1],
 	      lat*mpi_layout[2],
 	      lat*mpi_layout[3]});
-
+	std::cout << GridLogMessage<< latt_size <<std::endl;
 	GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 	RealD Nrank = Grid._Nprocessors;
 	RealD Nnode = Grid.NodeCount();
@@ -137,8 +137,8 @@ public:
 	for(int d=0;d<8;d++){
 	  xbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
 	  rbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
-	  //	  bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
-	  //	  bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	  bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	  bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
 	}
 
 	int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
@@ -189,11 +189,11 @@ public:
 	//	double rbytes    = dbytes*0.5;
 	double bidibytes = dbytes;
 
-	std::cout<<GridLogMessage << lat<<"\t"<<Ls<<"\t "
-		 << bytes << " \t "
-		 <<xbytes/timestat.mean<<" \t "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " \t "
+	std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+		 <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+		 <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
 		 <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
-		 << "\t\t"<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+		 << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
 		 << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
 	
 	    }
@@ -202,8 +202,6 @@ public:
     return;
   }
 
-
-  
   static void Memory(void)
   {
     const int Nvec=8;
@@ -224,7 +222,7 @@ public:
 
 
   uint64_t lmax=32;
-#define NLOOP (1000*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
+#define NLOOP (100*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
 
     GridSerialRNG          sRNG;      sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
     for(int lat=8;lat<=lmax;lat+=8){
@@ -249,6 +247,11 @@ public:
       double start=usecond();
       for(int i=0;i<Nloop;i++){
 	z=a*x-y;
+	autoView( x_v , x, CpuWrite);
+	autoView( y_v , y, CpuWrite);
+	autoView( z_v , z, CpuRead);
+        x_v[0]=z_v[0]; // force serial dependency to prevent optimise away
+        y_v[4]=z_v[4];
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
@@ -263,61 +266,6 @@ public:
   };
 
 
-  static void SU4(void)
-  {
-    const int Nc4=4;
-    typedef Lattice< iMatrix< vComplexF,Nc4> > LatticeSU4;
-
-    Coordinate simd_layout = GridDefaultSimd(Nd,vComplexF::Nsimd());
-    Coordinate mpi_layout  = GridDefaultMpi();
-    
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << "= Benchmarking z = y*x SU(4) bandwidth"<<std::endl;
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<< "\t\tGB/s / node"<<std::endl;
-    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
-  
-    uint64_t NN;
-
-
-    uint64_t lmax=32;
-#define NLOOP (1000*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
-
-    GridSerialRNG          sRNG;      sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
-    for(int lat=8;lat<=lmax;lat+=8){
-
-      Coordinate latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
-
-      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-
-      NN =Grid.NodeCount();
-
-
-      LatticeSU4 z(&Grid); z=Zero();
-      LatticeSU4 x(&Grid); x=Zero();
-      LatticeSU4 y(&Grid); y=Zero();
-      double a=2.0;
-
-      uint64_t Nloop=NLOOP;
-
-      double start=usecond();
-      for(int i=0;i<Nloop;i++){
-	z=x*y;
-      }
-      double stop=usecond();
-      double time = (stop-start)/Nloop*1000;
-     
-      double flops=vol*Nc4*Nc4*(6+(Nc4-1)*8);// mul,add
-      double bytes=3.0*vol*Nc4*Nc4*2*sizeof(RealF);
-      std::cout<<GridLogMessage<<std::setprecision(3) 
-	       << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.
-	       << "\t\t"<< bytes/time/NN <<std::endl;
-
-    }
-  };
-
-
   static double DWF(int Ls,int L)
   {
     RealD mass=0.1;
@@ -334,9 +282,8 @@ public:
     int threads = GridThread::GetThreads();
     Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
     Coordinate local({L,L,L,L});
-    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
 
-    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, 
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(Coordinate({72,72,72,72}), 
 								       GridDefaultSimd(Nd,vComplex::Nsimd()),
 								       GridDefaultMpi());
     uint64_t NP = TmpGrid->RankCount();
@@ -344,11 +291,11 @@ public:
     NN_global=NN;
     uint64_t SHM=NP/NN;
 
