diff --git a/Grid/GridCore.h b/Grid/GridCore.h
index 495a81e1..2209f960 100644
--- a/Grid/GridCore.h
+++ b/Grid/GridCore.h
@@ -47,10 +47,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
-#include <Grid/allocator/GridMemoryManager.h>
-#include <Grid/allocator/AlignedAllocator.h>
+#include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
-#include <Grid/threads/Threads.h>
+#include <Grid/threads/ThreadReduction.h>
 #include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 
diff --git a/Grid/GridStd.h b/Grid/GridStd.h
index 16cfcf50..ecb561ea 100644
--- a/Grid/GridStd.h
+++ b/Grid/GridStd.h
@@ -6,6 +6,7 @@
 ///////////////////
 #include <cassert>
 #include <complex>
+#include <memory>
 #include <vector>
 #include <array>
 #include <string>
diff --git a/Grid/Grid_Eigen_Dense.h b/Grid/Grid_Eigen_Dense.h
index 9d779e05..9556c03d 100644
--- a/Grid/Grid_Eigen_Dense.h
+++ b/Grid/Grid_Eigen_Dense.h
@@ -18,19 +18,20 @@
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")
+#undef __CUDA_ARCH__
 #undef __NVCC__
 #undef __CUDACC__
-#undef __CUDA_ARCH__
 #define __NVCC__REDEFINE__
 #endif 
 
 /* SYCL save and restore compile environment*/
-#ifdef __SYCL_DEVICE_ONLY__  
+#ifdef GRID_SYCL
 #pragma push
 #pragma push_macro("__SYCL_DEVICE_ONLY__")
 #undef __SYCL_DEVICE_ONLY__
-#undef EIGEN_USE_SYCL
 #define EIGEN_DONT_VECTORIZE
+//#undef EIGEN_USE_SYCL
+#define __SYCL__REDEFINE__
 #endif
 
 
@@ -41,7 +42,7 @@
 #ifdef __NVCC__REDEFINE__
 #pragma pop_macro("__CUDACC__")
 #pragma pop_macro("__NVCC__")
-#pragma pop_macro("__CUDA_ARCH__")
+#pragma pop_macro("GRID_SIMT")
 #pragma pop
 #endif
 
diff --git a/Grid/Makefile.am b/Grid/Makefile.am
index b88ea4f2..f1fa462e 100644
--- a/Grid/Makefile.am
+++ b/Grid/Makefile.am
@@ -21,7 +21,7 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
-all: version-cache
+all: version-cache Version.h
 
 version-cache:
 	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
@@ -42,7 +42,7 @@ version-cache:
 	fi;\
 	rm -f vertmp
 
-Version.h:
+Version.h: version-cache
 	cp version-cache Version.h
 
 .PHONY: version-cache
diff --git a/Grid/algorithms/Algorithms.h b/Grid/algorithms/Algorithms.h
index 48ea194b..7f27784b 100644
--- a/Grid/algorithms/Algorithms.h
+++ b/Grid/algorithms/Algorithms.h
@@ -29,9 +29,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_H
 #define GRID_ALGORITHMS_H
 
+NAMESPACE_CHECK(algorithms);
 #include <Grid/algorithms/SparseMatrix.h>
 #include <Grid/algorithms/LinearOperator.h>
 #include <Grid/algorithms/Preconditioner.h>
+NAMESPACE_CHECK(SparseMatrix);
 
 #include <Grid/algorithms/approx/Zolotarev.h>
 #include <Grid/algorithms/approx/Chebyshev.h>
@@ -41,10 +43,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/approx/Forecast.h>
 #include <Grid/algorithms/approx/RemezGeneral.h>
 #include <Grid/algorithms/approx/ZMobius.h>
-
+NAMESPACE_CHECK(approx);
 #include <Grid/algorithms/iterative/Deflation.h>
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
+NAMESPACE_CHECK(ConjGrad);
 #include <Grid/algorithms/iterative/BiCGSTAB.h>
+NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
@@ -62,7 +66,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/iterative/PowerMethod.h>
 
+NAMESPACE_CHECK(PowerMethod);
 #include <Grid/algorithms/CoarsenedMatrix.h>
+NAMESPACE_CHECK(CoarsendMatrix);
 #include <Grid/algorithms/FFT.h>
 
 #endif
diff --git a/Grid/algorithms/CoarsenedMatrix.h b/Grid/algorithms/CoarsenedMatrix.h
index 4c26f799..8d184aea 100644
--- a/Grid/algorithms/CoarsenedMatrix.h
+++ b/Grid/algorithms/CoarsenedMatrix.h
@@ -1,14 +1,3 @@
-    // blockZaxpy in bockPromote - 3s, 5%
-    // noncoalesced linalg in Preconditionoer ~ 3s 5%
-    // Lancos tuning or replace 10-20s ~ 25%, open ended
-    // setup tuning   5s  ~  8%
-    //    -- e.g. ordermin, orderstep tunables.
-    // MdagM path without norm in LinOp code.     few seconds
-
-    // Mdir calc blocking kernels
-    // Fuse kernels in blockMaskedInnerProduct
-    // preallocate Vectors in Cayley 5D ~ few percent few seconds
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -91,34 +80,7 @@ public:
     }
     directions   [2*_d]=0;
     displacements[2*_d]=0;
-      
-    //// report back
-    std::cout<<GridLogMessage<<"directions    :";
-    for(int d=0;d<npoint;d++) std::cout<< directions[d]<< " ";
-    std::cout<<std::endl;
-    std::cout<<GridLogMessage<<"displacements :";
-    for(int d=0;d<npoint;d++) std::cout<< displacements[d]<< " ";
-    std::cout<<std::endl;
   }
-  
-  /*
-  // Original cleaner code
-  Geometry(int _d) : dimension(_d), npoint(2*_d+1), directions(npoint), displacements(npoint) {
-  for(int d=0;d<dimension;d++){
-  directions[2*d  ] = d;
-  directions[2*d+1] = d;
-  displacements[2*d  ] = +1;
-  displacements[2*d+1] = -1;
-  }
-  directions   [2*dimension]=0;
-  displacements[2*dimension]=0;
-  }
-  std::vector<int> GetDelta(int point) {
-  std::vector<int> delta(dimension,0);
-  delta[directions[point]] = displacements[point];
-  return delta;
-  };
-  */    
 
 };
   
@@ -149,25 +111,7 @@ public:
     CoarseScalar InnerProd(CoarseGrid); 
     std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
     blockOrthogonalise(InnerProd,subspace);
-    //    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 2"<<std::endl; // Really have to do twice? Yuck
-    //    blockOrthogonalise(InnerProd,subspace);
-    //      std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
-    //      CheckOrthogonal();
   } 
-  void CheckOrthogonal(void){
-    CoarseVector iProj(CoarseGrid); 
-    CoarseVector eProj(CoarseGrid); 
-    for(int i=0;i<nbasis;i++){
-      blockProject(iProj,subspace[i],subspace);
-      eProj=Zero(); 
-      accelerator_for(ss, CoarseGrid->oSites(),1,{
-	eProj[ss](i)=CComplex(1.0);
-      });
-      eProj=eProj - iProj;
-      std::cout<<GridLogMessage<<"Orthog check error "<<i<<" " << norm2(eProj)<<std::endl;
-    }
-    std::cout<<GridLogMessage <<"CheckOrthog done"<<std::endl;
-  }
   void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
     blockProject(CoarseVec,FineVec,subspace);
   }
@@ -175,11 +119,6 @@ public:
     FineVec.Checkerboard() = subspace[0].Checkerboard();
     blockPromote(CoarseVec,FineVec,subspace);
   }
-  void CreateSubspaceRandom(GridParallelRNG &RNG){
-    for(int i=0;i<nbasis;i++){
-      random(RNG,subspace[i]);
-    }
-  }
 
   virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
 
@@ -190,12 +129,12 @@ public:
     FineField Mn(FineGrid);
 
     for(int b=0;b<nn;b++){
-	
+      
       subspace[b] = Zero();
       gaussian(RNG,noise);
       scale = std::pow(norm2(noise),-0.5); 
       noise=noise*scale;
-	
+      
       hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
 
       for(int i=0;i<1;i++){
@@ -218,7 +157,7 @@ public:
   // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
   // and this is the best I found
   ////////////////////////////////////////////////////////////////////////////////////////////////
-#if 1
+
   virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
 				       int nn,
 				       double hi,
@@ -280,10 +219,10 @@ public:
 	
 	hermop.HermOp(*Tn,y);
 
-	auto y_v = y.View();
-	auto Tn_v = Tn->View();
-	auto Tnp_v = Tnp->View();
-	auto Tnm_v = Tnm->View();
+	autoView( y_v , y, AcceleratorWrite);
+	autoView( Tn_v , (*Tn), AcceleratorWrite);
+	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
+	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
 	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
@@ -313,201 +252,6 @@ public:
     }
     assert(b==nn);
   }
-#endif
-#if 0
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-    FineField combined(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-#define FILTERb(llo,hhi,oorder)						\
-    {									\
-      Chebyshev<FineField> Cheb(llo,hhi,oorder);			\
-      Cheb(hermop,noise,Mn);						\
-      scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;			\
-      subspace[b]   = Mn;						\
-      hermop.Op(Mn,tmp);						\
-      std::cout<<GridLogMessage << oorder<< " Cheb filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
-      b++;								\
-    }									
-
-    //      JacobiPolynomial<FineField> Cheb(0.002,60.0,1500,-0.5,3.5);	\
-
-    RealD alpha=-0.8;
-    RealD beta =-0.8;
-#define FILTER(llo,hhi,oorder)						\
-    {									\
-      Chebyshev<FineField> Cheb(llo,hhi,oorder);			\
-      /* JacobiPolynomial<FineField> Cheb(0.0,60.0,oorder,alpha,beta);*/\
-      Cheb(hermop,noise,Mn);						\
-      scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;			\
-      subspace[b]   = Mn;						\
-      hermop.Op(Mn,tmp);						\
-      std::cout<<GridLogMessage << oorder<< "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
-      b++;								\
-    }									
-    
-#define FILTERc(llo,hhi,oorder)				\
-    {							\
-      Chebyshev<FineField> Cheb(llo,hhi,oorder);	\
-      Cheb(hermop,noise,combined);			\
-    }									
-
-    double node = 0.000;
-    FILTERb(lo,hi,orderfilter);// 0
-    //    FILTERc(node,hi,51);// 0
-    noise = Mn;
-    int base = 0;
-    int mult = 100;
-    FILTER(node,hi,base+1*mult);
-    FILTER(node,hi,base+2*mult);
-    FILTER(node,hi,base+3*mult);
-    FILTER(node,hi,base+4*mult);
-    FILTER(node,hi,base+5*mult);
-    FILTER(node,hi,base+6*mult);
-    FILTER(node,hi,base+7*mult);
-    FILTER(node,hi,base+8*mult);
-    FILTER(node,hi,base+9*mult);
-    FILTER(node,hi,base+10*mult);
-    FILTER(node,hi,base+11*mult);
-    FILTER(node,hi,base+12*mult);
-    FILTER(node,hi,base+13*mult);
-    FILTER(node,hi,base+14*mult);
-    FILTER(node,hi,base+15*mult);
-    assert(b==nn);
-  }
-#endif
-
-#if 0
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-    FineField combined(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-    {						
-      Chebyshev<FineField> JacobiPoly(0.005,60.,1500);
-      //      JacobiPolynomial<FineField> JacobiPoly(0.002,60.0,1500,-0.5,3.5);
-      //JacobiPolynomial<FineField> JacobiPoly(0.03,60.0,500,-0.5,3.5);
-      //      JacobiPolynomial<FineField> JacobiPoly(0.00,60.0,1000,-0.5,3.5);
-      JacobiPoly(hermop,noise,Mn);
-      scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;
-      subspace[b]   = Mn;
-      hermop.Op(Mn,tmp);
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; 
-      b++;
-      //      scale = std::pow(norm2(tmp),-0.5);     tmp=tmp*scale;
-      //      subspace[b]   = tmp;      b++;
-      //    }									
-    }									
-
-#define FILTER(lambda)						\
-    {								\
-      hermop.HermOp(subspace[0],tmp);				\
-      tmp = tmp - lambda *subspace[0];				\
-      scale = std::pow(norm2(tmp),-0.5);			\
-      tmp=tmp*scale;							\
-      subspace[b]   = tmp;						\
-      hermop.Op(subspace[b],tmp);					\
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
-      b++;								\
-    }									
-    //      scale = std::pow(norm2(tmp),-0.5);     tmp=tmp*scale;
-    //      subspace[b]   = tmp;      b++;
-    //    }									
-
-    FILTER(2.0e-5);
-    FILTER(2.0e-4);
-    FILTER(4.0e-4);
-    FILTER(8.0e-4);
-    FILTER(8.0e-4);
-
-    FILTER(2.0e-3);
-    FILTER(3.0e-3);
-    FILTER(4.0e-3);
-    FILTER(5.0e-3);
-    FILTER(6.0e-3);
-
-    FILTER(2.5e-3);
-    FILTER(3.5e-3);
-    FILTER(4.5e-3);
-    FILTER(5.5e-3);
-    FILTER(6.5e-3);
-
-    //    FILTER(6.0e-5);//6
-    //    FILTER(7.0e-5);//8
-    //    FILTER(8.0e-5);//9
-    //    FILTER(9.0e-5);//3
-
-    /*
-    //    FILTER(1.0e-4);//10
-    FILTER(2.0e-4);//11
-    //   FILTER(3.0e-4);//12
-    //    FILTER(4.0e-4);//13
-    FILTER(5.0e-4);//14
-
-    FILTER(6.0e-3);//4
-    FILTER(7.0e-4);//1
-    FILTER(8.0e-4);//7
-    FILTER(9.0e-4);//15
-    FILTER(1.0e-3);//2
-
-    FILTER(2.0e-3);//2
-    FILTER(3.0e-3);//2
-    FILTER(4.0e-3);//2
-    FILTER(5.0e-3);//2
-    FILTER(6.0e-3);//2
-
-    FILTER(7.0e-3);//2
-    FILTER(8.0e-3);//2
-    FILTER(1.0e-2);//2
-    */
-    std::cout << GridLogMessage <<"Jacobi filtering done" <<std::endl;
-    assert(b==nn);
-  }
-#endif
-
 
 };
 
@@ -535,7 +279,7 @@ public:
   CartesianStencil<siteVector,siteVector,int> Stencil; 
 
   std::vector<CoarseMatrix> A;
-      
+    
   ///////////////////////
   // Interface
   ///////////////////////
@@ -549,13 +293,13 @@ public:
     SimpleCompressor<siteVector> compressor;
 
     Stencil.HaloExchange(in,compressor);
-
-    auto in_v = in.View();
-    auto out_v = out.View();
+    autoView( in_v , in, AcceleratorRead);
+    autoView( out_v , out, AcceleratorWrite);
     typedef LatticeView<Cobj> Aview;
-
+      
     Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
+  
+    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();
@@ -572,24 +316,25 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      int lane=SIMTlane(Nsimd);
       for(int point=0;point<geom.npoint;point++){
 
 	SE=Stencil.GetEntry(ptype,point,ss);
 	  
 	if(SE->_is_local) { 
-	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute,lane);
+	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
 	} else {
-	  nbr = coalescedRead(Stencil.CommBuf()[SE->_offset],lane);
+	  nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
 	}
-	synchronise();
+	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,lane);
-    });
+      coalescedWrite(out_v[ss](b),res);
+      });
+
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   };
 
   void Mdag (const CoarseVector &in, CoarseVector &out)
@@ -617,11 +362,11 @@ public:
 
     typedef LatticeView<Cobj> Aview;
     Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
-    auto out_v = out.View();
-    auto in_v  = in.View();
+    autoView( out_v , out, AcceleratorWrite);
+    autoView( in_v  , in, AcceleratorRead);
 
     const int Nsimd = CComplex::Nsimd();
     typedef decltype(coalescedRead(in_v[0])) calcVector;
@@ -635,45 +380,21 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      int lane=SIMTlane(Nsimd);
       SE=Stencil.GetEntry(ptype,point,ss);
 	  
       if(SE->_is_local) { 
-	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute,lane);
+	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
       } else {
-	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset],lane);
+	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
       }
-      synchronise();
+      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,lane);
+      coalescedWrite(out_v[ss](b),res);
     });
-#if 0
-    accelerator_for(ss,Grid()->oSites(),1,{
-
-      siteVector res = Zero();
-      siteVector nbr;
-      int ptype;
-      StencilEntry *SE;
-      
-      SE=Stencil.GetEntry(ptype,point,ss);
-      
-      if(SE->_is_local&&SE->_permute) {
-	permute(nbr,in_v[SE->_offset],ptype);
-      } else if(SE->_is_local) {
-	nbr = in_v[SE->_offset];
-      } else {
-	nbr = Stencil.CommBuf()[SE->_offset];
-      }
-      synchronise();
-
-      res = res + Aview_p[point][ss]*nbr;
-      
-      out_v[ss]=res;
-    });
-#endif
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   }
   void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
   {
@@ -841,10 +562,10 @@ public:
 	    
 	    blockMaskedInnerProduct(oZProj,omask,Subspace.subspace[j],Mphi);
 	    
-	    auto iZProj_v = iZProj.View() ;
-	    auto oZProj_v = oZProj.View() ;
-	    auto A_p     =  A[p].View();
-	    auto A_self  = A[self_stencil].View();
+	    autoView( iZProj_v , iZProj, AcceleratorRead) ;
+	    autoView( oZProj_v , oZProj, AcceleratorRead) ;
+	    autoView( A_p     ,  A[p], AcceleratorWrite);
+	    autoView( A_self  , A[self_stencil], AcceleratorWrite);
 
 	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
 
@@ -860,11 +581,11 @@ public:
 	mult(tmp,phi,oddmask );  linop.Op(tmp,Mphio);
 
 	{
-	  auto tmp_      = tmp.View();
-	  auto evenmask_ = evenmask.View();
-	  auto oddmask_  =  oddmask.View();
-	  auto Mphie_    =  Mphie.View();
-	  auto Mphio_    =  Mphio.View();
+	  autoView( tmp_      , tmp, AcceleratorWrite);
+	  autoView( evenmask_ , evenmask, AcceleratorRead);
+	  autoView( oddmask_  ,  oddmask, AcceleratorRead);
+	  autoView( Mphie_    ,  Mphie, AcceleratorRead);
+	  autoView( Mphio_    ,  Mphio, AcceleratorRead);
 	  accelerator_for(ss, FineGrid->oSites(), Fobj::Nsimd(),{ 
 	      coalescedWrite(tmp_[ss],evenmask_(ss)*Mphie_(ss) + oddmask_(ss)*Mphio_(ss));
 	    });
@@ -872,8 +593,8 @@ public:
 
 	blockProject(SelfProj,tmp,Subspace.subspace);
 
-	auto SelfProj_ = SelfProj.View();
-	auto A_self  = A[self_stencil].View();
+	autoView( SelfProj_ , SelfProj, AcceleratorRead);
+	autoView( A_self  , A[self_stencil], AcceleratorWrite);
 
 	accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{
 	  for(int j=0;j<nbasis;j++){
@@ -887,33 +608,8 @@ public:
       std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
       ForceHermitian();
     }
-      // AssertHermitian();
-      // ForceDiagonal();
   }
 
-#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<< 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
-
-
   void ForceHermitian(void) {
     CoarseMatrix Diff  (Grid());
     for(int p=0;p<geom.npoint;p++){
@@ -933,27 +629,6 @@ public:
       }
     }
   }
-  void AssertHermitian(void) {
-    CoarseMatrix AA    (Grid());
-    CoarseMatrix AAc   (Grid());
-    CoarseMatrix Diff  (Grid());
-    for(int d=0;d<4;d++){
-	
-      int dd=d+1;
-      AAc = Cshift(A[2*d+1],dd,1);
-      AA  = A[2*d];
-	
-      Diff = AA - adj(AAc);
-
-      std::cout<<GridLogMessage<<"Norm diff dim "<<d<<" "<< norm2(Diff)<<std::endl;
-      std::cout<<GridLogMessage<<"Norm dim "<<d<<" "<< norm2(AA)<<std::endl;
-	  
-    }
-    Diff = A[8] - adj(A[8]);
-    std::cout<<GridLogMessage<<"Norm diff local "<< norm2(Diff)<<std::endl;
-    std::cout<<GridLogMessage<<"Norm local "<< norm2(A[8])<<std::endl;
-  }
-    
 };
 
 NAMESPACE_END(Grid);
diff --git a/Grid/algorithms/FFT.h b/Grid/algorithms/FFT.h
index 0d568c95..ad42f049 100644
--- a/Grid/algorithms/FFT.h
+++ b/Grid/algorithms/FFT.h
@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -37,7 +36,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif
 #endif
 
-
 NAMESPACE_BEGIN(Grid);
 
 template<class scalar> struct FFTW { };
@@ -191,7 +189,7 @@ public:
     typedef typename sobj::scalar_type   scalar;
       
     Lattice<sobj> pgbuf(&pencil_g);
-    auto pgbuf_v = pgbuf.View();
+    autoView(pgbuf_v , pgbuf, CpuWrite);
 
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -232,15 +230,18 @@ public:
     result = source;
     int pc = processor_coor[dim];
     for(int p=0;p<processors[dim];p++) {
-      thread_for(idx, sgrid->lSites(),{
+      {
+	autoView(r_v,result,CpuRead);
+	autoView(p_v,pgbuf,CpuWrite);
+	thread_for(idx, sgrid->lSites(),{
           Coordinate cbuf(Nd);
           sobj s;
 	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,result,cbuf);
+	  peekLocalSite(s,r_v,cbuf);
 	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  //            cbuf[dim]+=p*L;
-	  pokeLocalSite(s,pgbuf,cbuf);
-      });
+	  pokeLocalSite(s,p_v,cbuf);
+        });
+      }
       if (p != processors[dim] - 1) {
 	result = Cshift(result,dim,L);
       }
@@ -269,15 +270,19 @@ public:
     flops+= flops_call*NN;
       
     // writing out result
-    thread_for(idx,sgrid->lSites(),{
+    {
+      autoView(pgbuf_v,pgbuf,CpuRead);
+      autoView(result_v,result,CpuWrite);
+      thread_for(idx,sgrid->lSites(),{
 	Coordinate clbuf(Nd), cgbuf(Nd);
 	sobj s;
 	sgrid->LocalIndexToLocalCoor(idx,clbuf);
 	cgbuf = clbuf;
 	cgbuf[dim] = clbuf[dim]+L*pc;
-	peekLocalSite(s,pgbuf,cgbuf);
-	pokeLocalSite(s,result,clbuf);
-    });
+	peekLocalSite(s,pgbuf_v,cgbuf);
+	pokeLocalSite(s,result_v,clbuf);
+      });
+    }
     result = result*div;
       
     // destroying plan
diff --git a/Grid/algorithms/iterative/BiCGSTAB.h b/Grid/algorithms/iterative/BiCGSTAB.h
index 3a7be1ef..f4e5cdda 100644
--- a/Grid/algorithms/iterative/BiCGSTAB.h
+++ b/Grid/algorithms/iterative/BiCGSTAB.h
@@ -122,12 +122,14 @@ class BiCGSTAB : public OperatorFunction<Field>
 
         LinearCombTimer.Start();
         bo = beta * omega;
-        auto p_v = p.View();
-        auto r_v = r.View();
-        auto v_v = v.View();
-        accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
-        });
+	{
+	  autoView( p_v , p, AcceleratorWrite);
+	  autoView( r_v , r, AcceleratorRead);
+	  autoView( v_v , v, AcceleratorRead);
+	  accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
+	    });
+	}
         LinearCombTimer.Stop();
         LinalgTimer.Stop();
 
@@ -142,16 +144,20 @@ class BiCGSTAB : public OperatorFunction<Field>
         alpha = rho / Calpha.real();
 
         LinearCombTimer.Start();
-        auto h_v = h.View();
-        auto psi_v = psi.View();
-        accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
-        });
-        
-        auto s_v = s.View();
-        accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
-        });
+	{
+	  autoView( p_v , p, AcceleratorRead);
+	  autoView( r_v , r, AcceleratorRead);
+	  autoView( v_v , v, AcceleratorRead);
+	  autoView( psi_v,psi, AcceleratorRead);
+	  autoView( h_v  ,  h, AcceleratorWrite);
+	  autoView( s_v  ,  s, AcceleratorWrite);
+	  accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
+	    });
+	  accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
+ 	  });
+        }
         LinearCombTimer.Stop();
         LinalgTimer.Stop();
 
@@ -166,13 +172,19 @@ class BiCGSTAB : public OperatorFunction<Field>
         omega = Comega.real() / norm2(t);
 
         LinearCombTimer.Start();
-        auto t_v = t.View();
-        accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
-          coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
-        });
+	{
+	  autoView( psi_v,psi, AcceleratorWrite);
+	  autoView( r_v , r, AcceleratorWrite);
+	  autoView( h_v , h, AcceleratorRead);
+	  autoView( s_v , s, AcceleratorRead);
+	  autoView( t_v , t, AcceleratorRead);
+	  accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
+	      coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
+	    });
+	}
         LinearCombTimer.Stop();
-
+	
         cp = norm2(r);
         LinalgTimer.Stop();
 
diff --git a/Grid/algorithms/iterative/ConjugateGradient.h b/Grid/algorithms/iterative/ConjugateGradient.h
index 3a2544b5..14f3d306 100644
--- a/Grid/algorithms/iterative/ConjugateGradient.h
+++ b/Grid/algorithms/iterative/ConjugateGradient.h
@@ -140,13 +140,15 @@ public:
       b = cp / c;
 
       LinearCombTimer.Start();
-      auto psi_v = psi.View();
-      auto p_v   = p.View();
-      auto r_v   = r.View();
-      accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
-	  coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
-	  coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
-      });
+      {
+	autoView( psi_v , psi, AcceleratorWrite);
+	autoView( p_v   , p,   AcceleratorWrite);
+	autoView( r_v   , r,   AcceleratorWrite);
+	accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+	    coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+	    coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+	});
+      }
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
diff --git a/Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h b/Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h
new file mode 100644
index 00000000..22b7725e
--- /dev/null
+++ b/Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h
@@ -0,0 +1,241 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/PrecGeneralisedConjugateResidual.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_PREC_GCR_NON_HERM_H
+#define GRID_PREC_GCR_NON_HERM_H
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+//VPGCR Abe and Zhang, 2005.
+//INTERNATIONAL JOURNAL OF NUMERICAL ANALYSIS AND MODELING
+//Computing and Information Volume 2, Number 2, Pages 147-161
+//NB. Likely not original reference since they are focussing on a preconditioner variant.
+//    but VPGCR was nicely written up in their paper
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+
+#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" " 
+
+template<class Field>
+class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
+public:                                                
+
+  RealD   Tolerance;
+  Integer MaxIterations;
+  int verbose;
+  int mmax;
+  int nstep;
+  int steps;
+  int level;
+  GridStopWatch PrecTimer;
+  GridStopWatch MatTimer;
+  GridStopWatch LinalgTimer;
+
+  LinearFunction<Field>     &Preconditioner;
+  LinearOperatorBase<Field> &Linop;
+
+  void Level(int lv) { level=lv; };
+
+  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
+    Tolerance(tol), 
+    MaxIterations(maxit),
+    Linop(_Linop),
+    Preconditioner(Prec),
+    mmax(_mmax),
+    nstep(_nstep)
+  { 
+    level=1;
+    verbose=1;
+  };
+
+  void operator() (const Field &src, Field &psi){
+
+    psi=Zero();
+    RealD cp, ssq,rsq;
+    ssq=norm2(src);
+    rsq=Tolerance*Tolerance*ssq;
+      
+    Field r(src.Grid());
+
+    PrecTimer.Reset();
+    MatTimer.Reset();
+    LinalgTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    steps=0;
+    for(int k=0;k<MaxIterations;k++){
+
+      cp=GCRnStep(src,psi,rsq);
+
+      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
+
+      if(cp<rsq) {
+
+	SolverTimer.Stop();
+
+	Linop.Op(psi,r);
+	axpy(r,-1.0,src,r);
+	RealD tr = norm2(r);
+	GCRLogLevel<<"PGCR: Converged on iteration " <<steps
+		 << " computed residual "<<sqrt(cp/ssq)
+		 << " true residual "    <<sqrt(tr/ssq)
+		 << " target "           <<Tolerance <<std::endl;
+
+	GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
+	return;
+      }
+
+    }
+    GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
+    //    assert(0);
+  }
+
+  RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
+
+    RealD cp;
+    ComplexD a, b, zAz;
+    RealD zAAz;
+    ComplexD rq;
+
+    GridBase *grid = src.Grid();
+
+    Field r(grid);
+    Field z(grid);
+    Field tmp(grid);
+    Field ttmp(grid);
+    Field Az(grid);
+
+    ////////////////////////////////
+    // history for flexible orthog
+    ////////////////////////////////
+    std::vector<Field> q(mmax,grid);
+    std::vector<Field> p(mmax,grid);
+    std::vector<RealD> qq(mmax);
+      
+    GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
+
+    //////////////////////////////////
+    // initial guess x0 is taken as nonzero.
+    // r0=src-A x0 = src
+    //////////////////////////////////
+    MatTimer.Start();
+    Linop.Op(psi,Az);
+    zAz = innerProduct(Az,psi);
+    zAAz= norm2(Az);
+    MatTimer.Stop();
+    
+
+    LinalgTimer.Start();
+    r=src-Az;
+    LinalgTimer.Stop();
+    GCRLogLevel<< "PGCR true residual r = src - A psi   "<<norm2(r) <<std::endl;
+    
+    /////////////////////
+    // p = Prec(r)
+    /////////////////////
+
+    PrecTimer.Start();
+    Preconditioner(r,z);
+    PrecTimer.Stop();
+
+    MatTimer.Start();
+    Linop.Op(z,Az);
+    MatTimer.Stop();
+
+    LinalgTimer.Start();
+
+    zAz = innerProduct(Az,psi);
+    zAAz= norm2(Az);
+
+    //p[0],q[0],qq[0] 
+    p[0]= z;
+    q[0]= Az;
+    qq[0]= zAAz;
+    
+    cp =norm2(r);
+    LinalgTimer.Stop();
+
+    for(int k=0;k<nstep;k++){
+
+      steps++;
+
+      int kp     = k+1;
+      int peri_k = k %mmax;
+      int peri_kp= kp%mmax;
+
+      LinalgTimer.Start();
+      rq= innerProduct(q[peri_k],r); // what if rAr not real?
+      a = rq/qq[peri_k];
+
+      axpy(psi,a,p[peri_k],psi);         
+
+      cp = axpy_norm(r,-a,q[peri_k],r);
+      LinalgTimer.Stop();
+
+      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
+
+      if((k==nstep-1)||(cp<rsq)){
+	return cp;
+      }
+
+
+      PrecTimer.Start();
+      Preconditioner(r,z);// solve Az = r
+      PrecTimer.Stop();
+
+      MatTimer.Start();
+      Linop.Op(z,Az);
+      MatTimer.Stop();
+      zAz = innerProduct(Az,psi);
+      zAAz= norm2(Az);
+
+      LinalgTimer.Start();
+
+      q[peri_kp]=Az;
+      p[peri_kp]=z;
+
+      int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
+      for(int back=0;back<northog;back++){
+
+	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+
+	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
+	p[peri_kp]=p[peri_kp]+b*p[peri_back];
+	q[peri_kp]=q[peri_kp]+b*q[peri_back];
+
+      }
+      qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
+      LinalgTimer.Stop();
+    }
+    assert(0); // never reached
+    return cp;
+  }
+};
+NAMESPACE_END(Grid);
+#endif
diff --git a/Grid/allocator/AlignedAllocator.cc b/Grid/allocator/AlignedAllocator.cc
index ef6459ed..0d1707d9 100644
--- a/Grid/allocator/AlignedAllocator.cc
+++ b/Grid/allocator/AlignedAllocator.cc
@@ -6,93 +6,6 @@ NAMESPACE_BEGIN(Grid);
 MemoryStats *MemoryProfiler::stats = nullptr;
 bool         MemoryProfiler::debug = false;
 
-int PointerCache::NcacheSmall = PointerCache::NcacheSmallMax;
-#ifdef GRID_CUDA
-int PointerCache::Ncache      = 32;
-#else 
-int PointerCache::Ncache      = 8;
-#endif
-int PointerCache::Victim;
-int PointerCache::VictimSmall;
-PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::NcacheMax];
-PointerCache::PointerCacheEntry PointerCache::EntriesSmall[PointerCache::NcacheSmallMax];
-
-void PointerCache::Init(void)
-{
-  char * str;
-
-  str= getenv("GRID_ALLOC_NCACHE_LARGE");
-  if ( str ) Ncache = atoi(str);
-  if ( (Ncache<0) || (Ncache > NcacheMax)) Ncache = NcacheMax;
-
-  str= getenv("GRID_ALLOC_NCACHE_SMALL");
-  if ( str ) NcacheSmall = atoi(str);
-  if ( (NcacheSmall<0) || (NcacheSmall > NcacheSmallMax)) NcacheSmall = NcacheSmallMax;
-
-  //  printf("Aligned alloocator cache: large %d/%d small %d/%d\n",Ncache,NcacheMax,NcacheSmall,NcacheSmallMax);
-}
-void *PointerCache::Insert(void *ptr,size_t bytes) 
-{
-  if (bytes < GRID_ALLOC_SMALL_LIMIT ) 
-    return Insert(ptr,bytes,EntriesSmall,NcacheSmall,VictimSmall);
-  return Insert(ptr,bytes,Entries,Ncache,Victim);  
-}
-void *PointerCache::Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) 
-{
-#ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
-#endif 
-
-  void * ret = NULL;
-  int v = -1;
-
-  for(int e=0;e<ncache;e++) {
-    if ( entries[e].valid==0 ) {
-      v=e; 
-      break;
-    }
-  }
-
-  if ( v==-1 ) {
-    v=victim;
-    victim = (victim+1)%ncache;
-  }
-
-  if ( entries[v].valid ) {
-    ret = entries[v].address;
-    entries[v].valid = 0;
-    entries[v].address = NULL;
-    entries[v].bytes = 0;
-  }
-
-  entries[v].address=ptr;
-  entries[v].bytes  =bytes;
-  entries[v].valid  =1;
-
-  return ret;
-}
-
-void *PointerCache::Lookup(size_t bytes)
-{
-  if (bytes < GRID_ALLOC_SMALL_LIMIT ) 
-    return Lookup(bytes,EntriesSmall,NcacheSmall);
-  return Lookup(bytes,Entries,Ncache);
-}
-void *PointerCache::Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) 
-{
-#ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
-#endif 
-  for(int e=0;e<ncache;e++){
-    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
-      entries[e].valid = 0;
-      return entries[e].address;
-    }
-  }
-  return NULL;
-}
-
-
 void check_huge_pages(void *Buf,uint64_t BYTES)
 {
 #ifdef __linux__
diff --git a/Grid/allocator/AlignedAllocator.h b/Grid/allocator/AlignedAllocator.h
index 600b7097..ebb3162b 100644
--- a/Grid/allocator/AlignedAllocator.h
+++ b/Grid/allocator/AlignedAllocator.h
@@ -26,121 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_ALIGNED_ALLOCATOR_H
-#define GRID_ALIGNED_ALLOCATOR_H
-
-#ifdef HAVE_MALLOC_MALLOC_H
-#include <malloc/malloc.h>
-#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_MM_MALLOC_H
-#include <mm_malloc.h>
-#endif
-
-#define POINTER_CACHE
-#define GRID_ALLOC_ALIGN (2*1024*1024)
-#define GRID_ALLOC_SMALL_LIMIT (4096)
+#pragma once
 
 NAMESPACE_BEGIN(Grid);
 
-// Move control to configure.ac and Config.h?
-
-class PointerCache {
-private:
-/*Pinning pages is costly*/
-/*Could maintain separate large and small allocation caches*/
-/* Could make these configurable, perhaps up to a max size*/
-  static const int NcacheSmallMax=128; 
-  static const int NcacheMax=16;
-  static int NcacheSmall;
-  static int Ncache;
-
-  typedef struct { 
-    void *address;
-    size_t bytes;
-    int valid;
-  } PointerCacheEntry;
-    
-  static PointerCacheEntry Entries[NcacheMax];
-  static int Victim;
-  static PointerCacheEntry EntriesSmall[NcacheSmallMax];
-  static int VictimSmall;
-
-public:
-  static void Init(void);
-  static void *Insert(void *ptr,size_t bytes) ;
-  static void *Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) ;
-  static void *Lookup(size_t bytes) ;
-  static void *Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) ;
-};
-
-std::string sizeString(size_t bytes);
-
-struct MemoryStats
-{
-  size_t totalAllocated{0}, maxAllocated{0}, 
-    currentlyAllocated{0}, totalFreed{0};
-};
-    
-class MemoryProfiler
-{
-public:
-  static MemoryStats *stats;
-  static bool        debug;
-};
-
-#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
-#define profilerDebugPrint						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl; \
-      std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
-		<< std::endl;						\
-    }
-
-#define profilerAllocate(bytes)						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      s->totalAllocated     += (bytes);					\
-      s->currentlyAllocated += (bytes);					\
-      s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated); \
-    }									\
-  if (MemoryProfiler::debug)						\
-    {									\
-      std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl; \
-      profilerDebugPrint;						\
-    }
-
-#define profilerFree(bytes)						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      s->totalFreed         += (bytes);					\
-      s->currentlyAllocated -= (bytes);					\
-    }									\
-  if (MemoryProfiler::debug)						\
-    {									\
-      std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl; \
-      profilerDebugPrint;						\
-    }
-
-void check_huge_pages(void *Buf,uint64_t BYTES);
-
-////////////////////////////////////////////////////////////////////
-// A lattice of something, but assume the something is SIMDized.
-////////////////////////////////////////////////////////////////////
-
 template<typename _Tp>
 class alignedAllocator {
 public: 
@@ -164,71 +53,60 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
-
-
-#ifdef POINTER_CACHE
-    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
-#else
-    pointer ptr = nullptr;
-#endif
-
-#ifdef GRID_NVCC
-    ////////////////////////////////////
-    // Unified (managed) memory
-    ////////////////////////////////////
-    if ( ptr == (_Tp *) NULL ) {
-      //      printf(" alignedAllocater cache miss %ld bytes ",bytes);      BACKTRACEFP(stdout);
-      //      auto err = 
-gridMallocManaged((void **)&ptr,bytes);
-/*if( err != cudaSuccess ) {
-	ptr = (_Tp *) NULL;
-	std::cerr << " cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
-	assert(0);
-	}*/
-    } 
-    assert( ptr != (_Tp *)NULL);
-#else 
-    //////////////////////////////////////////////////////////////////////////////////////////
-    // 2MB align; could make option probably doesn't need configurability
-    //////////////////////////////////////////////////////////////////////////////////////////
-  #ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
-  #else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
-  #endif
-    assert( ptr != (_Tp *)NULL);
-
-    //////////////////////////////////////////////////
-    // First touch optimise in threaded loop 
-    //////////////////////////////////////////////////
-    uint64_t *cp = (uint64_t *)ptr;
-    thread_for(n,bytes/sizeof(uint64_t), { // need only one touch per page
-      cp[n]=0;
-    });
-#endif
+    _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
-  void deallocate(pointer __p, size_type __n) { 
+  void deallocate(pointer __p, size_type __n) 
+  { 
     size_type bytes = __n * sizeof(_Tp);
-
     profilerFree(bytes);
+    MemoryManager::CpuFree((void *)__p,bytes);
+  }
 
-#ifdef POINTER_CACHE
-    pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
-#else 
-    pointer __freeme = __p;
-#endif
+  // FIXME: hack for the copy constructor, eventually it must be avoided
+  //void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
+  void construct(pointer __p, const _Tp& __val) { assert(0);};
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
-#ifdef GRID_NVCC
-    if ( __freeme ) gridFree((void *)__freeme);
-#else 
-  #ifdef HAVE_MM_MALLOC_H
-    if ( __freeme ) _mm_free((void *)__freeme); 
-  #else
-    if ( __freeme ) free((void *)__freeme);
-  #endif
-#endif
+template<typename _Tp>
+class uvmAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef uvmAllocator<_Tp1> other; };
+  uvmAllocator() throw() { }
+  uvmAllocator(const uvmAllocator&) throw() { }
+  template<typename _Tp1> uvmAllocator(const uvmAllocator<_Tp1>&) throw() { }
+  ~uvmAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+  pointer allocate(size_type __n, const void* _p= 0)
+  { 
+    size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
+    _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
+    profilerFree(bytes);
+    MemoryManager::SharedFree((void *)__p,bytes);
   }
 
   // FIXME: hack for the copy constructor, eventually it must be avoided
@@ -237,17 +115,17 @@ gridMallocManaged((void **)&ptr,bytes);
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
-template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
-template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
+template<typename _Tp>  inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using commAllocator = alignedAllocator<T>;
-template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
-template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
+template<class T> using commAllocator = uvmAllocator<T>;
+template<class T> using Vector     = std::vector<T,uvmAllocator<T> >;           
+template<class T> using commVector = std::vector<T,uvmAllocator<T> >;
+//template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
 
 NAMESPACE_END(Grid);
 
-#endif
+
diff --git a/Grid/allocator/Allocator.h b/Grid/allocator/Allocator.h
new file mode 100644
index 00000000..589ea36f
--- /dev/null
+++ b/Grid/allocator/Allocator.h
@@ -0,0 +1,4 @@
+#pragma once
+#include <Grid/allocator/MemoryStats.h>
+#include <Grid/allocator/MemoryManager.h>
+#include <Grid/allocator/AlignedAllocator.h>
diff --git a/Grid/allocator/GridMemoryManager.cc b/Grid/allocator/GridMemoryManager.cc
deleted file mode 100644
index 369f72f7..00000000
--- a/Grid/allocator/GridMemoryManager.cc
+++ /dev/null
@@ -1,145 +0,0 @@
-#include <Grid/GridCore.h>
-
-NAMESPACE_BEGIN(Grid);
-
-#define _GRID_MEM_PAGE_SIZE 4096
-void* _grid_mem_base = 0;
-size_t _grid_mem_pages;
-struct _grid_mem_range {
-  size_t page_start, page_end;
-};
-std::vector<_grid_mem_range> _grid_mem_avail;
-std::map<void*,_grid_mem_range> _grid_mem_alloc;
-
-void gridMemoryInit() {
-#ifdef GRID_NVCC
-  size_t free,total;
-  cudaMemGetInfo(&free,&total);
-  
-  char* ev = getenv("GRID_DEVICE_BYTES_FOR_CACHE");
-  if (ev) {
-    long bytes;
-    assert(sscanf(ev,"%ld",&bytes)==1);
-    free -= bytes;
-  }
-
-  _grid_mem_pages = free / _GRID_MEM_PAGE_SIZE;
-  size_t sz = _grid_mem_pages * _GRID_MEM_PAGE_SIZE;
-
-  assert(cudaSuccess == cudaMallocManaged(&_grid_mem_base,sz));
-  
-  int target;
-  cudaGetDevice(&target);
-  cudaMemAdvise(_grid_mem_base,sz,cudaMemAdviseSetPreferredLocation,target);
-
-  assert(cudaSuccess == cudaMemset(_grid_mem_base,0,sz)); // touch on device
-  std::cout << GridLogMessage << "gridMemoryInit: " << sz << " bytes" << std::endl;
-
-  _grid_mem_avail.push_back( { 0, _grid_mem_pages } );
-#endif
-}
-
-void gridMallocManaged(void** pp, size_t sz) {
-#ifdef GRID_NVCC
-  if (_grid_mem_avail.empty())
-    gridMemoryInit();
-
-  size_t pages = (sz + _GRID_MEM_PAGE_SIZE - 1) / _GRID_MEM_PAGE_SIZE;
-  // find free block
-  size_t m;
-  for (m=0;m<_grid_mem_avail.size();m++) {
-    auto & b = _grid_mem_avail[m];
-    if (b.page_end - b.page_start >= pages)
-      break;
-  }
-  if (m == _grid_mem_avail.size()) {
-    std::cout << GridLogMessage << "Out of memory" << std::endl;
-    assert(0);
-  }
-  *pp = (char*)_grid_mem_base + _GRID_MEM_PAGE_SIZE*_grid_mem_avail[m].page_start;
-  _grid_mem_alloc[*pp] = { _grid_mem_avail[m].page_start, _grid_mem_avail[m].page_start + pages };
-  _grid_mem_avail[m].page_start += pages;
-#else
-  *pp = malloc(sz);
-#endif
-}
-
-void gridFree(void* p) {
-#ifdef GRID_NVCC
-  if (_grid_mem_avail.empty())
-    gridMemoryInit();
-
-  auto & alloc = _grid_mem_alloc[p];
-  if (alloc.page_start == alloc.page_end) {
-    free(p);
-    //cudaFreeHost(p);
-  } else {
-    // can we enlarge existing one?
-    for (size_t m=0;m<_grid_mem_avail.size();m++) {
-      auto & b = _grid_mem_avail[m];
-      if (b.page_start == alloc.page_end) {
-	b.page_start = alloc.page_start;
-	return;
-      }
-      if (b.page_end == alloc.page_start) {
-	b.page_end = alloc.page_end;
-	return;
-      }
-    }
-    // fragment memory
-    _grid_mem_avail.push_back( alloc );  
-  }
-  _grid_mem_alloc.erase(p);
-#else
-  free(p);
-#endif
-}
-
-void gridAcceleratorPrefetch(void* p, size_t sz) {
-#ifdef GRID_NVCC
-  auto & alloc = _grid_mem_alloc[p];
-  if (alloc.page_start == alloc.page_end) // pinned to host
-    return;
-
-  int target;
-  cudaGetDevice(&target);
-  cudaMemPrefetchAsync(p,sz,target);
-#endif
-}
-
-void gridMemGetInfo(size_t* pfree, size_t* ptotal) {
-#ifdef GRID_NVCC
-  if (_grid_mem_avail.empty())
-    gridMemoryInit();
-
-  *ptotal = _grid_mem_pages * _GRID_MEM_PAGE_SIZE;
-  *pfree = 0;
-  for (auto & a : _grid_mem_avail)
-    *pfree += (a.page_end - a.page_start) * _GRID_MEM_PAGE_SIZE;
-#else
-  *pfree = 0;
-  *ptotal = 0;
-#endif
-}
-
-void gridMoveToHost(void** pp) {
-#ifdef GRID_NVCC
-  if (_grid_mem_avail.empty())
-    gridMemoryInit();
-
-  auto & alloc = _grid_mem_alloc[*pp];
-  if (alloc.page_start == alloc.page_end) // already on host
-    return;
-
-  size_t sz = (alloc.page_end - alloc.page_start) * _GRID_MEM_PAGE_SIZE;
-  void*pn;
-  //assert(cudaSuccess == cudaMallocHost(&pn,sz));
-  pn = malloc(sz);
-  memcpy(pn,*pp,sz);
-  gridFree(*pp);
-  *pp = pn;
-  _grid_mem_alloc[pn] = { 0,0 };
-#endif
-}
-
-NAMESPACE_END(Grid);
diff --git a/Grid/allocator/GridMemoryManager.h b/Grid/allocator/GridMemoryManager.h
deleted file mode 100644
index 9e619301..00000000
--- a/Grid/allocator/GridMemoryManager.h
+++ /dev/null
@@ -1,42 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/GridMemoryManager.h
-
-    Copyright (C) 2020
-
-Author: Christoph Lehner <christoph@lhnr.de>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_MEMORY_MANAGER_H
-#define GRID_MEMORY_MANAGER_H
-
-NAMESPACE_BEGIN(Grid);
-
-void gridMemoryInit();
-void gridMallocManaged(void** pp, size_t sz);
-void gridMoveToHost(void** pp);
-void gridAcceleratorPrefetch(void* p, size_t sz);
-void gridMemGetInfo(size_t* pfree, size_t* ptotal);
-void gridFree(void* p);
-
-NAMESPACE_END(Grid);
-
-#endif
diff --git a/Grid/allocator/MemoryManager.cc b/Grid/allocator/MemoryManager.cc
new file mode 100644
index 00000000..a2866507
--- /dev/null
+++ b/Grid/allocator/MemoryManager.cc
@@ -0,0 +1,249 @@
+#include <Grid/GridCore.h>
+
+NAMESPACE_BEGIN(Grid);
+
+/*Allocation types, saying which pointer cache should be used*/
+#define Cpu      (0)
+#define CpuSmall (1)
+#define Acc      (2)
+#define AccSmall (3)
+#define Shared   (4)
+#define SharedSmall (5)
+uint64_t total_shared;
+uint64_t total_device;
+uint64_t total_host;;
+void MemoryManager::PrintBytes(void)
+{
+  std::cout << " MemoryManager : "<<total_shared<<" shared      bytes "<<std::endl;
+  std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
+  std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<std::endl;
+}
+
+//////////////////////////////////////////////////////////////////////
+// Data tables for recently freed pooiniter caches
+//////////////////////////////////////////////////////////////////////
+MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
+int MemoryManager::Victim[MemoryManager::NallocType];
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 8, 32, 8, 32, 8, 32 };
+
+//////////////////////////////////////////////////////////////////////
+// Actual allocation and deallocation utils
+//////////////////////////////////////////////////////////////////////
+void *MemoryManager::AcceleratorAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Acc);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocDevice(bytes);
+    total_device+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes)
+{
+  void *__freeme = Insert(ptr,bytes,Acc);
+  if ( __freeme ) {
+    acceleratorFreeDevice(__freeme);
+    total_device-=bytes;
+    //    PrintBytes();
+  }
+}
+void *MemoryManager::SharedAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Shared);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+    total_shared+=bytes;
+    //    std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
+    //    PrintBytes();
+  }
+  return ptr;
+}
+void  MemoryManager::SharedFree    (void *ptr,size_t bytes)
+{
+  void *__freeme = Insert(ptr,bytes,Shared);
+  if ( __freeme ) {
+    acceleratorFreeShared(__freeme);
+    total_shared-=bytes;
+    //    PrintBytes();
+  }
+}
+#ifdef GRID_UVM
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+    total_host+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeShared(__freeme);
+    total_host-=bytes;
+  }
+}
+#else
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocCpu(bytes);
+    total_host+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeCpu(__freeme);
+    total_host-=bytes;
+  }
+}
+#endif
+
+//////////////////////////////////////////
+// call only once
+//////////////////////////////////////////
+void MemoryManager::Init(void)
+{
+
+  char * str;
+  int Nc;
+  int NcS;
+  
+  str= getenv("GRID_ALLOC_NCACHE_LARGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[Cpu]=Nc;
+      Ncache[Acc]=Nc;
+      Ncache[Shared]=Nc;
+    }
+  }
+
+  str= getenv("GRID_ALLOC_NCACHE_SMALL");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuSmall]=Nc;
+      Ncache[AccSmall]=Nc;
+      Ncache[SharedSmall]=Nc;
+    }
+  }
+
+  // only root node delivers messages, this is called before communicator is initialized,
+  // so need a manual restriction
+  if ( CartesianCommunicator::RankWorld() == 0 ) {
+    std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
+#ifdef ALLOCATION_CACHE
+    std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
+#endif
+    
+#ifdef GRID_UVM
+    std::cout << GridLogMessage<< "MemoryManager::Init() Unified memory space"<<std::endl;
+#ifdef GRID_CUDA
+    std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMallocManaged"<<std::endl;
+#endif
+#ifdef GRID_HIP
+    std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMallocManaged"<<std::endl;
+#endif
+#ifdef GRID_SYCL
+    std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_shared"<<std::endl;
+#endif
+#else
+    std::cout << GridLogMessage<< "MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory"<<std::endl;
+#ifdef GRID_CUDA
+    std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMalloc"<<std::endl;
+#endif
+#ifdef GRID_HIP
+    std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMalloc"<<std::endl;
+#endif
+#ifdef GRID_SYCL
+    std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_device"<<std::endl;
+#endif
+#endif
+  }
+}
+
+void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
+{
+#ifdef ALLOCATION_CACHE
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type + small;
+  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);  
+#else
+  return ptr;
+#endif
+}
+
+void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) 
+{
+  assert(ncache>0);
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  void * ret = NULL;
+  int v = -1;
+
+  for(int e=0;e<ncache;e++) {
+    if ( entries[e].valid==0 ) {
+      v=e; 
+      break;
+    }
+  }
+
+  if ( v==-1 ) {
+    v=victim;
+    victim = (victim+1)%ncache;
+  }
+
+  if ( entries[v].valid ) {
+    ret = entries[v].address;
+    entries[v].valid = 0;
+    entries[v].address = NULL;
+    entries[v].bytes = 0;
+  }
+
+  entries[v].address=ptr;
+  entries[v].bytes  =bytes;
+  entries[v].valid  =1;
+
+  return ret;
+}
+
+void *MemoryManager::Lookup(size_t bytes,int type)
+{
+#ifdef ALLOCATION_CACHE
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type+small;
+  return Lookup(bytes,Entries[cache],Ncache[cache]);
+#else
+  return NULL;
+#endif
+}
+
+void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) 
+{
+  assert(ncache>0);
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+  for(int e=0;e<ncache;e++){
+    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
+      entries[e].valid = 0;
+      return entries[e].address;
+    }
+  }
+  return NULL;
+}
+
+
+NAMESPACE_END(Grid);
+
diff --git a/Grid/allocator/MemoryManager.h b/Grid/allocator/MemoryManager.h
new file mode 100644
index 00000000..f158bade
--- /dev/null
+++ b/Grid/allocator/MemoryManager.h
@@ -0,0 +1,181 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/MemoryManager.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+#include <list> 
+#include <unordered_map>  
+
+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*/
+////////////////////////////////////////////////////////////////////////////
+// Advise the LatticeAccelerator class
+////////////////////////////////////////////////////////////////////////////
+enum ViewAdvise {
+ AdviseDefault       = 0x0,    // Regular data
+ AdviseInfrequentUse = 0x1     // Advise that the data is used infrequently.  This can
+                               // significantly influence performance of bulk storage.
+ 
+ // AdviseTransient      = 0x2,   // Data will mostly be read.  On some architectures
+                               // enables read-only copies of memory to be kept on
+                               // host and device.
+
+ // AdviseAcceleratorWriteDiscard = 0x4  // Field will be written in entirety on device
+
+};
+
+////////////////////////////////////////////////////////////////////////////
+// View Access Mode
+////////////////////////////////////////////////////////////////////////////
+enum ViewMode {
+  AcceleratorRead  = 0x01,
+  AcceleratorWrite = 0x02,
+  AcceleratorWriteDiscard = 0x04,
+  CpuRead  = 0x08,
+  CpuWrite = 0x10,
+  CpuWriteDiscard = 0x10 // same for now
+};
+
+class MemoryManager {
+private:
+
+  ////////////////////////////////////////////////////////////
+  // For caching recently freed allocations
+  ////////////////////////////////////////////////////////////
+  typedef struct { 
+    void *address;
+    size_t bytes;
+    int valid;
+  } AllocationCacheEntry;
+
+  static const int NallocCacheMax=128; 
+  static const int NallocType=6;
+  static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
+  static int Victim[NallocType];
+  static int Ncache[NallocType];
+
+  /////////////////////////////////////////////////
+  // Free pool
+  /////////////////////////////////////////////////
+  static void *Insert(void *ptr,size_t bytes,int type) ;
+  static void *Lookup(size_t bytes,int type) ;
+  static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
+  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
+
+  static void *AcceleratorAllocate(size_t bytes);
+  static void  AcceleratorFree    (void *ptr,size_t bytes);
+  static void PrintBytes(void);
+ public:
+  static void Init(void);
+  static void *SharedAllocate(size_t bytes);
+  static void  SharedFree    (void *ptr,size_t bytes);
+  static void *CpuAllocate(size_t bytes);
+  static void  CpuFree    (void *ptr,size_t bytes);
+
+  ////////////////////////////////////////////////////////
+  // Footprint tracking
+  ////////////////////////////////////////////////////////
+  static uint64_t     DeviceBytes;
+  static uint64_t     DeviceLRUBytes;
+  static uint64_t     DeviceMaxBytes;
+  static uint64_t     HostToDeviceBytes;
+  static uint64_t     DeviceToHostBytes;
+  static uint64_t     HostToDeviceXfer;
+  static uint64_t     DeviceToHostXfer;
+ 
+ private:
+#ifndef GRID_UVM
+  //////////////////////////////////////////////////////////////////////
+  // Data tables for ViewCache
+  //////////////////////////////////////////////////////////////////////
+  typedef std::list<uint64_t> LRU_t;
+  typedef typename LRU_t::iterator LRUiterator;
+  typedef struct { 
+    int        LRU_valid;
+    LRUiterator LRU_entry;
+    uint64_t CpuPtr;
+    uint64_t AccPtr;
+    size_t   bytes;
+    uint32_t transient;
+    uint32_t state;
+    uint32_t accLock;
+    uint32_t cpuLock;
+  } AcceleratorViewEntry;
+  
+  typedef std::unordered_map<uint64_t,AcceleratorViewEntry> AccViewTable_t;
+  typedef typename AccViewTable_t::iterator AccViewTableIterator ;
+
+  static AccViewTable_t AccViewTable;
+  static LRU_t LRU;
+
+  /////////////////////////////////////////////////
+  // Device motion
+  /////////////////////////////////////////////////
+  static void  Create(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+  static void  EvictVictims(uint64_t bytes); // Frees up <bytes>
+  static void  Evict(AcceleratorViewEntry &AccCache);
+  static void  Flush(AcceleratorViewEntry &AccCache);
+  static void  Clone(AcceleratorViewEntry &AccCache);
+  static void  AccDiscard(AcceleratorViewEntry &AccCache);
+  static void  CpuDiscard(AcceleratorViewEntry &AccCache);
+
+  //  static void  LRUupdate(AcceleratorViewEntry &AccCache);
+  static void  LRUinsert(AcceleratorViewEntry &AccCache);
+  static void  LRUremove(AcceleratorViewEntry &AccCache);
+  
+  // manage entries in the table
+  static int                  EntryPresent(uint64_t CpuPtr);
+  static void                 EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+  static void                 EntryErase (uint64_t CpuPtr);
+  static AccViewTableIterator EntryLookup(uint64_t CpuPtr);
+  static void                 EntrySet   (uint64_t CpuPtr,AcceleratorViewEntry &entry);
+
+  static void     AcceleratorViewClose(uint64_t AccPtr);
+  static uint64_t AcceleratorViewOpen(uint64_t  CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+  static void     CpuViewClose(uint64_t Ptr);
+  static uint64_t CpuViewOpen(uint64_t  CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+#endif
+  static void NotifyDeletion(void * CpuPtr);
+
+ public:
+  static void Print(void);
+  static int   isOpen   (void* CpuPtr);
+  static void  ViewClose(void* CpuPtr,ViewMode mode);
+  static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+
+};
+
+NAMESPACE_END(Grid);
+
+
diff --git a/Grid/allocator/MemoryManagerCache.cc b/Grid/allocator/MemoryManagerCache.cc
new file mode 100644
index 00000000..5dd7575e
--- /dev/null
+++ b/Grid/allocator/MemoryManagerCache.cc
@@ -0,0 +1,468 @@
+#include <Grid/GridCore.h>
+
+#ifndef GRID_UVM
+
+#warning "Using explicit device memory copies"
+NAMESPACE_BEGIN(Grid);
+#define dprintf(...)
+
+////////////////////////////////////////////////////////////
+// For caching copies of data on device
+////////////////////////////////////////////////////////////
+MemoryManager::AccViewTable_t MemoryManager::AccViewTable;
+MemoryManager::LRU_t MemoryManager::LRU;
+  
+////////////////////////////////////////////////////////
+// Footprint tracking
+////////////////////////////////////////////////////////
+uint64_t  MemoryManager::DeviceBytes;
+uint64_t  MemoryManager::DeviceLRUBytes;
+uint64_t  MemoryManager::DeviceMaxBytes = 1024*1024*128;
+uint64_t  MemoryManager::HostToDeviceBytes;
+uint64_t  MemoryManager::DeviceToHostBytes;
+uint64_t  MemoryManager::HostToDeviceXfer;
+uint64_t  MemoryManager::DeviceToHostXfer;
+
+////////////////////////////////////
+// Priority ordering for unlocked entries
+//  Empty
+//  CpuDirty 
+//  Consistent
+//  AccDirty
+////////////////////////////////////
+#define Empty         (0x0)  /*Entry unoccupied  */
+#define CpuDirty      (0x1)  /*CPU copy is golden, Acc buffer MAY not be allocated*/
+#define Consistent    (0x2)  /*ACC copy AND CPU copy are valid */
+#define AccDirty      (0x4)  /*ACC copy is golden */
+#define EvictNext     (0x8)  /*Priority for eviction*/
+
+/////////////////////////////////////////////////
+// Mechanics of data table maintenance
+/////////////////////////////////////////////////
+int   MemoryManager::EntryPresent(uint64_t CpuPtr)
+{
+  if(AccViewTable.empty()) return 0;
+
+  auto count = AccViewTable.count(CpuPtr);  assert((count==0)||(count==1));
+  return count;
+}
+void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
+{
+  assert(!EntryPresent(CpuPtr));
+  AcceleratorViewEntry AccCache;
+  AccCache.CpuPtr = CpuPtr;
+  AccCache.AccPtr = (uint64_t)NULL;
+  AccCache.bytes  = bytes;
+  AccCache.state  = CpuDirty;
+  AccCache.LRU_valid=0;
+  AccCache.transient=0;
+  AccCache.accLock=0;
+  AccCache.cpuLock=0;
+  AccViewTable[CpuPtr] = AccCache;
+}
+MemoryManager::AccViewTableIterator MemoryManager::EntryLookup(uint64_t CpuPtr)
+{
+  assert(EntryPresent(CpuPtr));
+  auto AccCacheIterator = AccViewTable.find(CpuPtr);
+  assert(AccCacheIterator!=AccViewTable.end());
+  return AccCacheIterator;
+}
+void MemoryManager::EntryErase(uint64_t CpuPtr)
+{
+  auto AccCache = EntryLookup(CpuPtr);
+  AccViewTable.erase(CpuPtr);
+}
+void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
+{
+  assert(AccCache.LRU_valid==0);
+  if (AccCache.transient) { 
+    LRU.push_back(AccCache.CpuPtr);
+    AccCache.LRU_entry = --LRU.end();
+  } else {
+    LRU.push_front(AccCache.CpuPtr);
+    AccCache.LRU_entry = LRU.begin();
+  }
+  AccCache.LRU_valid = 1;
+  DeviceLRUBytes+=AccCache.bytes;
+}
+void  MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
+{
+  assert(AccCache.LRU_valid==1);
+  LRU.erase(AccCache.LRU_entry);
+  AccCache.LRU_valid = 0;
+  DeviceLRUBytes-=AccCache.bytes;
+}
+/////////////////////////////////////////////////
+// Accelerator cache motion & consistency logic
+/////////////////////////////////////////////////
+void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
+{
+  ///////////////////////////////////////////////////////////
+  // Remove from Accelerator, remove entry, without flush
+  // Cannot be locked. If allocated Must be in LRU pool.
+  ///////////////////////////////////////////////////////////
+  assert(AccCache.state!=Empty);
+  
+  //  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);
+  if(AccCache.AccPtr) {
+    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);  
+  }
+  uint64_t CpuPtr = AccCache.CpuPtr;
+  EntryErase(CpuPtr);
+}
+
+void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
+{
+  ///////////////////////////////////////////////////////////////////////////
+  // Make CPU consistent, remove from Accelerator, remove entry
+  // Cannot be locked. If allocated must be in LRU pool.
+  ///////////////////////////////////////////////////////////////////////////
+  assert(AccCache.state!=Empty);
+  
+  //  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) {
+    Flush(AccCache);
+  }
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr) {
+    AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceBytes   -=AccCache.bytes;
+    LRUremove(AccCache);
+    //    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+  }
+  uint64_t CpuPtr = AccCache.CpuPtr;
+  EntryErase(CpuPtr);
+}
+void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
+{
+  assert(AccCache.state==AccDirty);
+  assert(AccCache.cpuLock==0);
+  assert(AccCache.accLock==0);
+  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);
+  DeviceToHostBytes+=AccCache.bytes;
+  DeviceToHostXfer++;
+  AccCache.state=Consistent;
+}
+void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
+{
+  assert(AccCache.state==CpuDirty);
+  assert(AccCache.cpuLock==0);
+  assert(AccCache.accLock==0);
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr==(uint64_t)NULL){
+    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);
+  acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
+  HostToDeviceBytes+=AccCache.bytes;
+  HostToDeviceXfer++;
+  AccCache.state=Consistent;
+}
+
+void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
+{
+  assert(AccCache.state!=Empty);
+  assert(AccCache.cpuLock==0);
+  assert(AccCache.accLock==0);
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr==(uint64_t)NULL){
+    AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
+    DeviceBytes+=AccCache.bytes;
+  }
+  AccCache.state=AccDirty;
+}
+
+/////////////////////////////////////////////////////////////////////////////////
+// View management
+/////////////////////////////////////////////////////////////////////////////////
+void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
+{
+  if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    AcceleratorViewClose((uint64_t)Ptr);
+  } else if( (mode==CpuRead)||(mode==CpuWrite)){
+    CpuViewClose((uint64_t)Ptr);
+  } else { 
+    assert(0);
+  }
+}
+void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
+{
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+  if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
+  } else if( (mode==CpuRead)||(mode==CpuWrite)){
+    return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
+  } else { 
+    assert(0);
+    return NULL;
+  }
+}
+void  MemoryManager::EvictVictims(uint64_t bytes)
+{
+  while(bytes+DeviceLRUBytes > DeviceMaxBytes){
+    if ( DeviceLRUBytes > 0){
+      assert(LRU.size()>0);
+      uint64_t victim = LRU.back();
+      auto AccCacheIterator = EntryLookup(victim);
+      auto & AccCache = AccCacheIterator->second;
+      Evict(AccCache);
+    }
+  }
+}
+uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
+{
+  ////////////////////////////////////////////////////////////////////////////
+  // 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;
+  
+  assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
+
+  assert(AccCache.cpuLock==0);  // Programming error
+
+  if(AccCache.state!=Empty) {
+    assert(AccCache.CpuPtr == CpuPtr);
+    assert(AccCache.bytes  ==bytes);
+  }
+/*
+ *  State transitions and actions
+ *
+ *  Action  State   StateNext         Flush    Clone
+ *
+ *  AccRead  Empty   Consistent        -        Y
+ *  AccWrite Empty   AccDirty          -        Y
+ *  AccRead  CpuDirty Consistent       -        Y
+ *  AccWrite CpuDirty AccDirty         -        Y
+ *  AccRead  Consistent Consistent     -        - 
+ *  AccWrite Consistent AccDirty       -        - 
+ *  AccRead  AccDirty   AccDirty       -        - 
+ *  AccWrite AccDirty   AccDirty       -        - 
+ */
+  if(AccCache.state==Empty) {
+    assert(AccCache.LRU_valid==0);
+    AccCache.CpuPtr = CpuPtr;
+    AccCache.AccPtr = (uint64_t)NULL;
+    AccCache.bytes  = bytes;
+    AccCache.state  = CpuDirty;   // Cpu starts primary
+    if(mode==AcceleratorWriteDiscard){
+      CpuDiscard(AccCache);
+      AccCache.state  = AccDirty;   // Empty + AcceleratorWrite=> AccDirty
+    } else if(mode==AcceleratorWrite){
+      Clone(AccCache);
+      AccCache.state  = AccDirty;   // Empty + AcceleratorWrite=> AccDirty
+    } else {
+      Clone(AccCache);
+      AccCache.state  = Consistent; // Empty + AccRead => Consistent
+    }
+    AccCache.accLock= 1;
+  } else if(AccCache.state==CpuDirty ){
+    if(mode==AcceleratorWriteDiscard) {
+      CpuDiscard(AccCache);
+      AccCache.state  = AccDirty;   // CpuDirty + AcceleratorWrite=> AccDirty
+    } else if(mode==AcceleratorWrite) {
+      Clone(AccCache);
+      AccCache.state  = AccDirty;   // CpuDirty + AcceleratorWrite=> AccDirty
+    } else {
+      Clone(AccCache);
+      AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
+    }
+    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++;
+    //    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++;
+    //    printf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+  } else {
+    assert(0);
+  }
+
+  // If view is opened on device remove from LRU
+  if(AccCache.LRU_valid==1){
+    // must possibly remove from LRU as now locked on GPU
+    LRUremove(AccCache);
+  }
+
+  int transient =hint;
+  AccCache.transient= transient? EvictNext : 0;
+
+  return AccCache.AccPtr;
+}
+////////////////////////////////////
+// look up & decrement lock count
+////////////////////////////////////
+void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
+{
+  auto AccCacheIterator = EntryLookup(CpuPtr);
+  auto & AccCache = AccCacheIterator->second;
+
+  assert(AccCache.cpuLock==0);
+  assert(AccCache.accLock>0);
+
+  AccCache.accLock--;
+
+  // Move to LRU queue if not locked and close on device
+  if(AccCache.accLock==0) {
+    LRUinsert(AccCache);
+  }
+}
+void MemoryManager::CpuViewClose(uint64_t CpuPtr)
+{
+  auto AccCacheIterator = EntryLookup(CpuPtr);
+  auto & AccCache = AccCacheIterator->second;
+
+  assert(AccCache.cpuLock>0);
+  assert(AccCache.accLock==0);
+
+  AccCache.cpuLock--;
+}
+/*
+ *  Action  State   StateNext         Flush    Clone
+ *
+ *  CpuRead  Empty   CpuDirty          -        -
+ *  CpuWrite Empty   CpuDirty          -        -
+ *  CpuRead  CpuDirty CpuDirty         -        -
+ *  CpuWrite CpuDirty CpuDirty         -        - 
+ *  CpuRead  Consistent Consistent     -        - 
+ *  CpuWrite Consistent CpuDirty       -        - 
+ *  CpuRead  AccDirty   Consistent     Y        -
+ *  CpuWrite AccDirty   CpuDirty       Y        -
+ */
+uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise transient)
+{
+  ////////////////////////////////////////////////////////////////////////////
+  // 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;
+  
+  assert((mode==CpuRead)||(mode==CpuWrite));
+  assert(AccCache.accLock==0);  // Programming error
+
+  if(AccCache.state!=Empty) {
+    assert(AccCache.CpuPtr == CpuPtr);
+    assert(AccCache.bytes==bytes);
+  }
+
+  if(AccCache.state==Empty) {
+    AccCache.CpuPtr = CpuPtr;
+    AccCache.AccPtr = (uint64_t)NULL;
+    AccCache.bytes  = bytes;
+    AccCache.state  = CpuDirty; // Empty + CpuRead/CpuWrite => CpuDirty
+    AccCache.accLock= 0;
+    AccCache.cpuLock= 1;
+  } else if(AccCache.state==CpuDirty ){
+    // AccPtr dont care, deferred allocate
+    AccCache.state = CpuDirty; // CpuDirty +CpuRead/CpuWrite => CpuDirty
+    AccCache.cpuLock++;
+  } else if(AccCache.state==Consistent) {
+    assert(AccCache.AccPtr != (uint64_t)NULL);
+    if(mode==CpuWrite)
+      AccCache.state = CpuDirty;   // Consistent +CpuWrite => CpuDirty
+    else 
+      AccCache.state = Consistent; // Consistent +CpuRead  => Consistent
+    AccCache.cpuLock++;
+  } else if(AccCache.state==AccDirty) {
+    assert(AccCache.AccPtr != (uint64_t)NULL);
+    Flush(AccCache);
+    if(mode==CpuWrite) AccCache.state = CpuDirty;   // AccDirty +CpuWrite => CpuDirty, Flush
+    else            AccCache.state = Consistent; // AccDirty +CpuRead  => Consistent, Flush
+    AccCache.cpuLock++;
+  } else {
+    assert(0); // should be unreachable
+  }
+
+  AccCache.transient= transient? EvictNext : 0;
+
+  return AccCache.CpuPtr;
+}
+void  MemoryManager::NotifyDeletion(void *_ptr)
+{
+  // Look up in ViewCache
+  uint64_t ptr = (uint64_t)_ptr;
+  if(EntryPresent(ptr)) {
+    auto e = EntryLookup(ptr);
+    AccDiscard(e->second);
+  }
+}
+void  MemoryManager::Print(void)
+{
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+	      << "\t" << AccCache.cpuLock
+	      << "\t" << AccCache.accLock
+	      << "\t" << AccCache.LRU_valid<<std::endl;
+  }
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+
+};
+int   MemoryManager::isOpen   (void* _CpuPtr) 
+{ 
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    return AccCache.cpuLock+AccCache.accLock;
+  } else { 
+    return 0;
+  }
+}
+
+NAMESPACE_END(Grid);
+
+#endif
diff --git a/Grid/allocator/MemoryManagerShared.cc b/Grid/allocator/MemoryManagerShared.cc
new file mode 100644
index 00000000..537f7c32
--- /dev/null
+++ b/Grid/allocator/MemoryManagerShared.cc
@@ -0,0 +1,24 @@
+#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
+/////////////////////////////////////////////////////////////////////////////////
+uint64_t  MemoryManager::DeviceBytes;
+uint64_t  MemoryManager::DeviceLRUBytes;
+uint64_t  MemoryManager::DeviceMaxBytes = 1024*1024*128;
+uint64_t  MemoryManager::HostToDeviceBytes;
+uint64_t  MemoryManager::DeviceToHostBytes;
+uint64_t  MemoryManager::HostToDeviceXfer;
+uint64_t  MemoryManager::DeviceToHostXfer;
+
+void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
+void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
+int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
+void  MemoryManager::Print(void){};
+void  MemoryManager::NotifyDeletion(void *ptr){};
+
+NAMESPACE_END(Grid);
+#endif
diff --git a/Grid/allocator/MemoryStats.cc b/Grid/allocator/MemoryStats.cc
new file mode 100644
index 00000000..0d1707d9
--- /dev/null
+++ b/Grid/allocator/MemoryStats.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);
+
diff --git a/Grid/allocator/MemoryStats.h b/Grid/allocator/MemoryStats.h
new file mode 100644
index 00000000..156c9747
--- /dev/null
+++ b/Grid/allocator/MemoryStats.h
@@ -0,0 +1,95 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/MemoryStats.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+
+NAMESPACE_BEGIN(Grid);
+
+std::string sizeString(size_t bytes);
+
+struct MemoryStats
+{
+  size_t totalAllocated{0}, maxAllocated{0}, 
+    currentlyAllocated{0}, totalFreed{0};
+};
+    
+class MemoryProfiler
+{
+public:
+  static MemoryStats *stats;
+  static bool        debug;
+};
+
+#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
+#define profilerDebugPrint						\
+  if (MemoryProfiler::stats)						\
+    {									\
+      auto s = MemoryProfiler::stats;					\
+      std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl; \
+      std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
+		<< std::endl;						\
+      std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
+		<< std::endl;						\
+      std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
+		<< std::endl;						\
+      std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
+		<< std::endl;						\
+    }
+
+#define profilerAllocate(bytes)						\
+  if (MemoryProfiler::stats)						\
+    {									\
+      auto s = MemoryProfiler::stats;					\
+      s->totalAllocated     += (bytes);					\
+      s->currentlyAllocated += (bytes);					\
+      s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated); \
+    }									\
+  if (MemoryProfiler::debug)						\
+    {									\
+      std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl; \
+      profilerDebugPrint;						\
+    }
+
+#define profilerFree(bytes)						\
+  if (MemoryProfiler::stats)						\
+    {									\
+      auto s = MemoryProfiler::stats;					\
+      s->totalFreed         += (bytes);					\
+      s->currentlyAllocated -= (bytes);					\
+    }									\
+  if (MemoryProfiler::debug)						\
+    {									\
+      std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl; \
+      profilerDebugPrint;						\
+    }
+
+void check_huge_pages(void *Buf,uint64_t BYTES);
+
+NAMESPACE_END(Grid);
+
diff --git a/Grid/cartesian/Cartesian_base.h b/Grid/cartesian/Cartesian_base.h
index 87472cc9..ae1fd1fd 100644
--- a/Grid/cartesian/Cartesian_base.h
+++ b/Grid/cartesian/Cartesian_base.h
@@ -81,6 +81,7 @@ public:
 
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
+  Coordinate _checker_dim_mask;
 
 public:
 
diff --git a/Grid/cartesian/Cartesian_full.h b/Grid/cartesian/Cartesian_full.h
index c083817b..31a67bf0 100644
--- a/Grid/cartesian/Cartesian_full.h
+++ b/Grid/cartesian/Cartesian_full.h
@@ -38,6 +38,7 @@ class GridCartesian: public GridBase {
 
 public:
   int dummy;
+  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -104,6 +105,7 @@ public:
     _ldimensions.resize(_ndimension);
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
+    _checker_dim_mask.resize(_ndimension);;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -114,6 +116,8 @@ public:
 
     for (int d = 0; d < _ndimension; d++)
       {
+	_checker_dim_mask[d]=0;
+
         _fdimensions[d] = dimensions[d];   // Global dimensions
         _gdimensions[d] = _fdimensions[d]; // Global dimensions
         _simd_layout[d] = simd_layout[d];
diff --git a/Grid/cartesian/Cartesian_red_black.h b/Grid/cartesian/Cartesian_red_black.h
index 34f763d2..b71981f5 100644
--- a/Grid/cartesian/Cartesian_red_black.h
+++ b/Grid/cartesian/Cartesian_red_black.h
@@ -35,12 +35,28 @@ static const int CbRed  =0;
 static const int CbBlack=1;
 static const int Even   =CbRed;
 static const int Odd    =CbBlack;
+
+accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex, Coordinate &rdim, Coordinate &chk_dim_msk)
+{
+  int nd=rdim.size();
+  Coordinate coor(nd);
+
+  Lexicographic::CoorFromIndex(coor,oindex,rdim);
+
+  int linear=0;
+  for(int d=0;d<nd;d++){
+    if(chk_dim_msk[d])
+      linear=linear+coor[d];
+  }
+  return (linear&0x1);
+}
+
     
 // Specialise this for red black grids storing half the data like a chess board.
 class GridRedBlackCartesian : public GridBase
 {
 public:
-  Coordinate _checker_dim_mask;
+  //  Coordinate _checker_dim_mask;
   int              _checker_dim;
   std::vector<int> _checker_board;
 
diff --git a/Grid/communicator/Communicator_mpi3.cc b/Grid/communicator/Communicator_mpi3.cc
index 0e525674..6130195d 100644
--- a/Grid/communicator/Communicator_mpi3.cc
+++ b/Grid/communicator/Communicator_mpi3.cc
@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/communicator/Communicator_mpi.cc
 
@@ -35,7 +35,7 @@ Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 ////////////////////////////////////////////
 // First initialise of comms system
 ////////////////////////////////////////////
-void CartesianCommunicator::Init(int *argc, char ***argv) 
+void CartesianCommunicator::Init(int *argc, char ***argv)
 {
 
   int flag;
@@ -43,8 +43,16 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
 
+#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
+    // other comms schemes are ok
+    MPI_Init_thread(argc,argv,MPI_THREAD_SERIALIZED,&provided);
+#else
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+#endif
     //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
     if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
       assert(0);
@@ -91,7 +99,7 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Initialises from communicator_world
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   MPI_Comm optimal_comm;
   ////////////////////////////////////////////////////
@@ -110,7 +118,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)    
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
 {
   _ndimension = processors.size();  assert(_ndimension>=1);
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
@@ -127,7 +135,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  //  int Nparent = parent._processors ; 
+  //  int Nparent = parent._processors ;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
 
@@ -149,13 +157,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
-  int crank;  
+  int crank;
   // Mpi uses the reverse Lexico convention to us; so reversed routines called
   Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
   Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
 
   MPI_Comm comm_split;
-  if ( Nchild > 1 ) { 
+  if ( Nchild > 1 ) {
 
     ////////////////////////////////////////////////////////////////
     // Split the communicator
@@ -180,11 +188,11 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   SetCommunicator(comm_split);
 
   ///////////////////////////////////////////////
-  // Free the temp communicator 
+  // Free the temp communicator
   ///////////////////////////////////////////////
   MPI_Comm_free(&comm_split);
 
-  if(0){ 
+  if(0){
     std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
     for(int d=0;d<processors.size();d++){
       std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
@@ -245,7 +253,7 @@ CartesianCommunicator::~CartesianCommunicator()
     for(int i=0;i<communicator_halo.size();i++){
       MPI_Comm_free(&communicator_halo[i]);
     }
-  }  
+  }
 }
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
@@ -314,7 +322,7 @@ void CartesianCommunicator::SendRecvPacket(void *xmit,
   if ( _processor == sender ) {
     MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
   }
-  if ( _processor == receiver ) { 
+  if ( _processor == receiver ) {
     MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
   }
 }
@@ -329,17 +337,17 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
   int myrank = _processor;
   int ierr;
 
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
     MPI_Request xrq;
     MPI_Request rrq;
 
     ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
     ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-    
+
     assert(ierr==0);
     list.push_back(xrq);
     list.push_back(rrq);
-  } else { 
+  } else {
     // Give the CPU to MPI immediately; can use threads to overlap optionally
     ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
 		      recv,bytes,MPI_CHAR,from, from,
@@ -367,7 +375,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int from,
 							 int bytes,int dir)
 {
-  int ncomm  =communicator_halo.size(); 
+  int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
 
   MPI_Request xrq;
@@ -397,7 +405,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
     off_node_bytes+=bytes;
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);
   }
 
@@ -436,8 +444,8 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-int CartesianCommunicator::RankWorld(void){ 
-  int r; 
+int CartesianCommunicator::RankWorld(void){
+  int r;
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
@@ -470,7 +478,7 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
   // (Turns up on 32^3 x 64 Gparity too)
   MPI_Datatype object;
-  int iwords; 
+  int iwords;
   int ibytes;
   iwords = words;
   ibytes = bytes;
@@ -483,5 +491,3 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 }
 
 NAMESPACE_END(Grid);
-
-
diff --git a/Grid/communicator/SharedMemoryMPI.cc b/Grid/communicator/SharedMemoryMPI.cc
index bd83ce52..45fefc71 100644
--- a/Grid/communicator/SharedMemoryMPI.cc
+++ b/Grid/communicator/SharedMemoryMPI.cc
@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
 #endif
 
@@ -420,7 +420,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -440,13 +440,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   //  cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
 
-#ifdef GRID_IBM_SUMMIT
-  // IBM Jsrun makes cuda Device numbering screwy and not match rank
-    std::cout << "IBM Summit or similar - NOT setting device to WorldShmRank"<<std::endl;
-#else
-    std::cout << "setting device to WorldShmRank"<<std::endl;
-    cudaSetDevice(WorldShmRank);
-#endif
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -684,7 +677,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 /////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
   cudaMemset(dest,0,bytes);
 #else
   bzero(dest,bytes);
@@ -692,7 +685,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 }
 void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
 {
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
   cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
 #else   
   bcopy(src,dest,bytes);
diff --git a/Grid/cshift/Cshift_common.h b/Grid/cshift/Cshift_common.h
index 954342cb..4de2bbe2 100644
--- a/Grid/cshift/Cshift_common.h
+++ b/Grid/cshift/Cshift_common.h
@@ -29,6 +29,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+extern Vector<std::pair<int,int> > Cshift_table; 
+
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
@@ -46,16 +48,16 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
   int e2=rhs.Grid()->_slice_block[dimension];
   int ent = 0;
 
-  static Vector<std::pair<int,int> > table; table.resize(e1*e2);
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
+
   int stride=rhs.Grid()->_slice_stride[dimension];
 
-  auto rhs_v = rhs.View();
   if ( cbmask == 0x3 ) { 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o  = n*stride;
 	int bo = n*e2;
-	table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
+	Cshift_table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
       }
     }
   } else { 
@@ -65,14 +67,19 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
 	 int o  = n*stride;
 	 int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	 if ( ocb &cbmask ) {
-	   table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
+	   Cshift_table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
 	 }
        }
      }
   }
-  thread_for(i,ent,{
-    buffer[table[i].first]=rhs_v[table[i].second];
-  });
+  {
+    autoView(rhs_v , rhs, AcceleratorRead);
+    auto buffer_p = & buffer[0];
+    auto table = &Cshift_table[0];
+    accelerator_for(i,ent,1,{
+      buffer_p[table[i].first]=rhs_v[table[i].second];
+    });
+  }
 }
 
 ///////////////////////////////////////////////////////////////////
@@ -95,36 +102,38 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int e2=rhs.Grid()->_slice_block[dimension];
   int n1=rhs.Grid()->_slice_stride[dimension];
 
-  auto rhs_v = rhs.View();
   if ( cbmask ==0x3){
-    thread_for_collapse(2,n,e1,{
-      for(int b=0;b<e2;b++){
-
+    autoView(rhs_v , rhs, AcceleratorRead);
+    accelerator_for2d(n,e1,b,e2,1,{
 	int o      =   n*n1;
 	int offset = b+n*e2;
 	
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
-      }
-    });
+      });
   } else { 
+    autoView(rhs_v , rhs, AcceleratorRead);
 
-    // Case of SIMD split AND checker dim cannot currently be hit, except in 
-    // Test_cshift_red_black code.
-    std::cout << " Dense packed buffer WARNING " <<std::endl;
-    thread_for_collapse(2,n,e1,{
-      for(int b=0;b<e2;b++){
+    Coordinate rdim=rhs.Grid()->_rdimensions;
+    Coordinate cdm =rhs.Grid()->_checker_dim_mask;
+    std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
+    accelerator_for2d(n,e1,b,e2,1,{
+
+	Coordinate coor;
 
 	int o=n*n1;
-	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
+	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);
 	}
-      }
-    });
+      });
   }
 }
 
@@ -145,7 +154,8 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
   int e2=rhs.Grid()->_slice_block[dimension];
   int stride=rhs.Grid()->_slice_stride[dimension];
 
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
+
   int ent    =0;
 
   if ( cbmask ==0x3 ) {
@@ -154,7 +164,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
       for(int b=0;b<e2;b++){
 	int o   =n*rhs.Grid()->_slice_stride[dimension];
 	int bo  =n*rhs.Grid()->_slice_block[dimension];
-	table[ent++] = std::pair<int,int>(so+o+b,bo+b);
+	Cshift_table[ent++] = std::pair<int,int>(so+o+b,bo+b);
       }
     }
 
@@ -165,16 +175,20 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
 	int o   =n*rhs.Grid()->_slice_stride[dimension];
 	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	if ( ocb & cbmask ) {
-	  table[ent++]=std::pair<int,int> (so+o+b,bo++);
+	  Cshift_table[ent++]=std::pair<int,int> (so+o+b,bo++);
 	}
       }
     }
   }
   
-  auto rhs_v = rhs.View();
-  thread_for(i,ent,{
-    rhs_v[table[i].first]=buffer[table[i].second];
-  });
+  {
+    autoView( rhs_v, rhs, AcceleratorWrite);
+    auto buffer_p = & buffer[0];
+    auto table = &Cshift_table[0];
+    accelerator_for(i,ent,1,{
+	rhs_v[table[i].first]=buffer_p[table[i].second];
+    });
+  }
 }
 