+    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
 
     ///////// Welcome message ////////////
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
     std::cout<<GridLogMessage << "Benchmark DWF on "<<L<<"^4 local volume "<<std::endl;
-    std::cout<<GridLogMessage << "* Nc             : "<<Nc<<std::endl;
     std::cout<<GridLogMessage << "* Global volume  : "<<GridCmdVectorIntToString(latt4)<<std::endl;
     std::cout<<GridLogMessage << "* Ls             : "<<Ls<<std::endl;
     std::cout<<GridLogMessage << "* ranks          : "<<NP  <<std::endl;
@@ -377,7 +324,7 @@ public:
     typedef LatticeGaugeFieldF Gauge;
     
     ///////// Source preparation ////////////
-    Gauge Umu(UGrid);  SU<Nc>::HotConfiguration(RNG4,Umu); 
+    Gauge Umu(UGrid);  SU3::HotConfiguration(RNG4,Umu); 
     Fermion src   (FGrid); random(RNG5,src);
     Fermion src_e (FrbGrid);
     Fermion src_o (FrbGrid);
@@ -422,7 +369,7 @@ public:
 	}
 	FGrid->Barrier();
 	double t1=usecond();
-	uint64_t ncall = 500;
+	uint64_t ncall = 50;
 
 	FGrid->Broadcast(0,&ncall,sizeof(ncall));
 
@@ -440,17 +387,7 @@ public:
 	FGrid->Barrier();
 	
 	double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
-
-	// Nc=3 gives
-	// 1344= 3*(2*8+6)*2*8 + 8*3*2*2 + 3*4*2*8
-	// 1344 = Nc* (6+(Nc-1)*8)*2*Nd + Nd*Nc*2*2  + Nd*Nc*Ns*2
-	//	double flops=(1344.0*volume)/2;
-#if 1
-	double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + Nd*Nc*Ns  + Nd*Nc*Ns*2;
-#else
-	double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + 2*Nd*Nc*Ns  + 2*Nd*Nc*Ns*2;
-#endif
-	double flops=(fps*volume)/2;
+	double flops=(1344.0*volume)/2;
 	double mf_hi, mf_lo, mf_err;
 
 	timestat.statistics(t_time);
@@ -465,7 +402,6 @@ public:
 	if ( mflops>mflops_best ) mflops_best = mflops;
 	if ( mflops<mflops_worst) mflops_worst= mflops;
 
-	std::cout<<GridLogMessage<< "Deo FlopsPerSite is "<<fps<<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank   "<< mflops/NP<<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node   "<< mflops/NN<<std::endl;
@@ -502,9 +438,8 @@ public:
     int threads = GridThread::GetThreads();
     Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
     Coordinate local({L,L,L,L});
-    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
     
-    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(Coordinate({72,72,72,72}), 
 								       GridDefaultSimd(Nd,vComplex::Nsimd()),
 								       GridDefaultMpi());
     uint64_t NP = TmpGrid->RankCount();
@@ -512,6 +447,7 @@ public:
     NN_global=NN;
     uint64_t SHM=NP/NN;
 
+    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
 
     ///////// Welcome message ////////////
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
@@ -542,7 +478,7 @@ public:
     typedef typename Action::FermionField Fermion; 
     typedef LatticeGaugeFieldF Gauge;
     
-    Gauge Umu(FGrid);  SU<Nc>::HotConfiguration(RNG4,Umu); 
+    Gauge Umu(FGrid);  SU3::HotConfiguration(RNG4,Umu); 
 
     typename Action::ImplParams params;
     Action Ds(Umu,Umu,*FGrid,*FrbGrid,mass,c1,c2,u0,params);
@@ -660,12 +596,11 @@ int main (int argc, char ** argv)
 #endif
   Benchmark::Decomposition();
 
-  int do_su4=1;
   int do_memory=1;
   int do_comms =1;
 
-  int sel=4;
-  std::vector<int> L_list({8,12,16,24,32});
+  int sel=2;
+  std::vector<int> L_list({16,24,32});
   int selm1=sel-1;
 
   std::vector<double> wilson;
@@ -689,6 +624,7 @@ int main (int argc, char ** argv)
     dwf4.push_back(result);
   }
 