 //////////////////////////////////////////////////////
@@ -194,21 +208,19 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
-    auto rhs_v = rhs.View();
-    thread_for_collapse(2,n,e1,{
-      for(int b=0;b<e2;b++){
+    autoView( rhs_v , rhs, AcceleratorWrite);
+    accelerator_for2d(n,e1,b,e2,1,{
 	int o      = n*rhs.Grid()->_slice_stride[dimension];
 	int offset = b+n*rhs.Grid()->_slice_block[dimension];
 	merge(rhs_v[so+o+b],pointers,offset);
-      }
-    });
+      });
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
     // Test_cshift_red_black code.
     //    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
-    auto rhs_v = rhs.View();
+    autoView( rhs_v, rhs, CpuWrite);
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o      = n*rhs.Grid()->_slice_stride[dimension];
@@ -225,6 +237,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 //////////////////////////////////////////////////////
 // local to node block strided copies
 //////////////////////////////////////////////////////
+
 template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
@@ -239,14 +252,16 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int e1=rhs.Grid()->_slice_nblock[dimension]; // clearly loop invariant for icpc
   int e2=rhs.Grid()->_slice_block[dimension];
   int stride = rhs.Grid()->_slice_stride[dimension];
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
+
   int ent=0;
 
   if(cbmask == 0x3 ){
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int o =n*stride+b;
-	table[ent++] = std::pair<int,int>(lo+o,ro+o);
+	Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
       }
     }
   } else { 
@@ -255,23 +270,24 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
         int o =n*stride+b;
         int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o);
         if ( ocb&cbmask ) {
-	  table[ent++] = std::pair<int,int>(lo+o,ro+o);
+	  Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
 	}
       }
     }
   }
 
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for(i,ent,{
-    lhs_v[table[i].first]=rhs_v[table[i].second];
-  });
-
+  {
+    autoView(rhs_v , rhs, AcceleratorRead);
+    autoView(lhs_v , lhs, AcceleratorWrite);
+    auto table = &Cshift_table[0];
+    accelerator_for(i,ent,1,{
+      lhs_v[table[i].first]=rhs_v[table[i].second];
+    });
+  }
 }
 
 template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask,int permute_type)
 {
- 
   int rd = rhs.Grid()->_rdimensions[dimension];
 
   if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
@@ -285,29 +301,33 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int e2=rhs.Grid()->_slice_block [dimension];
   int stride = rhs.Grid()->_slice_stride[dimension];
 
-  static std::vector<std::pair<int,int> > table;  table.resize(e1*e2);
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
+
   int ent=0;
 
   if ( cbmask == 0x3 ) {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
-      table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+      Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
   } else {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
       int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o+b);
-      if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+      if ( ocb&cbmask ) Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
   }
 
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for(i,ent,{
-    permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-  });
+  {
+    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);
+    });
+  }
 }
 
 //////////////////////////////////////////////////////
diff --git a/Grid/cshift/Cshift_table.cc b/Grid/cshift/Cshift_table.cc
new file mode 100644
index 00000000..d46e51c0
--- /dev/null
+++ b/Grid/cshift/Cshift_table.cc
@@ -0,0 +1,4 @@
+#include <Grid/GridCore.h>       
+NAMESPACE_BEGIN(Grid);
+Vector<std::pair<int,int> > Cshift_table; 
+NAMESPACE_END(Grid);
diff --git a/Grid/lattice/Lattice.h b/Grid/lattice/Lattice.h
index 036633b4..a3017198 100644
--- a/Grid/lattice/Lattice.h
+++ b/Grid/lattice/Lattice.h
@@ -26,6 +26,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     *************************************************************************************/
     /*  END LEGAL */
 #pragma once
+#include <Grid/lattice/Lattice_view.h>
 #include <Grid/lattice/Lattice_base.h>
 #include <Grid/lattice/Lattice_conformable.h>
 #include <Grid/lattice/Lattice_ET.h>
diff --git a/Grid/lattice/Lattice_ET.h b/Grid/lattice/Lattice_ET.h
index da63d5e6..91b456d9 100644
--- a/Grid/lattice/Lattice_ET.h
+++ b/Grid/lattice/Lattice_ET.h
@@ -92,12 +92,18 @@ const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg)
 {
   return arg[ss];
 }
+
+// What needs this?
+// Cannot be legal on accelerator
+// Comparison must convert
+#if 1
 template <class lobj> accelerator_inline 
 const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg) 
 {
-  auto view = arg.AcceleratorView(ViewRead);
+  auto view = arg.View(AcceleratorRead);
   return view[ss];
 }
+#endif
 
 ///////////////////////////////////////////////////
 // handle nodes in syntax tree- eval one operand
@@ -180,16 +186,12 @@ inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
   cb = lat.Checkerboard();
 }
 template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-inline void CBFromExpression(int &cb, const T1 &notlat)  // non-lattice leaf
-{
-}
-
+inline void CBFromExpression(int &cb, const T1 &notlat) {} // non-lattice leaf
 template <typename Op, typename T1> inline 
 void CBFromExpression(int &cb,const LatticeUnaryExpression<Op, T1> &expr) 
 {
   CBFromExpression(cb, expr.arg1);  // recurse AST
 }
-
 template <typename Op, typename T1, typename T2> inline 
 void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr) 
 {
@@ -204,6 +206,68 @@ inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2,
   CBFromExpression(cb, expr.arg3);  // recurse AST
 }
 
+
+//////////////////////////////////////////////////////////////////////////
+// ViewOpen
+//////////////////////////////////////////////////////////////////////////
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+inline void ExpressionViewOpen(T1 &lat)  // Lattice leaf
+{
+  lat.ViewOpen(AcceleratorRead);
+}
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+  inline void ExpressionViewOpen(T1 &notlat) {}
+
+template <typename Op, typename T1> inline 
+void ExpressionViewOpen(LatticeUnaryExpression<Op, T1> &expr) 
+{  
+  ExpressionViewOpen(expr.arg1); // recurse AST
+}
+
+template <typename Op, typename T1, typename T2> inline 
+void ExpressionViewOpen(LatticeBinaryExpression<Op, T1, T2> &expr) 
+{
+  ExpressionViewOpen(expr.arg1);  // recurse AST
+  ExpressionViewOpen(expr.arg2);  // recurse AST
+}
+template <typename Op, typename T1, typename T2, typename T3>
+inline void ExpressionViewOpen(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
+{
+  ExpressionViewOpen(expr.arg1);  // recurse AST
+  ExpressionViewOpen(expr.arg2);  // recurse AST
+  ExpressionViewOpen(expr.arg3);  // recurse AST
+}
+
+//////////////////////////////////////////////////////////////////////////
+// ViewClose
+//////////////////////////////////////////////////////////////////////////
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+inline void ExpressionViewClose( T1 &lat)  // Lattice leaf
+{
+  lat.ViewClose();
+}
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+inline void ExpressionViewClose(T1 &notlat) {}
+
+template <typename Op, typename T1> inline 
+void ExpressionViewClose(LatticeUnaryExpression<Op, T1> &expr) 
+{  
+  ExpressionViewClose(expr.arg1); // recurse AST
+}
+template <typename Op, typename T1, typename T2> inline 
+void ExpressionViewClose(LatticeBinaryExpression<Op, T1, T2> &expr) 
+{
+  ExpressionViewClose(expr.arg1);  // recurse AST
+  ExpressionViewClose(expr.arg2);  // recurse AST
+}
+template <typename Op, typename T1, typename T2, typename T3>
+inline void ExpressionViewClose(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
+{
+  ExpressionViewClose(expr.arg1);  // recurse AST
+  ExpressionViewClose(expr.arg2);  // recurse AST
+  ExpressionViewClose(expr.arg3);  // recurse AST
+}
+
 ////////////////////////////////////////////
 // Unary operators and funcs
 ////////////////////////////////////////////
diff --git a/Grid/lattice/Lattice_arith.h b/Grid/lattice/Lattice_arith.h
index c4a67620..a3ae1f28 100644
--- a/Grid/lattice/Lattice_arith.h
+++ b/Grid/lattice/Lattice_arith.h
@@ -37,9 +37,9 @@ NAMESPACE_BEGIN(Grid);
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   conformable(ret,rhs);
   conformable(lhs,rhs);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
@@ -56,9 +56,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  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);
@@ -73,9 +73,9 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  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);
@@ -89,9 +89,9 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  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);
@@ -108,8 +108,8 @@ template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     mult(&tmp,&lhs_v(ss),&rhs);
@@ -121,8 +121,8 @@ template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -135,8 +135,8 @@ template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -148,8 +148,8 @@ template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -165,8 +165,8 @@ template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -179,8 +179,8 @@ template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -193,8 +193,8 @@ template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -206,8 +206,8 @@ template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -221,9 +221,9 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto x_v = x.AcceleratorView(ViewRead);
-  auto y_v = y.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( x_v , x, AcceleratorRead);
+  autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
     auto tmp = a*x_v(ss)+y_v(ss);
     coalescedWrite(ret_v[ss],tmp);
@@ -234,9 +234,9 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto x_v = x.AcceleratorView(ViewRead);
-  auto y_v = y.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( x_v , x, AcceleratorRead);
+  autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
     auto tmp = a*x_v(ss)+b*y_v(ss);
     coalescedWrite(ret_v[ss],tmp);
diff --git a/Grid/lattice/Lattice_base.h b/Grid/lattice/Lattice_base.h
index 42e9e50a..73b1b6a1 100644
--- a/Grid/lattice/Lattice_base.h
+++ b/Grid/lattice/Lattice_base.h
@@ -29,6 +29,7 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
+
 #pragma once 
 
 #define STREAMING_STORES
@@ -37,161 +38,6 @@ NAMESPACE_BEGIN(Grid);
 
 extern int GridCshiftPermuteMap[4][16];
 
-///////////////////////////////////////////////////////////////////
-// Base class which can be used by traits to pick up behaviour
-///////////////////////////////////////////////////////////////////
-class LatticeBase {};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Conformable checks; same instance of Grid required
-/////////////////////////////////////////////////////////////////////////////////////////
-void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
-{
-  assert(lhs == rhs);
-}
-
-////////////////////////////////////////////////////////////////////////////
-// Advise the LatticeAccelerator class
-////////////////////////////////////////////////////////////////////////////
-enum LatticeAcceleratorAdvise {
-  AdviseInfrequentUse = 0x1,    // Advise that the data is used infrequently.  This can
-                                // significantly influence performance of bulk storage.
-  AdviseReadMostly = 0x2,       // Data will mostly be read.  On some architectures
-                                // enables read-only copies of memory to be kept on
-                                // host and device.
-};
-
-////////////////////////////////////////////////////////////////////////////
-// View Access Mode
-////////////////////////////////////////////////////////////////////////////
-enum ViewMode {
-  ViewRead = 0x1,
-  ViewWrite = 0x2,
-  ViewReadWrite = 0x3
-};
-
-////////////////////////////////////////////////////////////////////////////
-// Minimal base class containing only data valid to access from accelerator
-// _odata will be a managed pointer in CUDA
-////////////////////////////////////////////////////////////////////////////
-// Force access to lattice through a view object.
-// prevents writing of code that will not offload to GPU, but perhaps annoyingly
-// strict since host could could in principle direct access through the lattice object
-// Need to decide programming model.
-#define LATTICE_VIEW_STRICT
-template<class vobj> class LatticeAccelerator : public LatticeBase
-{
-protected:
-  GridBase *_grid;
-  int checkerboard;
-  vobj     *_odata;    // A managed pointer
-  uint64_t _odata_size;    
-public:
-  accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr) { }; 
-  accelerator_inline uint64_t oSites(void) const { return _odata_size; };
-  accelerator_inline int  Checkerboard(void) const { return checkerboard; };
-  accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
-  accelerator_inline void Conformable(GridBase * &grid) const
-  { 
-    if (grid) conformable(grid, _grid);
-    else      grid = _grid;
-  };
-
-  accelerator_inline void AcceleratorPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
-    gridAcceleratorPrefetch(_odata,_odata_size*sizeof(vobj));
-  };
-
-  accelerator_inline void HostPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    //cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),cudaCpuDeviceId);
-#endif
-#endif
-  };
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// A View class which provides accessor to the data.
-// 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
-/////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> 
-class LatticeView : public LatticeAccelerator<vobj>
-{
-public:
-
-
-  // Rvalue
-#ifdef __CUDA_ARCH__
-  accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { return coalescedRead(this->_odata[i]); }
-#else 
-  accelerator_inline const vobj & operator()(size_t i) const { 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 uint64_t begin(void) const { return 0;};
-  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
-  accelerator_inline uint64_t size(void)  const { return this->_odata_size; };
-
-  LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me)
-  {
-  }
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Lattice expression types used by ET to assemble the AST
-// 
-// Need to be able to detect code paths according to the whether a lattice object or not
-// so introduce some trait type things
-/////////////////////////////////////////////////////////////////////////////////////////
-
-class LatticeExpressionBase {};
-
-template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
-template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
-
-template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
-template<class T>                 struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
-template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
-
-template <typename Op, typename _T1>                           
-class LatticeUnaryExpression : public  LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  Op op;
-  T1 arg1;
-  LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
-};
-
-template <typename Op, typename _T1, typename _T2>              
-class LatticeBinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
-};
-
-template <typename Op, typename _T1, typename _T2, typename _T3> 
-class LatticeTrinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  typedef typename ViewMap<_T3>::Type T3;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  T3 arg3;
-  LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
-};
-
 /////////////////////////////////////////////////////////////////////////////////////////
 // The real lattice class, with normal copy and assignment semantics.
 // This contains extra (host resident) grid pointer data that may be accessed by host code
@@ -235,44 +81,25 @@ private:
   }
 public:
 
-  void Advise(int advise) {
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    if (advise & AdviseInfrequentUse) {
-      gridMoveToHost((void**)&this->_odata);
-    }
-    if (advise & AdviseReadMostly) {
-      //cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetReadMostly,-1);
-    }
-#endif
-#endif
-  };
-
+  /////////////////////////////////////////////////////////////////////////////////
+  // Can use to make accelerator dirty without copy from host ; useful for temporaries "dont care" prev contents
+  /////////////////////////////////////////////////////////////////////////////////
+  void SetViewMode(ViewMode mode) {
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
+    accessor.ViewClose();
+  }
   /////////////////////////////////////////////////////////////////////////////////
   // Return a view object that may be dereferenced in site loops.
   // The view is trivially copy constructible and may be copied to an accelerator device
   // in device lambdas
   /////////////////////////////////////////////////////////////////////////////////
-  LatticeView<vobj> View (void) const // deprecated, should pick AcceleratorView for accelerator_for
-  {                                   //                     and HostView        for thread_for
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
+
+  LatticeView<vobj> View (ViewMode mode) const 
+  {
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
     return accessor;
   }
 
-  LatticeView<vobj> AcceleratorView(int mode = ViewReadWrite) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
-    //accessor.AcceleratorPrefetch(mode);
-    return accessor;
-  }
-
-  LatticeView<vobj> HostView(int mode = ViewReadWrite) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
-    //accessor.HostPrefetch(mode);
-    return accessor;
-  }
-  
   ~Lattice() { 
     if ( this->_odata_size ) {
       dealloc();
@@ -292,12 +119,16 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-
-    auto me  = AcceleratorView(ViewWrite);
+    
+    auto exprCopy = expr;
+    ExpressionViewOpen(exprCopy);
+    auto me  = View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
+      auto tmp = eval(ss,exprCopy);
       vstream(me[ss],tmp);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
@@ -312,11 +143,15 @@ public:
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
-    auto me  = AcceleratorView(ViewWrite);
+    auto exprCopy = expr;
+    ExpressionViewOpen(exprCopy);
+    auto me  = View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
+      auto tmp = eval(ss,exprCopy);
       vstream(me[ss],tmp);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
@@ -330,11 +165,15 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-    auto me  = AcceleratorView(ViewWrite);
+    auto exprCopy = expr;
+    ExpressionViewOpen(exprCopy);
+    auto me  = View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
+      auto tmp = eval(ss,exprCopy);
       vstream(me[ss],tmp);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   //GridFromExpression is tricky to do
@@ -385,10 +224,11 @@ public:
   }
 
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
-    auto me  = View();
+    auto me  = View(CpuWrite);
     thread_for(ss,me.size(),{
-      me[ss] = r;
+	me[ss]= r;
     });
+    me.ViewClose();
     return *this;
   }
 
@@ -398,11 +238,12 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid) { 
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
     this->checkerboard=0;
+    SetViewMode(mode);
   }
   
   //  virtual ~Lattice(void) = default;
@@ -440,11 +281,12 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   AcceleratorView(ViewWrite);
-    auto him= r.AcceleratorView(ViewRead);
+    auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
+    me.ViewClose();    him.ViewClose();
     return *this;
   }
 
@@ -454,11 +296,12 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   AcceleratorView(ViewWrite);
-    auto him= r.AcceleratorView(ViewRead);
+    auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
+    me.ViewClose();    him.ViewClose();
     return *this;
   }
   ///////////////////////////////////////////
diff --git a/Grid/lattice/Lattice_basis.h b/Grid/lattice/Lattice_basis.h
index f1126936..b930e018 100644
--- a/Grid/lattice/Lattice_basis.h
+++ b/Grid/lattice/Lattice_basis.h
@@ -51,34 +51,39 @@ template<class VField, class Matrix>
 void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm) 
 {
   typedef decltype(basis[0]) Field;
-  typedef decltype(basis[0].View()) View;
-  auto tmp_v = basis[0].AcceleratorView(ViewReadWrite);
-  Vector<View> basis_v(basis.size(),tmp_v);
-  typedef typename std::remove_reference<decltype(tmp_v[0])>::type vobj;
+  typedef decltype(basis[0].View(AcceleratorRead)) View;
+
+  Vector<View> basis_v; basis_v.reserve(basis.size());
+  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
+  typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
   GridBase* grid = basis[0].Grid();
       
   for(int k=0;k<basis.size();k++){
-    basis_v[k] = basis[k].AcceleratorView(ViewReadWrite);
+    basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
-#ifndef GRID_NVCC
+#ifdef GRID_OMP
+  int max_threads = thread_max();
+  Vector < vobj > Bt(Nm * max_threads);
   thread_region
-  {
-    std::vector < vobj > B(Nm); // Thread private
-    thread_for_in_region(ss, grid->oSites(),{
-	for(int j=j0; j<j1; ++j) B[j]=0.;
+    {
+      vobj* B = &Bt[Nm * thread_num()];
+      thread_for_in_region(ss, grid->oSites(),{
+	  for(int j=j0; j<j1; ++j) B[j]=0.;
       
-	for(int j=j0; j<j1; ++j){
-	  for(int k=k0; k<k1; ++k){
-	    B[j] +=Qt(j,k) * basis_v[k][ss];
+	  for(int j=j0; j<j1; ++j){
+	    for(int k=k0; k<k1; ++k){
+	      B[j] +=Qt(j,k) * basis_v[k][ss];
+	    }
 	  }
-	}
-	for(int j=j0; j<j1; ++j){
-	  basis_v[j][ss] = B[j];
-	}
-      });
-  }
+	  for(int j=j0; j<j1; ++j){
+	    basis_v[j][ss] = B[j];
+	  }
+	});
+    }
 #else
+  View *basis_vp = &basis_v[0];
+
   int nrot = j1-j0;
   if (!nrot) // edge case not handled gracefully by Cuda
     return;
@@ -90,13 +95,13 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  Vector<double> Qt_jv(Nm*Nm);
-  double *Qt_p = & Qt_jv[0];
+  Vector<Coeff_t> Qt_jv(Nm*Nm);
+  Coeff_t *Qt_p = & Qt_jv[0];
   thread_for(i,Nm*Nm,{
       int j = i/Nm;
       int k = i%Nm;
       Qt_p[i]=Qt(j,k);
-    });
+  });
 
   // Block the loop to keep storage footprint down
   for(uint64_t s=0;s<oSites;s+=siteBlock){
@@ -133,26 +138,30 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
       });
   }
 #endif
+
+  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
 }
 
 // Extract a single rotated vector
 template<class Field>
 void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
 {
-  typedef decltype(basis[0].AcceleratorView()) View;
+  typedef decltype(basis[0].View(AcceleratorRead)) View;
   typedef typename Field::vector_object vobj;
   GridBase* grid = basis[0].Grid();
 
   result.Checkerboard() = basis[0].Checkerboard();
-  auto result_v=result.AcceleratorView(ViewWrite);
-  Vector<View> basis_v(basis.size(),result_v);
+
+  Vector<View> basis_v; basis_v.reserve(basis.size());
   for(int k=0;k<basis.size();k++){
-    basis_v[k] = basis[k].AcceleratorView(ViewRead);
+    basis_v.push_back(basis[k].View(AcceleratorRead));
   }
   vobj zz=Zero();
   Vector<double> Qt_jv(Nm);
   double * Qt_j = & Qt_jv[0];
   for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
+
+  autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
     auto B=coalescedRead(zz);
     for(int k=k0; k<k1; ++k){
@@ -160,6 +169,7 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
     }
     coalescedWrite(result_v[ss], B);
   });
+  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
 }
 
 template<class Field>
diff --git a/Grid/lattice/Lattice_comparison.h b/Grid/lattice/Lattice_comparison.h
index bbed2ef5..6a29be94 100644
--- a/Grid/lattice/Lattice_comparison.h
+++ b/Grid/lattice/Lattice_comparison.h
@@ -78,9 +78,9 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
 {
   Lattice<vPredicate> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
       ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
   });
@@ -93,8 +93,8 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
 {
   Lattice<vPredicate> ret(lhs.Grid());
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, lhs_v.size(), {
     ret_v[ss]=op(lhs_v[ss],rhs);
   });
@@ -107,8 +107,8 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
 {
   Lattice<vPredicate> ret(rhs.Grid());
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( rhs_v, rhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
     ret_v[ss]=op(lhs,rhs_v[ss]);
   });
diff --git a/Grid/lattice/Lattice_coordinate.h b/Grid/lattice/Lattice_coordinate.h
index a1abe58d..cd0f11ee 100644
--- a/Grid/lattice/Lattice_coordinate.h
+++ b/Grid/lattice/Lattice_coordinate.h
@@ -37,7 +37,7 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
   GridBase *grid = l.Grid();
   int Nsimd = grid->iSites();
 
-  auto l_v = l.View();
+  autoView(l_v, l, CpuWrite);
   thread_for( o, grid->oSites(), {
     vector_type vI;
     Coordinate gcoor;
@@ -51,23 +51,5 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
   });
 };
 
-// LatticeCoordinate();
-// FIXME for debug; deprecate this; made obscelete by 
-template<class vobj> void lex_sites(Lattice<vobj> &l){
-  auto l_v = l.View();
-  Real *v_ptr = (Real *)&l_v[0];
-  size_t o_len = l.Grid()->oSites();
-  size_t v_len = sizeof(vobj)/sizeof(vRealF);
-  size_t vec_len = vRealF::Nsimd();
-
-  for(int i=0;i<o_len;i++){
-    for(int j=0;j<v_len;j++){
-      for(int vv=0;vv<vec_len;vv+=2){
-	v_ptr[i*v_len*vec_len+j*vec_len+vv  ]= i+vv*500;
-	v_ptr[i*v_len*vec_len+j*vec_len+vv+1]= i+vv*500;
-      }
-    }}
-}
-
 NAMESPACE_END(Grid);
 
diff --git a/Grid/lattice/Lattice_local.h b/Grid/lattice/Lattice_local.h
index 99455e24..1b31e9b3 100644
--- a/Grid/lattice/Lattice_local.h
+++ b/Grid/lattice/Lattice_local.h
@@ -43,8 +43,8 @@ template<class vobj>
 inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
 {
   Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
   });
@@ -56,9 +56,9 @@ template<class vobj>
 inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
 {
   Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
   });
@@ -73,9 +73,9 @@ inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Latt
   typedef decltype(coalescedRead(ll())) sll;
   typedef decltype(coalescedRead(rr())) srr;
   Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),1,{
     // FIXME had issues with scalar version of outer 
     // Use vector [] operator and don't read coalesce this loop
diff --git a/Grid/lattice/Lattice_matrix_reduction.h b/Grid/lattice/Lattice_matrix_reduction.h
index 0980ad8a..7c470fef 100644
--- a/Grid/lattice/Lattice_matrix_reduction.h
+++ b/Grid/lattice/Lattice_matrix_reduction.h
@@ -51,9 +51,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int block =FullGrid->_slice_block [Orthog];
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
-  auto X_v = X.View();
-  auto Y_v = Y.View();
-  auto R_v = R.View();
+  autoView( X_v , X, CpuRead);
+  autoView( Y_v , Y, CpuRead);
+  autoView( R_v , R, CpuWrite);
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -97,8 +97,8 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  auto X_v = X.View();
-  auto R_v = R.View();
+  autoView( X_v , X, CpuRead);
+  autoView( R_v , R, CpuWrite);
 
   thread_region
   {
@@ -156,8 +156,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   int ostride=FullGrid->_ostride[Orthog];
 
   typedef typename vobj::vector_typeD vector_typeD;
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( lhs_v , lhs, CpuRead);
+  autoView( rhs_v , rhs, CpuRead);
   thread_region {
     std::vector<vobj> Left(Nblock);
     std::vector<vobj> Right(Nblock);
diff --git a/Grid/lattice/Lattice_peekpoke.h b/Grid/lattice/Lattice_peekpoke.h
index 8f649bd7..c79becf2 100644
--- a/Grid/lattice/Lattice_peekpoke.h
+++ b/Grid/lattice/Lattice_peekpoke.h
@@ -46,9 +46,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Ind
 {
   Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
   ret.Checkerboard()=lhs.Checkerboard();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
+  accelerator_for( ss, lhs_v.size(), 1, {
     ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
   });
   return ret;
@@ -58,9 +58,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
 {
   Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
   ret.Checkerboard()=lhs.Checkerboard();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
+  accelerator_for( ss, lhs_v.size(), 1, {
     ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
   });
   return ret;
@@ -72,18 +72,18 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
 template<int Index,class vobj>  
 void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
 {
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( rhs_v, rhs, AcceleratorRead);
+  autoView( lhs_v, lhs, AcceleratorWrite);
+  accelerator_for( ss, lhs_v.size(), 1, {
     pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
   });
 }
 template<int Index,class vobj> 
 void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
 {
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( rhs_v, rhs, AcceleratorRead);
+  autoView( lhs_v, lhs, AcceleratorWrite);
+  accelerator_for( ss, lhs_v.size(), 1, {
     pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
   });
 }
@@ -111,7 +111,7 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   // extract-modify-merge cycle is easiest way and this is not perf critical
   ExtractBuffer<sobj> buf(Nsimd);
-  auto l_v = l.View();
+  autoView( l_v , l, CpuWrite);
   if ( rank == grid->ThisRank() ) {
     extract(l_v[odx],buf);
     buf[idx] = s;
@@ -141,7 +141,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
 
   ExtractBuffer<sobj> buf(Nsimd);
-  auto l_v = l.View();
+  autoView( l_v , l, CpuWrite);
   extract(l_v[odx],buf);
 
   s = buf[idx];
@@ -151,21 +151,21 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
   return;
 };
 
-
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+// Must be CPU read view
 template<class vobj,class sobj>
-inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
-        
-  GridBase *grid = l.Grid();
-
+inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
+{
+  GridBase *grid = l.getGrid();
+  assert(l.mode==CpuRead);
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
+  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -173,8 +173,7 @@ inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  auto l_v = l.View();
-  scalar_type * vp = (scalar_type *)&l_v[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
@@ -183,18 +182,19 @@ inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
       
   return;
 };
-
+// Must be CPU write view
 template<class vobj,class sobj>
-inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
-
-  GridBase *grid=l.Grid();
+inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
+{
+  GridBase *grid=l.getGrid();
+  assert(l.mode==CpuWrite);
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
+  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -202,13 +202,11 @@ inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  auto l_v = l.View();
-  scalar_type * vp = (scalar_type *)&l_v[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
   for(int w=0;w<words;w++){
     vp[idx+w*Nsimd] = pt[w];
   }
-
   return;
 };
 
diff --git a/Grid/lattice/Lattice_reality.h b/Grid/lattice/Lattice_reality.h
index 96af8dab..61491d6b 100644
--- a/Grid/lattice/Lattice_reality.h
+++ b/Grid/lattice/Lattice_reality.h
@@ -40,9 +40,11 @@ NAMESPACE_BEGIN(Grid);
 
 template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
   ret.Checkerboard()=lhs.Checkerboard();
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], adj(lhs_v(ss)));
   });
@@ -51,9 +53,11 @@ template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
 
 template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
   ret.Checkerboard() = lhs.Checkerboard();
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite( ret_v[ss] , conjugate(lhs_v(ss)));
   });
diff --git a/Grid/lattice/Lattice_reduction.h b/Grid/lattice/Lattice_reduction.h
index 3147823d..c2955485 100644
--- a/Grid/lattice/Lattice_reduction.h
+++ b/Grid/lattice/Lattice_reduction.h
@@ -25,7 +25,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #include <Grid/Grid_Eigen_Dense.h>
 
 
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
 
@@ -39,7 +39,36 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
 {
   typedef typename vobj::scalar_object  sobj;
 
-  const int Nsimd = vobj::Nsimd();
+  //  const int Nsimd = vobj::Nsimd();
+  const int nthread = GridThread::GetThreads();
+
+  Vector<sobj> sumarray(nthread);
+  for(int i=0;i<nthread;i++){
+    sumarray[i]=Zero();
+  }
+  
+  thread_for(thr,nthread, {
+    int nwork, mywork, myoff;
+    nwork = osites;
+    GridThread::GetWork(nwork,thr,mywork,myoff);
+    vobj vvsum=Zero();
+    for(int ss=myoff;ss<mywork+myoff; ss++){
+      vvsum = vvsum + arg[ss];
+    }
+    sumarray[thr]=Reduce(vvsum);
+  });
+  
+  sobj ssum=Zero();  // sum across threads
+  for(int i=0;i<nthread;i++){
+    ssum = ssum+sumarray[i];
+  } 
+  return ssum;
+}
+template<class vobj>
+inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
+{
+  typedef typename vobj::scalar_objectD  sobj;
+
   const int nthread = GridThread::GetThreads();
 
   Vector<sobj> sumarray(nthread);
@@ -63,23 +92,43 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
     ssum = ssum+sumarray[i];
   } 
   
-  return ssum;
+  typedef typename vobj::scalar_object ssobj;
+  ssobj ret = ssum;
+  return ret;
 }
+
+
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
 #endif  
 }
+template<class vobj>
+inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  return sumD_gpu(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
+
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-  auto arg_v = arg.View();
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum(&arg_v[0],osites);
+  auto ssum= sum_gpu(&arg_v[0],osites);
+#else
+  autoView(arg_v, arg, CpuRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif  
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -101,43 +150,30 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   ComplexD  nrm;
   
   GridBase *grid = left.Grid();
-  
-  // Might make all code paths go this way.
-  auto left_v = left.AcceleratorView(ViewRead);
-  auto right_v=right.AcceleratorView(ViewRead);
 
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-#ifdef GRID_NVCC
-  // GPU - SIMT lane compliance...
-  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
+  // Might make all code paths go this way.
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  
+    
+  {
+    autoView( left_v , left, AcceleratorRead);
+    autoView( right_v,right, AcceleratorRead);
 
-  accelerator_for( ss, sites, nsimd,{
-      auto x_l = left_v(ss);
-      auto y_l = right_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
-  })
+    // GPU - SIMT lane compliance...
+    accelerator_for( ss, sites, 1,{
+	auto x_l = left_v[ss];
+	auto y_l = right_v[ss];
+	inner_tmp_v[ss]=innerProductD(x_l,y_l);
+    });
+  }
 
   // This is in single precision and fails some tests
-  // Need a sumD that sums in double
-  nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));  
-#else
-  // CPU 
-  typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  
-  accelerator_for( ss, sites, nsimd,{
-      auto x_l = left_v[ss];
-      auto y_l = right_v[ss];
-      inner_tmp_v[ss]=innerProductD(x_l,y_l);
-  })
-  nrm = TensorRemove(sum(inner_tmp_v,sites));
-#endif
+  auto anrm = sum(inner_tmp_v,sites);  
+  nrm = anrm;
   return nrm;
 }
 
@@ -175,40 +211,24 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   
   GridBase *grid = x.Grid();
 
-  auto x_v=x.AcceleratorView(ViewRead);
-  auto y_v=y.AcceleratorView(ViewRead);
-  auto z_v=z.AcceleratorView(ViewWrite);
-
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-#ifdef GRID_NVCC
   // GPU
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
 
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-
-  nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
-#else
-  // CPU 
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
+
+  accelerator_for( ss, sites, 1,{
+      auto tmp = a*x_v[ss]+b*y_v[ss];
       inner_tmp_v[ss]=innerProductD(tmp,tmp);
       z_v[ss]=tmp;
   });
-  // Already promoted to double
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -224,47 +244,29 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
 
   GridBase *grid = left.Grid();
 
-  auto left_v=left.AcceleratorView(ViewRead);
-  auto right_v=right.AcceleratorView(ViewRead);
-
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
 
-#ifdef GRID_NVCC
   // GPU
-  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
-  typedef decltype(innerProduct(left_v[0],left_v[0])) norm_t;
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
+  typedef decltype(innerProductD(vobj(),vobj())) norm_t;
   Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t> norm_tmp(sites);
+  Vector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
+  {
+    autoView(left_v,left, AcceleratorRead);
+    autoView(right_v,right,AcceleratorRead);
+    accelerator_for( ss, sites, 1,{
+	auto left_tmp = left_v[ss];
+	inner_tmp_v[ss]=innerProductD(left_tmp,right_v[ss]);
+        norm_tmp_v [ss]=innerProductD(left_tmp,left_tmp);
+      });
+  }
 
-  accelerator_for( ss, sites, nsimd,{
-      auto left_tmp = left_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(left_tmp,right_v(ss)));
-      coalescedWrite(norm_tmp_v[ss],innerProduct(left_tmp,left_tmp));
-  });
-
-  tmp[0] = TensorRemove(sumD_gpu(inner_tmp_v,sites));
-  tmp[1] = TensorRemove(sumD_gpu(norm_tmp_v,sites));
-#else
-  // CPU
-  typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
-  typedef decltype(innerProductD(left_v[0],left_v[0])) norm_t;
-  Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t> norm_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  auto norm_tmp_v = &norm_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto left_tmp = left_v(ss);
-      inner_tmp_v[ss] = innerProductD(left_tmp,right_v(ss));
-      norm_tmp_v[ss] = innerProductD(left_tmp,left_tmp);
-  });
-  // Already promoted to double
   tmp[0] = TensorRemove(sum(inner_tmp_v,sites));
   tmp[1] = TensorRemove(sum(norm_tmp_v,sites));
-#endif
+
   grid->GlobalSumVector(&tmp[0],2); // keep norm Complex -> can use GlobalSumVector
   ip = tmp[0];
   nrm = real(tmp[1]);
@@ -335,7 +337,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
 
   // sum over reduced dimension planes, breaking out orthog dir
   // Parallel over orthog direction
-  auto Data_v=Data.View();
+  autoView( Data_v, Data, CpuRead);
   thread_for( r,rd, {
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
     for(int n=0;n<e1;n++){
@@ -413,8 +415,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
 
-  auto lhv=lhs.View();
-  auto rhv=rhs.View();
+  autoView( lhv, lhs, CpuRead);
+  autoView( rhv, rhs, CpuRead);
   thread_for( r,rd,{
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
@@ -521,14 +523,12 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
 
     tensor_reduced at; at=av;
 
-    auto Rv=R.View();
-    auto Xv=X.View();
-    auto Yv=Y.View();
-    thread_for_collapse(2, n, e1, {
-      for(int b=0;b<e2;b++){
+    autoView( Rv, R, CpuWrite);
+    autoView( Xv, X, CpuRead);
+    autoView( Yv, Y, CpuRead);
+    thread_for2d( n, e1, b,e2, {
 	int ss= so+n*stride+b;
 	Rv[ss] = at*Xv[ss]+Yv[ss];
-      }
     });
   }
 };
@@ -581,9 +581,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  auto X_v=X.View();
-  auto Y_v=Y.View();
-  auto R_v=R.View();
+  autoView( X_v, X, CpuRead);
+  autoView( Y_v, Y, CpuRead);
+  autoView( R_v, R, CpuWrite);
   thread_region
   {
     Vector<vobj> s_x(Nblock);
@@ -628,13 +628,14 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   //  int nl=1;
 
   //FIXME package in a convenient iterator
+  // thread_for2d_in_region
   //Should loop over a plane orthogonal to direction "Orthog"
   int stride=FullGrid->_slice_stride[Orthog];
   int block =FullGrid->_slice_block [Orthog];
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
-  auto R_v = R.View();
-  auto X_v = X.View();
+  autoView( R_v, R, CpuWrite);
+  autoView( X_v, X, CpuRead);
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -692,8 +693,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
 
   typedef typename vobj::vector_typeD vector_typeD;
 
-  auto lhs_v=lhs.View();
-  auto rhs_v=rhs.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
   thread_region
   {
     std::vector<vobj> Left(Nblock);
diff --git a/Grid/lattice/Lattice_reduction_gpu.h b/Grid/lattice/Lattice_reduction_gpu.h
index c5d75356..5f490507 100644
--- a/Grid/lattice/Lattice_reduction_gpu.h
+++ b/Grid/lattice/Lattice_reduction_gpu.h
@@ -1,7 +1,13 @@
 NAMESPACE_BEGIN(Grid);
 
-#define WARP_SIZE 32
+#ifdef GRID_HIP
+extern hipDeviceProp_t *gpu_props;
+#endif
+#ifdef GRID_CUDA
 extern cudaDeviceProp *gpu_props;
+#endif
+
+#define WARP_SIZE 32
 __device__ unsigned int retirementCount = 0;
 
 template <class Iterator>
@@ -19,7 +25,12 @@ template <class Iterator>
 void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
   
   int device;
+#ifdef GRID_CUDA
   cudaGetDevice(&device);
+#endif
+#ifdef GRID_HIP
+  hipGetDevice(&device);
+#endif
   
   Iterator warpSize            = gpu_props[device].warpSize;
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
@@ -147,7 +158,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     sobj *smem = (sobj *)shmem_pointer;
     
     // wait until all outstanding memory instructions in this thread are finished
-    __threadfence();
+    acceleratorFence();
     
     if (tid==0) {
       unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
@@ -156,8 +167,8 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     }
     
     // each thread must read the correct value of amLast
-    __syncthreads();
-    
+    acceleratorSynchroniseAll();
+
     if (amLast) {
       // reduce buffer[0], ..., buffer[gridDim.x-1]
       Iterator i = tid;
@@ -199,13 +210,7 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   sobj *buffer_v = &buffer[0];
   
   reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
-  cudaDeviceSynchronize();
-  
-  cudaError err = cudaGetLastError();
-  if ( cudaSuccess != err ) {
-    printf("Cuda error %s\n",cudaGetErrorString( err ));
-    exit(0);
-  }
+  accelerator_barrier();
   auto result = buffer_v[0];
   return result;
 }
diff --git a/Grid/lattice/Lattice_rng.h b/Grid/lattice/Lattice_rng.h
index 1bb1f087..e5e63716 100644
--- a/Grid/lattice/Lattice_rng.h
+++ b/Grid/lattice/Lattice_rng.h
@@ -375,7 +375,7 @@ public:
     int osites = _grid->oSites();  // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
     int words  = sizeof(scalar_object) / sizeof(scalar_type);
 
-    auto l_v = l.View();
+    autoView(l_v, l, CpuWrite);
     thread_for( ss, osites, {
       ExtractBuffer<scalar_object> buf(Nsimd);
       for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
@@ -461,8 +461,8 @@ public:
     }
 
     {
-      // Obtain one reseeded generator per thread
-      int Nthread = GridThread::GetThreads();
+      // Obtain one reseeded generator per thread      
+      int Nthread = 32; // Hardwire a good level or parallelism
       std::vector<RngEngine> seeders(Nthread);
       for(int t=0;t<Nthread;t++){
 	seeders[t] = Reseed(master_engine);
diff --git a/Grid/lattice/Lattice_trace.h b/Grid/lattice/Lattice_trace.h
index 6b5f67d2..5fe5173d 100644
--- a/Grid/lattice/Lattice_trace.h
+++ b/Grid/lattice/Lattice_trace.h
@@ -42,8 +42,8 @@ template<class vobj>
 inline auto trace(const Lattice<vobj> &lhs)  -> Lattice<decltype(trace(vobj()))>
 {
   Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView(ret_v , ret, AcceleratorWrite);
+  autoView(lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
   });
@@ -58,8 +58,8 @@ template<int Index,class vobj>
 inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
 {
   Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
   });
diff --git a/Grid/lattice/Lattice_transfer.h b/Grid/lattice/Lattice_transfer.h
index 435b7851..beceecc9 100644
--- a/Grid/lattice/Lattice_transfer.h
+++ b/Grid/lattice/Lattice_transfer.h
@@ -47,11 +47,12 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // remove and insert a half checkerboard
 ////////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
+template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full)
+{
   half.Checkerboard() = cb;
 
-  auto half_v = half.View();
-  auto full_v = full.View();
+  autoView( half_v, half, CpuWrite);
+  autoView( full_v, full, CpuRead);
   thread_for(ss, full.Grid()->oSites(),{
     int cbos;
     Coordinate coor;
@@ -64,11 +65,11 @@ template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,con
     }
   });
 }
-
-template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
+template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half)
+{
   int cb = half.Checkerboard();
-  auto half_v = half.View();
-  auto full_v = full.View();
+  autoView( half_v , half, CpuRead);
+  autoView( full_v , full, CpuWrite);
   thread_for(ss,full.Grid()->oSites(),{
 
     Coordinate coor;
@@ -96,15 +97,15 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> &
   out = in;
 }
 
-#ifdef GRID_NVCC
+#ifdef GRID_SIMT
 accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
-  ((ComplexF*)&out)[SIMTlane(vComplexF::Nsimd())] = in;
+  ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
 }
 accelerator_inline void convertType(vComplexD & out, const ComplexD & in) {
-  ((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd())] = in;
+  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd())] = in;
 }
 accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
-  ((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd()*2)] = in;
+  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd()*2)] = in;
 }
 #endif
 
@@ -151,12 +152,11 @@ accelerator_inline void convertType(T & out, const T & in) {
 
 template<typename T1,typename T2>
 accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
-  auto out_v = out.AcceleratorView(ViewWrite);
-  auto in_v  = in.AcceleratorView(ViewRead);
-
+  autoView( out_v , out,AcceleratorWrite);
+  autoView( in_v  , in ,AcceleratorRead);
   accelerator_for(ss,out_v.size(),T1::Nsimd(),{
       convertType(out_v[ss],in_v(ss));
-    });
+  });
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////
@@ -164,19 +164,20 @@ accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
 ////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj>
 inline auto localInnerProductD(const Lattice<vobj> &lhs,const Lattice<vobj> &rhs)
--> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View()[0],rhs.View()[0])))>>
+-> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View(CpuRead)[0],rhs.View(CpuRead)[0])))>>
 {
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
 
   typedef decltype(TensorRemove(innerProductD2(lhs_v[0],rhs_v[0]))) t_inner;
   Lattice<iScalar<t_inner>> ret(lhs.Grid());
-  auto ret_v = ret.AcceleratorView(ViewWrite);
 
-  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
+  {
+    autoView(ret_v, ret,AcceleratorWrite);
+    accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
       convertType(ret_v[ss],innerProductD2(lhs_v(ss),rhs_v(ss)));
     });
-
+  }
   return ret;
 }
 
@@ -194,14 +195,13 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
   Lattice<iScalar<CComplex>> ip(coarse);
   Lattice<vobj>     fineDataRed = fineData;
 
-  //  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.AcceleratorView(ViewWrite);
-  auto ip_         = ip.AcceleratorView(ViewReadWrite);
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( ip_         , ip,         AcceleratorWrite);
   for(int v=0;v<nbasis;v++) {
     blockInnerProductD(ip,Basis[v],fineDataRed); // ip = <basis|fine>
     accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
 	convertType(coarseData_[sc](v),ip_[sc]);
-      });
+    });
 
     // improve numerical stability of projection
     // |fine> = |fine> - <basis|fine> |basis>
@@ -210,68 +210,6 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
   }
 }
 
-template<class vobj,class CComplex,int nbasis>
-inline void blockProject1(Lattice<iVector<CComplex,nbasis > > &coarseData,
-			 const             Lattice<vobj>   &fineData,
-			 const std::vector<Lattice<vobj> > &Basis)
-{
-  typedef iVector<CComplex,nbasis > coarseSiteData;
-  coarseSiteData elide;
-  typedef decltype(coalescedRead(elide)) ScalarComplex;
-  GridBase * fine  = fineData.Grid();
-  GridBase * coarse= coarseData.Grid();
-  int  _ndimension = coarse->_ndimension;
-
-  // checks
-  assert( nbasis == Basis.size() );
-  subdivides(coarse,fine); 
-  for(int i=0;i<nbasis;i++){
-    conformable(Basis[i],fineData);
-  }
-
-  Coordinate block_r      (_ndimension);
-  
-  for(int d=0 ; d<_ndimension;d++){
-    block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
-    assert(block_r[d]*coarse->_rdimensions[d] == fine->_rdimensions[d]);
-  }
-  int blockVol = fine->oSites()/coarse->oSites();
-
-  coarseData=Zero();
-
-  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
-  ////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // To make this lock free, loop over coars parallel, and then loop over fine associated with coarse.
-  // Otherwise do fine inner product per site, and make the update atomic
-  ////////////////////////////////////////////////////////////////////////////////////////////////////////
-  accelerator_for( sci, nbasis*coarse->oSites(), vobj::Nsimd(), {
-
-    auto sc=sci/nbasis;
-    auto i=sci%nbasis;
-    auto Basis_      = Basis[i].View();
-
-    Coordinate coor_c(_ndimension);
-    Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions);  // Block coordinate
-
-    int sf;
-    decltype(innerProduct(Basis_(sf),fineData_(sf))) reduce=Zero();
-
-    for(int sb=0;sb<blockVol;sb++){
-
-      Coordinate coor_b(_ndimension);
-      Coordinate coor_f(_ndimension);
-
-      Lexicographic::CoorFromIndex(coor_b,sb,block_r);
-      for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d]+coor_b[d];
-      Lexicographic::IndexFromCoor(coor_f,sf,fine->_rdimensions);
-      
-      reduce=reduce+innerProduct(Basis_(sf),fineData_(sf));
-    }
-    coalescedWrite(coarseData_[sc](i),reduce);
-  });
-  return;
-}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -298,10 +236,10 @@ template<class vobj,class vobj2,class CComplex>
     assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
-  auto fineZ_  = fineZ.AcceleratorView(ViewWrite);
-  auto fineX_  = fineX.AcceleratorView(ViewRead);
-  auto fineY_  = fineY.AcceleratorView(ViewRead);
-  auto coarseA_= coarseA.AcceleratorView(ViewRead);
+  autoView( fineZ_  , fineZ, AcceleratorWrite);
+  autoView( fineX_  , fineX, AcceleratorRead);
+  autoView( fineY_  , fineY, AcceleratorRead);
+  autoView( coarseA_, coarseA, AcceleratorRead);
 
   accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
 
@@ -314,7 +252,7 @@ template<class vobj,class vobj2,class CComplex>
       Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
       // z = A x + y
-#ifdef __CUDA_ARCH__
+#ifdef GRID_SIMT
       typename vobj2::tensor_reduced::scalar_object cA;
       typename vobj::scalar_object cAx;
 #else
@@ -344,15 +282,16 @@ template<class vobj,class CComplex>
   Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
   Lattice<dotp> coarse_inner(coarse);
 
-  auto CoarseInner_  = CoarseInner.AcceleratorView(ViewWrite);
-  auto coarse_inner_ = coarse_inner.AcceleratorView(ViewReadWrite);
-
   // Precision promotion
-  fine_inner = localInnerProductD(fineX,fineY);
+  fine_inner = localInnerProductD<vobj>(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  accelerator_for(ss, coarse->oSites(), 1, {
+  {
+    autoView( CoarseInner_  , CoarseInner,AcceleratorWrite);
+    autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
+    accelerator_for(ss, coarse->oSites(), 1, {
       convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
     });
+  }
  
 }
 
@@ -370,14 +309,15 @@ inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
   Lattice<dotp> coarse_inner(coarse);
 
   // Precision promotion?
-  auto CoarseInner_  = CoarseInner.AcceleratorView(ViewWrite);
-  auto coarse_inner_ = coarse_inner.AcceleratorView(ViewReadWrite);
-
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  accelerator_for(ss, coarse->oSites(), 1, {
-      convertType(CoarseInner_[ss],coarse_inner_[ss]);
+  {
+    autoView( CoarseInner_  , CoarseInner, AcceleratorWrite);
+    autoView( coarse_inner_ , coarse_inner, AcceleratorRead);
+    accelerator_for(ss, coarse->oSites(), 1, {
+	CoarseInner_[ss] = coarse_inner_[ss];
     });
+  }
 }
 
 template<class vobj,class CComplex>
@@ -408,8 +348,10 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   }
   int blockVol = fine->oSites()/coarse->oSites();
 
-  auto coarseData_ = coarseData.AcceleratorView(ViewReadWrite);
-  auto fineData_   = fineData.AcceleratorView(ViewRead);
+  // Turn this around to loop threaded over sc and interior loop 
+  // over sf would thread better
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( fineData_   , fineData, AcceleratorRead);
 
   accelerator_for(sc,coarse->oSites(),1,{
 
@@ -510,8 +452,8 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
-  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
+  autoView( fineData_   , fineData, AcceleratorWrite);
+  autoView( coarseData_ , coarseData, AcceleratorRead);
 
   // Loop with a cache friendly loop ordering
   accelerator_for(sf,fine->oSites(),1,{
@@ -524,7 +466,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
     Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
     for(int i=0;i<nbasis;i++) {
-      auto basis_ = Basis[i].View();
+      /*      auto basis_ = Basis[i],  );*/
       if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
       else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
     }
@@ -543,7 +485,14 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   fineData=Zero();
   for(int i=0;i<nbasis;i++) {
     Lattice<iScalar<CComplex> > ip = PeekIndex<0>(coarseData,i);
-    auto  ip_ =  ip.AcceleratorView(ViewRead);
+
+    //Lattice<CComplex> cip(coarse);
+    //autoView( cip_ , cip, AcceleratorWrite);
+    //autoView(  ip_ ,  ip, AcceleratorRead);
+    //accelerator_forNB(sc,coarse->oSites(),CComplex::Nsimd(),{
+    //	coalescedWrite(cip_[sc], ip_(sc)());
+    //  });
+    //blockZAXPY<vobj,CComplex >(fineData,cip,Basis[i],fineData);
     blockZAXPY(fineData,ip,Basis[i],fineData);
   }
 }
@@ -571,15 +520,17 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
     assert(ig->lSites() == og->lSites());
   }
 
+  autoView(in_v,in,CpuRead);
+  autoView(out_v,out,CpuWrite);
   thread_for(idx, ig->lSites(),{
     sobj s;
     ssobj ss;
 
     Coordinate lcoor(ni);
     ig->LocalIndexToLocalCoor(idx,lcoor);
-    peekLocalSite(s,in,lcoor);
+    peekLocalSite(s,in_v,lcoor);
     ss=s;
-    pokeLocalSite(ss,out,lcoor);
+    pokeLocalSite(ss,out_v,lcoor);
   });
 }
 
@@ -614,8 +565,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate rdt = Tg->_rdimensions;
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
-  auto t_v = To.AcceleratorView(ViewWrite);
-  auto f_v = From.AcceleratorView(ViewRead);
+
+  autoView( t_v , To, AcceleratorWrite);
+  autoView( f_v , From, AcceleratorRead);
   accelerator_for(idx,Fg->lSites(),1,{
     sobj s;
     Coordinate Fcoor(nd);
@@ -638,8 +590,6 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       for(int w=0;w<words;w++){
 	tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd];  // FIXME IF RRII layout, type pun no worke
       }
-      //      peekLocalSite(s,From,Fcoor);
-      //      pokeLocalSite(s,To  ,Tcoor);
     }
   });
 }
@@ -670,6 +620,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuRead);
+  autoView(higherDimv,higherDim,CpuWrite);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -682,8 +634,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,lowDim,lcoor);
-    pokeLocalSite(s,higherDim,hcoor);
+    peekLocalSite(s,lowDimv,lcoor);
+    pokeLocalSite(s,higherDimv,hcoor);
   });
 }
 
@@ -711,6 +663,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     }
   }
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuWrite);
+  autoView(higherDimv,higherDim,CpuRead);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -723,8 +677,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,higherDim,hcoor);
-    pokeLocalSite(s,lowDim,lcoor);
+    peekLocalSite(s,higherDimv,hcoor);
+    pokeLocalSite(s,lowDimv,lcoor);
   });
 
 }
@@ -752,6 +706,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuRead);
+  autoView(higherDimv,higherDim,CpuWrite);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -760,8 +716,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,lowDim,lcoor);
-      pokeLocalSite(s,higherDim,hcoor);
+      peekLocalSite(s,lowDimv,lcoor);
+      pokeLocalSite(s,higherDimv,hcoor);
     }
   });
 }
@@ -789,6 +745,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuWrite);
+  autoView(higherDimv,higherDim,CpuRead);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -797,8 +755,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,higherDim,hcoor);
-      pokeLocalSite(s,lowDim,lcoor);
+      peekLocalSite(s,higherDimv,hcoor);
+      pokeLocalSite(s,lowDimv,lcoor);
     }
   });
 }
@@ -862,7 +820,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
   }
 
   //loop over outer index
-  auto in_v  = in.View();
+  autoView( in_v  , in, CpuRead);
   thread_for(in_oidx,in_grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> out_ptrs(in_nsimd);
@@ -955,7 +913,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
     icoor[lane].resize(ndim);
     grid->iCoorFromIindex(icoor[lane],lane);
   }
-  auto out_v = out.View();
+  autoView( out_v , out, CpuWrite);
   thread_for(oidx, grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> ptrs(nsimd);
@@ -1058,7 +1016,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   std::vector<SobjOut> in_slex_conv(in_grid->lSites());
   unvectorizeToLexOrdArray(in_slex_conv, in);
     
-  auto out_v = out.View();
+  autoView( out_v , out, CpuWrite);
   thread_for(out_oidx,out_grid->oSites(),{
     Coordinate out_ocoor(ndim);
     out_grid->oCoorFromOindex(out_ocoor, out_oidx);
diff --git a/Grid/lattice/Lattice_transpose.h b/Grid/lattice/Lattice_transpose.h
index 6fe08c10..adfe3380 100644
--- a/Grid/lattice/Lattice_transpose.h
+++ b/Grid/lattice/Lattice_transpose.h
@@ -42,8 +42,8 @@ NAMESPACE_BEGIN(Grid);
 template<class vobj>
 inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
   });
@@ -58,8 +58,8 @@ template<int Index,class vobj>
 inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
 {
   Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
   });
diff --git a/Grid/lattice/Lattice_unary.h b/Grid/lattice/Lattice_unary.h
index 591afe72..07424b3d 100644
--- a/Grid/lattice/Lattice_unary.h
+++ b/Grid/lattice/Lattice_unary.h
@@ -35,8 +35,8 @@ NAMESPACE_BEGIN(Grid);
 
 template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs, rhs_i, AcceleratorRead);
+  autoView( ret, ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),1,{
       ret[ss]=pow(rhs[ss],y);
@@ -45,8 +45,8 @@ template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
 }
 template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs , rhs_i, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],mod(rhs(ss),y));
@@ -56,8 +56,8 @@ template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
 
 template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto ret = ret_i.View();
-  auto rhs = rhs_i.View();
+  autoView( ret , ret_i, AcceleratorWrite);
+  autoView( rhs , rhs_i, AcceleratorRead);
   ret.Checkerboard() = rhs_i.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],div(rhs(ss),y));
@@ -67,8 +67,8 @@ template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
 
 template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs , rhs_i, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));
diff --git a/Grid/lattice/Lattice_view.h b/Grid/lattice/Lattice_view.h
new file mode 100644
index 00000000..3b76b921
--- /dev/null
+++ b/Grid/lattice/Lattice_view.h
@@ -0,0 +1,168 @@
+#pragma once
+NAMESPACE_BEGIN(Grid);
+///////////////////////////////////////////////////////////////////
+// Base class which can be used by traits to pick up behaviour
+///////////////////////////////////////////////////////////////////
+class LatticeBase {};
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// Conformable checks; same instance of Grid required
+/////////////////////////////////////////////////////////////////////////////////////////
+void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
+{
+  assert(lhs == rhs);
+}
+
+////////////////////////////////////////////////////////////////////////////
+// Minimal base class containing only data valid to access from accelerator
+// _odata will be a managed pointer in CUDA
+////////////////////////////////////////////////////////////////////////////
+// Force access to lattice through a view object.
+// prevents writing of code that will not offload to GPU, but perhaps annoyingly
+// strict since host could could in principle direct access through the lattice object
+// Need to decide programming model.
+#define LATTICE_VIEW_STRICT
+template<class vobj> class LatticeAccelerator : public LatticeBase
+{
+protected:
+  //public:
+  GridBase *_grid;
+  int checkerboard;
+  vobj     *_odata;    // A managed pointer
+  uint64_t _odata_size;    
+  ViewAdvise advise;
+public:
+  accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr), advise(AdviseDefault) { }; 
+  accelerator_inline uint64_t oSites(void) const { return _odata_size; };
+  accelerator_inline int  Checkerboard(void) const { return checkerboard; };
+  accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
+  accelerator_inline ViewAdvise Advise(void) const { return advise; };
+  accelerator_inline ViewAdvise &Advise(void) { return this->advise; }; // can assign advise on a container, not a view
+  accelerator_inline void Conformable(GridBase * &grid) const
+  { 
+    if (grid) conformable(grid, _grid);
+    else      grid = _grid;
+  };
+  // Host only
+  GridBase * getGrid(void) const { return _grid; };
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// A View class which provides accessor to the data.
+// 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
+/////////////////////////////////////////////////////////////////////////////////////////
+template<class vobj> 
+class LatticeView : public LatticeAccelerator<vobj>
+{
+public:
+  // Rvalue
+  ViewMode mode;
+  void * cpu_ptr;
+#ifdef GRID_SIMT
+  accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { 
+    return coalescedRead(this->_odata[i]); 
+  }
+#else 
+  accelerator_inline const vobj & operator()(size_t i) const { 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 uint64_t begin(void) const { return 0;};
+  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
+  accelerator_inline uint64_t size(void)  const { return this->_odata_size; };
+
+  LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me){}
+  LatticeView(const LatticeView<vobj> &refer_to_me) = default; // Trivially copyable
+  LatticeView(const LatticeAccelerator<vobj> &refer_to_me,ViewMode mode) : LatticeAccelerator<vobj> (refer_to_me)
+  {
+    this->ViewOpen(mode);
+  }
+
+  // Host functions
+  void ViewOpen(ViewMode mode)
+  { // Translate the pointer, could save a copy. Could use a "Handle" and not save _odata originally in base
+    //    std::cout << "View Open"<<std::hex<<this->_odata<<std::dec <<std::endl;
+    this->cpu_ptr = (void *)this->_odata;
+    this->mode    = mode;
+    this->_odata  =(vobj *)
+      MemoryManager::ViewOpen(this->cpu_ptr,
+				this->_odata_size*sizeof(vobj),
+				mode,
+				this->advise);    
+  }
+  void ViewClose(void)
+  { // Inform the manager
+    //    std::cout << "View Close"<<std::hex<<this->cpu_ptr<<std::dec <<std::endl;
+    MemoryManager::ViewClose(this->cpu_ptr,this->mode);    
+  }
+
+};
+// Little autoscope assister
+template<class View> 
+class ViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+ public:
+  ViewCloser(View &_v) : v(_v) {};
+  ~ViewCloser() { v.ViewClose(); }
+};
+
+#define autoView(l_v,l,mode)				\
+	  auto l_v = l.View(mode);			\
+	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v);
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// Lattice expression types used by ET to assemble the AST
+// 
+// Need to be able to detect code paths according to the whether a lattice object or not
+// so introduce some trait type things
+/////////////////////////////////////////////////////////////////////////////////////////
+
+class LatticeExpressionBase {};
+
+template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
+template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
+
+template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
+template<class T>                 struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
+template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
+
+template <typename Op, typename _T1>                           
+class LatticeUnaryExpression : public  LatticeExpressionBase 
+{
+public:
+  typedef typename ViewMap<_T1>::Type T1;
+  Op op;
+  T1 arg1;
+  LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
+};
+
+template <typename Op, typename _T1, typename _T2>              
+class LatticeBinaryExpression : public LatticeExpressionBase 
+{
+public:
+  typedef typename ViewMap<_T1>::Type T1;
+  typedef typename ViewMap<_T2>::Type T2;
+  Op op;
+  T1 arg1;
+  T2 arg2;
+  LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
+};
+
+template <typename Op, typename _T1, typename _T2, typename _T3> 
+class LatticeTrinaryExpression : public LatticeExpressionBase 
+{
+public:
+  typedef typename ViewMap<_T1>::Type T1;
+  typedef typename ViewMap<_T2>::Type T2;
+  typedef typename ViewMap<_T3>::Type T3;
+  Op op;
+  T1 arg1;
+  T2 arg2;
+  T3 arg3;
+  LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
+};
+NAMESPACE_END(Grid);
diff --git a/Grid/perfmon/PerfCount.h b/Grid/perfmon/PerfCount.h
index 1e2a9528..dd25b41e 100644
--- a/Grid/perfmon/PerfCount.h
+++ b/Grid/perfmon/PerfCount.h
@@ -44,7 +44,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <sys/syscall.h>
 #endif
 #ifdef __x86_64__
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 accelerator_inline uint64_t __rdtsc(void) {  return 0; }
 accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
 #else
@@ -112,7 +112,6 @@ class PerformanceCounter {
 private:
 
   typedef struct { 
-  public:
     uint32_t type;
     uint64_t config;
     const char *name;
diff --git a/Grid/pugixml/pugixml.cc b/Grid/pugixml/pugixml.cc
index e7b395ad..45e6496a 100644
--- a/Grid/pugixml/pugixml.cc
+++ b/Grid/pugixml/pugixml.cc
@@ -12773,7 +12773,7 @@ namespace pugi
 #undef PUGI__THROW_ERROR
 #undef PUGI__CHECK_ERROR
 
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #pragma pop
 #endif
 
diff --git a/Grid/qcd/action/fermion/DomainWallVec5dImpl.h b/Grid/qcd/action/fermion/DomainWallVec5dImpl.h
index 890c680b..0c8a0930 100644
--- a/Grid/qcd/action/fermion/DomainWallVec5dImpl.h
+++ b/Grid/qcd/action/fermion/DomainWallVec5dImpl.h
@@ -114,19 +114,22 @@ public:
       U = adj(Cshift(U, mu, -1));
       PokeIndex<LorentzIndex>(Uadj, U, mu);
     }
-    
-    for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
+
+    autoView(Umu_v,Umu,CpuRead);
+    autoView(Uadj_v,Uadj,CpuRead);
+    autoView(Uds_v,Uds,CpuWrite);
+    thread_for( lidx, GaugeGrid->lSites(), {
       Coordinate lcoor;
       GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
       
-      peekLocalSite(ScalarUmu, Umu, lcoor);
+      peekLocalSite(ScalarUmu, Umu_v, lcoor);
       for (int mu = 0; mu < 4; mu++) ScalarUds(mu) = ScalarUmu(mu);
       
-      peekLocalSite(ScalarUmu, Uadj, lcoor);
+      peekLocalSite(ScalarUmu, Uadj_v, lcoor);
       for (int mu = 0; mu < 4; mu++) ScalarUds(mu + 4) = ScalarUmu(mu);
       
-      pokeLocalSite(ScalarUds, Uds, lcoor);
-    }
+      pokeLocalSite(ScalarUds, Uds_v, lcoor);
+    });
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu) 
diff --git a/Grid/qcd/action/fermion/Fermion.h b/Grid/qcd/action/fermion/Fermion.h
index fb6f18bb..16252340 100644
--- a/Grid/qcd/action/fermion/Fermion.h
+++ b/Grid/qcd/action/fermion/Fermion.h
@@ -57,6 +57,7 @@ NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
 
+#include <Grid/qcd/action/fermion/NaiveStaggeredFermion.h>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion.h>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
 NAMESPACE_CHECK(Staggered);
@@ -282,11 +283,15 @@ typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
 typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
 
+typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR;
+typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
+typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
+
 typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
 typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
 typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
 
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
 typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR;
 typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF;
 typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD;
diff --git a/Grid/qcd/action/fermion/GparityWilsonImpl.h b/Grid/qcd/action/fermion/GparityWilsonImpl.h
index 0b147b3f..0b726db9 100644
--- a/Grid/qcd/action/fermion/GparityWilsonImpl.h
+++ b/Grid/qcd/action/fermion/GparityWilsonImpl.h
@@ -96,11 +96,11 @@ public:
     int sl        = St._simd_layout[direction];
     Coordinate icoor;
 
-#ifdef __CUDA_ARCH__
+#ifdef GRID_SIMT
     _Spinor tmp;
 
     const int Nsimd =SiteDoubledGaugeField::Nsimd();
-    int s = SIMTlane(Nsimd);
+    int s = acceleratorSIMTlane(Nsimd);
     St.iCoorFromIindex(icoor,s);
 
     int mmu = mu % Nd;
@@ -232,15 +232,17 @@ public:
       if ( Params.twists[mu] ) { 
 	Uconj = where(coor==neglink,-Uconj,Uconj);
       }
-	  
-      auto U_v = U.View();
-      auto Uds_v = Uds.View();
-      auto Uconj_v = Uconj.View();
-      auto Utmp_v= Utmp.View();
-      thread_foreach(ss,U_v,{
-	Uds_v[ss](0)(mu) = U_v[ss]();
-	Uds_v[ss](1)(mu) = Uconj_v[ss]();
-      });
+
+      {
+	autoView( U_v , U, CpuRead);
+	autoView( Uconj_v , Uconj, CpuRead);
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,U_v,{
+	    Uds_v[ss](0)(mu) = U_v[ss]();
+	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
+	  });
+      }
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
       Uconj = adj(Cshift(Uconj,mu,-1));
@@ -250,19 +252,25 @@ public:
 	Utmp = where(coor==0,Uconj,Utmp);
       }
 
-      thread_foreach(ss,Utmp_v,{
-	Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
-      });
-          
+      {
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,Utmp_v,{
+	    Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
+	  });
+      }
       Utmp = Uconj;
       if ( Params.twists[mu] ) { 
 	Utmp = where(coor==0,U,Utmp);
       }
-	  
-      thread_foreach(ss,Utmp_v,{
-        Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
-      });
-          
+
+      {	  
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,Utmp_v,{
+	    Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
+        });
+      }
     }
   }
       
@@ -272,11 +280,14 @@ public:
     GaugeLinkField link(mat.Grid());
     // use lorentz for flavour as hack.
     auto tmp = TraceIndex<SpinIndex>(outerProduct(Btilde, A));
-    auto link_v = link.View();
-    auto tmp_v = tmp.View();
-    thread_foreach(ss,tmp_v,{
-      link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
-    });
+
+    {
+      autoView( link_v , link, CpuWrite);
+      autoView( tmp_v , tmp, CpuRead);
+      thread_foreach(ss,tmp_v,{
+	  link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
+	});
+    }
     PokeIndex<LorentzIndex>(mat, link, mu);
     return;
   }
@@ -306,16 +317,18 @@ public:
         
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    auto tmp_v = tmp.View();
-    auto Atilde_v = Atilde.View();
-    auto Btilde_v = Btilde.View();
-    thread_for(ss,tmp.Grid()->oSites(),{
-      for (int s = 0; s < Ls; s++) {
-	int sF = s + Ls * ss;
-	auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
-	tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
-      }
-    });
+    {
+      autoView( tmp_v , tmp, CpuWrite);
+      autoView( Atilde_v , Atilde, CpuRead);
+      autoView( Btilde_v , Btilde, CpuRead);
+      thread_for(ss,tmp.Grid()->oSites(),{
+	  for (int s = 0; s < Ls; s++) {
+	    int sF = s + Ls * ss;
+	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+	  }
+	});
+    }
     PokeIndex<LorentzIndex>(mat, tmp, mu);
     return;
   }
diff --git a/Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h b/Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h
index d1bb0e9c..625eda63 100644
--- a/Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h
+++ b/Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h
@@ -61,8 +61,8 @@ public:
   double DhopCalls;
   double DhopCommTime;
   double DhopComputeTime;
-      double DhopComputeTime2;
-      double DhopFaceTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
diff --git a/Grid/qcd/action/fermion/NaiveStaggeredFermion.h b/Grid/qcd/action/fermion/NaiveStaggeredFermion.h
new file mode 100644
index 00000000..ca38a64f
--- /dev/null
+++ b/Grid/qcd/action/fermion/NaiveStaggeredFermion.h
@@ -0,0 +1,194 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/ImprovedStaggered.h
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#ifndef GRID_QCD_NAIVE_STAG_FERMION_H
+#define GRID_QCD_NAIVE_STAG_FERMION_H
+
+NAMESPACE_BEGIN(Grid);
+
+class NaiveStaggeredFermionStatic {
+public:
+  static const std::vector<int> directions;
+  static const std::vector<int> displacements;
+  static const int npoint = 8;
+};
+
+template <class Impl>
+class NaiveStaggeredFermion : public StaggeredKernels<Impl>, public NaiveStaggeredFermionStatic {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+  typedef StaggeredKernels<Impl> Kernels;
+
+  FermionField _tmp;
+  FermionField &tmp(void) { return _tmp; }
+
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
+  ///////////////////////////////////////////////////////////////
+  // Implement the abstract base
+  ///////////////////////////////////////////////////////////////
+  GridBase *GaugeGrid(void) { return _grid; }
+  GridBase *GaugeRedBlackGrid(void) { return _cbgrid; }
+  GridBase *FermionGrid(void) { return _grid; }
+  GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
+
+  //////////////////////////////////////////////////////////////////
+  // override multiply; cut number routines if pass dagger argument
+  // and also make interface more uniformly consistent
+  //////////////////////////////////////////////////////////////////
+  void M(const FermionField &in, FermionField &out);
+  void Mdag(const FermionField &in, FermionField &out);
+
+  /////////////////////////////////////////////////////////
+  // half checkerboard operations
+  /////////////////////////////////////////////////////////
+  void Meooe(const FermionField &in, FermionField &out);
+  void MeooeDag(const FermionField &in, FermionField &out);
+  void Mooee(const FermionField &in, FermionField &out);
+  void MooeeDag(const FermionField &in, FermionField &out);
+  void MooeeInv(const FermionField &in, FermionField &out);
+  void MooeeInvDag(const FermionField &in, FermionField &out);
+
+  ////////////////////////
+  // Derivative interface
+  ////////////////////////
+  // Interface calls an internal routine
+  void DhopDeriv  (GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+  void DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+  void DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+
+  ///////////////////////////////////////////////////////////////
+  // non-hermitian hopping term; half cb or both
+  ///////////////////////////////////////////////////////////////
+  void Dhop  (const FermionField &in, FermionField &out, int dag);
+  void DhopOE(const FermionField &in, FermionField &out, int dag);
+  void DhopEO(const FermionField &in, FermionField &out, int dag);
+
+  ///////////////////////////////////////////////////////////////
+  // Multigrid assistance; force term uses too
+  ///////////////////////////////////////////////////////////////
+  void Mdir(const FermionField &in, FermionField &out, int dir, int disp);
+  void MdirAll(const FermionField &in, std::vector<FermionField> &out);
+  void DhopDir(const FermionField &in, FermionField &out, int dir, int disp);
+
+  ///////////////////////////////////////////////////////////////
+  // Extra methods added by derived
+  ///////////////////////////////////////////////////////////////
+  void DerivInternal(StencilImpl &st, 
+		     DoubledGaugeField &U,
+		     GaugeField &mat, 
+		     const FermionField &A, const FermionField &B, int dag);
+
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
+  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+			       const FermionField &in, FermionField &out, int dag);
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+				   const FermionField &in, FermionField &out, int dag);
+
+  //////////////////////////////////////////////////////////////////////////
+  // Grid own interface Constructor
+  //////////////////////////////////////////////////////////////////////////
+  NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
+			GridRedBlackCartesian &Hgrid, RealD _mass,
+			RealD _c1, RealD _u0,
+			const ImplParams &p = ImplParams());
+  NaiveStaggeredFermion(GridCartesian &Fgrid,
+			GridRedBlackCartesian &Hgrid, RealD _mass,
+			RealD _c1, RealD _u0,
+			const ImplParams &p = ImplParams());
+
+  // DoubleStore impl dependent
+  void ImportGauge      (const GaugeField &_U );
+  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  void CopyGaugeCheckerboards(void);
+
+  ///////////////////////////////////////////////////////////////
+  // Data members require to support the functionality
+  ///////////////////////////////////////////////////////////////
+
+  //    protected:
+public:
+  // any other parameters of action ???
+  virtual int   isTrivialEE(void) { return 1; };
+  virtual RealD Mass(void) { return mass; }
+  RealD mass;
+  RealD u0;
+  RealD c1;
+
+  GridBase *_grid;
+  GridBase *_cbgrid;
+
+  // Defines the stencils for even and odd
+  StencilImpl Stencil;
+  StencilImpl StencilEven;
+  StencilImpl StencilOdd;
+
+  // Copy of the gauge field , with even and odd subsets
+  DoubledGaugeField Umu;
+  DoubledGaugeField UmuEven;
+  DoubledGaugeField UmuOdd;
+
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+  
+  ///////////////////////////////////////////////////////////////
+  // Conserved current utilities
+  ///////////////////////////////////////////////////////////////
+  void ContractConservedCurrent(PropagatorField &q_in_1,
+                                PropagatorField &q_in_2,
+                                PropagatorField &q_out,
+                                PropagatorField &src,
+                                Current curr_type,
+                                unsigned int mu);
+  void SeqConservedCurrent(PropagatorField &q_in,
+                           PropagatorField &q_out,
+                           PropagatorField &srct,
+                           Current curr_type,
+                           unsigned int mu, 
+                           unsigned int tmin,
+                           unsigned int tmax,
+			   ComplexField &lattice_cmplx);
+};
+
+typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
+typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
+
+NAMESPACE_END(Grid);
+
+#endif
diff --git a/Grid/qcd/action/fermion/StaggeredKernels.h b/Grid/qcd/action/fermion/StaggeredKernels.h
index 6ef0ab9d..30deee06 100644
--- a/Grid/qcd/action/fermion/StaggeredKernels.h
+++ b/Grid/qcd/action/fermion/StaggeredKernels.h
@@ -47,23 +47,34 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
   INHERIT_IMPL_TYPES(Impl);
   typedef FermionOperator<Impl> Base;
    
-public:
-    
-   void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
-		      int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+ public:
+
+  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
+		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
+		    const FermionField &in, FermionField &out, int dag, int interior,int exterior);
+  void DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
+		 DoubledGaugeField &U,
+		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
+  
+  void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+		     int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+ protected:    
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Generic Nc kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+   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);
-   void DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+   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);
-   void DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+   template<int Naik> accelerator_inline
+   void DhopSiteGenericExt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -71,15 +82,18 @@ public:
    ///////////////////////////////////////////////////////////////////////////////////////
    // Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+   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);
-   void DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+   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);
-   void DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+   template<int Naik> accelerator_inline
+   void DhopSiteHandExt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -87,27 +101,10 @@ public:
    ///////////////////////////////////////////////////////////////////////////////////////
    // Asm Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+   void DhopSiteAsm(StencilView &st, 
 		    DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		    SiteSpinor * buf, int LLs, int sU, 
 		    const FermionFieldView &in, FermionFieldView &out,int dag);
-   ///////////////////////////////////////////////////////////////////////////////////////////////////
-   // Generic interface; fan out to right routine
-   ///////////////////////////////////////////////////////////////////////////////////////////////////
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		 SiteSpinor * buf, int LLs, int sU,
-		 const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
-
-   void DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, 
-		    DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		    SiteSpinor * buf, int LLs, int sU,
-		    const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
-
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		 SiteSpinor * buf, int LLs, int sU,
-		 const FermionFieldView &in, FermionFieldView &out, int dag, int interior,int exterior);
   
 public:
 
diff --git a/Grid/qcd/action/fermion/StaggeredVec5dImpl.h b/Grid/qcd/action/fermion/StaggeredVec5dImpl.h
index 2d4de18e..18fe993c 100644
--- a/Grid/qcd/action/fermion/StaggeredVec5dImpl.h
+++ b/Grid/qcd/action/fermion/StaggeredVec5dImpl.h
@@ -113,20 +113,7 @@ public:
       
   inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
   {
-    GridBase *GaugeGrid = U_ds.Grid();
-    thread_for(lidx, GaugeGrid->lSites(),{
-
-	SiteScalarGaugeLink   ScalarU;
-	SiteDoubledGaugeField ScalarUds;
-	
-	Coordinate lcoor;
-	GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
-	peekLocalSite(ScalarUds, U_ds, lcoor);
-	
-	peekLocalSite(ScalarU, U, lcoor);
-	ScalarUds(mu) = ScalarU();
-	
-    });
+    assert(0);
   }
   inline void DoubleStore(GridBase *GaugeGrid,
 			  DoubledGaugeField &UUUds, // for Naik term
diff --git a/Grid/qcd/action/fermion/WilsonCloverFermion.h b/Grid/qcd/action/fermion/WilsonCloverFermion.h
index 4b25d00e..91ad6d6d 100644
--- a/Grid/qcd/action/fermion/WilsonCloverFermion.h
+++ b/Grid/qcd/action/fermion/WilsonCloverFermion.h
@@ -257,15 +257,16 @@ private:
   CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term Dag EO
   CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
 
+ public:
   // eventually these can be compressed into 6x6 blocks instead of the 12x12
   // using the DeGrand-Rossi basis for the gamma matrices
   CloverFieldType fillCloverYZ(const GaugeLinkField &F)
   {
     CloverFieldType T(F.Grid());
     T = Zero();
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
       T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
@@ -281,9 +282,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
     
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v, T,AcceleratorWrite);
+    autoView(F_v, F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = -F_v[i]()();
       T_v[i]()(1, 0) = F_v[i]()();
@@ -299,9 +300,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
       T_v[i]()(1, 1) = timesI(F_v[i]()());
@@ -317,9 +318,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v , T, AcceleratorWrite);
+    autoView( F_v , F, AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = timesI(F_v[i]()());
       T_v[i]()(1, 0) = timesI(F_v[i]()());
@@ -335,9 +336,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
     
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v ,T,AcceleratorWrite);
+    autoView( F_v ,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = -(F_v[i]()());
       T_v[i]()(1, 0) = (F_v[i]()());
@@ -354,9 +355,9 @@ private:
 
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v , T,AcceleratorWrite);
+    autoView( F_v , F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 0) = timesI(F_v[i]()());
       T_v[i]()(1, 1) = timesMinusI(F_v[i]()());
diff --git a/Grid/qcd/action/fermion/WilsonFermion.h b/Grid/qcd/action/fermion/WilsonFermion.h
index 1c4dd3cf..bf8926d0 100644
--- a/Grid/qcd/action/fermion/WilsonFermion.h
+++ b/Grid/qcd/action/fermion/WilsonFermion.h
@@ -50,14 +50,14 @@ public:
   double, nu);
 
   WilsonAnisotropyCoefficients():
-    isAnisotropic(false), 
-    t_direction(Nd-1), 
-    xi_0(1.0), 
+    isAnisotropic(false),
+    t_direction(Nd-1),
+    xi_0(1.0),
     nu(1.0){}
 };
 
 template <class Impl>
-class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic 
+class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic
 {
 public:
   INHERIT_IMPL_TYPES(Impl);
@@ -74,6 +74,20 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
+
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent
@@ -138,7 +152,7 @@ public:
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
                 GridRedBlackCartesian &Hgrid, RealD _mass,
-                const ImplParams &p = ImplParams(), 
+                const ImplParams &p = ImplParams(),
                 const WilsonAnisotropyCoefficients &anis = WilsonAnisotropyCoefficients() );
 
   // DoubleStore impl dependent
@@ -170,9 +184,9 @@ public:
 
   LebesgueOrder Lebesgue;
   LebesgueOrder LebesgueEvenOdd;
-  
+
   WilsonAnisotropyCoefficients anisotropyCoeff;
-  
+
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
   ///////////////////////////////////////////////////////////////
@@ -186,7 +200,7 @@ public:
                            PropagatorField &q_out,
                            PropagatorField &phys_src,
                            Current curr_type,
-                           unsigned int mu, 
+                           unsigned int mu,
                            unsigned int tmin,
 			   unsigned int tmax,
 			   ComplexField &lattice_cmplx);
@@ -196,5 +210,3 @@ typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
 NAMESPACE_END(Grid);
-
-
diff --git a/Grid/qcd/action/fermion/WilsonImpl.h b/Grid/qcd/action/fermion/WilsonImpl.h
index e78023cf..52e1ee00 100644
--- a/Grid/qcd/action/fermion/WilsonImpl.h
+++ b/Grid/qcd/action/fermion/WilsonImpl.h
@@ -106,10 +106,10 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
-    auto out_v= out.View();
-    auto phi_v= phi.View();
-    auto Umu_v= Umu.View();
-    thread_for(sss,out.Grid()->oSites(),{
+    autoView( out_v, out, AcceleratorWrite);
+    autoView( phi_v, phi, AcceleratorRead);
+    autoView( Umu_v, Umu, AcceleratorRead);
+    accelerator_for(sss,out.Grid()->oSites(),1,{
 	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
     });
   }
@@ -191,18 +191,19 @@ public:
     int Ls=Btilde.Grid()->_fdimensions[0];
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    auto tmp_v = tmp.View();
-    auto Btilde_v = Btilde.View();
-    auto Atilde_v = Atilde.View();
-    thread_for(sss,tmp.Grid()->oSites(),{
-      int sU=sss;
-      for(int s=0;s<Ls;s++){
-	int sF = s+Ls*sU;
-	tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
-      }
-    });
+    {
+      autoView( tmp_v , tmp, AcceleratorWrite);
+      autoView( Btilde_v , Btilde, AcceleratorRead);
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,tmp.Grid()->oSites(),1,{
+	  int sU=sss;
+	  for(int s=0;s<Ls;s++){
+	    int sF = s+Ls*sU;
+	    tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
+	  }
+	});
+    }
     PokeIndex<LorentzIndex>(mat,tmp,mu);
-      
   }
 };
 
diff --git a/Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h b/Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
index e9675b36..e79b64dc 100644
--- a/Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
@@ -180,7 +180,7 @@ template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
     std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
     std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls     : " << MooeeInvCalls   << std::endl;
     std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls            : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
     RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
     std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
     std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
@@ -642,7 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-#ifndef GRID_NVCC
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -826,7 +826,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
-#ifndef GRID_NVCC
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
diff --git a/Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h b/Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h
index dbdf134b..d2537ccf 100644
--- a/Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h
+++ b/Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h
@@ -50,9 +50,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(phi , phi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
   auto pdiag = &diag[0];
@@ -93,9 +93,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
 {
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(phi , phi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
   auto pdiag = &diag[0];
@@ -131,8 +131,8 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
 
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
 
   int Ls=this->Ls;
 
@@ -193,8 +193,8 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
   GridBase *grid=psi_i.Grid();
   int Ls=this->Ls;
 
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
 
   auto plee  = & lee [0];
   auto pdee  = & dee [0];
diff --git a/Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h b/Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h
index 034ce642..b54f63ad 100644
--- a/Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h
+++ b/Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h
@@ -65,9 +65,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi, psi_i,CpuRead);
+  autoView(phi, phi_i,CpuRead);
+  autoView(chi, chi_i,CpuWrite);
   int Ls   = this->Ls;
   int LLs  = grid->_rdimensions[0];
   const int nsimd= Simd::Nsimd();
@@ -213,9 +213,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi=psi_i.View();
-  auto phi=phi_i.View();
-  auto chi=chi_i.View();
+  autoView(psi,psi_i,CpuRead);
+  autoView(phi,phi_i,CpuRead);
+  autoView(chi,chi_i,CpuWrite);
   int Ls   = this->Ls;
   int LLs  = grid->_rdimensions[0];
   int nsimd= Simd::Nsimd();
@@ -357,8 +357,8 @@ CayleyFermion5D<Impl>::MooeeInternalAsm(const FermionField &psi_i, FermionField
 					Vector<iSinglet<Simd> > &Matm)
 {
   EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,CpuRead);
+  autoView(chi , chi_i,CpuWrite);
 #ifndef AVX512
   {
     SiteHalfSpinor BcastP;
@@ -535,8 +535,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
   EnableIf<Impl::LsVectorised,int> sfinae=0;
 #ifndef AVX512
   {
-    auto psi = psi_i.View();
-    auto chi = chi_i.View();
+    autoView(psi , psi_i,CpuRead);
+    autoView(chi , chi_i,CpuWrite);
 
     SiteHalfSpinor BcastP;
     SiteHalfSpinor BcastM;
@@ -586,8 +586,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
   }
 #else
   {
-    auto psi = psi_i.View();
-    auto chi = chi_i.View();
+    autoView(psi , psi_i,CpuRead);
+    autoView(chi , chi_i,CpuWrite);
     // pointers
     //  MASK_REGS;
 #define Chi_00 %zmm0
diff --git a/Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h b/Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h
index 46d3fa1f..9a8454ef 100644
--- a/Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h
+++ b/Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h
@@ -46,9 +46,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   chi_i.Checkerboard() = psi_i.Checkerboard();
   int Ls = this->Ls;
   GridBase* grid = psi_i.Grid();
-  auto phi = phi_i.View();
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView( phi , phi_i, AcceleratorRead);
+  autoView( psi , psi_i, AcceleratorRead);
+  autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
   auto pdiag = &diag[0];
   auto pupper = &upper[0];
@@ -82,9 +82,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
   GridBase* grid = psi_i.Grid();
   int Ls = this->Ls;
 
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView( psi , psi_i, AcceleratorRead);
+  autoView( phi , phi_i, AcceleratorRead);
+  autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
   auto pdiag = &diag[0];
   auto pupper = &upper[0];
@@ -116,8 +116,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase* grid = psi_i.Grid();
-  auto psi=psi_i.View();
-  auto chi=chi_i.View();
+  autoView( psi, psi_i, AcceleratorRead);
+  autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto plee  = & this->lee[0];
@@ -172,8 +172,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase* grid = psi_i.Grid();
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView( psi, psi_i, AcceleratorRead);
+  autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto plee  = & this->lee[0];
diff --git a/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h b/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h
index 44a201c1..888691c4 100644
--- a/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h
@@ -221,10 +221,10 @@ void ImprovedStaggeredFermion5D<Impl>::DhopDir(const FermionField &in, FermionFi
 
   Compressor compressor;
   Stencil.HaloExchange(in,compressor);
-  auto Umu_v   = Umu.View();
-  auto UUUmu_v = UUUmu.View();
-  auto in_v    = in.View();
-  auto out_v   = out.View();
+  autoView( Umu_v   ,   Umu, CpuRead);
+  autoView( UUUmu_v , UUUmu, CpuRead);
+  autoView( in_v    ,  in, CpuRead);
+  autoView( out_v   , out, CpuWrite);
   thread_for( ss,Umu.Grid()->oSites(),{
     for(int s=0;s<Ls;s++){
       int sU=ss;
@@ -281,11 +281,9 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
 						    const FermionField &in, FermionField &out,int dag)
 {
-#ifdef GRID_OMP
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-#endif
     DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 
@@ -294,9 +292,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
 								   DoubledGaugeField & U,DoubledGaugeField & UUU,
 								   const FermionField &in, FermionField &out,int dag)
 {
-#ifdef GRID_OMP
   //  assert((dag==DaggerNo) ||(dag==DaggerYes));
-
   Compressor compressor; 
 
   int LLs = in.Grid()->_rdimensions[0];
@@ -305,99 +301,42 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   DhopFaceTime-=usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
+
+  DhopCommTime -=usecond();
+  std::vector<std::vector<CommsRequest_t> > requests;
+  st.CommunicateBegin(requests);
+
   //  st.HaloExchangeOptGather(in,compressor); // Wilson compressor
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
   DhopFaceTime+=usecond();
 