+  /*
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
   std::cout<<GridLogMessage << " Improved Staggered dslash 4D vectorised" <<std::endl;
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
@@ -696,13 +632,14 @@ int main (int argc, char ** argv)
     double result = Benchmark::Staggered(L_list[l]) ;
     staggered.push_back(result);
   }
+  */
 
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
   std::cout<<GridLogMessage << " Summary table Ls="<<Ls <<std::endl;
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "L \t\t Wilson \t\t DWF4 \t\t Staggered" <<std::endl;
+  std::cout<<GridLogMessage << "L \t\t Wilson \t\t DWF4 \t\tt Staggered" <<std::endl;
   for(int l=0;l<L_list.size();l++){
-    std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t\t "<<dwf4[l] << " \t\t "<< staggered[l]<<std::endl;
+    std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t\t "<<dwf4[l] <<std::endl;
   }
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
 
@@ -714,13 +651,6 @@ int main (int argc, char ** argv)
     Benchmark::Memory();
   }
 
-  if ( do_su4 ) {
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << " Memory benchmark " <<std::endl;
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    Benchmark::SU4();
-  }
-  
   if ( do_comms && (NN>1) ) {
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
     std::cout<<GridLogMessage << " Communications benchmark " <<std::endl;
@@ -731,9 +661,9 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
     std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << " L \t\t Wilson\t\t DWF4\t\t Staggered " <<std::endl;
+    std::cout<<GridLogMessage << " L \t\t Wilson\t\t DWF4  " <<std::endl;
     for(int l=0;l<L_list.size();l++){
-      std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]/NN<<" \t "<<dwf4[l]/NN<< " \t "<<staggered[l]/NN<<std::endl;
+      std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]/NN<<" \t "<<dwf4[l]/NN<<std::endl;
     }
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
 

benchmarks/Benchmark_comms.cc

@@ -94,8 +94,8 @@ int main (int argc, char ** argv)
       RealD Nnode = Grid.NodeCount();
       RealD ppn = Nrank/Nnode;
 
-      std::vector<std::vector<HalfSpinColourVectorD> > xbuf(8);
-      std::vector<std::vector<HalfSpinColourVectorD> > rbuf(8);
+      std::vector<Vector<HalfSpinColourVectorD> > xbuf(8);
+      std::vector<Vector<HalfSpinColourVectorD> > rbuf(8);
 
       for(int mu=0;mu<8;mu++){
 	xbuf[mu].resize(lat*lat*lat*Ls);