-  double ctime=0;
-  double ptime=0;
-
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Ugly explicit thread mapping introduced for OPA reasons.
+  // Remove explicit thread mapping introduced for OPA reasons.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
+  DhopComputeTime-=usecond();
   {
-    int tid = omp_get_thread_num();
-    int nthreads = omp_get_num_threads();
-    int ncomms = CartesianCommunicator::nCommThreads;
-    if (ncomms == -1) ncomms = 1;
-    assert(nthreads > ncomms);
-    if (tid >= ncomms) {
-      double start = usecond();
-      nthreads -= ncomms;
-      int ttid  = tid - ncomms;
-      int n     = U.Grid()->oSites(); // 4d vol
-      int chunk = n / nthreads;
-      int rem   = n % nthreads;
-      int myblock, myn;
-      if (ttid < rem) {
-        myblock = ttid * chunk + ttid;
-        myn = chunk+1;
-      } else {
-        myblock = ttid*chunk + rem;
-        myn = chunk;
-      }
-
-      // do the compute
-      auto   U_v  =   U.View();
-      auto UUU_v  = UUU.View();
-      auto  in_v  =  in.View();
-      auto out_v  = out.View();
-
-      if (dag == DaggerYes) {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-          int sU = ss;
-	  // Interior = 1; Exterior = 0; must implement for staggered
-          Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<---------
-        }
-      } else {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-	  // Interior = 1; Exterior = 0;
-          int sU = ss;
-          Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<------------
-        }
-      }
-        ptime = usecond() - start;
-    } else {
-      double start = usecond();
-      st.CommunicateThreaded();
-      ctime = usecond() - start;
-    }
+    int interior=1;
+    int exterior=0;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
-  DhopCommTime += ctime;
-  DhopComputeTime+=ptime;
-
-  // First to enter, last to leave timing
-  st.CollateThreads();
+  DhopComputeTime+=usecond();
 
   DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
   DhopFaceTime+=usecond();
 
-  DhopComputeTime2-=usecond();
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
-  auto   U_v  =   U.View();
-  auto UUU_v  = UUU.View();
-  auto  in_v  =  in.View();
-  auto out_v  = out.View();
-  if (dag == DaggerYes) {
-    int sz=st.surface_list.size();
-    thread_for( ss,sz,{
-      int sU = st.surface_list[ss];
-      Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1); //<----------
-    });
-  } else {
-    int sz=st.surface_list.size();
-    thread_for( ss,sz,{
-      int sU = st.surface_list[ss];
-      Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1);//<----------
-    });
+  DhopComputeTime2-=usecond();
+  {
+    int interior=0;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime2+=usecond();
-#else
-  assert(0);
-#endif
-
 }
 
 template<class Impl>
@@ -408,8 +347,6 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   Compressor compressor;
   int LLs = in.Grid()->_rdimensions[0];
 
-
-
  //double t1=usecond();
   DhopTotalTime -= usecond();
   DhopCommTime -= usecond();
@@ -418,28 +355,13 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   
   DhopComputeTime -= usecond();
   // Dhop takes the 4d grid from U, and makes a 5d index for fermion
-  auto   U_v  =   U.View();
-  auto UUU_v  = UUU.View();
-  auto  in_v  =  in.View();
-  auto out_v  = out.View();
-  if (dag == DaggerYes) {
-    thread_for( ss,U.Grid()->oSites(),{
-      int sU=ss;
-      Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), LLs, sU,in_v, out_v);
-    });
-  } else {
-    thread_for( ss,U.Grid()->oSites(),{
-      int sU=ss;
-      Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v);
-    });
+  {
+    int interior=1;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime += usecond();
   DhopTotalTime   += usecond();
- //double t2=usecond();
- //std::cout << __FILE__ << " " << __func__  << " Total Time " << DhopTotalTime << std::endl;
- //std::cout << __FILE__ << " " << __func__  << " Total Time Org " << t2-t1 << std::endl;
- //std::cout << __FILE__ << " " << __func__  << " Comml Time " << DhopCommTime << std::endl;
- //std::cout << __FILE__ << " " << __func__  << " Compute Time " << DhopComputeTime << std::endl;
 
 }
 /*CHANGE END*/
diff --git a/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h b/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h
index 57f4cb89..05d9a17e 100644
--- a/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h
@@ -258,10 +258,10 @@ void ImprovedStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGauge
     ////////////////////////
     // Call the single hop
     ////////////////////////
-    auto U_v   = U.View();
-    auto UUU_v = UUU.View();
-    auto B_v   = B.View();
-    auto Btilde_v   = Btilde.View();
+    autoView( U_v   , U, CpuRead);
+    autoView( UUU_v , UUU, CpuRead);
+    autoView( B_v      , B, CpuWrite);
+    autoView( Btilde_v , Btilde, CpuWrite);
     thread_for(sss,B.Grid()->oSites(),{
       Kernels::DhopDirKernel(st, U_v, UUU_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
     });
@@ -386,10 +386,10 @@ void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionFiel
 
   Compressor compressor;
   Stencil.HaloExchange(in, compressor);
-  auto Umu_v   =   Umu.View();
-  auto UUUmu_v = UUUmu.View();
-  auto in_v    =  in.View();
-  auto out_v   = out.View();
+  autoView( Umu_v   ,   Umu, CpuRead);
+  autoView( UUUmu_v , UUUmu, CpuRead);
+  autoView( in_v    ,  in, CpuRead);
+  autoView( out_v   , out, CpuWrite);
   thread_for( sss, in.Grid()->oSites(),{
     Kernels::DhopDirKernel(Stencil, Umu_v, UUUmu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
   });
@@ -403,11 +403,9 @@ void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder
 						  const FermionField &in,
 						  FermionField &out, int dag) 
 {
-#ifdef GRID_OMP
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-#endif
     DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 template <class Impl>
@@ -417,7 +415,6 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
 								 const FermionField &in,
 								 FermionField &out, int dag) 
 {
-#ifdef GRID_OMP
   Compressor compressor; 
   int len =  U.Grid()->oSites();
 
@@ -426,60 +423,30 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   DhopFaceTime    -= usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
-  st.CommsMergeSHM(compressor);
   DhopFaceTime    += usecond();
 
+  DhopCommTime -=usecond();
+  std::vector<std::vector<CommsRequest_t> > requests;
+  st.CommunicateBegin(requests);
+
+  DhopFaceTime-=usecond();
+  st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
+
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Ugly explicit thread mapping introduced for OPA reasons.
+  // Removed explicit thread comms
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   DhopComputeTime    -= usecond();
-#pragma omp parallel 
   {
-    int tid = omp_get_thread_num();
-    int nthreads = omp_get_num_threads();
-    int ncomms = CartesianCommunicator::nCommThreads;
-    if (ncomms == -1) ncomms = 1;
-    assert(nthreads > ncomms);
-
-    if (tid >= ncomms) {
-      nthreads -= ncomms;
-      int ttid  = tid - ncomms;
-      int n     = len;
-      int chunk = n / nthreads;
-      int rem   = n % nthreads;
-      int myblock, myn;
-      if (ttid < rem) {
-        myblock = ttid * chunk + ttid;
-        myn = chunk+1;
-      } else {
-        myblock = ttid*chunk + rem;
-        myn = chunk;
-      }
-
-      // do the compute
-      auto U_v   = U.View();
-      auto UUU_v = UUU.View();
-      auto in_v  = in.View();
-      auto out_v = out.View();
-      if (dag == DaggerYes) {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-          int sU = ss;
-	  // Interior = 1; Exterior = 0; must implement for staggered
-          Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0); 
-        }
-      } else {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-	  // Interior = 1; Exterior = 0;
-          int sU = ss;
-          Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0);
-        }
-      }
-    } else {
-      st.CommunicateThreaded();
-    }
+    int interior=1;
+    int exterior=0;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime    += usecond();
 
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
+
   // First to enter, last to leave timing
   DhopFaceTime    -= usecond();
   st.CommsMerge(compressor);
@@ -487,28 +454,11 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
 
   DhopComputeTime2    -= usecond();
   {
-    auto U_v   = U.View();
-    auto UUU_v = UUU.View();
-    auto in_v  = in.View();
-    auto out_v = out.View();
-    if (dag == DaggerYes) {
-      int sz=st.surface_list.size();
-      thread_for(ss,sz,{
-	int sU = st.surface_list[ss];
-	Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
-      });
-    } else {
-      int sz=st.surface_list.size();
-      thread_for(ss,sz,{
-	int sU = st.surface_list[ss];
-	Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
-      });
-    }
+    int interior=0;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime2    += usecond();
-#else
-  assert(0);
-#endif
 }
 
 
@@ -528,19 +478,11 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
   st.HaloExchange(in, compressor);
   DhopCommTime    += usecond();
 
-  auto U_v   =   U.View();
-  auto UUU_v = UUU.View();
-  auto in_v  =  in.View();
-  auto out_v = out.View();
   DhopComputeTime -= usecond();
-  if (dag == DaggerYes) {
-    thread_for(sss, in.Grid()->oSites(),{
-      Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
-    });
-  } else {
-    thread_for(sss, in.Grid()->oSites(),{
-      Kernels::DhopSite(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
-    });
+  {
+    int interior=1;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime += usecond();
   DhopTotalTime   += usecond();
diff --git a/Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h b/Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
index f74c7a51..41b9170d 100644
--- a/Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
+++ b/Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
@@ -44,9 +44,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
@@ -84,9 +84,9 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto pm  = this->pm;
   int shift_s = (pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
@@ -132,9 +132,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
@@ -174,9 +174,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
   int shift_s = (this->pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
@@ -226,8 +226,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto plee = & this->lee [0];
   auto pdee = & this->dee [0];
@@ -286,8 +286,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto pm = this->pm;
   auto plee = & this->lee [0];
@@ -354,8 +354,8 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto plee = & this->lee [0];
   auto pdee = & this->dee [0];
@@ -410,8 +410,8 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto pm = this->pm;
diff --git a/Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h b/Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h
new file mode 100644
index 00000000..788e02cf
--- /dev/null
+++ b/Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h
@@ -0,0 +1,499 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+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>
+
+#pragma once 
+
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////
+// Constructor and gauge import
+/////////////////////////////////
+
+template <class Impl>
+NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid, 
+						   RealD _mass,
+						   RealD _c1, RealD _u0,
+						   const ImplParams &p)
+  : Kernels(p),
+    _grid(&Fgrid),
+    _cbgrid(&Hgrid),
+    Stencil(&Fgrid, npoint, Even, directions, displacements,p),
+    StencilEven(&Hgrid, npoint, Even, directions, displacements,p),  // source is Even
+    StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p),  // source is Odd
+    mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
+    Umu(&Fgrid),
+    UmuEven(&Hgrid),
+    UmuOdd(&Hgrid),
+    _tmp(&Hgrid)
+{
+  int vol4;
+  int LLs=1;
+  c1=_c1;
+  u0=_u0;
+  vol4= _grid->oSites();
+  Stencil.BuildSurfaceList(LLs,vol4);
+  vol4= _cbgrid->oSites();
+  StencilEven.BuildSurfaceList(LLs,vol4);
+  StencilOdd.BuildSurfaceList(LLs,vol4);
+}
+
+template <class Impl>
+NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
+						   GridRedBlackCartesian &Hgrid, RealD _mass,
+						   RealD _c1, RealD _u0,
+						   const ImplParams &p)
+  : NaiveStaggeredFermion(Fgrid,Hgrid,_mass,_c1,_u0,p)
+{
+  ImportGauge(_U);
+}
+
+////////////////////////////////////////////////////////////
+// Momentum space propagator should be 
+// https://arxiv.org/pdf/hep-lat/9712010.pdf
+//
+// mom space action.
+//   gamma_mu i ( c1 sin pmu + c2 sin 3 pmu ) + m
+//
+// must track through staggered flavour/spin reduction in literature to 
+// turn to free propagator for the one component chi field, a la page 4/5
+// of above link to implmement fourier based solver.
+////////////////////////////////////////////////////////////
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::CopyGaugeCheckerboards(void)
+{
+  pickCheckerboard(Even, UmuEven,  Umu);
+  pickCheckerboard(Odd,  UmuOdd ,  Umu);
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::ImportGauge(const GaugeField &_U) 
+{
+  GaugeLinkField U(GaugeGrid());
+  DoubledGaugeField _UUU(GaugeGrid());
+  ////////////////////////////////////////////////////////
+  // Double Store should take two fields for Naik and one hop separately.
+  // Discard teh Naik as Naive
+  ////////////////////////////////////////////////////////
+  Impl::DoubleStore(GaugeGrid(), _UUU, Umu, _U, _U );
+
+  ////////////////////////////////////////////////////////
+  // Apply scale factors to get the right fermion Kinetic term
+  // Could pass coeffs into the double store to save work.
+  // 0.5 ( U p(x+mu) - Udag(x-mu) p(x-mu) ) 
+  ////////////////////////////////////////////////////////
+  for (int mu = 0; mu < Nd; mu++) {
+
+    U = PeekIndex<LorentzIndex>(Umu, mu);
+    PokeIndex<LorentzIndex>(Umu, U*( 0.5*c1/u0), mu );
+    
+    U = PeekIndex<LorentzIndex>(Umu, mu+4);
+    PokeIndex<LorentzIndex>(Umu, U*(-0.5*c1/u0), mu+4);
+
+  }
+
+  CopyGaugeCheckerboards();
+}
+
+/////////////////////////////
+// Implement the interface
+/////////////////////////////
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  Dhop(in, out, DaggerNo);
+  axpy(out, mass, in, out);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  Dhop(in, out, DaggerYes);
+  axpy(out, mass, in, out);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerNo);
+  } else {
+    DhopOE(in, out, DaggerNo);
+  }
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerYes);
+  } else {
+    DhopOE(in, out, DaggerYes);
+  }
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  typename FermionField::scalar_type scal(mass);
+  out = scal * in;
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  Mooee(in, out);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  out = (1.0 / (mass)) * in;
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) 
+{
+  out.Checkerboard() = in.Checkerboard();
+  MooeeInv(in, out);
+}
+
+///////////////////////////////////
+// Internal
+///////////////////////////////////
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
+						GaugeField & mat,
+						const FermionField &A, const FermionField &B, int dag) 
+{
+  assert((dag == DaggerNo) || (dag == DaggerYes));
+
+  Compressor compressor;
+
+  FermionField Btilde(B.Grid());
+  FermionField Atilde(B.Grid());
+  Atilde = A;
+
+  st.HaloExchange(B, compressor);
+
+  for (int mu = 0; mu < Nd; mu++) {
+
+    ////////////////////////
+    // Call the single hop
+    ////////////////////////
+    autoView( U_v      , U, CpuRead);
+    autoView( B_v      , B, CpuWrite);
+    autoView( Btilde_v , Btilde, CpuWrite);
+    thread_for(sss,B.Grid()->oSites(),{
+      Kernels::DhopDirKernel(st, U_v, U_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
+    });
+
+    assert(0);// need to figure out the force interface with a blasted three link term.
+    
+  }
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
+
+  conformable(U.Grid(), _grid);
+  conformable(U.Grid(), V.Grid());
+  conformable(U.Grid(), mat.Grid());
+
+  mat.Checkerboard() = U.Checkerboard();
+
+  DerivInternal(Stencil, Umu, mat, U, V, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
+
+  conformable(U.Grid(), _cbgrid);
+  conformable(U.Grid(), V.Grid());
+  conformable(U.Grid(), mat.Grid());
+
+  assert(V.Checkerboard() == Even);
+  assert(U.Checkerboard() == Odd);
+  mat.Checkerboard() = Odd;
+
+  DerivInternal(StencilEven, UmuOdd, mat, U, V, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
+
+  conformable(U.Grid(), _cbgrid);
+  conformable(U.Grid(), V.Grid());
+  conformable(U.Grid(), mat.Grid());
+
+  assert(V.Checkerboard() == Odd);
+  assert(U.Checkerboard() == Even);
+  mat.Checkerboard() = Even;
+
+  DerivInternal(StencilOdd, UmuEven, mat, U, V, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) 
+{
+  DhopCalls+=2;
+  conformable(in.Grid(), _grid);  // verifies full grid
+  conformable(in.Grid(), out.Grid());
+
+  out.Checkerboard() = in.Checkerboard();
+
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) 
+{
+  DhopCalls+=1;
+  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, LebesgueEvenOdd, UmuOdd, in, out, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) 
+{
+  DhopCalls+=1;
+  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, LebesgueEvenOdd, UmuEven, in, out, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) 
+{
+  DhopDir(in, out, dir, disp);
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out) 
+{
+  assert(0); // Not implemented yet
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp) 
+{
+
+  Compressor compressor;
+  Stencil.HaloExchange(in, compressor);
+  autoView( Umu_v   ,  Umu, CpuRead);
+  autoView( in_v    ,  in, CpuRead);
+  autoView( out_v   , out, CpuWrite);
+  //  thread_for( sss, in.Grid()->oSites(),{
+  //    Kernels::DhopDirKernel(Stencil, Umu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
+  //  });
+  assert(0);
+};
+
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
+					       DoubledGaugeField &U,
+					       const FermionField &in,
+					       FermionField &out, int dag) 
+{
+  if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
+  else
+    DhopInternalSerialComms(st,lo,U,in,out,dag);
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
+							      DoubledGaugeField &U,
+							      const FermionField &in,
+							      FermionField &out, int dag) 
+{
+  Compressor compressor; 
+  int len =  U.Grid()->oSites();
+
+  DhopTotalTime   -= usecond();
+
+  DhopFaceTime    -= usecond();
+  st.Prepare();
+  st.HaloGather(in,compressor);
+  DhopFaceTime    += usecond();
+
+  DhopCommTime -=usecond();
+  std::vector<std::vector<CommsRequest_t> > requests;
+  st.CommunicateBegin(requests);
+
+  DhopFaceTime-=usecond();
+  st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Removed explicit thread comms
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime    -= usecond();
+  {
+    int interior=1;
+    int exterior=0;
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
+  }
+  DhopComputeTime    += usecond();
+
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
+
+  // First to enter, last to leave timing
+  DhopFaceTime    -= usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime    -= usecond();
+
+  DhopComputeTime2    -= usecond();
+  {
+    int interior=0;
+    int exterior=1;
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
+  }
+  DhopComputeTime2    += usecond();
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
+							  DoubledGaugeField &U,
+							  const FermionField &in,
+							  FermionField &out, int dag) 
+{
+  assert((dag == DaggerNo) || (dag == DaggerYes));
+
+  DhopTotalTime   -= usecond();
+
+  DhopCommTime    -= usecond();
+  Compressor compressor;
+  st.HaloExchange(in, compressor);
+  DhopCommTime    += usecond();
+
+  DhopComputeTime -= usecond();
+  {
+    int interior=1;
+    int exterior=1;
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
+  }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
+};
+
+  ////////////////////////////////////////////////////////////////
+  // Reporting
+  ////////////////////////////////////////////////////////////////
+template<class Impl>
+void NaiveStaggeredFermion<Impl>::Report(void) 
+{
+  Coordinate latt = _grid->GlobalDimensions();
+  RealD volume = 1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "NaiveStaggeredFermion Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _grid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "NaiveStaggeredFermion Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "NaiveStaggeredFermion StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "NaiveStaggeredFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void NaiveStaggeredFermion<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime   = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
+
+//////////////////////////////////////////////////////// 
+// Conserved current - not yet implemented.
+////////////////////////////////////////////////////////
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
+							      PropagatorField &q_in_2,
+							      PropagatorField &q_out,
+							      PropagatorField &src,
+							      Current curr_type,
+							      unsigned int mu)
+{
+  assert(0);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
+                                                         PropagatorField &q_out,
+                                                         PropagatorField &src,
+                                                         Current curr_type,
+                                                         unsigned int mu, 
+                                                         unsigned int tmin,
+                                              unsigned int tmax,
+					      ComplexField &lattice_cmplx)
+{
+  assert(0);
+
+}
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h b/Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
index 1a13e73a..63fd2a2f 100644
--- a/Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
+++ b/Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
@@ -618,10 +618,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+void StaggeredKernels<Impl>::DhopSiteAsm(StencilView &st,
 					 DoubledGaugeFieldView &U,
 					 DoubledGaugeFieldView &UUU,
-					 SiteSpinor *buf, int LLs,
+					 SiteSpinor *buf, int sF,
 					 int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   assert(0);
@@ -680,12 +680,13 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
   gauge2 =(uint64_t)&UU[sU]( Z );				\
   gauge3 =(uint64_t)&UU[sU]( T ); 
   
+
   // This is the single precision 5th direction vectorised kernel
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
 								    DoubledGaugeFieldView &U,
 								    DoubledGaugeFieldView &UUU,
-								    SiteSpinor *buf, int LLs,
+								    SiteSpinor *buf, int sF,
 								    int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
@@ -702,9 +703,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-   for(int s=0;s<LLs;s++){
+  //   for(int s=0;s<LLs;s++){
 
-    int sF=s+LLs*sU;
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHI(addr0,addr1,addr2,addr3);
@@ -736,10 +738,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView &st, 
 								    DoubledGaugeFieldView &U,
 								    DoubledGaugeFieldView &UUU,
-								    SiteSpinor *buf, int LLs,
+								    SiteSpinor *buf, int sF,
 								    int sU, const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
 #ifdef AVX512
@@ -756,8 +758,9 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHI(addr0,addr1,addr2,addr3);
@@ -821,10 +824,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
   // This is the single precision 5th direction vectorised kernel
 
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st, 
 							       DoubledGaugeFieldView &U,
 							       DoubledGaugeFieldView &UUU,
-							       SiteSpinor *buf, int LLs,
+							       SiteSpinor *buf, int sF,
 							       int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
@@ -841,9 +844,9 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-  for(int s=0;s<LLs;s++){
-    
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHIa(addr0,addr1);
@@ -890,10 +893,10 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilView &st, 
 							       DoubledGaugeFieldView &U,
 							       DoubledGaugeFieldView &UUU,
-							       SiteSpinor *buf, int LLs,
+							       SiteSpinor *buf, int sF,
 							       int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
@@ -910,9 +913,9 @@ template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st,
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-  for(int s=0;s<LLs;s++){
-    
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHIa(addr0,addr1);
diff --git a/Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h b/Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h
index 62370d75..ee53de49 100644
--- a/Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h
+++ b/Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h
@@ -146,9 +146,10 @@ NAMESPACE_BEGIN(Grid);
 
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
 					  DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
-					  SiteSpinor *buf, int LLs, int sU, 
+					  SiteSpinor *buf, int sF, int sU, 
 					  const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
@@ -181,8 +182,9 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
   StencilEntry *SE;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
 
     skew = 0;
     HAND_STENCIL_LEG_BEGIN(Xp,3,skew,even);  
@@ -193,6 +195,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG      (U,Ym,2,skew,odd);   
     HAND_STENCIL_LEG      (U,Zm,1,skew,even);  
     HAND_STENCIL_LEG      (U,Tm,0,skew,odd);  
+    if (Naik) {
     skew = 8;
     HAND_STENCIL_LEG(UUU,Xp,3,skew,even);  
     HAND_STENCIL_LEG(UUU,Yp,2,skew,odd);   
@@ -202,7 +205,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG(UUU,Ym,2,skew,odd);   
     HAND_STENCIL_LEG(UUU,Zm,1,skew,even);  
     HAND_STENCIL_LEG(UUU,Tm,0,skew,odd);  
-    
+    }    
     if ( dag ) {
       result()()(0) = - even_0 - odd_0;
       result()()(1) = - even_1 - odd_1;
@@ -218,9 +221,10 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
 
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st, 
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					     SiteSpinor *buf, int LLs, int sU, 
+					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
@@ -253,8 +257,9 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
   StencilEntry *SE;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
 
     even_0 = Zero();    even_1 = Zero();    even_2 = Zero();
      odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
@@ -268,6 +273,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_INT(U,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_INT(U,Zm,1,skew,even);  
     HAND_STENCIL_LEG_INT(U,Tm,0,skew,odd);  
+    if (Naik) {
     skew = 8;
     HAND_STENCIL_LEG_INT(UUU,Xp,3,skew,even);  
     HAND_STENCIL_LEG_INT(UUU,Yp,2,skew,odd);   
@@ -277,7 +283,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_INT(UUU,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_INT(UUU,Zm,1,skew,even);  
     HAND_STENCIL_LEG_INT(UUU,Tm,0,skew,odd);  
-
+    }
     // Assume every site must be connected to at least one interior point. No 1^4 subvols.
     if ( dag ) {
       result()()(0) = - even_0 - odd_0;
@@ -294,9 +300,10 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
 
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					     SiteSpinor *buf, int LLs, int sU, 
+					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
@@ -329,8 +336,9 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
   StencilEntry *SE;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
 
     even_0 = Zero();    even_1 = Zero();    even_2 = Zero();
      odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
@@ -344,6 +352,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_EXT(U,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_EXT(U,Zm,1,skew,even);  
     HAND_STENCIL_LEG_EXT(U,Tm,0,skew,odd);  
+    if (Naik) {
     skew = 8;
     HAND_STENCIL_LEG_EXT(UUU,Xp,3,skew,even);  
     HAND_STENCIL_LEG_EXT(UUU,Yp,2,skew,odd);   
@@ -353,7 +362,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_EXT(UUU,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_EXT(UUU,Zm,1,skew,even);  
     HAND_STENCIL_LEG_EXT(UUU,Tm,0,skew,odd);  
-
+    }
     // Add sum of all exterior connected stencil legs
     if ( nmu ) { 
       if ( dag ) {
@@ -370,6 +379,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
   }
 }
 
+/*
 #define DHOP_SITE_HAND_INSTANTIATE(IMPL)				\
   template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
 						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
@@ -385,7 +395,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
 						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
 						     SiteSpinor *buf, int LLs, int sU, \
 						     const FermionFieldView &in, FermionFieldView &out, int dag); \
-
+*/
 #undef LOAD_CHI
 
 NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h b/Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h
index d301556c..141725a7 100644
--- a/Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h
@@ -37,9 +37,9 @@ NAMESPACE_BEGIN(Grid);
   if (SE->_is_local ) {						\
     if (SE->_permute) {						\
       chi_p = &chi;						\
-      permute(chi,  in[SE->_offset], ptype);		\
+      permute(chi,  in[SE->_offset], ptype);			\
     } else {							\
-      chi_p = &in[SE->_offset];				\
+      chi_p = &in[SE->_offset];					\
     }								\
   } else {							\
     chi_p = &buf[SE->_offset];					\
@@ -51,15 +51,15 @@ NAMESPACE_BEGIN(Grid);
   if (SE->_is_local ) {						\
     if (SE->_permute) {						\
       chi_p = &chi;						\
-      permute(chi,  in[SE->_offset], ptype);		\
+      permute(chi,  in[SE->_offset], ptype);			\
     } else {							\
-      chi_p = &in[SE->_offset];				\
+      chi_p = &in[SE->_offset];					\
     }								\
   } else if ( st.same_node[Dir] ) {				\
     chi_p = &buf[SE->_offset];					\
   }								\
   if (SE->_is_local || st.same_node[Dir] ) {			\
-    multLink(Uchi, U[sU], *chi_p, Dir);			\
+    multLink(Uchi, U[sU], *chi_p, Dir);				\
   }
 
 #define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink)		\
@@ -67,7 +67,7 @@ NAMESPACE_BEGIN(Grid);
   if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
     chi_p = &buf[SE->_offset];					\
-    multLink(Uchi, U[sU], *chi_p, Dir);			\
+    multLink(Uchi, U[sU], *chi_p, Dir);				\
   }
 
 template <class Impl>
@@ -78,10 +78,12 @@ StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
 // Int, Ext, Int+Ext cases for comms overlap
 ////////////////////////////////////////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st, 
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					     SiteSpinor *buf, int LLs, int sU, 
-					     const FermionFieldView &in, FermionFieldView &out, int dag) {
+					     SiteSpinor *buf, int sF, int sU, 
+					     const FermionFieldView &in, FermionFieldView &out, int dag) 
+{
   const SiteSpinor *chi_p;
   SiteSpinor chi;
   SiteSpinor Uchi;
@@ -89,8 +91,10 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
   int ptype;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+s;
+  //  for(int s=0;s<LLs;s++){
+  //
+  //    int sF=LLs*sU+s;
+  {
     skew = 0;
     GENERIC_STENCIL_LEG(U,Xp,skew,Impl::multLink);
     GENERIC_STENCIL_LEG(U,Yp,skew,Impl::multLinkAdd);
@@ -100,6 +104,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
     GENERIC_STENCIL_LEG(U,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(U,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(U,Tm,skew,Impl::multLinkAdd);
+    if ( Naik ) {
     skew=8;
     GENERIC_STENCIL_LEG(UUU,Xp,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(UUU,Yp,skew,Impl::multLinkAdd);
@@ -109,6 +114,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
     GENERIC_STENCIL_LEG(UUU,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(UUU,Tm,skew,Impl::multLinkAdd);
+    }
     if ( dag ) { 
       Uchi = - Uchi;
     } 
@@ -120,9 +126,10 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
   // Only contributions from interior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-						SiteSpinor *buf, int LLs, int sU, 
+						SiteSpinor *buf, int sF, int sU, 
 						const FermionFieldView &in, FermionFieldView &out,int dag) {
   const SiteSpinor *chi_p;
   SiteSpinor chi;
@@ -131,8 +138,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
   int ptype;
   int skew ;
 
-  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+s;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=LLs*sU+s;
+  {
     skew = 0;
     Uchi=Zero();
     GENERIC_STENCIL_LEG_INT(U,Xp,skew,Impl::multLinkAdd);
@@ -143,6 +151,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_INT(U,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(U,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(U,Tm,skew,Impl::multLinkAdd);
+    if ( Naik ) {
     skew=8;
     GENERIC_STENCIL_LEG_INT(UUU,Xp,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(UUU,Yp,skew,Impl::multLinkAdd);
@@ -152,6 +161,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_INT(UUU,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(UUU,Tm,skew,Impl::multLinkAdd);
+    }
     if ( dag ) {
       Uchi = - Uchi;
     }
@@ -164,9 +174,10 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
   // Only contributions from exterior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik>
+void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-						SiteSpinor *buf, int LLs, int sU,
+						SiteSpinor *buf, int sF, int sU,
 						const FermionFieldView &in, FermionFieldView &out,int dag) {
   const SiteSpinor *chi_p;
   //  SiteSpinor chi;
@@ -176,8 +187,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
   int nmu=0;
   int skew ;
 
-  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+s;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=LLs*sU+s;
+  {
     skew = 0;
     Uchi=Zero();
     GENERIC_STENCIL_LEG_EXT(U,Xp,skew,Impl::multLinkAdd);
@@ -188,6 +200,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_EXT(U,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(U,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(U,Tm,skew,Impl::multLinkAdd);
+    if ( Naik ) {
     skew=8;
     GENERIC_STENCIL_LEG_EXT(UUU,Xp,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Yp,skew,Impl::multLinkAdd);
@@ -197,7 +210,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_EXT(UUU,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
-
+    }
     if ( nmu ) { 
       if ( dag ) { 
 	out[sF] = out[sF] - Uchi;
@@ -211,72 +224,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
 ////////////////////////////////////////////////////////////////////////////////////
 // Driving / wrapping routine to select right kernel
 ////////////////////////////////////////////////////////////////////////////////////
-
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					 SiteSpinor *buf, int LLs, int sU,
-					 const FermionFieldView &in, FermionFieldView &out,
-					 int interior,int exterior)
-{
-  int dag=1;
-  DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
-};
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-				      SiteSpinor *buf, int LLs, int sU,
-				      const FermionFieldView &in, FermionFieldView &out,
-				      int interior,int exterior)
-{
-  int dag=0;
-  DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
-};
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-				      SiteSpinor *buf, int LLs,
-				      int sU, const FermionFieldView &in, FermionFieldView &out,
-				      int dag,int interior,int exterior) 
-{
-  switch(Opt) {
-#ifdef AVX512
-  case OptInlineAsm:
-    if ( interior && exterior ) {
-      DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else { 
-      std::cout << GridLogError << "Cannot overlap comms and compute with Staggered assembly"<<std::endl;
-      assert(0);
-    }
-    break;
-#endif
-  case OptHandUnroll:
-    if ( interior && exterior ) {
-      DhopSiteHand   (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( interior ) {
-      DhopSiteHandInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( exterior ) {
-      DhopSiteHandExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    }
-    break;
-  case OptGeneric:
-    if ( interior && exterior ) {
-      DhopSiteGeneric   (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( interior ) {
-      DhopSiteGenericInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( exterior ) {
-      DhopSiteGenericExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    }
-    break;
-  default:
-    std::cout<<"Oops Opt = "<<Opt<<std::endl;
-    assert(0);
-    break;
-  }
-};
-
-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) 
+void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+					   int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp)
 {
   // Disp should be either +1,-1,+3,-3
   // What about "dag" ?
@@ -285,6 +235,108 @@ void StaggeredKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeFieldVi
   assert(0);
 }
 
+#define KERNEL_CALLNB(A,improved)					\
+  const uint64_t    NN = Nsite*Ls;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      ThisKernel:: template A<improved>(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag); \
+    });
+
+#define KERNEL_CALL(A,improved) KERNEL_CALLNB(A,improved); accelerator_barrier(); 
+
+#define ASM_CALL(A)							\
+  const uint64_t    NN = Nsite*Ls;					\
+  thread_for( ss, NN, {							\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      ThisKernel::A(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag);		\
+  });
+
+template <class Impl>
+void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
+					  DoubledGaugeField &U, DoubledGaugeField &UUU, 
+					  const FermionField &in, FermionField &out, int dag, int interior,int exterior)
+{
+  GridBase *FGrid=in.Grid();  
+  GridBase *UGrid=U.Grid();  
+  typedef StaggeredKernels<Impl> ThisKernel;
+  autoView( UUU_v , UUU, AcceleratorRead);
+  autoView( U_v   ,   U, AcceleratorRead);
+  autoView( in_v  ,  in, AcceleratorRead);
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( st_v  ,  st, AcceleratorRead);
+  SiteSpinor * buf = st.CommBuf();
+    
+  int Ls=1;
+  if(FGrid->Nd()==UGrid->Nd()+1){
+    Ls    = FGrid->_rdimensions[0];
+  }
+  int Nsite = UGrid->oSites();
+
+  if( interior && exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGeneric,1); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,1);    return;}
+    if (Opt == OptInlineAsm  ) {  ASM_CALL(DhopSiteAsm);     return;}
+#endif
+  } else if( interior ) {
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericInt,1); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,1);    return;}
+#endif
+  } else if( exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericExt,1); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,1);    return;}
+#endif
+  }
+  assert(0 && " Kernel optimisation case not covered ");
+}
+template <class Impl>
+void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
+				       DoubledGaugeField &U,
+				       const FermionField &in, FermionField &out, int dag, int interior,int exterior)
+{
+  GridBase *FGrid=in.Grid();  
+  GridBase *UGrid=U.Grid();  
+  typedef StaggeredKernels<Impl> ThisKernel;
+  autoView( UUU_v ,   U, AcceleratorRead);
+  autoView( U_v   ,   U, AcceleratorRead);
+  autoView( in_v  ,  in, AcceleratorRead);
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( st_v  ,  st, AcceleratorRead);
+  SiteSpinor * buf = st.CommBuf();
+
+  int Ls=1;
+  if(FGrid->Nd()==UGrid->Nd()+1){
+    Ls    = FGrid->_rdimensions[0];
+  }
+  int Nsite = UGrid->oSites();
+  
+  if( interior && exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGeneric,0); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,0);    return;}
+#endif
+  } else if( interior ) {
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericInt,0); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,0);    return;}
+#endif
+  } else if( exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericExt,0); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,0);    return;}
+#endif
+  }
+}
+
+
+#undef KERNEL_CALLNB
+#undef KERNEL_CALL
+#undef ASM_CALL
+
 NAMESPACE_END(Grid);
 
 
diff --git a/Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h b/Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
index 36447153..df1bce7c 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
@@ -98,32 +98,35 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   Coordinate lcoor;
   typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
 
-  for (int site = 0; site < lvol; site++)
   {
-    grid->LocalIndexToLocalCoor(site, lcoor);
-    EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-    peekLocalSite(Qx, CloverTerm, lcoor);
-    Qxinv = Zero();
-    //if (csw!=0){
-    for (int j = 0; j < Ns; j++)
-      for (int k = 0; k < Ns; k++)
-        for (int a = 0; a < DimRep; a++)
-          for (int b = 0; b < DimRep; b++){
-	    auto zz =  Qx()(j, k)(a, b);
-            EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
-	  }
-    //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
-
-    EigenInvCloverOp = EigenCloverOp.inverse();
-    //std::cout << EigenInvCloverOp << std::endl;
-    for (int j = 0; j < Ns; j++)
-      for (int k = 0; k < Ns; k++)
-        for (int a = 0; a < DimRep; a++)
-          for (int b = 0; b < DimRep; b++)
-            Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-    //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
-    //  }
-    pokeLocalSite(Qxinv, CloverTermInv, lcoor);
+    autoView(CTv,CloverTerm,CpuRead);
+    autoView(CTIv,CloverTermInv,CpuWrite);
+    for (int site = 0; site < lvol; site++) {
+      grid->LocalIndexToLocalCoor(site, lcoor);
+      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++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++){
+	      auto zz =  Qx()(j, k)(a, b);
+	      EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+	    }
+      //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
+      
+      EigenInvCloverOp = EigenCloverOp.inverse();
+      //std::cout << EigenInvCloverOp << std::endl;
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++)
+	      Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+      //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
+      //  }
+      pokeLocalSite(Qxinv, CTIv, lcoor);
+    }
   }
 
   // Separate the even and odd parts
diff --git a/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h b/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
index 2a202a77..2cc308cc 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
@@ -580,16 +580,21 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
   cosha = (one + W*W + sk) / (abs(W)*2.0);
 
   // FIXME Need a Lattice acosh
-  for(int idx=0;idx<_grid->lSites();idx++){
-    Coordinate lcoor(Nd);
-    Tcomplex cc;
-    //    RealD sgn;
-    _grid->LocalIndexToLocalCoor(idx,lcoor);
-    peekLocalSite(cc,cosha,lcoor);
-    assert((double)real(cc)>=1.0);
-    assert(fabs((double)imag(cc))<=1.0e-15);
-    cc = ScalComplex(::acosh(real(cc)),0.0);
-    pokeLocalSite(cc,a,lcoor);
+
+  {
+    autoView(cosha_v,cosha,CpuRead);
+    autoView(a_v,a,CpuWrite);
+    for(int idx=0;idx<_grid->lSites();idx++){
+      Coordinate lcoor(Nd);
+      Tcomplex cc;
+      //    RealD sgn;
+      _grid->LocalIndexToLocalCoor(idx,lcoor);
+      peekLocalSite(cc,cosha_v,lcoor);
+      assert((double)real(cc)>=1.0);
+      assert(fabs((double)imag(cc))<=1.0e-15);
+      cc = ScalComplex(::acosh(real(cc)),0.0);
+      pokeLocalSite(cc,a_v,lcoor);
+    }
   }
 
   Wea = ( exp( a) * abs(W)  );
@@ -775,17 +780,20 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
   cosha =  (one + W*W + sk) / (abs(W)*2.0);
 
   // FIXME Need a Lattice acosh
+  {
+  autoView(cosha_v,cosha,CpuRead);
+  autoView(a_v,a,CpuWrite);
   for(int idx=0;idx<_grid->lSites();idx++){
     Coordinate lcoor(Nd);
     Tcomplex cc;
     //    RealD sgn;
     _grid->LocalIndexToLocalCoor(idx,lcoor);
-    peekLocalSite(cc,cosha,lcoor);
+    peekLocalSite(cc,cosha_v,lcoor);
     assert((double)real(cc)>=1.0);
     assert(fabs((double)imag(cc))<=1.0e-15);
     cc = ScalComplex(::acosh(real(cc)),0.0);
-    pokeLocalSite(cc,a,lcoor);
-  }
+    pokeLocalSite(cc,a_v,lcoor);
+  }}
   
   Wea = ( exp( a) * abs(W)  );
   Wema= ( exp(-a) * abs(W)  );
diff --git a/Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h b/Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
index 5267e0c1..4977ea68 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
@@ -43,7 +43,7 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
                                    GridRedBlackCartesian &Hgrid, RealD _mass,
                                    const ImplParams &p,
                                    const WilsonAnisotropyCoefficients &anis)
-  : 
+  :
     Kernels(p),
     _grid(&Fgrid),
     _cbgrid(&Hgrid),
@@ -67,11 +67,101 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
     diag_mass = 4.0 + mass;
   }
 
-
+  int vol4;
+  vol4=Fgrid.oSites();
+  Stencil.BuildSurfaceList(1,vol4);
+  vol4=Hgrid.oSites();
+  StencilEven.BuildSurfaceList(1,vol4);
+  StencilOdd.BuildSurfaceList(1,vol4);
 }
 
+template<class Impl>
+void WilsonFermion<Impl>::Report(void)
+{
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+  RealD volume = 1;
+  Coordinate latt = _grid->GlobalDimensions();
+  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of DhopEO Calls   : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion TotalTime   /Calls        : " << DhopTotalTime   / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime    /Calls        : " << DhopCommTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion FaceTime    /Calls        : " << DhopFaceTime    / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime1/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime2/Calls        : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
+
+    // Average the compute time
+    _grid->GlobalSum(DhopComputeTime);
+    DhopComputeTime/=NP;
+    RealD mflops = 1320*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+
+    RealD Fullmflops = 1320*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+
+    // how to count flops here?
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call               ? : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node      ? : " << mflops/NP << std::endl;
+
+    // how to count flops here?
+    RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call (full)        ? : " << Fullmflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node (full) ? : " << Fullmflops/NP << std::endl;  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil"    <<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+  }
+  if ( DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion Stencil     Reporti()"    <<std::endl;  Stencil.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilEven Reporti()"<<std::endl;  StencilEven.Reporti(DhopCalls);
+    std::cout << GridLogMessage << "WilsonFermion StencilOdd  Reporti()" <<std::endl;  StencilOdd.Reporti(DhopCalls);
+  }
+}
+
+template<class Impl>
+void WilsonFermion<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0; // ok
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopComputeTime2= 0;
+  DhopFaceTime    = 0;
+  DhopTotalTime   = 0;
+
+  DerivCalls       = 0; // ok
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+  Stencil.ZeroCountersi();
+  StencilEven.ZeroCountersi();
+  StencilOdd.ZeroCountersi();
+}
+
+
 template <class Impl>
-void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu) 
+void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
 
@@ -102,7 +192,7 @@ void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 /////////////////////////////
 
 template <class Impl>
-void WilsonFermion<Impl>::M(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::M(const FermionField &in, FermionField &out)
 {
   out.Checkerboard() = in.Checkerboard();
   Dhop(in, out, DaggerNo);
@@ -110,7 +200,7 @@ void WilsonFermion<Impl>::M(const FermionField &in, FermionField &out)
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 {
   out.Checkerboard() = in.Checkerboard();
   Dhop(in, out, DaggerYes);
@@ -118,7 +208,7 @@ void WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out)
 {
   if (in.Checkerboard() == Odd) {
     DhopEO(in, out, DaggerNo);
@@ -128,7 +218,7 @@ void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out)
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out)
 {
   if (in.Checkerboard() == Odd) {
     DhopEO(in, out, DaggerYes);
@@ -136,9 +226,9 @@ void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out)
     DhopOE(in, out, DaggerYes);
   }
 }
-  
+
 template <class Impl>
-void WilsonFermion<Impl>::Mooee(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
 {
   out.Checkerboard() = in.Checkerboard();
   typename FermionField::scalar_type scal(diag_mass);
@@ -146,80 +236,80 @@ void WilsonFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
 {
   out.Checkerboard() = in.Checkerboard();
   Mooee(in, out);
 }
 
 template<class Impl>
-void WilsonFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
 {
   out.Checkerboard() = in.Checkerboard();
   out = (1.0/(diag_mass))*in;
 }
-  
+
 template<class Impl>
-void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
 {
   out.Checkerboard() = in.Checkerboard();
   MooeeInv(in,out);
 }
 template<class Impl>
 void WilsonFermion<Impl>::MomentumSpacePropagator(FermionField &out, const FermionField &in,RealD _m,std::vector<double> twist)
-{  
+{
   typedef typename FermionField::vector_type vector_type;
   typedef typename FermionField::scalar_type ScalComplex;
   typedef Lattice<iSinglet<vector_type> > LatComplex;
-  
-  // what type LatticeComplex 
+
+  // what type LatticeComplex
   conformable(_grid,out.Grid());
-  
+
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
     Gamma::Algebra::GammaZ,
     Gamma::Algebra::GammaT
   };
-  
+
   Coordinate latt_size   = _grid->_fdimensions;
-  
+
   FermionField   num  (_grid); num  = Zero();
   LatComplex    wilson(_grid); wilson= Zero();
   LatComplex     one  (_grid); one = ScalComplex(1.0,0.0);
-  
+
   LatComplex denom(_grid); denom= Zero();
-  LatComplex kmu(_grid); 
+  LatComplex kmu(_grid);
   ScalComplex ci(0.0,1.0);
   // momphase = n * 2pi / L
   for(int mu=0;mu<Nd;mu++) {
-    
+
     LatticeCoordinate(kmu,mu);
-    
+
     RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
-    
+
     kmu = TwoPiL * kmu;
     kmu = kmu + TwoPiL * one * twist[mu];//momentum for twisted boundary conditions
-    
+
     wilson = wilson + 2.0*sin(kmu*0.5)*sin(kmu*0.5); // Wilson term
-    
+
     num = num - sin(kmu)*ci*(Gamma(Gmu[mu])*in);    // derivative term
-    
+
     denom=denom + sin(kmu)*sin(kmu);
   }
-  
+
   wilson = wilson + _m;     // 2 sin^2 k/2 + m
-  
+
   num   = num + wilson*in;     // -i gmu sin k + 2 sin^2 k/2 + m
-  
+
   denom= denom+wilson*wilson; // sin^2 k + (2 sin^2 k/2 + m)^2
-  
+
   denom= one/denom;
-  
+
   out = num*denom; // [ -i gmu sin k + 2 sin^2 k/2 + m] / [ sin^2 k + (2 sin^2 k/2 + m)^2 ]
-  
+
 }
-  
+
 
 ///////////////////////////////////
 // Internal
@@ -229,6 +319,7 @@ template <class Impl>
 void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
                                         GaugeField &mat, const FermionField &A,
                                         const FermionField &B, int dag) {
+  DerivCalls++;
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -237,8 +328,11 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
   FermionField Atilde(B.Grid());
   Atilde = A;
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B, compressor);
+  DerivCommTime+=usecond();
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma (1+g)<->(1-g) if dag
@@ -246,6 +340,7 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     int gamma = mu;
     if (!dag) gamma += Nd;
 
+    DerivDhopComputeTime -= usecond();
     int Ls=1;
     Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, B.Grid()->oSites(), B, Btilde, mu, gamma);
 
@@ -253,11 +348,13 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
     // spin trace outer product
     //////////////////////////////////////////////////
     Impl::InsertForce4D(mat, Btilde, Atilde, mu);
+    DerivDhopComputeTime += usecond();
   }
+  DerivComputeTime += usecond();
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) 
+void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   conformable(U.Grid(), _grid);
   conformable(U.Grid(), V.Grid());
@@ -269,13 +366,13 @@ void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, cons
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) 
+void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   conformable(U.Grid(), _cbgrid);
   conformable(U.Grid(), V.Grid());
   //conformable(U.Grid(), mat.Grid()); not general, leaving as a comment (Guido)
   // Motivation: look at the SchurDiff operator
-  
+
   assert(V.Checkerboard() == Even);
   assert(U.Checkerboard() == Odd);
   mat.Checkerboard() = Odd;
@@ -284,7 +381,7 @@ void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, co
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) 
+void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   conformable(U.Grid(), _cbgrid);
   conformable(U.Grid(), V.Grid());
@@ -298,7 +395,7 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) 
+void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
 {
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
@@ -309,7 +406,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) 
+void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
 {
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
@@ -321,7 +418,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag) 
+void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
@@ -333,18 +430,18 @@ void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int d
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) 
+void WilsonFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp)
 {
   DhopDir(in, out, dir, disp);
 }
 template <class Impl>
-void WilsonFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out) 
+void WilsonFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out)
 {
   DhopDirAll(in, out);
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp) 
+void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp)
 {
   Compressor compressor(DaggerNo);
   Stencil.HaloExchange(in, compressor);
@@ -356,12 +453,12 @@ void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int
   DhopDirCalc(in, out, dirdisp, gamma, DaggerNo);
 };
 template <class Impl>
-void WilsonFermion<Impl>::DhopDirAll(const FermionField &in, std::vector<FermionField> &out) 
+void WilsonFermion<Impl>::DhopDirAll(const FermionField &in, std::vector<FermionField> &out)
 {
   Compressor compressor(DaggerNo);
   Stencil.HaloExchange(in, compressor);
 
-  assert((out.size()==8)||(out.size()==9)); 
+  assert((out.size()==8)||(out.size()==9));
   for(int dir=0;dir<Nd;dir++){
     for(int disp=-1;disp<=1;disp+=2){
 
@@ -374,7 +471,7 @@ void WilsonFermion<Impl>::DhopDirAll(const FermionField &in, std::vector<Fermion
   }
 }
 template <class Impl>
-void WilsonFermion<Impl>::DhopDirCalc(const FermionField &in, FermionField &out,int dirdisp, int gamma, int dag) 
+void WilsonFermion<Impl>::DhopDirCalc(const FermionField &in, FermionField &out,int dirdisp, int gamma, int dag)
 {
   int Ls=1;
   uint64_t Nsite=in.oSites();
@@ -385,22 +482,23 @@ template <class Impl>
 void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        DoubledGaugeField &U,
                                        const FermionField &in,
-                                       FermionField &out, int dag) 
+                                       FermionField &out, int dag)
 {
+  DhopTotalTime-=usecond();
 #ifdef GRID_OMP
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
-#endif 
+#endif
     DhopInternalSerial(st,lo,U,in,out,dag);
-
+  DhopTotalTime+=usecond();
 }
 
 template <class Impl>
 void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 						      DoubledGaugeField &U,
 						      const FermionField &in,
-						      FermionField &out, int dag) 
+						      FermionField &out, int dag)
 {
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
@@ -412,38 +510,53 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
   /////////////////////////////
   std::vector<std::vector<CommsRequest_t> > requests;
   st.Prepare();
+  DhopFaceTime-=usecond();
   st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
+
+  DhopCommTime -=usecond();
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);
+  DhopFaceTime+=usecond();
 
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt;
+  DhopComputeTime-=usecond();
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
   } else {
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
-  } 
+  }
+  DhopComputeTime+=usecond();
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
+  DhopCommTime   +=usecond();
+
+  DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
+
+  DhopComputeTime2-=usecond();
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   } else {
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
   }
+  DhopComputeTime2+=usecond();
 };
 
 
@@ -451,24 +564,28 @@ template <class Impl>
 void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
                                        DoubledGaugeField &U,
                                        const FermionField &in,
-                                       FermionField &out, int dag) 
+                                       FermionField &out, int dag)
 {
   assert((dag == DaggerNo) || (dag == DaggerYes));
   Compressor compressor(dag);
+  DhopCommTime-=usecond();
   st.HaloExchange(in, compressor);
+  DhopCommTime+=usecond();
 
+  DhopComputeTime-=usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   } else {
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
   }
+  DhopComputeTime+=usecond();
 };
 /*Change ends */
 
 /*******************************************************************************
  * Conserved current utilities for Wilson fermions, for contracting propagators
- * to make a conserved current sink or inserting the conserved current 
+ * to make a conserved current sink or inserting the conserved current
  * sequentially.
  ******************************************************************************/
 template <class Impl>
@@ -483,103 +600,23 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
   conformable(_grid, q_in_1.Grid());
   conformable(_grid, q_in_2.Grid());
   conformable(_grid, q_out.Grid());
-#if 0
-  PropagatorField tmp1(_grid), tmp2(_grid);
-  q_out = Zero();
-
-  // Forward, need q1(x + mu), q2(x). Backward, need q1(x), q2(x + mu).
-  // Inefficient comms method but not performance critical.
-  tmp1 = Cshift(q_in_1, mu, 1);
-  tmp2 = Cshift(q_in_2, mu, 1);
-  auto tmp1_v  =  tmp1.View();
-  auto tmp2_v  =  tmp2.View();
-  auto q_in_1_v=q_in_1.View();
-  auto q_in_2_v=q_in_2.View();
-  auto q_out_v = q_out.View();
-  auto Umu_v   =   Umu.View();
-  thread_for(sU, Umu.Grid()->oSites(),{
-      Kernels::ContractConservedCurrentSiteFwd(tmp1_v[sU],
-					       q_in_2_v[sU],
-					       q_out_v[sU],
-					       Umu_v, sU, mu);
-      Kernels::ContractConservedCurrentSiteBwd(q_in_1_v[sU],
-					       tmp2_v[sU],
-					       q_out_v[sU],
-					       Umu_v, sU, mu);
-  });
-#else
-#endif
+  assert(0);
 }
 
 
 template <class Impl>
-void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in, 
+void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
                                               PropagatorField &q_out,
                                               PropagatorField &src,
                                               Current curr_type,
                                               unsigned int mu,
-                                              unsigned int tmin, 
+                                              unsigned int tmin,
                                               unsigned int tmax,
 					      ComplexField &lattice_cmplx)
 {
   conformable(_grid, q_in.Grid());
   conformable(_grid, q_out.Grid());
-#if 0
-
-  //  Lattice<iSinglet<Simd>> ph(_grid), coor(_grid);
-  Complex i(0.0,1.0);
-  PropagatorField tmpFwd(_grid), tmpBwd(_grid), tmp(_grid);
-  unsigned int tshift = (mu == Tp) ? 1 : 0;
-  unsigned int LLt    = GridDefaultLatt()[Tp];
-
-  q_out = Zero();
-  LatticeInteger coords(_grid);
-  LatticeCoordinate(coords, Tp);
-
-  // Need q(x + mu) and q(x - mu).
-  tmp    = Cshift(q_in, mu, 1);
-  tmpFwd = tmp*lattice_cmplx;
-  tmp    = lattice_cmplx*q_in;
-  tmpBwd = Cshift(tmp, mu, -1);
-
-  auto coords_v = coords.View();
-  auto tmpFwd_v = tmpFwd.View();
-  auto tmpBwd_v = tmpBwd.View();
-  auto Umu_v    = Umu.View();
-  auto q_out_v  = q_out.View();
-
-  thread_for(sU, Umu.Grid()->oSites(), {
-
-    // Compute the sequential conserved current insertion only if our simd
-    // object contains a timeslice we need.
-    vPredicate t_mask;
-    t_mask() = ((coords_v[sU] >= tmin) && (coords_v[sU] <= tmax));
-    Integer timeSlices = Reduce(t_mask());
-
-    if (timeSlices > 0) {
-      Kernels::SeqConservedCurrentSiteFwd(tmpFwd_v[sU], 
-					  q_out_v[sU], 
-					  Umu_v, sU, mu, t_mask);
-    }
-
-    // Repeat for backward direction.
-    t_mask()     = ((coords_v[sU] >= (tmin + tshift)) && 
-		    (coords_v[sU] <= (tmax + tshift)));
-    
-    //if tmax = LLt-1 (last timeslice) include timeslice 0 if the time is shifted (mu=3)	
-    unsigned int t0 = 0;
-    if((tmax==LLt-1) && (tshift==1)) t_mask() = (t_mask() || (coords_v[sU] == t0 ));
-    
-    timeSlices = Reduce(t_mask());
-
-    if (timeSlices > 0) {
-      Kernels::SeqConservedCurrentSiteBwd(tmpBwd_v[sU], 
-					  q_out_v[sU], 
-					  Umu_v, sU, mu, t_mask);
-    }
-  });
-#else
-#endif
+  assert(0);
 }
 
 NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h b/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
new file mode 100644
index 00000000..2e587dfa
--- /dev/null
+++ b/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
@@ -0,0 +1,574 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+
+
+    Source file: ./lib/qcd/action/fermion/WilsonKernelsAsmA64FX.h
+
+    Copyright (C) 2020
+
+Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+//#if defined(A64FXASM)
+#if defined(A64FX)
+
+// safety include
+#include <arm_sve.h>
+
+// undefine everything related to kernels
+#include <simd/Fujitsu_A64FX_undef.h>
+
+// enable A64FX body
+#define WILSONKERNELSASMBODYA64FX
+//#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
+
+    ///////////////////////////////////////////////////////////
+    // If we are A64FX specialise the single precision routine
+    ///////////////////////////////////////////////////////////
+#if defined(DSLASHINTRIN)
+//#pragma message ("A64FX Dslash: intrin")
+#include <simd/Fujitsu_A64FX_intrin_single.h>
+#else
+#pragma message ("A64FX Dslash: asm")
+#include <simd/Fujitsu_A64FX_asm_single.h>
+#endif
+
+/// Switch off the 5d vectorised code optimisations
+#undef DWFVEC5D
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, undag Kernel, single
+/////////////////////////////////////////////////////////////////
+#undef KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+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
+
+template<> void
+WilsonKernels<ZWilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, dag Kernel, single
+/////////////////////////////////////////////////////////////////
+#define KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+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
+
+template<> void
+WilsonKernels<ZWilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+#include <simd/Fujitsu_A64FX_undef.h>
+
+///////////////////////////////////////////////////////////
+// If we are A64FX specialise the double precision routine
+///////////////////////////////////////////////////////////
+
+#if defined(DSLASHINTRIN)
+#include <simd/Fujitsu_A64FX_intrin_double.h>
+#else
+#include <simd/Fujitsu_A64FX_asm_double.h>
+#endif
+
+// former KNL
+//#define MAYBEPERM(A,perm) if (perm) { A ; }
+//#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
+//#define COMPLEX_SIGNS(isigns) vComplexD *isigns = &signsD[0];
+
+
+#define INTERIOR_AND_EXTERIOR
+#undef  INTERIOR
+#undef  EXTERIOR
+
+/////////////////////////////////////////////////////////////////
+// XYZT vectorised, undag Kernel, double
+/////////////////////////////////////////////////////////////////
+#undef KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+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
+
+template<> void
+WilsonKernels<ZWilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+/////////////////////////////////////////////////////////////////
+#define KERNEL_DAG
+#define INTERIOR_AND_EXTERIOR
+#undef INTERIOR
+#undef EXTERIOR
+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
+
+template<> void
+WilsonKernels<ZWilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
+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
+
+template<> void
+WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
+						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
+#if defined (WILSONKERNELSASMBODYA64FX)
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
+#else
+#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
+#endif
+
+template<> void
+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
+
+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
diff --git a/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h b/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h
new file mode 100644
index 00000000..406e5c25
--- /dev/null
+++ b/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h
@@ -0,0 +1,380 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: WilsonKernelsAsmBodyA64FX.h
+
+    Copyright (C) 2020
+
+Author:  Nils Meyer  <nils.meyer@ur.de>  Regensburg University
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifdef KERNEL_DAG
+#define DIR0_PROJ    XP_PROJ
+#define DIR1_PROJ    YP_PROJ
+#define DIR2_PROJ    ZP_PROJ
+#define DIR3_PROJ    TP_PROJ
+#define DIR4_PROJ    XM_PROJ
+#define DIR5_PROJ    YM_PROJ
+#define DIR6_PROJ    ZM_PROJ
+#define DIR7_PROJ    TM_PROJ
+#define DIR0_RECON   XP_RECON
+#define DIR1_RECON   YP_RECON_ACCUM
+#define DIR2_RECON   ZP_RECON_ACCUM
+#define DIR3_RECON   TP_RECON_ACCUM
+#define DIR4_RECON   XM_RECON_ACCUM
+#define DIR5_RECON   YM_RECON_ACCUM
+#define DIR6_RECON   ZM_RECON_ACCUM
+#define DIR7_RECON   TM_RECON_ACCUM
+#else
+#define DIR0_PROJ    XM_PROJ
+#define DIR1_PROJ    YM_PROJ
+#define DIR2_PROJ    ZM_PROJ
+#define DIR3_PROJ    TM_PROJ
+#define DIR4_PROJ    XP_PROJ
+#define DIR5_PROJ    YP_PROJ
+#define DIR6_PROJ    ZP_PROJ
+#define DIR7_PROJ    TP_PROJ
+#define DIR0_RECON   XM_RECON
+#define DIR1_RECON   YM_RECON_ACCUM
+#define DIR2_RECON   ZM_RECON_ACCUM
+#define DIR3_RECON   TM_RECON_ACCUM
+#define DIR4_RECON   XP_RECON_ACCUM
+#define DIR5_RECON   YP_RECON_ACCUM
+#define DIR6_RECON   ZP_RECON_ACCUM
+#define DIR7_RECON   TP_RECON_ACCUM
+#endif
+
+//using namespace std;
+
+#undef SHOW
+//#define SHOW
+
+#undef WHERE
+
+#ifdef INTERIOR_AND_EXTERIOR
+#define WHERE "INT_AND_EXT"
+#endif
+
+#ifdef INTERIOR
+#define WHERE "INT"
+#endif
+
+#ifdef EXTERIOR
+#define WHERE "EXT"
+#endif
+
+//#pragma message("here")
+
+
+
+////////////////////////////////////////////////////////////////////////////////
+// Comms then compute kernel
+////////////////////////////////////////////////////////////////////////////////
+#ifdef INTERIOR_AND_EXTERIOR
+
+#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+      basep = st.GetPFInfo(nent,plocal); nent++;			\
+      if ( local ) {							            \
+    LOAD_CHIMU(base);                                       \
+    LOAD_TABLE(PERMUTE_DIR);                                \
+    PROJ;							                        \
+    MAYBEPERM(PERMUTE_DIR,perm);					        \
+      } else {								                \
+	LOAD_CHI(base);							                \
+      }									                    \
+      base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++;	\
+    MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    PREFETCH_CHIMU_L2(basep);                               \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+    MULT_2SPIN_2;					                        \
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
+  PREFETCH1_CHIMU(base);						            \
+  ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
+
+#define RESULT(base,basep) SAVE_RESULT(base,basep);
+
+#endif
+
+////////////////////////////////////////////////////////////////////////////////
+// Pre comms kernel -- prefetch like normal because it is mostly right
+////////////////////////////////////////////////////////////////////////////////
+#ifdef INTERIOR
+
+#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+      basep = st.GetPFInfo(nent,plocal); nent++;			\
+      if ( local ) {							            \
+    LOAD_CHIMU(base);                                       \
+    LOAD_TABLE(PERMUTE_DIR);                                \
+    PROJ;							                        \
+    MAYBEPERM(PERMUTE_DIR,perm);					        \
+      }else if ( st.same_node[Dir] ) {LOAD_CHI(base);}	    \
+      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);						\
+  ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
+
+#define RESULT(base,basep) SAVE_RESULT(base,basep);
+
+#endif
+////////////////////////////////////////////////////////////////////////////////
+// Post comms kernel
+////////////////////////////////////////////////////////////////////////////////
+#ifdef EXTERIOR
+
+
+#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)			\
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
+  if((!local)&&(!st.same_node[Dir]) ) {					    \
+    LOAD_CHI(base);							                \
+    MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+    MULT_2SPIN_2;					                        \
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    nmu++;								                    \
+  }
+
+#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)	    \
+  nmu=0;								                    \
+  base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;\
+  if((!local)&&(!st.same_node[Dir]) ) {					    \
+    LOAD_CHI(base);							                \
+    MULT_2SPIN_1(Dir);					                    \
+    PREFETCH_CHIMU(base);                                   \
+    /* PREFETCH_GAUGE_L1(NxtDir); */                        \
+    MULT_2SPIN_2;					                        \
+    if (s == 0) {                                           \
+      if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
+    }                                                       \
+    RECON;								                    \
+    nmu++;								                    \
+  }
+
+#define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
+
+#endif
+{
+  int nmu;
+  int local,perm, ptype;
+  uint64_t base;
+  uint64_t basep;
+  const uint64_t plocal =(uint64_t) & in[0];
+
+  MASK_REGS;
+  int nmax=U.oSites();
+  for(int site=0;site<Ns;site++) {
+#ifndef EXTERIOR
+    //    int sU =lo.Reorder(ssU);
+    int sU =ssU;
+    int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
+    //    int sUn=lo.Reorder(ssn);
+    int sUn=ssn;
+    LOCK_GAUGE(0);
+#else
+    int sU =ssU;
+    int ssn=ssU+1;     if(ssn>=nmax) ssn=0;
+    int sUn=ssn;
+#endif
+    for(int s=0;s<Ls;s++) {
+      ss =sU*Ls+s;
+      ssn=sUn*Ls+s;
+      int  ent=ss*8;// 2*Ndim
+      int nent=ssn*8;
+
+      uint64_t delta_base, delta_base_p;
+
+   ASM_LEG_XP(Xp,Yp,PERMUTE_DIR3,DIR0_PROJ,DIR0_RECON);
+
+#ifdef SHOW
+      float rescale = 64. * 12.;
+      std::cout << "=================================================================" << std::endl;
+      std::cout << "ss = " << ss << "   ssn = " << ssn << std::endl;
+      std::cout << "sU = " << sU << "   ssU = " << ssU << std::endl;
+      std::cout << " " << std::endl;
+
+
+      std::cout << "Dir = " << Xp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Xp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Yp,Zp,PERMUTE_DIR2,DIR1_PROJ,DIR1_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Yp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Yp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Zp,Tp,PERMUTE_DIR1,DIR2_PROJ,DIR2_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Zp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Zp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Tp,Xm,PERMUTE_DIR0,DIR3_PROJ,DIR3_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Tp << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Tp] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Xm,Ym,PERMUTE_DIR3,DIR4_PROJ,DIR4_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Xm << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Xm] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJ,DIR5_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Ym << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Ym] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJ,DIR6_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Zm << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Zm] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+      ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJ,DIR7_RECON);
+
+#ifdef SHOW
+      std::cout << "Dir = " << Tm << "        "  << WHERE<< std::endl;
+
+      std::cout << "ent  nent  local  perm       = " << ent << "  " << nent << "  " << local << "  "  << perm << std::endl;
+      std::cout << "st.same_node[Dir] = " << st.same_node[Tm] << std::endl;
+      std::cout << "base              = " << (base - plocal)/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+#ifdef EXTERIOR
+      if (nmu==0) break;
+      //      if (nmu!=0) std::cout << "EXT "<<sU<<std::endl;
+#endif
+      base = (uint64_t) &out[ss];
+      basep= st.GetPFInfo(nent,plocal); ent++;
+      basep = (uint64_t) &out[ssn];
+      RESULT(base,basep);
+
+#ifdef SHOW
+      std::cout << "Dir = FINAL        " <<  WHERE<< std::endl;;
+
+      base_ss = base;
+      std::cout << "base              = " << (base - (uint64_t) &out[0])/rescale << std::endl;
+      std::cout << "Basep             = " << (basep - plocal)/rescale << std::endl;
+      //printf("U                 = %llu\n", (uint64_t)&[sU](Dir));
+      std::cout << "----------------------------------------------------" << std::endl;
+#endif
+
+    }
+    ssU++;
+    UNLOCK_GAUGE(0);
+  }
+}
+
+#undef DIR0_PROJ
+#undef DIR1_PROJ
+#undef DIR2_PROJ
+#undef DIR3_PROJ
+#undef DIR4_PROJ
+#undef DIR5_PROJ
+#undef DIR6_PROJ
+#undef DIR7_PROJ
+#undef DIR0_RECON
+#undef DIR1_RECON
+#undef DIR2_RECON
+#undef DIR3_RECON
+#undef DIR4_RECON
+#undef DIR5_RECON
+#undef DIR6_RECON
+#undef DIR7_RECON
+#undef ASM_LEG
+#undef ASM_LEG_XP
+#undef RESULT
diff --git a/Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h.debug b/Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h.debug
new file mode 100644
index 00000000..6f3edbb5
--- /dev/null
+++ b/Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h.debug
@@ -0,0 +1,943 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonKernelsHand.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+#pragma once
+
+#include <Grid/qcd/action/fermion/FermionCore.h>
+
+
+#undef LOAD_CHIMU
+#undef LOAD_CHI
+#undef MULT_2SPIN
+#undef PERMUTE_DIR
+#undef XP_PROJ
+#undef YP_PROJ
+#undef ZP_PROJ
+#undef TP_PROJ
+#undef XM_PROJ
+#undef YM_PROJ
+#undef ZM_PROJ
+#undef TM_PROJ
+#undef XP_RECON
+#undef XP_RECON_ACCUM
+#undef XM_RECON
+#undef XM_RECON_ACCUM
+#undef YP_RECON_ACCUM
+#undef YM_RECON_ACCUM
+#undef ZP_RECON_ACCUM
+#undef ZM_RECON_ACCUM
+#undef TP_RECON_ACCUM
+#undef TM_RECON_ACCUM
+#undef ZERO_RESULT
+#undef Chimu_00
+#undef Chimu_01
+#undef Chimu_02
+#undef Chimu_10
+#undef Chimu_11
+#undef Chimu_12
+#undef Chimu_20
+#undef Chimu_21
+#undef Chimu_22
+#undef Chimu_30
+#undef Chimu_31
+#undef Chimu_32
+#undef HAND_STENCIL_LEG
+#undef HAND_STENCIL_LEG_INT
+#undef HAND_STENCIL_LEG_EXT
+#undef HAND_RESULT
+#undef HAND_RESULT_INT
+#undef HAND_RESULT_EXT
+
+#define REGISTER
+
+#define LOAD_CHIMU \
+  {const SiteSpinor & ref (in[offset]);	\
+    Chimu_00=ref()(0)(0);\
+    Chimu_01=ref()(0)(1);\
+    Chimu_02=ref()(0)(2);\
+    Chimu_10=ref()(1)(0);\
+    Chimu_11=ref()(1)(1);\
+    Chimu_12=ref()(1)(2);\
+    Chimu_20=ref()(2)(0);\
+    Chimu_21=ref()(2)(1);\
+    Chimu_22=ref()(2)(2);\
+    Chimu_30=ref()(3)(0);\
+    Chimu_31=ref()(3)(1);\
+    Chimu_32=ref()(3)(2);\
+    std::cout << std::endl << "DEBUG -- LOAD_CHIMU" << std::endl; \
+    std::cout << "Chimu_00 -- " <<  Chimu_00 << std::endl; \
+    std::cout << "Chimu_01 -- " <<  Chimu_01 << std::endl; \
+    std::cout << "Chimu_02 -- " <<  Chimu_02 << std::endl; \
+    std::cout << "Chimu_10 -- " <<  Chimu_10 << std::endl; \
+    std::cout << "Chimu_11 -- " <<  Chimu_11 << std::endl; \
+    std::cout << "Chimu_12 -- " <<  Chimu_12 << std::endl; \
+    std::cout << "Chimu_20 -- " <<  Chimu_20 << std::endl; \
+    std::cout << "Chimu_21 -- " <<  Chimu_21 << std::endl; \
+    std::cout << "Chimu_22 -- " <<  Chimu_22 << std::endl; \
+    std::cout << "Chimu_30 -- " <<  Chimu_30 << std::endl; \
+    std::cout << "Chimu_31 -- " <<  Chimu_31 << std::endl; \
+    std::cout << "Chimu_32 -- " <<  Chimu_32 << std::endl; \
+}
+
+#define LOAD_CHI\
+  {const SiteHalfSpinor &ref(buf[offset]);	\
+    Chi_00 = ref()(0)(0);\
+    Chi_01 = ref()(0)(1);\
+    Chi_02 = ref()(0)(2);\
+    Chi_10 = ref()(1)(0);\
+    Chi_11 = ref()(1)(1);\
+    Chi_12 = ref()(1)(2);\
+    std::cout << std::endl << "DEBUG -- LOAD_CHI" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl; \
+  }
+
+// To splat or not to splat depends on the implementation
+#define MULT_2SPIN(A)\
+  {auto & ref(U[sU](A));			\
+   Impl::loadLinkElement(U_00,ref()(0,0));	\
+   Impl::loadLinkElement(U_10,ref()(1,0));	\
+   Impl::loadLinkElement(U_20,ref()(2,0));	\
+   Impl::loadLinkElement(U_01,ref()(0,1));	\
+   Impl::loadLinkElement(U_11,ref()(1,1));	\
+   Impl::loadLinkElement(U_21,ref()(2,1));	\
+    UChi_00 = U_00*Chi_00;\
+    UChi_10 = U_00*Chi_10;\
+    UChi_01 = U_10*Chi_00;\
+    UChi_11 = U_10*Chi_10;\
+    UChi_02 = U_20*Chi_00;\
+    UChi_12 = U_20*Chi_10;\
+    UChi_00+= U_01*Chi_01;\
+    UChi_10+= U_01*Chi_11;\
+    UChi_01+= U_11*Chi_01;\
+    UChi_11+= U_11*Chi_11;\
+    UChi_02+= U_21*Chi_01;\
+    UChi_12+= U_21*Chi_11;\
+    Impl::loadLinkElement(U_00,ref()(0,2));	\
+    Impl::loadLinkElement(U_10,ref()(1,2));	\
+    Impl::loadLinkElement(U_20,ref()(2,2));	\
+    UChi_00+= U_00*Chi_02;\
+    UChi_10+= U_00*Chi_12;\
+    UChi_01+= U_10*Chi_02;\
+    UChi_11+= U_10*Chi_12;\
+    UChi_02+= U_20*Chi_02;\
+    UChi_12+= U_20*Chi_12;\
+    std::cout << std::endl << "DEBUG -- MULT_2SPIN" << std::endl; \
+    std::cout << "UChi_00 -- " <<  UChi_00 << std::endl; \
+    std::cout << "UChi_01 -- " <<  UChi_01 << std::endl; \
+    std::cout << "UChi_02 -- " <<  UChi_02 << std::endl; \
+    std::cout << "UChi_10 -- " <<  UChi_10 << std::endl; \
+    std::cout << "UChi_11 -- " <<  UChi_11 << std::endl; \
+    std::cout << "UChi_12 -- " <<  UChi_12 << std::endl; \
+    }
+
+
+#define PERMUTE_DIR(dir)			\
+std::cout << std::endl << "DEBUG -- PERM PRE" << std::endl; \
+std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+std::cout << "Chi_12 -- " <<  Chi_12 << std::endl; \
+      permute##dir(Chi_00,Chi_00);\
+      permute##dir(Chi_01,Chi_01);\
+      permute##dir(Chi_02,Chi_02);\
+      permute##dir(Chi_10,Chi_10);\
+      permute##dir(Chi_11,Chi_11);\
+      permute##dir(Chi_12,Chi_12);\
+  std::cout << std::endl << "DEBUG -- PERM POST" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+//      hspin(0)=fspin(0)+timesI(fspin(3));
+//      hspin(1)=fspin(1)+timesI(fspin(2));
+#define XP_PROJ \
+    Chi_00 = Chimu_00+timesI(Chimu_30);\
+    Chi_01 = Chimu_01+timesI(Chimu_31);\
+    Chi_02 = Chimu_02+timesI(Chimu_32);\
+    Chi_10 = Chimu_10+timesI(Chimu_20);\
+    Chi_11 = Chimu_11+timesI(Chimu_21);\
+    Chi_12 = Chimu_12+timesI(Chimu_22);\
+    std::cout << std::endl << "DEBUG -- XP_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define YP_PROJ \
+    Chi_00 = Chimu_00-Chimu_30;\
+    Chi_01 = Chimu_01-Chimu_31;\
+    Chi_02 = Chimu_02-Chimu_32;\
+    Chi_10 = Chimu_10+Chimu_20;\
+    Chi_11 = Chimu_11+Chimu_21;\
+    Chi_12 = Chimu_12+Chimu_22;\
+    std::cout << std::endl << "DEBUG -- YP_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define ZP_PROJ \
+  Chi_00 = Chimu_00+timesI(Chimu_20);		\
+  Chi_01 = Chimu_01+timesI(Chimu_21);		\
+  Chi_02 = Chimu_02+timesI(Chimu_22);		\
+  Chi_10 = Chimu_10-timesI(Chimu_30);		\
+  Chi_11 = Chimu_11-timesI(Chimu_31);		\
+  Chi_12 = Chimu_12-timesI(Chimu_32);\
+  std::cout << std::endl << "DEBUG -- ZP_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define TP_PROJ \
+  Chi_00 = Chimu_00+Chimu_20;		\
+  Chi_01 = Chimu_01+Chimu_21;		\
+  Chi_02 = Chimu_02+Chimu_22;		\
+  Chi_10 = Chimu_10+Chimu_30;		\
+  Chi_11 = Chimu_11+Chimu_31;		\
+  Chi_12 = Chimu_12+Chimu_32;\
+  std::cout << std::endl << "DEBUG -- TP_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+
+//      hspin(0)=fspin(0)-timesI(fspin(3));
+//      hspin(1)=fspin(1)-timesI(fspin(2));
+#define XM_PROJ \
+    Chi_00 = Chimu_00-timesI(Chimu_30);\
+    Chi_01 = Chimu_01-timesI(Chimu_31);\
+    Chi_02 = Chimu_02-timesI(Chimu_32);\
+    Chi_10 = Chimu_10-timesI(Chimu_20);\
+    Chi_11 = Chimu_11-timesI(Chimu_21);\
+    Chi_12 = Chimu_12-timesI(Chimu_22);\
+    std::cout << std::endl << "DEBUG -- XM_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define YM_PROJ \
+    Chi_00 = Chimu_00+Chimu_30;\
+    Chi_01 = Chimu_01+Chimu_31;\
+    Chi_02 = Chimu_02+Chimu_32;\
+    Chi_10 = Chimu_10-Chimu_20;\
+    Chi_11 = Chimu_11-Chimu_21;\
+    Chi_12 = Chimu_12-Chimu_22;\
+    std::cout << std::endl << "DEBUG -- YM_PROJ" << std::endl; \
+    std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+    std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+    std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+    std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+    std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+    std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define ZM_PROJ \
+  Chi_00 = Chimu_00-timesI(Chimu_20);		\
+  Chi_01 = Chimu_01-timesI(Chimu_21);		\
+  Chi_02 = Chimu_02-timesI(Chimu_22);		\
+  Chi_10 = Chimu_10+timesI(Chimu_30);		\
+  Chi_11 = Chimu_11+timesI(Chimu_31);		\
+  Chi_12 = Chimu_12+timesI(Chimu_32);\
+  std::cout << std::endl << "DEBUG -- ZM_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+#define TM_PROJ \
+  Chi_00 = Chimu_00-Chimu_20;		\
+  Chi_01 = Chimu_01-Chimu_21;		\
+  Chi_02 = Chimu_02-Chimu_22;		\
+  Chi_10 = Chimu_10-Chimu_30;		\
+  Chi_11 = Chimu_11-Chimu_31;		\
+  Chi_12 = Chimu_12-Chimu_32;\
+  std::cout << std::endl << "DEBUG -- TM_PROJ" << std::endl; \
+  std::cout << "Chi_00 -- " <<  Chi_00 << std::endl; \
+  std::cout << "Chi_01 -- " <<  Chi_01 << std::endl; \
+  std::cout << "Chi_02 -- " <<  Chi_02 << std::endl; \
+  std::cout << "Chi_10 -- " <<  Chi_10 << std::endl; \
+  std::cout << "Chi_11 -- " <<  Chi_11 << std::endl; \
+  std::cout << "Chi_12 -- " <<  Chi_12 << std::endl;
+
+//      fspin(0)=hspin(0);
+//      fspin(1)=hspin(1);
+//      fspin(2)=timesMinusI(hspin(1));
+//      fspin(3)=timesMinusI(hspin(0));
+#define XP_RECON\
+  result_00 = UChi_00;\
+  result_01 = UChi_01;\
+  result_02 = UChi_02;\
+  result_10 = UChi_10;\
+  result_11 = UChi_11;\
+  result_12 = UChi_12;\
+  result_20 = timesMinusI(UChi_10);\
+  result_21 = timesMinusI(UChi_11);\
+  result_22 = timesMinusI(UChi_12);\
+  result_30 = timesMinusI(UChi_00);\
+  result_31 = timesMinusI(UChi_01);\
+  result_32 = timesMinusI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XP_RECON" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define XP_RECON_ACCUM\
+  result_00+=UChi_00;\
+  result_01+=UChi_01;\
+  result_02+=UChi_02;\
+  result_10+=UChi_10;\
+  result_11+=UChi_11;\
+  result_12+=UChi_12;\
+  result_20-=timesI(UChi_10);\
+  result_21-=timesI(UChi_11);\
+  result_22-=timesI(UChi_12);\
+  result_30-=timesI(UChi_00);\
+  result_31-=timesI(UChi_01);\
+  result_32-=timesI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define XM_RECON\
+  result_00 = UChi_00;\
+  result_01 = UChi_01;\
+  result_02 = UChi_02;\
+  result_10 = UChi_10;\
+  result_11 = UChi_11;\
+  result_12 = UChi_12;\
+  result_20 = timesI(UChi_10);\
+  result_21 = timesI(UChi_11);\
+  result_22 = timesI(UChi_12);\
+  result_30 = timesI(UChi_00);\
+  result_31 = timesI(UChi_01);\
+  result_32 = timesI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XM_RECON" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define XM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= timesI(UChi_10);\
+  result_21+= timesI(UChi_11);\
+  result_22+= timesI(UChi_12);\
+  result_30+= timesI(UChi_00);\
+  result_31+= timesI(UChi_01);\
+  result_32+= timesI(UChi_02);\
+  std::cout << std::endl << "DEBUG -- XM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define YP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= UChi_10;\
+  result_21+= UChi_11;\
+  result_22+= UChi_12;\
+  result_30-= UChi_00;\
+  result_31-= UChi_01;\
+  result_32-= UChi_02;\
+  std::cout << std::endl << "DEBUG -- YP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define YM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= UChi_10;\
+  result_21-= UChi_11;\
+  result_22-= UChi_12;\
+  result_30+= UChi_00;\
+  result_31+= UChi_01;\
+  result_32+= UChi_02;\
+  std::cout << std::endl << "DEBUG -- YM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define ZP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= timesI(UChi_00);			\
+  result_21-= timesI(UChi_01);			\
+  result_22-= timesI(UChi_02);			\
+  result_30+= timesI(UChi_10);			\
+  result_31+= timesI(UChi_11);			\
+  result_32+= timesI(UChi_12);\
+  std::cout << std::endl << "DEBUG -- ZP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define ZM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= timesI(UChi_00);			\
+  result_21+= timesI(UChi_01);			\
+  result_22+= timesI(UChi_02);			\
+  result_30-= timesI(UChi_10);			\
+  result_31-= timesI(UChi_11);			\
+  result_32-= timesI(UChi_12);\
+  std::cout << std::endl << "DEBUG -- ZM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define TP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= UChi_00;			\
+  result_21+= UChi_01;			\
+  result_22+= UChi_02;			\
+  result_30+= UChi_10;			\
+  result_31+= UChi_11;			\
+  result_32+= UChi_12;\
+  std::cout << std::endl << "DEBUG -- TP_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define TM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= UChi_00;	\
+  result_21-= UChi_01;	\
+  result_22-= UChi_02;	\
+  result_30-= UChi_10;	\
+  result_31-= UChi_11;	\
+  result_32-= UChi_12;\
+  std::cout << std::endl << "DEBUG -- TM_RECON_ACCUM" << std::endl; \
+  std::cout << "result_00 -- " <<  result_00 << std::endl; \
+  std::cout << "result_01 -- " <<  result_01 << std::endl; \
+  std::cout << "result_02 -- " <<  result_02 << std::endl; \
+  std::cout << "result_10 -- " <<  result_10 << std::endl; \
+  std::cout << "result_11 -- " <<  result_11 << std::endl; \
+  std::cout << "result_12 -- " <<  result_12 << std::endl; \
+  std::cout << "result_20 -- " <<  result_20 << std::endl; \
+  std::cout << "result_21 -- " <<  result_21 << std::endl; \
+  std::cout << "result_22 -- " <<  result_22 << std::endl; \
+  std::cout << "result_30 -- " <<  result_30 << std::endl; \
+  std::cout << "result_31 -- " <<  result_31 << std::endl; \
+  std::cout << "result_32 -- " <<  result_32 << std::endl;
+
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU;					\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else {					\
+    LOAD_CHI;					\
+  }						\
+  MULT_2SPIN(DIR);				\
+  RECON;
+
+#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON)	\
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU;					\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else if ( st.same_node[DIR] ) {		\
+    LOAD_CHI;					\
+  }						\
+  if (local || st.same_node[DIR] ) {		\
+    MULT_2SPIN(DIR);				\
+    RECON;					\
+  }
+
+#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON)	\
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {	\
+    LOAD_CHI;					\
+    MULT_2SPIN(DIR);				\
+    RECON;					\
+    nmu++;					\
+  }
+
+#define HAND_RESULT(ss)				\
+  {						\
+    SiteSpinor & ref (out[ss]);		\
+    vstream(ref()(0)(0),result_00);		\
+    vstream(ref()(0)(1),result_01);		\
+    vstream(ref()(0)(2),result_02);		\
+    vstream(ref()(1)(0),result_10);		\
+    vstream(ref()(1)(1),result_11);		\
+    vstream(ref()(1)(2),result_12);		\
+    vstream(ref()(2)(0),result_20);		\
+    vstream(ref()(2)(1),result_21);		\
+    vstream(ref()(2)(2),result_22);		\
+    vstream(ref()(3)(0),result_30);		\
+    vstream(ref()(3)(1),result_31);		\
+    vstream(ref()(3)(2),result_32);		\
+    std::cout << std::endl << "DEBUG -- RESULT" << std::endl; \
+    std::cout << "result_00 -- " <<  result_00 << std::endl; \
+    std::cout << "result_01 -- " <<  result_01 << std::endl; \
+    std::cout << "result_02 -- " <<  result_02 << std::endl; \
+    std::cout << "result_10 -- " <<  result_10 << std::endl; \
+    std::cout << "result_11 -- " <<  result_11 << std::endl; \
+    std::cout << "result_12 -- " <<  result_12 << std::endl; \
+    std::cout << "result_20 -- " <<  result_20 << std::endl; \
+    std::cout << "result_21 -- " <<  result_21 << std::endl; \
+    std::cout << "result_22 -- " <<  result_22 << std::endl; \
+    std::cout << "result_30 -- " <<  result_30 << std::endl; \
+    std::cout << "result_31 -- " <<  result_31 << std::endl; \
+    std::cout << "result_32 -- " <<  result_32 << std::endl;\
+  }
+
+#define HAND_RESULT_EXT(ss)			\
+  if (nmu){					\
+    SiteSpinor & ref (out[ss]);		\
+    ref()(0)(0)+=result_00;		\
+    ref()(0)(1)+=result_01;		\
+    ref()(0)(2)+=result_02;		\
+    ref()(1)(0)+=result_10;		\
+    ref()(1)(1)+=result_11;		\
+    ref()(1)(2)+=result_12;		\
+    ref()(2)(0)+=result_20;		\
+    ref()(2)(1)+=result_21;		\
+    ref()(2)(2)+=result_22;		\
+    ref()(3)(0)+=result_30;		\
+    ref()(3)(1)+=result_31;		\
+    ref()(3)(2)+=result_32;		\
+    std::cout << std::endl << "DEBUG -- RESULT EXT" << std::endl; \
+    std::cout << "result_00 -- " <<  result_00 << std::endl; \
+    std::cout << "result_01 -- " <<  result_01 << std::endl; \
+    std::cout << "result_02 -- " <<  result_02 << std::endl; \
+    std::cout << "result_10 -- " <<  result_10 << std::endl; \
+    std::cout << "result_11 -- " <<  result_11 << std::endl; \
+    std::cout << "result_12 -- " <<  result_12 << std::endl; \
+    std::cout << "result_20 -- " <<  result_20 << std::endl; \
+    std::cout << "result_21 -- " <<  result_21 << std::endl; \
+    std::cout << "result_22 -- " <<  result_22 << std::endl; \
+    std::cout << "result_30 -- " <<  result_30 << std::endl; \
+    std::cout << "result_31 -- " <<  result_31 << std::endl; \
+    std::cout << "result_32 -- " <<  result_32 << std::endl;\
+  }
+
+
+#define HAND_DECLARATIONS(a)			\
+  Simd result_00;				\
+  Simd result_01;				\
+  Simd result_02;				\
+  Simd result_10;				\
+  Simd result_11;				\
+  Simd result_12;				\
+  Simd result_20;				\
+  Simd result_21;				\
+  Simd result_22;				\
+  Simd result_30;				\
+  Simd result_31;				\
+  Simd result_32;				\
+  Simd Chi_00;					\
+  Simd Chi_01;					\
+  Simd Chi_02;					\
+  Simd Chi_10;					\
+  Simd Chi_11;					\
+  Simd Chi_12;					\
+  Simd UChi_00;					\
+  Simd UChi_01;					\
+  Simd UChi_02;					\
+  Simd UChi_10;					\
+  Simd UChi_11;					\
+  Simd UChi_12;					\
+  Simd U_00;					\
+  Simd U_10;					\
+  Simd U_20;					\
+  Simd U_01;					\
+  Simd U_11;					\
+  Simd U_21;\
+  Simd debugreg;\
+  svbool_t pg1;        \
+  pg1 = svptrue_b64();        \
+
+#define ZERO_RESULT				\
+  result_00=Zero();				\
+  result_01=Zero();				\
+  result_02=Zero();				\
+  result_10=Zero();				\
+  result_11=Zero();				\
+  result_12=Zero();				\
+  result_20=Zero();				\
+  result_21=Zero();				\
+  result_22=Zero();				\
+  result_30=Zero();				\
+  result_31=Zero();				\
+  result_32=Zero();
+
+#define Chimu_00 Chi_00
+#define Chimu_01 Chi_01
+#define Chimu_02 Chi_02
+#define Chimu_10 Chi_10
+#define Chimu_11 Chi_11
+#define Chimu_12 Chi_12
+#define Chimu_20 UChi_00
+#define Chimu_21 UChi_01
+#define Chimu_22 UChi_02
+#define Chimu_30 UChi_10
+#define Chimu_31 UChi_11
+#define Chimu_32 UChi_12
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Impl> void
+WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+				  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+
+  HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
+  HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset,local,perm, ptype;
+
+  HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
+  HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl> void
+WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset,local,perm, ptype;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT(ss);
+}
+
+template<class Impl> void
+WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset, ptype;
+  StencilEntry *SE;
+  int nmu=0;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT_EXT(ss);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset, ptype;
+  int nmu=0;
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT_EXT(ss);
+}
+
+////////////// Wilson ; uses this implementation /////////////////////
+
+NAMESPACE_END(Grid);
+#undef LOAD_CHIMU
+#undef LOAD_CHI
+#undef MULT_2SPIN
+#undef PERMUTE_DIR
+#undef XP_PROJ
+#undef YP_PROJ
+#undef ZP_PROJ
+#undef TP_PROJ
+#undef XM_PROJ
+#undef YM_PROJ
+#undef ZM_PROJ
+#undef TM_PROJ
+#undef XP_RECON
+#undef XP_RECON_ACCUM
+#undef XM_RECON
+#undef XM_RECON_ACCUM
+#undef YP_RECON_ACCUM
+#undef YM_RECON_ACCUM
+#undef ZP_RECON_ACCUM
+#undef ZM_RECON_ACCUM
+#undef TP_RECON_ACCUM
+#undef TM_RECON_ACCUM
+#undef ZERO_RESULT
+#undef Chimu_00
+#undef Chimu_01
+#undef Chimu_02
+#undef Chimu_10
+#undef Chimu_11
+#undef Chimu_12
+#undef Chimu_20
+#undef Chimu_21
+#undef Chimu_22
+#undef Chimu_30
+#undef Chimu_31
+#undef Chimu_32
+#undef HAND_STENCIL_LEG
+#undef HAND_STENCIL_LEG_INT
+#undef HAND_STENCIL_LEG_EXT
+#undef HAND_RESULT
+#undef HAND_RESULT_INT
+#undef HAND_RESULT_EXT
diff --git a/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h b/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
index 0ff72789..c2b62416 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
@@ -39,19 +39,21 @@ NAMESPACE_BEGIN(Grid);
 // Generic implementation; move to different file?
 ////////////////////////////////////////////
 