benchmarks/Benchmark_comms_host_device.cc

@@ -1,260 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./benchmarks/Benchmark_comms.cc
-
-    Copyright (C) 2015
-
-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 */
-#include <Grid/Grid.h>
-
-using namespace std;
-using namespace Grid;
-
-struct time_statistics{
-  double mean;
-  double err;
-  double min;
-  double max;
-
-  void statistics(std::vector<double> v){
-      double sum = std::accumulate(v.begin(), v.end(), 0.0);
-      mean = sum / v.size();
-
-      std::vector<double> diff(v.size());
-      std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
-      double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
-      err = std::sqrt(sq_sum / (v.size()*(v.size() - 1)));
-
-      auto result = std::minmax_element(v.begin(), v.end());
-      min = *result.first;
-      max = *result.second;
-}
-};
-
-void header(){
-  std::cout <<GridLogMessage << " L  "<<"\t"<<" Ls  "<<"\t"
-            <<std::setw(11)<<"bytes\t\t"<<"MB/s uni (err/min/max)"<<"\t\t"<<"MB/s bidi (err/min/max)"<<std::endl;
-};
-
-int main (int argc, char ** argv)
-{
-  Grid_init(&argc,&argv);
-
-  Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
-  Coordinate mpi_layout  = GridDefaultMpi();
-  int threads = GridThread::GetThreads();
-  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
-
-  int Nloop=250;
-  int nmu=0;
-  int maxlat=32;
-  for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
-
-  std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
-  std::vector<double> t_time(Nloop);
-  time_statistics timestat;
-
-  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from host memory "<<std::endl;
-  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  header();
-
-  for(int lat=8;lat<=maxlat;lat+=4){
-    for(int Ls=8;Ls<=8;Ls*=2){
-
-      Coordinate latt_size  ({lat*mpi_layout[0],
-	                      lat*mpi_layout[1],
-      			      lat*mpi_layout[2],
-      			      lat*mpi_layout[3]});
-
-      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      RealD Nrank = Grid._Nprocessors;
-      RealD Nnode = Grid.NodeCount();
-      RealD ppn = Nrank/Nnode;
-
-      std::vector<std::vector<HalfSpinColourVectorD> > xbuf(8);
-      std::vector<std::vector<HalfSpinColourVectorD> > rbuf(8);
-
-      for(int mu=0;mu<8;mu++){
-	xbuf[mu].resize(lat*lat*lat*Ls);
-	rbuf[mu].resize(lat*lat*lat*Ls);
-      }
-      uint64_t bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
-
-      int ncomm;
-
-      for(int mu=0;mu<4;mu++){
-	if (mpi_layout[mu]>1 ) {
-	double start=usecond();
-	for(int i=0;i<Nloop;i++){
-
-	  ncomm=0;
-	
-	  
-	    ncomm++;
-	    int comm_proc=1;
-	    int xmit_to_rank;
-	    int recv_from_rank;
-	    
-	    {
-	      std::vector<CommsRequest_t> requests;
-	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	      Grid.SendToRecvFrom((void *)&xbuf[mu][0],
-				  xmit_to_rank,
-				  (void *)&rbuf[mu][0],
-				  recv_from_rank,
-				  bytes);
-	    }
-
-	    comm_proc = mpi_layout[mu]-1;
-	    {
-	      std::vector<CommsRequest_t> requests;
-	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	      Grid.SendToRecvFrom((void *)&xbuf[mu+4][0],
-				  xmit_to_rank,
-				  (void *)&rbuf[mu+4][0],
-				  recv_from_rank,
-				  bytes);
-	    }
-	}
-	Grid.Barrier();
-	double stop=usecond();
-        double mean=(stop-start)/Nloop;      
-      double dbytes    = bytes*ppn;
-      double xbytes    = dbytes*2.0*ncomm;
-      double rbytes    = xbytes;
-      double bidibytes = xbytes+rbytes;
-
-      std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
-               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)<<" "
-               <<std::right<< xbytes/mean<<"  "
-               << "\t\t"<<std::setw(7)<< bidibytes/mean<< std::endl;
-
-
-	
-	}
-      }
-
-
-      
-    }
-  }
-
-  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from GPU memory "<<std::endl;
-  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  header();
-
-  for(int lat=8;lat<=maxlat;lat+=4){
-    for(int Ls=8;Ls<=8;Ls*=2){
-
-      Coordinate latt_size  ({lat*mpi_layout[0],
-	                      lat*mpi_layout[1],
-      			      lat*mpi_layout[2],
-      			      lat*mpi_layout[3]});
-
-      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      RealD Nrank = Grid._Nprocessors;
-      RealD Nnode = Grid.NodeCount();
-      RealD ppn = Nrank/Nnode;
-
-
-      std::vector<HalfSpinColourVectorD *> xbuf(8);
-      std::vector<HalfSpinColourVectorD *> rbuf(8);
-
-      uint64_t bytes = lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
-      for(int d=0;d<8;d++){
-	xbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
-	rbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
-      }
-
-      int ncomm;
-
-      for(int mu=0;mu<4;mu++){
-	if (mpi_layout[mu]>1 ) {
-	double start=usecond();
-	for(int i=0;i<Nloop;i++){
-
-	  ncomm=0;
-	
-	  
-	    ncomm++;
-	    int comm_proc=1;
-	    int xmit_to_rank;
-	    int recv_from_rank;
-	    
-	    {
-	      std::vector<CommsRequest_t> requests;
-	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	      Grid.SendToRecvFrom((void *)&xbuf[mu][0],
-				  xmit_to_rank,
-				  (void *)&rbuf[mu][0],
-				  recv_from_rank,
-				  bytes);
-	    }
-
-	    comm_proc = mpi_layout[mu]-1;
-	    {
-	      std::vector<CommsRequest_t> requests;
-	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	      Grid.SendToRecvFrom((void *)&xbuf[mu+4][0],
-				  xmit_to_rank,
-				  (void *)&rbuf[mu+4][0],
-				  recv_from_rank,
-				  bytes);
-	    }
-	}
-	Grid.Barrier();
-	double stop=usecond();
-        double mean=(stop-start)/Nloop;      
-      double dbytes    = bytes*ppn;
-      double xbytes    = dbytes*2.0*ncomm;
-      double rbytes    = xbytes;
-      double bidibytes = xbytes+rbytes;
-
-      std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
-               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)<<" "
-               <<std::right<< xbytes/mean<<"  "
-               << "\t\t"<<std::setw(7)<< bidibytes/mean<< std::endl;
-
-
-	
-	}
-      }
-
-      for(int d=0;d<8;d++){
-	acceleratorFreeDevice(xbuf[d]);
-	acceleratorFreeDevice(rbuf[d]);
-      }
-
-      
-    }
-  }
-
-
-  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "= All done; Bye Bye"<<std::endl;
-  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-
-  Grid_finalize();
-}

benchmarks/Benchmark_dwf.cc

@@ -108,7 +108,7 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeField Umu(UGrid);
-  SU<Nc>::HotConfiguration(RNG4,Umu);
+  SU3::HotConfiguration(RNG4,Umu);
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
 #if 0
   Umu=1.0;

benchmarks/Benchmark_dwf_fp32.cc

@@ -1,364 +0,0 @@
- /*************************************************************************************
-    Grid physics library, www.github.com/paboyle/Grid
-    Source file: ./benchmarks/Benchmark_dwf.cc
-    Copyright (C) 2015
-
-    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/Grid.h>
-#ifdef GRID_CUDA
-#define CUDA_PROFILE
-#endif
-
-#ifdef CUDA_PROFILE
-#include <cuda_profiler_api.h>
-#endif
-
-using namespace std;
-using namespace Grid;
-
-template<class d>
-struct scal {
-  d internal;
-};
-
-  Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT
-  };
-
-
-int main (int argc, char ** argv)
-{
-  Grid_init(&argc,&argv);
-
-
-  int threads = GridThread::GetThreads();
-
-  Coordinate latt4 = GridDefaultLatt();
-  int Ls=8;
-  for(int i=0;i<argc;i++)
-    if(std::string(argv[i]) == "-Ls"){
-      std::stringstream ss(argv[i+1]); ss >> Ls;
-    }
-
-  GridLogLayout();
-
-  long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
-
-
-  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
-  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
-  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
-
-  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
-  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
-  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
-  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
-  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
-
-  std::vector<int> seeds4({1,2,3,4});
-  std::vector<int> seeds5({5,6,7,8});
-
-  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
-  GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
-  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
-  GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
-  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
-
-  LatticeFermionF src   (FGrid); random(RNG5,src);
-#if 0
-  src = Zero();
-  {
-    Coordinate origin({0,0,0,latt4[2]-1,0});
-    SpinColourVectorF tmp;
-    tmp=Zero();
-    tmp()(0)(0)=Complex(-2.0,0.0);
-    std::cout << " source site 0 " << tmp<<std::endl;
-    pokeSite(tmp,src,origin);
-  }
-#else
-  RealD N2 = 1.0/::sqrt(norm2(src));
-  src = src*N2;
-#endif
-
-
-  LatticeFermionF result(FGrid); result=Zero();
-  LatticeFermionF    ref(FGrid);    ref=Zero();
-  LatticeFermionF    tmp(FGrid);
-  LatticeFermionF    err(FGrid);
-
-  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
-  LatticeGaugeFieldF Umu(UGrid);
-  SU<Nc>::HotConfiguration(RNG4,Umu);
-  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
-#if 0
-  Umu=1.0;
-  for(int mu=0;mu<Nd;mu++){
-    LatticeColourMatrixF ttmp(UGrid);
-    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
-    //    if (mu !=2 ) ttmp = 0;
-    //    ttmp = ttmp* pow(10.0,mu);
-    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
-  }
-  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
-#endif
-
-  ////////////////////////////////////
-  // Naive wilson implementation
-  ////////////////////////////////////
-  // replicate across fifth dimension
-  LatticeGaugeFieldF Umu5d(FGrid);
-  std::vector<LatticeColourMatrixF> U(4,FGrid);
-  {
-    autoView( Umu5d_v, Umu5d, CpuWrite);
-    autoView( Umu_v  , Umu  , CpuRead);
-    for(int ss=0;ss<Umu.Grid()->oSites();ss++){
-      for(int s=0;s<Ls;s++){
-	Umu5d_v[Ls*ss+s] = Umu_v[ss];
-      }
-    }
-  }
-  for(int mu=0;mu<Nd;mu++){
-    U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
-  }
-  std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
-
-  if (1)
-  {
-    ref = Zero();