+/*
 accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
 {
-#ifdef __CUDA_ARCH__
-  static_assert(sizeof(StencilEntry)==sizeof(uint4),"Unexpected Stencil Entry Size"); 
+#ifdef GRID_SIMT
+  static_assert(sizeof(StencilEntry)==sizeof(uint4),"Unexpected Stencil Entry Size");
   uint4 * mem_pun  = (uint4 *)mem; // force 128 bit loads
   uint4 * chip_pun = (uint4 *)&chip;
   * chip_pun = * mem_pun;
-#else 
+#else
   chip = *mem;
 #endif
   return;
 }
-  
+*/
+
 #define GENERIC_STENCIL_LEG(Dir,spProj,Recon)			\
   SE = st.GetEntry(ptype, Dir, sF);				\
   if (SE->_is_local) {						\
@@ -61,10 +63,10 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   } else {							\
     chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  synchronise();						\
+  acceleratorSynchronise();						\
   Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
   Recon(result, Uchi);
-  
+
 #define GENERIC_STENCIL_LEG_INT(Dir,spProj,Recon)		\
   SE = st.GetEntry(ptype, Dir, sF);				\
   if (SE->_is_local) {						\
@@ -74,12 +76,12 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   } else if ( st.same_node[Dir] ) {				\
     chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  synchronise();						\
+  acceleratorSynchronise();						\
   if (SE->_is_local || st.same_node[Dir] ) {			\
     Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
     Recon(result, Uchi);					\
   }								\
-  synchronise();						
+  acceleratorSynchronise();
 
 #define GENERIC_STENCIL_LEG_EXT(Dir,spProj,Recon)		\
   SE = st.GetEntry(ptype, Dir, sF);				\
@@ -89,7 +91,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
     Recon(result, Uchi);					\
     nmu++;							\
   }								\
-  synchronise();						
+  acceleratorSynchronise();
 
 #define GENERIC_DHOPDIR_LEG_BODY(Dir,spProj,Recon)		\
     if (SE->_is_local ) {					\
@@ -99,9 +101,9 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
     } else {							\
       chi = coalescedRead(buf[SE->_offset],lane);		\
     }								\
-    synchronise();						\
+    acceleratorSynchronise();					\
     Impl::multLink(Uchi, U[sU], chi, dir, SE, st);		\
-    Recon(result, Uchi);					
+    Recon(result, Uchi);
 
 #define GENERIC_DHOPDIR_LEG(Dir,spProj,Recon)			\
   if (gamma == Dir) {						\
@@ -126,7 +128,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldV
   StencilEntry *SE;
   int ptype;
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
   GENERIC_STENCIL_LEG(Xp,spProjXp,spReconXp);
   GENERIC_STENCIL_LEG(Yp,spProjYp,accumReconYp);
   GENERIC_STENCIL_LEG(Zp,spProjZp,accumReconZp);
@@ -141,7 +143,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldV
 template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U,
 					  SiteHalfSpinor *buf, int sF,
-					  int sU, const FermionFieldView &in, FermionFieldView &out) 
+					  int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -153,7 +155,7 @@ void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView
   int ptype;
 
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
   GENERIC_STENCIL_LEG(Xm,spProjXp,spReconXp);
   GENERIC_STENCIL_LEG(Ym,spProjYp,accumReconYp);
   GENERIC_STENCIL_LEG(Zm,spProjZp,accumReconZp);
@@ -181,7 +183,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFi
   StencilEntry *SE;
   int ptype;
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   result=Zero();
   GENERIC_STENCIL_LEG_INT(Xp,spProjXp,accumReconXp);
@@ -198,12 +200,12 @@ void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st,  DoubledGaugeFi
 template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeFieldView &U,
 							 SiteHalfSpinor *buf, int sF,
-							 int sU, const FermionFieldView &in, FermionFieldView &out) 
+							 int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   calcHalfSpinor chi;
   //  calcHalfSpinor *chi_p;
@@ -239,7 +241,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFi
   int ptype;
   int nmu=0;
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
   result=Zero();
   GENERIC_STENCIL_LEG_EXT(Xp,spProjXp,accumReconXp);
   GENERIC_STENCIL_LEG_EXT(Yp,spProjYp,accumReconYp);
@@ -249,7 +251,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFi
   GENERIC_STENCIL_LEG_EXT(Ym,spProjYm,accumReconYm);
   GENERIC_STENCIL_LEG_EXT(Zm,spProjZm,accumReconZm);
   GENERIC_STENCIL_LEG_EXT(Tm,spProjTm,accumReconTm);
-  if ( nmu ) { 
+  if ( nmu ) {
     auto out_t = coalescedRead(out[sF],lane);
     out_t = out_t + result;
     coalescedWrite(out[sF],out_t,lane);
@@ -259,7 +261,7 @@ void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st,  DoubledGaugeFi
 template <class Impl>
 void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeFieldView &U,
 					     SiteHalfSpinor *buf, int sF,
-					     int sU, const FermionFieldView &in, FermionFieldView &out) 
+					     int sU, const FermionFieldView &in, FermionFieldView &out)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -270,7 +272,7 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeField
   int ptype;
   int nmu=0;
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
   result=Zero();
   GENERIC_STENCIL_LEG_EXT(Xm,spProjXp,accumReconXp);
   GENERIC_STENCIL_LEG_EXT(Ym,spProjYp,accumReconYp);
@@ -280,7 +282,7 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeField
   GENERIC_STENCIL_LEG_EXT(Yp,spProjYm,accumReconYm);
   GENERIC_STENCIL_LEG_EXT(Zp,spProjZm,accumReconZm);
   GENERIC_STENCIL_LEG_EXT(Tp,spProjTm,accumReconTm);
-  if ( nmu ) { 
+  if ( nmu ) {
     auto out_t = coalescedRead(out[sF],lane);
     out_t = out_t + result;
     coalescedWrite(out[sF],out_t,lane);
@@ -300,12 +302,12 @@ void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st,  DoubledGaugeField
   StencilEntry *SE;							\
   int ptype;								\
   const int Nsimd = SiteHalfSpinor::Nsimd();				\
-  const int lane=SIMTlane(Nsimd);					\
+  const int lane=acceleratorSIMTlane(Nsimd);					\
 									\
   SE = st.GetEntry(ptype, dir, sF);					\
   GENERIC_DHOPDIR_LEG_BODY(Dir,spProj,spRecon);				\
   coalescedWrite(out[sF], result,lane);					\
-  }									
+  }
 
 DhopDirMacro(Xp,spProjXp,spReconXp);
 DhopDirMacro(Yp,spProjYp,spReconYp);
@@ -316,9 +318,9 @@ DhopDirMacro(Ym,spProjYm,spReconYm);
 DhopDirMacro(Zm,spProjZm,spReconZm);
 DhopDirMacro(Tm,spProjTm,spReconTm);
 
-template <class Impl> 
+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) 
+				    int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int gamma)
 {
   typedef decltype(coalescedRead(buf[0])) calcHalfSpinor;
   typedef decltype(coalescedRead(in[0]))  calcSpinor;
@@ -328,7 +330,7 @@ void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,Si
   StencilEntry *SE;
   int ptype;
   const int Nsimd = SiteHalfSpinor::Nsimd();
-  const int lane=SIMTlane(Nsimd);
+  const int lane=acceleratorSIMTlane(Nsimd);
 
   SE = st.GetEntry(ptype, dir, sF);
   GENERIC_DHOPDIR_LEG(Xp,spProjXp,spReconXp);
@@ -344,54 +346,55 @@ void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,Si
 
 template <class Impl>
 void WilsonKernels<Impl>::DhopDirAll( StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor *buf, int Ls,
-				      int Nsite, const FermionField &in, std::vector<FermionField> &out) 
+				      int Nsite, const FermionField &in, std::vector<FermionField> &out)
 {
-   auto U_v   = U.View();
-   auto in_v  = in.View();
-   auto st_v  = st.View();
+   autoView(U_v  ,U,AcceleratorRead);
+   autoView(in_v ,in,AcceleratorRead);
+   autoView(st_v ,st,AcceleratorRead);
 
-   auto out_Xm = out[0].View();
-   auto out_Ym = out[1].View();
-   auto out_Zm = out[2].View();
-   auto out_Tm = out[3].View();
-   auto out_Xp = out[4].View();
-   auto out_Yp = out[5].View();
-   auto out_Zp = out[6].View();
-   auto out_Tp = out[7].View();
-
-   accelerator_forNB(sss,Nsite*Ls,Simd::Nsimd(),{
-      int sU=sss/Ls;				
-      int sF =sss;				
-      DhopDirXm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Xm,0);
-      DhopDirYm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Ym,1);
-      DhopDirZm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Zm,2);
-      DhopDirTm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Tm,3);
-      DhopDirXp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Xp,4);
-      DhopDirYp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Yp,5);
-      DhopDirZp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Zp,6);
-      DhopDirTp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Tp,7);
+   autoView(out_Xm,out[0],AcceleratorWrite);
+   autoView(out_Ym,out[1],AcceleratorWrite);
+   autoView(out_Zm,out[2],AcceleratorWrite);
+   autoView(out_Tm,out[3],AcceleratorWrite);
+   autoView(out_Xp,out[4],AcceleratorWrite);
+   autoView(out_Yp,out[5],AcceleratorWrite);
+   autoView(out_Zp,out[6],AcceleratorWrite);
+   autoView(out_Tp,out[7],AcceleratorWrite);
+   auto CBp=st.CommBuf();
+   accelerator_for(sss,Nsite*Ls,Simd::Nsimd(),{
+      int sU=sss/Ls;
+      int sF =sss;
+      DhopDirXm(st_v,U_v,CBp,sF,sU,in_v,out_Xm,0);
+      DhopDirYm(st_v,U_v,CBp,sF,sU,in_v,out_Ym,1);
+      DhopDirZm(st_v,U_v,CBp,sF,sU,in_v,out_Zm,2);
+      DhopDirTm(st_v,U_v,CBp,sF,sU,in_v,out_Tm,3);
+      DhopDirXp(st_v,U_v,CBp,sF,sU,in_v,out_Xp,4);
+      DhopDirYp(st_v,U_v,CBp,sF,sU,in_v,out_Yp,5);
+      DhopDirZp(st_v,U_v,CBp,sF,sU,in_v,out_Zp,6);
+      DhopDirTp(st_v,U_v,CBp,sF,sU,in_v,out_Tp,7);
    });
 }
 
 
 template <class Impl>
 void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor *buf, int Ls,
-					 int Nsite, const FermionField &in, FermionField &out, int dirdisp, int gamma) 
+					 int Nsite, const FermionField &in, FermionField &out, int dirdisp, int gamma)
 {
   assert(dirdisp<=7);
   assert(dirdisp>=0);
 
-   auto U_v   = U.View();
-   auto in_v  = in.View();
-   auto out_v = out.View();
-   auto st_v  = st.View();
+   autoView(U_v  ,U  ,AcceleratorRead);
+   autoView(in_v ,in ,AcceleratorRead);
+   autoView(out_v,out,AcceleratorWrite);
+   autoView(st_v ,st ,AcceleratorRead);
+   auto CBp=st.CommBuf();
 #define LoopBody(Dir)				\
-   case Dir :			\
-     accelerator_forNB(ss,Nsite,Simd::Nsimd(),{	\
+   case Dir :					\
+     accelerator_for(ss,Nsite,Simd::Nsimd(),{	\
        for(int s=0;s<Ls;s++){			\
 	 int sU=ss;				\
 	 int sF = s+Ls*sU;						\
-	 DhopDir##Dir(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_v,dirdisp);\
+	 DhopDir##Dir(st_v,U_v,CBp,sF,sU,in_v,out_v,dirdisp);\
        }							       \
        });							       \
      break;
@@ -411,7 +414,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
      break;
    }
 #undef LoopBody
-} 
+}
 
 #define KERNEL_CALLNB(A) \
   const uint64_t    NN = Nsite*Ls;					\
@@ -421,7 +424,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
   });
 
-#define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier(); 
+#define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
@@ -433,28 +436,28 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 template <class Impl>
 void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 				     int Ls, int Nsite, const FermionField &in, FermionField &out,
-				     int interior,int exterior) 
+				     int interior,int exterior)
 {
-    auto U_v   =   U.View();
-    auto in_v  =  in.View();
-    auto out_v = out.View();
-    auto st_v  =  st.View();
+    autoView(U_v  ,  U,AcceleratorRead);
+    autoView(in_v , in,AcceleratorRead);
+    autoView(out_v,out,AcceleratorWrite);
+    autoView(st_v , st,AcceleratorRead);
 
-   if( interior && exterior ) { 
+   if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteInt); return;}
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteInt);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
-   } else if( exterior ) { 
+   } else if( exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
@@ -464,28 +467,28 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
   template <class Impl>
   void WilsonKernels<Impl>::DhopDagKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 					  int Ls, int Nsite, const FermionField &in, FermionField &out,
-					  int interior,int exterior) 
+					  int interior,int exterior)
   {
-    auto U_v   = U.View();
-    auto in_v  = in.View();
-    auto out_v = out.View();
-    auto st_v  = st.View();
+    autoView(U_v  ,U,AcceleratorRead);
+    autoView(in_v ,in,AcceleratorRead);
+    autoView(out_v,out,AcceleratorWrite);
+    autoView(st_v ,st,AcceleratorRead);
 
-   if( interior && exterior ) { 
+   if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDag); return;}
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDag);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
-   } else if( exterior ) { 
+   } else if( exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
@@ -493,5 +496,8 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    assert(0 && " Kernel optimisation case not covered ");
   }
 
-NAMESPACE_END(Grid);
+#undef KERNEL_CALLNB
+#undef KERNEL_CALL
+#undef ASM_CALL
 
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/NaiveStaggeredFermionInstantiation.cc b/Grid/qcd/action/fermion/instantiation/NaiveStaggeredFermionInstantiation.cc
new file mode 100644
index 00000000..c424cb2d
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/NaiveStaggeredFermionInstantiation.cc
@@ -0,0 +1,36 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+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>
+
+NAMESPACE_BEGIN(Grid);
+
+const std::vector<int> NaiveStaggeredFermionStatic::directions({0, 1, 2, 3, 0, 1, 2, 3});
+const std::vector<int> NaiveStaggeredFermionStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1});
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/NaiveStaggeredFermionInstantiation.cc.master b/Grid/qcd/action/fermion/instantiation/NaiveStaggeredFermionInstantiation.cc.master
new file mode 100644
index 00000000..75b75678
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/NaiveStaggeredFermionInstantiation.cc.master
@@ -0,0 +1,37 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/NaiveStaggeredFermion.cc
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h>
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class NaiveStaggeredFermion<IMPLEMENTATION>; 
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/StaggeredImplD/NaiveStaggeredFermionInstantiationStaggeredImplD.cc b/Grid/qcd/action/fermion/instantiation/StaggeredImplD/NaiveStaggeredFermionInstantiationStaggeredImplD.cc
new file mode 120000
index 00000000..42057f56
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/StaggeredImplD/NaiveStaggeredFermionInstantiationStaggeredImplD.cc
@@ -0,0 +1 @@
+../NaiveStaggeredFermionInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/StaggeredImplF/NaiveStaggeredFermionInstantiationStaggeredImplF.cc b/Grid/qcd/action/fermion/instantiation/StaggeredImplF/NaiveStaggeredFermionInstantiationStaggeredImplF.cc
new file mode 120000
index 00000000..42057f56
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/StaggeredImplF/NaiveStaggeredFermionInstantiationStaggeredImplF.cc
@@ -0,0 +1 @@
+../NaiveStaggeredFermionInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc b/Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiation.cc.master b/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiation.cc.master
index 9af5ed85..f0b15e3b 100644
--- a/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiation.cc.master
+++ b/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiation.cc.master
@@ -4,11 +4,12 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
 
-Copyright (C) 2015
+Copyright (C) 2015, 2020
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -34,14 +35,17 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>; 
+template class WilsonKernels<IMPLEMENTATION>;
 
 NAMESPACE_END(Grid);
-
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiationAsm.cc b/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiationAsm.cc
index f6f235c8..a8e9e6d9 100644
--- a/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiationAsm.cc
+++ b/Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiationAsm.cc
@@ -37,6 +37,7 @@ directory
 ////////////////////////////////////////////////////////////////////////
 NAMESPACE_BEGIN(Grid);
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmAvx512.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmQPX.h>
 NAMESPACE_END(Grid);
 
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc
deleted file mode 120000
index 01c35e7b..00000000
--- a/Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc
+++ /dev/null
@@ -1 +0,0 @@
-../WilsonKernelsInstantiation.cc.master
\ No newline at end of file
diff --git a/Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc b/Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc
new file mode 100644
index 00000000..f0b15e3b
--- /dev/null
+++ b/Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc
@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/qcd/action/fermion/instantiation/generate_instantiations.sh b/Grid/qcd/action/fermion/instantiation/generate_instantiations.sh
index 330dcfa8..72a9eaf9 100755
--- a/Grid/qcd/action/fermion/instantiation/generate_instantiations.sh
+++ b/Grid/qcd/action/fermion/instantiation/generate_instantiations.sh
@@ -88,6 +88,7 @@ done
 CC_LIST=" \
   ImprovedStaggeredFermion5DInstantiation \
   ImprovedStaggeredFermionInstantiation \
+  NaiveStaggeredFermionInstantiation \
   StaggeredKernelsInstantiation "
 
 for impl in $STAG_IMPL_LIST
diff --git a/Grid/qcd/action/gauge/GaugeImplTypes.h b/Grid/qcd/action/gauge/GaugeImplTypes.h
index b9a5296d..9b7d5a60 100644
--- a/Grid/qcd/action/gauge/GaugeImplTypes.h
+++ b/Grid/qcd/action/gauge/GaugeImplTypes.h
@@ -86,9 +86,9 @@ public:
 
   // Move this elsewhere? FIXME
   static inline void AddLink(Field &U, LinkField &W, int mu) { // U[mu] += W
-    auto U_v = U.View();
-    auto W_v = W.View();
-    thread_for( ss, U.Grid()->oSites(), {
+    autoView(U_v,U,AcceleratorWrite);
+    autoView(W_v,W,AcceleratorRead);
+    accelerator_for( ss, U.Grid()->oSites(), 1, {
       U_v[ss](mu) = U_v[ss](mu) + W_v[ss]();
     });
   }
@@ -131,15 +131,14 @@ public:
     //static std::chrono::duration<double> diff;
 
     //auto start = std::chrono::high_resolution_clock::now();
-    auto U_v = U.View();
-    auto P_v = P.View();
-    thread_for(ss, P.Grid()->oSites(),{
+    autoView(U_v,U,AcceleratorWrite);
+    autoView(P_v,P,AcceleratorRead);
+    accelerator_for(ss, P.Grid()->oSites(),1,{
       for (int mu = 0; mu < Nd; mu++) {
         U_v[ss](mu) = ProjectOnGroup(Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu));
       }
     });
-    
-    //auto end = std::chrono::high_resolution_clock::now();
+   //auto end = std::chrono::high_resolution_clock::now();
    // diff += end - start;
    // std::cout << "Time to exponentiate matrix " << diff.count() << " s\n";
   }
diff --git a/Grid/qcd/action/scalar/ScalarImpl.h b/Grid/qcd/action/scalar/ScalarImpl.h
index febb315e..14675b11 100644
--- a/Grid/qcd/action/scalar/ScalarImpl.h
+++ b/Grid/qcd/action/scalar/ScalarImpl.h
@@ -1,5 +1,13 @@
 #pragma once
 
+#define CPS_MD_TIME 
+
+#ifdef CPS_MD_TIME
+#define HMC_MOMENTUM_DENOMINATOR (2.0)
+#else
+#define HMC_MOMENTUM_DENOMINATOR (1.0)
+#endif
+
 NAMESPACE_BEGIN(Grid);
 
 template <class S>
@@ -20,7 +28,9 @@ public:
   typedef Field              PropagatorField;
     
   static inline void generate_momenta(Field& P, GridParallelRNG& pRNG){
+    RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR); // CPS/UKQCD momentum rescaling
     gaussian(pRNG, P);
+    P *= scale; 
   }
 
   static inline Field projectForce(Field& P){return P;}
@@ -66,7 +76,7 @@ public:
   }
     
   static void FreePropagator(const Field &in, Field &out,
-			     const Field &momKernel)
+           const Field &momKernel)
   {
     FFT   fft((GridCartesian *)in.Grid());
     Field inFT(in.Grid());
@@ -139,14 +149,17 @@ public:
 
     static inline void generate_momenta(Field &P, GridParallelRNG &pRNG)
     {
+      RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR); // CPS/UKQCD momentum rescaling
 #ifndef USE_FFT_ACCELERATION
     Group::GaussianFundamentalLieAlgebraMatrix(pRNG, P);
+    
 #else
 
       Field Pgaussian(P.Grid()), Pp(P.Grid());
       ComplexField p2(P.Grid()); p2 = zero;
       RealD M = FFT_MASS;
-      
+
+
       Group::GaussianFundamentalLieAlgebraMatrix(pRNG, Pgaussian);
 
       FFT theFFT((GridCartesian*)P.Grid());
@@ -156,17 +169,17 @@ public:
       p2 = sqrt(p2);
       Pp *= p2;
       theFFT.FFT_all_dim(P, Pp, FFT::backward);
-
 #endif //USE_FFT_ACCELERATION
+      P *= scale; 
   }
 
-  static inline Field projectForce(Field& P) {return P;}
+  static inline Field projectForce(Field& P) {return Ta(P);}
 
     static inline void update_field(Field &P, Field &U, double ep)
     {
 #ifndef USE_FFT_ACCELERATION
       double t0=usecond(); 
-    U += P*ep;
+      U += P*ep;
       double t1=usecond();
       double total_time = (t1-t0)/1e6;
       std::cout << GridLogIntegrator << "Total time for updating field (s)       : " << total_time << std::endl; 
diff --git a/Grid/qcd/action/scalar/ScalarInteractionAction.h b/Grid/qcd/action/scalar/ScalarInteractionAction.h
index 3be84480..5a5f9251 100644
--- a/Grid/qcd/action/scalar/ScalarInteractionAction.h
+++ b/Grid/qcd/action/scalar/ScalarInteractionAction.h
@@ -89,8 +89,8 @@ public:
     action = (2.0 * Ndim + mass_square) * phisquared - lambda * phisquared * phisquared;
     
     
-    auto p_v = p.View();
-    auto action_v = action.View();
+    autoView( p_v      , p, CpuRead);
+    autoView( action_v , action, CpuWrite);
     for (int mu = 0; mu < Ndim; mu++)
     {
       //  pshift = Cshift(p, mu, +1);  // not efficient, implement with stencils
@@ -146,8 +146,8 @@ public:
     for (int point = 0; point < npoint; point++)
     {
 
-      auto p_v = p.View();
-      auto force_v = force.View();
+      autoView( p_v , p, CpuRead);
+      autoView( force_v , force, CpuWrite);
             
       int permute_type;
       StencilEntry *SE;
diff --git a/Grid/qcd/modules/Registration.h b/Grid/qcd/modules/Registration.h
index 459e1d0b..28a9fdae 100644
--- a/Grid/qcd/modules/Registration.h
+++ b/Grid/qcd/modules/Registration.h
@@ -80,10 +80,11 @@ static Registrar<OneFlavourRatioEOFModule<FermionImplementationPolicy>,
 
 static Registrar< ConjugateGradientModule<WilsonFermionR::FermionField>,   
                   HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CGWFmodXMLInit("ConjugateGradient"); 
-//static Registrar< BiCGSTABModule<WilsonFermionR::FermionField>,   
-//                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CGWFmodXMLInit("BiCGSTAB"); 
-//static Registrar< ConjugateResidualModule<WilsonFermionR::FermionField>,   
-//                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CRWFmodXMLInit("ConjugateResidual"); 
+
+static Registrar< BiCGSTABModule<WilsonFermionR::FermionField>,   
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __BiCGWFmodXMLInit("BiCGSTAB"); 
+static Registrar< ConjugateResidualModule<WilsonFermionR::FermionField>,   
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CRWFmodXMLInit("ConjugateResidual"); 
 
 // add the staggered, scalar versions here
 
diff --git a/Grid/qcd/smearing/GaugeConfiguration.h b/Grid/qcd/smearing/GaugeConfiguration.h
index f4d00c72..0ff7fc25 100644
--- a/Grid/qcd/smearing/GaugeConfiguration.h
+++ b/Grid/qcd/smearing/GaugeConfiguration.h
@@ -49,7 +49,7 @@ public:
 
 private:
   const unsigned int smearingLevels;
-  Smear_Stout<Gimpl> StoutSmearing;
+  Smear_Stout<Gimpl> *StoutSmearing;
   std::vector<GaugeField> SmearedSet;
 
   // Member functions
@@ -72,7 +72,7 @@ private:
       previous_u = *ThinLinks;
       for (int smearLvl = 0; smearLvl < smearingLevels; ++smearLvl)
       {
-        StoutSmearing.smear(SmearedSet[smearLvl], previous_u);
+        StoutSmearing->smear(SmearedSet[smearLvl], previous_u);
         previous_u = SmearedSet[smearLvl];
 
         // For debug purposes
@@ -93,7 +93,7 @@ private:
     GaugeLinkField SigmaKPrime_mu(grid);
     GaugeLinkField GaugeKmu(grid), Cmu(grid);
 
-    StoutSmearing.BaseSmear(C, GaugeK);
+    StoutSmearing->BaseSmear(C, GaugeK);
     SigmaK = Zero();
     iLambda = Zero();
 
@@ -107,7 +107,7 @@ private:
       pokeLorentz(SigmaK, SigmaKPrime_mu * e_iQ + adj(Cmu) * iLambda_mu, mu);
       pokeLorentz(iLambda, iLambda_mu, mu);
     }
-    StoutSmearing.derivative(SigmaK, iLambda,
+    StoutSmearing->derivative(SigmaK, iLambda,
                              GaugeK);  // derivative of SmearBase
     return SigmaK;
   }
@@ -144,14 +144,14 @@ private:
     // Exponential
     iQ2 = iQ * iQ;
     iQ3 = iQ * iQ2;
-    StoutSmearing.set_uw(u, w, iQ2, iQ3);
-    StoutSmearing.set_fj(f0, f1, f2, u, w);
+    StoutSmearing->set_uw(u, w, iQ2, iQ3);
+    StoutSmearing->set_fj(f0, f1, f2, u, w);
     e_iQ = f0 * unity + timesMinusI(f1) * iQ - f2 * iQ2;
 
     // Getting B1, B2, Gamma and Lambda
     // simplify this part, reduntant calculations in set_fj
-    xi0 = StoutSmearing.func_xi0(w);
-    xi1 = StoutSmearing.func_xi1(w);
+    xi0 = StoutSmearing->func_xi0(w);
+    xi1 = StoutSmearing->func_xi1(w);
     u2 = u * u;
     w2 = w * w;
     cosw = cos(w);
@@ -219,7 +219,7 @@ public:
   /* Standard constructor */
   SmearedConfiguration(GridCartesian* UGrid, unsigned int Nsmear,
                        Smear_Stout<Gimpl>& Stout)
-      : smearingLevels(Nsmear), StoutSmearing(Stout), ThinLinks(NULL)
+      : smearingLevels(Nsmear), StoutSmearing(&Stout), ThinLinks(NULL)
   {
     for (unsigned int i = 0; i < smearingLevels; ++i)
       SmearedSet.push_back(*(new GaugeField(UGrid)));
@@ -227,7 +227,7 @@ public:
 
   /*! For just thin links */
   SmearedConfiguration()
-    : smearingLevels(0), StoutSmearing(), SmearedSet(), ThinLinks(NULL) {}
+    : smearingLevels(0), StoutSmearing(nullptr), SmearedSet(), ThinLinks(NULL) {}
 
   // attach the smeared routines to the thin links U and fill the smeared set
   void set_Field(GaugeField &U)
diff --git a/Grid/qcd/utils/A2Autils.h b/Grid/qcd/utils/A2Autils.h
index c7c7d329..b63d8571 100644
--- a/Grid/qcd/utils/A2Autils.h
+++ b/Grid/qcd/utils/A2Autils.h
@@ -185,13 +185,14 @@ void A2Autils<FImpl>::MesonField(TensorType &mat,
 
 	for(int i=0;i<Lblock;i++){
 
-	  auto lhs_v = lhs_wi[i].View();
+	  // Recreate view potentially expensive outside fo UVM mode
+	  autoView(lhs_v,lhs_wi[i],CpuRead);
 	  auto left = conjugate(lhs_v[ss]);
-
 	  for(int j=0;j<Rblock;j++){
 
 	    SpinMatrix_v vv;
-	    auto rhs_v = rhs_vj[j].View();
+	    // Recreate view potentially expensive outside fo UVM mode
+	    autoView(rhs_v,rhs_vj[j],CpuRead);
 	    auto right = rhs_v[ss];
 	    for(int s1=0;s1<Ns;s1++){
 	    for(int s2=0;s2<Ns;s2++){
@@ -204,11 +205,10 @@ void A2Autils<FImpl>::MesonField(TensorType &mat,
 	    int base = Nmom*i+Nmom*Lblock*j+Nmom*Lblock*Rblock*r;
 	    for ( int m=0;m<Nmom;m++){
 	      int idx = m+base;
-	      auto mom_v = mom[m].View();
+	      autoView(mom_v,mom[m],CpuRead);
 	      auto phase = mom_v[ss];
 	      mac(&lvSum[idx],&vv,&phase);
 	    }
-	  
 	  }
 	}
       }
@@ -371,7 +371,7 @@ void A2Autils<FImpl>::PionFieldXX(Eigen::Tensor<ComplexD,3> &mat,
 
 	for(int i=0;i<Lblock;i++){
 
-	  auto wi_v = wi[i].View();
+	  autoView(wi_v,wi[i],CpuRead);
 	  auto w = conjugate(wi_v[ss]);
 	  if (g5) {
 	    w()(2)(0) = - w()(2)(0);
@@ -383,7 +383,7 @@ void A2Autils<FImpl>::PionFieldXX(Eigen::Tensor<ComplexD,3> &mat,
 	  }
 	  for(int j=0;j<Rblock;j++){
 	    
-	    auto vj_v=vj[j].View();
+	    autoView(vj_v,vj[j],CpuRead);
 	    auto v  = vj_v[ss];
 	    auto vv = v()(0)(0);
 
@@ -518,12 +518,12 @@ void A2Autils<FImpl>::PionFieldWVmom(Eigen::Tensor<ComplexD,4> &mat,
 
 	for(int i=0;i<Lblock;i++){
 
-	  auto wi_v = wi[i].View();
+	  autoView(wi_v,wi[i],CpuRead);
 	  auto w = conjugate(wi_v[ss]);
 
 	  for(int j=0;j<Rblock;j++){
-	    
-	    auto vj_v = vj[j].View();
+
+	    autoView(vj_v,vj[j],CpuRead);
 	    auto v = vj_v[ss];
 
 	    auto vv = w()(0)(0) * v()(0)(0)// Gamma5 Dirac basis explicitly written out
@@ -544,7 +544,7 @@ void A2Autils<FImpl>::PionFieldWVmom(Eigen::Tensor<ComplexD,4> &mat,
 	    int base = Nmom*i+Nmom*Lblock*j+Nmom*Lblock*Rblock*r;
 	    for ( int m=0;m<Nmom;m++){
 	      int idx = m+base;
-	      auto mom_v = mom[m].View();
+	      autoView(mom_v,mom[m],CpuRead);
 	      auto phase = mom_v[ss];
 	      mac(&lvSum[idx],&vv,&phase()()());
 	    }
@@ -730,13 +730,13 @@ void A2Autils<FImpl>::AslashField(TensorType &mat,
 
             for(int i=0;i<Lblock;i++)
             {
-  	        auto wi_v = lhs_wi[i].View();
+  	        autoView(wi_v,lhs_wi[i],CpuRead);
                 auto left = conjugate(wi_v[ss]);
 
                 for(int j=0;j<Rblock;j++)
                 {
                     SpinMatrix_v vv;
-		    auto vj_v  = rhs_vj[j].View();
+		    autoView(vj_v,rhs_vj[j],CpuRead);
                     auto right = vj_v[ss];
 
                     for(int s1=0;s1<Ns;s1++)
@@ -752,8 +752,8 @@ void A2Autils<FImpl>::AslashField(TensorType &mat,
 
                     for ( int m=0;m<Nem;m++)
                     {
-  		        auto emB0_v = emB0[m].View();
-  		        auto emB1_v = emB1[m].View();
+  		        autoView(emB0_v,emB0[m],CpuRead);
+		        autoView(emB1_v,emB1[m],CpuRead);
                         int idx  = m+base;
                         auto b0  = emB0_v[ss];
                         auto b1  = emB1_v[ss];
@@ -1014,21 +1014,21 @@ A2Autils<FImpl>::ContractWWVV(std::vector<PropagatorField> &WWVV,
     for(int d_o=0;d_o<N_d;d_o+=d_unroll){
       for(int t=0;t<N_t;t++){
       for(int s=0;s<N_s;s++){
-  auto vs_v = vs[s].View();
-  auto tmp1 = vs_v[ss];
-  vobj tmp2 = Zero();
-  vobj tmp3 = Zero();
-  for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
-    auto vd_v = vd[d].View();
-    Scalar_v coeff = WW_sd(t,s,d);
-    tmp3 = conjugate(vd_v[ss]);
-    mac(&tmp2, &coeff, &tmp3);
-  }
+	autoView(vs_v,vs[s],CpuRead);
+	auto tmp1 = vs_v[ss];
+	vobj tmp2 = Zero();
+	vobj tmp3 = Zero();
+	for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
+	  autoView(vd_v,vd[d],CpuRead);
+	  Scalar_v coeff = WW_sd(t,s,d);
+	  tmp3 = conjugate(vd_v[ss]);
+	  mac(&tmp2, &coeff, &tmp3);
+	}
 
-  //////////////////////////
-  // Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
-  //////////////////////////
-  OuterProductWWVV(WWVV[t], tmp1, tmp2, Ns, ss);
+	//////////////////////////
+	// Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
+	//////////////////////////
+	OuterProductWWVV(WWVV[t], tmp1, tmp2, Ns, ss);
 
       }}
     }
@@ -1067,21 +1067,20 @@ A2Autils<FImpl>::ContractWWVV(std::vector<PropagatorField> &WWVV,
     thread_for(ss,grid->oSites(),{
       for(int d_o=0;d_o<N_d;d_o+=d_unroll){
         for(int s=0;s<N_s;s++){
-    auto vs_v = vs[s].View();
-    auto tmp1 = vs_v[ss];
-    vobj tmp2 = Zero();
-    vobj tmp3 = Zero();
-    for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
-      auto vd_v = vd[d].View();
-      Scalar_v coeff = buf(s,d);
-      tmp3 = conjugate(vd_v[ss]);
-      mac(&tmp2, &coeff, &tmp3);
-    }
-
-    //////////////////////////
-    // Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
-    //////////////////////////
-    OuterProductWWVV(WWVV[t], tmp1, tmp2, Ns, ss);
+	  autoView(vs_v,vs[s],CpuRead);
+	  auto tmp1 = vs_v[ss];
+	  vobj tmp2 = Zero();
+	  vobj tmp3 = Zero();
+	  for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
+	    autoView(vd_v,vd[d],CpuRead);
+	    Scalar_v coeff = buf(s,d);
+	    tmp3 = conjugate(vd_v[ss]);
+	    mac(&tmp2, &coeff, &tmp3);
+	  }
+	  //////////////////////////
+	  // Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
+	  //////////////////////////
+	  OuterProductWWVV(WWVV[t], tmp1, tmp2, Ns, ss);
       }}
     });
   }
@@ -1093,7 +1092,7 @@ inline void A2Autils<FImpl>::OuterProductWWVV(PropagatorField &WWVV,
                                              const vobj &rhs,
                                              const int Ns, const int ss)
 {
-  auto WWVV_v = WWVV.View();
+  autoView(WWVV_v,WWVV,CpuWrite);
   for (int s1 = 0; s1 < Ns; s1++){
     for (int s2 = 0; s2 < Ns; s2++){
       WWVV_v[ss]()(s1,s2)(0, 0) += lhs()(s1)(0) * rhs()(s2)(0);
@@ -1122,10 +1121,10 @@ void A2Autils<FImpl>::ContractFourQuarkColourDiagonal(const PropagatorField &WWV
 
   GridBase *grid = WWVV0.Grid();
 
-  auto WWVV0_v = WWVV0.View();
-  auto WWVV1_v = WWVV1.View();
-  auto O_trtr_v= O_trtr.View();
-  auto O_fig8_v= O_fig8.View();
+  autoView(WWVV0_v , WWVV0,CpuRead);
+  autoView(WWVV1_v , WWVV1,CpuRead);
+  autoView(O_trtr_v, O_trtr,CpuWrite);
+  autoView(O_fig8_v, O_fig8,CpuWrite);
   thread_for(ss,grid->oSites(),{
 
     typedef typename ComplexField::vector_object vobj;
@@ -1166,10 +1165,10 @@ void A2Autils<FImpl>::ContractFourQuarkColourMix(const PropagatorField &WWVV0,
 
   GridBase *grid = WWVV0.Grid();
 
-  auto WWVV0_v = WWVV0.View();
-  auto WWVV1_v = WWVV1.View();
-  auto O_trtr_v= O_trtr.View();
-  auto O_fig8_v= O_fig8.View();
+  autoView( WWVV0_v , WWVV0,CpuRead);
+  autoView( WWVV1_v , WWVV1,CpuRead);
+  autoView( O_trtr_v, O_trtr,CpuWrite);
+  autoView( O_fig8_v, O_fig8,CpuWrite);
 
   thread_for(ss,grid->oSites(),{
 
diff --git a/Grid/qcd/utils/BaryonUtils.h b/Grid/qcd/utils/BaryonUtils.h
index fa2f3376..b268b684 100644
--- a/Grid/qcd/utils/BaryonUtils.h
+++ b/Grid/qcd/utils/BaryonUtils.h
@@ -7,6 +7,7 @@
  Copyright (C) 2019
  
  Author: Felix Erben <felix.erben@ed.ac.uk>
+ Author: Raoul Hodgson <raoul.hodgson@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
@@ -46,7 +47,7 @@ public:
   typedef typename SpinMatrixField::vector_object sobj;
 
   static const int epsilon[6][3] ;
-  static const Complex epsilon_sgn[6];
+  static const Real epsilon_sgn[6];
 
   private: 
   template <class mobj, class robj>
@@ -58,9 +59,12 @@ public:
 				 const Gamma GammaA_right,
 				 const Gamma GammaB_right,
 				 const int parity,
-				 const int * wick_contractions,
+				 const bool * wick_contractions,
   				 robj &result);
   public:
+  static void Wick_Contractions(std::string qi, 
+                 std::string qf, 
+                 bool* wick_contractions);
   static void ContractBaryons(const PropagatorField &q1_left,
 				 const PropagatorField &q2_left,
 				 const PropagatorField &q3_left,
@@ -68,8 +72,7 @@ public:
 				 const Gamma GammaB_left,
 				 const Gamma GammaA_right,
 				 const Gamma GammaB_right,
-				 const char * quarks_left,
-				 const char * quarks_right,
+				 const bool* wick_contractions,
 				 const int parity,
 				 ComplexField &baryon_corr);
   template <class mobj, class robj>
@@ -80,10 +83,59 @@ public:
 				 const Gamma GammaB_left,
 				 const Gamma GammaA_right,
 				 const Gamma GammaB_right,
-				 const char * quarks_left,
-				 const char * quarks_right,
+				 const bool* wick_contractions,
 				 const int parity,
+				 const int nt,
 				 robj &result);
+  private:
+  template <class mobj, class mobj2, class robj>
+  static void Baryon_Gamma_3pt_Group1_Site(
+           const mobj &Dq1_ti,
+           const mobj2 &Dq2_spec,
+           const mobj2 &Dq3_spec,
+           const mobj &Dq4_tf,
+                   const Gamma GammaJ,
+                   const Gamma GammaBi,
+                   const Gamma GammaBf,
+           int wick_contraction,
+           robj &result);
+
+  template <class mobj, class mobj2, class robj>
+  static void Baryon_Gamma_3pt_Group2_Site(
+           const mobj2 &Dq1_spec,
+           const mobj &Dq2_ti,
+           const mobj2 &Dq3_spec,
+           const mobj &Dq4_tf,
+                   const Gamma GammaJ,
+                   const Gamma GammaBi,
+                   const Gamma GammaBf,
+           int wick_contraction,
+           robj &result);
+
+  template <class mobj, class mobj2, class robj>
+  static void Baryon_Gamma_3pt_Group3_Site(
+           const mobj2 &Dq1_spec,
+           const mobj2 &Dq2_spec,
+           const mobj &Dq3_ti,
+           const mobj &Dq4_tf,
+                   const Gamma GammaJ,
+                   const Gamma GammaBi,
+                   const Gamma GammaBf,
+           int wick_contraction,
+           robj &result);
+  public:
+  template <class mobj>
+  static void Baryon_Gamma_3pt(
+           const PropagatorField &q_ti,
+           const mobj &Dq_spec1,
+           const mobj &Dq_spec2,
+           const PropagatorField &q_tf,
+           int group,
+           int wick_contraction,
+                   const Gamma GammaJ,
+                   const Gamma GammaBi,
+                   const Gamma GammaBf,
+           SpinMatrixField &stn_corr);
   private: 
   template <class mobj, class mobj2, class robj>
   static void Sigma_to_Nucleon_Q1_Eye_site(const mobj &Dq_loop,
@@ -151,119 +203,152 @@ public:
 
 template <class FImpl> 
 const int BaryonUtils<FImpl>::epsilon[6][3] = {{0,1,2},{1,2,0},{2,0,1},{0,2,1},{2,1,0},{1,0,2}};
-template <class FImpl> 
+/*template <class FImpl> 
 const Complex BaryonUtils<FImpl>::epsilon_sgn[6] = {Complex(1),
 						    Complex(1),
 						    Complex(1),
 						    Complex(-1),
 						    Complex(-1),
 						    Complex(-1)};
+*/
+template <class FImpl> 
+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>::baryon_site(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 int parity,
-						 const int * wick_contraction,
-						 robj &result)
+                const mobj &D2,
+                const mobj &D3,
+                         const Gamma GammaA_i,
+                         const Gamma GammaB_i,
+                         const Gamma GammaA_f,
+                         const Gamma GammaB_f,
+                const int parity,
+                const bool * wick_contraction,
+                robj &result)
 {
 
-  Gamma g4(Gamma::Algebra::GammaT); //needed for parity P_\pm = 0.5*(1 \pm \gamma_4)
+    Gamma g4(Gamma::Algebra::GammaT); //needed for parity P_\pm = 0.5*(1 \pm \gamma_4)
+    
+    auto D1_GAi =  D1 * GammaA_i;
+    auto D1_GAi_g4 = D1_GAi * g4;
+    auto D1_GAi_P = 0.5*(D1_GAi + (Real)parity * D1_GAi_g4);
+    auto GAf_D1_GAi_P = GammaA_f * D1_GAi_P;
+    auto GBf_D1_GAi_P = GammaB_f * D1_GAi_P;
 
-    auto gD1a = GammaA_left * GammaA_right * D1;
-    auto gD1b = GammaA_left * g4 * GammaA_right * D1;
-    auto pD1 = 0.5* (gD1a + (double)parity * gD1b);
-    auto gD3 = GammaB_right * D3;
+    auto D2_GBi = D2 * GammaB_i;
+    auto GBf_D2_GBi = GammaB_f * D2_GBi;
+    auto GAf_D2_GBi = GammaA_f * D2_GBi;
 
-    auto D2g = D2 * GammaB_left;
-    auto pD1g = pD1 * GammaB_left;
-    auto gD3g = gD3 * GammaB_left;
+    auto GBf_D3 = GammaB_f * D3;
+    auto GAf_D3 = GammaA_f * D3;
 
-    for (int ie_left=0; ie_left < 6 ; ie_left++){
-      int a_left = epsilon[ie_left][0]; //a
-      int b_left = epsilon[ie_left][1]; //b
-      int c_left = epsilon[ie_left][2]; //c
-      for (int ie_right=0; ie_right < 6 ; ie_right++){
-        int a_right = epsilon[ie_right][0]; //a'
-        int b_right = epsilon[ie_right][1]; //b'
-        int c_right = epsilon[ie_right][2]; //c'
-	Complex ee = epsilon_sgn[ie_left] * epsilon_sgn[ie_right];
+    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 gamma_left=0; gamma_left<Ns; gamma_left++){
-            auto eepD1 = ee * pD1()(gamma_left,gamma_left)(c_right,c_left);
-	  for (int alpha_right=0; alpha_right<Ns; alpha_right++){
-	  for (int beta_left=0; beta_left<Ns; beta_left++){
-	    auto D2g_ab = D2g()(alpha_right,beta_left)(a_right,a_left);
-	    auto gD3_ab = gD3()(alpha_right,beta_left)(b_right,b_left);
-	        result()()() += eepD1*D2g_ab*gD3_ab;
-          }}}
-  	}	  
+        if (wick_contraction[0]){
+            for (int rho=0; rho<Ns; rho++){
+                auto GAf_D1_GAi_P_rr_cc = GAf_D1_GAi_P()(rho,rho)(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()()() += ee  * GAf_D1_GAi_P_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 gamma_left=0; gamma_left<Ns; gamma_left++){
-	  for (int alpha_right=0; alpha_right<Ns; alpha_right++){
-            auto gD3_ag = gD3()(alpha_right,gamma_left)(b_right,c_left);
-	  for (int beta_left=0; beta_left<Ns; beta_left++){
-            auto eepD1g_gb = ee * pD1g()(gamma_left,beta_left)(c_right,a_left);
-	    auto D2_ab = D2()(alpha_right,beta_left)(a_right,b_left);
-		result()()() += eepD1g_gb*D2_ab*gD3_ag;
-          }}}
-        }	  
+        if (wick_contraction[1]){
+            for (int rho=0; rho<Ns; rho++){
+            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+                auto D1_GAi_P_ar_ac = D1_GAi_P()(alpha_f,rho)(a_f,c_i);
+                for (int beta_i=0; beta_i<Ns; beta_i++){
+                    result()()() += ee  * D1_GAi_P_ar_ac
+                                        * GBf_D2_GBi    ()(alpha_f,beta_i)(b_f,a_i)
+                                        * GAf_D3        ()(rho,beta_i)(c_f,b_i);
+                }
+            }}
+        }   
         //This is the \delta_{456}^{312} part
-	if (wick_contraction[2]){
-	  for (int gamma_left=0; gamma_left<Ns; gamma_left++){
-	  for (int alpha_right=0; alpha_right<Ns; alpha_right++){
-	    auto D2_ag = D2()(alpha_right,gamma_left)(a_right,c_left);
-	  for (int beta_left=0; beta_left<Ns; beta_left++){
-            auto eepD1_gb = ee * pD1()(gamma_left,beta_left)(c_right,b_left);
-	    auto gD3g_ab = gD3g()(alpha_right,beta_left)(b_right,a_left);
-		result()()() += eepD1_gb*D2_ag*gD3g_ab;
-          }}}
-        }	  
+        if (wick_contraction[2]){
+            for (int rho=0; rho<Ns; rho++){
+                for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+                auto GBf_D1_GAi_P_ar_bc = GBf_D1_GAi_P()(alpha_f,rho)(b_f,c_i);
+                for (int beta_i=0; beta_i<Ns; beta_i++){
+                    result()()() += ee  * GBf_D1_GAi_P_ar_bc
+                                        * GAf_D2_GBi    ()(rho,beta_i)(c_f,a_i)
+                                        * D3            ()(alpha_f,beta_i)(a_f,b_i);
+                }
+            }}
+        }   
         //This is the \delta_{456}^{132} part
-	if (wick_contraction[3]){
-	  for (int gamma_left=0; gamma_left<Ns; gamma_left++){
-            auto eepD1 = ee * pD1()(gamma_left,gamma_left)(c_right,c_left);
-	  for (int alpha_right=0; alpha_right<Ns; alpha_right++){
-	  for (int beta_left=0; beta_left<Ns; beta_left++){
-	    auto D2_ab = D2()(alpha_right,beta_left)(a_right,b_left);
-	    auto gD3g_ab = gD3g()(alpha_right,beta_left)(b_right,a_left);
-    		result()()() -= eepD1*D2_ab*gD3g_ab;
-          }}}
-        }	  
+        if (wick_contraction[3]){
+            for (int rho=0; rho<Ns; rho++){
+                auto GAf_D1_GAi_P_rr_cc = GAf_D1_GAi_P()(rho,rho)(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()()() -= ee  * GAf_D1_GAi_P_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 gamma_left=0; gamma_left<Ns; gamma_left++){
-	  for (int alpha_right=0; alpha_right<Ns; alpha_right++){
-            auto gD3_ag = gD3()(alpha_right,gamma_left)(b_right,c_left);
-	  for (int beta_left=0; beta_left<Ns; beta_left++){
-            auto eepD1_gb = ee * pD1()(gamma_left,beta_left)(c_right,b_left);
-	    auto D2g_ab = D2g()(alpha_right,beta_left)(a_right,a_left);
-		result()()() -= eepD1_gb*D2g_ab*gD3_ag;
-          }}}
-        }	  
+        if (wick_contraction[4]){
+            for (int rho=0; rho<Ns; rho++){
+            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+                auto GBf_D1_GAi_P_ar_bc = GBf_D1_GAi_P()(alpha_f,rho)(b_f,c_i);
+                for (int beta_i=0; beta_i<Ns; beta_i++){
+                    result()()() -= ee  * GBf_D1_GAi_P_ar_bc
+                                        * D2_GBi    ()(alpha_f,beta_i)(a_f,a_i)
+                                        * GAf_D3    ()(rho,beta_i)(c_f,b_i);
+                }
+            }}
+        }   
         //This is the \delta_{456}^{213} part
-	if (wick_contraction[5]){
-	  for (int gamma_left=0; gamma_left<Ns; gamma_left++){
-	  for (int alpha_right=0; alpha_right<Ns; alpha_right++){
-	    auto D2_ag = D2()(alpha_right,gamma_left)(a_right,c_left);
-	  for (int beta_left=0; beta_left<Ns; beta_left++){
-            auto eepD1g_gb = ee * pD1g()(gamma_left,beta_left)(c_right,a_left);
-	    auto gD3_ab = gD3()(alpha_right,beta_left)(b_right,b_left);
-    	        result()()() -= eepD1g_gb*D2_ag*gD3_ab;
-          }}}
-        }	  
-      }
+        if (wick_contraction[5]){
+            for (int rho=0; rho<Ns; rho++){
+            for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+                auto D1_GAi_P_ar_ac = D1_GAi_P()(alpha_f,rho)(a_f,c_i);
+                for (int beta_i=0; beta_i<Ns; beta_i++){
+                    result()()() -= ee  * D1_GAi_P_ar_ac
+                                        * GAf_D2_GBi    ()(rho,beta_i)(c_f,a_i)
+                                        * GBf_D3        ()(alpha_f,beta_i)(b_f,b_i);
+                }
+            }}
+        }
+    }}
+}
+
+/* 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>::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
+                                    && qi[0] == qf[epsilon[ie][0]] 
+                                    && qi[1] == qf[epsilon[ie][1]] 
+                                    && qi[2] == qf[epsilon[ie][2]]);
     }
 }
 
+/* The array wick_contractions must be of length 6. The order     * 
+ * corresponds to the to that shown in the Hadrons documentation  *
+ * at https://aportelli.github.io/Hadrons-doc/#/mcontraction      *
+ * This can be computed from the quark flavours using the         *
+ * Wick_Contractions function above                               */
 template<class FImpl>
 void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
 						 const PropagatorField &q2_left,
@@ -272,8 +357,7 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
 				                 const Gamma GammaB_left,
 				                 const Gamma GammaA_right,
 		                 		 const Gamma GammaB_right,
-						 const char * quarks_left,
-						 const char * quarks_right,
+						 const bool* wick_contractions,
 						 const int parity,
 						 ComplexField &baryon_corr)
 {
@@ -281,38 +365,53 @@ 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 << "Contraction <" << quarks_right[0] << quarks_right[1] << quarks_right[2] << "|" << quarks_left[0] << quarks_left[1] << quarks_left[2] << ">" << std::endl;
-    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;
+  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");
 
   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);
 
-  int wick_contraction[6];
-  for (int ie=0; ie < 6 ; ie++)
-    wick_contraction[ie] = (quarks_left[0] == quarks_right[epsilon[ie][0]] && quarks_left[1] == quarks_right[epsilon[ie][1]] && quarks_left[2] == quarks_right[epsilon[ie][2]]) ? 1 : 0;
-
-  auto vbaryon_corr= baryon_corr.View();
-  auto v1 = q1_left.View();
-  auto v2 = q2_left.View();
-  auto v3 = q3_left.View();
-
- // accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
-  thread_for(ss,grid->oSites(),{
-  //for(int ss=0; ss < grid->oSites(); ss++){
+  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];
-
     vobj result=Zero();
-    baryon_site(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,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 << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s " << std::endl;
+
 }
+
+/* The array wick_contractions must be of length 6. The order     * 
+ * corresponds to the to that shown in the Hadrons documentation  *
+ * at https://aportelli.github.io/Hadrons-doc/#/mcontraction      *
+ * This can also be computed from the quark flavours using the    *
+ * Wick_Contractions function above                               */
 template <class FImpl>
 template <class mobj, class robj>
 void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
@@ -322,34 +421,363 @@ void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
 				                 const Gamma GammaB_left,
 				                 const Gamma GammaA_right,
 		                 		 const Gamma GammaB_right,
-						 const char * quarks_left,
-						 const char * quarks_right,
+						 const bool* wick_contractions,
 						 const int parity,
+						 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");
 
-  std::cout << "Contraction <" << quarks_right[0] << quarks_right[1] << quarks_right[2] << "|" << quarks_left[0] << quarks_left[1] << quarks_left[2] << ">" << std::endl;
-    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;
+  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");
 
-  int wick_contraction[6];
-  for (int ie=0; ie < 6 ; ie++)
-    wick_contraction[ie] = (quarks_left[0] == quarks_right[epsilon[ie][0]] && quarks_left[1] == quarks_right[epsilon[ie][1]] && quarks_left[2] == quarks_right[epsilon[ie][2]]) ? 1 : 0;
-
-     result=Zero();
-     baryon_site<decltype(D1),decltype(result)>(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,result);
+  for (int t=0; t<nt; t++) {
+    baryon_site(D1[t],D2[t],D3[t],GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contractions,result[t]);
+  }
 }
 
 /***********************************************************************
  * End of Baryon 2pt-function code.                                    *
  *                                                                     *
+ * The following code is for baryonGamma3pt function                   *
+ **********************************************************************/
+
+/* Dq1_ti is a quark line from t_i to t_J
+ * Dq2_spec is a quark line from t_i to t_f
+ * Dq3_spec is a quark line from t_i to t_f
+ * 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>::Baryon_Gamma_3pt_Group1_Site(
+                        const mobj &Dq1_ti,
+                        const mobj2 &Dq2_spec,
+                        const mobj2 &Dq3_spec,
+                        const mobj &Dq4_tf,
+                                const Gamma GammaJ,
+                                const Gamma GammaBi,
+                                const Gamma GammaBf,
+                        int wick_contraction,
+                        robj &result)
+{
+    Gamma g5(Gamma::Algebra::Gamma5); 
+
+    auto adjD4_g_D1     = g5 * adj(Dq4_tf) * g5 * GammaJ * Dq1_ti;
+    auto Gf_adjD4_g_D1  = GammaBf * adjD4_g_D1;
+    auto D2_Gi          = Dq2_spec * GammaBi;
+    auto Gf_D2_Gi       = GammaBf * D2_Gi;
+    auto Gf_D3          = GammaBf * Dq3_spec;
+
+    int a_f, b_f, c_f;
+    int a_i, b_i, c_i;
+
+    Real ee;
+
+    for (int ie_f=0; ie_f < 6 ; ie_f++){
+        a_f = epsilon[ie_f][0]; //a
+        b_f = epsilon[ie_f][1]; //b
+        c_f = epsilon[ie_f][2]; //c
+    for (int ie_i=0; ie_i < 6 ; ie_i++){
+        a_i = epsilon[ie_i][0]; //a'
+        b_i = epsilon[ie_i][1]; //b'
+        c_i = epsilon[ie_i][2]; //c'
+
+        ee = epsilon_sgn[ie_f] * epsilon_sgn[ie_i];
+
+        for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+        for (int beta_i=0; beta_i<Ns; beta_i++){
+            auto D2_Gi_ab_aa        = D2_Gi     ()(alpha_f,beta_i)(a_f,a_i);
+            auto Gf_D3_ab_bb        = Gf_D3     ()(alpha_f,beta_i)(b_f,b_i);
+            auto Gf_D2_Gi_ab_ba     = Gf_D2_Gi  ()(alpha_f,beta_i)(b_f,a_i);
+            auto Dq3_spec_ab_ab     = Dq3_spec  ()(alpha_f,beta_i)(a_f,b_i);
+
+            for (int gamma_i=0; gamma_i<Ns; gamma_i++){
+                auto ee_adjD4_g_D1_ag_ac        = ee * adjD4_g_D1   ()(alpha_f,gamma_i)(a_f,c_i);
+                auto ee_Gf_adjD4_g_D1_ag_bc     = ee * Gf_adjD4_g_D1()(alpha_f,gamma_i)(b_f,c_i);
+            for (int gamma_f=0; gamma_f<Ns; gamma_f++){
+                auto ee_adjD4_g_D1_gg_cc        = ee * adjD4_g_D1   ()(gamma_f,gamma_i)(c_f,c_i);
+                auto Dq3_spec_gb_cb             = Dq3_spec          ()(gamma_f,beta_i)(c_f,b_i);
+                auto D2_Gi_gb_ca                = D2_Gi             ()(gamma_f,beta_i)(c_f,a_i);
+
+
+                if(wick_contraction == 1) { // Do contraction I1
+                    result()(gamma_f,gamma_i)() -= ee_adjD4_g_D1_gg_cc
+                                                        * D2_Gi_ab_aa
+                                                        * Gf_D3_ab_bb;
+                }
+                if(wick_contraction == 2) { // Do contraction I2
+                    result()(gamma_f,gamma_i)() -= ee_adjD4_g_D1_ag_ac
+                                                        * Gf_D2_Gi_ab_ba
+                                                        * Dq3_spec_gb_cb;
+                }
+                if(wick_contraction == 3) { // Do contraction I3
+                    result()(gamma_f,gamma_i)() -= ee_Gf_adjD4_g_D1_ag_bc
+                                                        * D2_Gi_gb_ca
+                                                        * Dq3_spec_ab_ab;
+                }
+                if(wick_contraction == 4) { // Do contraction I4
+                    result()(gamma_f,gamma_i)() += ee_adjD4_g_D1_gg_cc
+                                                        * Gf_D2_Gi_ab_ba
+                                                        * Dq3_spec_ab_ab;
+                }
+                if(wick_contraction == 5) { // Do contraction I5
+                    result()(gamma_f,gamma_i)() += ee_Gf_adjD4_g_D1_ag_bc
+                                                        * D2_Gi_ab_aa
+                                                        * Dq3_spec_gb_cb;
+                }
+                if(wick_contraction == 6) { // Do contraction I6
+                    result()(gamma_f,gamma_i)() += ee_adjD4_g_D1_ag_ac
+                                                        * D2_Gi_gb_ca
+                                                        * Gf_D3_ab_bb;
+                }
+            }}
+        }}
+    }}
+}
+
+/* Dq1_spec is a quark line from t_i to t_f
+ * Dq2_ti is a quark line from t_i to t_J
+ * Dq3_spec is a quark line from t_i to t_f
+ * 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>::Baryon_Gamma_3pt_Group2_Site(
+                        const mobj2 &Dq1_spec,
+                        const mobj &Dq2_ti,
+                        const mobj2 &Dq3_spec,
+                        const mobj &Dq4_tf,
+                                const Gamma GammaJ,
+                                const Gamma GammaBi,
+                                const Gamma GammaBf,
+                        int wick_contraction,
+                        robj &result)
+{
+    Gamma g5(Gamma::Algebra::Gamma5); 
+
+    auto adjD4_g_D2_Gi      = g5 * adj(Dq4_tf) * g5 * GammaJ * Dq2_ti * GammaBi;
+    auto Gf_adjD4_g_D2_Gi   = GammaBf * adjD4_g_D2_Gi;
+    auto Gf_D1              = GammaBf * Dq1_spec;
+    auto Gf_D3              = GammaBf * Dq3_spec;
+
+    int a_f, b_f, c_f;
+    int a_i, b_i, c_i;
+
+    Real ee;
+
+    for (int ie_f=0; ie_f < 6 ; ie_f++){
+        a_f = epsilon[ie_f][0]; //a
+        b_f = epsilon[ie_f][1]; //b
+        c_f = epsilon[ie_f][2]; //c
+    for (int ie_i=0; ie_i < 6 ; ie_i++){
+        a_i = epsilon[ie_i][0]; //a'
+        b_i = epsilon[ie_i][1]; //b'
+        c_i = epsilon[ie_i][2]; //c'
+
+        ee = epsilon_sgn[ie_f] * epsilon_sgn[ie_i];
+
+        for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+        for (int beta_i=0; beta_i<Ns; beta_i++){
+            auto adjD4_g_D2_Gi_ab_aa        = adjD4_g_D2_Gi     ()(alpha_f,beta_i)(a_f,a_i);
+            auto Gf_D3_ab_bb                = Gf_D3             ()(alpha_f,beta_i)(b_f,b_i);
+            auto Gf_adjD4_g_D2_Gi_ab_ba     = Gf_adjD4_g_D2_Gi  ()(alpha_f,beta_i)(b_f,a_i);
+            auto Dq3_spec_ab_ab             = Dq3_spec          ()(alpha_f,beta_i)(a_f,b_i);
+
+            for (int gamma_i=0; gamma_i<Ns; gamma_i++){ 
+                auto ee_Dq1_spec_ag_ac      = ee * Dq1_spec     ()(alpha_f,gamma_i)(a_f,c_i);
+                auto ee_Gf_D1_ag_bc         = ee * Gf_D1        ()(alpha_f,gamma_i)(b_f,c_i);
+            for (int gamma_f=0; gamma_f<Ns; gamma_f++){
+                auto ee_Dq1_spec_gg_cc      = ee * Dq1_spec     ()(gamma_f,gamma_i)(c_f,c_i);
+                auto Dq3_spec_gb_cb         = Dq3_spec          ()(gamma_f,beta_i)(c_f,b_i);
+                auto adjD4_g_D2_Gi_gb_ca    = adjD4_g_D2_Gi     ()(gamma_f,beta_i)(c_f,a_i);
+
+                if(wick_contraction == 1) { // Do contraction II1
+                    result()(gamma_f,gamma_i)() -= ee_Dq1_spec_gg_cc
+                                                        * adjD4_g_D2_Gi_ab_aa
+                                                        * Gf_D3_ab_bb;
+                }
+                if(wick_contraction == 2) { // Do contraction II2
+                    result()(gamma_f,gamma_i)() -= ee_Dq1_spec_ag_ac
+                                                        * Gf_adjD4_g_D2_Gi_ab_ba
+                                                        * Dq3_spec_gb_cb;
+                }
+                if(wick_contraction == 3) { // Do contraction II3
+                    result()(gamma_f,gamma_i)() -= ee_Gf_D1_ag_bc
+                                                        * adjD4_g_D2_Gi_gb_ca
+                                                        * Dq3_spec_ab_ab;
+                }
+                if(wick_contraction == 4) { // Do contraction II4
+                    result()(gamma_f,gamma_i)() += ee_Dq1_spec_gg_cc
+                                                        * Gf_adjD4_g_D2_Gi_ab_ba
+                                                        * Dq3_spec_ab_ab;
+                }
+                if(wick_contraction == 5) { // Do contraction II5
+                    result()(gamma_f,gamma_i)() += ee_Gf_D1_ag_bc
+                                                        * adjD4_g_D2_Gi_ab_aa
+                                                        * Dq3_spec_gb_cb;
+                }
+                if(wick_contraction == 6) { // Do contraction II6
+                    result()(gamma_f,gamma_i)() += ee_Dq1_spec_ag_ac
+                                                        * adjD4_g_D2_Gi_gb_ca
+                                                        * Gf_D3_ab_bb;
+                }
+            }}
+        }}
+    }}
+}
+
+/* Dq1_spec is a quark line from t_i to t_f
+ * Dq2_spec is a quark line from t_i to t_f
+ * Dq3_ti is a quark line from t_i to t_J
+ * 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>::Baryon_Gamma_3pt_Group3_Site(
+                        const mobj2 &Dq1_spec,
+                        const mobj2 &Dq2_spec,
+                        const mobj &Dq3_ti,
+                        const mobj &Dq4_tf,
+                                const Gamma GammaJ,
+                                const Gamma GammaBi,
+                                const Gamma GammaBf,
+                        int wick_contraction,
+                        robj &result)
+{
+    Gamma g5(Gamma::Algebra::Gamma5);
+
+    auto adjD4_g_D3     = g5 * adj(Dq4_tf) * g5 * GammaJ * Dq3_ti;
+    auto Gf_adjD4_g_D3  = GammaBf * adjD4_g_D3;
+    auto Gf_D1          = GammaBf * Dq1_spec;
+    auto D2_Gi          = Dq2_spec * GammaBi;
+    auto Gf_D2_Gi       = GammaBf * D2_Gi;
+
+    int a_f, b_f, c_f;
+    int a_i, b_i, c_i;
+
+    Real ee;
+
+    for (int ie_f=0; ie_f < 6 ; ie_f++){
+        a_f = epsilon[ie_f][0]; //a
+        b_f = epsilon[ie_f][1]; //b
+        c_f = epsilon[ie_f][2]; //c
+    for (int ie_i=0; ie_i < 6 ; ie_i++){
+        a_i = epsilon[ie_i][0]; //a'
+        b_i = epsilon[ie_i][1]; //b'
+        c_i = epsilon[ie_i][2]; //c'
+
+        ee = epsilon_sgn[ie_f] * epsilon_sgn[ie_i];
+
+        for (int alpha_f=0; alpha_f<Ns; alpha_f++){
+        for (int beta_i=0; beta_i<Ns; beta_i++){
+            auto D2_Gi_ab_aa            = D2_Gi         ()(alpha_f,beta_i)(a_f,a_i);
+            auto Gf_adjD4_g_D3_ab_bb    = Gf_adjD4_g_D3 ()(alpha_f,beta_i)(b_f,b_i);
+            auto Gf_D2_Gi_ab_ba         = Gf_D2_Gi      ()(alpha_f,beta_i)(b_f,a_i);
+            auto adjD4_g_D3_ab_ab       = adjD4_g_D3    ()(alpha_f,beta_i)(a_f,b_i);
+
+            for (int gamma_i=0; gamma_i<Ns; gamma_i++) {
+                auto ee_Dq1_spec_ag_ac  = ee * Dq1_spec ()(alpha_f,gamma_i)(a_f,c_i);
+                auto ee_Gf_D1_ag_bc     = ee * Gf_D1    ()(alpha_f,gamma_i)(b_f,c_i);
+            for (int gamma_f=0; gamma_f<Ns; gamma_f++) {
+                auto ee_Dq1_spec_gg_cc  = ee * Dq1_spec ()(gamma_f,gamma_i)(c_f,c_i);
+                auto adjD4_g_D3_gb_cb   = adjD4_g_D3    ()(gamma_f,beta_i)(c_f,b_i);
+                auto D2_Gi_gb_ca        = D2_Gi         ()(gamma_f,beta_i)(c_f,a_i);
+
+                if(wick_contraction == 1) { // Do contraction III1
+                    result()(gamma_f,gamma_i)() -= ee_Dq1_spec_gg_cc
+                                                        * D2_Gi_ab_aa
+                                                        * Gf_adjD4_g_D3_ab_bb;
+                }
+                if(wick_contraction == 2) { // Do contraction III2
+                    result()(gamma_f,gamma_i)() -= ee_Dq1_spec_ag_ac
+                                                        * Gf_D2_Gi_ab_ba
+                                                        * adjD4_g_D3_gb_cb;
+                }
+                if(wick_contraction == 3) { // Do contraction III3
+                    result()(gamma_f,gamma_i)() -= ee_Gf_D1_ag_bc
+                                                        * D2_Gi_gb_ca
+                                                        * adjD4_g_D3_ab_ab;
+                }
+                if(wick_contraction == 4) { // Do contraction III4
+                    result()(gamma_f,gamma_i)() += ee_Dq1_spec_gg_cc
+                                                        * Gf_D2_Gi_ab_ba
+                                                        * adjD4_g_D3_ab_ab;
+                }
+                if(wick_contraction == 5) { // Do contraction III5
+                    result()(gamma_f,gamma_i)() += ee_Gf_D1_ag_bc
+                                                        * D2_Gi_ab_aa
+                                                        * adjD4_g_D3_gb_cb;
+                }
+                if(wick_contraction == 6) { // Do contraction III6
+                    result()(gamma_f,gamma_i)() += ee_Dq1_spec_ag_ac
+                                                        * D2_Gi_gb_ca
+                                                        * Gf_adjD4_g_D3_ab_bb;
+                }
+            }}
+        }}
+    }}
+}
+
+/* The group indicates which inital state quarks the current is  * 
+ * connected to. It must be in the range 1-3.                    *
+ * The wick_contraction must be in the range 1-6 correspond to   *
+ * the contractions given in the Hadrons documentation at        *
+ * https://aportelli.github.io/Hadrons-doc/#/mcontraction        */
+template<class FImpl>
+template <class mobj>
+void BaryonUtils<FImpl>::Baryon_Gamma_3pt(
+                        const PropagatorField &q_ti,
+                        const mobj &Dq_spec1,
+                        const mobj &Dq_spec2,
+                        const PropagatorField &q_tf,
+                        int group,
+                        int wick_contraction,
+                                const Gamma GammaJ,
+                                const Gamma GammaBi,
+                                const Gamma GammaBf,
+                        SpinMatrixField &stn_corr)
+{
+    GridBase *grid = q_tf.Grid();
+
+    autoView( vcorr, stn_corr, CpuWrite);
+    autoView( vq_ti , q_ti, CpuRead);
+    autoView( vq_tf , q_tf, CpuRead);
+
+    if (group == 1) {
+        accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
+            auto Dq_ti = vq_ti[ss];
+            auto Dq_tf = vq_tf[ss];
+            sobj result=Zero();
+            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) {
+        accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
+            auto Dq_ti = vq_ti[ss];
+            auto Dq_tf = vq_tf[ss];
+            sobj result=Zero();
+            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) {
+        accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
+            auto Dq_ti = vq_ti[ss];
+            auto Dq_tf = vq_tf[ss];
+            sobj result=Zero();
+            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
+    }
+}
+
+
+/***********************************************************************
+ * End of BaryonGamma3pt-function code.                                *
+ *																	   *
  * The following code is for Sigma -> N rare hypeon decays             *
  **********************************************************************/
 
@@ -590,13 +1018,12 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
 
   GridBase *grid = qs_ti.Grid();
 
-  auto vcorr= stn_corr.View();
-  auto vq_loop = qq_loop.View();
-  auto vd_tf = qd_tf.View();
-  auto vs_ti = qs_ti.View();
+  autoView( vcorr, stn_corr, CpuWrite);
+  autoView( vq_loop , qq_loop, CpuRead);
+  autoView( vd_tf , qd_tf, CpuRead);
+  autoView( vs_ti , qs_ti, CpuRead);
 
- // accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
-  thread_for(ss,grid->oSites(),{
+  accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
     auto Dq_loop = vq_loop[ss];
     auto Dd_tf = vd_tf[ss];
     auto Ds_ti = vs_ti[ss];
@@ -631,12 +1058,11 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
 
   GridBase *grid = qs_ti.Grid();
 
-  auto vcorr= stn_corr.View();
-  auto vq_ti = qq_ti.View();
-  auto vq_tf = qq_tf.View();
-  auto vd_tf = qd_tf.View();
-  auto vs_ti = qs_ti.View();
-
+  autoView( vcorr , stn_corr, CpuWrite);
+  autoView( vq_ti , qq_ti, CpuRead);
+  autoView( vq_tf , qq_tf, CpuRead);
+  autoView( vd_tf , qd_tf, CpuRead);
+  autoView( vs_ti , qs_ti, CpuRead);
  // accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
   thread_for(ss,grid->oSites(),{
     auto Dq_ti = vq_ti[ss];
diff --git a/Grid/qcd/utils/LinalgUtils.h b/Grid/qcd/utils/LinalgUtils.h
index 56f8f164..1e016e4e 100644
--- a/Grid/qcd/utils/LinalgUtils.h
+++ b/Grid/qcd/utils/LinalgUtils.h
@@ -47,8 +47,8 @@ void axpibg5x(Lattice<vobj> &z,const Lattice<vobj> &x,Coeff a,Coeff b)
   GridBase *grid=x.Grid();
 
   Gamma G5(Gamma::Algebra::Gamma5);
-  auto x_v = x.View();
-  auto z_v = z.View();
+  autoView(x_v, x, AcceleratorRead);
+  autoView(z_v, z, AcceleratorWrite);
   accelerator_for( ss, x_v.size(),vobj::Nsimd(), {
     auto tmp = a*x_v(ss) + G5*(b*timesI(x_v(ss)));
     coalescedWrite(z_v[ss],tmp);
@@ -63,9 +63,9 @@ void axpby_ssp(Lattice<vobj> &z, Coeff a,const Lattice<vobj> &x,Coeff b,const La
   conformable(x,z);
   GridBase *grid=x.Grid();
   int Ls = grid->_rdimensions[0];
-  auto x_v = x.View();
-  auto y_v = y.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   // FIXME -- need a new class of accelerator_loop to implement this
   //
   uint64_t nloop = grid->oSites()/Ls;
@@ -85,9 +85,9 @@ void ag5xpby_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const L
   GridBase *grid=x.Grid();
   int Ls = grid->_rdimensions[0];
   Gamma G5(Gamma::Algebra::Gamma5);
-  auto x_v = x.View();
-  auto y_v = y.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,vobj::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -104,9 +104,9 @@ void axpbg5y_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const L
   conformable(x,z);
   GridBase *grid=x.Grid();
   int Ls = grid->_rdimensions[0];
-  auto x_v = x.View();
-  auto y_v = y.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   Gamma G5(Gamma::Algebra::Gamma5);
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,vobj::Nsimd(),{
@@ -125,9 +125,9 @@ void ag5xpbg5y_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const
   GridBase *grid=x.Grid();
   int Ls = grid->_rdimensions[0];
 
-  auto x_v = x.View();
-  auto y_v = y.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   Gamma G5(Gamma::Algebra::Gamma5);
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,vobj::Nsimd(),{
@@ -147,9 +147,9 @@ void axpby_ssp_pminus(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,co
   GridBase *grid=x.Grid();
   int Ls = grid->_rdimensions[0];
 
-  auto x_v = x.View();
-  auto y_v = y.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,vobj::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -168,9 +168,9 @@ void axpby_ssp_pplus(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,con
   conformable(x,z);
   GridBase *grid=x.Grid();
   int Ls = grid->_rdimensions[0];
-  auto x_v = x.View();
-  auto y_v = y.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,vobj::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -189,8 +189,8 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
   conformable(x,z);
   int Ls = grid->_rdimensions[0];
   Gamma G5(Gamma::Algebra::Gamma5);
-  auto x_v = x.View();
-  auto z_v = z.View();
+  autoView( x_v, x, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,vobj::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -222,8 +222,8 @@ void G5C(Lattice<iVector<CComplex, nbasis>> &z, const Lattice<iVector<CComplex,
   static_assert(nbasis % 2 == 0, "");
   int nb = nbasis / 2;
 
-  auto z_v = z.View();
-  auto x_v = x.View();
+  autoView( z_v, z, AcceleratorWrite);
+  autoView( x_v, x, AcceleratorRead);
   accelerator_for(ss,grid->oSites(),CComplex::Nsimd(),
   {
     for(int n = 0; n < nb; ++n) {
diff --git a/Grid/qcd/utils/SUn.h b/Grid/qcd/utils/SUn.h
index 7ad80d00..0cc0cc1a 100644
--- a/Grid/qcd/utils/SUn.h
+++ b/Grid/qcd/utils/SUn.h
@@ -222,11 +222,11 @@ public:
     conformable(subgroup, Determinant);
     int i0, i1;
     su2SubGroupIndex(i0, i1, su2_index);
-    auto subgroup_v = subgroup.View();
-    auto source_v   = source.View();
-    auto Determinant_v = Determinant.View();
 
-    thread_for(ss, grid->oSites(), {
+    autoView( subgroup_v , subgroup,AcceleratorWrite);
+    autoView( source_v   , source,AcceleratorRead);
+    autoView( Determinant_v , Determinant,AcceleratorWrite);
+    accelerator_for(ss, grid->oSites(), 1, {
 
       subgroup_v[ss]()()(0, 0) = source_v[ss]()()(i0, i0);
       subgroup_v[ss]()()(0, 1) = source_v[ss]()()(i0, i1);
@@ -257,15 +257,16 @@ public:
     su2SubGroupIndex(i0, i1, su2_index);
 
     dest = 1.0;  // start out with identity
-    auto dest_v = dest.View();
-    auto subgroup_v = subgroup.View();
-    thread_for(ss, grid->oSites(),
+    autoView( dest_v , dest, AcceleratorWrite);
+    autoView( subgroup_v, subgroup, AcceleratorRead);
+    accelerator_for(ss, grid->oSites(),1,
     {
       dest_v[ss]()()(i0, i0) = subgroup_v[ss]()()(0, 0);
       dest_v[ss]()()(i0, i1) = subgroup_v[ss]()()(0, 1);
       dest_v[ss]()()(i1, i0) = subgroup_v[ss]()()(1, 0);
       dest_v[ss]()()(i1, i1) = subgroup_v[ss]()()(1, 1);
     });
+
   }
 
   ///////////////////////////////////////////////
@@ -608,8 +609,8 @@ public:
 
   // reunitarise??
   template <typename LatticeMatrixType>
-  static void LieRandomize(GridParallelRNG &pRNG, LatticeMatrixType &out,
-                           double scale = 1.0) {
+  static void LieRandomize(GridParallelRNG &pRNG, LatticeMatrixType &out, double scale = 1.0) 
+  {
     GridBase *grid = out.Grid();
 
     typedef typename LatticeMatrixType::vector_type vector_type;
@@ -618,8 +619,7 @@ public:
     typedef iSinglet<vector_type> vTComplexType;
 
     typedef Lattice<vTComplexType> LatticeComplexType;
-    typedef typename GridTypeMapper<
-      typename LatticeMatrixType::vector_object>::scalar_object MatrixType;
+    typedef typename GridTypeMapper<typename LatticeMatrixType::vector_object>::scalar_object MatrixType;
 
     LatticeComplexType ca(grid);
     LatticeMatrixType lie(grid);
@@ -629,6 +629,7 @@ public:
     MatrixType ta;
 
     lie = Zero();
+
     for (int a = 0; a < AdjointDimension; a++) {
       random(pRNG, ca);
 
@@ -640,6 +641,7 @@ public:
       la = ci * ca * ta;
 
       lie = lie + la;  // e^{i la ta}
+
     }
     taExp(lie, out);
   }
diff --git a/Grid/serialisation/Serialisation.h b/Grid/serialisation/Serialisation.h
index c95226b1..e14120af 100644
--- a/Grid/serialisation/Serialisation.h
+++ b/Grid/serialisation/Serialisation.h
@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include "BinaryIO.h"
 #include "TextIO.h"
 #include "XmlIO.h"
-#ifndef GRID_NVCC
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
 #include "JSON_IO.h"
 #endif
 
diff --git a/Grid/simd/Fujitsu_A64FX_asm_double.h b/Grid/simd/Fujitsu_A64FX_asm_double.h
new file mode 100644
index 00000000..76c556d7
--- /dev/null
+++ b/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" \
+); 
+
diff --git a/Grid/simd/Fujitsu_A64FX_asm_single.h b/Grid/simd/Fujitsu_A64FX_asm_single.h
new file mode 100644
index 00000000..d809f83b
--- /dev/null
+++ b/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" \
+); 
+
diff --git a/Grid/simd/Fujitsu_A64FX_intrin_double.h b/Grid/simd/Fujitsu_A64FX_intrin_double.h
new file mode 100644
index 00000000..232610f2
--- /dev/null
+++ b/Grid/simd/Fujitsu_A64FX_intrin_double.h
@@ -0,0 +1,601 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Fujitsu_A64FX_intrin_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} };\
+    svfloat64_t result_00;        \
+    svfloat64_t result_01;        \
+    svfloat64_t result_02;        \
+    svfloat64_t result_10;        \
+    svfloat64_t result_11;        \
+    svfloat64_t result_12;        \
+    svfloat64_t result_20;        \
+    svfloat64_t result_21;        \
+    svfloat64_t result_22;        \
+    svfloat64_t result_30;        \
+    svfloat64_t result_31;        \
+    svfloat64_t result_32;        \
+    svfloat64_t Chi_00;        \
+    svfloat64_t Chi_01;        \
+    svfloat64_t Chi_02;        \
+    svfloat64_t Chi_10;        \
+    svfloat64_t Chi_11;        \
+    svfloat64_t Chi_12;        \
+    svfloat64_t UChi_00;        \
+    svfloat64_t UChi_01;        \
+    svfloat64_t UChi_02;        \
+    svfloat64_t UChi_10;        \
+    svfloat64_t UChi_11;        \
+    svfloat64_t UChi_12;        \
+    svfloat64_t U_00;        \
+    svfloat64_t U_10;        \
+    svfloat64_t U_20;        \
+    svfloat64_t U_01;        \
+    svfloat64_t U_11;        \
+    svfloat64_t U_21;        \
+    svbool_t pg1;        \
+    pg1 = svptrue_b64();        \
+    svuint64_t table0; \
+    svfloat64_t zero0;        \
+    zero0 = svdup_f64(0.); 
+
+#define Chimu_00 Chi_00  
+#define Chimu_01 Chi_01  
+#define Chimu_02 Chi_02  
+#define Chimu_10 Chi_10  
+#define Chimu_11 Chi_11  
+#define Chimu_12 Chi_12  
+#define Chimu_20 UChi_00  
+#define Chimu_21 UChi_01  
+#define Chimu_22 UChi_02  
+#define Chimu_30 UChi_10  
+#define Chimu_31 UChi_11  
+#define Chimu_32 UChi_12  
+// RESULT
+#define RESULT_A64FXd(base)  \
+{ \
+    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(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(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)); 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)); 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(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(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(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(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  \
+    table0 = svld1(pg1, (uint64_t*)&lut[0]);  
+
+// LOAD_TABLE1
+#define LOAD_TABLE1  \
+    table0 = svld1(pg1, (uint64_t*)&lut[1]);  
+
+// LOAD_TABLE2
+#define LOAD_TABLE2  \
+    table0 = svld1(pg1, (uint64_t*)&lut[2]);  
+
+// LOAD_TABLE3
+#define LOAD_TABLE3  \
+    table0 = svld1(pg1, (uint64_t*)&lut[3]);  
+
+// PERMUTE
+#define PERMUTE_A64FXd  \
+    Chi_00 = svtbl(Chi_00, table0);    \
+    Chi_01 = svtbl(Chi_01, table0);    \
+    Chi_02 = svtbl(Chi_02, table0);    \
+    Chi_10 = svtbl(Chi_10, table0);    \
+    Chi_11 = svtbl(Chi_11, table0);    \
+    Chi_12 = svtbl(Chi_12, table0);    
+
+// LOAD_GAUGE
+#define LOAD_GAUGE  \
+    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));  \
+}
+// MULT_2SPIN
+#define MULT_2SPIN_1_A64FXd(A)  \
+{ \
+    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
+    U_00 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -6 * 64));  \
+    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); \
+    UChi_11 = svcmla_x(pg1, zero0, U_10, Chi_10, 0); \
+    UChi_02 = svcmla_x(pg1, zero0, U_20, Chi_00, 0); \
+    UChi_12 = svcmla_x(pg1, zero0, U_20, Chi_10, 0); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_00, Chi_00, 90); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_00, Chi_10, 90); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_10, Chi_00, 90); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_10, 90); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_00, 90); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_10, 90); \
+    U_00 = svld1(pg1, (float64_t*)(baseU + 2 * 3 * 64 + -4 * 64));  \
+    U_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  \
+{ \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_01, Chi_01, 0); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_01, Chi_11, 0); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_11, Chi_01, 0); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_11, Chi_11, 0); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_21, Chi_01, 0); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_21, Chi_11, 0); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_01, Chi_01, 90); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_01, Chi_11, 90); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_11, Chi_01, 90); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_11, Chi_11, 90); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_21, Chi_01, 90); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_21, Chi_11, 90); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_00, Chi_02, 0); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_00, Chi_12, 0); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_10, Chi_02, 0); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_12, 0); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_02, 0); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_12, 0); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_00, Chi_02, 90); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_00, Chi_12, 90); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_10, Chi_02, 90); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_12, 90); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_02, 90); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_12, 90); \
+}
+// XP_PROJ
+#define XP_PROJ_A64FXd  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_30, 90);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_31, 90);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_32, 90);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_20, 90);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_21, 90);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_22, 90);   \
+}
+// XP_RECON
+#define XP_RECON_A64FXd  \
+    result_20 = svcadd_x(pg1, zero0, UChi_10, 270);   \
+    result_21 = svcadd_x(pg1, zero0, UChi_11, 270);   \
+    result_22 = svcadd_x(pg1, zero0, UChi_12, 270);   \
+    result_30 = svcadd_x(pg1, zero0, UChi_00, 270);   \
+    result_31 = svcadd_x(pg1, zero0, UChi_01, 270);   \
+    result_32 = svcadd_x(pg1, zero0, UChi_02, 270);   \
+    result_00 = UChi_00;        \
+    result_01 = UChi_01;        \
+    result_02 = UChi_02;        \
+    result_10 = UChi_10;        \
+    result_11 = UChi_11;        \
+    result_12 = UChi_12;        
+
+// XP_RECON_ACCUM
+#define XP_RECON_ACCUM_A64FXd  \
+    result_30 = svcadd_x(pg1, result_30, UChi_00, 270);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_31 = svcadd_x(pg1, result_31, UChi_01, 270);   \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_32 = svcadd_x(pg1, result_32, UChi_02, 270);   \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_20 = svcadd_x(pg1, result_20, UChi_10, 270);   \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_21 = svcadd_x(pg1, result_21, UChi_11, 270);   \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_22 = svcadd_x(pg1, result_22, UChi_12, 270);   \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// YP_PROJ
+#define YP_PROJ_A64FXd  \
+{ \
+    Chi_00 = svsub_x(pg1, Chimu_00, Chimu_30);  \
+    Chi_01 = svsub_x(pg1, Chimu_01, Chimu_31);  \
+    Chi_02 = svsub_x(pg1, Chimu_02, Chimu_32);  \
+    Chi_10 = svadd_x(pg1, Chimu_10, Chimu_20);  \
+    Chi_11 = svadd_x(pg1, Chimu_11, Chimu_21);  \
+    Chi_12 = svadd_x(pg1, Chimu_12, Chimu_22);  \
+}
+// ZP_PROJ
+#define ZP_PROJ_A64FXd  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_20, 90);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_21, 90);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_22, 90);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_30, 270);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_31, 270);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_32, 270);   \
+}
+// TP_PROJ
+#define TP_PROJ_A64FXd  \
+{ \
+    Chi_00 = svadd_x(pg1, Chimu_00, Chimu_20);  \
+    Chi_01 = svadd_x(pg1, Chimu_01, Chimu_21);  \
+    Chi_02 = svadd_x(pg1, Chimu_02, Chimu_22);  \
+    Chi_10 = svadd_x(pg1, Chimu_10, Chimu_30);  \
+    Chi_11 = svadd_x(pg1, Chimu_11, Chimu_31);  \
+    Chi_12 = svadd_x(pg1, Chimu_12, Chimu_32);  \
+}
+// XM_PROJ
+#define XM_PROJ_A64FXd  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_30, 270);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_31, 270);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_32, 270);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_20, 270);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_21, 270);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_22, 270);   \
+}
+// XM_RECON
+#define XM_RECON_A64FXd  \
+    result_20 = svcadd_x(pg1, zero0, UChi_10, 90);   \
+    result_21 = svcadd_x(pg1, zero0, UChi_11, 90);   \
+    result_22 = svcadd_x(pg1, zero0, UChi_12, 90);   \
+    result_30 = svcadd_x(pg1, zero0, UChi_00, 90);   \
+    result_31 = svcadd_x(pg1, zero0, UChi_01, 90);   \
+    result_32 = svcadd_x(pg1, zero0, UChi_02, 90);   \
+    result_00 = UChi_00;        \
+    result_01 = UChi_01;        \
+    result_02 = UChi_02;        \
+    result_10 = UChi_10;        \
+    result_11 = UChi_11;        \
+    result_12 = UChi_12;        
+
+// YM_PROJ
+#define YM_PROJ_A64FXd  \
+{ \
+    Chi_00 = svadd_x(pg1, Chimu_00, Chimu_30);  \
+    Chi_01 = svadd_x(pg1, Chimu_01, Chimu_31);  \
+    Chi_02 = svadd_x(pg1, Chimu_02, Chimu_32);  \
+    Chi_10 = svsub_x(pg1, Chimu_10, Chimu_20);  \
+    Chi_11 = svsub_x(pg1, Chimu_11, Chimu_21);  \
+    Chi_12 = svsub_x(pg1, Chimu_12, Chimu_22);  \
+}
+// ZM_PROJ
+#define ZM_PROJ_A64FXd  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_20, 270);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_21, 270);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_22, 270);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_30, 90);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_31, 90);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_32, 90);   \
+}
+// TM_PROJ
+#define TM_PROJ_A64FXd  \
+{ \
+    Chi_00 = svsub_x(pg1, Chimu_00, Chimu_20);  \
+    Chi_01 = svsub_x(pg1, Chimu_01, Chimu_21);  \
+    Chi_02 = svsub_x(pg1, Chimu_02, Chimu_22);  \
+    Chi_10 = svsub_x(pg1, Chimu_10, Chimu_30);  \
+    Chi_11 = svsub_x(pg1, Chimu_11, Chimu_31);  \
+    Chi_12 = svsub_x(pg1, Chimu_12, Chimu_32);  \
+}
+// XM_RECON_ACCUM
+#define XM_RECON_ACCUM_A64FXd  \
+    result_30 = svcadd_x(pg1, result_30, UChi_00, 90);   \
+    result_31 = svcadd_x(pg1, result_31, UChi_01, 90);   \
+    result_32 = svcadd_x(pg1, result_32, UChi_02, 90);   \
+    result_20 = svcadd_x(pg1, result_20, UChi_10, 90);   \
+    result_21 = svcadd_x(pg1, result_21, UChi_11, 90);   \
+    result_22 = svcadd_x(pg1, result_22, UChi_12, 90);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// YP_RECON_ACCUM
+#define YP_RECON_ACCUM_A64FXd  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_30 = svsub_x(pg1, result_30, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_31 = svsub_x(pg1, result_31, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_32 = svsub_x(pg1, result_32, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_20 = svadd_x(pg1, result_20, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_21 = svadd_x(pg1, result_21, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_22 = svadd_x(pg1, result_22, UChi_12); 
+
+// YM_RECON_ACCUM
+#define YM_RECON_ACCUM_A64FXd  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_30 = svadd_x(pg1, result_30, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_31 = svadd_x(pg1, result_31, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_32 = svadd_x(pg1, result_32, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_20 = svsub_x(pg1, result_20, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_21 = svsub_x(pg1, result_21, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_22 = svsub_x(pg1, result_22, UChi_12); 
+
+// ZP_RECON_ACCUM
+#define ZP_RECON_ACCUM_A64FXd  \
+    result_20 = svcadd_x(pg1, result_20, UChi_00, 270);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_21 = svcadd_x(pg1, result_21, UChi_01, 270);   \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_22 = svcadd_x(pg1, result_22, UChi_02, 270);   \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_30 = svcadd_x(pg1, result_30, UChi_10, 90);   \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_31 = svcadd_x(pg1, result_31, UChi_11, 90);   \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_32 = svcadd_x(pg1, result_32, UChi_12, 90);   \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// ZM_RECON_ACCUM
+#define ZM_RECON_ACCUM_A64FXd  \
+    result_20 = svcadd_x(pg1, result_20, UChi_00, 90);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_21 = svcadd_x(pg1, result_21, UChi_01, 90);   \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_22 = svcadd_x(pg1, result_22, UChi_02, 90);   \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_30 = svcadd_x(pg1, result_30, UChi_10, 270);   \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_31 = svcadd_x(pg1, result_31, UChi_11, 270);   \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_32 = svcadd_x(pg1, result_32, UChi_12, 270);   \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// TP_RECON_ACCUM
+#define TP_RECON_ACCUM_A64FXd  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_20 = svadd_x(pg1, result_20, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_21 = svadd_x(pg1, result_21, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_22 = svadd_x(pg1, result_22, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_30 = svadd_x(pg1, result_30, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_31 = svadd_x(pg1, result_31, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_32 = svadd_x(pg1, result_32, UChi_12); 
+
+// TM_RECON_ACCUM
+#define TM_RECON_ACCUM_A64FXd  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_20 = svsub_x(pg1, result_20, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_21 = svsub_x(pg1, result_21, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_22 = svsub_x(pg1, result_22, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_30 = svsub_x(pg1, result_30, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_31 = svsub_x(pg1, result_31, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_32 = svsub_x(pg1, result_32, UChi_12); 
+
+// ZERO_PSI
+#define ZERO_PSI_A64FXd  \
+    result_00 = svdup_f64(0.); \
+    result_01 = svdup_f64(0.); \
+    result_02 = svdup_f64(0.); \
+    result_10 = svdup_f64(0.); \
+    result_11 = svdup_f64(0.); \
+    result_12 = svdup_f64(0.); \
+    result_20 = svdup_f64(0.); \
+    result_21 = svdup_f64(0.); \
+    result_22 = svdup_f64(0.); \
+    result_30 = svdup_f64(0.); \
+    result_31 = svdup_f64(0.); \
+    result_32 = svdup_f64(0.); 
+
+// PREFETCH_RESULT_L2_STORE (prefetch store to L2)
+#define PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXd(base)  \
+{ \
+    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)  \
+{ \
+    svprfd(pg1, (int64_t*)(base + 0), SV_PSTL1STRM); \
+    svprfd(pg1, (int64_t*)(base + 256), SV_PSTL1STRM); \
+    svprfd(pg1, (int64_t*)(base + 512), SV_PSTL1STRM); \
+}
+// ADD_RESULT_INTERNAL
+#define ADD_RESULT_INTERNAL_A64FXd  \
+    result_00 = svadd_x(pg1, result_00, Chimu_00); \
+    result_01 = svadd_x(pg1, result_01, Chimu_01); \
+    result_02 = svadd_x(pg1, result_02, Chimu_02); \
+    result_10 = svadd_x(pg1, result_10, Chimu_10); \
+    result_11 = svadd_x(pg1, result_11, Chimu_11); \
+    result_12 = svadd_x(pg1, result_12, Chimu_12); \
+    result_20 = svadd_x(pg1, result_20, Chimu_20); \
+    result_21 = svadd_x(pg1, result_21, Chimu_21); \
+    result_22 = svadd_x(pg1, result_22, Chimu_22); \
+    result_30 = svadd_x(pg1, result_30, Chimu_30); \
+    result_31 = svadd_x(pg1, result_31, Chimu_31); \
+    result_32 = svadd_x(pg1, result_32, Chimu_32); 
+
diff --git a/Grid/simd/Fujitsu_A64FX_intrin_single.h b/Grid/simd/Fujitsu_A64FX_intrin_single.h
new file mode 100644
index 00000000..180e5f4f
--- /dev/null
+++ b/Grid/simd/Fujitsu_A64FX_intrin_single.h
@@ -0,0 +1,601 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Fujitsu_A64FX_intrin_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} }; \
+    svfloat32_t result_00;        \
+    svfloat32_t result_01;        \
+    svfloat32_t result_02;        \
+    svfloat32_t result_10;        \
+    svfloat32_t result_11;        \
+    svfloat32_t result_12;        \
+    svfloat32_t result_20;        \
+    svfloat32_t result_21;        \
+    svfloat32_t result_22;        \
+    svfloat32_t result_30;        \
+    svfloat32_t result_31;        \
+    svfloat32_t result_32;        \
+    svfloat32_t Chi_00;        \
+    svfloat32_t Chi_01;        \
+    svfloat32_t Chi_02;        \
+    svfloat32_t Chi_10;        \
+    svfloat32_t Chi_11;        \
+    svfloat32_t Chi_12;        \
+    svfloat32_t UChi_00;        \
+    svfloat32_t UChi_01;        \
+    svfloat32_t UChi_02;        \
+    svfloat32_t UChi_10;        \
+    svfloat32_t UChi_11;        \
+    svfloat32_t UChi_12;        \
+    svfloat32_t U_00;        \
+    svfloat32_t U_10;        \
+    svfloat32_t U_20;        \
+    svfloat32_t U_01;        \
+    svfloat32_t U_11;        \
+    svfloat32_t U_21;        \
+    svbool_t pg1;        \
+    pg1 = svptrue_b32();        \
+    svuint32_t table0; \
+    svfloat32_t zero0;        \
+    zero0 = svdup_f32(0.); 
+
+#define Chimu_00 Chi_00  
+#define Chimu_01 Chi_01  
+#define Chimu_02 Chi_02  
+#define Chimu_10 Chi_10  
+#define Chimu_11 Chi_11  
+#define Chimu_12 Chi_12  
+#define Chimu_20 UChi_00  
+#define Chimu_21 UChi_01  
+#define Chimu_22 UChi_02  
+#define Chimu_30 UChi_10  
+#define Chimu_31 UChi_11  
+#define Chimu_32 UChi_12  
+// RESULT
+#define RESULT_A64FXf(base)  \
+{ \
+    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(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(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)); 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)); 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(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(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(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(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  \
+    table0 = svld1(pg1, (uint32_t*)&lut[0]);  
+
+// LOAD_TABLE1
+#define LOAD_TABLE1  \
+    table0 = svld1(pg1, (uint32_t*)&lut[1]);  
+
+// LOAD_TABLE2
+#define LOAD_TABLE2  \
+    table0 = svld1(pg1, (uint32_t*)&lut[2]);  
+
+// LOAD_TABLE3
+#define LOAD_TABLE3  \
+    table0 = svld1(pg1, (uint32_t*)&lut[3]);  
+
+// PERMUTE
+#define PERMUTE_A64FXf  \
+    Chi_00 = svtbl(Chi_00, table0);    \
+    Chi_01 = svtbl(Chi_01, table0);    \
+    Chi_02 = svtbl(Chi_02, table0);    \
+    Chi_10 = svtbl(Chi_10, table0);    \
+    Chi_11 = svtbl(Chi_11, table0);    \
+    Chi_12 = svtbl(Chi_12, table0);    
+
+// LOAD_GAUGE
+#define LOAD_GAUGE  \
+    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));  \
+}
+// MULT_2SPIN
+#define MULT_2SPIN_1_A64FXf(A)  \
+{ \
+    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \
+    U_00 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -6 * 64));  \
+    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); \
+    UChi_11 = svcmla_x(pg1, zero0, U_10, Chi_10, 0); \
+    UChi_02 = svcmla_x(pg1, zero0, U_20, Chi_00, 0); \
+    UChi_12 = svcmla_x(pg1, zero0, U_20, Chi_10, 0); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_00, Chi_00, 90); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_00, Chi_10, 90); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_10, Chi_00, 90); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_10, 90); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_00, 90); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_10, 90); \
+    U_00 = svld1(pg1, (float32_t*)(baseU + 2 * 3 * 64 + -4 * 64));  \
+    U_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  \
+{ \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_01, Chi_01, 0); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_01, Chi_11, 0); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_11, Chi_01, 0); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_11, Chi_11, 0); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_21, Chi_01, 0); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_21, Chi_11, 0); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_01, Chi_01, 90); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_01, Chi_11, 90); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_11, Chi_01, 90); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_11, Chi_11, 90); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_21, Chi_01, 90); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_21, Chi_11, 90); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_00, Chi_02, 0); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_00, Chi_12, 0); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_10, Chi_02, 0); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_12, 0); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_02, 0); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_12, 0); \
+    UChi_00 = svcmla_x(pg1, UChi_00, U_00, Chi_02, 90); \
+    UChi_10 = svcmla_x(pg1, UChi_10, U_00, Chi_12, 90); \
+    UChi_01 = svcmla_x(pg1, UChi_01, U_10, Chi_02, 90); \
+    UChi_11 = svcmla_x(pg1, UChi_11, U_10, Chi_12, 90); \
+    UChi_02 = svcmla_x(pg1, UChi_02, U_20, Chi_02, 90); \
+    UChi_12 = svcmla_x(pg1, UChi_12, U_20, Chi_12, 90); \
+}
+// XP_PROJ
+#define XP_PROJ_A64FXf  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_30, 90);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_31, 90);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_32, 90);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_20, 90);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_21, 90);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_22, 90);   \
+}
+// XP_RECON
+#define XP_RECON_A64FXf  \
+    result_20 = svcadd_x(pg1, zero0, UChi_10, 270);   \
+    result_21 = svcadd_x(pg1, zero0, UChi_11, 270);   \
+    result_22 = svcadd_x(pg1, zero0, UChi_12, 270);   \
+    result_30 = svcadd_x(pg1, zero0, UChi_00, 270);   \
+    result_31 = svcadd_x(pg1, zero0, UChi_01, 270);   \
+    result_32 = svcadd_x(pg1, zero0, UChi_02, 270);   \
+    result_00 = UChi_00;        \
+    result_01 = UChi_01;        \
+    result_02 = UChi_02;        \
+    result_10 = UChi_10;        \
+    result_11 = UChi_11;        \
+    result_12 = UChi_12;        
+
+// XP_RECON_ACCUM
+#define XP_RECON_ACCUM_A64FXf  \
+    result_30 = svcadd_x(pg1, result_30, UChi_00, 270);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_31 = svcadd_x(pg1, result_31, UChi_01, 270);   \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_32 = svcadd_x(pg1, result_32, UChi_02, 270);   \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_20 = svcadd_x(pg1, result_20, UChi_10, 270);   \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_21 = svcadd_x(pg1, result_21, UChi_11, 270);   \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_22 = svcadd_x(pg1, result_22, UChi_12, 270);   \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// YP_PROJ
+#define YP_PROJ_A64FXf  \
+{ \
+    Chi_00 = svsub_x(pg1, Chimu_00, Chimu_30);  \
+    Chi_01 = svsub_x(pg1, Chimu_01, Chimu_31);  \
+    Chi_02 = svsub_x(pg1, Chimu_02, Chimu_32);  \
+    Chi_10 = svadd_x(pg1, Chimu_10, Chimu_20);  \
+    Chi_11 = svadd_x(pg1, Chimu_11, Chimu_21);  \
+    Chi_12 = svadd_x(pg1, Chimu_12, Chimu_22);  \
+}
+// ZP_PROJ
+#define ZP_PROJ_A64FXf  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_20, 90);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_21, 90);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_22, 90);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_30, 270);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_31, 270);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_32, 270);   \
+}
+// TP_PROJ
+#define TP_PROJ_A64FXf  \
+{ \
+    Chi_00 = svadd_x(pg1, Chimu_00, Chimu_20);  \
+    Chi_01 = svadd_x(pg1, Chimu_01, Chimu_21);  \
+    Chi_02 = svadd_x(pg1, Chimu_02, Chimu_22);  \
+    Chi_10 = svadd_x(pg1, Chimu_10, Chimu_30);  \
+    Chi_11 = svadd_x(pg1, Chimu_11, Chimu_31);  \
+    Chi_12 = svadd_x(pg1, Chimu_12, Chimu_32);  \
+}
+// XM_PROJ
+#define XM_PROJ_A64FXf  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_30, 270);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_31, 270);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_32, 270);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_20, 270);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_21, 270);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_22, 270);   \
+}
+// XM_RECON
+#define XM_RECON_A64FXf  \
+    result_20 = svcadd_x(pg1, zero0, UChi_10, 90);   \
+    result_21 = svcadd_x(pg1, zero0, UChi_11, 90);   \
+    result_22 = svcadd_x(pg1, zero0, UChi_12, 90);   \
+    result_30 = svcadd_x(pg1, zero0, UChi_00, 90);   \
+    result_31 = svcadd_x(pg1, zero0, UChi_01, 90);   \
+    result_32 = svcadd_x(pg1, zero0, UChi_02, 90);   \
+    result_00 = UChi_00;        \
+    result_01 = UChi_01;        \
+    result_02 = UChi_02;        \
+    result_10 = UChi_10;        \
+    result_11 = UChi_11;        \
+    result_12 = UChi_12;        
+
+// YM_PROJ
+#define YM_PROJ_A64FXf  \
+{ \
+    Chi_00 = svadd_x(pg1, Chimu_00, Chimu_30);  \
+    Chi_01 = svadd_x(pg1, Chimu_01, Chimu_31);  \
+    Chi_02 = svadd_x(pg1, Chimu_02, Chimu_32);  \
+    Chi_10 = svsub_x(pg1, Chimu_10, Chimu_20);  \
+    Chi_11 = svsub_x(pg1, Chimu_11, Chimu_21);  \
+    Chi_12 = svsub_x(pg1, Chimu_12, Chimu_22);  \
+}
+// ZM_PROJ
+#define ZM_PROJ_A64FXf  \
+{ \
+    Chi_00 = svcadd_x(pg1, Chimu_00, Chimu_20, 270);   \
+    Chi_01 = svcadd_x(pg1, Chimu_01, Chimu_21, 270);   \
+    Chi_02 = svcadd_x(pg1, Chimu_02, Chimu_22, 270);   \
+    Chi_10 = svcadd_x(pg1, Chimu_10, Chimu_30, 90);   \
+    Chi_11 = svcadd_x(pg1, Chimu_11, Chimu_31, 90);   \
+    Chi_12 = svcadd_x(pg1, Chimu_12, Chimu_32, 90);   \
+}
+// TM_PROJ
+#define TM_PROJ_A64FXf  \
+{ \
+    Chi_00 = svsub_x(pg1, Chimu_00, Chimu_20);  \
+    Chi_01 = svsub_x(pg1, Chimu_01, Chimu_21);  \
+    Chi_02 = svsub_x(pg1, Chimu_02, Chimu_22);  \
+    Chi_10 = svsub_x(pg1, Chimu_10, Chimu_30);  \
+    Chi_11 = svsub_x(pg1, Chimu_11, Chimu_31);  \
+    Chi_12 = svsub_x(pg1, Chimu_12, Chimu_32);  \
+}
+// XM_RECON_ACCUM
+#define XM_RECON_ACCUM_A64FXf  \
+    result_30 = svcadd_x(pg1, result_30, UChi_00, 90);   \
+    result_31 = svcadd_x(pg1, result_31, UChi_01, 90);   \
+    result_32 = svcadd_x(pg1, result_32, UChi_02, 90);   \
+    result_20 = svcadd_x(pg1, result_20, UChi_10, 90);   \
+    result_21 = svcadd_x(pg1, result_21, UChi_11, 90);   \
+    result_22 = svcadd_x(pg1, result_22, UChi_12, 90);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// YP_RECON_ACCUM
+#define YP_RECON_ACCUM_A64FXf  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_30 = svsub_x(pg1, result_30, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_31 = svsub_x(pg1, result_31, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_32 = svsub_x(pg1, result_32, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_20 = svadd_x(pg1, result_20, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_21 = svadd_x(pg1, result_21, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_22 = svadd_x(pg1, result_22, UChi_12); 
+
+// YM_RECON_ACCUM
+#define YM_RECON_ACCUM_A64FXf  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_30 = svadd_x(pg1, result_30, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_31 = svadd_x(pg1, result_31, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_32 = svadd_x(pg1, result_32, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_20 = svsub_x(pg1, result_20, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_21 = svsub_x(pg1, result_21, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_22 = svsub_x(pg1, result_22, UChi_12); 
+
+// ZP_RECON_ACCUM
+#define ZP_RECON_ACCUM_A64FXf  \
+    result_20 = svcadd_x(pg1, result_20, UChi_00, 270);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_21 = svcadd_x(pg1, result_21, UChi_01, 270);   \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_22 = svcadd_x(pg1, result_22, UChi_02, 270);   \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_30 = svcadd_x(pg1, result_30, UChi_10, 90);   \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_31 = svcadd_x(pg1, result_31, UChi_11, 90);   \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_32 = svcadd_x(pg1, result_32, UChi_12, 90);   \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// ZM_RECON_ACCUM
+#define ZM_RECON_ACCUM_A64FXf  \
+    result_20 = svcadd_x(pg1, result_20, UChi_00, 90);   \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_21 = svcadd_x(pg1, result_21, UChi_01, 90);   \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_22 = svcadd_x(pg1, result_22, UChi_02, 90);   \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_30 = svcadd_x(pg1, result_30, UChi_10, 270);   \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_31 = svcadd_x(pg1, result_31, UChi_11, 270);   \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_32 = svcadd_x(pg1, result_32, UChi_12, 270);   \
+    result_12 = svadd_x(pg1, result_12, UChi_12); 
+
+// TP_RECON_ACCUM
+#define TP_RECON_ACCUM_A64FXf  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_20 = svadd_x(pg1, result_20, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_21 = svadd_x(pg1, result_21, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_22 = svadd_x(pg1, result_22, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_30 = svadd_x(pg1, result_30, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_31 = svadd_x(pg1, result_31, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_32 = svadd_x(pg1, result_32, UChi_12); 
+
+// TM_RECON_ACCUM
+#define TM_RECON_ACCUM_A64FXf  \
+    result_00 = svadd_x(pg1, result_00, UChi_00); \
+    result_20 = svsub_x(pg1, result_20, UChi_00); \
+    result_01 = svadd_x(pg1, result_01, UChi_01); \
+    result_21 = svsub_x(pg1, result_21, UChi_01); \
+    result_02 = svadd_x(pg1, result_02, UChi_02); \
+    result_22 = svsub_x(pg1, result_22, UChi_02); \
+    result_10 = svadd_x(pg1, result_10, UChi_10); \
+    result_30 = svsub_x(pg1, result_30, UChi_10); \
+    result_11 = svadd_x(pg1, result_11, UChi_11); \
+    result_31 = svsub_x(pg1, result_31, UChi_11); \
+    result_12 = svadd_x(pg1, result_12, UChi_12); \
+    result_32 = svsub_x(pg1, result_32, UChi_12); 
+
+// ZERO_PSI
+#define ZERO_PSI_A64FXf  \
+    result_00 = svdup_f32(0.); \
+    result_01 = svdup_f32(0.); \
+    result_02 = svdup_f32(0.); \
+    result_10 = svdup_f32(0.); \
+    result_11 = svdup_f32(0.); \
+    result_12 = svdup_f32(0.); \
+    result_20 = svdup_f32(0.); \
+    result_21 = svdup_f32(0.); \
+    result_22 = svdup_f32(0.); \
+    result_30 = svdup_f32(0.); \
+    result_31 = svdup_f32(0.); \
+    result_32 = svdup_f32(0.); 
+
+// PREFETCH_RESULT_L2_STORE (prefetch store to L2)
+#define PREFETCH_RESULT_L2_STORE_INTERNAL_A64FXf(base)  \
+{ \
+    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)  \
+{ \
+    svprfd(pg1, (int64_t*)(base + 0), SV_PSTL1STRM); \
+    svprfd(pg1, (int64_t*)(base + 256), SV_PSTL1STRM); \
+    svprfd(pg1, (int64_t*)(base + 512), SV_PSTL1STRM); \
+}
+// ADD_RESULT_INTERNAL
+#define ADD_RESULT_INTERNAL_A64FXf  \
+    result_00 = svadd_x(pg1, result_00, Chimu_00); \
+    result_01 = svadd_x(pg1, result_01, Chimu_01); \
+    result_02 = svadd_x(pg1, result_02, Chimu_02); \
+    result_10 = svadd_x(pg1, result_10, Chimu_10); \
+    result_11 = svadd_x(pg1, result_11, Chimu_11); \
+    result_12 = svadd_x(pg1, result_12, Chimu_12); \
+    result_20 = svadd_x(pg1, result_20, Chimu_20); \
+    result_21 = svadd_x(pg1, result_21, Chimu_21); \
+    result_22 = svadd_x(pg1, result_22, Chimu_22); \
+    result_30 = svadd_x(pg1, result_30, Chimu_30); \
+    result_31 = svadd_x(pg1, result_31, Chimu_31); \
+    result_32 = svadd_x(pg1, result_32, Chimu_32); 
+
diff --git a/Grid/simd/Fujitsu_A64FX_undef.h b/Grid/simd/Fujitsu_A64FX_undef.h
new file mode 100644
index 00000000..81eec37a
--- /dev/null
+++ b/Grid/simd/Fujitsu_A64FX_undef.h
@@ -0,0 +1,76 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Fujitsu_A64FX_undef.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 */
+
+#undef LOAD_CHIMU
+#undef PREFETCH_CHIMU_L1
+#undef PREFETCH_GAUGE_L1
+#undef PREFETCH_CHIMU_L2
+#undef PREFETCH_GAUGE_L2
+#undef PREFETCH_GAUGE_L1_INTERNAL
+#undef PREFETCH1_CHIMU
+#undef PREFETCH_CHIMU
+#undef PREFETCH_RESULT_L2_STORE
+#undef PREFETCH_RESULT_L1_STORE
+#undef LOAD_GAUGE
+#undef LOCK_GAUGE
+#undef UNLOCK_GAUGE
+#undef MASK_REGS
+#undef SAVE_RESULT
+#undef ADD_RESULT
+#undef MULT_2SPIN_1
+#undef MULT_2SPIN_2
+#undef MAYBEPERM
+#undef LOAD_CHI
+#undef XP_PROJ
+#undef YP_PROJ
+#undef ZP_PROJ
+#undef TP_PROJ
+#undef XM_PROJ
+#undef YM_PROJ
+#undef ZM_PROJ
+#undef TM_PROJ
+#undef XP_RECON
+#undef XM_RECON
+#undef XM_RECON_ACCUM
+#undef YM_RECON_ACCUM
+#undef ZM_RECON_ACCUM
+#undef TM_RECON_ACCUM
+#undef XP_RECON_ACCUM
+#undef YP_RECON_ACCUM
+#undef ZP_RECON_ACCUM
+#undef TP_RECON_ACCUM
+#undef PERMUTE
+#undef PERMUTE_DIR0
+#undef PERMUTE_DIR1
+#undef PERMUTE_DIR2
+#undef PERMUTE_DIR3
+#undef LOAD_TABLE
+#undef LOAD_TABLE0
+#undef LOAD_TABLE1
+#undef LOAD_TABLE2
+#undef LOAD_TABLE3
diff --git a/Grid/simd/Grid_a64fx-2.h b/Grid/simd/Grid_a64fx-2.h
new file mode 100644
index 00000000..2ad8591c
--- /dev/null
+++ b/Grid/simd/Grid_a64fx-2.h
@@ -0,0 +1,942 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Grid_a64fx-2.h
+
+    Copyright (C) 2020
+
+    Author: Nils Meyer          <nils.meyer@ur.de>
+
+    with support from Arm
+
+    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 */
+
+/////////////////////////////////////////////////////
+// Using SVE ACLE
+/////////////////////////////////////////////////////
+
+static_assert(GEN_SIMD_WIDTH % 64u == 0, "A64FX SIMD vector size is 64 bytes");
+
+NAMESPACE_BEGIN(Grid);
+NAMESPACE_BEGIN(Optimization);
+
+  // type traits giving the number of elements for each vector type
+  template <typename T> struct W;
+  template <> struct W<double> {
+    constexpr static unsigned int c = GEN_SIMD_WIDTH/16u;
+    constexpr static unsigned int r = GEN_SIMD_WIDTH/8u;
+  };
+  template <> struct W<float> {
+    constexpr static unsigned int c = GEN_SIMD_WIDTH/8u;
+    constexpr static unsigned int r = GEN_SIMD_WIDTH/4u;
+  };
+  template <> struct W<Integer> {
+    constexpr static unsigned int r = GEN_SIMD_WIDTH/4u;
+  };
+  template <> struct W<uint16_t> {
+    constexpr static unsigned int c = GEN_SIMD_WIDTH/4u;
+    constexpr static unsigned int r = GEN_SIMD_WIDTH/2u;
+  };
+  template <> struct W<uint64_t> {
+    constexpr static unsigned int c = GEN_SIMD_WIDTH/16u;
+    constexpr static unsigned int r = GEN_SIMD_WIDTH/8u;
+  };
+
+  #ifdef ARMCLANGCOMPAT
+  // SIMD vector immediate types
+  template <typename T>
+  struct vec_imm {
+    alignas(GEN_SIMD_WIDTH) T v[W<T>::r];
+  };
+
+  // SIMD vector types
+  template <typename T>
+  struct vec {
+    alignas(GEN_SIMD_WIDTH) T v[W<T>::r];
+    vec() = default;
+    vec(const vec &rhs) { this->operator=(rhs); }
+    vec(const vec_imm<T> &rhs) {
+      // v = rhs.v
+      svst1(svptrue_b8(), (T*)this, svld1(svptrue_b8(), (T*)rhs.v));
+    }
+
+    inline vec &operator=(const vec &rhs) {
+      // v = rhs.v
+      svst1(svptrue_b8(), (T*)this, svld1(svptrue_b8(), (T*)rhs.v));
+      return *this;
+    };
+  };
+
+  #else //  no ARMCLANGCOMPAT
+  #define vec_imm vec
+  // SIMD vector types
+  template <typename T>
+  struct vec {
+    alignas(GEN_SIMD_WIDTH) T v[W<T>::r];
+  };
+  #endif
+
+  typedef vec<float>     vecf;
+  typedef vec<double>    vecd;
+  typedef vec<uint16_t>  vech; // half precision comms
+  typedef vec<Integer>   veci;
+
+NAMESPACE_END(Optimization)
+NAMESPACE_END(Grid)
+
+// low-level API
+NAMESPACE_BEGIN(Grid);
+NAMESPACE_BEGIN(Optimization);
+
+template <typename T>
+struct acle{};
+
+template <>
+struct acle<double>{
+  typedef svfloat64_t vt;
+  typedef svfloat64x2_t vt2;
+  typedef svfloat64x4_t vt4;
+  typedef float64_t pt;
+  typedef uint64_t uint;
+  typedef svuint64_t svuint;
+
+  static inline svbool_t pg1(){return svptrue_b64();}
+  static inline svbool_t pg2(){return svptrue_pat_b64(SV_VL4);}
+  static inline svbool_t pg4(){return svptrue_pat_b64(SV_VL2);}
+  static inline vec<uint64_t> tbl_swap(){
+    //const vec<uint64_t> t = {1, 0, 3, 2, 5, 4, 7, 6};
+    const vec_imm<uint64_t> t = {1, 0, 3, 2, 5, 4, 7, 6};
+    return t;
+  }
+  static inline vec<uint64_t> tbl0(){
+    //const vec<uint64_t> t = {4, 5, 6, 7, 0, 1, 2, 3};
+    const vec_imm<uint64_t> t = {4, 5, 6, 7, 0, 1, 2, 3};
+    return t;
+  }
+  static inline vec<uint64_t> tbl1(){
+    //const vec<uint64_t> t = {2, 3, 0, 1, 6, 7, 4, 5};
+    const vec_imm<uint64_t> t = {2, 3, 0, 1, 6, 7, 4, 5};
+    return t;
+  }
+  static inline vec<uint64_t> tbl_exch1a(){ // Exchange1
+    //const vec<uint64_t> t = {0, 1, 4, 5, 2, 3, 6, 7};
+    const vec_imm<uint64_t> t = {0, 1, 4, 5, 2, 3, 6, 7};
+    return t;
+  }
+  static inline vec<uint64_t> tbl_exch1b(){ // Exchange1
+    //const vec<uint64_t> t = {2, 3, 6, 7, 0, 1, 4, 5};
+    const vec_imm<uint64_t> t = {2, 3, 6, 7, 0, 1, 4, 5};
+    return t;
+  }
+  static inline vec<uint64_t> tbl_exch1c(){ // Exchange1
+    //const vec<uint64_t> t = {4, 5, 0, 1, 6, 7, 2, 3};
+    const vec_imm<uint64_t> t = {4, 5, 0, 1, 6, 7, 2, 3};
+    return t;
+  }
+  static inline svbool_t pg_even(){return svzip1_b64(svptrue_b64(), svpfalse_b());}
+  static inline svbool_t pg_odd() {return svzip1_b64(svpfalse_b(), svptrue_b64());}
+  static inline svfloat64_t zero(){return svdup_f64(0.);}
+};
+
+template <>
+struct acle<float>{
+  typedef svfloat32_t vt;
+  typedef svfloat32x2_t vt2;
+  typedef float32_t pt;
+  typedef uint32_t uint;
+  typedef svuint32_t svuint;
+
+  static inline svbool_t pg1(){return svptrue_b32();}
+  static inline svbool_t pg2(){return svptrue_pat_b32(SV_VL8);}
+  // exchange neighboring elements
+  static inline vec<uint32_t> tbl_swap(){
+    //const vec<uint32_t> t = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
+    const vec_imm<uint32_t> t = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
+    return t;
+  }
+  static inline vec<uint32_t> tbl0(){
+    //const vec<uint32_t> t = {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7};
+    const vec_imm<uint32_t> t = {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7};
+    return t;
+  }
+  static inline vec<uint32_t> tbl1(){
+    //const vec<uint32_t> t = {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11};
+    const vec_imm<uint32_t> t = {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11};
+    return t;
+  }
+  static inline vec<uint32_t> tbl2(){
+    //const vec<uint32_t> t = {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13};
+    const vec_imm<uint32_t> t = {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13};
+    return t;
+  }
+  static inline vec<uint32_t> tbl_exch1a(){ // Exchange1
+    //const vec<uint32_t> t = {0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 };
+    const vec_imm<uint32_t> t = {0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 };
+    return t;
+  }
+  static inline vec<uint32_t> tbl_exch1b(){ // Exchange1
+    //const vec<uint32_t> t = {4, 5, 6, 7, 12, 13, 14, 15, 0, 1, 2, 3, 8, 9, 10, 11 };
+    const vec_imm<uint32_t> t = {4, 5, 6, 7, 12, 13, 14, 15, 0, 1, 2, 3, 8, 9, 10, 11 };
+    return t;
+  }
+  static inline vec<uint32_t> tbl_exch1c(){ // Exchange1
+    //const vec<uint32_t> t = {8, 9, 10, 11, 0, 1, 2, 3, 12, 13, 14, 15, 4, 5, 6, 7};
+    const vec_imm<uint32_t> t = {8, 9, 10, 11, 0, 1, 2, 3, 12, 13, 14, 15, 4, 5, 6, 7};
+    return t;
+  }
+  static inline svbool_t pg_even(){return svzip1_b32(svptrue_b32(), svpfalse_b());}
+  static inline svbool_t pg_odd() {return svzip1_b32(svpfalse_b(), svptrue_b32());}
+  static inline svfloat32_t zero(){return svdup_f32(0.);}
+};
+
+template <>
+struct acle<uint16_t>{
+  typedef svfloat16_t vt;
+  typedef float16_t pt;
+  typedef uint16_t uint;
+  typedef svuint16_t svuint;
+
+  static inline svbool_t pg1(){return svptrue_b16();}
+  static inline svbool_t pg2(){return svptrue_pat_b16(SV_VL16);}
+  static inline svbool_t pg_even(){return svzip1_b16(svptrue_b16(), svpfalse_b());}
+  static inline svbool_t pg_odd() {return svzip1_b16(svpfalse_b(), svptrue_b16());}
+  static inline svfloat16_t zero(){return svdup_f16(0.);}
+};
+
+template <>
+struct acle<Integer>{
+  typedef svuint32_t vt;
+  typedef svuint32x2_t vt2;
+  typedef Integer pt;
+  typedef uint32_t uint;
+  typedef svuint32_t svuint;
+
+  //static inline svbool_t pg1(){return svptrue_b16();}
+  static inline svbool_t pg1(){return svptrue_b32();}
+  static inline svbool_t pg2(){return svptrue_pat_b32(SV_VL8);}
+  static inline svbool_t pg_even(){return svzip1_b32(svptrue_b32(), svpfalse_b());}
+  static inline svbool_t pg_odd() {return svzip1_b32(svpfalse_b(), svptrue_b32());}
+};
+
+// ---------------------------------------------------
+
+struct Vsplat{
+  // Complex float
+  inline vecf operator()(float a, float b){
+    vecf out;
+    svbool_t pg1 = acle<float>::pg1();
+    typename acle<float>::vt a_v = svdup_f32(a);
+    typename acle<float>::vt b_v = svdup_f32(b);
+    typename acle<float>::vt r_v = svzip1(a_v, b_v);
+    svst1(pg1, out.v, r_v);
+    return out;
+  }
+
+  // Real float
+  inline vecf operator()(float a){
+    vecf out;
+    svbool_t pg1 = acle<float>::pg1();
+    typename acle<float>::vt r_v = svdup_f32(a);
+    svst1(pg1, out.v, r_v);
+    return out;
+  }
+
+ // Complex double
+  inline vecd operator()(double a, double b){
+    vecd out;
+    svbool_t pg1 = acle<double>::pg1();
+    typename acle<double>::vt a_v = svdup_f64(a);
+    typename acle<double>::vt b_v = svdup_f64(b);
+    typename acle<double>::vt r_v = svzip1(a_v, b_v);
+    svst1(pg1, out.v, r_v);
+    return out;
+  }
+
+  // Real double
+  inline vecd operator()(double a){
+    vecd out;
+    svbool_t pg1 = acle<double>::pg1();
+    typename acle<double>::vt r_v = svdup_f64(a);
+    svst1(pg1, out.v, r_v);
+    return out;
+  }
+
+  // Integer
+  inline vec<Integer> operator()(Integer a){
+    vec<Integer> out;
+    svbool_t pg1 = acle<Integer>::pg1();
+    // Add check whether Integer is really a uint32_t???
+    typename acle<Integer>::vt r_v = svdup_u32(a);
+    svst1(pg1, out.v, r_v);
+    return out;
+  }
+};
+
+struct Vstore{
+  // Real
+  template <typename T>
+  inline void operator()(vec<T> a, T *D){
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, (typename acle<T>::pt*)&a.v);
+    svst1(pg1, D, a_v);
+  }
+};
+
+struct Vstream{
+  // Real
+  template <typename T>
+  inline void operator()(T * a, vec<T> b){
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    svstnt1(pg1, a, b_v);
+    //svst1(pg1, a, b_v);
+  }
+};
+
+  struct Vset{
+    // Complex
+    template <typename T>
+    inline vec<T> operator()(std::complex<T> *a){
+      vec<T> out;
+      svbool_t pg1 = acle<T>::pg1();
+      typename acle<T>::vt a_v = svld1(pg1, (T*)a);
+      svst1(pg1, out.v, a_v);
+
+      return out;
+    }
+
+    // Real
+    template <typename T>
+    inline vec<T> operator()(T *a){
+      vec<T> out;
+      svbool_t pg1 = acle<T>::pg1();
+      typename acle<T>::vt a_v = svld1(pg1, a);
+      svst1(pg1, out.v, a_v);
+
+      return out;
+    }
+  };
+
+/////////////////////////////////////////////////////
+// Arithmetic operations
+/////////////////////////////////////////////////////
+
+struct Sum{
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt r_v = svadd_x(pg1, a_v, b_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct Sub{
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt r_v = svsub_x(pg1, a_v, b_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct Mult{
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt c_v = svld1(pg1, c.v);
+    typename acle<T>::vt r_v = svmla_x(pg1, c_v, a_v, b_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt r_v = svmul_x(pg1, a_v, b_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct MultRealPart{
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v  = svld1(pg1, a.v);
+    typename acle<T>::vt b_v  = svld1(pg1, b.v);
+
+    // using FCMLA
+    typename acle<T>::vt z_v = acle<T>::zero();
+    typename acle<T>::vt r_v = svcmla_x(pg1, z_v, a_v, b_v, 0);
+
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct MaddRealPart{
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v  = svld1(pg1, a.v);
+    typename acle<T>::vt b_v  = svld1(pg1, b.v);
+    typename acle<T>::vt c_v  = svld1(pg1, c.v);
+
+    // using FCMLA
+    typename acle<T>::vt r_v = svcmla_x(pg1, c_v, a_v, b_v, 0);
+
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct MultComplex{
+  // Complex a*b
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt z_v = acle<T>::zero();
+
+    // using FCMLA
+    typename acle<T>::vt r_v = svcmla_x(pg1, z_v, a_v, b_v, 0);
+    r_v = svcmla_x(pg1, r_v, a_v, b_v, 90);
+
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct MultAddComplex{
+  // Complex a*b+c
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt c_v = svld1(pg1, c.v);;
+
+    // using FCMLA
+    typename acle<T>::vt r_v = svcmla_x(pg1, c_v, a_v, b_v, 0);
+    r_v = svcmla_x(pg1, r_v, a_v, b_v, 90);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct Div{
+  // Real
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt r_v = svdiv_x(pg1, a_v, b_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct Conj{
+  // Complex
+  template <typename T>
+  inline vec<T> operator()(vec<T> a){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    svbool_t pg_odd = acle<T>::pg_odd();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    //typename acle<T>::vt r_v = svneg_x(pg_odd, a_v);
+    typename acle<T>::vt r_v = svneg_m(a_v, pg_odd, a_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct TimesMinusI{
+  // Complex
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    const vec<typename acle<T>::uint> tbl_swap = acle<T>::tbl_swap();
+    svbool_t pg1 = acle<T>::pg1();
+    svbool_t pg_odd = acle<T>::pg_odd();
+
+    typename acle<T>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    a_v = svtbl(a_v, tbl_swap_v);
+    typename acle<T>::vt r_v = svneg_m(a_v, pg_odd, a_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct TimesI{
+  // Complex
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b){
+    vec<T> out;
+    const vec<typename acle<T>::uint> tbl_swap = acle<T>::tbl_swap();
+    svbool_t pg1 = acle<T>::pg1();
+    svbool_t pg_even = acle<T>::pg_even();
+
+    typename acle<T>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    a_v = svtbl(a_v, tbl_swap_v);
+    //typename acle<T>::vt r_v = svneg_x(pg_even, a_v);
+    typename acle<T>::vt r_v = svneg_m(a_v, pg_even, a_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+};
+
+struct PrecisionChange {
+  static inline vech StoH (const vecf &sa,const vecf &sb) {
+    vech ret;
+    svbool_t pg1s = acle<float>::pg1();
+    svbool_t pg1h = acle<uint16_t>::pg1();
+    typename acle<float>::vt sa_v = svld1(pg1s, sa.v);
+    typename acle<float>::vt sb_v = svld1(pg1s, sb.v);
+    typename acle<uint16_t>::vt ha_v = svcvt_f16_x(pg1s, sa_v);
+    typename acle<uint16_t>::vt hb_v = svcvt_f16_x(pg1s, sb_v);
+    typename acle<uint16_t>::vt r_v = svuzp1(ha_v, hb_v);
+    svst1(pg1h, (typename acle<uint16_t>::pt*)&ret.v, r_v);
+
+    return ret;
+  }
+  static inline void HtoS(vech h,vecf &sa,vecf &sb) {
+    svbool_t pg1h = acle<uint16_t>::pg1();
+    svbool_t pg1s = acle<float>::pg1();
+    typename acle<uint16_t>::vt h_v = svld1(pg1h, (typename acle<uint16_t>::pt*)&h.v);
+    typename acle<uint16_t>::vt ha_v = svzip1(h_v, h_v);
+    typename acle<uint16_t>::vt hb_v = svzip2(h_v, h_v);
+    typename acle<float>::vt sa_v = svcvt_f32_x(pg1s, ha_v);
+    typename acle<float>::vt sb_v = svcvt_f32_x(pg1s, hb_v);
+    svst1(pg1s, sa.v, sa_v);
+    svst1(pg1s, sb.v, sb_v);
+  }
+  static inline vecf DtoS (vecd a,vecd b) {
+    vecf ret;
+    svbool_t pg1d = acle<double>::pg1();
+    svbool_t pg1s = acle<float>::pg1();
+    typename acle<double>::vt a_v = svld1(pg1d, a.v);
+    typename acle<double>::vt b_v = svld1(pg1d, b.v);
+    typename acle<float>::vt sa_v = svcvt_f32_x(pg1d, a_v);
+    typename acle<float>::vt sb_v = svcvt_f32_x(pg1d, b_v);
+    typename acle<float>::vt r_v = svuzp1(sa_v, sb_v);
+    svst1(pg1s, ret.v, r_v);
+
+    return ret;
+  }
+  static inline void StoD (vecf s,vecd &a,vecd &b) {
+    svbool_t pg1s = acle<float>::pg1();
+    svbool_t pg1d = acle<double>::pg1();
+    typename acle<float>::vt s_v = svld1(pg1s, s.v);
+    typename acle<float>::vt sa_v = svzip1(s_v, s_v);
+    typename acle<float>::vt sb_v = svzip2(s_v, s_v);
+    typename acle<double>::vt a_v = svcvt_f64_x(pg1d, sa_v);
+    typename acle<double>::vt b_v = svcvt_f64_x(pg1d, sb_v);
+    svst1(pg1d, a.v, a_v);
+    svst1(pg1d, b.v, b_v);
+  }
+  static inline vech DtoH (vecd a,vecd b,vecd c,vecd d) {
+    vech ret;
+    svbool_t pg1d = acle<double>::pg1();
+    svbool_t pg1h = acle<uint16_t>::pg1();
+    typename acle<double>::vt a_v = svld1(pg1d, a.v);
+    typename acle<double>::vt b_v = svld1(pg1d, b.v);
+    typename acle<double>::vt c_v = svld1(pg1d, c.v);
+    typename acle<double>::vt d_v = svld1(pg1d, d.v);
+    typename acle<uint16_t>::vt ha_v = svcvt_f16_x(pg1d, a_v);
+    typename acle<uint16_t>::vt hb_v = svcvt_f16_x(pg1d, b_v);
+    typename acle<uint16_t>::vt hc_v = svcvt_f16_x(pg1d, c_v);
+    typename acle<uint16_t>::vt hd_v = svcvt_f16_x(pg1d, d_v);
+    typename acle<uint16_t>::vt hab_v = svuzp1(ha_v, hb_v);
+    typename acle<uint16_t>::vt hcd_v = svuzp1(hc_v, hd_v);
+    typename acle<uint16_t>::vt r_v = svuzp1(hab_v, hcd_v);
+    svst1(pg1h, (typename acle<uint16_t>::pt*)&ret.v, r_v);
+
+    return ret;
+/*
+    vecf sa,sb;
+    sa = DtoS(a,b);
+    sb = DtoS(c,d);
+    return StoH(sa,sb);
+*/
+  }
+  static inline void HtoD(vech h,vecd &a,vecd &b,vecd &c,vecd &d) {
+    svbool_t pg1h = acle<uint16_t>::pg1();
+    svbool_t pg1d = acle<double>::pg1();
+    typename acle<uint16_t>::vt h_v = svld1(pg1h, (typename acle<uint16_t>::pt*)&h.v);
+    typename acle<uint16_t>::vt sa_v = svzip1(h_v, h_v);
+    typename acle<uint16_t>::vt sb_v = svzip2(h_v, h_v);
+    typename acle<uint16_t>::vt da_v = svzip1(sa_v, sa_v);
+    typename acle<uint16_t>::vt db_v = svzip2(sa_v, sa_v);
+    typename acle<uint16_t>::vt dc_v = svzip1(sb_v, sb_v);
+    typename acle<uint16_t>::vt dd_v = svzip2(sb_v, sb_v);
+    typename acle<double>::vt a_v = svcvt_f64_x(pg1d, da_v);
+    typename acle<double>::vt b_v = svcvt_f64_x(pg1d, db_v);
+    typename acle<double>::vt c_v = svcvt_f64_x(pg1d, dc_v);
+    typename acle<double>::vt d_v = svcvt_f64_x(pg1d, dd_v);
+    svst1(pg1d, a.v, a_v);
+    svst1(pg1d, b.v, b_v);
+    svst1(pg1d, c.v, c_v);
+    svst1(pg1d, d.v, d_v);
+/*
+    vecf sa,sb;
+    HtoS(h,sa,sb);
+    StoD(sa,a,b);
+    StoD(sb,c,d);
+*/
+  }
+};
+
+struct Exchange{
+
+  // Exchange0 is valid for arbitrary SVE vector length
+  template <typename T>
+  static inline void Exchange0(vec<T> &out1, vec<T> &out2, const vec<T> &in1, const vec<T> &in2){
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a1_v = svld1(pg1, in1.v);
+    typename acle<T>::vt a2_v = svld1(pg1, in2.v);
+    typename acle<T>::vt r1_v = svext(a1_v, a1_v, (uint64_t)W<T>::c);
+    r1_v = svext(r1_v, a2_v, (uint64_t)W<T>::c);
+    typename acle<T>::vt r2_v = svext(a2_v, a2_v, (uint64_t)W<T>::c);
+    r2_v = svext(a1_v, r2_v, (uint64_t)W<T>::c);
+    svst1(pg1, out1.v, r1_v);
+    svst1(pg1, out2.v, r2_v);
+  }
+
+  template <typename T>
+  static inline void Exchange1(vec<T> &out1, vec<T> &out2, const vec<T> &in1, const vec<T> &in2){
+    // this one is tricky; svtrn2q* from SVE2 fits best, but it is not available in SVE1
+    // alternative: use 4-el structure; expect translation into ldp + stp -> SFI
+    svbool_t pg1 = acle<T>::pg1();
+    const vec<typename acle<T>::uint> tbl_exch1a = acle<T>::tbl_exch1a();
+    const vec<typename acle<T>::uint> tbl_exch1b = acle<T>::tbl_exch1b();
+    const vec<typename acle<T>::uint> tbl_exch1c = acle<T>::tbl_exch1c();
+
+    typename acle<T>::svuint tbl_exch1a_v = svld1(pg1, tbl_exch1a.v);
+    typename acle<T>::svuint tbl_exch1b_v = svld1(pg1, tbl_exch1b.v);
+    typename acle<T>::svuint tbl_exch1c_v = svld1(pg1, tbl_exch1c.v);
+
+    typename acle<T>::vt in1_v  = svld1(pg1, in1.v);
+    typename acle<T>::vt in2_v  = svld1(pg1, in2.v);
+
+    typename acle<T>::vt a1_v   = svtbl(in1_v, tbl_exch1a_v);
+    typename acle<T>::vt a2_v   = svtbl(in2_v, tbl_exch1b_v);
+    typename acle<T>::vt b1_v   = svext(a2_v, a1_v, (uint64_t)(W<T>::r / 2u));
+    typename acle<T>::vt b2_v   = svext(a1_v, a2_v, (uint64_t)(W<T>::r / 2u));
+    typename acle<T>::vt out1_v = svtbl(b1_v, tbl_exch1c_v);
+    typename acle<T>::vt out2_v = svtbl(b2_v, tbl_exch1a_v);
+
+    svst1(pg1, out1.v, out1_v);
+    svst1(pg1, out2.v, out2_v);
+  }
+
+  template <typename T>
+  static inline void Exchange2(vec<T> &out1, vec<T> &out2, const vec<T> &in1, const vec<T> &in2){
+    svbool_t pg1 = acle<double>::pg1();
+    typename acle<double>::vt a1_v = svld1(pg1, (typename acle<double>::pt*)in1.v);
+    typename acle<double>::vt a2_v = svld1(pg1, (typename acle<double>::pt*)in2.v);
+    typename acle<double>::vt r1_v = svtrn1(a1_v, a2_v);
+    typename acle<double>::vt r2_v = svtrn2(a1_v, a2_v);
+    svst1(pg1, (typename acle<double>::pt*)out1.v, r1_v);
+    svst1(pg1, (typename acle<double>::pt*)out2.v, r2_v);
+  }
+
+  static inline void Exchange3(vecf &out1, vecf &out2, const vecf &in1, const vecf &in2){
+    svbool_t pg1 = acle<float>::pg1();
+    typename acle<float>::vt a1_v = svld1(pg1, in1.v);
+    typename acle<float>::vt a2_v = svld1(pg1, in2.v);
+    typename acle<float>::vt r1_v = svtrn1(a1_v, a2_v);
+    typename acle<float>::vt r2_v = svtrn2(a1_v, a2_v);
+    svst1(pg1, out1.v, r1_v);
+    svst1(pg1, out2.v, r2_v);
+  }
+
+  static inline void Exchange3(vecd &out1, vecd &out2, const vecd &in1, const vecd &in2){
+    assert(0);
+    return;
+  }
+};
+
+struct Permute{
+
+  // Permute0 is valid for any SVE vector width
+  template <typename T>
+  static inline vec<T> Permute0(vec<T> in) {
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, in.v);
+    typename acle<T>::vt r_v = svext(a_v, a_v, (uint64_t)(W<T>::r / 2u));
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  static inline vecd Permute1(vecd in) {
+    vecd out;
+    const vec<typename acle<double>::uint> tbl_swap = acle<double>::tbl1();
+    svbool_t pg1 = acle<double>::pg1();
+    typename acle<double>::vt a_v = svld1(pg1, in.v);
+    typename acle<double>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<double>::vt r_v = svtbl(a_v, tbl_swap_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  static inline vecf Permute1(vecf in) {
+    vecf out;
+    const vec<typename acle<float>::uint> tbl_swap = acle<float>::tbl1();
+    svbool_t pg1 = acle<float>::pg1();
+    typename acle<float>::vt a_v = svld1(pg1, in.v);
+    typename acle<float>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<float>::vt r_v = svtbl(a_v, tbl_swap_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  static inline vecd Permute2(vecd in) {
+    vecd out;
+    const vec<typename acle<double>::uint> tbl_swap = acle<double>::tbl_swap();
+    svbool_t pg1 = acle<double>::pg1();
+    typename acle<double>::vt a_v = svld1(pg1, in.v);
+    typename acle<double>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<double>::vt r_v = svtbl(a_v, tbl_swap_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  static inline vecf Permute2(vecf in) {
+    vecf out;
+    const vec<typename acle<float>::uint> tbl_swap = acle<float>::tbl2();
+    svbool_t pg1 = acle<float>::pg1();
+    typename acle<float>::vt a_v = svld1(pg1, in.v);
+    typename acle<float>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<float>::vt r_v = svtbl(a_v, tbl_swap_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  static inline vecf Permute3(vecf in) {
+    vecf out;
+    const vec<typename acle<float>::uint> tbl_swap = acle<float>::tbl_swap();
+    svbool_t pg1 = acle<float>::pg1();
+    typename acle<float>::vt a_v = svld1(pg1, in.v);
+    typename acle<float>::svuint tbl_swap_v = svld1(pg1, tbl_swap.v);
+    typename acle<float>::vt r_v = svtbl(a_v, tbl_swap_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  static inline vecd Permute3(vecd in) {
+    return in;
+  }
+
+};
+
+struct Rotate{
+
+  template <int n, typename T> static inline vec<T> tRotate(vec<T> in){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, in.v);
+    typename acle<T>::vt r_v = svext(a_v, a_v, (uint64_t)(n%W<T>::r));
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
+
+  template <typename T>
+  static inline vec<T> rotate(vec<T> in, int n){
+
+    switch(n){
+    case 0:  return tRotate<0,  T>(in); break;
+    case 1:  return tRotate<1,  T>(in); break;
+    case 2:  return tRotate<2,  T>(in); break;
+    case 3:  return tRotate<3,  T>(in); break;
+    case 4:  return tRotate<4,  T>(in); break;
+    case 5:  return tRotate<5,  T>(in); break;
+    case 6:  return tRotate<6,  T>(in); break;
+    case 7:  return tRotate<7,  T>(in); break;
+
+    case 8:  return tRotate<8,  T>(in); break;
+    case 9:  return tRotate<9,  T>(in); break;
+    case 10: return tRotate<10, T>(in); break;
+    case 11: return tRotate<11, T>(in); break;
+    case 12: return tRotate<12, T>(in); break;
+    case 13: return tRotate<13, T>(in); break;
+    case 14: return tRotate<14, T>(in); break;
+    case 15: return tRotate<15, T>(in); break;
+    default: assert(0);
+    }
+  }
+};
+
+// tree-based reduction
+#define svred(pg, v)\
+svaddv(pg, v);
+
+// left-to-right reduction
+// #define svred(pg, v)\
+// svadda(pg, 0, v)
+
+template <typename Out_type, typename In_type>
+struct Reduce{
+  //Need templated class to overload output type
+  //General form must generate error if compiled
+  inline Out_type operator()(In_type in){
+    printf("Error, using wrong Reduce function\n");
+    exit(1);
+    return 0;
+  }
+};
+
+//Complex float Reduce
+template <>
+inline Grid::ComplexF Reduce<Grid::ComplexF, vecf>::operator()(vecf in){
+  svbool_t pg1 = acle<float>::pg1();
+  svbool_t pg_even = acle<float>::pg_even();
+  svbool_t pg_odd  = acle<float>::pg_odd();
+  typename acle<float>::vt a_v = svld1(pg1, in.v);
+  float a = svred(pg_even, a_v);
+  float b = svred(pg_odd, a_v);
+
+  return Grid::ComplexF(a, b);
+
+}
+
+//Real float Reduce
+template <>
+inline Grid::RealF Reduce<Grid::RealF, vecf>::operator()(vecf in){
+  svbool_t pg1 = acle<float>::pg1();
+  typename acle<float>::vt a_v = svld1(pg1, in.v);
+  float a = svred(pg1, a_v);
+
+  return a;
+}
+
+//Complex double Reduce
+template <>
+inline Grid::ComplexD Reduce<Grid::ComplexD, vecd>::operator()(vecd in){
+  svbool_t pg1 = acle<double>::pg1();
+  svbool_t pg_even = acle<double>::pg_even();
+  svbool_t pg_odd  = acle<double>::pg_odd();
+  typename acle<double>::vt a_v = svld1(pg1, in.v);
+  double a = svred(pg_even, a_v);
+  double b = svred(pg_odd, a_v);
+
+  return Grid::ComplexD(a, b);
+}
+
+//Real double Reduce
+template <>
+inline Grid::RealD Reduce<Grid::RealD, vecd>::operator()(vecd in){
+  svbool_t pg1 = acle<double>::pg1();
+  typename acle<double>::vt a_v = svld1(pg1, in.v);
+  double a = svred(pg1, a_v);
+
+  return a;
+}
+
+//Integer Reduce
+template <>
+inline Integer Reduce<Integer, veci>::operator()(veci in){
+  svbool_t pg1 = acle<Integer>::pg1();
+  typename acle<Integer>::vt a_v = svld1(pg1, in.v);
+  Integer a = svred(pg1, a_v);
+
+  return a;
+}
+
+#undef svred
+#undef vec_imm
+
+NAMESPACE_END(Optimization)
+
+//////////////////////////////////////////////////////////////////////////////////////
+// Here assign types
+
+typedef Optimization::vech SIMD_Htype; // Reduced precision type
+typedef Optimization::vecf SIMD_Ftype; // Single precision type
+typedef Optimization::vecd SIMD_Dtype; // Double precision type
+typedef Optimization::veci SIMD_Itype; // Integer type
+
+// prefetch utilities
+inline void v_prefetch0(int size, const char *ptr){};
+inline void prefetch_HINT_T0(const char *ptr){};
+
+// Function name aliases
+typedef Optimization::Vsplat   VsplatSIMD;
+typedef Optimization::Vstore   VstoreSIMD;
+typedef Optimization::Vset     VsetSIMD;
+typedef Optimization::Vstream  VstreamSIMD;
+template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
+
+// Arithmetic operations
+typedef Optimization::Sum            SumSIMD;
+typedef Optimization::Sub            SubSIMD;
+typedef Optimization::Div            DivSIMD;
+typedef Optimization::Mult           MultSIMD;
+typedef Optimization::MultComplex    MultComplexSIMD;
+typedef Optimization::MultAddComplex MultAddComplexSIMD;
+typedef Optimization::MultRealPart   MultRealPartSIMD;
+typedef Optimization::MaddRealPart   MaddRealPartSIMD;
+typedef Optimization::Conj           ConjSIMD;
+typedef Optimization::TimesMinusI    TimesMinusISIMD;
+typedef Optimization::TimesI         TimesISIMD;
+
+NAMESPACE_END(Grid)
diff --git a/Grid/simd/Grid_a64fx-fixedsize.h b/Grid/simd/Grid_a64fx-fixedsize.h
new file mode 100644
index 00000000..6b450012
--- /dev/null
+++ b/Grid/simd/Grid_a64fx-fixedsize.h
@@ -0,0 +1,769 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Grid_a64fx-fixedsize.h
+
+    Copyright (C) 2020
+
+    Author: Nils Meyer         <nils.meyer@ur.de>           Regensburg University
+
+    with support from Arm
+            Richard Sandiford  <richard.sandiford@arm.com>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+/////////////////////////////////////////////////////
+// Using SVE ACLE with fixed-size data types
+/////////////////////////////////////////////////////
+
+
+// gcc 10 features
+#if __ARM_FEATURE_SVE_BITS==512
+/* gcc 10.0.1 and gcc 10.1 bug using ACLE data types  CAS-159553-Y1K4C6
+   workaround: use gcc's internal data types, bugfix expected for gcc 10.2
+typedef svbool_t    pred __attribute__((arm_sve_vector_bits(512)));
+typedef svfloat16_t vech __attribute__((arm_sve_vector_bits(512)));
+typedef svfloat32_t vecf __attribute__((arm_sve_vector_bits(512)));
+typedef svfloat64_t vecd __attribute__((arm_sve_vector_bits(512)));
+typedef svuint32_t  veci __attribute__((arm_sve_vector_bits(512)));
+typedef svuint32_t  lutf __attribute__((arm_sve_vector_bits(512))); // LUTs for float
+typedef svuint64_t  lutd __attribute__((arm_sve_vector_bits(512))); // LUTs for double
+*/
+typedef __SVBool_t    pred __attribute__((arm_sve_vector_bits(512)));
+typedef __SVFloat16_t vech __attribute__((arm_sve_vector_bits(512)));
+typedef __SVFloat32_t vecf __attribute__((arm_sve_vector_bits(512)));
+typedef __SVFloat64_t vecd __attribute__((arm_sve_vector_bits(512)));
+typedef __SVUint32_t  veci __attribute__((arm_sve_vector_bits(512)));
+typedef __SVUint32_t  lutf __attribute__((arm_sve_vector_bits(512))); // LUTs for float
+typedef __SVUint64_t  lutd __attribute__((arm_sve_vector_bits(512))); // LUTs for double
+#else
+#pragma error("Oops. Illegal SVE vector size!?")
+#endif /* __ARM_FEATURE_SVE_BITS */
+
+// low-level API
+NAMESPACE_BEGIN(Grid);
+NAMESPACE_BEGIN(Optimization);
+
+// convenience union types for tables eliminating loads
+union ulutf {
+  lutf v;
+  uint32_t s[16];
+};
+union ulutd {
+  lutd v;
+  uint64_t s[8];
+};
+
+template <typename T>
+struct acle{};
+
+template <>
+struct acle<double>{
+  static inline lutd tbl_swap(){
+    const ulutd t = { .s = {1, 0, 3, 2, 5, 4, 7, 6} };
+    return t.v;
+  }
+  static inline lutd tbl0(){
+    const ulutd t = { .s = {4, 5, 6, 7, 0, 1, 2, 3} };
+    return t.v;
+  }
+  static inline lutd tbl1(){
+    const ulutd t = { .s = {2, 3, 0, 1, 6, 7, 4, 5} };
+    return t.v;
+  }
+  static inline lutd tbl_exch1a(){ // Exchange1
+    const ulutd t = { .s = {0, 1, 4, 5, 2, 3, 6, 7} };
+    return t.v;
+  }
+  static inline lutd tbl_exch1b(){ // Exchange1
+    const ulutd t = { .s = {2, 3, 6, 7, 0, 1, 4, 5} };
+    return t.v;
+  }
+  static inline lutd tbl_exch1c(){ // Exchange1
+    const ulutd t = { .s = {4, 5, 0, 1, 6, 7, 2, 3} };
+    return t.v;
+  }
+  static inline pred pg1(){return svptrue_b64();}
+  static inline pred pg_even(){return svzip1_b64(svptrue_b64(), svpfalse_b());}
+  static inline pred pg_odd() {return svzip1_b64(svpfalse_b(), svptrue_b64());}
+  static inline vecd zero(){return svdup_f64(0.);}
+};
+
+template <>
+struct acle<float>{
+  // exchange neighboring elements
+  static inline lutf tbl_swap(){
+    const ulutf t = { .s = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14} };
+    return t.v;
+  }
+  static inline lutf tbl0(){
+    const ulutf t = { .s = {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7} };
+    return t.v;
+  }
+  static inline lutf tbl1(){
+    const ulutf t = { .s = {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11} };
+    return t.v;
+  }
+  static inline lutf tbl2(){
+    const ulutf t = { .s = {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13} };
+    return t.v;
+  }
+  static inline lutf tbl_exch1a(){ // Exchange1
+    const ulutf t = { .s = {0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 } };
+    return t.v;
+  }
+  static inline lutf tbl_exch1b(){ // Exchange1
+    const ulutf t = { .s = {4, 5, 6, 7, 12, 13, 14, 15, 0, 1, 2, 3, 8, 9, 10, 11 } };
+    return t.v;
+  }
+  static inline lutf tbl_exch1c(){ // Exchange1
+    const ulutf t = { .s = {8, 9, 10, 11, 0, 1, 2, 3, 12, 13, 14, 15, 4, 5, 6, 7} };
+    return t.v;
+  }
+  static inline pred pg1(){return svptrue_b32();}
+  static inline pred pg_even(){return svzip1_b32(svptrue_b32(), svpfalse_b());}
+  static inline pred pg_odd() {return svzip1_b32(svpfalse_b(), svptrue_b32());}
+  static inline vecf zero(){return svdup_f32(0.);}
+};
+
+template <>
+struct acle<uint16_t>{
+  static inline pred pg1(){return svptrue_b16();}
+  static inline pred pg_even(){return svzip1_b16(svptrue_b16(), svpfalse_b());}
+  static inline pred pg_odd() {return svzip1_b16(svpfalse_b(), svptrue_b16());}
+  static inline vech zero(){return svdup_f16(0.);}
+};
+
+template <>
+struct acle<Integer>{
+  //static inline svbool_t pg1(){return svptrue_b16();}
+  static inline pred pg1(){return svptrue_b32();}
+  static inline pred pg_even(){return svzip1_b32(svptrue_b32(), svpfalse_b());}
+  static inline pred pg_odd() {return svzip1_b32(svpfalse_b(), svptrue_b32());}
+};
+
+// ---------------------------------------------------
+
+struct Vsplat{
+  // Complex float
+  inline vecf operator()(float a, float b){
+    vecf a_v = svdup_f32(a);
+    vecf b_v = svdup_f32(b);
+    return svzip1(a_v, b_v);
+  }
+  // Real float
+  inline vecf operator()(float a){
+    return svdup_f32(a);
+  }
+  // Complex double
+  inline vecd operator()(double a, double b){
+    vecd a_v = svdup_f64(a);
+    vecd b_v = svdup_f64(b);
+    return svzip1(a_v, b_v);
+  }
+  // Real double
+  inline vecd operator()(double a){
+    return svdup_f64(a);
+  }
+  // Integer
+  inline veci operator()(Integer a){
+    return svdup_u32(a);
+  }
+};
+
+struct Vstore{
+  // Real float
+  inline void operator()(vecf a, float *D){
+    pred pg1 = acle<float>::pg1();
+    svst1(pg1, D, a);
+  }
+  // Real double
+  inline void operator()(vecd a, double *D){
+    pred pg1 = acle<double>::pg1();
+    svst1(pg1, D, a);
+  }
+  // Real float
+  inline void operator()(veci a, Integer *D){
+    pred pg1 = acle<Integer>::pg1();
+    svst1(pg1, D, a);
+  }
+};
+
+struct Vstream{
+  // Real float
+  inline void operator()(float * a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    svstnt1(pg1, a, b);
+    //svst1(pg1, a, b);
+  }
+  // Real double
+  inline void operator()(double * a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    svstnt1(pg1, a, b);
+    //svst1(pg1, a, b);
+  }
+};
+
+struct Vset{
+  // Complex float
+  inline vecf operator()(Grid::ComplexF *a){
+    pred pg1 = acle<float>::pg1();
+    return svld1(pg1, (float*)a);
+  }
+  // Complex double
+  inline vecd operator()(Grid::ComplexD *a){
+    pred pg1 = acle<double>::pg1();
+    return svld1(pg1, (double*)a);
+  }
+  // Real float
+  inline vecf operator()(float *a){
+    pred pg1 = acle<float>::pg1();
+    return svld1(pg1, a);
+  }
+  // Real double
+  inline vecd operator()(double *a){
+    pred pg1 = acle<double>::pg1();
+    return svld1(pg1, a);
+  }
+  // Integer
+  inline veci operator()(Integer *a){
+    pred pg1 = acle<Integer>::pg1();
+    return svld1(pg1, a);
+  }
+};
+
+/////////////////////////////////////////////////////
+// Arithmetic operations
+/////////////////////////////////////////////////////
+
+struct Sum{
+  // Complex/real float
+  inline vecf operator()(vecf a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    return svadd_x(pg1, a, b);
+  }
+  // Complex/real double
+  inline vecd operator()(vecd a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    return svadd_x(pg1, a, b);
+  }
+  // Integer
+  inline veci operator()(veci a, veci b){
+    pred pg1 = acle<Integer>::pg1();
+    return svadd_x(pg1, a, b);
+  }
+};
+
+struct Sub{
+  // Complex/real float
+  inline vecf operator()(vecf a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    return svsub_x(pg1, a, b);
+  }
+  // Complex/real double
+  inline vecd operator()(vecd a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    return svsub_x(pg1, a, b);
+  }
+  // Integer
+  inline veci operator()(veci a, veci b){
+    pred pg1 = acle<Integer>::pg1();
+    return svsub_x(pg1, a, b);
+  }
+
+};
+
+struct Mult{
+  // Real float fma
+  inline vecf operator()(vecf a, vecf b, vecf c){
+    pred pg1 = acle<float>::pg1();
+    return svmad_x(pg1, b, c, a);
+  }
+  // Real double fma
+  inline vecd operator()(vecd a, vecd b, vecd c){
+    pred pg1 = acle<double>::pg1();
+    return svmad_x(pg1, b, c, a);
+  }
+  // Real float
+  inline vecf operator()(vecf a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    return svmul_x(pg1, a, b);
+  }
+  // Real double
+  inline vecd operator()(vecd a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    return svmul_x(pg1, a, b);
+  }
+  // Integer
+  inline veci operator()(veci a, veci b){
+    pred pg1 = acle<Integer>::pg1();
+    return svmul_x(pg1, a, b);
+  }
+};
+
+struct MultRealPart{
+  // Complex float
+  inline vecf operator()(vecf a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    // using FCMLA
+    vecf z_v = acle<float>::zero();
+    return svcmla_x(pg1, z_v, a, b, 0);
+  }
+  // Complex double
+  inline vecd operator()(vecd a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    // using FCMLA
+    vecd z_v = acle<double>::zero();
+    return svcmla_x(pg1, z_v, a, b, 0);
+  }
+};
+
+struct MaddRealPart{
+  // Complex float
+  inline vecf operator()(vecf a, vecf b, vecf c){
+    pred pg1 = acle<float>::pg1();
+    // using FCMLA
+    return svcmla_x(pg1, c, a, b, 0);
+  }
+  // Complex double
+  inline vecd operator()(vecd a, vecd b, vecd c){
+    pred pg1 = acle<double>::pg1();
+    // using FCMLA
+    return svcmla_x(pg1, c, a, b, 0);
+  }
+};
+
+struct MultComplex{
+  // Complex a*b
+  // Complex float
+  inline vecf operator()(vecf a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    vecf z = acle<float>::zero();
+    // using FCMLA
+    vecf r_v = svcmla_x(pg1, z, a, b, 0);
+    return svcmla_x(pg1, r_v, a, b, 90);
+  }
+  // Complex double
+  inline vecd operator()(vecd a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    vecd z = acle<double>::zero();
+    // using FCMLA
+    vecd r_v = svcmla_x(pg1, z, a, b, 0);
+    return svcmla_x(pg1, r_v, a, b, 90);
+  }
+};
+
+struct MultAddComplex{
+  // Complex a*b+c
+  // Complex float
+  inline vecf operator()(vecf a, vecf b, vecf c){
+    pred pg1 = acle<float>::pg1();
+    // using FCMLA
+    vecf r_v = svcmla_x(pg1, c, a, b, 0);
+    return svcmla_x(pg1, r_v, a, b, 90);
+  }
+  // Complex double
+  inline vecd operator()(vecd a, vecd b, vecd c){
+    pred pg1 = acle<double>::pg1();
+    // using FCMLA
+    vecd r_v = svcmla_x(pg1, c, a, b, 0);
+    return svcmla_x(pg1, r_v, a, b, 90);
+  }
+};
+
+struct Div{
+  // Real float
+  inline vecf operator()(vecf a, vecf b){
+    pred pg1 = acle<float>::pg1();
+    return svdiv_x(pg1, a, b);
+  }
+  // Real double
+  inline vecd operator()(vecd a, vecd b){
+    pred pg1 = acle<double>::pg1();
+    return svdiv_x(pg1, a, b);
+  }
+};
+
+struct Conj{
+  // Complex float
+  inline vecf operator()(vecf a){
+    pred pg_odd = acle<float>::pg_odd();
+    //return svneg_x(pg_odd, a);  this is unsafe
+    return svneg_m(a, pg_odd, a);
+  }
+  // Complex double
+  inline vecd operator()(vecd a){
+    pred pg_odd = acle<double>::pg_odd();
+    //return svneg_x(pg_odd, a);  this is unsafe
+    return svneg_m(a, pg_odd, a);
+  }
+};
+
+struct TimesMinusI{
+  // Complex float
+  inline vecf operator()(vecf a, vecf b){
+    lutf tbl_swap = acle<float>::tbl_swap();
+    pred pg1 = acle<float>::pg1();
+    pred pg_odd = acle<float>::pg_odd();
+
+    vecf a_v = svtbl(a, tbl_swap);
+    //return svneg_x(pg_odd, a_v);  this is unsafe
+    return svneg_m(a_v, pg_odd, a_v);
+  }
+  // Complex double
+  inline vecd operator()(vecd a, vecd b){
+    lutd tbl_swap = acle<double>::tbl_swap();
+    pred pg1 = acle<double>::pg1();
+    pred pg_odd = acle<double>::pg_odd();
+
+    vecd a_v = svtbl(a, tbl_swap);
+    //return svneg_x(pg_odd, a_v);  this is unsafe
+    return svneg_m(a_v, pg_odd, a_v);
+  }
+};
+
+struct TimesI{
+  // Complex float
+  inline vecf operator()(vecf a, vecf b){
+    lutf tbl_swap = acle<float>::tbl_swap();
+    pred pg1 = acle<float>::pg1();
+    pred pg_even = acle<float>::pg_even();
+
+    vecf a_v = svtbl(a, tbl_swap);
+    //return svneg_x(pg_even, a_v);  this is unsafe
+    return svneg_m(a_v, pg_even, a_v);
+  }
+  // Complex double
+  inline vecd operator()(vecd a, vecd b){
+    lutd tbl_swap = acle<double>::tbl_swap();
+    pred pg1 = acle<double>::pg1();
+    pred pg_even = acle<double>::pg_even();
+
+    vecd a_v = svtbl(a, tbl_swap);
+    //return svneg_x(pg_even, a_v);  this is unsafe
+    return svneg_m(a_v, pg_even, a_v);
+  }
+};
+
+struct PrecisionChange {
+  static inline vech StoH (vecf sa, vecf sb) {
+    pred pg1s = acle<float>::pg1();
+    vech ha_v = svcvt_f16_x(pg1s, sa);
+    vech hb_v = svcvt_f16_x(pg1s, sb);
+    return svuzp1(ha_v, hb_v);
+  }
+  static inline void HtoS(vech h,vecf &sa,vecf &sb) {
+    pred pg1s = acle<float>::pg1();
+    vech ha_v = svzip1(h, h);
+    vech hb_v = svzip2(h, h);
+    sa = svcvt_f32_x(pg1s, ha_v);
+    sb = svcvt_f32_x(pg1s, hb_v);
+  }
+  static inline vecf DtoS (vecd a,vecd b) {
+    pred pg1d = acle<double>::pg1();
+    vecf sa_v = svcvt_f32_x(pg1d, a);
+    vecf sb_v = svcvt_f32_x(pg1d, b);
+    return svuzp1(sa_v, sb_v);
+  }
+  static inline void StoD (vecf s,vecd &a,vecd &b) {
+    pred pg1d = acle<double>::pg1();
+    vecf sa_v = svzip1(s, s);
+    vecf sb_v = svzip2(s, s);
+    a = svcvt_f64_x(pg1d, sa_v);
+    b = svcvt_f64_x(pg1d, sb_v);
+  }
+  static inline vech DtoH (vecd a,vecd b,vecd c,vecd d) {
+    pred pg1d = acle<double>::pg1();
+    pred pg1h = acle<uint16_t>::pg1();
+    vech ha_v = svcvt_f16_x(pg1d, a);
+    vech hb_v = svcvt_f16_x(pg1d, b);
+    vech hc_v = svcvt_f16_x(pg1d, c);
+    vech hd_v = svcvt_f16_x(pg1d, d);
+    vech hab_v = svuzp1(ha_v, hb_v);
+    vech hcd_v = svuzp1(hc_v, hd_v);
+    return svuzp1(hab_v, hcd_v);
+
+/*
+    vecf sa,sb;
+    sa = DtoS(a,b);
+    sb = DtoS(c,d);
+    return StoH(sa,sb);
+*/
+  }
+  static inline void HtoD(vech h,vecd &a,vecd &b,vecd &c,vecd &d) {
+    pred pg1h = acle<uint16_t>::pg1();
+    pred pg1d = acle<double>::pg1();
+    vech sa_v = svzip1(h, h);
+    vech sb_v = svzip2(h, h);
+    vech da_v = svzip1(sa_v, sa_v);
+    vech db_v = svzip2(sa_v, sa_v);
+    vech dc_v = svzip1(sb_v, sb_v);
+    vech dd_v = svzip2(sb_v, sb_v);
+    a = svcvt_f64_x(pg1d, da_v);
+    b = svcvt_f64_x(pg1d, db_v);
+    c = svcvt_f64_x(pg1d, dc_v);
+    d = svcvt_f64_x(pg1d, dd_v);
+
+/*
+    vecf sa,sb;
+    HtoS(h,sa,sb);
+    StoD(sa,a,b);
+    StoD(sb,c,d);
+*/
+  }
+};
+
+struct Exchange{
+  // float
+  static inline void Exchange0(vecf &out1, vecf &out2, vecf in1, vecf in2){
+    vecf r1_v = svext(in1, in1, (uint64_t)8u);
+    vecf r2_v = svext(in2, in2, (uint64_t)8u);
+    out1 = svext(r1_v, in2, (uint64_t)8u);
+    out2 = svext(in1, r2_v, (uint64_t)8u);
+  }
+  static inline void Exchange1(vecf &out1, vecf &out2, vecf in1, vecf in2){
+    // this one is tricky; svtrn2q* from SVE2 fits best, but it is not available in SVE1
+    // alternative: use 4-el structure; expect translation into 4x ldp + 4x stp -> SFI
+    lutf tbl_exch1a = acle<float>::tbl_exch1a();
+    lutf tbl_exch1b = acle<float>::tbl_exch1b();
+    lutf tbl_exch1c = acle<float>::tbl_exch1c();
+
+    vecf a1_v = svtbl(in1, tbl_exch1a);
+    vecf a2_v = svtbl(in2, tbl_exch1b);
+    vecf b1_v  = svext(a2_v, a1_v, (uint64_t)8u);
+    vecf b2_v  = svext(a1_v, a2_v, (uint64_t)8u);
+    out1 = svtbl(b1_v, tbl_exch1c);
+    out2 = svtbl(b2_v, tbl_exch1a);
+  }
+  static inline void Exchange2(vecf &out1, vecf &out2, vecf in1, vecf in2){
+    out1 = (vecf)svtrn1((vecd)in1, (vecd)in2);
+    out2 = (vecf)svtrn2((vecd)in1, (vecd)in2);
+  }
+  static inline void Exchange3(vecf &out1, vecf &out2, vecf in1, vecf in2){
+    out1 = svtrn1(in1, in2);
+    out2 = svtrn2(in1, in2);
+  }
+
+  // double
+  static inline void Exchange0(vecd &out1, vecd &out2, vecd in1, vecd in2){
+    vecd r1_v = svext(in1, in1, (uint64_t)4u);
+    vecd r2_v = svext(in2, in2, (uint64_t)4u);
+    out1 = svext(r1_v, in2, (uint64_t)4u);
+    out2 = svext(in1, r2_v, (uint64_t)4u);
+  }
+  static inline void Exchange1(vecd &out1, vecd &out2, vecd in1, vecd in2){
+    // this one is tricky; svtrn2q* from SVE2 fits best, but it is not available in SVE1
+    // alternative: use 4-el structure; expect translation into 4x ldp + 4x stp -> SFI
+    lutd tbl_exch1a = acle<double>::tbl_exch1a();
+    lutd tbl_exch1b = acle<double>::tbl_exch1b();
+    lutd tbl_exch1c = acle<double>::tbl_exch1c();
+
+    vecd a1_v = svtbl(in1, tbl_exch1a);
+    vecd a2_v = svtbl(in2, tbl_exch1b);
+    vecd b1_v = svext(a2_v, a1_v, (uint64_t)4u);
+    vecd b2_v = svext(a1_v, a2_v, (uint64_t)4u);
+    out1 = svtbl(b1_v, tbl_exch1c);
+    out2 = svtbl(b2_v, tbl_exch1a);
+  }
+  static inline void Exchange2(vecd &out1, vecd &out2, vecd in1, vecd in2){
+    out1 = svtrn1(in1, in2);
+    out2 = svtrn2(in1, in2);
+  }
+  static inline void Exchange3(vecd &out1, vecd &out2, vecd in1, vecd in2){
+    assert(0);
+    return;
+  }
+};
+
+#undef VECTOR_FOR
+
+struct Permute{
+  // float
+  static inline vecf Permute0(vecf in) {
+    return svext(in, in, (uint64_t)8u);
+  }
+  static inline vecf Permute1(vecf in) {
+    lutf tbl_swap = acle<float>::tbl1();
+    return svtbl(in, tbl_swap);
+  }
+  static inline vecf Permute2(vecf in) {
+    lutf tbl_swap = acle<float>::tbl2();
+    return svtbl(in, tbl_swap);
+  }
+  static inline vecf Permute3(vecf in) {
+    lutf tbl_swap = acle<float>::tbl_swap();
+    return svtbl(in, tbl_swap);
+  }
+
+  // double
+  static inline vecd Permute0(vecd in) {
+    return svext(in, in, (uint64_t)(8u / 2u));
+  }
+  static inline vecd Permute1(vecd in) {
+    lutd tbl_swap = acle<double>::tbl1();
+    return svtbl(in, tbl_swap);
+  }
+  static inline vecd Permute2(vecd in) {
+    lutd tbl_swap = acle<double>::tbl_swap();
+    return svtbl(in, tbl_swap);
+  }
+  static inline vecd Permute3(vecd in) {
+    return in;
+  }
+};
+
+struct Rotate{
+
+  static inline vecf rotate(vecf in, int n){
+    switch(n){
+    case 0:  return tRotate<0>(in); break;
+    case 1:  return tRotate<1>(in); break;
+    case 2:  return tRotate<2>(in); break;
+    case 3:  return tRotate<3>(in); break;
+    case 4:  return tRotate<4>(in); break;
+    case 5:  return tRotate<5>(in); break;
+    case 6:  return tRotate<6>(in); break;
+    case 7:  return tRotate<7>(in); break;
+
+    case 8:  return tRotate<8>(in); break;
+    case 9:  return tRotate<9>(in); break;
+    case 10: return tRotate<10>(in); break;
+    case 11: return tRotate<11>(in); break;
+    case 12: return tRotate<12>(in); break;
+    case 13: return tRotate<13>(in); break;
+    case 14: return tRotate<14>(in); break;
+    case 15: return tRotate<15>(in); break;
+    default: assert(0);
+    }
+  }
+  static inline vecd rotate(vecd in, int n){
+    switch(n){
+    case 0:  return tRotate<0>(in); break;
+    case 1:  return tRotate<1>(in); break;
+    case 2:  return tRotate<2>(in); break;
+    case 3:  return tRotate<3>(in); break;
+    case 4:  return tRotate<4>(in); break;
+    case 5:  return tRotate<5>(in); break;
+    case 6:  return tRotate<6>(in); break;
+    case 7:  return tRotate<7>(in); break;
+    default: assert(0);
+    }
+  }
+
+  template <int n> static inline vecf tRotate(vecf in){
+    return svext(in, in, (uint64_t)n);
+  }
+  template <int n> static inline vecd tRotate(vecd in){
+    return svext(in, in, (uint64_t)n);
+  }
+};
+
+// tree-based reduction
+#define svred(pg, v)\
+svaddv(pg, v);
+
+// left-to-right reduction
+// #define svred(pg, v)\
+// svadda(pg, 0, v)
+
+template <typename Out_type, typename In_type>
+struct Reduce{
+  //Need templated class to overload output type
+  //General form must generate error if compiled
+  inline Out_type operator()(In_type in){
+    printf("Error, using wrong Reduce function\n");
+    //exit(1);
+    return 0;
+  }
+};
+//Complex float Reduce
+template <>
+inline Grid::ComplexF Reduce<Grid::ComplexF, vecf>::operator()(vecf in){
+  pred pg_even = acle<float>::pg_even();
+  pred pg_odd  = acle<float>::pg_odd();
+  float a = svred(pg_even, in);
+  float b = svred(pg_odd, in);
+  return Grid::ComplexF(a, b);
+}
+//Real float Reduce
+template <>
+inline Grid::RealF Reduce<Grid::RealF, vecf>::operator()(vecf in){
+  pred pg1 = acle<float>::pg1();
+  return svred(pg1, in);
+}
+//Complex double Reduce
+template <>
+inline Grid::ComplexD Reduce<Grid::ComplexD, vecd>::operator()(vecd in){
+  pred pg_even = acle<double>::pg_even();
+  pred pg_odd  = acle<double>::pg_odd();
+  double a = svred(pg_even, in);
+  double b = svred(pg_odd, in);
+  return Grid::ComplexD(a, b);
+}
+//Real double Reduce
+template <>
+inline Grid::RealD Reduce<Grid::RealD, vecd>::operator()(vecd in){
+  pred pg1 = acle<double>::pg1();
+  return svred(pg1, in);
+}
+//Integer Reduce
+template <>
+inline Integer Reduce<Integer, veci>::operator()(veci in){
+  pred pg1 = acle<Integer>::pg1();
+  return svred(pg1, in);
+}
+
+#undef svred
+
+NAMESPACE_END(Optimization);
+
+//////////////////////////////////////////////////////////////////////////////////////
+// Here assign types
+
+typedef vech SIMD_Htype; // Reduced precision type
+typedef vecf SIMD_Ftype; // Single precision type
+typedef vecd SIMD_Dtype; // Double precision type
+typedef veci SIMD_Itype; // Integer type
+
+// prefetch utilities
+inline void v_prefetch0(int size, const char *ptr){};
+inline void prefetch_HINT_T0(const char *ptr){};
+
+// Function name aliases
+typedef Optimization::Vsplat   VsplatSIMD;
+typedef Optimization::Vstore   VstoreSIMD;
+typedef Optimization::Vset     VsetSIMD;
+typedef Optimization::Vstream  VstreamSIMD;
+template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
+
+// Arithmetic operations
+typedef Optimization::Sum            SumSIMD;
+typedef Optimization::Sub            SubSIMD;
+typedef Optimization::Div            DivSIMD;
+typedef Optimization::Mult           MultSIMD;
+typedef Optimization::MultComplex    MultComplexSIMD;
+typedef Optimization::MultAddComplex MultAddComplexSIMD;
+typedef Optimization::MultRealPart   MultRealPartSIMD;
+typedef Optimization::MaddRealPart   MaddRealPartSIMD;
+typedef Optimization::Conj           ConjSIMD;
+typedef Optimization::TimesMinusI    TimesMinusISIMD;
+typedef Optimization::TimesI         TimesISIMD;
+
+NAMESPACE_END(Grid);
diff --git a/Grid/simd/Grid_gpu_vec.h b/Grid/simd/Grid_gpu_vec.h
index 4584fb36..b9c6a81b 100644
--- a/Grid/simd/Grid_gpu_vec.h
+++ b/Grid/simd/Grid_gpu_vec.h
@@ -32,7 +32,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 */
 //----------------------------------------------------------------------
 