-    for(int mu=0;mu<Nd;mu++){
-
-      tmp = U[mu]*Cshift(src,mu+1,1);
-      ref=ref + tmp - Gamma(Gmu[mu])*tmp;
-
-      tmp =adj(U[mu])*src;
-      tmp =Cshift(tmp,mu+1,-1);
-      ref=ref + tmp + Gamma(Gmu[mu])*tmp;
-    }
-    ref = -0.5*ref;
-  }
-
-  RealD mass=0.1;
-  RealD M5  =1.8;
-
-  RealD NP = UGrid->_Nprocessors;
-  RealD NN = UGrid->NodeCount();
-
-  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
-  std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
-  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
-  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
-  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionR::Dhop                  "<<std::endl;
-  std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
-  std::cout << GridLogMessage<< "* VComplexF size is "<<sizeof(vComplexF)<< " B"<<std::endl;
-  if ( sizeof(RealF)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
-  if ( sizeof(RealF)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
-#ifdef GRID_OMP
-  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
-  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
-#endif
-  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
-  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
-  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
-  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
-
-  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  int ncall =1000;
-
-  if (1) {
-    FGrid->Barrier();
-    Dw.ZeroCounters();
-    Dw.Dhop(src,result,0);
-    std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
-    double t0=usecond();
-    for(int i=0;i<ncall;i++){
-      __SSC_START;
-      Dw.Dhop(src,result,0);
-      __SSC_STOP;
-    }
-    double t1=usecond();
-    FGrid->Barrier();
-
-    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
-    double flops=single_site_flops*volume*ncall;
-
-    auto nsimd = vComplex::Nsimd();
-    auto simdwidth = sizeof(vComplex);
-
-    // RF: Nd Wilson * Ls, Nd gauge * Ls, Nc colors
-    double data_rf = volume * ((2*Nd+1)*Nd*Nc + 2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
-
-    // mem: Nd Wilson * Ls, Nd gauge, Nc colors
-    double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
-
-    std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
-    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
-    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
-    err = ref-result;
-    std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-    //exit(0);
-
-    if(( norm2(err)>1.0e-4) ) {
-      /*
-      std::cout << "RESULT\n " << result<<std::endl;
-      std::cout << "REF   \n " << ref   <<std::endl;
-      std::cout << "ERR   \n " << err   <<std::endl;
-      */
-      std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
-      FGrid->Barrier();
-      exit(-1);
-    }
-    assert (norm2(err)< 1.0e-4 );
-    Dw.Report();
-  }
-
-  if (1)
-  { // Naive wilson dag implementation
-    ref = Zero();
-    for(int mu=0;mu<Nd;mu++){
-
-      //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
-      tmp = U[mu]*Cshift(src,mu+1,1);
-      {
-	autoView( ref_v, ref, CpuWrite);
-	autoView( tmp_v, tmp, CpuRead);
-	for(int i=0;i<ref_v.size();i++){
-	  ref_v[i]+= tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]; ;
-	}
-      }
-
-      tmp =adj(U[mu])*src;
-      tmp =Cshift(tmp,mu+1,-1);
-      {
-	autoView( ref_v, ref, CpuWrite);
-	autoView( tmp_v, tmp, CpuRead);
-	for(int i=0;i<ref_v.size();i++){
-	  ref_v[i]+= tmp_v[i] - Gamma(Gmu[mu])*tmp_v[i]; ;
-	}
-      }
-    }
-    ref = -0.5*ref;
-  }
-  //  dump=1;
-  Dw.Dhop(src,result,1);
-  std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
-  std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
-  std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
-  std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
-  err = ref-result;
-  std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-/*
-	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
-	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
-	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
-*/
-  }
-  LatticeFermionF src_e (FrbGrid);
-  LatticeFermionF src_o (FrbGrid);
-  LatticeFermionF r_e   (FrbGrid);
-  LatticeFermionF r_o   (FrbGrid);
-  LatticeFermionF r_eo  (FGrid);
-
-  std::cout<<GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"<<std::endl;
-  pickCheckerboard(Even,src_e,src);
-  pickCheckerboard(Odd,src_o,src);
-
-  std::cout<<GridLogMessage << "src_e"<<norm2(src_e)<<std::endl;
-  std::cout<<GridLogMessage << "src_o"<<norm2(src_o)<<std::endl;
-
-
-  // S-direction is INNERMOST and takes no part in the parity.
-  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
-  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionF::DhopEO                "<<std::endl;
-  std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
-  if ( sizeof(RealF)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
-  if ( sizeof(RealF)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
-#ifdef GRID_OMP
-  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
-  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
-#endif
-  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
-  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
-  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
-  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
-  {
-    Dw.ZeroCounters();
-    FGrid->Barrier();
-    Dw.DhopEO(src_o,r_e,DaggerNo);
-    double t0=usecond();
-    for(int i=0;i<ncall;i++){
-#ifdef CUDA_PROFILE
-      if(i==10) cudaProfilerStart();
-#endif
-      Dw.DhopEO(src_o,r_e,DaggerNo);
-#ifdef CUDA_PROFILE
-      if(i==20) cudaProfilerStop();
-#endif
-    }
-    double t1=usecond();
-    FGrid->Barrier();
-
-    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
-    double flops=(single_site_flops*volume*ncall)/2.0;
-
-    std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
-    std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
-    Dw.Report();
-  }
-  Dw.DhopEO(src_o,r_e,DaggerNo);
-  Dw.DhopOE(src_e,r_o,DaggerNo);
-  Dw.Dhop  (src  ,result,DaggerNo);
-
-  std::cout<<GridLogMessage << "r_e"<<norm2(r_e)<<std::endl;
-  std::cout<<GridLogMessage << "r_o"<<norm2(r_o)<<std::endl;
-  std::cout<<GridLogMessage << "res"<<norm2(result)<<std::endl;
-
-  setCheckerboard(r_eo,r_o);
-  setCheckerboard(r_eo,r_e);
-
-  err = r_eo-result;
-  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-    /*
-	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
-	std::cout<< "Deo REF\n " <<result  << std::endl;
-	std::cout<< "Deo ERR   \n " << err <<std::endl;
-    */
-  }
-
-  pickCheckerboard(Even,src_e,err);
-  pickCheckerboard(Odd,src_o,err);
-  std::cout<<GridLogMessage << "norm diff even  "<< norm2(src_e)<<std::endl;
-  std::cout<<GridLogMessage << "norm diff odd   "<< norm2(src_o)<<std::endl;
-
-  assert(norm2(src_e)<1.0e-4);
-  assert(norm2(src_o)<1.0e-4);
-  Grid_finalize();
-  exit(0);
-}

benchmarks/Benchmark_gparity.cc

@@ -24,7 +24,7 @@ typedef typename GparityDomainWallFermionD::FermionField GparityLatticeFermionD;
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
-#ifdef ENABLE_GPARITY
+
   int Ls=16;
   for(int i=0;i<argc;i++)
     if(std::string(argv[i]) == "-Ls"){
@@ -63,7 +63,7 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeFieldF Umu(UGrid); 
-  SU<Nc>::HotConfiguration(RNG4,Umu); 
+  SU3::HotConfiguration(RNG4,Umu); 
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
 
   RealD mass=0.1;
@@ -184,7 +184,7 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
     DwD.Report();
   }
-#endif
+
   Grid_finalize();
 }