+#ifdef GRID_CUDA
 #include <cuda_fp16.h>
+#endif
+#ifdef GRID_HIP
+#include <hip/hip_fp16.h>
+#endif
 
 namespace Grid {
 
@@ -142,7 +147,7 @@ typedef GpuVector<NSIMD_Integer,  Integer     > GpuVectorI;
 accelerator_inline float half2float(half h)
 {
   float f;
-#ifdef __CUDA_ARCH__
+#ifdef GRID_SIMT
   f = __half2float(h);
 #else 
   //f = __half2float(h);
@@ -156,7 +161,7 @@ accelerator_inline float half2float(half h)
 accelerator_inline half float2half(float f)
 {
   half h;
-#ifdef __CUDA_ARCH__
+#ifdef GRID_SIMT
   h = __float2half(f);
 #else
   Grid_half hh = sfw_float_to_half(f);
diff --git a/Grid/simd/Grid_vector_types.h b/Grid/simd/Grid_vector_types.h
index e2b1fd07..c07077a3 100644
--- a/Grid/simd/Grid_vector_types.h
+++ b/Grid/simd/Grid_vector_types.h
@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/simd/Grid_vector_types.h
 
@@ -73,7 +73,7 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
   const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
   unsigned int sign_mask = 0x80000000u;
   Grid_half o;
-    
+
   o.x = static_cast<unsigned short>(0x0u);
   unsigned int sign = f.u & sign_mask;
   f.u ^= sign;
@@ -93,7 +93,7 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
       o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
     } else {
       unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
-	
+
       // update exponent, rounding bias part 1
       f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
       // rounding bias part 2
@@ -101,7 +101,7 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
       // take the bits!
       o.x = static_cast<unsigned short>(f.u >> 13);
     }
-  } 
+  }
   o.x |= static_cast<unsigned short>(sign >> 16);
   return o;
 }
@@ -110,9 +110,63 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
 #ifdef GPU_VEC
 #include "Grid_gpu_vec.h"
 #endif
+/*
 #ifdef GEN
 #include "Grid_generic.h"
 #endif
+*/
+
+#ifdef GEN
+  #if defined(A64FX) || defined(A64FXFIXEDSIZE) // breakout A64FX SVE ACLE here
+    #include <arm_sve.h>
+    #if defined(A64FX) // VLA
+      #pragma message("building A64FX / SVE ACLE VLA")
+      #if defined(ARMCLANGCOMPAT)
+        #pragma message("applying data types patch")
+      #endif
+      #include "Grid_a64fx-2.h"
+    #endif
+    #if defined(A64FXFIXEDSIZE) // fixed size data types
+      #pragma message("building for A64FX / SVE ACLE fixed size")
+      #include "Grid_a64fx-fixedsize.h"
+    #endif
+  #else
+    //#pragma message("building GEN") // generic
+    #include "Grid_generic.h"
+  #endif
+#endif
+
+#ifdef A64FX
+  #include <arm_sve.h>
+  #ifdef __ARM_FEATURE_SVE_BITS
+    //#pragma message("building A64FX SVE VLS")
+    #include "Grid_a64fx-fixedsize.h"
+  #else
+    #pragma message("building A64FX SVE VLA")
+    #if defined(ARMCLANGCOMPAT)
+      #pragma message("applying data types patch")
+    #endif
+    #include "Grid_a64fx-2.h"
+  #endif
+#endif
+
+/*
+#ifdef A64FXVLA
+#pragma message("building A64FX VLA")
+#if defined(ARMCLANGCOMPAT)
+  #pragma message("applying data types patch")
+#endif
+#include <arm_sve.h>
+#include "Grid_a64fx-2.h"
+#endif
+
+#ifdef A64FXVLS
+#pragma message("building A64FX VLS")
+#include <arm_sve.h>
+#include "Grid_a64fx-fixedsize.h"
+#endif
+*/
+
 #ifdef SSE4
 #include "Grid_sse4.h"
 #endif
@@ -163,6 +217,12 @@ template <typename T> struct is_complex : public std::false_type {};
 template <> struct is_complex<ComplexD> : public std::true_type {};
 template <> struct is_complex<ComplexF> : public std::true_type {};
 
+template <typename T> struct is_ComplexD : public std::false_type {};
+template <> struct is_ComplexD<ComplexD> : public std::true_type {};
+
+template <typename T> struct is_ComplexF : public std::false_type {};
+template <> struct is_ComplexF<ComplexF> : public std::true_type {};
+
 template<typename T, typename V=void> struct is_real : public std::false_type {};
 template<typename T> struct is_real<T, typename std::enable_if<std::is_floating_point<T>::value,
   void>::type> : public std::true_type {};
@@ -170,7 +230,7 @@ template<typename T> struct is_real<T, typename std::enable_if<std::is_floating_
 template<typename T, typename V=void> struct is_integer : public std::false_type {};
 template<typename T> struct is_integer<T, typename std::enable_if<std::is_integral<T>::value,
   void>::type> : public std::true_type {};
-  
+
 template <typename T>              using IfReal    = Invoke<std::enable_if<is_real<T>::value, int> >;
 template <typename T>              using IfComplex = Invoke<std::enable_if<is_complex<T>::value, int> >;
 template <typename T>              using IfInteger = Invoke<std::enable_if<is_integer<T>::value, int> >;
@@ -223,6 +283,69 @@ public:
     return sizeof(Vector_type) / sizeof(Scalar_type);
   }
 
+  #ifdef ARMCLANGCOMPAT
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<!is_complex<S>::value, S>::type, Vector_type> &&rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b8(), (Scalar_type*)this, svld1(svptrue_b8(), (Scalar_type*)&(rhs.v)));
+      return *this;
+    };
+
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<!is_complex<S>::value, S>::type, Vector_type> &rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b8(), (Scalar_type*)this, svld1(svptrue_b8(), (Scalar_type*)&(rhs.v)));
+      return *this;
+    };
+
+    /*
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<is_complex<S>::value, S>::type, Vector_type> &&rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b8(), (int8_t*)this, svld1(svptrue_b8(), (int8_t*)&(rhs.v)));
+      return *this;
+    };
+
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<is_complex<S>::value, S>::type, Vector_type> &rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b8(), (int8_t*)this, svld1(svptrue_b8(), (int8_t*)&(rhs.v)));
+      return *this;
+    };
+    */
+
+    // ComplexF
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<is_ComplexF<S>::value, S>::type, Vector_type> &&rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b32(), (float*)this, svld1(svptrue_b32(), (float*)&(rhs.v)));
+      return *this;
+    };
+
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<is_ComplexF<S>::value, S>::type, Vector_type> &rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b32(), (float*)this, svld1(svptrue_b32(), (float*)&(rhs.v)));
+      return *this;
+    };
+
+    // ComplexD
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<is_ComplexD<S>::value, S>::type, Vector_type> &&rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b64(), (double*)this, svld1(svptrue_b64(), (double*)&(rhs.v)));
+      return *this;
+    };
+
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd &operator=(const Grid_simd<typename std::enable_if<is_ComplexD<S>::value, S>::type, Vector_type> &rhs) {
+      //v = rhs.v;
+      svst1(svptrue_b64(), (double*)this, svld1(svptrue_b64(), (double*)&(rhs.v)));
+      return *this;
+    };
+
+  #else
+
   accelerator_inline Grid_simd &operator=(const Grid_simd &&rhs) {
     v = rhs.v;
     return *this;
@@ -232,10 +355,23 @@ public:
     return *this;
   };  // faster than not declaring it and leaving to the compiler
 
+  #endif
 
   accelerator Grid_simd() = default;
-  accelerator_inline Grid_simd(const Grid_simd &rhs) : v(rhs.v){};  // compiles in movaps
-  accelerator_inline Grid_simd(const Grid_simd &&rhs) : v(rhs.v){};
+
+  #ifdef ARMCLANGCOMPAT
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd(const Grid_simd<typename std::enable_if<!is_complex<S>::value, S>::type, Vector_type> &rhs) { this->operator=(rhs); }
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd(const Grid_simd<typename std::enable_if<!is_complex<S>::value, S>::type, Vector_type> &&rhs) { this->operator=(rhs); }
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd(const Grid_simd<typename std::enable_if<is_complex<S>::value, S>::type, Vector_type> &rhs) { this->operator=(rhs); }
+    template <class S = Scalar_type>
+    accelerator_inline Grid_simd(const Grid_simd<typename std::enable_if<is_complex<S>::value, S>::type, Vector_type> &&rhs) { this->operator=(rhs); }
+  #else
+    accelerator_inline Grid_simd(const Grid_simd &rhs) : v(rhs.v){};  // compiles in movaps
+    accelerator_inline Grid_simd(const Grid_simd &&rhs) : v(rhs.v){};
+  #endif
   accelerator_inline Grid_simd(const Real a) { vsplat(*this, Scalar_type(a)); };
   // Enable if complex type
   template <typename S = Scalar_type> accelerator_inline
@@ -258,12 +394,21 @@ public:
   ///////////////////////////////////////////////
 
   // FIXME -- alias this to an accelerator_inline MAC struct.
+
+  #if defined(A64FX) || defined(A64FXFIXEDSIZE)
+  friend accelerator_inline void mac(Grid_simd *__restrict__ y,
+				     const Grid_simd *__restrict__ a,
+				     const Grid_simd *__restrict__ x) {
+    *y = fxmac((*a), (*x), (*y));
+  };
+  #else
   friend accelerator_inline void mac(Grid_simd *__restrict__ y,
 				     const Grid_simd *__restrict__ a,
 				     const Grid_simd *__restrict__ x) {
     *y = (*a) * (*x) + (*y);
   };
-  
+  #endif
+
   friend accelerator_inline void mult(Grid_simd *__restrict__ y,
 				      const Grid_simd *__restrict__ l,
 				      const Grid_simd *__restrict__ r) {
@@ -412,7 +557,7 @@ public:
     Grid_simd ret;
     Grid_simd::conv_t conv;
     Grid_simd::scalar_type s;
-    
+
     conv.v = v.v;
     for (int i = 0; i < Nsimd(); i++) {
       s = conv.s[i];
@@ -441,7 +586,7 @@ public:
     return ret;
   }
   ///////////////////////
-  // Exchange 
+  // Exchange
   // Al Ah , Bl Bh -> Al Bl Ah,Bh
   ///////////////////////
   friend accelerator_inline void exchange(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2,int n)
@@ -452,20 +597,20 @@ public:
       Optimization::Exchange::Exchange2(out1.v,out2.v,in1.v,in2.v);
     } else if(n==1) {
       Optimization::Exchange::Exchange1(out1.v,out2.v,in1.v,in2.v);
-    } else if(n==0) { 
+    } else if(n==0) {
       Optimization::Exchange::Exchange0(out1.v,out2.v,in1.v,in2.v);
     }
   }
-  friend accelerator_inline void exchange0(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){    
+  friend accelerator_inline void exchange0(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){
     Optimization::Exchange::Exchange0(out1.v,out2.v,in1.v,in2.v);
   }
-  friend accelerator_inline void exchange1(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){    
+  friend accelerator_inline void exchange1(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){
     Optimization::Exchange::Exchange1(out1.v,out2.v,in1.v,in2.v);
   }
-  friend accelerator_inline void exchange2(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){    
+  friend accelerator_inline void exchange2(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){
     Optimization::Exchange::Exchange2(out1.v,out2.v,in1.v,in2.v);
   }
-  friend accelerator_inline void exchange3(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){    
+  friend accelerator_inline void exchange3(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2){
     Optimization::Exchange::Exchange3(out1.v,out2.v,in1.v,in2.v);
   }
   ////////////////////////////////////////////////////////////////////
@@ -490,7 +635,7 @@ public:
       int dist = perm & 0xF;
       y = rotate(b, dist);
       return;
-    } 
+    }
     else if(perm==3) permute3(y, b);
     else if(perm==2) permute2(y, b);
     else if(perm==1) permute1(y, b);
@@ -564,29 +709,29 @@ accelerator_inline Grid_simd<S, V> rotate(Grid_simd<S, V> b, int nrot) {
   ret.v = Optimization::Rotate::rotate(b.v, 2 * nrot);
   return ret;
 }
-template <class S, class V, IfNotComplex<S> =0> 
+template <class S, class V, IfNotComplex<S> =0>
 accelerator_inline void rotate( Grid_simd<S,V> &ret,Grid_simd<S,V> b,int nrot)
 {
   nrot = nrot % Grid_simd<S,V>::Nsimd();
   ret.v = Optimization::Rotate::rotate(b.v,nrot);
 }
-template <class S, class V, IfComplex<S> =0> 
+template <class S, class V, IfComplex<S> =0>
 accelerator_inline void rotate(Grid_simd<S,V> &ret,Grid_simd<S,V> b,int nrot)
 {
   nrot = nrot % Grid_simd<S,V>::Nsimd();
   ret.v = Optimization::Rotate::rotate(b.v,2*nrot);
 }
 
-template <class S, class V> 
+template <class S, class V>
 accelerator_inline void vbroadcast(Grid_simd<S,V> &ret,const Grid_simd<S,V> &src,int lane){
   S* typepun =(S*) &src;
   vsplat(ret,typepun[lane]);
-}    
-template <class S, class V, IfComplex<S> =0> 
+}
+template <class S, class V, IfComplex<S> =0>
 accelerator_inline void rbroadcast(Grid_simd<S,V> &ret,const Grid_simd<S,V> &src,int lane){
   S* typepun =(S*) &src;
   ret.v = unary<V>(real(typepun[lane]), VsplatSIMD());
-}    
+}
 
 
 
@@ -741,6 +886,27 @@ accelerator_inline Grid_simd<S, V> operator*(Grid_simd<S, V> a, Grid_simd<S, V>
   return ret;
 };
 
+// ---------------- A64FX MAC -------------------
+// Distinguish between complex types and others
+#if defined(A64FX) || defined(A64FXFIXEDSIZE)
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline Grid_simd<S, V> fxmac(Grid_simd<S, V> a, Grid_simd<S, V> b, Grid_simd<S, V> c) {
+  Grid_simd<S, V> ret;
+  ret.v = trinary<V>(a.v, b.v, c.v, MultAddComplexSIMD());
+  return ret;
+};
+
+// Real/Integer types
+template <class S, class V, IfNotComplex<S> = 0>
+accelerator_inline Grid_simd<S, V> fxmac(Grid_simd<S, V> a, Grid_simd<S, V> b, Grid_simd<S, V> c) {
+  Grid_simd<S, V> ret;
+  ret.v = trinary<V>(a.v, b.v, c.v, MultSIMD());
+  return ret;
+};
+#endif
+// ----------------------------------------------
+
+
 // Distinguish between complex types and others
 template <class S, class V, IfComplex<S> = 0>
 accelerator_inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
@@ -877,7 +1043,7 @@ accelerator_inline typename toComplexMapper<Rsimd>::Complexified toComplex(const
 
   conv.v = in.v;
   for (int i = 0; i < Rsimd::Nsimd(); i += 2) {
-    assert(conv.s[i + 1] == conv.s[i]);  
+    assert(conv.s[i + 1] == conv.s[i]);
     // trap any cases where real was not duplicated
     // indicating the SIMD grids of real and imag assignment did not correctly
     // match
@@ -919,6 +1085,14 @@ accelerator_inline void precisionChange(vRealD    *out,vRealF    *in,int nvec)
   for(int m=0;m*2<nvec;m++){
     int n=m*2;
     Optimization::PrecisionChange::StoD(in[m].v,out[n].v,out[n+1].v);
+    // Bug in gcc 10.0.1 and gcc 10.1 using fixed-size SVE ACLE data types  CAS-159553-Y1K4C6
+    // function call results in compile-time error:
+    // In function ‘void Grid::precisionChange(Grid::vRealD*, Grid::vRealF*, int)’:
+    // .../Grid_vector_types.h:961:56: error:
+    // cannot bind non-const lvalue reference of type ‘vecd&’ {aka ‘svfloat64_t&’}
+    // to an rvalue of type ‘vecd’ {aka ‘svfloat64_t’}
+    // 961 |     Optimization::PrecisionChange::StoD(in[m].v,out[n].v,out[n+1].v);
+    //  |                                                 ~~~~~~~^
   }
 }
 accelerator_inline void precisionChange(vRealD    *out,vRealH    *in,int nvec)
diff --git a/Grid/simd/Simd.h b/Grid/simd/Simd.h
index bc8cd2fd..37aee2ed 100644
--- a/Grid/simd/Simd.h
+++ b/Grid/simd/Simd.h
@@ -31,7 +31,7 @@ directory
 #ifndef GRID_SIMD_H
 #define GRID_SIMD_H
 
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA) || defined(GRID_HIP)
 #include <thrust/complex.h>
 #endif
 
@@ -65,7 +65,7 @@ typedef RealD   Real;
 typedef RealF  Real;
 #endif
 
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA) || defined(GRID_HIP)
 typedef thrust::complex<RealF> ComplexF;
 typedef thrust::complex<RealD> ComplexD;
 typedef thrust::complex<Real>  Complex;
diff --git a/Grid/simd/gridverter.py b/Grid/simd/gridverter.py
new file mode 100755
index 00000000..f00a5019
--- /dev/null
+++ b/Grid/simd/gridverter.py
@@ -0,0 +1,2377 @@
+#!/usr/bin/python3
+
+import re
+import argparse
+import sys
+
+# Grid for A64FX
+#
+# * should align std::vector to (multiples of) cache block size = 256 bytes
+
+# place benchmark runtime in cycles here !
+measured_cycles = 690 #1500 #775 #1500
+
+
+# command line parser
+parser = argparse.ArgumentParser(description="Dslash generator.")
+parser.add_argument("--single", action="store_true", default="False")
+parser.add_argument("--double", action="store_true", default="True")
+parser.add_argument("--debug", action="store_true", default="False")
+parser.add_argument("--gridbench", action="store_true", default="False")
+args = parser.parse_args()
+
+print(args)
+
+ASM_LOAD_CHIMU = True       # load chimu
+ASM_LOAD_GAUGE = True       # load gauge
+ASM_LOAD_TABLE = True       # load table
+ASM_STORE = True            # store result
+
+# Disable all loads and stores in asm for benchmarking purposes
+#DISABLE_ASM_LOAD_STORE = True
+DISABLE_ASM_LOAD_STORE = False
+
+if DISABLE_ASM_LOAD_STORE:
+    ASM_LOAD_CHIMU = True       # load chimu
+    ASM_LOAD_GAUGE = True       # load gauge
+    ASM_LOAD_TABLE = True       # load table
+    ASM_STORE = False            # store result
+
+# Alternative implementation using PROJ specific loads works,
+# but be careful with predication
+
+ALTERNATIVE_LOADS = False
+#ALTERNATIVE_LOADS = not ALTERNATIVE_LOADS      # True
+
+# Alternative register mapping,
+# must use with my_wilson4.h and my_wilson4pf.h
+
+ALTERNATIVE_REGISTER_MAPPING = False
+#ALTERNATIVE_REGISTER_MAPPING = not ALTERNATIVE_REGISTER_MAPPING
+
+if ALTERNATIVE_REGISTER_MAPPING == True:
+    ALTERNATIVE_LOADS = False
+
+# use movprfx
+MOVPRFX = False
+MOVPRFX = not MOVPRFX
+
+
+PREFETCH = False
+PREFETCH = not PREFETCH # True
+
+PRECISION = 'double'   # DP by default
+PRECSUFFIX = 'A64FXd'
+if args.single == True:
+    PRECISION = 'single'
+    PRECSUFFIX = 'A64FXf'
+
+_DEBUG = False #True       # insert debugging output
+if args.debug == True:
+    _DEBUG = True
+
+GRIDBENCH = False
+if args.gridbench == True:
+    GRIDBENCH = True
+
+print("PRECISION                    = ", PRECISION)
+print("DEBUG                        = ", _DEBUG)
+print("ALTERNATIVE_LOADS            = ", ALTERNATIVE_LOADS)
+print("ALTERNATIVE_REGISTER_MAPPING = ", ALTERNATIVE_REGISTER_MAPPING)
+print("MOVPRFX                      = ", MOVPRFX)
+print("DISABLE_ASM_LOAD_STORE       = ", DISABLE_ASM_LOAD_STORE)
+print("GRIDBENCH                    = ", GRIDBENCH)
+
+print("")
+
+#sys.exit(0)
+
+
+#_DEBUG = True       # insert debugging output
+
+FETCH_BASE_PTR_COLOR_OFFSET = 2  # offset for scalar plus signed immediate addressing
+STORE_BASE_PTR_COLOR_OFFSET = 2
+
+# 64-bit gp register usage !!! armclang 20.0 complains about the register choice !!!
+# table address: x30
+# data address:  x29
+# store address: x28
+# debug address: r8
+
+# Max performance of complex FMA using FCMLA instruction
+# is 25% peak.
+#
+# Issue latency of FCMLA is 2 cycles.
+# Need 2 FCMLA instructions for complex FMA.
+# Complete complex FMA takes 4 cycles.
+# Peak throughput is 4 * 8 Flops DP = 32 Flops DP in 4 cycles.
+# A64FX FMA throughput is 4 * 8 * 2 * 2 = 132 Flops DP in 4 cycles.
+# -> 25% peak FMA
+#
+# In:  3x 512 bits = 192 bytes
+# Out: 1x 512 bits = 64 bytes
+# Tot: 4x 512 bits = 256 bytes
+#
+# 256 bytes * 2.2 GHz = 563.2 GB/s (base 10), 524 GB/s (base 2)
+
+OPT = """
+* interleave prefetching and compute in MULT_2SPIN
+* could test storing U's in MULT_2SPIN to L1d for cache line update
+* structure reordering: MAYBEPERM after MULT_2SPIN ?
+"""
+
+filename = 'XXX'
+LEGAL = """/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: {}
+
+    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 */
+"""
+
+class Register:
+
+    def __init__(self, variable, asmreg='X', predication=False):
+        global d
+        x = 'Y'
+        if predication == False:
+            x = asmreg # + d['asmsuffix']
+        else:
+            x = asmreg
+        self.asmreg = x
+        self.asmregwithsuffix = asmreg + d['asmsuffix']
+        self.asmregbyte = asmreg + '.b'
+        self.name = variable
+        self.asmname = variable
+        self.asmnamebyte = variable + '.b'
+        self.predication = predication
+
+        d['registers'] += 1
+
+    def define(self, statement):
+        global d
+        d['C'] += F'#define {self.name} {statement}'
+        #d['A'] += F'#define {self.name} {statement}'
+
+    def declare(self, predication=False):
+        global d
+
+        if self.predication == False:
+            d['C'] += F'    Simd {self.name};               \\\n'
+
+            predtype = 'svfloat64_t'
+            if PRECISION == 'single':
+                predtype = 'svfloat32_t'
+
+            d['I'] += F'    {predtype} {self.name};        \\\n'
+        else:
+            d['I'] += F'    svbool_t {self.name};        \\\n'
+        #d['A'] += F'#define {self.name} {self.asmreg} \n'
+
+    def loadpredication(self, target='A'):
+        global d
+        if (target == 'A'):
+            d['A'] += F'    "ptrue {self.asmregwithsuffix} \\n\\t" \\\n'
+            d['asmclobber'].append(F'"{self.asmreg}"')
+
+    def loadtable(self, t):
+        global d
+        d['load'] += d['factor']
+        gpr = d['asmtableptr']
+
+        cast = 'uint64_t'
+        #asm_opcode = 'ld1d'
+        #if PRECISION == 'single':
+        #   asm_opcode = 'ld1w'
+        #    cast = 'uint32_t'
+        asm_opcode = 'ldr'
+        if PRECISION == 'single':
+            asm_opcode = 'ldr'
+            cast = 'uint32_t'
+
+        d['I'] += F'    {self.name} = svld1(pg1, ({cast}*)&lut[{t}]);  \\\n'
+
+        # using immediate index break-out works
+        if asm_opcode == 'ldr':
+            # ldr version
+            d['A'] += F'    "{asm_opcode} {self.asmreg}, [%[tableptr], %[index], mul vl] \\n\\t" \\\n'
+        else:
+            # ld1 version
+            d['A'] += F'    "{asm_opcode} {{ {self.asmregwithsuffix} }}, {pg1.asmreg}/z, [%[tableptr], %[index], mul vl] \\n\\t" \\\n'
+
+        d['asminput'].append(F'[tableptr] "r" (&lut[0])')
+        d['asminput'].append(F'[index] "i" ({t})')
+        d['asmclobber'].append(F'"memory"')
+        d['asmclobber'].append(F'"cc"')
+
+    def load(self, address, target='ALL', cast='float64_t', colors=3, offset=FETCH_BASE_PTR_COLOR_OFFSET):
+        global d
+        d['load'] += d['factor']
+        indices = re.findall(r'\d+', address)
+        index = (int(indices[0]) - offset) * colors + int(indices[1])
+
+        #asm_opcode = 'ld1d'
+        #if PRECISION == 'single':
+        #asm_opcode = 'ld1w'
+        #    cast = 'float32_t'
+
+        asm_opcode = 'ldr'
+        if PRECISION == 'single':
+            asm_opcode = 'ldr'
+            cast = 'float32_t'
+
+        gpr = d['asmfetchbaseptr']
+        intrinfetchbase = d['intrinfetchbase']
+        if (target in ['ALL', 'C']):
+            d['C'] += F'    {self.name} = {address};        \\\n'
+        if (target in ['ALL', 'I']):
+#            d['I'] += F'    {self.name} = svldnt1(pg1, ({cast}*)({intrinfetchbase} + {index} * 64));  \\\n'
+            d['I'] += F'    {self.name} = svld1(pg1, ({cast}*)({intrinfetchbase} + {index} * 64));  \\\n'
+        if (target in ['ALL', 'A']):
+            if asm_opcode == 'ldr':
+                d['A'] += F'    "{asm_opcode} {self.asmreg}, [%[fetchptr], {index}, mul vl] \\n\\t" \\\n'
+            else:
+                d['A'] += F'    "{asm_opcode} {{ {self.asmregwithsuffix} }}, {pg1.asmreg}/z, [%[fetchptr], {index}, mul vl] \\n\\t" \\\n'
+
+    def store(self, address, cast='float64_t', colors=3, offset=STORE_BASE_PTR_COLOR_OFFSET):
+        global d
+        d['store'] += d['factor']
+        indices = re.findall(r'\d+', address)
+        index = (int(indices[0]) - offset) * colors + int(indices[1])
+
+        #asm_opcode = 'stnt1d'
+        #if PRECISION == 'single':
+        #    asm_opcode = 'stnt1w'
+        #    cast = 'float32_t'
+        asm_opcode = 'str'
+        if PRECISION == 'single':
+            asm_opcode = 'str'
+            cast = 'float32_t'
+
+        intrinstorebase = d['intrinstorebase']
+
+        d['C'] += F'    {address} = {self.name};        \\\n'
+        #d['I'] += F'    svstnt1(pg1, ({cast}*)({intrinstorebase} + {index} * 64), {self.name});  \\\n'
+        d['I'] += F'    svst1(pg1, ({cast}*)({intrinstorebase} + {index} * 64), {self.name});  \\\n'
+        if asm_opcode == 'str':
+            d['A'] += F'    "{asm_opcode} {self.asmreg}, [%[storeptr], {index}, mul vl] \\n\\t" \\\n'
+        else:
+            d['A'] += F'    "{asm_opcode} {{ {self.asmregwithsuffix} }}, {pg1.asmreg}, [%[storeptr], {index}, mul vl] \\n\\t" \\\n'
+
+    def movestr(self, str):
+        global d
+        #d['move'] += d['factor']
+        d['I'] += F'    {self.name} = {str};        \\\n'
+
+    def move(self, op1):
+        global d
+        d['move'] += d['factor']
+        d['C'] += F'    {self.name} = {op1.name};   \\\n'
+        d['I'] += F'    {self.name} = {op1.name};        \\\n'
+        d['A'] += F'    "mov {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix} \\n\\t" \\\n'
+
+    # a = a + b , a = b + c
+    def add(self, op1, op2=None):
+        global d
+        d['add'] += d['factor']
+        if op2 is None:
+            d['C'] += F'    {self.name} = {self.name} + {op1.name};   \\\n'
+            d['I'] += F'    {self.name} = svadd_x(pg1, {self.name}, {op1.name}); \\\n'
+            d['A'] += F'    "fadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {op1.asmregwithsuffix} \\n\\t"  \\\n'
+        else:
+            d['C'] += F'    {self.name} = {op1.name} + {op2.name};    \\\n'
+            d['I'] += F'    {self.name} = svadd_x(pg1, {op1.name}, {op2.name});  \\\n'
+            d['A'] += F'    "fadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix} \\n\\t"  \\\n'
+
+    # a = a -b , a = b - c
+    def sub(self, op1, op2=None):
+        global d
+        d['sub'] += d['factor']
+        if op2 is None:
+            d['C'] += F'    {self.name} = {self.name} - {op1.name};    \\\n'
+            d['I'] += F'    {self.name} = svsub_x(pg1, {self.name}, {op1.name}); \\\n'
+            d['A'] += F'    "fsub {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {op1.asmregwithsuffix} \\n\\t" \\\n'
+        else:
+            d['C'] += F'    {self.name} = {op1.name} - {op2.name};     \\\n'
+            d['I'] += F'    {self.name} = svsub_x(pg1, {op1.name}, {op2.name});  \\\n'
+            d['A'] += F'    "fsub {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix} \\n\\t" \\\n'
+
+    # a = a * b , a = b * c
+    def mul(self, op1, op2):
+        global d
+        d['mul'] += 2 * d['factor']
+        d['C'] += F'    {self.name} = {op1.name} * {op2.name};                \\\n'
+        d['I'] += F'    {self.name} = __svzero({self.name}); \\\n'
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 0); \\\n'
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 90); \\\n'
+        d['A'] += F'    "mov {self.asmregwithsuffix} , 0 \\n\\t" \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 0 \\n\\t" \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 90 \\n\\t" \\\n'
+
+    def mul0(self, op1, op2, op3=None, constructive=False):
+        global d
+        d['mul'] += d['factor']
+
+        # no movprfx intrinsics support
+        if constructive == True:
+            d['movprfx'] += d['factor']
+            d['I'] += F'    {self.name} = svcmla_x(pg1, {op1.name}, {op2.name}, {op3.name}, 0); \\\n'
+            d['A'] += F'    "movprfx {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix} \\n\\t" \\\n'
+            d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op2.asmregwithsuffix}, {op3.asmregwithsuffix}, 0 \\n\\t" \\\n'
+        else:
+            d['C'] += F'    {self.name} = {op1.name} * {op2.name};                \\\n'
+            d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 0); \\\n'
+            d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 0 \\n\\t" \\\n'
+
+    def mul1(self, op1, op2):
+        global d
+        d['mul'] += d['factor']
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 90); \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 90 \\n\\t" \\\n'
+
+    def mac(self, op1, op2):
+        global d
+        d['mac'] += 2 * d['factor']
+        d['C'] += F'    {self.name} = {self.name} + {op1.name} * {op2.name};    \\\n'
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 0); \\\n'
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 90); \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 0 \\n\\t" \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 90 \\n\\t" \\\n'
+
+    def mac0(self, op1, op2):
+        global d
+        d['mac'] += d['factor']
+        d['C'] += F'    {self.name} = {self.name} + {op1.name} * {op2.name};    \\\n'
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 0); \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 0 \\n\\t" \\\n'
+
+    def mac1(self, op1, op2):
+        global d
+        d['mac'] += d['factor']
+        d['I'] += F'    {self.name} = svcmla_x(pg1, {self.name}, {op1.name}, {op2.name}, 90); \\\n'
+        d['A'] += F'    "fcmla {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 90 \\n\\t" \\\n'
+
+    def zero(self, zeroreg=False):
+        d['zero'] += d['factor']
+        d['C'] += F'    {self.name} = 0; \\\n'
+        #d['I'] += F'    {self.name} = __svzero({self.name}); \\\n'   only armclang
+
+        if PRECISION == 'double':
+            d['I'] += F'    {self.name} = svdup_f64(0.); \\\n'
+        else:
+            d['I'] += F'    {self.name} = svdup_f32(0.); \\\n'
+
+        if zeroreg == True:
+            d['A'] += F'    "fmov {self.asmregwithsuffix} , 0 \\n\\t" \\\n'
+        else:
+            #using mov z, zero0      issue 1c, FLA, latency 6c
+            #d['A'] += F'    "mov {self.asmregwithsuffix} , {zero0.asmregwithsuffix} \\n\\t" \\\n'
+
+            #using mov z, 0      issue 1c, FLA, latency 6c
+            d['A'] += F'    "fmov {self.asmregwithsuffix} , 0 \\n\\t" \\\n'
+
+            #using xor z, z, z   issue 0.5c, FL*, latency 4c
+            #d['A'] += F'    "eor  {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {self.asmregwithsuffix} \\n\\t" \\\n'
+
+            #using and z, z, zero0  issue 0.5c, FL*, latency 4c
+            #d['A'] += F'    "and {self.asmregwithsuffix}, {self.asmregwithsuffix} , {zero0.asmregwithsuffix} \\n\\t" \\\n'
+
+            #using sub z, z, z    issue 0.5c, FL*, latency 9c
+            #d['A'] += F'    "sub  {self.asmregwithsuffix}, {self.asmregwithsuffix}, {self.asmregwithsuffix} \\n\\t" \\\n'
+
+    # without table
+    def timesI(self, op1, tempreg=None, tablereg=None):
+        global d
+        d['timesI'] += d['factor']
+        d['C'] += F'    {self.name} = timesI({op1.name});    \\\n'
+        # correct if DEBUG enabled, wrong if DEBUG disabled; no idea what's causing this
+        #table.load('table2', target='I', cast='uint64_t')
+        #d['I'] += F'    {self.name} = svtbl({op1.name}, {tablereg.name});  \\\n'
+        #d['I'] += F'    {self.name} = svneg_x(pg2, {self.name});  \\\n'
+        # timesI using trn tested, works but tbl should be faster
+        d['I'] += F'    {tempreg.name} = svtrn2({op1.name}, {op1.name});   \\\n'
+        d['I'] += F'    {tempreg.name} = svneg_x(pg1, {tempreg.name});   \\\n'
+        d['I'] += F'    {self.name} = svtrn1({tempreg.name}, {op1.name});   \\\n'
+        d['A'] += F'    "trn2 {tempreg.asmregwithsuffix}, {op1.asmregwithsuffix}, {op1.asmregwithsuffix} \\n\\t"   \\\n'
+        d['A'] += F'    "fneg {tempreg.asmregwithsuffix}, {pg1.asmreg}/m, {tempreg.asmregwithsuffix} \\n\\t"   \\\n'
+        d['A'] += F'    "trn1 {self.asmregwithsuffix}, {tempreg.asmregwithsuffix}, {op1.asmregwithsuffix} \\n\\t"   \\\n'
+
+    def addTimesI(self, op1, op2=None, constructive=False):
+        global d
+        d['addTimesI'] += d['factor']
+
+        if op2 is None:
+            d['C'] += F'    {self.name} = {self.name} + timesI({op1.name});    \\\n'
+        else:
+            d['C'] += F'    {self.name} = {op1.name} + timesI({op2.name});    \\\n'
+
+        # no movprfx intrinsics support
+        if constructive == True:
+            d['movprfx'] += d['factor']
+            d['I'] += F'    {self.name} = svcadd_x(pg1, {op1.name}, {op2.name}, 90);   \\\n'
+            d['A'] += F'    "movprfx {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix} \\n\\t" \\\n'
+            d['A'] += F'    "fcadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {op2.asmregwithsuffix}, 90 \\n\\t" \\\n'
+        else:
+            if op2 is None:
+                d['C'] += F'    {self.name} = {self.name} + timesI({op1.name});    \\\n'
+                d['I'] += F'    {self.name} = svcadd_x(pg1, {self.name}, {op1.name}, 90);   \\\n'
+                d['A'] += F'    "fcadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {op1.asmregwithsuffix}, 90 \\n\\t" \\\n'
+            else:
+                d['C'] += F'    {self.name} = {op1.name} + timesI({op2.name});    \\\n'
+                d['I'] += F'    {self.name} = svcadd_x(pg1, {op1.name}, {op2.name}, 90);   \\\n'
+                d['A'] += F'    "fcadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 90 \\n\\t" \\\n'
+
+    def subTimesI(self, op1, op2=None, constructive=False):
+        global d
+        d['subTimesI'] += d['factor']
+
+        # no movprfx intrinsics support
+        if constructive == True:
+            d['movprfx'] += d['factor']
+            d['I'] += F'    {self.name} = svcadd_x(pg1, {op1.name}, {op2.name}, 270);   \\\n'
+            d['A'] += F'    "movprfx {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix} \\n\\t" \\\n'
+            d['A'] += F'    "fcadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {op2.asmregwithsuffix}, 270 \\n\\t" \\\n'
+        else:
+            if op2 is None:
+                d['C'] += F'    {self.name} = {self.name} - timesI({op1.name});    \\\n'
+                d['I'] += F'    {self.name} = svcadd_x(pg1, {self.name}, {op1.name}, 270);   \\\n'
+                d['A'] += F'    "fcadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {self.asmregwithsuffix}, {op1.asmregwithsuffix}, 270 \\n\\t" \\\n'
+            else:
+                d['C'] += F'    {self.name} = {op1.name} - timesI({op2.name});    \\\n'
+                d['I'] += F'    {self.name} = svcadd_x(pg1, {op1.name}, {op2.name}, 270);   \\\n'
+                d['A'] += F'    "fcadd {self.asmregwithsuffix}, {pg1.asmreg}/m, {op1.asmregwithsuffix}, {op2.asmregwithsuffix}, 270 \\n\\t" \\\n'
+
+    # timesMinusI is not used, def is probably wrong !!!!  OPTIMIZATION with table
+    def timesMinusI(self, op1):
+        global d
+        d['timesMinusI'] += d['factor']
+        d['C'] += F'    {self.name} = timesMinusI({self.name});    \\\n'
+        d['I'] += F'    {self.name} = svtrn1({op1.name}, {op1.name});   \\\n'
+        d['I'] += F'    {self.name} = svneg_x(pg1, {self.name});   \\\n'
+        d['I'] += F'    {self.name} = svtrn1({op1.name}, {self.name});   \\\n'
+
+    def permute(self, dir, tablereg=None):
+        global d
+        d['permutes'] += d['factor']
+
+        d['C'] += F'    permute{dir}({self.name}, {self.name});    \\\n'
+
+        d['I'] += F'    {self.name} = svtbl({self.name}, {tablereg.name});    \\\n'
+        d['A'] += F'    "tbl {self.asmregwithsuffix}, {{ {self.asmregwithsuffix} }}, {tablereg.asmregwithsuffix} \\n\\t"  \\\n'
+
+        # if dir == 0:
+        #     d['I'] += F'    {self.name} = svext({self.name}, {self.name}, 4); \\\n'
+        #     # this might not work, see intrinsics assembly
+        #     # d['A'] += F'    ext {self.name}, {self.name}, {self.name}, #4 \\\n'
+        #     # use registers directly
+        #     d['A'] += F'    "ext {self.asmregbyte}, {self.asmregbyte}, {self.asmregbyte}, 32 \\n\\t" \\\n'
+        #
+        # elif dir in [1, 2]:
+        #     d['I'] += F'    {self.name} = svtbl({self.name}, {tablereg.name});    \\\n'
+        #     d['A'] += F'    "tbl {self.asmregwithsuffix}, {{ {self.asmregwithsuffix} }}, {tablereg.asmregwithsuffix} \\n\\t"  \\\n'
+
+    def debug(self):
+        global d
+        typecast = d['cfloat']
+        gpr = d['asmdebugptr']
+        vregs = d['asmclobberlist']
+        if (d['debug'] == True):
+            d['C'] += F'std::cout << "{self.name} -- " <<  {self.name} << std::endl; \\\n'
+
+            d['I'] += F'svst1(pg1, ({typecast}*)&debugreg.v, {self.name}); \\\n'
+            d['I'] += F'std::cout << "{self.name} -- " <<  debugreg << std::endl; \\\n'
+            #d['I'] += F'std::cout << "{self.name} -- " <<  {self.name} << std::endl; \\\n'
+
+            d['A'] += F'asm ( \\\n'
+            d['A'] += F'    " DMB SY \\n\\t " " DSB SY \\n\\t " " ISB SY \\n\\t " \\\n'   # memory barrier
+            d['A'] += F'    "str {self.asmreg}, [%[ptr]] \\n\\t" \\\n'
+            d['A'] += F'    " DMB SY \\n\\t " " DSB SY \\n\\t " " ISB SY \\n\\t " \\\n'   # memory barrier
+            d['A'] += F'    : "=m" (debugreg.v)  \\\n'
+            d['A'] += F'    : [ptr] "r" (&debugreg.v)  \\\n'
+            d['A'] += F'    : "p5", "cc", "memory" \\\n'
+            d['A'] += F'); \\\n'
+            d['A'] += F'std::cout << "{self.name} -- " <<  debugreg << std::endl; \\\n'
+            # this form of addressing is not valid!
+            #d['A'] += F'    "str {self.asmreg}, %[ptr] \\n\\t" \\\n'
+# end Register
+
+def define(s, target='ALL'):
+    x = F'#define {s}  \n'
+    global d
+    if (target in ['ALL', 'C']):
+        d['C'] += x
+    if (target in ['ALL', 'I']):
+        d['I'] += x
+    if (target in ['ALL', 'A']):
+        d['A'] += x
+
+def definemultiline(s):
+    x = F'#define {s}  \\\n'
+    global d
+    d['C'] += x
+    d['I'] += x
+    d['A'] += x
+
+def write(s, target='ALL'):
+    x = F'{s}\n'
+    global d
+    if (target in ['ALL', 'C']):
+        d['C'] += x
+    if (target in ['ALL', 'I']):
+        d['I'] += x
+    if (target in ['ALL', 'A']):
+        d['A'] += x
+
+def curlyopen():
+    write(F'{{ \\')
+
+def curlyclose():
+    write(F'}}')
+
+def newline(target='ALL'):
+    global d
+
+    if target == 'A':
+        if d['A'][-2:] == '\\\n':
+            d['A'] = d['A'][:-2] + '\n\n'
+    else:
+        if d['C'][-2:] == '\\\n':
+            d['C'] = d['C'][:-2] + '\n\n'
+        if d['I'][-2:] == '\\\n':
+            d['I'] = d['I'][:-2] + '\n\n'
+        if d['A'][-2:] == '\\\n':
+            d['A'] = d['A'][:-2] + '\n\n'
+
+# load the base pointer for fetches
+def fetch_base_ptr(address, target='A'):
+    global d
+    #d['load'] += d['factor']
+
+    # DEBUG
+    #colors=3
+    #indices = re.findall(r'\d+', address)
+    #index = (int(indices[0]) - FETCH_BASE_PTR_COLOR_OFFSET) * colors + int(indices[1])
+    #print(F'{address}                      (base)')
+
+    vregs = d['asmclobberlist']
+    if target == 'A':
+        d['asminput'].append(F'[fetchptr] "r" ({address})')
+        d['asmclobber'].extend(vregs)
+        d['asmclobber'].append(F'"memory"')
+        d['asmclobber'].append(F'"cc"')
+    if target == 'I':
+        #print("intrinfetchbase = ", address)
+        d['intrinfetchbase'] = address
+
+# load the base pointer for stores
+def store_base_ptr(address, target='A'):
+    global d
+    #d['load'] += d['factor']
+    gpr = d['asmstorebaseptr']
+    vregs = d['asmclobberlist']
+    if target == 'A':
+        d['asminput'].append(F'[storeptr] "r" ({address})')
+        d['asmclobber'].extend(vregs)
+        d['asmclobber'].append(F'"memory"')
+        d['asmclobber'].append(F'"cc"')
+    if target == 'I':
+        d['intrinstorebase'] = address
+
+def prefetch_L1(address, offset):
+    global d
+    multiplier = 4  # offset in CL, have to multiply by 4
+    policy = "PLDL1STRM"     # weak
+    #policy = "PLDL1KEEP"     # strong
+
+    d['I'] += F'    svprfd(pg1, (int64_t*)({address} + {offset * multiplier * 64}), SV_{policy}); \\\n'
+    d['A'] += F'    "prfd {policy}, {pg1.asmreg}, [%[fetchptr], {offset * multiplier}, mul vl] \\n\\t" \\\n'
+
+def prefetch_L2(address, offset):
+    global d
+    multiplier = 4  # offset in CL, have to multiply by 4
+    policy = "PLDL2STRM"     # weak
+    #policy = "PLDL2KEEP"     # strong
+
+    d['I'] += F'    svprfd(pg1, (int64_t*)({address} + {offset * multiplier * 64}), SV_{policy}); \\\n'
+    d['A'] += F'    "prfd {policy}, {pg1.asmreg}, [%[fetchptr], {offset * multiplier}, mul vl] \\n\\t" \\\n'
+    #d['A'] +=
+
+def prefetch_L2_store(address, offset):
+    global d
+    multiplier = 4  # offset in CL, have to multiply by 4
+    policy = "PSTL2STRM"     # weak
+    #policy = "PSTL2KEEP"     # strong
+
+    d['I'] += F'    svprfd(pg1, (int64_t*)({address} + {offset * multiplier * 64}), SV_{policy}); \\\n'
+    d['A'] += F'    "prfd {policy}, {pg1.asmreg}, [%[fetchptr], {offset * multiplier}, mul vl] \\n\\t" \\\n'
+
+def prefetch_L1_store(address, offset):
+    global d
+    multiplier = 4  # offset in CL, have to multiply by 4
+    policy = "PSTL1STRM"     # weak
+    #policy = "PSTL2KEEP"     # strong
+
+    d['I'] += F'    svprfd(pg1, (int64_t*)({address} + {offset * multiplier * 64}), SV_{policy}); \\\n'
+    d['A'] += F'    "prfd {policy}, {pg1.asmreg}, [%[fetchptr], {offset * multiplier}, mul vl] \\n\\t" \\\n'
+
+
+def asmopen():
+    #write('asm volatile ( \\', target='A')
+    write('asm ( \\', target='A')
+
+    # DEBUG
+    #write(F'    " DMB SY \\n\\t " " DSB SY \\n\\t " " ISB SY \\n\\t " \\', target='A')   # memory barrier
+    #write('asm volatile ( \\', target='A')
+
+def asmclose():
+    global d
+
+    #print(d['asminput'])
+
+    asmin  = d['asminput']
+    asmin_s = ''
+    if len(asmin) > 0:
+        asmin = list(dict.fromkeys(asmin))   # remove duplicates
+        #print(asmin)
+        for el in asmin:
+            asmin_s += el + ','
+        asmin_s = asmin_s[:-1]
+        #print("-> ", asmin_s)
+
+    d['asminput'] = []
+
+    asmout = d['asmoutput']
+    asmout_s = ''
+    if len(asmout) > 0:
+        asmout = list(dict.fromkeys(asmout))   # remove duplicates
+        for el in asmout:
+            asmout_s += el + ','
+        asmout_s = asmout_s[:-1]
+
+    d['asmoutput'] = []
+
+    # DEBUG  put all regs into clobber by default
+    d['asmclobber'].extend(d['asmclobberlist'])
+
+    asmclobber = d['asmclobber']
+    asmclobber_s = ''
+    #print(asmclobber)
+    if len(asmclobber) > 0:
+        asmclobber = list(dict.fromkeys(asmclobber))   # remove duplicates
+        for el in asmclobber:
+            asmclobber_s += el + ','
+        asmclobber_s = asmclobber_s[:-1]
+
+    d['asmclobber'] = []
+
+    # DEBUG
+    #write(F'    " DMB SY \\n\\t " " DSB SY \\n\\t " " ISB SY \\n\\t " \\', target='A')   # memory barrier
+
+
+    write(F'    : {asmout_s} \\', target='A')
+    write(F'    : {asmin_s} \\', target='A')
+    write(F'    : {asmclobber_s} \\', target='A')
+    write('); \\', target='A')
+
+# --------------------------------------------------------------------------------
+
+# string of vector registers to be used in clobber list
+#clobberlist = ['"p0"']
+clobberlist = ['"p5"']
+clobberlist.append('"cc"')
+for i in range(0, 32):
+    clobberlist.append(F'"z{i}"')
+
+d = {
+'debug': _DEBUG,
+'C': '',
+'I': '',
+'A': '',
+'asmsuffix': '.d',      # double precision by default
+'cfloat': 'float64_t',
+'registers': 0,
+'load': 0,
+'store': 0,
+'move': 0,
+'movprfx': 0,
+'zero': 0,
+'add': 0,
+'sub': 0,
+'mul': 0,
+'mac': 0,
+'permutes': 0,
+'neg': 0,
+'addTimesI': 0,
+'subTimesI': 0,
+'timesI': 0,
+'timesMinusI': 0,
+'flops': 0,
+'factor': 1,                    # multiplicity
+'asmtableptr': 'x30',
+'asmfetchbaseptr': 'x29',
+'asmstorebaseptr': 'x28',
+'asmdebugptr': 'r12',
+'asminput': [],
+'asmoutput': [],
+'asmclobber': [],
+'asmclobberlist': clobberlist,
+'intrinfetchbase': '',
+'intrinstorebase': '',
+'cycles_LOAD_CHIMU': 0,
+'cycles_PROJ': 0,
+'cycles_PERM': 0,
+'cycles_MULT_2SPIN': 0,
+'cycles_RECON': 0,
+'cycles_RESULT': 0,
+'cycles_ZERO_PSI': 0,
+'cycles_PREFETCH_L1': 0,
+'cycles_PREFETCH_L2': 0
+}
+
+if PRECISION == 'single':
+    d['asmsuffix'] = '.s'
+    d['cfloat'] = 'float32_t'
+
+# --------------------------------------------------------------------------------
+# Grid
+# --------------------------------------------------------------------------------
+
+# Variables / Registers
+result_00 = Register('result_00', asmreg='z0')
+result_01 = Register('result_01', asmreg='z1')
+result_02 = Register('result_02', asmreg='z2')
+result_10 = Register('result_10', asmreg='z3')
+result_11 = Register('result_11', asmreg='z4')
+result_12 = Register('result_12', asmreg='z5')
+result_20 = Register('result_20', asmreg='z6')
+result_21 = Register('result_21', asmreg='z7')
+result_22 = Register('result_22', asmreg='z8')
+result_30 = Register('result_30', asmreg='z9')
+result_31 = Register('result_31', asmreg='z10')
+result_32 = Register('result_32', asmreg='z11')       # 12 Regs
+Chi_00	  = Register('Chi_00', asmreg='z12')
+Chi_01	  = Register('Chi_01', asmreg='z13')
+Chi_02	  = Register('Chi_02', asmreg='z14')
+Chi_10	  = Register('Chi_10', asmreg='z15')
+Chi_11	  = Register('Chi_11', asmreg='z16')
+Chi_12	  = Register('Chi_12', asmreg='z17')          # 6
+UChi_00   = Register('UChi_00', asmreg='z18')
+UChi_01   = Register('UChi_01', asmreg='z19')
+UChi_02   = Register('UChi_02', asmreg='z20')
+UChi_10   = Register('UChi_10', asmreg='z21')
+UChi_11   = Register('UChi_11', asmreg='z22')
+UChi_12   = Register('UChi_12', asmreg='z23')         # 6
+U_00	  = Register('U_00', asmreg='z24')
+U_10	  = Register('U_10', asmreg='z25')
+U_20	  = Register('U_20', asmreg='z26')
+U_01	  = Register('U_01', asmreg='z27')
+U_11	  = Register('U_11', asmreg='z28')
+U_21      = Register('U_21', asmreg='z29')            # 6    -> 30 Registers
+
+table0    = Register('table0', asmreg='z30')
+zero0     = Register('zero0', asmreg='z31')           # 2    -> 32 Registers
+# can't overload temp1 / table due to type mismatch using intrinsics :(
+# typecasting SVE intrinsics variables is not allowed
+
+pg1       = Register('pg1', predication=True, asmreg='p5')
+#pg2       = Register('pg2', predication=True, asmreg='p1')
+
+# Overloaded with Chi_* and UChi_*
+Chimu_00  = Register('Chimu_00', asmreg=Chi_00.asmreg)
+Chimu_01  = Register('Chimu_01', asmreg=Chi_01.asmreg)
+Chimu_02  = Register('Chimu_02', asmreg=Chi_02.asmreg)
+Chimu_10  = Register('Chimu_10', asmreg=Chi_10.asmreg)
+Chimu_11  = Register('Chimu_11', asmreg=Chi_11.asmreg)
+Chimu_12  = Register('Chimu_12', asmreg=Chi_12.asmreg)
+if ALTERNATIVE_REGISTER_MAPPING == False:
+    Chimu_20  = Register('Chimu_20', asmreg=UChi_00.asmreg)
+    Chimu_21  = Register('Chimu_21', asmreg=UChi_01.asmreg)
+    Chimu_22  = Register('Chimu_22', asmreg=UChi_02.asmreg)
+    Chimu_30  = Register('Chimu_30', asmreg=UChi_10.asmreg)
+    Chimu_31  = Register('Chimu_31', asmreg=UChi_11.asmreg)
+    Chimu_32  = Register('Chimu_32', asmreg=UChi_12.asmreg)        # 12 Registers
+else: # wilson4.h
+    Chimu_20  = Register('Chimu_20', asmreg=U_00.asmreg)
+    Chimu_21  = Register('Chimu_21', asmreg=U_10.asmreg)
+    Chimu_22  = Register('Chimu_22', asmreg=U_20.asmreg)
+    Chimu_30  = Register('Chimu_30', asmreg=U_01.asmreg)
+    Chimu_31  = Register('Chimu_31', asmreg=U_11.asmreg)
+    Chimu_32  = Register('Chimu_32', asmreg=U_21.asmreg)
+
+# debugging output
+def debugall(msg=None, group='ALL'):
+    global d
+    if (d['debug'] == False):
+        return
+    write(F'std::cout << std::endl << "DEBUG -- {msg}" << std::endl; \\')
+    if (group in ['ALL', 'result']):
+        result_00.debug()
+        result_01.debug()
+        result_02.debug()
+        result_10.debug()
+        result_11.debug()
+        result_12.debug()
+        result_20.debug()
+        result_21.debug()
+        result_22.debug()
+        result_30.debug()
+        result_31.debug()
+        result_32.debug()
+    if (group in ['ALL', 'Chi']):
+        Chi_00.debug()
+        Chi_01.debug()
+        Chi_02.debug()
+        Chi_10.debug()
+        Chi_11.debug()
+        Chi_12.debug()
+    if (group in ['ALL', 'UChi']):
+        UChi_00.debug()
+        UChi_01.debug()
+        UChi_02.debug()
+        UChi_10.debug()
+        UChi_11.debug()
+        UChi_12.debug()
+    if (group in ['ALL', 'U']):
+        U_00.debug()
+        U_10.debug()
+        U_20.debug()
+        U_01.debug()
+        U_11.debug()
+        U_21.debug()
+    if (group in ['ALL', 'Chimu']):
+        Chimu_00.debug()
+        Chimu_01.debug()
+        Chimu_02.debug()
+        Chimu_10.debug()
+        Chimu_11.debug()
+        Chimu_12.debug()
+        Chimu_20.debug()
+        Chimu_21.debug()
+        Chimu_22.debug()
+        Chimu_30.debug()
+        Chimu_31.debug()
+        Chimu_32.debug()
+
+# --------------------------------------------------------------------------------
+# Output
+# --------------------------------------------------------------------------------
+
+if ALTERNATIVE_LOADS == True:
+    define(F'LOAD_CHIMU_0213_PLUG    LOAD_CHIMU_0213_{PRECSUFFIX}')
+    define(F'LOAD_CHIMU_0312_PLUG    LOAD_CHIMU_0312_{PRECSUFFIX}')
+    define(F'LOAD_CHIMU(x)')
+else:
+    #define(F'LOAD_CHIMU_{PRECSUFFIX}(x)           LOAD_CHIMU_INTERLEAVED_{PRECSUFFIX}(x)')
+    define(F'LOAD_CHIMU(base)               LOAD_CHIMU_INTERLEAVED_{PRECSUFFIX}(base)')
+
+if PREFETCH:
+    define(F'PREFETCH_CHIMU_L1(A)           PREFETCH_CHIMU_L1_INTERNAL_{PRECSUFFIX}(A)')
+    define(F'PREFETCH_GAUGE_L1(A)           PREFETCH_GAUGE_L1_INTERNAL_{PRECSUFFIX}(A)')
+    define(F'PREFETCH_CHIMU_L2(A)           PREFETCH_CHIMU_L2_INTERNAL_{PRECSUFFIX}(A)')
+    define(F'PREFETCH_GAUGE_L2(A)           PREFETCH_GAUGE_L2_INTERNAL_{PRECSUFFIX}(A)')
+    define(F'PF_GAUGE(A)')
+    define(F'PREFETCH_RESULT_L2_STORE(A)    PREFETCH_RESULT_L2_STORE_INTERNAL_{PRECSUFFIX}(A)')
+    define(F'PREFETCH_RESULT_L1_STORE(A)    PREFETCH_RESULT_L1_STORE_INTERNAL_{PRECSUFFIX}(A)')
+    define(F'PREFETCH1_CHIMU(A)             PREFETCH_CHIMU_L1(A)')
+#    define(F'PREFETCH1_CHIMU(A)')
+    define(F'PREFETCH_CHIMU(A)              PREFETCH_CHIMU_L1(A)')
+#    define(F'PREFETCH_CHIMU(A)')
+else:
+    define(F'PREFETCH_CHIMU_L1(A)')
+    define(F'PREFETCH_GAUGE_L1(A)')
+    define(F'PREFETCH_CHIMU_L2(A)')
+    define(F'PREFETCH_GAUGE_L2(A)')
+    define(F'PF_GAUGE(A)')
+    define(F'PREFETCH1_CHIMU(A)')
+    define(F'PREFETCH_CHIMU(A)')
+    define(F'PREFETCH_RESULT_L2_STORE(A)')
+
+# standard defines
+define(F'LOCK_GAUGE(A)')
+define(F'UNLOCK_GAUGE(A)')
+define(F'MASK_REGS                      DECLARATIONS_{PRECSUFFIX}')
+define(F'SAVE_RESULT(A,B)               RESULT_{PRECSUFFIX}(A); PREFETCH_RESULT_L2_STORE(B)')
+define(F'MULT_2SPIN_1(Dir)              MULT_2SPIN_1_{PRECSUFFIX}(Dir)')
+define(F'MULT_2SPIN_2                   MULT_2SPIN_2_{PRECSUFFIX}')
+define(F'LOAD_CHI(base)                 LOAD_CHI_{PRECSUFFIX}(base)')
+# don't need zero psi, everything is done in recons
+#define(F'ZERO_PSI                       ZERO_PSI_{PRECSUFFIX}')
+define(F'ADD_RESULT(base,basep)         LOAD_CHIMU(base); ADD_RESULT_INTERNAL_{PRECSUFFIX}; RESULT_{PRECSUFFIX}(base)')
+# loads projections
+define(F'XP_PROJ                        XP_PROJ_{PRECSUFFIX}')
+define(F'YP_PROJ                        YP_PROJ_{PRECSUFFIX}')
+define(F'ZP_PROJ                        ZP_PROJ_{PRECSUFFIX}')
+define(F'TP_PROJ                        TP_PROJ_{PRECSUFFIX}')
+define(F'XM_PROJ                        XM_PROJ_{PRECSUFFIX}')
+define(F'YM_PROJ                        YM_PROJ_{PRECSUFFIX}')
+define(F'ZM_PROJ                        ZM_PROJ_{PRECSUFFIX}')
+define(F'TM_PROJ                        TM_PROJ_{PRECSUFFIX}')
+# recons
+define(F'XP_RECON                       XP_RECON_{PRECSUFFIX}')
+define(F'XM_RECON                       XM_RECON_{PRECSUFFIX}')
+define(F'XM_RECON_ACCUM                 XM_RECON_ACCUM_{PRECSUFFIX}')
+define(F'YM_RECON_ACCUM                 YM_RECON_ACCUM_{PRECSUFFIX}')
+define(F'ZM_RECON_ACCUM                 ZM_RECON_ACCUM_{PRECSUFFIX}')
+define(F'TM_RECON_ACCUM                 TM_RECON_ACCUM_{PRECSUFFIX}')
+define(F'XP_RECON_ACCUM                 XP_RECON_ACCUM_{PRECSUFFIX}')
+define(F'YP_RECON_ACCUM                 YP_RECON_ACCUM_{PRECSUFFIX}')
+define(F'ZP_RECON_ACCUM                 ZP_RECON_ACCUM_{PRECSUFFIX}')
+define(F'TP_RECON_ACCUM                 TP_RECON_ACCUM_{PRECSUFFIX}')
+# new permutes
+define(F'PERMUTE_DIR0                   0')
+define(F'PERMUTE_DIR1                   1')
+define(F'PERMUTE_DIR2                   2')
+define(F'PERMUTE_DIR3                   3')
+define(F'PERMUTE                        PERMUTE_{PRECSUFFIX};')
+# load table
+#define(F'MAYBEPERM(A,perm)              if (perm) {{ A ; }}')
+if PRECISION == 'double':
+    define(F'LOAD_TABLE(Dir)                if (Dir == 0) {{ LOAD_TABLE0; }} else if (Dir == 1) {{ LOAD_TABLE1; }} else if (Dir == 2) {{ LOAD_TABLE2; }}')
+    define(F'MAYBEPERM(Dir,perm)            if (Dir != 3) {{ if (perm) {{ PERMUTE; }} }}')
+else:
+    define(F'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(F'MAYBEPERM(A,perm)              if (perm) {{ PERMUTE; }}')
+
+
+
+write('// DECLARATIONS')
+definemultiline(F'DECLARATIONS_{PRECSUFFIX}')
+# debugging register
+if d['debug'] == True:
+    write('    Simd debugreg; \\')
+# perm tables
+if PRECISION == 'double':
+    write('    const uint64_t lut[4][8] = { \\')
+    write('        {4, 5, 6, 7, 0, 1, 2, 3}, \\')  #0 = swap register halves
+    write('        {2, 3, 0, 1, 6, 7, 4, 5}, \\')  #1 = swap halves of halves
+    write('        {1, 0, 3, 2, 5, 4, 7, 6}, \\')  #2 = swap re/im
+    write('        {0, 1, 2, 4, 5, 6, 7, 8} };\\')  #3 = identity
+else:
+    write('    const uint32_t lut[4][16] = { \\')
+    write('        {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7}, \\')  #0 = swap register halves
+    write('        {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11}, \\')  #1 = swap halves of halves
+    write('        {2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13}, \\')  #2 = swap halves of halves of halves
+    write('        {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14} }; \\') #3 = swap re/im
+
+#newline(target='A')
+result_00.declare()
+result_01.declare()
+result_02.declare()
+result_10.declare()
+result_11.declare()
+result_12.declare()
+result_20.declare()
+result_21.declare()
+result_22.declare()
+result_30.declare()
+result_31.declare()
+result_32.declare()     # 12
+Chi_00.declare()
+Chi_01.declare()
+Chi_02.declare()
+Chi_10.declare()
+Chi_11.declare()
+Chi_12.declare()        # 6
+UChi_00.declare()
+UChi_01.declare()
+UChi_02.declare()
+UChi_10.declare()
+UChi_11.declare()
+UChi_12.declare()       # 6
+U_00.declare()
+U_10.declare()
+U_20.declare()
+U_01.declare()
+U_11.declare()
+U_21.declare()          # 6   -> 30 regs
+
+# all predications true
+pg1.declare()
+if PRECISION == 'double':
+    pg1.movestr('svptrue_b64()')
+else:
+    pg1.movestr('svptrue_b32()')
+
+# tables
+if PRECISION == 'double':
+    write('    svuint64_t table0; \\', target='I')   #      -> 31 regs
+else:
+    write('    svuint32_t table0; \\', target='I')   #      -> 31 regs
+
+zero0.declare()
+
+# zero register
+asmopen()
+zero0.zero(zeroreg=True)
+asmclose()
+newline()
+
+define('Chimu_00 Chi_00', target='I')
+define('Chimu_01 Chi_01', target='I')
+define('Chimu_02 Chi_02', target='I')
+define('Chimu_10 Chi_10', target='I')
+define('Chimu_11 Chi_11', target='I')
+define('Chimu_12 Chi_12', target='I')
+if ALTERNATIVE_REGISTER_MAPPING == False:
+    define('Chimu_20 UChi_00', target='I')
+    define('Chimu_21 UChi_01', target='I')
+    define('Chimu_22 UChi_02', target='I')
+    define('Chimu_30 UChi_10', target='I')
+    define('Chimu_31 UChi_11', target='I')
+    define('Chimu_32 UChi_12', target='I')
+else: # wilson4.h
+    define('Chimu_20 U_00', target='I')
+    define('Chimu_21 U_10', target='I')
+    define('Chimu_22 U_20', target='I')
+    define('Chimu_30 U_01', target='I')
+    define('Chimu_31 U_11', target='I')
+    define('Chimu_32 U_21', target='I')
+newline()
+
+
+d['cycles_RESULT'] += 12
+write('// RESULT')
+definemultiline(F'RESULT_{PRECSUFFIX}(base)')
+if ASM_STORE:
+    curlyopen()
+    #write('    SiteSpinor & ref(out[ss]); \\')
+    asmopen()
+    #pg1.loadpredication()
+    #store_base_ptr("&ref[0][0]")
+    #store_base_ptr(F"&ref[{STORE_BASE_PTR_COLOR_OFFSET}][0]")
+    store_base_ptr(F"base + {STORE_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='I')
+    store_base_ptr(F"base + {STORE_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='A')
+    result_00.store("ref[0][0]")
+    result_01.store("ref[0][1]")
+    result_02.store("ref[0][2]")
+    result_10.store("ref[1][0]")
+    result_11.store("ref[1][1]")
+    result_12.store("ref[1][2]")
+    result_20.store("ref[2][0]")
+    result_21.store("ref[2][1]")
+    result_22.store("ref[2][2]")
+    result_30.store("ref[3][0]")
+    result_31.store("ref[3][1]")
+    result_32.store("ref[3][2]")
+    asmclose()
+    debugall('RESULT', group='result')
+    curlyclose()
+newline()
+
+# prefetch spinors from memory into L2 cache
+d['factor'] = 0
+d['cycles_PREFETCH_L2'] += 0 * d['factor']
+write('// PREFETCH_CHIMU_L2 (prefetch to L2)')
+definemultiline(F'PREFETCH_CHIMU_L2_INTERNAL_{PRECSUFFIX}(base)')
+curlyopen()
+fetch_base_ptr(F"base")
+asmopen()
+#pg1.loadpredication()
+#fetch_base_ptr(F"&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]")
+fetch_base_ptr(F"base", target='A')
+prefetch_L2(F"base", 0)
+prefetch_L2(F"base", 1)
+prefetch_L2(F"base", 2)
+asmclose()
+curlyclose()
+newline()
+
+# prefetch spinors from memory into L1 cache
+d['factor'] = 0
+d['cycles_PREFETCH_L1'] += 0 * d['factor']
+write('// PREFETCH_CHIMU_L1 (prefetch to L1)')
+definemultiline(F'PREFETCH_CHIMU_L1_INTERNAL_{PRECSUFFIX}(base)')
+curlyopen()
+fetch_base_ptr(F"base")
+asmopen()
+#pg1.loadpredication()
+fetch_base_ptr(F"base", target='A')
+prefetch_L1(F"base", 0)
+prefetch_L1(F"base", 1)
+prefetch_L1(F"base", 2)
+asmclose()
+curlyclose()
+newline()
+
+# prefetch gauge from memory into L2 cache
+d['factor'] = 0
+d['cycles_PREFETCH_L2'] += 0 * d['factor']
+write('// PREFETCH_GAUGE_L2 (prefetch to L2)')
+definemultiline(F'PREFETCH_GAUGE_L2_INTERNAL_{PRECSUFFIX}(A)')
+curlyopen()
+if GRIDBENCH:   # referencing differs in Grid and GridBench
+    write('    const auto & ref(U[sUn][A]); uint64_t baseU = (uint64_t)&ref + 3 * 3 * 64; \\')
+else:
+    write('    const auto & ref(U[sUn](A)); uint64_t baseU = (uint64_t)&ref + 3 * 3 * 64; \\')
+asmopen()
+#pg1.loadpredication()
+#fetch_base_ptr(F"&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]")
+fetch_base_ptr(F"baseU", target='A')
+prefetch_L2(F"baseU", -1)
+prefetch_L2(F"baseU", 0)
+prefetch_L2(F"baseU", 1)
+prefetch_L2(F"baseU", 2)
+prefetch_L2(F"baseU", 3)
+prefetch_L2(F"baseU", 4)
+prefetch_L2(F"baseU", 5)
+prefetch_L2(F"baseU", 6)
+prefetch_L2(F"baseU", 7)
+#prefetch_L2(F"baseU", 8)
+asmclose()
+curlyclose()
+newline()
+
+# prefetch gauge from memory into L1 cache
+d['factor'] = 0
+d['cycles_PREFETCH_L1'] += 0 * d['factor']
+write('// PREFETCH_GAUGE_L1 (prefetch to L1)')
+definemultiline(F'PREFETCH_GAUGE_L1_INTERNAL_{PRECSUFFIX}(A)')
+curlyopen()
+if GRIDBENCH:   # referencing differs in Grid and GridBench
+    write('    const auto & ref(U[sU][A]); uint64_t baseU = (uint64_t)&ref; \\')
+else:
+    write('    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \\')
+asmopen()
+#pg1.loadpredication()
+#fetch_base_ptr(F"&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]")
+fetch_base_ptr(F"baseU", target='A')
+prefetch_L1(F"baseU", 0)
+prefetch_L1(F"baseU", 1)
+prefetch_L1(F"baseU", 2)
+asmclose()
+curlyclose()
+newline()
+
+d['factor'] = 0
+write('// LOAD_CHI')
+definemultiline(F'LOAD_CHI_{PRECSUFFIX}(base)')
+if ASM_LOAD_CHIMU:
+    curlyopen()
+    #write('    const SiteSpinor & ref(in[offset]); \\')
+    asmopen()
+    #fetch_base_ptr(F"base + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='I')
+    #fetch_base_ptr(F"base + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='A')
+    fetch_base_ptr(F"base", target='I')
+    fetch_base_ptr(F"base", target='A')
+
+    Chi_00.load("ref[0][0]", offset=0)
+    Chi_01.load("ref[0][1]", offset=0)
+    Chi_02.load("ref[0][2]", offset=0)
+    Chi_10.load("ref[1][0]", offset=0)
+    Chi_11.load("ref[1][1]", offset=0)
+    Chi_12.load("ref[1][2]", offset=0)
+    asmclose()
+    debugall('LOAD_CHI', group='Chi')
+    curlyclose()
+newline()
+
+
+
+d['factor'] = 8
+# 12 loads = 12 issues, load latency = 8+1 cycles
+# (not perfectly clear to me from docs)
+d['cycles_LOAD_CHIMU'] += 11 * d['factor']
+write('// LOAD_CHIMU')
+definemultiline(F'LOAD_CHIMU_INTERLEAVED_{PRECSUFFIX}(base)')
+if ASM_LOAD_CHIMU:
+    curlyopen()
+    #write('    const SiteSpinor & ref(in[offset]); \\')
+    asmopen()
+    pg1.loadpredication()
+    #fetch_base_ptr("&ref[0][0]")
+    #fetch_base_ptr(F"&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]")
+    fetch_base_ptr(F"base + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='I')
+    fetch_base_ptr(F"base + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='A')
+    # Chimu_00.load("ref[0][0]")
+    # Chimu_01.load("ref[0][1]")
+    # Chimu_02.load("ref[0][2]")
+    # Chimu_10.load("ref[1][0]")
+    # Chimu_11.load("ref[1][1]")
+    # Chimu_12.load("ref[1][2]")
+    # Chimu_20.load("ref[2][0]")
+    # Chimu_21.load("ref[2][1]")
+    # Chimu_22.load("ref[2][2]")
+    # Chimu_30.load("ref[3][0]")
+    # Chimu_31.load("ref[3][1]")
+    # Chimu_32.load("ref[3][2]")
+
+    Chimu_00.load("ref[0][0]")  # minimum penalty for all directions
+    Chimu_30.load("ref[3][0]")
+    Chimu_10.load("ref[1][0]")
+    Chimu_20.load("ref[2][0]")
+
+    Chimu_01.load("ref[0][1]")
+    Chimu_31.load("ref[3][1]")
+    Chimu_11.load("ref[1][1]")
+    Chimu_21.load("ref[2][1]")
+
+    Chimu_02.load("ref[0][2]")
+    Chimu_32.load("ref[3][2]")
+    Chimu_12.load("ref[1][2]")
+    Chimu_22.load("ref[2][2]")
+    asmclose()
+    debugall('LOAD_CHIMU', group='Chimu')
+    curlyclose()
+newline()
+
+# alternative load chimu: dirac order 0213
+# placed into asm (...)
+d['factor'] = 0
+d['cycles_LOAD_CHIMU'] += 11 * d['factor']
+write('// LOAD_CHIMU_0213')
+definemultiline(F'LOAD_CHIMU_0213_{PRECSUFFIX}')
+if ASM_LOAD_CHIMU:
+    curlyopen()
+    write('    const SiteSpinor & ref(in[offset]); \\')
+    asmopen()
+    pg1.loadpredication()
+    fetch_base_ptr(F"&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]")
+    Chimu_00.load("ref[0][0]")  # reordered
+    Chimu_20.load("ref[2][0]")
+
+    Chimu_01.load("ref[0][1]")
+    Chimu_21.load("ref[2][1]")
+
+    Chimu_02.load("ref[0][2]")
+    Chimu_22.load("ref[2][2]")
+
+    Chimu_10.load("ref[1][0]")
+    Chimu_30.load("ref[3][0]")
+
+    Chimu_11.load("ref[1][1]")
+    Chimu_31.load("ref[3][1]")
+
+    Chimu_12.load("ref[1][2]")
+    Chimu_32.load("ref[3][2]")
+    asmclose()
+    debugall('LOAD_CHIMU_0213', group='Chimu')
+    curlyclose()
+newline()
+
+# alternative load chimu: dirac order 0312
+# placed into asm (...)
+d['factor'] = 0
+d['cycles_LOAD_CHIMU'] += 11 * d['factor']
+write('// LOAD_CHIMU_0312')
+definemultiline(F'LOAD_CHIMU_0312_{PRECSUFFIX}')
+if ASM_LOAD_CHIMU:
+    curlyopen()
+    write('    const SiteSpinor & ref(in[offset]); \\')
+    asmopen()
+    pg1.loadpredication()
+    fetch_base_ptr(F"&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]")
+    Chimu_00.load("ref[0][0]")  # reordered
+    Chimu_30.load("ref[3][0]")
+
+    Chimu_01.load("ref[0][1]")
+    Chimu_31.load("ref[3][1]")
+
+    Chimu_02.load("ref[0][2]")
+    Chimu_32.load("ref[3][2]")
+
+    Chimu_10.load("ref[1][0]")
+    Chimu_20.load("ref[2][0]")
+
+    Chimu_11.load("ref[1][1]")
+    Chimu_21.load("ref[2][1]")
+
+    Chimu_12.load("ref[1][2]")
+    Chimu_22.load("ref[2][2]")
+    asmclose()
+    debugall('LOAD_CHIMU_0312', group='Chimu')
+    curlyclose()
+newline()
+
+d['factor'] = 2
+d['cycles_PERM'] += 1 * d['factor']
+write('// LOAD_TABLE0')
+definemultiline(F'LOAD_TABLE0')
+asmopen()
+table0.loadtable(0)
+asmclose()
+newline()
+
+d['factor'] = 2
+d['cycles_PERM'] += 1 * d['factor']
+write('// LOAD_TABLE1')
+definemultiline(F'LOAD_TABLE1')
+asmopen()
+table0.loadtable(1)
+asmclose()
+newline()
+
+d['factor'] = 2
+d['cycles_PERM'] += 1 * d['factor']
+write('// LOAD_TABLE2')
+definemultiline(F'LOAD_TABLE2')
+asmopen()
+table0.loadtable(2)
+asmclose()
+newline()
+
+d['factor'] = 0
+d['cycles_PERM'] += 1 * d['factor']
+write('// LOAD_TABLE3')
+definemultiline(F'LOAD_TABLE3')
+asmopen()
+table0.loadtable(3)
+asmclose()
+newline()
+
+d['factor'] = 2     # factor is 2
+d['cycles_PERM'] += 6 * d['factor']
+write('// PERMUTE')
+definemultiline(F'PERMUTE_{PRECSUFFIX}')
+debugall('PERM PRE', group='Chi')
+asmopen()
+#table0.loadtable(2)
+Chi_00.permute(2, table0)
+Chi_01.permute(2, table0)
+Chi_02.permute(2, table0)
+Chi_10.permute(2, table0)
+Chi_11.permute(2, table0)
+Chi_12.permute(2, table0)
+asmclose()
+debugall('PERM POST', group='Chi')
+newline()
+
+write('// LOAD_GAUGE')
+definemultiline(F'LOAD_GAUGE')
+if GRIDBENCH:   # referencing differs in Grid and GridBench
+    write('    const auto & ref(U[sU][A]); uint64_t baseU = (uint64_t)&ref; \\')
+else:
+    write('    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \\')
+curlyopen()
+asmopen()
+pg1.loadpredication()
+fetch_base_ptr(F"baseU + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='I')
+if ASM_LOAD_GAUGE:
+    fetch_base_ptr(F"baseU + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='A')
+    U_00.load("ref[0][0]")
+    U_10.load("ref[1][0]")
+    U_20.load("ref[2][0]")
+    U_01.load("ref[0][1]")
+    U_11.load("ref[1][1]")
+    U_21.load("ref[2][1]")
+asmclose()
+curlyclose()
+newline()
+
+d['factor'] = 8    # MULT_2SPIN executes 1 time per direction = 8 times total
+# assume all U loads are hidden
+# FCMLA issue latency = 2 cycles
+# measurement: latency = 16 cycles if FULLY pipelined !?
+# spec says 6+6+9 cycles
+# 6 rounds of FCMLA, each with 6 FCMLA -> 21 - 6*2 = 9
+d['cycles_MULT_2SPIN'] += 6 * 21 * d['factor']
+write('// MULT_2SPIN')
+definemultiline(F'MULT_2SPIN_1_{PRECSUFFIX}(A)')
+curlyopen()
+#write('    const auto & ref(U[sU][A]); \\')
+if GRIDBENCH:   # referencing differs in Grid and GridBench
+    write('    const auto & ref(U[sU][A]); uint64_t baseU = (uint64_t)&ref; \\')
+else:
+    write('    const auto & ref(U[sU](A)); uint64_t baseU = (uint64_t)&ref; \\')
+asmopen()
+#pg1.loadpredication()
+#fetch_base_ptr("&ref[0][0]")
+fetch_base_ptr(F"baseU + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='I')
+fetch_base_ptr(F"baseU + {FETCH_BASE_PTR_COLOR_OFFSET} * 3 * 64", target='A')
+#fetch_base_ptr(F"(uint64_t)&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]", target='I')
+#fetch_base_ptr(F"(uint64_t)&ref[{FETCH_BASE_PTR_COLOR_OFFSET}][0]", target='A')
+#fetch_base_ptr(F"&ref[0][{FETCH_BASE_PTR_COLOR_OFFSET}]")
+if ASM_LOAD_GAUGE:
+    U_00.load("ref[0][0]")
+    U_10.load("ref[1][0]")
+    U_20.load("ref[2][0]")
+    U_01.load("ref[0][1]")
+    U_11.load("ref[1][1]")
+    U_21.load("ref[2][1]")
+
+if MOVPRFX == False:
+    UChi_00.zero()  # implementation specific
+    UChi_10.zero()
+    UChi_01.zero()
+    UChi_11.zero()
+    UChi_02.zero()
+    UChi_12.zero()
+
+    # round 1
+    UChi_00.mul0(U_00, Chi_00) # FCMLA latency is 6+6+9 cycles
+    UChi_10.mul0(U_00, Chi_10)
+    UChi_01.mul0(U_10, Chi_00)
+    UChi_11.mul0(U_10, Chi_10)
+    UChi_02.mul0(U_20, Chi_00)
+    UChi_12.mul0(U_20, Chi_10)
+else:
+    # round 1
+    UChi_00.mul0(zero0, U_00, Chi_00, constructive=True) # FCMLA latency is 6+6+9 cycles
+    UChi_10.mul0(zero0, U_00, Chi_10, constructive=True)
+    UChi_01.mul0(zero0, U_10, Chi_00, constructive=True)
+    UChi_11.mul0(zero0, U_10, Chi_10, constructive=True)
+    UChi_02.mul0(zero0, U_20, Chi_00, constructive=True)
+    UChi_12.mul0(zero0, U_20, Chi_10, constructive=True)
+
+# round 2
+UChi_00.mul1(U_00, Chi_00)
+UChi_10.mul1(U_00, Chi_10)
+UChi_01.mul1(U_10, Chi_00)
+UChi_11.mul1(U_10, Chi_10)
+UChi_02.mul1(U_20, Chi_00)
+UChi_12.mul1(U_20, Chi_10)  # Chi_00 and Chi_10 available from here
+
+if ASM_LOAD_GAUGE:
+    U_00.load("ref[0][2]")      # U_00, U_10, U_20 overloaded
+    U_10.load("ref[1][2]")      # early load
+    U_20.load("ref[2][2]")      # A -->
+asmclose()
+debugall('MULT_2SPIN_1', group='UChi')
+curlyclose()
+newline()
+
+write('// MULT_2SPIN_BACKEND')
+definemultiline(F'MULT_2SPIN_2_{PRECSUFFIX}')
+curlyopen()
+asmopen()
+# round 3
+UChi_00.mac0(U_01, Chi_01)  # armclang separates fcmla(..., 0) and
+UChi_10.mac0(U_01, Chi_11)  #                    fcmla(..., 90)
+UChi_01.mac0(U_11, Chi_01)  # autonomously using intrinsics
+UChi_11.mac0(U_11, Chi_11)
+UChi_02.mac0(U_21, Chi_01)
+UChi_12.mac0(U_21, Chi_11)
+# round 4
+UChi_00.mac1(U_01, Chi_01)
+UChi_10.mac1(U_01, Chi_11)
+UChi_01.mac1(U_11, Chi_01)
+UChi_11.mac1(U_11, Chi_11)
+UChi_02.mac1(U_21, Chi_01)
+UChi_12.mac1(U_21, Chi_11)
+# round 5
+UChi_00.mac0(U_00, Chi_02)  # <-- A
+UChi_10.mac0(U_00, Chi_12)
+UChi_01.mac0(U_10, Chi_02)
+UChi_11.mac0(U_10, Chi_12)
+UChi_02.mac0(U_20, Chi_02)
+UChi_12.mac0(U_20, Chi_12)
+# round 6
+UChi_00.mac1(U_00, Chi_02)
+UChi_10.mac1(U_00, Chi_12)
+UChi_01.mac1(U_10, Chi_02)
+UChi_11.mac1(U_10, Chi_12)
+UChi_02.mac1(U_20, Chi_02)
+UChi_12.mac1(U_20, Chi_12)
+asmclose()
+debugall('MULT_2SPIN_2', group='UChi')
+curlyclose()
+newline()
+
+
+#//      hspin(0)=fspin(0)+timesI(fspin(3));
+#//      hspin(1)=fspin(1)+timesI(fspin(2));
+d['factor'] = 1
+# FCADD issue latency = 1, latency is 6+9
+d['cycles_PROJ'] += 15 * d['factor']
+write('// XP_PROJ')
+definemultiline(F'XP_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0312_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.addTimesI(Chimu_00, Chimu_30)
+Chi_01.addTimesI(Chimu_01, Chimu_31)
+Chi_02.addTimesI(Chimu_02, Chimu_32)
+Chi_10.addTimesI(Chimu_10, Chimu_20)
+Chi_11.addTimesI(Chimu_11, Chimu_21)
+Chi_12.addTimesI(Chimu_12, Chimu_22)
+asmclose()
+debugall('XP_PROJ', group='Chi')
+curlyclose()
+newline()
+
+#//      fspin(0)=hspin(0);
+#//      fspin(1)=hspin(1);
+#//      fspin(2)=timesMinusI(hspin(1));
+#//      fspin(3)=timesMinusI(hspin(0));
+# does not occur in GridBench
+d['factor'] = 0
+d['cycles_RECON'] += 15 * d['factor']
+write('// XP_RECON')
+definemultiline(F'XP_RECON_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+if MOVPRFX == False:
+    result_20.zero()
+    result_21.zero()
+    result_22.zero()
+    result_30.zero()
+    result_31.zero()
+    result_32.zero()
+
+    result_20.subTimesI(UChi_10)
+    result_21.subTimesI(UChi_11)
+    result_22.subTimesI(UChi_12)
+    result_30.subTimesI(UChi_00)
+    result_31.subTimesI(UChi_01)
+    result_32.subTimesI(UChi_02)
+else:
+    result_20.subTimesI(zero0, UChi_10, constructive=True)
+    result_21.subTimesI(zero0, UChi_11, constructive=True)
+    result_22.subTimesI(zero0, UChi_12, constructive=True)
+    result_30.subTimesI(zero0, UChi_00, constructive=True)
+    result_31.subTimesI(zero0, UChi_01, constructive=True)
+    result_32.subTimesI(zero0, UChi_02, constructive=True)
+
+result_00.move(UChi_00) # don't reorder !
+result_01.move(UChi_01)
+result_02.move(UChi_02)
+result_10.move(UChi_10)
+result_11.move(UChi_11)
+result_12.move(UChi_12)
+
+# result_00.add(UChi_00)   # faster than move?
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+asmclose()
+debugall('XP_RECON', group='result')
+newline()
+
+
+d['factor'] = 1
+# FCADD issue latency = 1, latency is 6+9
+d['cycles_RECON'] += 15 * d['factor']
+write('// XP_RECON_ACCUM')
+definemultiline(F'XP_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_20.subTimesI(UChi_10)
+# result_21.subTimesI(UChi_11)
+# result_22.subTimesI(UChi_12)
+# result_30.subTimesI(UChi_00)
+# result_31.subTimesI(UChi_01)
+# result_32.subTimesI(UChi_02)
+#
+# result_00.add(UChi_00) # reordered
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+
+result_30.subTimesI(UChi_00)    # reordered
+result_00.add(UChi_00)
+
+result_31.subTimesI(UChi_01)
+result_01.add(UChi_01)
+
+result_32.subTimesI(UChi_02)
+result_02.add(UChi_02)
+
+result_20.subTimesI(UChi_10)
+result_10.add(UChi_10)
+
+result_21.subTimesI(UChi_11)
+result_11.add(UChi_11)
+
+result_22.subTimesI(UChi_12)
+result_12.add(UChi_12)
+asmclose()
+debugall('XP_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 1
+# add/sub issue latency = 1, latency is 9
+d['cycles_PROJ'] += 9 * d['factor']
+write('// YP_PROJ')
+definemultiline(F'YP_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0312_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.sub(Chimu_00, Chimu_30)
+Chi_01.sub(Chimu_01, Chimu_31)
+Chi_02.sub(Chimu_02, Chimu_32)
+Chi_10.add(Chimu_10, Chimu_20)
+Chi_11.add(Chimu_11, Chimu_21)
+Chi_12.add(Chimu_12, Chimu_22)
+asmclose()
+debugall('YP_PROJ', group='Chi')
+curlyclose()
+newline()
+
+d['factor'] = 1
+# FCADD issue latency = 1, latency is 6+9
+d['cycles_PROJ'] += 15 * d['factor']
+write('// ZP_PROJ')
+definemultiline(F'ZP_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0213_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.addTimesI(Chimu_00, Chimu_20)
+Chi_01.addTimesI(Chimu_01, Chimu_21)
+Chi_02.addTimesI(Chimu_02, Chimu_22)
+Chi_10.subTimesI(Chimu_10, Chimu_30)
+Chi_11.subTimesI(Chimu_11, Chimu_31)
+Chi_12.subTimesI(Chimu_12, Chimu_32)
+asmclose()
+debugall('ZP_PROJ', group='Chi')
+curlyclose()
+newline()
+
+d['factor'] = 1
+# add/sub issue latency = 1, latency is 9
+d['cycles_PROJ'] += 9 * d['factor']
+write('// TP_PROJ')
+definemultiline(F'TP_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0213_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.add(Chimu_00, Chimu_20)
+Chi_01.add(Chimu_01, Chimu_21)
+Chi_02.add(Chimu_02, Chimu_22)
+Chi_10.add(Chimu_10, Chimu_30)
+Chi_11.add(Chimu_11, Chimu_31)
+Chi_12.add(Chimu_12, Chimu_32)
+asmclose()
+debugall('TP_PROJ', group='Chi')
+curlyclose()
+newline()
+
+#//      hspin(0)=fspin(0)-timesI(fspin(3));
+#//      hspin(1)=fspin(1)-timesI(fspin(2));
+
+d['factor'] = 1
+# FCADD issue latency = 1, latency is 6+9
+d['cycles_PROJ'] += 15 * d['factor']
+write('// XM_PROJ')
+definemultiline(F'XM_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0312_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.subTimesI(Chimu_00, Chimu_30)
+Chi_01.subTimesI(Chimu_01, Chimu_31)
+Chi_02.subTimesI(Chimu_02, Chimu_32)
+Chi_10.subTimesI(Chimu_10, Chimu_20)
+Chi_11.subTimesI(Chimu_11, Chimu_21)
+Chi_12.subTimesI(Chimu_12, Chimu_22)
+asmclose()
+debugall('XM_PROJ sub', group='Chi')
+curlyclose()
+newline()
+
+d['factor'] = 1
+d['cycles_RECON'] += 15 * d['factor']
+write('// XM_RECON')
+definemultiline(F'XM_RECON_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+
+# only necessary if not zeroed before
+if MOVPRFX == False:
+    result_20.zero()
+    result_21.zero()
+    result_22.zero()
+    result_30.zero()
+    result_31.zero()
+    result_32.zero()
+
+    result_20.addTimesI(UChi_10) # <--
+    result_21.addTimesI(UChi_11)
+    result_22.addTimesI(UChi_12)
+    result_30.addTimesI(UChi_00)
+    result_31.addTimesI(UChi_01)
+    result_32.addTimesI(UChi_02)
+else:
+    result_20.addTimesI(zero0, UChi_10, constructive=True) # <--
+    result_21.addTimesI(zero0, UChi_11, constructive=True)
+    result_22.addTimesI(zero0, UChi_12, constructive=True)
+    result_30.addTimesI(zero0, UChi_00, constructive=True)
+    result_31.addTimesI(zero0, UChi_01, constructive=True)
+    result_32.addTimesI(zero0, UChi_02, constructive=True)
+
+result_00.move(UChi_00)
+result_01.move(UChi_01)
+result_02.move(UChi_02)
+result_10.move(UChi_10)
+result_11.move(UChi_11)
+result_12.move(UChi_12)
+asmclose()
+debugall('XM_RECON result', group='result')
+newline()
+
+d['factor'] = 1
+# add/sub issue latency = 1, latency is 9
+d['cycles_PROJ'] += 9 * d['factor']
+write('// YM_PROJ')
+definemultiline(F'YM_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0312_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.add(Chimu_00, Chimu_30)
+Chi_01.add(Chimu_01, Chimu_31)
+Chi_02.add(Chimu_02, Chimu_32)
+Chi_10.sub(Chimu_10, Chimu_20)
+Chi_11.sub(Chimu_11, Chimu_21)
+Chi_12.sub(Chimu_12, Chimu_22)
+asmclose()
+debugall('YM_PROJ', group='Chi')
+curlyclose()
+newline()
+
+d['factor'] = 1
+# FCADD issue latency = 1, latency is 6+9
+d['cycles_PROJ'] += 15 * d['factor']
+write('// ZM_PROJ')
+definemultiline(F'ZM_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0213_PLUG \\')
+curlyopen()
+asmopen()
+#pg1.loadpredication()
+Chi_00.subTimesI(Chimu_00, Chimu_20)
+Chi_01.subTimesI(Chimu_01, Chimu_21)
+Chi_02.subTimesI(Chimu_02, Chimu_22)
+Chi_10.addTimesI(Chimu_10, Chimu_30)
+Chi_11.addTimesI(Chimu_11, Chimu_31)
+Chi_12.addTimesI(Chimu_12, Chimu_32)
+asmclose()
+debugall('ZM_PROJ', group='Chi')
+curlyclose()
+newline()
+
+d['factor'] = 1
+# add/sub issue latency = 1, latency is 9
+d['cycles_PROJ'] += 9 * d['factor']
+write('// TM_PROJ')
+definemultiline(F'TM_PROJ_{PRECSUFFIX}')
+if ALTERNATIVE_LOADS == True:
+    write('    LOAD_CHIMU_0213_PLUG \\')
+curlyopen()
+asmopen()
+pg1.loadpredication()
+Chi_00.sub(Chimu_00, Chimu_20)
+Chi_01.sub(Chimu_01, Chimu_21)
+Chi_02.sub(Chimu_02, Chimu_22)
+Chi_10.sub(Chimu_10, Chimu_30)
+Chi_11.sub(Chimu_11, Chimu_31)
+Chi_12.sub(Chimu_12, Chimu_32)
+asmclose()
+debugall('TM_PROJ', group='Chi')
+curlyclose()
+newline()
+
+# does not occur in GridBench
+d['factor'] = 0
+# add/sub issue latency = 1, latency is 9
+d['cycles_RECON'] += 15 * d['factor']
+write('// XM_RECON_ACCUM')
+definemultiline(F'XM_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+# result_20.addTimesI(UChi_10)
+# result_21.addTimesI(UChi_11)
+# result_22.addTimesI(UChi_12)
+# result_30.addTimesI(UChi_00)
+# result_31.addTimesI(UChi_01)
+# result_32.addTimesI(UChi_02)
+#
+# # result_00.move(UChi_00)
+# # result_01.move(UChi_01)
+# # result_02.move(UChi_02)
+# # result_10.move(UChi_10)
+# # result_11.move(UChi_11)
+# # result_12.move(UChi_12)
+#
+# # faster than move ?
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+
+result_30.addTimesI(UChi_00)    # reordered
+result_31.addTimesI(UChi_01)
+result_32.addTimesI(UChi_02)
+
+result_20.addTimesI(UChi_10)
+result_21.addTimesI(UChi_11)
+result_22.addTimesI(UChi_12)
+
+result_00.add(UChi_00)
+result_01.add(UChi_01)
+result_02.add(UChi_02)
+result_10.add(UChi_10)
+result_11.add(UChi_11)
+result_12.add(UChi_12)
+asmclose()
+debugall('XM_RECON_ACCUM', group='result')
+newline()
+
+
+
+d['factor'] = 1
+d['cycles_RECON'] += 9 * d['factor']
+write('// YP_RECON_ACCUM')
+definemultiline(F'YP_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+# result_20.add(UChi_10)
+# result_21.add(UChi_11)
+# result_22.add(UChi_12)
+# result_30.sub(UChi_00)
+# result_31.sub(UChi_01)
+# result_32.sub(UChi_02)
+
+result_00.add(UChi_00)  # reordered
+result_30.sub(UChi_00)
+
+result_01.add(UChi_01)
+result_31.sub(UChi_01)
+
+result_02.add(UChi_02)
+result_32.sub(UChi_02)
+
+result_10.add(UChi_10)
+result_20.add(UChi_10)
+
+result_11.add(UChi_11)
+result_21.add(UChi_11)
+
+result_12.add(UChi_12)
+result_22.add(UChi_12)
+asmclose()
+debugall('YP_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 1
+d['cycles_RECON'] += 9 * d['factor']
+write('// YM_RECON_ACCUM')
+definemultiline(F'YM_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+# result_20.sub(UChi_10)
+# result_21.sub(UChi_11)
+# result_22.sub(UChi_12)
+# result_30.add(UChi_00)
+# result_31.add(UChi_01)
+# result_32.add(UChi_02)
+
+result_00.add(UChi_00)  # reordered
+result_30.add(UChi_00)
+
+result_01.add(UChi_01)
+result_31.add(UChi_01)
+
+result_02.add(UChi_02)
+result_32.add(UChi_02)
+
+result_10.add(UChi_10)
+result_20.sub(UChi_10)
+
+result_11.add(UChi_11)
+result_21.sub(UChi_11)
+
+result_12.add(UChi_12)
+result_22.sub(UChi_12)
+asmclose()
+debugall('YM_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 1
+d['cycles_RECON'] += 15 * d['factor']
+write('// ZP_RECON_ACCUM')
+definemultiline(F'ZP_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_20.subTimesI(UChi_00)
+# result_21.subTimesI(UChi_01)
+# result_22.subTimesI(UChi_02)
+# result_30.addTimesI(UChi_10)
+# result_31.addTimesI(UChi_11)
+# result_32.addTimesI(UChi_12)
+#
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+result_20.subTimesI(UChi_00)    # reordered
+result_00.add(UChi_00)
+
+result_21.subTimesI(UChi_01)
+result_01.add(UChi_01)
+
+result_22.subTimesI(UChi_02)
+result_02.add(UChi_02)
+
+result_30.addTimesI(UChi_10)
+result_10.add(UChi_10)
+
+result_31.addTimesI(UChi_11)
+result_11.add(UChi_11)
+
+result_32.addTimesI(UChi_12)
+result_12.add(UChi_12)
+asmclose()
+debugall('ZP_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 1
+d['cycles_RECON'] += 15 * d['factor']
+write('// ZM_RECON_ACCUM')
+definemultiline(F'ZM_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_20.addTimesI(UChi_00)
+# result_21.addTimesI(UChi_01)
+# result_22.addTimesI(UChi_02)
+# result_30.subTimesI(UChi_10)
+# result_31.subTimesI(UChi_11)
+# result_32.subTimesI(UChi_12)
+#
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+result_20.addTimesI(UChi_00)    # reordered
+result_00.add(UChi_00)
+
+result_21.addTimesI(UChi_01)
+result_01.add(UChi_01)
+
+result_22.addTimesI(UChi_02)
+result_02.add(UChi_02)
+
+result_30.subTimesI(UChi_10)
+result_10.add(UChi_10)
+
+result_31.subTimesI(UChi_11)
+result_11.add(UChi_11)
+
+result_32.subTimesI(UChi_12)
+result_12.add(UChi_12)
+asmclose()
+debugall('ZM_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 1
+d['cycles_RECON'] += 9 * d['factor']
+write('// TP_RECON_ACCUM')
+definemultiline(F'TP_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+# result_20.add(UChi_00)
+# result_21.add(UChi_01)
+# result_22.add(UChi_02)
+# result_30.add(UChi_10)
+# result_31.add(UChi_11)
+# result_32.add(UChi_12)
+
+result_00.add(UChi_00)  # reordered
+result_20.add(UChi_00)
+
+result_01.add(UChi_01)
+result_21.add(UChi_01)
+
+result_02.add(UChi_02)
+result_22.add(UChi_02)
+
+result_10.add(UChi_10)
+result_30.add(UChi_10)
+
+result_11.add(UChi_11)
+result_31.add(UChi_11)
+
+result_12.add(UChi_12)
+result_32.add(UChi_12)
+asmclose()
+debugall('TP_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 1
+d['cycles_RECON'] += 9 * d['factor']
+write('// TM_RECON_ACCUM')
+definemultiline(F'TM_RECON_ACCUM_{PRECSUFFIX}')
+asmopen()
+#pg1.loadpredication()
+# result_00.add(UChi_00)
+# result_01.add(UChi_01)
+# result_02.add(UChi_02)
+# result_10.add(UChi_10)
+# result_11.add(UChi_11)
+# result_12.add(UChi_12)
+# result_20.sub(UChi_00)
+# result_21.sub(UChi_01)
+# result_22.sub(UChi_02)
+# result_30.sub(UChi_10)
+# result_31.sub(UChi_11)
+# result_32.sub(UChi_12)
+
+result_00.add(UChi_00)  # reordered
+result_20.sub(UChi_00)
+
+result_01.add(UChi_01)
+result_21.sub(UChi_01)
+
+result_02.add(UChi_02)
+result_22.sub(UChi_02)
+
+result_10.add(UChi_10)
+result_30.sub(UChi_10)
+
+result_11.add(UChi_11)
+result_31.sub(UChi_11)
+
+result_12.add(UChi_12)
+result_32.sub(UChi_12)
+asmclose()
+debugall('TM_RECON_ACCUM', group='result')
+newline()
+
+d['factor'] = 0
+# have 12 instructions
+# picking dual issue versions
+d['cycles_ZERO_PSI'] += 6 * d['factor']
+write('// ZERO_PSI')
+definemultiline(F'ZERO_PSI_{PRECSUFFIX}')
+asmopen()
+pg1.loadpredication()
+result_00.zero()
+result_01.zero()
+result_02.zero()
+result_10.zero()
+result_11.zero()
+result_12.zero()
+result_20.zero()
+result_21.zero()
+result_22.zero()
+result_30.zero()
+result_31.zero()
+result_32.zero()
+asmclose()
+#debugall('ZERO_PSI', group='result')
+newline()
+
+# prefetch store spinors to L2 cache
+d['factor'] = 0
+d['cycles_PREFETCH_L2'] += 0 * d['factor']
+write('// PREFETCH_RESULT_L2_STORE (prefetch store to L2)')
+definemultiline(F'PREFETCH_RESULT_L2_STORE_INTERNAL_{PRECSUFFIX}(base)')
+curlyopen()
+fetch_base_ptr(F"base")
+asmopen()
+fetch_base_ptr(F"base", target='A')
+prefetch_L2_store(F"base", 0)
+prefetch_L2_store(F"base", 1)
+prefetch_L2_store(F"base", 2)
+asmclose()
+curlyclose()
+newline()
+
+# prefetch store spinors to L1 cache
+d['factor'] = 0
+d['cycles_PREFETCH_L1'] += 0 * d['factor']
+write('// PREFETCH_RESULT_L1_STORE (prefetch store to L1)')
+definemultiline(F'PREFETCH_RESULT_L1_STORE_INTERNAL_{PRECSUFFIX}(base)')
+curlyopen()
+fetch_base_ptr(F"base")
+asmopen()
+fetch_base_ptr(F"base", target='A')
+prefetch_L1_store(F"base", 0)
+prefetch_L1_store(F"base", 1)
+prefetch_L1_store(F"base", 2)
+asmclose()
+curlyclose()
+newline()
+
+
+d['factor'] = 0
+write('// ADD_RESULT_INTERNAL')
+definemultiline(F'ADD_RESULT_INTERNAL_{PRECSUFFIX}')
+asmopen()
+result_00.add(Chimu_00)
+result_01.add(Chimu_01)
+result_02.add(Chimu_02)
+result_10.add(Chimu_10)
+result_11.add(Chimu_11)
+result_12.add(Chimu_12)
+result_20.add(Chimu_20)
+result_21.add(Chimu_21)
+result_22.add(Chimu_22)
+result_30.add(Chimu_30)
+result_31.add(Chimu_31)
+result_32.add(Chimu_32)
+asmclose()
+#debugall('ZERO_PSI', group='result')
+newline()
+
+# --------------------------------------------------------------------------------
+
+# C
+f = open('w.h', 'w')
+f.write(d['C'])
+f.close()
+
+# intrin
+f = open('wi.h', 'w')
+f.write(d['I'])
+f.close()
+
+filename = ''
+if PRECISION == 'double':
+    filename = "Fujitsu_A64FX_intrin_double.h"
+else:
+    filename = "Fujitsu_A64FX_intrin_single.h"
+f = open(filename, 'w')
+f.write(LEGAL.format(filename))
+f.write(d['I'])
+f.close()
+
+
+# asm
+f = open('wa.h', 'w')
+f.write(d['A'])
+f.close()
+
+filename = ''
+if PRECISION == 'double':
+    filename = "Fujitsu_A64FX_asm_double.h"
+else:
+    filename = "Fujitsu_A64FX_asm_single.h"
+f = open(filename, 'w')
+f.write(LEGAL.format(filename))
+f.write(d['A'])
+f.close()
+
+
+# arithmetics instruction count, mul/mac = 2 instructions each
+d['acount'] = d['add'] + d['sub'] + \
+    d['mul'] + d['mac'] + d['addTimesI'] + d['subTimesI']
+
+# permutations
+d['permutes'] += 2*d['timesI'] + 1*d['timesMinusI']
+d['neg'] = 1*d['timesI'] + 1*d['timesMinusI']
+
+# instruction count,  mul/mac = 2 instructions each, +/- *i = 3 instructions each
+d['icount'] = d['load'] + d['store'] + d['move'] + d['add'] + d['sub'] + \
+    d['mul'] + d['mac'] + d['permutes'] + d['neg'] + \
+    d['addTimesI'] + d['subTimesI'] + d['zero'] + d['movprfx']
+
+# flops
+d['flops'] = 4*d['mac'] + 3*d['mul'] + d['add'] + d['sub'] + \
+    d['addTimesI'] + d['subTimesI']
+
+
+
+
+
+print('Statistics')
+print('')
+print('Type                     Occurences      Total / Arith instructions')
+print('-------------------------------------------------------------------')
+print('Variables                {:4d}'.format(d['registers']))
+print('')
+print('load                     {:4d}'.format(d['load']))
+print('store                    {:4d}'.format(d['store']))
+print('move                     {:4d}'.format(d['move']))
+print('movprfx                  {:4d}'.format(d['movprfx']))
+print('zero                     {:4d}'.format(d['zero']))
+print('negate                   {:4d}'.format(d['neg']))
+
+
+print('add                      {:4d}              {:0.2f} / {:0.2f}'.\
+    format(d['add'], d['add'] /   d['icount'], d['add'] /   d['acount']))
+print('sub                      {:4d}              {:0.2f} / {:0.2f}'.\
+    format(d['sub'], d['sub'] /   d['icount'], d['sub'] /   d['acount']))
+print('mul                      {:4d}              {:0.2f} / {:0.2f}'.\
+    format(d['mul'], 2*d['mul'] / d['icount'], 2*d['mul'] /   d['acount']))
+print('mac                      {:4d}              {:0.2f} / {:0.2f}'.\
+    format(d['mac'], 2*d['mac'] / d['icount'], 2*d['mac'] /   d['acount']))
+print('addTimesI                {:4d}              {:0.2f} / {:0.2f}'.\
+    format(d['addTimesI'], 2*d['addTimesI'] / d['icount'], 2*d['addTimesI'] / d['acount']))
+print('subTimesI                {:4d}              {:0.2f} / {:0.2f}'.\
+    format(d['subTimesI'], 2*d['subTimesI'] / d['icount'], 2*d['subTimesI'] / d['acount']))
+
+print('timesI                   {:4d}'.format(d['timesI']))
+print('timesMinusI              {:4d}'.format(d['timesMinusI']))
+print('permutes                 {:4d}              {:0.2f}'.\
+    format(d['permutes'], d['permutes'] / d['icount']))
+print('')
+print('flops                    {:4d}'.format(d['flops']))
+print('instruction count        {:4d}'.format(d['icount']))
+print('arith. instruction count {:4d}              {:0.2f}'.\
+    format(d['acount'], d['acount'] / d['icount']))
+
+
+# ---- static pipeline resources consumption ----
+FLA = 0
+FLA += 2 * d['mac'] + 2 * d['mul']
+FLA += 1 * d['addTimesI'] + 1 * d['subTimesI']
+FLA += 1 * d['move']
+FLA += 1 * d['permutes']
+FLA += 1 * d['store']
+FLA += 1 * d['zero']
+
+FLB = 0
+FLB += 1 * d['addTimesI'] + 1 * d['subTimesI']
+
+FLAB = 0
+FLAB += 1 * d['mac'] + 1 * d['mul']
+FLAB += 1 * d['add'] + 1 * d['sub']
+FLAB += 1 * d['neg'] + 1 * d['movprfx']
+#FLAB += 1 * d['zero']
+
+
+FL_slots = 2 * d['icount']
+FL_micro_ops = FLA + FLB + FLAB
+
+print('')
+print('------------------------------------------------------------------')
+print('')
+print('Static FL slot usage')
+print('')
+print('  FLA                      {:4d}'.format(FLA))
+print('  FLB                      {:4d}'.format(FLB))
+print('  FLA/B                    {:4d}'.format(FLAB))
+
+print('')
+print('Static FL slot efficiency')
+print('')
+print('  Total FL slots           {:4d}'.format(FL_slots))
+print('  FL slots occupied        {:4d}'.format(FL_micro_ops))
+print('  FL slot efficiency       {:0.2f}'.format(FL_micro_ops / FL_slots))
+
+cycles_total = d['cycles_ZERO_PSI'] + d['cycles_LOAD_CHIMU'] + \
+    d['cycles_PROJ'] + d['cycles_PERM'] + d['cycles_MULT_2SPIN'] + \
+    d['cycles_RECON'] + d['cycles_RESULT']
+cycles_total_hidden = d['cycles_ZERO_PSI'] + \
+    d['cycles_PROJ'] + d['cycles_MULT_2SPIN'] + \
+    d['cycles_RECON']
+
+# ---- dynamic estimate ----
+
+print('')
+print('Dynamic cycles estimate (incl. latencies)')
+print('')
+print('  ZERO_PSI                 {:4d}'.format(d['cycles_ZERO_PSI']))
+print('  LOAD_CHIMU               {:4d}'.format(d['cycles_LOAD_CHIMU']))
+print('  PROJ                     {:4d}'.format(d['cycles_PROJ']))
+print('  PERM                     {:4d}'.format(d['cycles_PERM']))
+print('  MULT_2SPIN               {:4d}'.format(d['cycles_MULT_2SPIN']))
+print('  RECON                    {:4d}'.format(d['cycles_RECON']))
+print('  STORE                    {:4d}'.format(d['cycles_RESULT']))
+print('')
+print('  Sum                      {:4d}'.format(cycles_total))
+print('')
+print('  Sum*                     {:4d}'.format(cycles_total_hidden))
+print('  Total FL slots*          {:4d}'.format(cycles_total_hidden * 2))
+print('  FL slots occupied*       {:4d}'.format(FL_micro_ops))
+print('  FL slot efficiency*      {:0.2f}'.format(FL_micro_ops / (2*cycles_total_hidden)))
+print('')
+print('  *load/store/PERM hidden')
+
+estimated_cycles = cycles_total_hidden
+# Estimate percent peak DP; dual issue, fma
+pp = 100 * 4 * d['flops'] / (2*2*8*estimated_cycles)
+print('')
+print('Model prediction')
+print('')
+print('  Cycles*                  {:4d}'.format(estimated_cycles))
+print('  Percent peak*            {:4.1f} %'.format(pp))
+
+# estimated RF throughput in GB/s @ 2.2 GHz
+tp10 = (d['load'] + d['store']) * 64 * 2.2 / estimated_cycles
+tp2  = (d['load'] + d['store']) * 64 * 1000.**3 * 2.2 / 1024.**3 / estimated_cycles
+print('')
+print('  Estimated RF throughput* {:4.1f}      GB/s'.\
+    format(tp10))
+print('  Estimated RF throughput* {:4.1f}      GiB/s'.\
+    format(tp2))
+
+# ---- dynamic pipeline resources consumption ----
+
+runtime = measured_cycles  # runtime in cycles
+pp_runtime = 100 * 4 * d['flops'] / (2*2*8*runtime)
+runtime_FL_slots = 2 * runtime
+delta = runtime - estimated_cycles
+
+
+print('')
+print('------------------------------------------------------------------')
+print('')
+print('Dynamic runtime analysis (cycles from measurements)')
+print('')
+print('  Cycles                   {:4d}'.format(runtime))
+print('  Percent peak             {:4.1f} %'.format(pp_runtime))
+print('  Deviation from estimate  {:4d}               {:4.2f} %'.\
+    format(delta, 100. * abs(delta/runtime)))
+print('  Deviation per direction  {:4.1f}'.format(delta/8))
+
+# estimated RF throughput in GB/s @ 2.2 GHz
+tp10_rt = (d['load'] + d['store']) * 64 * 2.2 / runtime
+tp2_rt  = (d['load'] + d['store']) * 64 * 1000.**3 * 2.2 / 1024.**3 / runtime
+print('')
+print('  RF throughput            {:4.1f}      GB/s'.\
+    format(tp10_rt))
+print('  RF throughput            {:4.1f}      GiB/s'.\
+    format(tp2_rt))
+print('')
+print('  Total FL slots           {:4d}'.format(runtime_FL_slots))
+print('  FL slots occupied        {:4d}'.format(FL_micro_ops))
+print('  FL slot efficiency       {:0.2f}'.format(FL_micro_ops / runtime_FL_slots))
+print('')
diff --git a/Grid/stencil/Stencil.h b/Grid/stencil/Stencil.h
index 37b866cb..1e198972 100644
--- a/Grid/stencil/Stencil.h
+++ b/Grid/stencil/Stencil.h
@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-     Grid physics library, www.github.com/paboyle/Grid 
+     Grid physics library, www.github.com/paboyle/Grid
 
      Source file: ./lib/Stencil.h
 
@@ -41,13 +41,13 @@
 // Stencil based code will exchange haloes and use a table lookup for neighbours.
 // This will be done with generality to allow easier efficient implementations.
 // Overlap of comms and compute is enabled by tabulating off-node connected,
-// 
+//
 // Generic services
 // 0) Prebuild neighbour tables
 // 1) Compute sizes of all haloes/comms buffers; allocate them.
 // 2) Gather all faces, and communicate.
 // 3) Loop over result sites, giving nbr index/offnode info for each
-// 
+//
 //////////////////////////////////////////////////////////////////////////////////////////
 
 NAMESPACE_BEGIN(Grid);
@@ -59,15 +59,16 @@ NAMESPACE_BEGIN(Grid);
 void Gather_plane_table_compute (GridBase *grid,int dimension,int plane,int cbmask,
 				 int off,Vector<std::pair<int,int> > & table);
 
-template<class vobj,class cobj,class compressor> 
+template<class vobj,class cobj,class compressor>
 void Gather_plane_simple_table (Vector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so)   __attribute__((noinline));
 
-template<class vobj,class cobj,class compressor> 
+template<class vobj,class cobj,class compressor>
 void Gather_plane_simple_table (Vector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so)
 {
   int num=table.size();
   std::pair<int,int> *table_v = & table[0];
-  auto rhs_v = rhs.View();
+
+  auto rhs_v = rhs.View(AcceleratorRead);
   accelerator_forNB( i,num, vobj::Nsimd(), {
     typedef decltype(coalescedRead(buffer[0])) compressed_t;
     compressed_t   tmp_c;
@@ -75,6 +76,7 @@ void Gather_plane_simple_table (Vector<std::pair<int,int> >& table,const Lattice
     compress.Compress(&tmp_c,0,rhs_v(so+table_v[i].second));
     coalescedWrite(buffer[off+o],tmp_c);
   });
+  rhs_v.ViewClose();
 // Further optimisatoin: i) software prefetch the first element of the next table entry, prefetch the table
 }
 
@@ -92,37 +94,38 @@ void Gather_plane_exchange_table(Vector<std::pair<int,int> >& table,const Lattic
 {
   assert( (table.size()&0x1)==0);
   int num=table.size()/2;
-  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
-  auto rhs_v = rhs.View();
+  auto rhs_v = rhs.View(AcceleratorRead);
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
-  accelerator_forNB(j, num, 1, { 
+  accelerator_forNB(j, num, 1, {
       compress.CompressExchange(p0,p1, &rhs_v[0], j,
 			      so+tp[2*j  ].second,
 			      so+tp[2*j+1].second,
 			      type);
   });
+  rhs_v.ViewClose();
 }
 
-struct StencilEntry { 
-#ifdef GRID_NVCC
-  uint64_t _byte_offset;       // 8 bytes 
-  uint32_t _offset;            // 4 bytes 
+struct StencilEntry {
+#ifdef GRID_CUDA
+  uint64_t _byte_offset;       // 8 bytes
+  uint32_t _offset;            // 4 bytes
 #else
-  uint64_t _byte_offset;       // 8 bytes 
+  uint64_t _byte_offset;       // 8 bytes
   uint64_t _offset;            // 8 bytes (8 ever required?)
 #endif
-  uint8_t _is_local;           // 1 bytes 
+  uint8_t _is_local;           // 1 bytes
   uint8_t _permute;            // 1 bytes
   uint8_t _around_the_world;   // 1 bytes
   uint8_t _pad;   // 1 bytes
 };
-// Could pack to 8 + 4 + 4 = 128 bit and use 
+// Could pack to 8 + 4 + 4 = 128 bit and use
 
 template<class vobj,class cobj,class Parameters>
-class CartesianStencilView {
+class CartesianStencilAccelerator {
  public:
   typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;
 
@@ -130,14 +133,15 @@ class CartesianStencilView {
   ////////////////////////////////////////
   // Basic Grid and stencil info
   ////////////////////////////////////////
-  int                               _checkerboard;
-  int                               _npoints; // Move to template param?
+  int           _checkerboard;
+  int           _npoints; // Move to template param?
+  int           _osites;
   StencilVector _directions;
   StencilVector _distances;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
-  Coordinate                         _simd_layout;
+  Coordinate    _simd_layout;
   Parameters    parameters;
   StencilEntry*  _entries_p;
   cobj* u_recv_buf_p;
@@ -145,18 +149,18 @@ class CartesianStencilView {
 
   accelerator_inline cobj *CommBuf(void) { return u_recv_buf_p; }
 
-  accelerator_inline int GetNodeLocal(int osite,int point) { 
+  accelerator_inline int GetNodeLocal(int osite,int point) {
     return this->_entries_p[point+this->_npoints*osite]._is_local;
   }
-  accelerator_inline StencilEntry * GetEntry(int &ptype,int point,int osite) { 
-    ptype = this->_permute_type[point]; return & this->_entries_p[point+this->_npoints*osite]; 
+  accelerator_inline StencilEntry * GetEntry(int &ptype,int point,int osite) {
+    ptype = this->_permute_type[point]; return & this->_entries_p[point+this->_npoints*osite];
   }
 
   accelerator_inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) {
     uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
     local = this->_entries_p[ent]._is_local;
     perm  = this->_entries_p[ent]._permute;
-    if (perm)  ptype = this->_permute_type[point]; 
+    if (perm)  ptype = this->_permute_type[point];
     if (local) {
       return  base + this->_entries_p[ent]._byte_offset;
     } else {
@@ -171,17 +175,47 @@ class CartesianStencilView {
     else       return cbase + this->_entries_p[ent]._byte_offset;
   }
 
-  accelerator_inline void iCoorFromIindex(Coordinate &coor,int lane) 
+  accelerator_inline void iCoorFromIindex(Coordinate &coor,int lane)
   {
     Lexicographic::CoorFromIndex(coor,lane,this->_simd_layout);
   }
+};
+
+template<class vobj,class cobj,class Parameters>
+class CartesianStencilView : public CartesianStencilAccelerator<vobj,cobj,Parameters>
+{
+ private:
+  int *closed;
+  StencilEntry *cpu_ptr;
+  ViewMode      mode;
+ public:
+  // default copy constructor
+  CartesianStencilView (const CartesianStencilView &refer_to_me) = default;
+
+  CartesianStencilView (const CartesianStencilAccelerator<vobj,cobj,Parameters> &refer_to_me,ViewMode _mode)
+    : CartesianStencilAccelerator<vobj,cobj,Parameters>(refer_to_me),
+    cpu_ptr(this->_entries_p),
+    mode(_mode)
+  {
+    this->_entries_p =(StencilEntry *)
+      MemoryManager::ViewOpen(this->_entries_p,
+			      this->_npoints*this->_osites*sizeof(StencilEntry),
+			      mode,
+			      AdviseDefault);
+  }
+
+  void ViewClose(void)
+  {
+    MemoryManager::ViewClose(this->cpu_ptr,this->mode);
+  }
 
 };
+
 ////////////////////////////////////////
 // The Stencil Class itself
 ////////////////////////////////////////
 template<class vobj,class cobj,class Parameters>
-class CartesianStencil : public CartesianStencilView<vobj,cobj,Parameters> { // Stencil runs along coordinate axes only; NO diagonal fill in.
+class CartesianStencil : public CartesianStencilAccelerator<vobj,cobj,Parameters> { // Stencil runs along coordinate axes only; NO diagonal fill in.
 public:
 
   typedef typename cobj::vector_type vector_type;
@@ -211,12 +245,12 @@ public:
     cobj * mpi_p;
     Integer buffer_size;
   };
-  
+
 
 protected:
   GridBase *                        _grid;
 
-public: 
+public:
   GridBase *Grid(void) const { return _grid; }
 
   ////////////////////////////////////////////////////////////////////////
@@ -226,11 +260,11 @@ public:
   // Generalise as required later if needed
   ////////////////////////////////////////////////////////////////////////
 
-  View_type View(void) const {
-    View_type accessor(*( (View_type *) this));
+  View_type View(ViewMode mode) const {
+    View_type accessor(*( (View_type *) this),mode);
     return accessor;
   }
-  
+
   int face_table_computed;
   std::vector<Vector<std::pair<int,int> > > face_table ;
   Vector<int> surface_list;
@@ -280,7 +314,7 @@ public:
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
-  inline int SameNode(int point) { 
+  inline int SameNode(int point) {
 
     int dimension    = this->_directions[point];
     int displacement = this->_distances[point];
@@ -304,7 +338,7 @@ public:
     // FIXME  this logic needs to be sorted for three link term
     //    assert( (displacement==1) || (displacement==-1));
     // Present hack only works for >= 4^4 subvol per node
-    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
+    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
 
     void *shm = (void *) _grid->ShmBufferTranslate(recv_from_rank,this->u_recv_buf_p);
 
@@ -344,7 +378,7 @@ public:
       comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
     }
   }
-  
+
   void CollateThreads(void)
   {
     int nthreads = CartesianCommunicator::nCommThreads;
@@ -368,7 +402,7 @@ public:
       if ( (t0 > 0.0) && ( t0 < first ) ) first = t0;   // min time seen
 
       if ( t1 > last ) last = t1;                       // max time seen
-      
+
     }
     commtime+= last-first;
   }
@@ -430,30 +464,30 @@ public:
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
     }
-  }    
-  
-  template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress) 
+  }
+
+  template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
   {
     Prepare();
     HaloGather(source,compress);
     Communicate();
-    CommsMergeSHM(compress); 
-    CommsMerge(compress); 
+    CommsMergeSHM(compress);
+    CommsMerge(compress);
   }
-  
+
   template<class compressor> int HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point,int & face_idx)
   {
     int dimension    = this->_directions[point];
     int displacement = this->_distances[point];
-    
+
     int fd = _grid->_fdimensions[dimension];
     int rd = _grid->_rdimensions[dimension];
-    
+
     // Map to always positive shift modulo global full dimension.
     int shift = (displacement+fd)%fd;
 
     assert (source.Checkerboard()== this->_checkerboard);
-    
+
     // the permute type
     int simd_layout     = _grid->_simd_layout[dimension];
     int comm_dim        = _grid->_processors[dimension] >1 ;
@@ -471,7 +505,7 @@ public:
 	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
 	  is_same_node = is_same_node && tmp;
 	  splicetime+=usecond();
-	} else { 
+	} else {
 	  nosplicetime-=usecond();
 	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
 	  is_same_node = is_same_node && tmp;
@@ -497,7 +531,7 @@ public:
     }
     return is_same_node;
   }
-  
+
   template<class compressor>
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
@@ -508,9 +542,9 @@ public:
     // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
     halogtime-=usecond();
-    
+
     u_comm_offset=0;
-    
+
     // Gather all comms buffers
     int face_idx=0;
     for(int point = 0 ; point < this->_npoints; point++) {
@@ -523,16 +557,16 @@ public:
     accelerator_barrier();
     halogtime+=usecond();
   }
- 
+
   /////////////////////////
   // Implementation
   /////////////////////////
   void Prepare(void)
   {
-    Decompressions.resize(0); 
-    DecompressionsSHM.resize(0); 
-    Mergers.resize(0); 
-    MergersSHM.resize(0); 
+    Decompressions.resize(0);
+    DecompressionsSHM.resize(0);
+    Mergers.resize(0);
+    MergersSHM.resize(0);
     Packets.resize(0);
     calls++;
   }
@@ -561,22 +595,22 @@ public:
     mv.push_back(m);
   }
   template<class decompressor>  void CommsMerge(decompressor decompress)    {
-    CommsMerge(decompress,Mergers,Decompressions); 
+    CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    mpi3synctime-=usecond();    
+    mpi3synctime-=usecond();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime+=usecond();    
-    shmmergetime-=usecond();    
+    mpi3synctime+=usecond();
+    shmmergetime-=usecond();
     CommsMerge(decompress,MergersSHM,DecompressionsSHM);
-    shmmergetime+=usecond();    
+    shmmergetime+=usecond();
   }
 
   template<class decompressor>
-  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) { 
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
 
     mergetime-=usecond();
-    for(int i=0;i<mm.size();i++){	
+    for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
       auto vp1= &mm[i].vpointers[1][0];
@@ -588,7 +622,7 @@ public:
     mergetime+=usecond();
 
     decompresstime-=usecond();
-    for(int i=0;i<dd.size();i++){	
+    for(int i=0;i<dd.size();i++){
       auto kp = dd[i].kernel_p;
       auto mp = dd[i].mpi_p;
       accelerator_forNB(o,dd[i].buffer_size,1,{
@@ -604,7 +638,7 @@ public:
     for(int i=0;i<_entries.size();i++){
       if( _entries[i]._is_local ) {
 	_entries[i]._byte_offset = _entries[i]._offset*sizeof(vobj);
-      } else { 
+      } else {
 	_entries[i]._byte_offset = _entries[i]._offset*sizeof(cobj);
       }
     }
@@ -619,15 +653,15 @@ public:
     for(int point=0;point<this->_npoints;point++){
       this->same_node[point] = this->SameNode(point);
     }
-    
+
     for(int site = 0 ;site< vol4;site++){
       int local = 1;
       for(int point=0;point<this->_npoints;point++){
-	if( (!this->GetNodeLocal(site*Ls,point)) && (!this->same_node[point]) ){ 
+	if( (!this->GetNodeLocal(site*Ls,point)) && (!this->same_node[point]) ){
 	  local = 0;
 	}
       }
-      if(local == 0) { 
+      if(local == 0) {
 	surface_list.push_back(site);
       }
     }
@@ -638,11 +672,11 @@ public:
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
-		   Parameters p) 
-    : shm_bytes_thr(npoints), 
-      comm_bytes_thr(npoints), 
+		   Parameters p)
+    : shm_bytes_thr(npoints),
+      comm_bytes_thr(npoints),
       comm_enter_thr(npoints),
-      comm_leave_thr(npoints), 
+      comm_leave_thr(npoints),
       comm_time_thr(npoints)
   {
     face_table_computed=0;
@@ -653,7 +687,7 @@ public:
     /////////////////////////////////////
     this->_npoints = npoints;
     this->_comm_buf_size.resize(npoints),
-    this->_permute_type.resize(npoints), 
+    this->_permute_type.resize(npoints),
     this->_simd_layout = _grid->_simd_layout; // copy simd_layout to give access to Accelerator Kernels
     this->_directions = StencilVector(directions);
     this->_distances  = StencilVector(distances);
@@ -662,25 +696,25 @@ public:
     _unified_buffer_size=0;
     surface_list.resize(0);
 
-    int osites  = _grid->oSites();
-    
-    _entries.resize(this->_npoints* osites);
+    this->_osites  = _grid->oSites();
+
+    _entries.resize(this->_npoints* this->_osites);
     this->_entries_p = &_entries[0];
     for(int ii=0;ii<npoints;ii++){
-      
+
       int i = ii; // reverse direction to get SIMD comms done first
       int point = i;
-      
+
       int dimension    = directions[i];
       int displacement = distances[i];
       int shift = displacement;
-      
+
       int fd = _grid->_fdimensions[dimension];
       int rd = _grid->_rdimensions[dimension];
       this->_permute_type[point]=_grid->PermuteType(dimension);
-      
+
       this->_checkerboard = checkerboard;
-      
+
       //////////////////////////
       // the permute type
       //////////////////////////
@@ -690,25 +724,25 @@ public:
       int rotate_dim      = _grid->_simd_layout[dimension]>2;
 
       assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
-      
+
       int sshift[2];
-      
+
       //////////////////////////
       // Underlying approach. For each local site build
-      // up a table containing the npoint "neighbours" and whether they 
+      // up a table containing the npoint "neighbours" and whether they
       // live in lattice or a comms buffer.
       //////////////////////////
       if ( !comm_dim ) {
 	sshift[0] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Even);
 	sshift[1] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Odd);
-	
+
 	if ( sshift[0] == sshift[1] ) {
 	  Local(point,dimension,shift,0x3);
 	} else {
 	  Local(point,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
 	  Local(point,dimension,shift,0x2);// both with block stride loop iteration
 	}
-      } else { 
+      } else {
 	// All permute extract done in comms phase prior to Stencil application
 	//        So tables are the same whether comm_dim or splice_dim
 	sshift[0] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Even);
@@ -750,23 +784,23 @@ public:
     int ld = _grid->_ldimensions[dimension];
     int gd = _grid->_gdimensions[dimension];
     int ly = _grid->_simd_layout[dimension];
-    
+
     // Map to always positive shift modulo global full dimension.
     int shift = (shiftpm+fd)%fd;
 
     // the permute type
     int permute_dim =_grid->PermuteDim(dimension);
-    
-    for(int x=0;x<rd;x++){       
-      
+
+    for(int x=0;x<rd;x++){
+
       //      int o   = 0;
       int bo  = x * _grid->_ostride[dimension];
-      
+
       int cb= (cbmask==0x2)? Odd : Even;
-      
+
       int sshift = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,cb);
       int sx     = (x+sshift)%rd;
-      
+
       int wraparound=0;
       if ( (shiftpm==-1) && (sx>x)  ) {
 	wraparound = 1;
@@ -774,7 +808,7 @@ public:
       if ( (shiftpm== 1) && (sx<x)  ) {
 	wraparound = 1;
       }
-      
+
       int permute_slice=0;
       if(permute_dim){
 	int wrap = sshift/rd; wrap=wrap % ly; // but it is local anyway
@@ -782,66 +816,66 @@ public:
 	if ( x< rd-num ) permute_slice=wrap;
 	else permute_slice = (wrap+1)%ly;
       }
-      
+
       CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-      
+
     }
   }
-  
+
   void Comms     (int point,int dimension,int shiftpm,int cbmask)
   {
     GridBase *grid=_grid;
     const int Nsimd = grid->Nsimd();
-    
+
     int fd              = _grid->_fdimensions[dimension];
     int ld              = _grid->_ldimensions[dimension];
     int rd              = _grid->_rdimensions[dimension];
     int pd              = _grid->_processors[dimension];
     int simd_layout     = _grid->_simd_layout[dimension];
     int comm_dim        = _grid->_processors[dimension] >1 ;
-    
+
     assert(comm_dim==1);
     int shift = (shiftpm + fd) %fd;
     assert(shift>=0);
     assert(shift<fd);
-    
+
     // done in reduced dims, so SIMD factored
-    int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension]; 
+    int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
 
     this->_comm_buf_size[point] = buffer_size; // Size of _one_ plane. Multiple planes may be gathered and
 
     // send to one or more remote nodes.
-    
+
     int cb= (cbmask==0x2)? Odd : Even;
     int sshift= _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,cb);
-    
-    for(int x=0;x<rd;x++){       
-      
+
+    for(int x=0;x<rd;x++){
+
       int permute_type=grid->PermuteType(dimension);
-      
+
       int sx        =  (x+sshift)%rd;
-      
+
       int offnode = 0;
       if ( simd_layout > 1 ) {
-	
+
 	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
-	  
-	  if ( nbr_proc ) { 
+
+	  if ( nbr_proc ) {
 	    offnode =1;
 	  }
 	}
-	
-      } else { 
+
+      } else {
 	int comm_proc = ((x+sshift)/rd)%pd;
 	offnode = (comm_proc!= 0);
       }
-      
+
       int wraparound=0;
       if ( (shiftpm==-1) && (sx>x) && (grid->_processor_coor[dimension]==0) ) {
 	wraparound = 1;
@@ -850,24 +884,24 @@ public:
 	wraparound = 1;
       }
       if (!offnode) {
-	
+
 	int permute_slice=0;
-	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound); 
-	
+	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
+
       } else {
 
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
-	
+
 	//	int rank           = grid->_processor;
 	//	int recv_from_rank;
 	//	int xmit_to_rank;
-	
+
 	int unified_buffer_offset = _unified_buffer_size;
 	_unified_buffer_size    += words;
-	
+
 	ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
-	
+
       }
     }
   }
@@ -875,13 +909,13 @@ public:
   void CopyPlane(int point, int dimension,int lplane,int rplane,int cbmask,int permute,int wrap)
   {
     int rd = _grid->_rdimensions[dimension];
-    
+
     if ( !_grid->CheckerBoarded(dimension) ) {
-      
+
       int o   = 0;                                     // relative offset to base within plane
-      int ro  = rplane*_grid->_ostride[dimension]; // base offset for start of plane 
+      int ro  = rplane*_grid->_ostride[dimension]; // base offset for start of plane
       int lo  = lplane*_grid->_ostride[dimension]; // offset in buffer
-      
+
       // Simple block stride gather of SIMD objects
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
@@ -893,18 +927,18 @@ public:
 	}
 	o +=_grid->_slice_stride[dimension];
       }
-      
+
     } else {
-      
-      int ro  = rplane*_grid->_ostride[dimension]; // base offset for start of plane 
-      int lo  = lplane*_grid->_ostride[dimension]; // base offset for start of plane 
+
+      int ro  = rplane*_grid->_ostride[dimension]; // base offset for start of plane
+      int lo  = lplane*_grid->_ostride[dimension]; // base offset for start of plane
       int o   = 0;                                     // relative offset to base within plane
-      
+
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
-	  
+
 	  int ocb=1<<_grid->CheckerBoardFromOindex(o+b);
-	  
+
 	  if ( ocb&cbmask ) {
 	    int idx = point+(lo+o+b)*this->_npoints;
 	    _entries[idx]._offset =ro+o+b;
@@ -912,24 +946,24 @@ public:
 	    _entries[idx]._permute=permute;
 	    _entries[idx]._around_the_world=wrap;
 	  }
-	  
+
 	}
 	o +=_grid->_slice_stride[dimension];
       }
-      
+
     }
   }
   // Routine builds up integer table for each site in _offsets, _is_local, _permute
   void ScatterPlane (int point,int dimension,int plane,int cbmask,int offset, int wrap)
   {
     int rd = _grid->_rdimensions[dimension];
-    
+
     if ( !_grid->CheckerBoarded(dimension) ) {
-      
-      int so  = plane*_grid->_ostride[dimension]; // base offset for start of plane 
+
+      int so  = plane*_grid->_ostride[dimension]; // base offset for start of plane
       int o   = 0;                                    // relative offset to base within plane
       int bo  = 0;                                    // offset in buffer
-      
+
       // Simple block stride gather of SIMD objects
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
@@ -941,16 +975,16 @@ public:
 	}
 	o +=_grid->_slice_stride[dimension];
       }
-      
-    } else { 
-      
-      int so  = plane*_grid->_ostride[dimension]; // base offset for start of plane 
+
+    } else {
+
+      int so  = plane*_grid->_ostride[dimension]; // base offset for start of plane
       int o   = 0;                                      // relative offset to base within plane
       int bo  = 0;                                      // offset in buffer
-      
+
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
-	  
+
 	  int ocb=1<<_grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	  if ( ocb & cbmask ) {
 	    int idx = point+(so+o+b)*this->_npoints;
@@ -964,16 +998,16 @@ public:
       }
     }
   }
-  
+
   template<class compressor>
   int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
   {
     typedef typename cobj::vector_type vector_type;
     typedef typename cobj::scalar_type scalar_type;
-    
+
     assert(rhs.Grid()==_grid);
     //	  conformable(_grid,rhs.Grid());
-    
+
     int fd              = _grid->_fdimensions[dimension];
     int rd              = _grid->_rdimensions[dimension];
     int pd              = _grid->_processors[dimension];
@@ -985,37 +1019,37 @@ public:
     assert(shift<fd);
 
     int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
-    
+
     int cb= (cbmask==0x2)? Odd : Even;
     int sshift= _grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-    
+
     int shm_receive_only = 1;
-    for(int x=0;x<rd;x++){       
-      
+    for(int x=0;x<rd;x++){
+
       int sx        = (x+sshift)%rd;
       int comm_proc = ((x+sshift)/rd)%pd;
-      
+
       if (comm_proc) {
 
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
-	
+
 	int bytes =  words * compress.CommDatumSize();
-	
-	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
+
+	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 	if ( !face_table_computed ) {
 	  face_table.resize(face_idx+1);
 	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table[face_idx]);
 	}
-	
+
 	//      	int rank           = _grid->_processor;
 	int recv_from_rank;
 	int xmit_to_rank;
 	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-	
+
 	assert (xmit_to_rank   != _grid->ThisRank());
 	assert (recv_from_rank != _grid->ThisRank());
-	
+
 	/////////////////////////////////////////////////////////
 	// try the direct copy if possible
 	/////////////////////////////////////////////////////////
@@ -1028,13 +1062,13 @@ public:
 	}
 
 	send_buf = (cobj *)_grid->ShmBufferTranslate(xmit_to_rank,recv_buf);
-	if ( send_buf==NULL ) { 
+	if ( send_buf==NULL ) {
 	  send_buf = this->u_send_buf_p;
-	} 
-	
+	}
+
 	// Find out if we get the direct copy.
 	void *success = (void *) _grid->ShmBufferTranslate(recv_from_rank,this->u_send_buf_p);
-	if (success==NULL) { 
+	if (success==NULL) {
 	  // we found a packet that comes from MPI and contributes to this leg of stencil
 	  shm_receive_only = 0;
 	}
@@ -1043,9 +1077,9 @@ public:
 	assert(send_buf!=NULL);
 	Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so);  face_idx++;
 	gathertime+=usecond();
-	
+
 	if ( compress.DecompressionStep() ) {
-	  
+
 	  if ( shm_receive_only ) { // Early decompress before MPI is finished is possible
 	    AddDecompress(&this->u_recv_buf_p[u_comm_offset],
 			  &recv_buf[u_comm_offset],
@@ -1074,7 +1108,7 @@ public:
     }
     return shm_receive_only;
   }
-  
+
   template<class compressor>
   int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
   {
@@ -1102,7 +1136,7 @@ public:
     ///////////////////////////////////////////////
     int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
     //    int words = sizeof(cobj)/sizeof(vector_type);
-    
+
     assert(cbmask==0x3); // Fixme think there is a latent bug if not true
                          // This assert will trap it if ever hit. Not hit normally so far
     int reduced_buffer_size = buffer_size;
@@ -1118,22 +1152,22 @@ public:
     ///////////////////////////////////////////
     // 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
     int shm_receive_only = 1;
-    for(int x=0;x<rd;x++){       
-      
+    for(int x=0;x<rd;x++){
+
       int any_offnode = ( ((x+sshift)%fd) >= rd );
 
       if ( any_offnode ) {
-	
-	for(int i=0;i<maxl;i++){       
+
+	for(int i=0;i<maxl;i++){
 	  spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
 	}
-	
+
 	int sx   = (x+sshift)%rd;
 
 	if ( !face_table_computed ) {
@@ -1168,13 +1202,13 @@ public:
 
 	    int recv_from_rank;
 	    int xmit_to_rank;
-	    
-	    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
- 
+
+	    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
+
 	    // shm == receive pointer         if offnode
 	    // shm == Translate[send pointer] if on node -- my view of his send pointer
 	    cobj *shm = (cobj *) _grid->ShmBufferTranslate(recv_from_rank,sp);
-	    if (shm==NULL) { 
+	    if (shm==NULL) {
 	      shm = rp;
 	      // we found a packet that comes from MPI and contributes to this shift.
 	      // is_same_node is only used in the WilsonStencil, and gets set for this point in the stencil.
@@ -1188,15 +1222,15 @@ public:
 
 	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
 
-	    
-	  } else { 
-	    
+
+	  } else {
+
 	    rpointers[i] = sp;
-	    
+
 	  }
 	}
 
-	if ( shm_receive_only ) { 
+	if ( shm_receive_only ) {
 	  AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,MergersSHM);
 	} else {
 	  AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
@@ -1231,9 +1265,9 @@ public:
     shm_bytes = 0.;
     calls = 0.;
   };
-  
+
   void Report(void) {
-#define AVERAGE(A) 
+#define AVERAGE(A)
 #define PRINTIT(A) AVERAGE(A); std::cout << GridLogMessage << " Stencil " << #A << " "<< A/calls<<std::endl;
     RealD NP = _grid->_Nprocessors;
     RealD NN = _grid->NodeCount();
@@ -1250,7 +1284,7 @@ public:
       }
     }
     if (threaded) commtime += t;
-    
+
     _grid->GlobalSum(commtime);    commtime/=NP;
     if ( calls > 0. ) {
       std::cout << GridLogMessage << " Stencil calls "<<calls<<std::endl;
@@ -1273,8 +1307,8 @@ public:
 	std::cout << GridLogMessage << " Stencil SHM " << (shm_bytes)/gatheralltime/1000.*NP/NN << " GB/s per node"<<std::endl;
 
 	auto all_bytes = comms_bytes+shm_bytes;
-	std::cout << GridLogMessage << " Stencil SHM all" << (all_bytes)/gatheralltime/1000. << " GB/s per rank"<<std::endl;
-	std::cout << GridLogMessage << " Stencil SHM all" << (all_bytes)/gatheralltime/1000.*NP/NN << " GB/s per node"<<std::endl;
+	std::cout << GridLogMessage << " Stencil SHM all " << (all_bytes)/gatheralltime/1000. << " GB/s per rank"<<std::endl;
+	std::cout << GridLogMessage << " Stencil SHM all " << (all_bytes)/gatheralltime/1000.*NP/NN << " GB/s per node"<<std::endl;
 
 	auto membytes = (shm_bytes + comms_bytes/2) // read/write
 	              + (shm_bytes+comms_bytes)/2 * sizeof(vobj)/sizeof(cobj);
@@ -1292,7 +1326,7 @@ public:
 #undef PRINTIT
 #undef AVERAGE
   };
-  
+
 };
 NAMESPACE_END(Grid);
 
diff --git a/Grid/tensors/Tensor_SIMT.h b/Grid/tensors/Tensor_SIMT.h
index c18b8484..ec57a679 100644
--- a/Grid/tensors/Tensor_SIMT.h
+++ b/Grid/tensors/Tensor_SIMT.h
@@ -31,22 +31,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-//accelerator_inline void SIMTsynchronise(void) 
-accelerator_inline void synchronise(void) 
-{
-#ifdef __CUDA_ARCH__
-//  __syncthreads();
-  __syncwarp();
-#endif
-  return;
-}
 
-#ifndef __CUDA_ARCH__
+#ifndef GRID_SIMT
 //////////////////////////////////////////
 // Trivial mapping of vectors on host
 //////////////////////////////////////////
-accelerator_inline int SIMTlane(int Nsimd) { return 0; } // CUDA specific
-
 template<class vobj> accelerator_inline
 vobj coalescedRead(const vobj & __restrict__ vec,int lane=0)
 {
@@ -66,7 +55,6 @@ vobj coalescedReadPermute(const vobj & __restrict__ vec,int ptype,int doperm,int
 template<class vobj> accelerator_inline
 void coalescedWrite(vobj & __restrict__ vec,const vobj & __restrict__ extracted,int lane=0)
 {
-  //  vstream(vec, extracted);
   vec = extracted;
 }
 template<class vobj> accelerator_inline
@@ -75,25 +63,24 @@ void coalescedWriteNonTemporal(vobj & __restrict__ vec,const vobj & __restrict__
   vstream(vec, extracted);
 }
 #else
-accelerator_inline int SIMTlane(int Nsimd) { return threadIdx.y; } // CUDA specific
 
 //////////////////////////////////////////
 // Extract and insert slices on the GPU
 //////////////////////////////////////////
 template<class vobj> accelerator_inline
-typename vobj::scalar_object coalescedRead(const vobj & __restrict__ vec,int lane=SIMTlane(vobj::Nsimd()))
+typename vobj::scalar_object coalescedRead(const vobj & __restrict__ vec,int lane=acceleratorSIMTlane(vobj::Nsimd()))
 {
   return extractLane(lane,vec);
 }
 template<class vobj> accelerator_inline
-typename vobj::scalar_object coalescedReadPermute(const vobj & __restrict__ vec,int ptype,int doperm,int lane=SIMTlane(vobj::Nsimd()))
+typename vobj::scalar_object coalescedReadPermute(const vobj & __restrict__ vec,int ptype,int doperm,int lane=acceleratorSIMTlane(vobj::Nsimd()))
 {
   int mask = vobj::Nsimd() >> (ptype + 1);		
   int plane= doperm ? lane ^ mask : lane;
   return extractLane(plane,vec);
 }
 template<class vobj> accelerator_inline
-void coalescedWrite(vobj & __restrict__ vec,const typename vobj::scalar_object & __restrict__ extracted,int lane=SIMTlane(vobj::Nsimd()))
+void coalescedWrite(vobj & __restrict__ vec,const typename vobj::scalar_object & __restrict__ extracted,int lane=acceleratorSIMTlane(vobj::Nsimd()))
 {
   insertLane(lane,vec,extracted);
 }
diff --git a/Grid/tensors/Tensor_class.h b/Grid/tensors/Tensor_class.h
index dbcbae8d..36becc49 100644
--- a/Grid/tensors/Tensor_class.h
+++ b/Grid/tensors/Tensor_class.h
@@ -59,6 +59,20 @@ class GridTensorBase {};
   using DoublePrecision2= typename Traits::DoublePrecision2; \
   static constexpr int TensorLevel = Traits::TensorLevel
 
+///////////////////////////////////////////////////////////
+// Allows to turn scalar<scalar<scalar<double>>>> back to double.
+///////////////////////////////////////////////////////////
+template <class T>
+accelerator_inline typename std::enable_if<!isGridTensor<T>::value, T>::type
+TensorRemove(T arg) {
+  return arg;
+}
+template <class vtype>
+accelerator_inline auto TensorRemove(iScalar<vtype> arg)
+  -> decltype(TensorRemove(arg._internal)) {
+  return TensorRemove(arg._internal);
+}
+
 template <class vtype>
 class iScalar {
 public:
@@ -135,9 +149,10 @@ public:
   operator ComplexD() const {
     return (TensorRemove(_internal));
   }
+  //             instantiation of "Grid::iScalar<vtype>::operator Grid::RealD() const [with vtype=Grid::Real, U=Grid::Real, V=Grid::RealD, <unnamed>=0, <unnamed>=0U]" 
   template <class U = vtype, class V = scalar_type, IfReal<V> = 0,IfNotSimd<U> = 0> accelerator_inline
   operator RealD() const {
-    return TensorRemove(_internal);
+    return (RealD) TensorRemove(_internal);
   }
   template <class U = vtype, class V = scalar_type, IfInteger<V> = 0, IfNotSimd<U> = 0> accelerator_inline
   operator Integer() const {
@@ -169,20 +184,6 @@ public:
   strong_inline       scalar_type * end()         { return begin() + Traits::count; }
 };
 
-///////////////////////////////////////////////////////////
-// Allows to turn scalar<scalar<scalar<double>>>> back to double.
-///////////////////////////////////////////////////////////
-template <class T>
-accelerator_inline typename std::enable_if<!isGridTensor<T>::value, T>::type
-TensorRemove(T arg) {
-  return arg;
-}
-template <class vtype>
-accelerator_inline auto TensorRemove(iScalar<vtype> arg)
-  -> decltype(TensorRemove(arg._internal)) {
-  return TensorRemove(arg._internal);
-}
-
 template <class vtype, int N>
 class iVector {
 public:
diff --git a/Grid/tensors/Tensor_exp.h b/Grid/tensors/Tensor_exp.h
index 11d37f9c..0a1d6389 100644
--- a/Grid/tensors/Tensor_exp.h
+++ b/Grid/tensors/Tensor_exp.h
@@ -55,7 +55,7 @@ template<class vtype, int N> accelerator_inline iVector<vtype, N> Exponentiate(c
 
 
 // Specialisation: Cayley-Hamilton exponential for SU(3)
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
 template<class vtype, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0>::type * =nullptr> 
 accelerator_inline iMatrix<vtype,3> Exponentiate(const iMatrix<vtype,3> &arg, RealD alpha  , Integer Nexp = DEFAULT_MAT_EXP )
 {
diff --git a/Grid/threads/Accelerator.cc b/Grid/threads/Accelerator.cc
new file mode 100644
index 00000000..2c4ad9df
--- /dev/null
+++ b/Grid/threads/Accelerator.cc
@@ -0,0 +1,207 @@
+#include <Grid/GridCore.h>
+
+NAMESPACE_BEGIN(Grid);
+uint32_t accelerator_threads=2;
+uint32_t acceleratorThreads(void)       {return accelerator_threads;};
+void     acceleratorThreads(uint32_t t) {accelerator_threads = t;};
+
+#ifdef GRID_CUDA
+cudaDeviceProp *gpu_props;
+void acceleratorInit(void)
+{
+  int nDevices = 1;
+  cudaGetDeviceCount(&nDevices);
+  gpu_props = new cudaDeviceProp[nDevices];
+
+  char * localRankStr = NULL;
+  int rank = 0, world_rank=0; 
+#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
+#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_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)
+  {
+    rank = atoi(localRankStr);		
+  }
+  if ((localRankStr = getenv(ENV_LOCAL_RANK_MVAPICH)) != NULL)
+  {
+    rank = atoi(localRankStr);		
+  }
+  if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
+  if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
+
+  size_t totalDeviceMem=0;
+  for (int i = 0; i < nDevices; i++) {
+
+#define GPU_PROP_FMT(canMapHostMemory,FMT)     printf("AcceleratorCudaInit:   " #canMapHostMemory ": " FMT" \n",prop.canMapHostMemory);
+#define GPU_PROP(canMapHostMemory)             GPU_PROP_FMT(canMapHostMemory,"%d");
+    cudaGetDeviceProperties(&gpu_props[i], i);
+    cudaDeviceProp prop; 
+    prop = gpu_props[i];
+    totalDeviceMem = prop.totalGlobalMem;
+    if ( world_rank == 0) {
+      printf("AcceleratorCudaInit: ========================\n");
+      printf("AcceleratorCudaInit: Device Number    : %d\n", i);
+      printf("AcceleratorCudaInit: ========================\n");
+      printf("AcceleratorCudaInit: Device identifier: %s\n", prop.name);
+
+      GPU_PROP_FMT(totalGlobalMem,"%lld");
+      GPU_PROP(managedMemory);
+      GPU_PROP(isMultiGpuBoard);
+      GPU_PROP(warpSize);
+      //      GPU_PROP(unifiedAddressing);
+      //      GPU_PROP(l2CacheSize);
+      //      GPU_PROP(singleToDoublePrecisionPerfRatio);
+    }
+  }
+  MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
+#undef GPU_PROP_FMT    
+#undef GPU_PROP
+
+#ifdef GRID_IBM_SUMMIT
+  // IBM Jsrun makes cuda Device numbering screwy and not match rank
+  if ( world_rank == 0 )  printf("AcceleratorCudaInit: IBM Summit or similar - NOT setting device to node rank\n");
+#else
+  if ( world_rank == 0 )  printf("AcceleratorCudaInit: setting device to node rank\n");
+  cudaSetDevice(rank);
+#endif
+  if ( world_rank == 0 )  printf("AcceleratorCudaInit: ================================================\n");
+}
+#endif
+
+#ifdef GRID_HIP
+hipDeviceProp_t *gpu_props;
+void acceleratorInit(void)
+{
+  int nDevices = 1;
+  hipGetDeviceCount(&nDevices);
+  gpu_props = new hipDeviceProp_t[nDevices];
+
+  char * localRankStr = NULL;
+  int rank = 0, world_rank=0; 
+#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
+#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_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)
+  {
+    rank = atoi(localRankStr);		
+  }
+  if ((localRankStr = getenv(ENV_LOCAL_RANK_MVAPICH)) != NULL)
+  {
+    rank = atoi(localRankStr);		
+  }
+  if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
+  if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
+
+  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);
+    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(managedMemory);
+      GPU_PROP(isMultiGpuBoard);
+      GPU_PROP(warpSize);
+      //      GPU_PROP(unifiedAddressing);
+      //      GPU_PROP(l2CacheSize);
+      //      GPU_PROP(singleToDoublePrecisionPerfRatio);
+    }
+  }
+#undef GPU_PROP_FMT    
+#undef GPU_PROP
+#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: setting device to node rank\n");
+  hipSetDevice(rank);
+#endif
+  if ( world_rank == 0 )  printf("AcceleratorHipInit: ================================================\n");
+}
+#endif
+
+
+#ifdef GRID_SYCL
+
+cl::sycl::queue *theGridAccelerator;
+
+void acceleratorInit(void)
+{
+  int nDevices = 1;
+  cl::sycl::gpu_selector selector;
+  cl::sycl::device selectedDevice { selector };
+  theGridAccelerator = new sycl::queue (selectedDevice);
+
+  char * localRankStr = NULL;
+  int rank = 0, world_rank=0; 
+#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
+#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_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)
+  {
+    rank = atoi(localRankStr);		
+  }
+  if ((localRankStr = getenv(ENV_LOCAL_RANK_MVAPICH)) != NULL)
+  {
+    rank = atoi(localRankStr);		
+  }
+  if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
+  if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
+
+  auto devices = cl::sycl::device::get_devices();
+  for(int d = 0;d<devices.size();d++){
+
+#define GPU_PROP_STR(prop) \
+    printf("AcceleratorSyclInit:   " #prop ": %s \n",devices[d].get_info<cl::sycl::info::device::prop>().c_str());
+
+#define GPU_PROP_FMT(prop,FMT) \
+    printf("AcceleratorSyclInit:   " #prop ": " FMT" \n",devices[d].get_info<cl::sycl::info::device::prop>());
+
+#define GPU_PROP(prop)             GPU_PROP_FMT(prop,"%ld");
+
+    GPU_PROP_STR(vendor);
+    GPU_PROP_STR(version);
+    //    GPU_PROP_STR(device_type);
+    /*
+    GPU_PROP(max_compute_units);
+    GPU_PROP(native_vector_width_char);
+    GPU_PROP(native_vector_width_short);
+    GPU_PROP(native_vector_width_int);
+    GPU_PROP(native_vector_width_long);
+    GPU_PROP(native_vector_width_float);
+    GPU_PROP(native_vector_width_double);
+    GPU_PROP(native_vector_width_half);
+    GPU_PROP(address_bits);
+    GPU_PROP(half_fp_config);
+    GPU_PROP(single_fp_config);
+    */
+    //    GPU_PROP(double_fp_config);
+    GPU_PROP(global_mem_size);
+
+  }
+  if ( world_rank == 0 ) {
+    auto name = theGridAccelerator->get_device().get_info<sycl::info::device::name>();
+    printf("AcceleratorSyclInit: Selected device is %s\n",name.c_str());
+    printf("AcceleratorSyclInit: ================================================\n");
+  }
+}
+#endif
+
+#if (!defined(GRID_CUDA)) && (!defined(GRID_SYCL))&& (!defined(GRID_HIP))
+void acceleratorInit(void){}
+#endif
+
+NAMESPACE_END(Grid);
diff --git a/Grid/threads/Accelerator.h b/Grid/threads/Accelerator.h
new file mode 100644
index 00000000..74a3ea22
--- /dev/null
+++ b/Grid/threads/Accelerator.h
@@ -0,0 +1,426 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/Accelerator.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include <string.h>
+
+#ifdef HAVE_MALLOC_MALLOC_H
+#include <malloc/malloc.h>
+#endif
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
+#ifdef HAVE_MM_MALLOC_H
+#include <mm_malloc.h>
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+//////////////////////////////////////////////////////////////////////////////////
+// Accelerator primitives; fall back to threading if not CUDA or SYCL
+//////////////////////////////////////////////////////////////////////////////////
+//
+// Function attributes
+//
+//    accelerator
+//    accelerator_inline
+//
+// Parallel looping
+// 
+//    accelerator_for
+//    accelerator_forNB 
+//    uint32_t accelerator_barrier();         // device synchronise
+//
+// Parallelism control: Number of threads in thread block is acceleratorThreads*Nsimd
+//
+//    uint32_t acceleratorThreads(void);   
+//    void     acceleratorThreads(uint32_t);
+//
+// Warp control and info:
+//
+//    acceleratorInit;
+//    void     acceleratorSynchronise(void); // synch warp etc..
+//    int      acceleratorSIMTlane(int Nsimd);
+//
+// Memory management:
+//
+//    void *acceleratorAllocShared(size_t bytes);
+//    void acceleratorFreeShared(void *ptr);
+//
+//    void *acceleratorAllocDevice(size_t bytes);
+//    void acceleratorFreeDevice(void *ptr);
+//
+//    void *acceleratorCopyToDevice(void *from,void *to,size_t bytes);
+//    void *acceleratorCopyFromDevice(void *from,void *to,size_t bytes);
+//
+//////////////////////////////////////////////////////////////////////////////////
+
+uint32_t acceleratorThreads(void);   
+void     acceleratorThreads(uint32_t);
+void     acceleratorInit(void);
+
+//////////////////////////////////////////////
+// CUDA acceleration
+//////////////////////////////////////////////
+
+#ifdef GRID_CUDA
+
+#ifdef __CUDA_ARCH__
+#define GRID_SIMT
+#endif
+
+#define accelerator        __host__ __device__
+#define accelerator_inline __host__ __device__ inline
+
+accelerator_inline int acceleratorSIMTlane(int Nsimd) {
+#ifdef GRID_SIMT
+  return threadIdx.z; 
+#else
+  return 0;
+#endif
+} // CUDA specific
+
+#define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
+  {									\
+    typedef uint64_t Iterator;						\
+    auto lambda = [=] accelerator					\
+      (Iterator iter1,Iterator iter2,Iterator lane) mutable {		\
+      __VA_ARGS__;							\
+    };									\
+    int nt=acceleratorThreads();					\
+    dim3 cu_threads(acceleratorThreads(),1,nsimd);			\
+    dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
+    LambdaApply<<<cu_blocks,cu_threads>>>(num1,num2,nsimd,lambda);	\
+  }
+
+template<typename lambda>  __global__
+void LambdaApply(uint64_t num1, uint64_t num2, uint64_t num3, lambda Lambda)
+{
+  uint64_t x = threadIdx.x + blockDim.x*blockIdx.x;
+  uint64_t y = threadIdx.y + blockDim.y*blockIdx.y;
+  uint64_t z = threadIdx.z;
+  if ( (x < num1) && (y<num2) && (z<num3) ) {
+    Lambda(x,y,z);
+  }
+}
+
+#define accelerator_barrier(dummy)					\
+  {									\
+    cudaDeviceSynchronize();						\
+    cudaError err = cudaGetLastError();					\
+    if ( cudaSuccess != err ) {						\
+      printf("Cuda error %s \n", cudaGetErrorString( err ));		\
+      puts(__FILE__);							\
+      printf("Line %d\n",__LINE__);					\
+    }									\
+  }
+
+inline void *acceleratorAllocShared(size_t bytes)
+{
+  void *ptr=NULL;
+  auto err = cudaMallocManaged((void **)&ptr,bytes);
+  if( err != cudaSuccess ) {
+    ptr = (void *) NULL;
+    printf(" cudaMallocManaged failed for %d %s \n",bytes,cudaGetErrorString(err));
+  }
+  return ptr;
+};
+inline void *acceleratorAllocDevice(size_t bytes)
+{
+  void *ptr=NULL;
+  auto err = cudaMalloc((void **)&ptr,bytes);
+  if( err != cudaSuccess ) {
+    ptr = (void *) NULL;
+    printf(" cudaMalloc failed for %d %s \n",bytes,cudaGetErrorString(err));
+  }
+  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 acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
+#endif
+
+//////////////////////////////////////////////
+// SyCL acceleration
+//////////////////////////////////////////////
+
+#ifdef GRID_SYCL
+NAMESPACE_END(Grid);
+#include <CL/sycl.hpp>
+#include <CL/sycl/usm.hpp>
+NAMESPACE_BEGIN(Grid);
+
+extern cl::sycl::queue *theGridAccelerator;
+
+#ifdef __SYCL_DEVICE_ONLY__
+#define GRID_SIMT
+#endif
+
+#define accelerator 
+#define accelerator_inline strong_inline
+
+accelerator_inline int acceleratorSIMTlane(int Nsimd) {
+#ifdef GRID_SIMT
+ return __spirv::initLocalInvocationId<3, cl::sycl::id<3>>()[2]; 
+#else
+ return 0;
+#endif
+} // SYCL specific
+
+#define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {		\
+      unsigned long nt=acceleratorThreads();				\
+      unsigned long unum1 = num1;					\
+      unsigned long unum2 = num2;					\
+      cl::sycl::range<3> local {nt,1,nsimd};				\
+      cl::sycl::range<3> global{unum1,unum2,nsimd};			\
+      cgh.parallel_for<class dslash>(					\
+      cl::sycl::nd_range<3>(global,local), \
+      [=] (cl::sycl::nd_item<3> item) mutable {       \
+      auto iter1    = item.get_global_id(0);	      \
+      auto iter2    = item.get_global_id(1);	      \
+      auto lane     = item.get_global_id(2);	      \
+      { __VA_ARGS__ };				      \
+     });	   			              \
+    });
+
+#define accelerator_barrier(dummy) theGridAccelerator->wait();
+
+inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
+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 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();}
+
+#endif
+
+//////////////////////////////////////////////
+// HIP acceleration
+//////////////////////////////////////////////
+#ifdef GRID_HIP
+NAMESPACE_END(Grid);
+#include <hip/hip_runtime.h>
+NAMESPACE_BEGIN(Grid);
+
+#ifdef __HIP_DEVICE_COMPILE__
+#define GRID_SIMT
+#endif
+
+#define accelerator        __host__ __device__
+#define accelerator_inline __host__ __device__ inline
+
+/*These routines define mapping from thread grid to loop & vector lane indexing */
+accelerator_inline int acceleratorSIMTlane(int Nsimd) {
+#ifdef GRID_SIMT
+  return hipThreadIdx_z; 
+#else
+  return 0;
+#endif
+} // HIP specific
+
+#define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
+  {									\
+    typedef uint64_t Iterator;						\
+    auto lambda = [=] accelerator					\
+      (Iterator iter1,Iterator iter2,Iterator lane ) mutable {		\
+      { __VA_ARGS__;}							\
+    };									\
+    int nt=acceleratorThreads();					\
+    dim3 hip_threads(nt,1,nsimd);					\
+    dim3 hip_blocks ((num1+nt-1)/nt,num2,1);				\
+    hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
+		       0,0,						\
+		       num1,num2,nsimd,lambda);				\
+  }
+
+template<typename lambda>  __global__
+void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
+{
+  uint64_t x = hipThreadIdx_x + hipBlockDim_x*hipBlockIdx_x;
+  uint64_t y = hipThreadIdx_y + hipBlockDim_y*hipBlockIdx_y;
+  uint64_t z = hipThreadIdx_z ;//+ hipBlockDim_z*hipBlockIdx_z;
+  if ( (x < numx) && (y<numy) && (z<numz) ) {
+    Lambda(x,y,z);
+  }
+}
+
+#define accelerator_barrier(dummy)				\
+  {								\
+    hipDeviceSynchronize();					\
+    auto err = hipGetLastError();				\
+    if ( err != hipSuccess ) {					\
+      printf("After hipDeviceSynchronize() : HIP error %s \n", hipGetErrorString( err )); \
+      puts(__FILE__);							\
+      printf("Line %d\n",__LINE__);				\
+      exit(0);							\
+    }								\
+  }
+
+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 %d %s \n",bytes,hipGetErrorString(err));
+  }
+  return ptr;
+#else
+  return malloc(bytes);
+#endif
+};
+
+inline void *acceleratorAllocDevice(size_t bytes)
+{
+  void *ptr=NULL;
+  auto err = hipMalloc((void **)&ptr,bytes);
+  if( err != hipSuccess ) {
+    ptr = (void *) NULL;
+    printf(" hipMalloc failed for %d %s \n",bytes,hipGetErrorString(err));
+  }
+  return 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);}
+
+#endif
+
+//////////////////////////////////////////////
+// Common on all GPU targets
+//////////////////////////////////////////////
+#if defined(GRID_SYCL) || defined(GRID_CUDA) || defined(GRID_HIP)
+#define accelerator_forNB( iter1, num1, nsimd, ... ) accelerator_for2dNB( iter1, num1, iter2, 1, nsimd, {__VA_ARGS__} );
+
+#define accelerator_for( iter, num, nsimd, ... )		\
+  accelerator_forNB(iter, num, nsimd, { __VA_ARGS__ } );	\
+  accelerator_barrier(dummy);
+
+#define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... )	\
+  accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, { __VA_ARGS__ } ); \
+  accelerator_barrier(dummy);
+
+#endif
+
+//////////////////////////////////////////////
+// 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
+
+#define accelerator 
+#define accelerator_inline strong_inline
+#define accelerator_for(iterator,num,nsimd, ... )   thread_for(iterator, num, { __VA_ARGS__ });
+#define accelerator_forNB(iterator,num,nsimd, ... ) thread_for(iterator, num, { __VA_ARGS__ });
+#define accelerator_barrier(dummy) 
+#define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) thread_for2d(iter1,num1,iter2,num2,{ __VA_ARGS__ });
+
+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);}
+
+#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);};
+inline void acceleratorFreeShared(void *ptr){_mm_free(ptr);};
+inline void acceleratorFreeDevice(void *ptr){_mm_free(ptr);};
+#else
+inline void *acceleratorAllocShared(size_t bytes){return memalign(GRID_ALLOC_ALIGN,bytes);};
+inline void *acceleratorAllocDevice(size_t bytes){return memalign(GRID_ALLOC_ALIGN,bytes);};
+inline void acceleratorFreeShared(void *ptr){free(ptr);};
+inline void acceleratorFreeDevice(void *ptr){free(ptr);};
+#endif
+
+#endif // CPU target
+
+#ifdef HAVE_MM_MALLOC_H
+inline void *acceleratorAllocCpu(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
+inline void acceleratorFreeCpu  (void *ptr){_mm_free(ptr);};
+#else
+inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,bytes);};
+inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
+#endif
+
+///////////////////////////////////////////////////
+// Synchronise across local threads for divergence resynch
+///////////////////////////////////////////////////
+accelerator_inline void acceleratorSynchronise(void) 
+{
+#ifdef GRID_SIMT
+#ifdef GRID_CUDA
+  __syncwarp();
+#endif
+#ifdef GRID_SYCL
+  // No barrier call on SYCL??  // Option get __spir:: stuff to do warp barrier
+#endif
+#ifdef GRID_HIP
+  __syncthreads();
+#endif
+#endif
+  return;
+}
+accelerator_inline void acceleratorSynchroniseAll(void) 
+{
+#ifdef GRID_SIMT
+#ifdef GRID_CUDA
+  __syncthreads();
+#endif
+#ifdef GRID_SYCL
+  // No barrier call on SYCL??  // Option get __spir:: stuff to do warp barrier
+#endif
+#ifdef GRID_HIP
+  __syncthreads();
+#endif
+#endif
+  return;
+}
+accelerator_inline void acceleratorFence(void) 
+{
+#ifdef GRID_SIMT
+#ifdef GRID_CUDA
+  __threadfence();
+#endif
+#ifdef GRID_SYCL
+  // FIXMEE
+#endif
+#ifdef GRID_HIP
+  __threadfence();
+#endif
+#endif
+  return;
+}
+
+NAMESPACE_END(Grid);
diff --git a/Grid/threads/Pragmas.h b/Grid/threads/Pragmas.h
index d05f8ee9..f311bf9b 100644
--- a/Grid/threads/Pragmas.h
+++ b/Grid/threads/Pragmas.h
@@ -2,7 +2,7 @@
 
     Grid physics library, www.github.com/paboyle/Grid 
 
-    Source file: ./lib/Threads.h
+    Source file: ./lib/Pragmas.h
 
     Copyright (C) 2015
 
@@ -28,107 +28,5 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once
 
-#ifndef MAX
-#define MAX(x,y) ((x)>(y)?(x):(y))
-#define MIN(x,y) ((x)>(y)?(y):(x))
-#endif
-
-#define strong_inline     __attribute__((always_inline)) inline
-#define UNROLL  _Pragma("unroll")
-
-//////////////////////////////////////////////////////////////////////////////////
-// New primitives; explicit host thread calls, and accelerator data parallel calls
-//////////////////////////////////////////////////////////////////////////////////
-
-#ifdef _OPENMP
-#define GRID_OMP
-#include <omp.h>
-#endif
-
-#ifdef GRID_OMP
-#define DO_PRAGMA_(x) _Pragma (#x)
-#define DO_PRAGMA(x) DO_PRAGMA_(x)
-#define thread_num(a) omp_get_thread_num()
-#define thread_max(a) omp_get_max_threads()
-#else 
-#define DO_PRAGMA_(x) 
-#define DO_PRAGMA(x) 
-#define thread_num(a) (0)
-#define thread_max(a) (1)
-#endif
-
-#define thread_for( i, num, ... )                           DO_PRAGMA(omp parallel for schedule(static)) for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-#define thread_foreach( i, container, ... )                 DO_PRAGMA(omp parallel for schedule(static)) for ( uint64_t i=container.begin();i<container.end();i++) { __VA_ARGS__ } ;
-#define thread_for_in_region( i, num, ... )                 DO_PRAGMA(omp for schedule(static))          for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-#define thread_for_collapse2( i, num, ... )                 DO_PRAGMA(omp parallel for collapse(2))      for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-#define thread_for_collapse( N , i, num, ... )              DO_PRAGMA(omp parallel for collapse ( N ) )  for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-#define thread_for_collapse_in_region( N , i, num, ... )    DO_PRAGMA(omp for collapse ( N ))            for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-#define thread_region                                       DO_PRAGMA(omp parallel)
-#define thread_critical                                     DO_PRAGMA(omp critical)
-
-
-//////////////////////////////////////////////////////////////////////////////////
-// Accelerator primitives; fall back to threading
-//////////////////////////////////////////////////////////////////////////////////
-#ifdef __NVCC__
-#define GRID_NVCC
-#endif
-
-#ifdef GRID_NVCC
-
-extern uint32_t gpu_threads;
-
-#define accelerator        __host__ __device__
-#define accelerator_inline __host__ __device__ inline
-
-template<typename lambda>  __global__
-void LambdaApplySIMT(uint64_t Isites, uint64_t Osites, lambda Lambda)
-{
-  uint64_t isite = threadIdx.y;
-  uint64_t osite = threadIdx.x+blockDim.x*blockIdx.x;
-  if ( (osite <Osites) && (isite<Isites) ) {
-    Lambda(isite,osite);
-  }
-}
-
-/////////////////////////////////////////////////////////////////
-// Internal only really... but need to call when 
-/////////////////////////////////////////////////////////////////
-#define accelerator_barrier(dummy)				\
-  {								\
-    cudaDeviceSynchronize();					\
-    cudaError err = cudaGetLastError();				\
-    if ( cudaSuccess != err ) {					\
-      printf("Cuda error %s \n", cudaGetErrorString( err )); \
-      puts(__FILE__); \
-      printf("Line %d\n",__LINE__);					\
-      exit(0);							\
-    }								\
-  }
-
-// Copy the for_each_n style ; Non-blocking variant
-#define accelerator_forNB( iterator, num, nsimd, ... )			\
-  {									\
-    typedef uint64_t Iterator;						\
-    auto lambda = [=] accelerator (Iterator lane,Iterator iterator) mutable { \
-      __VA_ARGS__;							\
-    };									\
-    dim3 cu_threads(gpu_threads,nsimd);					\
-    dim3 cu_blocks ((num+gpu_threads-1)/gpu_threads);			\
-    LambdaApplySIMT<<<cu_blocks,cu_threads>>>(nsimd,num,lambda);	\
-  }
-
-// Copy the for_each_n style ; Non-blocking variant (default
-#define accelerator_for( iterator, num, nsimd, ... )		\
-  accelerator_forNB(iterator, num, nsimd, { __VA_ARGS__ } );	\
-  accelerator_barrier(dummy);
-
-#else
-
-#define accelerator 
-#define accelerator_inline strong_inline
-#define accelerator_for(iterator,num,nsimd, ... )   thread_for(iterator, num, { __VA_ARGS__ });
-#define accelerator_forNB(iterator,num,nsimd, ... ) thread_for(iterator, num, { __VA_ARGS__ });
-#define accelerator_barrier(dummy) 
-
-#endif
+#include <Grid/threads/Threads.h>
+#include <Grid/threads/Accelerator.h>
diff --git a/Grid/threads/ThreadReduction.h b/Grid/threads/ThreadReduction.h
new file mode 100644
index 00000000..f0d24d50
--- /dev/null
+++ b/Grid/threads/ThreadReduction.h
@@ -0,0 +1,127 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/ThreadReduction.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once 
+
+// Introduce a class to gain deterministic bit reproducible reduction.
+// make static; perhaps just a namespace is required.
+NAMESPACE_BEGIN(Grid);
+
+class GridThread {
+public:
+  static int _threads;
+  static int _hyperthreads;
+  static int _cores;
+
+  static void SetCores(int cr) { 
+#ifdef GRID_OMP
+    _cores = cr;
+#else 
+    _cores = 1;
+#endif
+  }
+  static void SetThreads(int thr) { 
+#ifdef GRID_OMP
+    _threads = MIN(thr,omp_get_max_threads()) ;
+    omp_set_num_threads(_threads);
+#else 
+    _threads = 1;
+#endif
+  };
+  static void SetMaxThreads(void) { 
+#ifdef GRID_OMP
+    _threads = omp_get_max_threads();
+    omp_set_num_threads(_threads);
+#else 
+    _threads = 1;
+#endif
+  };
+  static int GetHyperThreads(void) { assert(_threads%_cores ==0); return _threads/_cores; };
+  static int GetCores(void)   { return _cores; };
+  static int GetThreads(void) { return _threads; };
+  static int SumArraySize(void) {return _threads;};
+
+  static void GetWork(int nwork, int me, int & mywork, int & myoff){
+    GetWork(nwork,me,mywork,myoff,_threads);
+  }
+  static void GetWork(int nwork, int me, int & mywork, int & myoff,int units){
+    int basework = nwork/units;
+    int backfill = units-(nwork%units);
+    if ( me >= units ) { 
+      mywork = myoff = 0;
+    } else { 
+      mywork = (nwork+me)/units;
+      myoff  = basework * me;
+      if ( me > backfill ) 
+	myoff+= (me-backfill);
+    }
+    return;
+  };
+
+  static void GetWorkBarrier(int nwork, int &me, int & mywork, int & myoff){
+    me     = ThreadBarrier();
+    GetWork(nwork,me,mywork,myoff);
+  };
+
+  static int  ThreadBarrier(void) {
+#ifdef GRID_OMP
+#pragma omp barrier
+    return omp_get_thread_num();
+#else
+    return 0;
+#endif
+  };
+  
+  template<class obj> static void ThreadSum( std::vector<obj> &sum_array,obj &val,int me){
+    sum_array[me] = val;
+    val=Zero();
+    ThreadBarrier();
+    for(int i=0;i<_threads;i++) val+= sum_array[i];
+    ThreadBarrier();
+  }
+
+  static void bcopy(const void *src, void *dst, size_t len) {
+#ifdef GRID_OMP
+#pragma omp parallel 
+    {
+      const char *c_src =(char *) src;
+      char *c_dest=(char *) dst;
+      int me,mywork,myoff;
+      GridThread::GetWorkBarrier(len,me, mywork,myoff);
+      bcopy(&c_src[myoff],&c_dest[myoff],mywork);
+    }
+#else 
+    bcopy(src,dst,len);
+#endif
+  }
+
+
+};
+
+NAMESPACE_END(Grid);
+
diff --git a/Grid/threads/Threads.h b/Grid/threads/Threads.h
index 29cae060..a9fa13ea 100644
--- a/Grid/threads/Threads.h
+++ b/Grid/threads/Threads.h
@@ -28,101 +28,47 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once 
 
+#ifndef MAX
+#define MAX(x,y) ((x)>(y)?(x):(y))
+#define MIN(x,y) ((x)>(y)?(y):(x))
+#endif
 
-// Introduce a class to gain deterministic bit reproducible reduction.
-// make static; perhaps just a namespace is required.
-NAMESPACE_BEGIN(Grid);
+#define strong_inline     __attribute__((always_inline)) inline
+#define UNROLL  _Pragma("unroll")
 
-class GridThread {
-public:
-  static int _threads;
-  static int _hyperthreads;
-  static int _cores;
+//////////////////////////////////////////////////////////////////////////////////
+// New primitives; explicit host thread calls, and accelerator data parallel calls
+//////////////////////////////////////////////////////////////////////////////////
+
+#ifdef _OPENMP
+#define GRID_OMP
+#include <omp.h>
+#endif
 
-  static void SetCores(int cr) { 
 #ifdef GRID_OMP
-    _cores = cr;
+#define DO_PRAGMA_(x) _Pragma (#x)
+#define DO_PRAGMA(x) DO_PRAGMA_(x)
+#define thread_num(a) omp_get_thread_num()
+#define thread_max(a) omp_get_max_threads()
 #else 
-    _cores = 1;
+#define DO_PRAGMA_(x) 
+#define DO_PRAGMA(x) 
+#define thread_num(a) (0)
+#define thread_max(a) (1)
 #endif
-  }
-  static void SetThreads(int thr) { 
-#ifdef GRID_OMP
-    _threads = MIN(thr,omp_get_max_threads()) ;
-    omp_set_num_threads(_threads);
-#else 
-    _threads = 1;
-#endif
-  };
-  static void SetMaxThreads(void) { 
-#ifdef GRID_OMP
-    _threads = omp_get_max_threads();
-    omp_set_num_threads(_threads);
-#else 
-    _threads = 1;
-#endif
-  };
-  static int GetHyperThreads(void) { assert(_threads%_cores ==0); return _threads/_cores; };
-  static int GetCores(void)   { return _cores; };
-  static int GetThreads(void) { return _threads; };
-  static int SumArraySize(void) {return _threads;};
 
-  static void GetWork(int nwork, int me, int & mywork, int & myoff){
-    GetWork(nwork,me,mywork,myoff,_threads);
-  }
-  static void GetWork(int nwork, int me, int & mywork, int & myoff,int units){
-    int basework = nwork/units;
-    int backfill = units-(nwork%units);
-    if ( me >= units ) { 
-      mywork = myoff = 0;
-    } else { 
-      mywork = (nwork+me)/units;
-      myoff  = basework * me;
-      if ( me > backfill ) 
-	myoff+= (me-backfill);
-    }
-    return;
-  };
-
-  static void GetWorkBarrier(int nwork, int &me, int & mywork, int & myoff){
-    me     = ThreadBarrier();
-    GetWork(nwork,me,mywork,myoff);
-  };
-
-  static int  ThreadBarrier(void) {
-#ifdef GRID_OMP
-#pragma omp barrier
-    return omp_get_thread_num();
-#else
-    return 0;
-#endif
-  };
-  
-  template<class obj> static void ThreadSum( std::vector<obj> &sum_array,obj &val,int me){
-    sum_array[me] = val;
-    val=Zero();
-    ThreadBarrier();
-    for(int i=0;i<_threads;i++) val+= sum_array[i];
-    ThreadBarrier();
-  }
-
-  static void bcopy(const void *src, void *dst, size_t len) {
-#ifdef GRID_OMP
-#pragma omp parallel 
-    {
-      const char *c_src =(char *) src;
-      char *c_dest=(char *) dst;
-      int me,mywork,myoff;
-      GridThread::GetWorkBarrier(len,me, mywork,myoff);
-      bcopy(&c_src[myoff],&c_dest[myoff],mywork);
-    }
-#else 
-    bcopy(src,dst,len);
-#endif
-  }
-
-
-};
-
-NAMESPACE_END(Grid);
+#define thread_for( i, num, ... )                           DO_PRAGMA(omp parallel for schedule(static)) for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
+#define thread_for2d( i1, n1,i2,n2, ... )  \
+  DO_PRAGMA(omp parallel for collapse(2))  \
+  for ( uint64_t i1=0;i1<n1;i1++) {	   \
+  for ( uint64_t i2=0;i2<n2;i2++) {	   \
+  { __VA_ARGS__ } ;			   \
+  }}
+#define thread_foreach( i, container, ... )                 DO_PRAGMA(omp parallel for schedule(static)) for ( uint64_t i=container.begin();i<container.end();i++) { __VA_ARGS__ } ;
+#define thread_for_in_region( i, num, ... )                 DO_PRAGMA(omp for schedule(static))          for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
+#define thread_for_collapse2( i, num, ... )                 DO_PRAGMA(omp parallel for collapse(2))      for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
+#define thread_for_collapse( N , i, num, ... )              DO_PRAGMA(omp parallel for collapse ( N ) )  for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
+#define thread_for_collapse_in_region( N , i, num, ... )    DO_PRAGMA(omp for collapse ( N ))            for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
+#define thread_region                                       DO_PRAGMA(omp parallel)
+#define thread_critical                                     DO_PRAGMA(omp critical)
 
diff --git a/Grid/util/Coordinate.h b/Grid/util/Coordinate.h
index c319a14d..7f1d31c0 100644
--- a/Grid/util/Coordinate.h
+++ b/Grid/util/Coordinate.h
@@ -52,14 +52,14 @@ public:
   accelerator_inline size_type size(void) const { return _size; };
   accelerator_inline void  clear(void) { resize(0);}
   accelerator_inline void  resize(size_type sz) {
+#ifndef GRID_HIP
     assert(sz>=0);
     assert(sz<=MaxEntries);
+#endif
     _size = sz;
   }
   accelerator_inline void  resize(size_type sz,const value &val) {
-    assert(sz>=0);
-    assert(sz<=MaxEntries);
-    _size = sz;
+    resize(sz);
     for(int s=0;s<sz;s++) _data[s]=val;
   }
   accelerator_inline pointer begin(void)                   { return &_data[0]; } 
@@ -67,7 +67,7 @@ public:
   accelerator_inline pointer end  (void)                   { return &_data[_size]; } 
   accelerator_inline const_pointer end  (void) const       { return &_data[_size]; } 
   accelerator_inline void push_back(const value &val)      { resize(_size+1); _data[_size-1] = val;}
-  accelerator_inline AcceleratorVector()                   { _size = 0; }
+  accelerator_inline AcceleratorVector()                   { resize(0); }
   accelerator_inline AcceleratorVector(size_type sz)           { resize(sz); }
   accelerator_inline AcceleratorVector(size_type sz,const value &val) { resize(sz,val); }
   AcceleratorVector(const std::vector<value> &copyme) { 
diff --git a/Grid/util/Init.cc b/Grid/util/Init.cc
index 1b672141..6a84a2f2 100644
--- a/Grid/util/Init.cc
+++ b/Grid/util/Init.cc
@@ -73,8 +73,6 @@ feenableexcept (unsigned int excepts)
 }
 #endif
 
-uint32_t gpu_threads=8;
-
 NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////
@@ -192,16 +190,12 @@ void GridParseLayout(char **argv,int argc,
     assert(ompthreads.size()==1);
     GridThread::SetThreads(ompthreads[0]);
   }
-  if( GridCmdOptionExists(argv,argv+argc,"--gpu-threads") ){
+  if( GridCmdOptionExists(argv,argv+argc,"--accelerator-threads") ){
     std::vector<int> gputhreads(0);
-#ifndef GRID_NVCC
-    std::cout << GridLogWarning << "'--gpu-threads' option used but Grid was"
-              << " not compiled with GPU support" << std::endl;
-#endif
-    arg= GridCmdOptionPayload(argv,argv+argc,"--gpu-threads");
+    arg= GridCmdOptionPayload(argv,argv+argc,"--accelerator-threads");
     GridCmdOptionIntVector(arg,gputhreads);
     assert(gputhreads.size()==1);
-    gpu_threads=gputhreads[0];
+    acceleratorThreads(gputhreads[0]);
   }
 
   if( GridCmdOptionExists(argv,argv+argc,"--cores") ){
@@ -241,8 +235,6 @@ static int Grid_is_initialised;
 /////////////////////////////////////////////////////////
 void GridBanner(void)
 {
-  static int printed =0;
-  if( !printed ) {
     std::cout <<std::endl;
     std::cout  << "__|__|__|__|__|__|__|__|__|__|__|__|__|__|__"<<std::endl; 
     std::cout  << "__|__|__|__|__|__|__|__|__|__|__|__|__|__|__"<<std::endl; 
@@ -278,67 +270,6 @@ void GridBanner(void)
     std::cout << "Build " << GRID_BUILD_STR(GRID_BUILD_REF) << std::endl;
 #endif
     std::cout << std::endl;
-    printed=1;
-  }
-}
-#ifdef GRID_NVCC
-cudaDeviceProp *gpu_props;
-#endif
-void GridGpuInit(void)
-{
-#ifdef GRID_NVCC
-  int nDevices = 1;
-  cudaGetDeviceCount(&nDevices);
-  gpu_props = new cudaDeviceProp[nDevices];
-
-  char * localRankStr = NULL;
-  int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_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)
-  {
-    rank = atoi(localRankStr);		
-  }
-  if ((localRankStr = getenv(ENV_LOCAL_RANK_MVAPICH)) != NULL)
-  {
-    rank = atoi(localRankStr);		
-  }
-  if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
-  if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
-
-  if ( world_rank == 0 ) {
-    GridBanner();
-  }
-
-  for (int i = 0; i < nDevices; i++) {
-
-#define GPU_PROP_FMT(canMapHostMemory,FMT)     printf("GpuInit:   " #canMapHostMemory ": " FMT" \n",prop.canMapHostMemory);
-#define GPU_PROP(canMapHostMemory)             GPU_PROP_FMT(canMapHostMemory,"%d");
-    
-    cudaGetDeviceProperties(&gpu_props[i], i);
-    if ( world_rank == 0) {
-      cudaDeviceProp prop; 
-      prop = gpu_props[i];
-      printf("GpuInit: ========================\n");
-      printf("GpuInit: Device Number    : %d\n", i);
-      printf("GpuInit: ========================\n");
-      printf("GpuInit: Device identifier: %s\n", prop.name);
-
-      GPU_PROP(managedMemory);
-      GPU_PROP(isMultiGpuBoard);
-      GPU_PROP(warpSize);
-      //      GPU_PROP(unifiedAddressing);
-      //      GPU_PROP(l2CacheSize);
-      //      GPU_PROP(singleToDoublePrecisionPerfRatio);
-    }
-  }
-  if ( world_rank == 0 ) {
-    printf("GpuInit: ================================================\n");
-  }
-#endif
 }
 
 void Grid_init(int *argc,char ***argv)
@@ -353,9 +284,7 @@ void Grid_init(int *argc,char ***argv)
   //////////////////////////////////////////////////////////
   // Early intialisation necessities without rank knowledge
   //////////////////////////////////////////////////////////
-  GridGpuInit(); // Must come first to set device prior to MPI init
-
-  PointerCache::Init();
+  acceleratorInit(); // Must come first to set device prior to MPI init due to Omnipath Driver
 
   if( GridCmdOptionExists(*argv,*argv+*argc,"--shm") ){
     int MB;
@@ -365,6 +294,14 @@ void Grid_init(int *argc,char ***argv)
     GlobalSharedMemory::MAX_MPI_SHM_BYTES = MB64*1024LL*1024LL;
   }
 
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--device-mem") ){
+    int MB;
+    arg= GridCmdOptionPayload(*argv,*argv+*argc,"--device-mem");
+    GridCmdOptionInt(arg,MB);
+    uint64_t MB64 = MB;
+    MemoryManager::DeviceMaxBytes = MB64*1024LL*1024LL;
+  }
+
   if( GridCmdOptionExists(*argv,*argv+*argc,"--hypercube") ){
     int enable;
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--hypercube");
@@ -381,6 +318,18 @@ void Grid_init(int *argc,char ***argv)
     Grid_debug_handler_init();
   }
 
+#if defined(A64FX)
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-overlap") ){
+    std::cout << "Option --comms-overlap currently not supported on QPACE4. Exiting." << std::endl;
+    exit(EXIT_FAILURE);
+  }
+#endif
+
+  //////////////////////////////////////////////////////////
+  // Memory manager
+  //////////////////////////////////////////////////////////
+  MemoryManager::Init();
+
   //////////////////////////////////////////////////////////
   // MPI initialisation
   //////////////////////////////////////////////////////////
@@ -419,11 +368,18 @@ void Grid_init(int *argc,char ***argv)
   std::cout << GridLogMessage << "MPI is initialised and logging filters activated "<<std::endl;
   std::cout << GridLogMessage << "================================================ "<<std::endl;
 
+
+  /////////////////////////////////////////////////////////
+  // Reporting
+  /////////////////////////////////////////////////////////
   std::cout << GridLogMessage << "Requested "<< GlobalSharedMemory::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
   if ( GlobalSharedMemory::Hugepages) {
     std::cout << GridLogMessage << "Mapped stencil comms buffers as MAP_HUGETLB "<<std::endl;
   }
 
+#ifndef GRID_UVM
+  std::cout << GridLogMessage << "MemoryManager Cache "<< MemoryManager::DeviceMaxBytes <<" bytes "<<std::endl;
+#endif
 
   if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-mem") ){
     MemoryProfiler::debug = true;
diff --git a/SVE_README.txt b/SVE_README.txt
new file mode 100644
index 00000000..0c167c4a
--- /dev/null
+++ b/SVE_README.txt
@@ -0,0 +1,89 @@
+* gcc 10.1 prebuild, QPACE4 interactive login w/ MPI
+
+scl enable gcc-toolset-10 bash
+module load mpi/openmpi-aarch64
+
+../configure --enable-simd=A64FX --enable-comms=mpi3 --enable-shm=shmget CXX=mpicxx CC=mpicc
+
+
+================================== deprecated ================================================
+
+* gcc 10.1 prebuild, QPACE4 interactive login
+
+scl enable gcc-toolset-10 bash
+
+../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-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-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-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-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
+
+export OMPI_CC=gcc-10.0.1
+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-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-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-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=$HOME/local --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-precision=double --enable-comms=none --enable-openmp CXX=armclang++ CC=armclang CXXFLAGS="-std=c++11 -mcpu=a64fx -DARMCLANGCOMPAT -DA64FX -DA64FXASM -DDSLASHINTRIN -DTOFU"
+
+TODO check ARMCLANGCOMPAT
+
+
+* armclang 20.1 VLA w/MPI (fjt cluster)
+
+../configure --with-lime=$HOME/local --without-hdf5 --enable-gen-simd-width=64 --enable-simd=GEN --enable-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 ?
+
+--------------------------------------------------------
+
+* Fujitsu fcc
+
+../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-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"
diff --git a/benchmarks/Benchmark_ITT.cc b/benchmarks/Benchmark_ITT.cc
index 1bb77aff..dc09549c 100644
--- a/benchmarks/Benchmark_ITT.cc
+++ b/benchmarks/Benchmark_ITT.cc
@@ -237,9 +237,9 @@ public:
 
       Vec rn ; random(sRNG,rn);
 
-      LatticeVec z(&Grid); z=rn;
-      LatticeVec x(&Grid); x=rn;
-      LatticeVec y(&Grid); y=rn;
+      LatticeVec z(&Grid); z=Zero();
+      LatticeVec x(&Grid); x=Zero();
+      LatticeVec y(&Grid); y=Zero();
       double a=2.0;
 
       uint64_t Nloop=NLOOP;
@@ -247,9 +247,9 @@ public:
       double start=usecond();
       for(int i=0;i<Nloop;i++){
 	z=a*x-y;
-	auto x_v = x.View();
-	auto y_v = y.View();
-	auto z_v = z.View();
+	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];
       }
diff --git a/benchmarks/Benchmark_dwf.cc b/benchmarks/Benchmark_dwf.cc
index 4d6b026f..2ef5921d 100644
--- a/benchmarks/Benchmark_dwf.cc
+++ b/benchmarks/Benchmark_dwf.cc
@@ -1,5 +1,5 @@
  /*************************************************************************************
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
     Source file: ./benchmarks/Benchmark_dwf.cc
     Copyright (C) 2015
 
@@ -21,7 +21,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #define CUDA_PROFILE
 #endif
 
@@ -77,7 +77,7 @@ int main (int argc, char ** argv)
 
   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;
@@ -107,8 +107,8 @@ int main (int argc, char ** argv)
   LatticeFermion    err(FGrid);
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
-  LatticeGaugeField Umu(UGrid); 
-  SU3::HotConfiguration(RNG4,Umu); 
+  LatticeGaugeField Umu(UGrid);
+  SU3::HotConfiguration(RNG4,Umu);
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
 #if 0
   Umu=1.0;
@@ -126,11 +126,11 @@ int main (int argc, char ** argv)
   // Naive wilson implementation
   ////////////////////////////////////
   // replicate across fifth dimension
-  LatticeGaugeField Umu5d(FGrid); 
+  LatticeGaugeField Umu5d(FGrid);
   std::vector<LatticeColourMatrix> U(4,FGrid);
   {
-    auto Umu5d_v = Umu5d.View();
-    auto Umu_v = Umu.View();
+    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];
@@ -197,24 +197,38 @@ int main (int argc, char ** argv)
     }
     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;
-    err = ref-result; 
+    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) ) { 
+    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);
     }
@@ -235,7 +249,7 @@ int main (int argc, char ** argv)
     }
     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;
 
@@ -243,7 +257,7 @@ int main (int argc, char ** argv)
     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;
-    err = ref-result; 
+    err = ref-result;
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
 
     assert (norm2(err)< 1.0e-3 );
@@ -258,8 +272,8 @@ int main (int argc, char ** argv)
       //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
       tmp = U[mu]*Cshift(src,mu+1,1);
       {
-	auto ref_v = ref.View();
-	auto tmp_v = tmp.View();
+	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]; ;
 	}
@@ -268,8 +282,8 @@ int main (int argc, char ** argv)
       tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu+1,-1);
       {
-	auto ref_v = ref.View();
-	auto tmp_v = tmp.View();
+	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]; ;
 	}
@@ -283,12 +297,14 @@ int main (int argc, char ** argv)
   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; 
+  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;
+*/
   }
   LatticeFermion src_e (FrbGrid);
   LatticeFermion src_o (FrbGrid);
@@ -335,7 +351,7 @@ int main (int argc, char ** argv)
     }
     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;
 
@@ -355,12 +371,14 @@ int main (int argc, char ** argv)
   setCheckerboard(r_eo,r_o);
   setCheckerboard(r_eo,r_e);
 
-  err = r_eo-result; 
+  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);
@@ -373,4 +391,3 @@ int main (int argc, char ** argv)
   Grid_finalize();
   exit(0);
 }
-
diff --git a/benchmarks/Benchmark_dwf_sweep.cc b/benchmarks/Benchmark_dwf_sweep.cc
index 3f108b16..2f13b667 100644
--- a/benchmarks/Benchmark_dwf_sweep.cc
+++ b/benchmarks/Benchmark_dwf_sweep.cc
@@ -130,11 +130,13 @@ void benchDw(std::vector<int> & latt4, int Ls, int threads,int report )
   LatticeGaugeField Umu5d(FGrid); 
 
   // replicate across fifth dimension
-  auto Umu5d_v = Umu5d.View();
-  auto Umu_v   = Umu.View();
-  for(int ss=0;ss<Umu.Grid()->oSites();ss++){
-    for(int s=0;s<Ls;s++){
-      Umu5d_v[Ls*ss+s] = Umu_v[ss];
+  {
+    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];
+      }
     }
   }
 
diff --git a/benchmarks/Benchmark_memory_asynch.cc b/benchmarks/Benchmark_memory_asynch.cc
index a29d281b..97825144 100644
--- a/benchmarks/Benchmark_memory_asynch.cc
+++ b/benchmarks/Benchmark_memory_asynch.cc
@@ -79,7 +79,7 @@ int main (int argc, char ** argv)
 
     double start=usecond();
     thread_for(t,threads,{
-      auto x_t = x[t].View();
+      autoView( x_t , x[t],CpuRead);
       sum[t] = x_t[0];
       for(int i=0;i<Nloop;i++){
 	for(auto ss=x_t.begin();ss<x_t.end();ss++){
diff --git a/benchmarks/Benchmark_memory_bandwidth.cc b/benchmarks/Benchmark_memory_bandwidth.cc
index 9b921a77..10fd48a1 100644
--- a/benchmarks/Benchmark_memory_bandwidth.cc
+++ b/benchmarks/Benchmark_memory_bandwidth.cc
@@ -177,9 +177,7 @@ int main (int argc, char ** argv)
       Real nn;      
       double start=usecond();
       for(int i=0;i<Nloop;i++){
-	auto x_v = x.View();
 	nn=norm2(x);
-	vsplat(x_v[0]._internal[0],nn);
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
diff --git a/benchmarks/Benchmark_meson_field.cc b/benchmarks/Benchmark_meson_field.cc
index d37acb7f..cca33716 100644
--- a/benchmarks/Benchmark_meson_field.cc
+++ b/benchmarks/Benchmark_meson_field.cc
@@ -85,11 +85,11 @@ void sliceInnerProductMesonField(std::vector< std::vector<ComplexD> > &mat,
       for(int b=0;b<e2;b++){
 	int ss= so+n*stride+b;
 	for(int i=0;i<Lblock;i++){
-	  auto lhs_v = lhs[i].View();
+	  autoView(lhs_v, lhs[i], CpuRead);
 	  auto left = conjugate(lhs_v[ss]);
 	  for(int j=0;j<Rblock;j++){
 	    int idx = i+Lblock*j+Lblock*Rblock*r;
-	    auto rhs_v = rhs[j].View();
+	    autoView(rhs_v, rhs[j], CpuRead);
 	    auto right = rhs_v[ss];
 	    vector_type vv = left()(0)(0) * right()(0)(0)
 	      +              left()(0)(1) * right()(0)(1)
@@ -221,12 +221,12 @@ void sliceInnerProductMesonFieldGamma(std::vector< std::vector<ComplexD> > &mat,
       for(int b=0;b<e2;b++){
 	int ss= so+n*stride+b;
 	for(int i=0;i<Lblock;i++){
- 	  auto lhs_v=lhs[i].View();
+ 	  autoView(lhs_v,lhs[i],CpuRead);
 	  auto left = conjugate(lhs_v[ss]);
 	  for(int j=0;j<Rblock;j++){
 	  for(int mu=0;mu<Ngamma;mu++){
 	    
-	    auto rhs_v = rhs[j].View();
+	    autoView(rhs_v,rhs[j],CpuRead);
 	    auto right = Gamma(gammas[mu])*rhs_v[ss];
 
 	      vector_type vv = left()(0)(0) * right()(0)(0)
@@ -370,12 +370,12 @@ void sliceInnerProductMesonFieldGamma1(std::vector< std::vector<ComplexD> > &mat
 	int ss= so+n*stride+b;
 	for(int i=0;i<Lblock;i++){
 	  
-	  auto lhs_v=lhs[i].View();
+	  autoView(lhs_v,lhs[i],CpuRead);
 	  auto left = conjugate(lhs_v[ss]);
 	  for(int j=0;j<Rblock;j++){
 
 	    SpinMatrix_v vv;
-	    auto rhs_v = rhs[j].View();
+	    autoView(rhs_v,rhs[j],CpuRead);
 	    auto right = rhs_v[ss];
 	    for(int s1=0;s1<Ns;s1++){
 	    for(int s2=0;s2<Ns;s2++){
@@ -518,12 +518,12 @@ void sliceInnerProductMesonFieldGammaMom(std::vector< std::vector<ComplexD> > &m
 
 	for(int i=0;i<Lblock;i++){
 
-	  auto lhs_v = lhs[i].View();
+	  autoView(lhs_v,lhs[i],CpuRead);
 	  auto left = conjugate(lhs_v[ss]);
 	  for(int j=0;j<Rblock;j++){
 
 	    SpinMatrix_v vv;
-	    auto rhs_v = rhs[j].View();
+	    autoView(rhs_v,rhs[j],CpuRead);
 	    auto right = rhs_v[ss];
 	    for(int s1=0;s1<Ns;s1++){
 	    for(int s2=0;s2<Ns;s2++){
@@ -537,7 +537,7 @@ void sliceInnerProductMesonFieldGammaMom(std::vector< std::vector<ComplexD> > &m
 	    // Trigger unroll
 	    for ( int m=0;m<Nmom;m++){
 	      int idx = m+base;
-	      auto mom_v = mom[m].View();
+	      autoView(mom_v,mom[m],CpuRead);
 	      auto phase = mom_v[ss];
 	      mac(&lvSum[idx],&vv,&phase);
 	    }
diff --git a/benchmarks/Benchmark_staggered.cc b/benchmarks/Benchmark_staggered.cc
index 17b73c57..34e1e470 100644
--- a/benchmarks/Benchmark_staggered.cc
+++ b/benchmarks/Benchmark_staggered.cc
@@ -106,10 +106,7 @@ int main (int argc, char ** argv)
   
   std::cout<<GridLogMessage << "Called Ds"<<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;
-  err = ref-result; 
-  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
 
   Grid_finalize();
 }
diff --git a/benchmarks/Benchmark_staggeredF.cc b/benchmarks/Benchmark_staggeredF.cc
new file mode 100644
index 00000000..f7beed2d
--- /dev/null
+++ b/benchmarks/Benchmark_staggeredF.cc
@@ -0,0 +1,114 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_staggered.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>
+
+using namespace std;
+using namespace Grid;
+ ;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd,vComplexF::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALF"<< sizeof(RealF)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALD"<< sizeof(RealD)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REAL"<< sizeof(Real)<<std::endl;
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);
+  pRNG.SeedFixedIntegers(seeds);
+  //  pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+
+  typedef typename ImprovedStaggeredFermionF::FermionField FermionField; 
+  typename ImprovedStaggeredFermionF::ImplParams params; 
+
+  FermionField src   (&Grid); random(pRNG,src);
+  FermionField result(&Grid); result=Zero();
+  FermionField    ref(&Grid);    ref=Zero();
+  FermionField    tmp(&Grid);    tmp=Zero();
+  FermionField    err(&Grid);    tmp=Zero();
+  LatticeGaugeFieldF Umu(&Grid); random(pRNG,Umu);
+  std::vector<LatticeColourMatrixF> U(4,&Grid);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
+  // Only one non-zero (y)
+#if 0
+  Umu=Zero();
+  Complex cone(1.0,0.0);
+  for(int nn=0;nn<Nd;nn++){
+    random(pRNG,U[nn]);
+    if(1) {
+      if (nn!=2) { U[nn]=Zero(); std::cout<<GridLogMessage << "zeroing gauge field in dir "<<nn<<std::endl; }
+      //      else       { U[nn]= cone;std::cout<<GridLogMessage << "unit gauge field in dir "<<nn<<std::endl; }
+      else       { std::cout<<GridLogMessage << "random gauge field in dir "<<nn<<std::endl; }
+    }
+    PokeIndex<LorentzIndex>(Umu,U[nn],nn);
+  }
+#endif
+
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+
+  RealD mass=0.1;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionF Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0,params);
+  
+  std::cout<<GridLogMessage << "Calling Ds"<<std::endl;
+  int ncall=1000;
+  for(int i=0;i<ncall;i++){
+    Ds.Dhop(src,result,0);
+  }
+  double t0=usecond();
+  for(int i=0;i<ncall;i++){
+    Ds.Dhop(src,result,0);
+  }
+  double t1=usecond();
+  double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 +  == 1146
+  
+  std::cout<<GridLogMessage << "Called Ds"<<std::endl;
+  std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+
+  Grid_finalize();
+}
diff --git a/benchmarks/Benchmark_su3_gpu.cc b/benchmarks/Benchmark_su3_gpu.cc
index c87f501e..93592ca2 100644
--- a/benchmarks/Benchmark_su3_gpu.cc
+++ b/benchmarks/Benchmark_su3_gpu.cc
@@ -41,7 +41,7 @@ int main (int argc, char ** argv)
 #define LADD (8)
 
   int64_t Nwarm=20;
-  int64_t Nloop=500;
+  int64_t Nloop=50;
 
   Coordinate simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   Coordinate mpi_layout  = GridDefaultMpi();
@@ -66,9 +66,9 @@ int main (int argc, char ** argv)
       LatticeColourMatrix x(&Grid);// random(pRNG,x);
       LatticeColourMatrix y(&Grid);// random(pRNG,y);
 
-      auto x_v = x.View();
-      auto y_v = y.View();
-      auto z_v = z.View();
+      autoView( x_v , x, AcceleratorRead);
+      autoView( y_v , y, AcceleratorRead);
+      autoView( z_v , z, AcceleratorWrite);
       const uint64_t Nsite = x_v.size();
       const uint64_t nsimd = vComplex::Nsimd();
       for(int64_t i=0;i<Nwarm;i++){
@@ -116,9 +116,9 @@ int main (int argc, char ** argv)
       LatticeColourMatrix x(&Grid);// random(pRNG,x);
       LatticeColourMatrix y(&Grid);// random(pRNG,y);
 
-      auto x_v = x.View();
-      auto y_v = y.View();
-      auto z_v = z.View();
+      autoView( x_v , x, AcceleratorWrite);
+      autoView( y_v , y, AcceleratorRead);
+      autoView( z_v , z, AcceleratorRead);
       const uint64_t Nsite = x_v.size();
       const uint64_t nsimd = vComplex::Nsimd();
       for(int64_t i=0;i<Nwarm;i++){
@@ -167,9 +167,9 @@ int main (int argc, char ** argv)
       LatticeColourMatrix x(&Grid);// random(pRNG,x);
       LatticeColourMatrix y(&Grid);// random(pRNG,y);
 
-      auto x_v = x.View();
-      auto y_v = y.View();
-      auto z_v = z.View();
+      autoView( x_v , x, AcceleratorRead);
+      autoView( y_v , y, AcceleratorRead);
+      autoView( z_v , z, AcceleratorWrite);
       const uint64_t Nsite = x_v.size();
       const uint64_t nsimd = vComplex::Nsimd();
       for(int64_t i=0;i<Nwarm;i++){
@@ -220,10 +220,10 @@ int main (int argc, char ** argv)
       LatticeColourMatrix y(&Grid);// random(pRNG,y);
       LatticeColourMatrix w(&Grid);// random(pRNG,y);
 
-      auto x_v = x.View();
-      auto y_v = y.View();
-      auto z_v = z.View();
-      auto w_v = z.View();
+      autoView( x_v , x, AcceleratorRead);
+      autoView( y_v , y, AcceleratorRead);
+      autoView( z_v , z, AcceleratorRead);
+      autoView( w_v , w, AcceleratorWrite);
       const uint64_t Nsite = x_v.size();
       const uint64_t nsimd = vComplex::Nsimd();
       for(int64_t i=0;i<Nwarm;i++){
diff --git a/benchmarks/Benchmark_wilson.cc b/benchmarks/Benchmark_wilson.cc
index 49939620..b834e887 100644
--- a/benchmarks/Benchmark_wilson.cc
+++ b/benchmarks/Benchmark_wilson.cc
@@ -1,6 +1,6 @@
     /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./benchmarks/Benchmark_wilson.cc
 
@@ -98,7 +98,7 @@ int main (int argc, char ** argv)
   double volume=1;
   for(int mu=0;mu<Nd;mu++){
     volume=volume*latt_size[mu];
-  }  
+  }
 
   // Only one non-zero (y)
 #if 0
@@ -118,15 +118,15 @@ int main (int argc, char ** argv)
   for(int mu=0;mu<Nd;mu++){
     U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
   }
-  
+
   { // Naive wilson 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);
       {
-	auto ref_v = ref.View();
-	auto tmp_v = tmp.View();
+	autoView( ref_v, ref, CpuWrite);
+	autoView( tmp_v, tmp, CpuWrite);
 	for(int i=0;i<ref_v.size();i++){
 	  ref_v[i]+= tmp_v[i] - Gamma(Gmu[mu])*tmp_v[i]; ;
 	}
@@ -135,8 +135,8 @@ int main (int argc, char ** argv)
       tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu,-1);
       {
-	auto ref_v = ref.View();
-	auto tmp_v = tmp.View();
+	autoView( ref_v, ref, CpuWrite);
+	autoView( tmp_v, tmp, CpuWrite);
 	for(int i=0;i<ref_v.size();i++){
 	  ref_v[i]+= tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]; ;
 	}
@@ -146,22 +146,32 @@ int main (int argc, char ** argv)
   ref = -0.5*ref;
   RealD mass=0.1;
 
-  typename WilsonFermionR::ImplParams params; 
+  typename WilsonFermionR::ImplParams params;
 
   WilsonFermionR Dw(Umu,Grid,RBGrid,mass,params);
-  
+
   std::cout<<GridLogMessage << "Calling Dw"<<std::endl;
   int ncall=1000;
+  //int ncall=1;
+
+  // Counters
+  Dw.ZeroCounters();
+  Grid.Barrier();
+
   double t0=usecond();
   for(int i=0;i<ncall;i++){
     Dw.Dhop(src,result,0);
   }
+
+  // Counters
+  Grid.Barrier();
+
   double t1=usecond();
   double flops=single_site_flops*volume*ncall;
-  
+
   if (perfProfiling){
   std::cout<<GridLogMessage << "Profiling Dw with perf"<<std::endl;
-    
+
   System::profile("kernel", [&]() {
     for(int i=0;i<ncall;i++){
       Dw.Dhop(src,result,0);
@@ -173,22 +183,35 @@ int main (int argc, char ** argv)
 
   }
 
-  
+  auto nsimd = vComplex::Nsimd();
+  auto simdwidth = sizeof(vComplex);
+
+  std::cout<<GridLogMessage << "Nsimd "<< nsimd << std::endl;
+  std::cout<<GridLogMessage << "Simd width "<< simdwidth << std::endl;
+
+  // RF: Nd Wilson, Nd gauge, Nc colors
+  double data = volume * ((2*Nd+1)*Nd*Nc + 2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
+
   std::cout<<GridLogMessage << "Called Dw"<<std::endl;
   std::cout<<GridLogMessage << "flops per site " << single_site_flops << 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;
-  err = ref-result; 
+  std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data/(t1-t0)<<std::endl;
+  err = ref-result;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
 
+  Dw.Report();
+  
+  // guard
+  double err0 = norm2(err);
 
   //  for(int ss=0;ss<10;ss++ ){
   for(int ss=0;ss<0;ss++ ){
     for(int i=0;i<Ns;i++){
       for(int j=0;j<Nc;j++){
-	auto ref_v = ref.View();
-	auto result_v = result.View();
+	autoView( ref_v, ref, CpuWrite);
+	autoView( result_v, result, CpuWrite);
 	ComplexF * ref_p = (ComplexF *)&ref_v[ss]()(i)(j);
 	ComplexF * res_p = (ComplexF *)&result_v[ss]()(i)(j);
 	std::cout<<GridLogMessage << ss<< " "<<i<<" "<<j<<" "<< (*ref_p)<<" " <<(*res_p)<<std::endl;
@@ -204,8 +227,8 @@ int main (int argc, char ** argv)
       //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
       tmp = U[mu]*Cshift(src,mu,1);
       {
-	auto ref_v = ref.View();
-	auto tmp_v = tmp.View();
+	autoView( ref_v, ref, CpuWrite);
+	autoView( tmp_v, tmp, CpuWrite);
 	for(int i=0;i<ref_v.size();i++){
 	  ref_v[i]+= tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]; ;
 	}
@@ -214,8 +237,8 @@ int main (int argc, char ** argv)
       tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu,-1);
       {
-	auto ref_v = ref.View();
-	auto tmp_v = tmp.View();
+	autoView( ref_v, ref, CpuWrite);
+	autoView( tmp_v, tmp, CpuWrite);
 	for(int i=0;i<ref_v.size();i++){
 	  ref_v[i]+= tmp_v[i] - Gamma(Gmu[mu])*tmp_v[i]; ;
 	}
@@ -227,8 +250,13 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
   std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
   std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
-  err = ref-result; 
+  err = ref-result;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
 
+  // guard
+  double err1 = norm2(err);
+  assert(fabs(err0) < 1.0e-3);
+  assert(fabs(err1) < 1.0e-3);
+
   Grid_finalize();
 }
diff --git a/benchmarks/Benchmark_wilson_sweep.cc b/benchmarks/Benchmark_wilson_sweep.cc
index a01a9e1f..f1473b44 100644
--- a/benchmarks/Benchmark_wilson_sweep.cc
+++ b/benchmarks/Benchmark_wilson_sweep.cc
@@ -132,7 +132,10 @@ void bench_wilson (
   for(int i=0; i<ncall; i++) { Dw.Dhop(src,result,dag); }
   double t1    = usecond();
   double flops = single_site_flops * volume * ncall;
+  double data_tp = (volume * 180 * 64 * ncall) / 1000.; // / (1024.*1024.*1024.);
+  //std::cout << flops/(t1-t0) << " (" << data_tp/(t1-t0) << " MB/s)  \t";
   std::cout << flops/(t1-t0) << "\t\t";
+
 }
 
 void bench_wilson_eo (
diff --git a/configure.ac b/configure.ac
index 6cacb813..f0bea6a4 100644
--- a/configure.ac
+++ b/configure.ac
@@ -138,7 +138,7 @@ esac
 ############### SUMMIT JSRUN
 AC_ARG_ENABLE([summit],
     [AC_HELP_STRING([--enable-summit=yes|no], [enable IBMs jsrun resource manager for SUMMIT])],
-    [ac_JSRUN=${enable_summit}], [ac_SUMMIT=no])
+    [ac_SUMMIT=${enable_summit}], [ac_SUMMIT=no])
 case ${ac_SUMMIT} in
     no);;
     yes)
@@ -147,6 +147,41 @@ case ${ac_SUMMIT} in
       AC_DEFINE([GRID_IBM_SUMMIT],[1],[Let JSRUN manage the GPU device allocation]);;
 esac
 
+############### SYCL/CUDA/HIP/none
+AC_ARG_ENABLE([accelerator],
+    [AC_HELP_STRING([--enable-accelerator=cuda|sycl|hip|none], [enable none,cuda,sycl,hip acceleration])],
+    [ac_ACCELERATOR=${enable_accelerator}], [ac_ACCELERATOR=none])
+case ${ac_ACCELERATOR} in
+    cuda)
+      echo CUDA acceleration
+      AC_DEFINE([GRID_CUDA],[1],[Use CUDA offload]);;
+    sycl)
+      echo SYCL acceleration
+      AC_DEFINE([GRID_SYCL],[1],[Use SYCL offload]);;
+    hip)
+      echo HIP acceleration
+      AC_DEFINE([GRID_HIP],[1],[Use HIP offload]);;
+    none)
+      echo NO acceleration
+    ;;
+    *)
+      AC_MSG_ERROR(["Acceleration not suppoorted ${ac_ACCELERATOR}"]);;
+esac
+
+############### UNIFIED MEMORY
+AC_ARG_ENABLE([unified],
+    [AC_HELP_STRING([--enable-unified=yes|no], [enable unified address space for accelerator loops])],
+    [ac_UNIFIED=${enable_unified}], [ac_UNIFIED=yes])
+case ${ac_UNIFIED} in
+    yes)
+      echo Unified memory for accelerator loops
+      AC_DEFINE([GRID_UVM],[1],[Use unified address space]);;
+    no)
+      echo Manual memory copy for accelerator loops;;
+    *)
+      AC_MSG_ERROR(["Unified virtual memory option not suppoorted ${ac_UNIFIED}"]);;
+esac
+
 ############### Intel libraries
 AC_ARG_ENABLE([mkl],
     [AC_HELP_STRING([--enable-mkl=yes|no|prefix], [enable Intel MKL for LAPACK & FFTW])],
@@ -199,7 +234,7 @@ if test "${ac_MKL}x" != "nox"; then
 fi
 
 if test "${ac_IPP}x" != "nox"; then
-     AC_SEARCH_LIBS([ippsCRC32C_8u], [ippdc], 
+     AC_SEARCH_LIBS([ippsCRC32C_8u], [ippdc],
                    [LIBS="${LIBS} -lippdc -lippvm -lipps -lippcore"],
                    [AC_MSG_ERROR("Intel IPP enabled but library not found")])
 fi
@@ -280,10 +315,10 @@ CXXBASE=${CXX}
 CXXTEST=${CXX}
 if echo "${CXX}" | grep -q "nvcc"; then
   CXXTEST="nvcc"
-fi   
+fi
 
-case ${CXXTEST} in 
-  nvcc) 
+case ${CXXTEST} in
+  nvcc)
 #    CXX="nvcc -keep -v -x cu "
 #    CXXLD="nvcc -v -link"
     CXX="${CXXBASE} -x cu "
@@ -294,6 +329,13 @@ case ${CXXTEST} in
        CXXFLAGS="$CXXFLAGS -Xcompiler -fopenmp"
     fi
     ;;
+  hipcc)
+    CXXFLAGS="$CXXFLAGS -Xcompiler -fno-strict-aliasing --expt-extended-lambda --expt-relaxed-constexpr"
+    CXXLD=${CXX}
+    if test $ac_openmp = yes; then
+       CXXFLAGS="$CXXFLAGS -Xcompiler -fopenmp"
+    fi
+    ;;
   *)
     CXXLD=${CXX}
     CXXFLAGS="$CXXFLAGS -fno-strict-aliasing"
@@ -350,6 +392,15 @@ case ${ax_cv_cxx_compiler_vendor} in
                            [generic SIMD vector width (in bytes)])
         SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)"
         SIMD_FLAGS='';;
+      A64FX)
+        case ${ax_cv_cxx_compiler_vendor} in
+          gnu)
+            AC_DEFINE([A64FX],[1],[A64FX / 512-bit SVE VLS])
+            SIMD_FLAGS='-march=armv8.2-a+sve -msve-vector-bits=512 -fno-gcse -DDSLASHINTRIN';;
+          clang)
+            AC_DEFINE([A64FX],[1],[A64FX / 512-bit SVE VLA])
+            SIMD_FLAGS='-mcpu=a64fx -DARMCLANGCOMPAT -DDSLASHINTRIN';;
+        esac;;
       NEONv8)
         AC_DEFINE([NEONV8],[1],[ARMv8 NEON])
         SIMD_FLAGS='-march=armv8-a';;
@@ -428,6 +479,14 @@ case ${ac_SHM} in
 
      shmopen)
      AC_DEFINE([GRID_MPI3_SHMOPEN],[1],[GRID_MPI3_SHMOPEN] )
+     CXXFLAGS_CPY=$CXXFLAGS
+     LDFLAGS_CPY=$LDFLAGS
+     CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
+     LDFLAGS="$AM_LDFLAGS $LDFLAGS"
+     AC_SEARCH_LIBS([shm_unlink], [rt], [],
+                    [AC_MSG_ERROR("no library found for shm_unlink")])
+     CXXFLAGS=$CXXFLAGS_CPY
+     LDFLAGS=$LDFLAGS_CPY
      ;;
 
      shmget)
@@ -594,6 +653,8 @@ compiler version            : ${ax_cv_gxx_version}
 ----- BUILD OPTIONS -----------------------------------
 SIMD                        : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
 Threading                   : ${ac_openmp}
+Acceleration                : ${ac_ACCELERATOR}
+Unified virtual memory      : ${ac_UNIFIED}
 Communications type         : ${comms_type}
 Shared memory allocator     : ${ac_SHM}
 Shared memory mmap path     : ${ac_SHMPATH}
@@ -640,4 +701,3 @@ AC_OUTPUT
 echo ""
 cat grid.configure.summary
 echo ""
-
diff --git a/documentation/Grid.pdf b/documentation/Grid.pdf
index 8b9f2be1..c1475d63 100644
Binary files a/documentation/Grid.pdf and b/documentation/Grid.pdf differ
diff --git a/documentation/conf.py b/documentation/conf.py
index dc055c4f..c45d5e75 100644
--- a/documentation/conf.py
+++ b/documentation/conf.py
@@ -20,7 +20,7 @@
 # -- Project information -----------------------------------------------------
 
 project = 'Grid'
-copyright = '2018, Peter Boyle, Guido Cossu, Antonin Portelli, Azusa Yamaguchi'
+copyright = '2019, Peter Boyle, Guido Cossu, Antonin Portelli, Azusa Yamaguchi'
 author = 'Peter Boyle, Guido Cossu, Antonin Portelli, Azusa Yamaguchi'
 
 # The short X.Y version
diff --git a/tests/Test_dwf_mixedcg_prec.cc b/tests/Test_dwf_mixedcg_prec.cc
index 0b8463d9..be881db9 100644
--- a/tests/Test_dwf_mixedcg_prec.cc
+++ b/tests/Test_dwf_mixedcg_prec.cc
@@ -29,7 +29,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 using namespace std;
 using namespace Grid;
- ;
 
 template<class d>
 struct scal {
@@ -51,6 +50,7 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "::::: NB: to enable a quick bit reproducibility check use the --checksums flag. " << std::endl;
 
+  { 
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
@@ -100,6 +100,8 @@ int main (int argc, char ** argv)
   ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
   CG(HermOpEO,src_o,result_o_2);
 
+  MemoryManager::Print();
+
   LatticeFermionD diff_o(FrbGrid);
   RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
 
@@ -130,7 +132,9 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << " CG checksums "<<std::hex << scidac_csuma << " "<<scidac_csumb<<std::endl;
   }
   #endif
-
+  }
   
+  MemoryManager::Print();
+
   Grid_finalize();
 }
diff --git a/tests/Test_general_stencil.cc b/tests/Test_general_stencil.cc
index f03677d0..74b6fb4f 100644
--- a/tests/Test_general_stencil.cc
+++ b/tests/Test_general_stencil.cc
@@ -107,9 +107,8 @@ int main(int argc, char ** argv)
 	// Implement a stencil code that should agree with cshift!
 	for(int i=0;i<Check.Grid()->oSites();i++){
 	  auto SE = gStencil.GetEntry(0,i);
-	  auto check = Check.View();
-	  auto foo   = Foo.View();
-
+	  autoView(check, Check, CpuWrite);
+	  autoView(  foo, Foo, CpuRead);
 	  // Encapsulate in a general wrapper
 	  check[i] = foo[SE->_offset];                   auto tmp=check[i];
 	  if (SE->_permute & 0x1 ) { permute(check[i],tmp,0); tmp=check[i];}
@@ -147,8 +146,8 @@ int main(int argc, char ** argv)
 	}}}}
 
 	if (nrm > 1.0e-4) {
-	  auto check = Check.View();
-	  auto bar   = Bar.View();
+	  autoView( check , Check, CpuRead);
+	  autoView(   bar ,   Bar, CpuRead);
 	  for(int i=0;i<check.size();i++){
 	    std::cout << i<<" Check "<<check[i]<< "\n"<<i<<" Bar "<<bar[i]<<std::endl;
 	  }
diff --git a/tests/Test_simd.cc b/tests/Test_simd.cc
index 9b0fa02b..468bc982 100644
--- a/tests/Test_simd.cc
+++ b/tests/Test_simd.cc
@@ -101,14 +101,14 @@ public:
 // FIXME still to test:
 //
 //  innerProduct,
-//  norm2, 
+//  norm2,
 //  Reduce,
 //
 //  mac,mult,sub,add, vone,vzero,vcomplex_i, =Zero(),
 //  vset,vsplat,vstore,vstream,vload, scalar*vec, vec*scalar
 //  unary -,
 //  *= , -=, +=
-//  outerproduct, 
+//  outerproduct,
 //  zeroit
 //  permute
 class funcReduce {
@@ -119,12 +119,12 @@ template<class reduce,class scal>   void sfunc(reduce &rr,scal &i1,scal &i2) con
   std::string name(void) const { return std::string("Reduce"); }
 };
 
-template<class scal, class vec,class functor > 
+template<class scal, class vec,class functor >
 void Tester(const functor &func)
 {
   GridSerialRNG          sRNG;
   sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
-  
+
   int Nsimd = vec::Nsimd();
 
   ExtractBuffer<scal> input1(Nsimd);
@@ -172,6 +172,8 @@ void Tester(const functor &func)
   }
   if ( ok==0 ) {
     std::cout<<GridLogMessage << " OK!" <<std::endl;
+  } else {
+    std::cout<<GridLogMessage << " wrong!" <<std::endl;
   }
   assert(ok==0);
 }
@@ -229,17 +231,19 @@ void IntTester(const functor &func)
   }
   if ( ok==0 ) {
     std::cout<<GridLogMessage << " OK!" <<std::endl;
+  } else {
+    std::cout<<GridLogMessage << " wrong!" <<std::endl;
   }
   assert(ok==0);
 }
 
 
-template<class reduced,class scal, class vec,class functor > 
+template<class reduced,class scal, class vec,class functor >
 void ReductionTester(const functor &func)
 {
   GridSerialRNG          sRNG;
   sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
-  
+
   int Nsimd = vec::Nsimd();
 
   ExtractBuffer<scal> input1(Nsimd);
@@ -278,12 +282,14 @@ void ReductionTester(const functor &func)
   }
   if ( ok==0 ) {
     std::cout<<GridLogMessage << " OK!" <<std::endl;
+  } else {
+    std::cout<<GridLogMessage << " wrong!" <<std::endl;
   }
   assert(ok==0);
 }
 
 
-template<class reduced,class scal, class vec,class functor > 
+template<class reduced,class scal, class vec,class functor >
 void IntReductionTester(const functor &func)
 {
   int Nsimd = vec::Nsimd();
@@ -323,6 +329,8 @@ void IntReductionTester(const functor &func)
   }
   if ( ok==0 ) {
     std::cout<<GridLogMessage << " OK!" <<std::endl;
+  } else {
+    std::cout<<GridLogMessage << " wrong!" <<std::endl;
   }
   assert(ok==0);
 }
@@ -333,7 +341,7 @@ public:
   int n;
   funcPermute(int _n) { n=_n;};
   template<class vec>    void operator()(vec &rr,vec &i1,vec &i2) const { permute(rr,i1,n);}
-  template<class scal>   void apply(ExtractBuffer<scal> &rr,ExtractBuffer<scal> &in)  const { 
+  template<class scal>   void apply(ExtractBuffer<scal> &rr,ExtractBuffer<scal> &in)  const {
     int sz=in.size();
     int msk = sz>>(n+1);
     for(int i=0;i<sz;i++){
@@ -351,8 +359,8 @@ public:
   template<class scal>   void apply(ExtractBuffer<scal> &r1,
 				    ExtractBuffer<scal> &r2,
 				    ExtractBuffer<scal> &in1,
-				    ExtractBuffer<scal> &in2)  const 
-  { 
+				    ExtractBuffer<scal> &in2)  const
+  {
     int sz=in1.size();
     int msk = sz>>(n+1);
 
@@ -364,7 +372,7 @@ public:
 
       if  ( (i&msk) == 0 ) { r2[i]=in1[j2];}
       else                 { r2[i]=in2[j2];}
-    }      
+    }
   }
   std::string name(void) const { return std::string("Exchange"); }
 };
@@ -374,7 +382,7 @@ public:
   int n;
   funcRotate(int _n) { n=_n;};
   template<class vec>    void operator()(vec &rr,vec &i1,vec &i2) const { rr=rotate(i1,n);}
-  template<class scal>   void apply(ExtractBuffer<scal> &rr,ExtractBuffer<scal> &in)  const { 
+  template<class scal>   void apply(ExtractBuffer<scal> &rr,ExtractBuffer<scal> &in)  const {
     int sz = in.size();
     for(int i=0;i<sz;i++){
       rr[i] = in[(i+n)%sz];
@@ -384,12 +392,12 @@ public:
 };
 
 
-template<class scal, class vec,class functor > 
+template<class scal, class vec,class functor >
 void PermTester(const functor &func)
 {
   GridSerialRNG          sRNG;
   sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
-  
+
   int Nsimd = vec::Nsimd();
 
   ExtractBuffer<scal> input1(Nsimd);
@@ -425,37 +433,39 @@ void PermTester(const functor &func)
     for(int i=0;i<Nsimd;i++){
       std::cout<< input1[i]<<" ";
     }
-    std::cout <<std::endl; 
+    std::cout <<std::endl;
     for(int i=0;i<Nsimd;i++){
       std::cout<< result[i]<<" ";
     }
-    std::cout <<std::endl; 
+    std::cout <<std::endl;
     for(int i=0;i<Nsimd;i++){
       std::cout<< reference[i]<<" ";
     }
-    std::cout <<std::endl; 
+    std::cout <<std::endl;
     std::cout<<GridLogMessage<< "*****" << std::endl;
   }
   for(int i=0;i<Nsimd;i++){
     if ( abs(reference[i]-result[i])>1.0e-7){
-      std::cout<<GridLogMessage<< "*****" << std::endl;      
+      std::cout<<GridLogMessage<< "*****" << std::endl;
       std::cout<<GridLogMessage<< "["<<i<<"] "<< abs(reference[i]-result[i]) << " " <<reference[i]<< " " << result[i]<<std::endl;
       ok++;
     }
   }
   if ( ok==0 ) {
     std::cout<<GridLogMessage << " OK!" <<std::endl;
+  } else {
+    std::cout<<GridLogMessage << " wrong!" <<std::endl;
   }
   assert(ok==0);
 }
 
 
-template<class scal, class vec,class functor > 
+template<class scal, class vec,class functor >
 void ExchangeTester(const functor &func)
 {
   GridSerialRNG          sRNG;
   sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
-  
+
   int Nsimd = vec::Nsimd();
 
   ExtractBuffer<scal> input1(Nsimd);
@@ -566,7 +576,7 @@ int main (int argc, char ** argv)
     std::cout << " Test {1,2,3,4} " << Test <<std::endl;
   }
   */
-    
+
   GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
   std::vector<int> seeds({1,2,3,4});
 
@@ -742,7 +752,7 @@ int main (int argc, char ** argv)
   for(int r=0;r<vComplexD::Nsimd();r++){
     PermTester<ComplexD,vComplexD>(funcRotate(r));
   }
-  
+
   std::cout<<GridLogMessage << "==================================="<<  std::endl;
   std::cout<<GridLogMessage << "Testing vInteger                   "<<  std::endl;
   std::cout<<GridLogMessage << "==================================="<<  std::endl;
@@ -773,13 +783,13 @@ int main (int argc, char ** argv)
     precisionChange(&DD[0],&F[0],Ndp);
     std::cout << GridLogMessage<<"Double to single";
     for(int i=0;i<Ndp;i++){
-      //      std::cout << "DD["<<i<<"] = "<< DD[i]<<" "<<D[i]<<" "<<DD[i]-D[i] <<std::endl; 
+      //      std::cout << "DD["<<i<<"] = "<< DD[i]<<" "<<D[i]<<" "<<DD[i]-D[i] <<std::endl;
       DD[i] = DD[i] - D[i];
       decltype(innerProduct(DD[0],DD[0])) nrm;
       nrm = innerProduct(DD[i],DD[i]);
       auto tmp = Reduce(nrm);
       //      std::cout << tmp << std::endl;
-      assert( tmp < 1.0e-14 ); 
+      assert( tmp < 1.0e-14 );
     }
     std::cout <<" OK ! "<<std::endl;
 
@@ -788,13 +798,13 @@ int main (int argc, char ** argv)
     precisionChange(&H[0],&D[0],Ndp);
     precisionChange(&DD[0],&H[0],Ndp);
     for(int i=0;i<Ndp;i++){
-      //      std::cout << "DD["<<i<<"] = "<< DD[i]<<" "<<D[i]<<" "<<DD[i]-D[i]<<std::endl; 
+      //      std::cout << "DD["<<i<<"] = "<< DD[i]<<" "<<D[i]<<" "<<DD[i]-D[i]<<std::endl;
       DD[i] = DD[i] - D[i];
       decltype(innerProduct(DD[0],DD[0])) nrm;
       nrm = innerProduct(DD[i],DD[i]);
       auto tmp = Reduce(nrm);
       //      std::cout << tmp << std::endl;
-      assert( tmp < 1.0e-3 ); 
+      assert( tmp < 1.0e-3 );
     }
     std::cout <<" OK ! "<<std::endl;
 
@@ -803,13 +813,13 @@ int main (int argc, char ** argv)
     precisionChange(&H[0] ,&F[0],Nsp);
     precisionChange(&FF[0],&H[0],Nsp);
     for(int i=0;i<Nsp;i++){
-      //      std::cout << "FF["<<i<<"] = "<< FF[i]<<" "<<F[i]<<" "<<FF[i]-F[i]<<std::endl; 
+      //      std::cout << "FF["<<i<<"] = "<< FF[i]<<" "<<F[i]<<" "<<FF[i]-F[i]<<std::endl;
       FF[i] = FF[i] - F[i];
       decltype(innerProduct(FF[0],FF[0])) nrm;
       nrm = innerProduct(FF[i],FF[i]);
       auto tmp = Reduce(nrm);
       //      std::cout << tmp << std::endl;
-      assert( tmp < 1.0e-3 ); 
+      assert( tmp < 1.0e-3 );
     }
     std::cout <<" OK ! "<<std::endl;
 
diff --git a/tests/Test_stencil.cc b/tests/Test_stencil.cc
index 0bd97fea..16a9138a 100644
--- a/tests/Test_stencil.cc
+++ b/tests/Test_stencil.cc
@@ -109,8 +109,8 @@ int main(int argc, char ** argv) {
 	  StencilEntry *SE;
 	  SE = myStencil.GetEntry(permute_type,0,i);
 	  
-	  auto check = Check.View();
-	  auto foo   = Foo.View();
+	  autoView( check , Check, CpuWrite);
+	  autoView( foo   , Foo, CpuRead);
 	  if ( SE->_is_local && SE->_permute )
 	    permute(check[i],foo[SE->_offset],permute_type);
 	  else if (SE->_is_local)
@@ -151,8 +151,8 @@ int main(int argc, char ** argv) {
 	}}}}
 
 	if (nrm > 1.0e-4) {
-	  auto check = Check.View();
-	  auto bar   = Bar.View();
+	  autoView( check , Check, CpuRead);
+	  autoView( bar   , Bar, CpuRead);
 	  for(int i=0;i<check.size();i++){
 	    std::cout << i<<" Check "<<check[i]<< "\n"<<i<<" Bar "<<bar[i]<<std::endl;
 	  }
@@ -210,8 +210,8 @@ int main(int argc, char ** argv) {
 	  SE = EStencil.GetEntry(permute_type,0,i);
 	  //	  std::cout << "Even source "<< i<<" -> " <<SE->_offset << " "<< SE->_is_local<<std::endl;
 
-	  auto ocheck = OCheck.View();
-	  auto efoo   = EFoo.View();
+	  autoView( ocheck , OCheck, CpuWrite);
+	  autoView( efoo   , EFoo, CpuRead);
 	  if ( SE->_is_local && SE->_permute )
 	    permute(ocheck[i],efoo[SE->_offset],permute_type);
 	  else if (SE->_is_local)
@@ -226,8 +226,8 @@ int main(int argc, char ** argv) {
 	  SE = OStencil.GetEntry(permute_type,0,i);
 	  //	  std::cout << "ODD source "<< i<<" -> " <<SE->_offset << " "<< SE->_is_local<<std::endl;
 
-	  auto echeck = ECheck.View();
-	  auto ofoo   = OFoo.View();
+	  autoView( echeck , ECheck, CpuWrite);
+	  autoView( ofoo   , OFoo, CpuRead);
 	  if ( SE->_is_local && SE->_permute )
 	    permute(echeck[i],ofoo[SE->_offset],permute_type);
 	  else if (SE->_is_local)
diff --git a/tests/core/Test_cshift_red_black.cc b/tests/core/Test_cshift_red_black.cc
index 34325072..4fdd5fc0 100644
--- a/tests/core/Test_cshift_red_black.cc
+++ b/tests/core/Test_cshift_red_black.cc
@@ -82,7 +82,7 @@ int main (int argc, char ** argv)
   pickCheckerboard(Odd,Uo,U);
 
   //  std::cout<<GridLogMessage << U<<std::endl;
-  std::cout<<GridLogMessage<< U <<std::endl;
+  //  std::cout<<GridLogMessage<< U <<std::endl;
   std::cout<<GridLogMessage << "U " <<norm2(U)<<std::endl;
   std::cout<<GridLogMessage << "Ue " <<norm2(Ue)<<std::endl;
   std::cout<<GridLogMessage << "Uo " <<norm2(Uo)<<std::endl;
diff --git a/tests/core/Test_cshift_rotate.cc b/tests/core/Test_cshift_rotate.cc
index 9e572b70..80c4813c 100644
--- a/tests/core/Test_cshift_rotate.cc
+++ b/tests/core/Test_cshift_rotate.cc
@@ -69,6 +69,7 @@ int main (int argc, char ** argv)
 
 	ShiftU  = Cshift(U,dir,shift);    // Shift everything
 
+	std::cout<<GridLogMessage<<"Shifted by "<<shift<<" in direction"<<dir<<" checking the AE35 unit" <<std::endl;
 	/*
 	std::cout << "U[0]" << U[0]<<std::endl;
 	std::cout << "U[1]" << U[1]<<std::endl;
diff --git a/tests/core/Test_main.cc b/tests/core/Test_main.cc
index cd2373ce..08752a46 100644
--- a/tests/core/Test_main.cc
+++ b/tests/core/Test_main.cc
@@ -73,7 +73,7 @@ int main(int argc, char **argv) {
     omp_set_num_threads(omp);
 #endif
 
-    for (int lat = 8; lat <= 16; lat += 40) {
+    for (int lat = 16; lat <= 16; lat += 40) {
       std::cout << "Lat " << lat << std::endl;
 
       latt_size[0] = lat;
@@ -159,15 +159,17 @@ int main(int argc, char **argv) {
       LatticeColourMatrix newFoo = Foo; 
       // confirm correctness of copy constructor
       Bar = Foo - newFoo;
-      std::cout << "Copy constructor diff check: "; 
+      std::cout << "Copy constructor diff check: \n"; 
       double test_cc = norm2(Bar);
       if (test_cc < 1e-5){
         std::cout << "OK\n";
-    }
-      else{
+      } else{
+	std::cout << "Foo\n"<<Foo<<std::endl;
+	std::cout << "newFoo\n"<<newFoo<<std::endl;
+	std::cout << "Bar\n"<<Bar<<std::endl;
         std::cout << "fail\n";
         abort();
-    }
+      }
 
       // Norm2 check
       LatticeReal BarReal(&Fine);
diff --git a/tests/core/Test_staggered5D.cc b/tests/core/Test_staggered5D.cc
index 68b8e537..3d175890 100644
--- a/tests/core/Test_staggered5D.cc
+++ b/tests/core/Test_staggered5D.cc
@@ -88,14 +88,15 @@ int main (int argc, char ** argv)
   // replicate across fifth dimension
   ////////////////////////////////////
   LatticeGaugeField Umu5d(FGrid); 
-  auto umu5d = Umu5d.View();
-  auto umu   = Umu.View();
-  for(int ss=0;ss<Umu.Grid()->oSites();ss++){
-    for(int s=0;s<Ls;s++){
-      umu5d[Ls*ss+s] = umu[ss];
+  {
+    autoView(umu5d, Umu5d, CpuWrite);
+    autoView(  umu, Umu  , CpuRead);
+    for(int ss=0;ss<Umu.Grid()->oSites();ss++){
+      for(int s=0;s<Ls;s++){
+	umu5d[Ls*ss+s] = umu[ss];
+      }
     }
   }
-
   std::vector<LatticeColourMatrix> U(4,FGrid);
 
   for(int mu=0;mu<Nd;mu++){
diff --git a/tests/core/Test_staggered_naive.cc b/tests/core/Test_staggered_naive.cc
new file mode 100644
index 00000000..f96bac93
--- /dev/null
+++ b/tests/core/Test_staggered_naive.cc
@@ -0,0 +1,282 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_wilson.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>
+
+using namespace std;
+using namespace Grid;
+ ;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALF"<< sizeof(RealF)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALD"<< sizeof(RealD)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REAL"<< sizeof(Real)<<std::endl;
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);
+  pRNG.SeedFixedIntegers(seeds);
+  //  pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+
+  typedef typename NaiveStaggeredFermionR::FermionField FermionField; 
+  typedef typename NaiveStaggeredFermionR::ComplexField ComplexField; 
+  typename NaiveStaggeredFermionR::ImplParams params; 
+
+  FermionField src   (&Grid); random(pRNG,src);
+  FermionField result(&Grid); result=Zero();
+  FermionField    ref(&Grid);    ref=Zero();
+  FermionField    tmp(&Grid);    tmp=Zero();
+  FermionField    err(&Grid);    tmp=Zero();
+  FermionField phi   (&Grid); random(pRNG,phi);
+  FermionField chi   (&Grid); random(pRNG,chi);
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+  std::vector<LatticeColourMatrix> U(4,&Grid);
+
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
+  // Only one non-zero (y)
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  /* Debug force unit
+    U[mu] = 1.0;
+    PokeIndex<LorentzIndex>(Umu,U[mu],mu);
+  */
+  }
+
+  ref = Zero();
+
+  RealD mass=0.1;
+  RealD c1=9.0/8.0;
+  RealD u0=1.0;
+
+  { // Simple improved staggered implementation
+    ref = Zero();
+    RealD c1tad = 0.5*c1/u0;
+
+    Lattice<iScalar<vInteger> > coor(&Grid);
+
+    Lattice<iScalar<vInteger> > x(&Grid); LatticeCoordinate(x,0);
+    Lattice<iScalar<vInteger> > y(&Grid); LatticeCoordinate(y,1);
+    Lattice<iScalar<vInteger> > z(&Grid); LatticeCoordinate(z,2);
+    Lattice<iScalar<vInteger> > t(&Grid); LatticeCoordinate(t,3);
+
+    Lattice<iScalar<vInteger> > lin_z(&Grid); lin_z=x+y;
+    Lattice<iScalar<vInteger> > lin_t(&Grid); lin_t=x+y+z;
+
+    for(int mu=0;mu<Nd;mu++){
+
+      // Staggered Phase.
+      ComplexField phases(&Grid);	phases=1.0;
+  
+      if ( mu == 1 ) phases = where( mod(x    ,2)==(Integer)0, phases,-phases);
+      if ( mu == 2 ) phases = where( mod(lin_z,2)==(Integer)0, phases,-phases);
+      if ( mu == 3 ) phases = where( mod(lin_t,2)==(Integer)0, phases,-phases);
+
+      tmp = PeriodicBC::CovShiftForward(U[mu],mu,src);
+      ref = ref +c1tad*tmp*phases; // Forward 1 hop
+
+      tmp = PeriodicBC::CovShiftBackward(U[mu],mu,src);
+      ref = ref -c1tad*tmp*phases; // Backward 1 hop
+
+    }
+    //    ref = ref + mass * src;
+  }
+
+  NaiveStaggeredFermionR Ds(Umu,Grid,RBGrid,mass,c1,u0,params);
+  
+
+  std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Testing Dhop against cshift implementation         "<<std::endl;
+  std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
+
+  std::cout<<GridLogMessage << "Calling Ds"<<std::endl;
+  int ncall=1000;
+  double t0=usecond();
+  for(int i=0;i<ncall;i++){
+    Ds.Dhop(src,result,0);
+  }
+  double t1=usecond();
+  double t2;
+  double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 +  == 1146
+  
+  std::cout<<GridLogMessage << "Called Ds"<<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;
+
+  err = ref-result; 
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+  std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Testing that Deo + Doe = Dunprec "<<std::endl;
+  std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
+
+  FermionField src_e   (&RBGrid);
+  FermionField src_o   (&RBGrid);
+  FermionField r_e   (&RBGrid);
+  FermionField r_o   (&RBGrid);
+  FermionField r_eo  (&Grid);
+  pickCheckerboard(Even,src_e,src);
+  pickCheckerboard(Odd,src_o,src);
+
+  Ds.Meooe(src_e,r_o);  std::cout<<GridLogMessage<<"Applied Meo"<<std::endl;
+  Ds.Meooe(src_o,r_e);  std::cout<<GridLogMessage<<"Applied Moe"<<std::endl;
+  Ds.Dhop (src,ref,DaggerNo);
+
+  setCheckerboard(r_eo,r_o);
+  setCheckerboard(r_eo,r_e);
+
+  err= ref - r_eo;
+  std::cout<<GridLogMessage << "EO norm diff   "<< norm2(err)<< " "<<norm2(ref)<< " " << norm2(r_eo) <<std::endl;
+
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Test Ddagger is the dagger of D by requiring                "<<std::endl;
+  std::cout<<GridLogMessage<<"=  < phi | Deo | chi > * = < chi | Deo^dag| phi>  "<<std::endl;
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+
+  FermionField chi_e   (&RBGrid);
+  FermionField chi_o   (&RBGrid);
+
+  FermionField dchi_e  (&RBGrid);
+  FermionField dchi_o  (&RBGrid);
+
+  FermionField phi_e   (&RBGrid);
+  FermionField phi_o   (&RBGrid);
+
+  FermionField dphi_e  (&RBGrid);
+  FermionField dphi_o  (&RBGrid);
+
+  pickCheckerboard(Even,chi_e,chi);
+  pickCheckerboard(Odd ,chi_o,chi);
+  pickCheckerboard(Even,phi_e,phi);
+  pickCheckerboard(Odd ,phi_o,phi);
+
+  Ds.Meooe(chi_e,dchi_o);
+  Ds.Meooe(chi_o,dchi_e);
+  Ds.MeooeDag(phi_e,dphi_o);
+  Ds.MeooeDag(phi_o,dphi_e);
+
+  ComplexD pDce = innerProduct(phi_e,dchi_e);
+  ComplexD pDco = innerProduct(phi_o,dchi_o);
+  ComplexD cDpe = innerProduct(chi_e,dphi_e);
+  ComplexD cDpo = innerProduct(chi_o,dphi_o);
+
+  std::cout<<GridLogMessage <<"e "<<pDce<<" "<<cDpe <<std::endl;
+  std::cout<<GridLogMessage <<"o "<<pDco<<" "<<cDpo <<std::endl;
+
+  std::cout<<GridLogMessage <<"pDce - conj(cDpo) "<< pDce-conj(cDpo) <<std::endl;
+  std::cout<<GridLogMessage <<"pDco - conj(cDpe) "<< pDco-conj(cDpe) <<std::endl;
+  std::cout<<GridLogMessage <<"e "<<pDce<<" "<<cDpe <<std::endl;
+  std::cout<<GridLogMessage <<"o "<<pDco<<" "<<cDpo <<std::endl;
+
+  std::cout<<GridLogMessage <<"pDce - conj(cDpo) "<< pDce-conj(cDpo) <<std::endl;
+  std::cout<<GridLogMessage <<"pDco - conj(cDpe) "<< pDco-conj(cDpe) <<std::endl;
+
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Test MeeInv Mee = 1                                         "<<std::endl;
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+
+  pickCheckerboard(Even,chi_e,chi);
+  pickCheckerboard(Odd ,chi_o,chi);
+
+  Ds.Mooee(chi_e,src_e);
+  Ds.MooeeInv(src_e,phi_e);
+
+  Ds.Mooee(chi_o,src_o);
+  Ds.MooeeInv(src_o,phi_o);
+  
+  setCheckerboard(phi,phi_e);
+  setCheckerboard(phi,phi_o);
+
+  err = phi-chi;
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<< std::endl;
+
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Test MeeInvDag MeeDag = 1                                   "<<std::endl;
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+
+  pickCheckerboard(Even,chi_e,chi);
+  pickCheckerboard(Odd ,chi_o,chi);
+
+  Ds.MooeeDag(chi_e,src_e);
+  Ds.MooeeInvDag(src_e,phi_e);
+
+  Ds.MooeeDag(chi_o,src_o);
+  Ds.MooeeInvDag(src_o,phi_o);
+  
+  setCheckerboard(phi,phi_e);
+  setCheckerboard(phi,phi_o);
+
+  err = phi-chi;
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<< std::endl;
+
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Test MpcDagMpc is Hermitian              "<<std::endl;
+  std::cout<<GridLogMessage<<"=============================================================="<<std::endl;
+  
+  random(pRNG,phi);
+  random(pRNG,chi);
+  pickCheckerboard(Even,chi_e,chi);
+  pickCheckerboard(Odd ,chi_o,chi);
+  pickCheckerboard(Even,phi_e,phi);
+  pickCheckerboard(Odd ,phi_o,phi);
+
+  SchurDiagMooeeOperator<NaiveStaggeredFermionR,FermionField> HermOpEO(Ds);
+  HermOpEO.MpcDagMpc(chi_e,dchi_e,t1,t2);
+  HermOpEO.MpcDagMpc(chi_o,dchi_o,t1,t2);
+
+  HermOpEO.MpcDagMpc(phi_e,dphi_e,t1,t2);
+  HermOpEO.MpcDagMpc(phi_o,dphi_o,t1,t2);
+
+  pDce = innerProduct(phi_e,dchi_e);
+  pDco = innerProduct(phi_o,dchi_o);
+  cDpe = innerProduct(chi_e,dphi_e);
+  cDpo = innerProduct(chi_o,dphi_o);
+
+  std::cout<<GridLogMessage <<"e "<<pDce<<" "<<cDpe <<std::endl;
+  std::cout<<GridLogMessage <<"o "<<pDco<<" "<<cDpo <<std::endl;
+
+  std::cout<<GridLogMessage <<"pDce - conj(cDpo) "<< pDco-conj(cDpo) <<std::endl;
+  std::cout<<GridLogMessage <<"pDco - conj(cDpe) "<< pDce-conj(cDpe) <<std::endl;
+
+  Grid_finalize();
+}
diff --git a/tests/debug/Test_cayley_mres.cc b/tests/debug/Test_cayley_mres.cc
index 1bf6f2dc..2ad605b8 100644
--- a/tests/debug/Test_cayley_mres.cc
+++ b/tests/debug/Test_cayley_mres.cc
@@ -570,8 +570,8 @@ void  TestConserved1(Action & Ddwf, Action & Ddwfrev,
       LatticeGaugeField Umu5d(FGrid); 
       std::vector<LatticeColourMatrix> U(4,FGrid);
       {
-	auto Umu5d_v = Umu5d.View();
-	auto Umu_v = Umu.View();
+        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];
@@ -597,8 +597,8 @@ void  TestConserved1(Action & Ddwf, Action & Ddwfrev,
       {
 	RealD diag = 5.0 - Ddwf.M5;
 	mass = Ddwf.mass;
-	auto psi=result5.View();
-	auto chi=tmp.View();
+	autoView( psi,result5,CpuRead);
+	autoView( chi,tmp,   CpuWrite);
 	thread_for(sss,UGrid->oSites(),{
 	  uint64_t ss= sss*Ls;
 	  typedef vSpinColourVector spinor;
diff --git a/tests/forces/Test_contfrac_force.cc b/tests/forces/Test_contfrac_force.cc
index 4eeb8c27..cb30faad 100644
--- a/tests/forces/Test_contfrac_force.cc
+++ b/tests/forces/Test_contfrac_force.cc
@@ -98,9 +98,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto mom_v = mom.View();
-    auto Uprime_v = Uprime.View();
-    auto U_v = U.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
     thread_foreach( i,mom_v,{
       Uprime_v[i](mu) =
 	  U_v[i](mu)
diff --git a/tests/forces/Test_dwf_force.cc b/tests/forces/Test_dwf_force.cc
index 009f50b3..81a1b8c4 100644
--- a/tests/forces/Test_dwf_force.cc
+++ b/tests/forces/Test_dwf_force.cc
@@ -100,9 +100,9 @@ int main (int argc, char ** argv)
 
     // fourth order exponential approx
 
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach( i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/forces/Test_dwf_force_eofa.cc b/tests/forces/Test_dwf_force_eofa.cc
index 670e7589..0b0ba346 100644
--- a/tests/forces/Test_dwf_force_eofa.cc
+++ b/tests/forces/Test_dwf_force_eofa.cc
@@ -110,9 +110,9 @@ int main (int argc, char** argv)
     PokeIndex<LorentzIndex>(mom, mommu, mu);
 
     // fourth order exponential approx
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/forces/Test_dwf_gpforce.cc b/tests/forces/Test_dwf_gpforce.cc
index d762e22a..b39fdd14 100644
--- a/tests/forces/Test_dwf_gpforce.cc
+++ b/tests/forces/Test_dwf_gpforce.cc
@@ -119,9 +119,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/forces/Test_dwf_gpforce_eofa.cc b/tests/forces/Test_dwf_gpforce_eofa.cc
index 66ae9dcf..58258a5e 100644
--- a/tests/forces/Test_dwf_gpforce_eofa.cc
+++ b/tests/forces/Test_dwf_gpforce_eofa.cc
@@ -114,9 +114,9 @@ int main (int argc, char** argv)
     PokeIndex<LorentzIndex>(mom, mommu, mu);
 
     // fourth order exponential approx
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/forces/Test_gp_plaq_force.cc b/tests/forces/Test_gp_plaq_force.cc
index c4e214bb..21f0b9d0 100644
--- a/tests/forces/Test_gp_plaq_force.cc
+++ b/tests/forces/Test_gp_plaq_force.cc
@@ -85,9 +85,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto Uprime_v = Uprime.View();
-    auto U_v = U.View();
-    auto mom_v = mom.View();
+    autoView(Uprime_v, Uprime, CpuWrite);
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
     thread_foreach(i,mom_v,{ // exp(pmu dt) * Umu
       Uprime_v[i](mu) = U_v[i](mu) + mom_v[i](mu)*U_v[i](mu)*dt ;
     });
diff --git a/tests/forces/Test_gp_rect_force.cc b/tests/forces/Test_gp_rect_force.cc
index 2573af6a..bb4ea6de 100644
--- a/tests/forces/Test_gp_rect_force.cc
+++ b/tests/forces/Test_gp_rect_force.cc
@@ -87,9 +87,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto mom_v   = mom.View();
-    auto Uprime_v= Uprime.View();
-    auto U_v     = U.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
+    autoView( U_v , U, CpuRead);
     thread_foreach(i,mom_v,{ // exp(pmu dt) * Umu
       Uprime_v[i](mu) = U_v[i](mu) + mom_v[i](mu)*U_v[i](mu)*dt ;
     });
diff --git a/tests/forces/Test_gpdwf_force.cc b/tests/forces/Test_gpdwf_force.cc
index 09a1dc4b..bdc332d9 100644
--- a/tests/forces/Test_gpdwf_force.cc
+++ b/tests/forces/Test_gpdwf_force.cc
@@ -105,9 +105,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto U_v = U.View();
-    auto mom_v = mom.View();
-    auto Uprime_v = Uprime.View();
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) = U_v[i](mu)
 	+ mom_v[i](mu)*U_v[i](mu)*dt 
diff --git a/tests/forces/Test_gpwilson_force.cc b/tests/forces/Test_gpwilson_force.cc
index cd30d898..1c85a5d9 100644
--- a/tests/forces/Test_gpwilson_force.cc
+++ b/tests/forces/Test_gpwilson_force.cc
@@ -99,9 +99,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/forces/Test_mobius_force.cc b/tests/forces/Test_mobius_force.cc
index a1c4e930..11e69652 100644
--- a/tests/forces/Test_mobius_force.cc
+++ b/tests/forces/Test_mobius_force.cc
@@ -101,9 +101,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto U_v      = U.View();
-    auto mom_v    = mom.View();
-    auto Uprime_v = Uprime.View();
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =  U_v[i](mu)
 	+ mom_v[i](mu)*U_v[i](mu)*dt 
diff --git a/tests/forces/Test_mobius_force_eofa.cc b/tests/forces/Test_mobius_force_eofa.cc
index f71e2d41..f85501fa 100644
--- a/tests/forces/Test_mobius_force_eofa.cc
+++ b/tests/forces/Test_mobius_force_eofa.cc
@@ -112,9 +112,9 @@ int main (int argc, char** argv)
     PokeIndex<LorentzIndex>(mom, mommu, mu);
 
     // fourth order exponential approx
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/forces/Test_mobius_gpforce_eofa.cc b/tests/forces/Test_mobius_gpforce_eofa.cc
index 4975f36d..68163e63 100644
--- a/tests/forces/Test_mobius_gpforce_eofa.cc
+++ b/tests/forces/Test_mobius_gpforce_eofa.cc
@@ -115,9 +115,9 @@ int main (int argc, char** argv)
     SU3::GaussianFundamentalLieAlgebraMatrix(RNG4, mommu); // Traceless antihermitian momentum; gaussian in lie alg
 
     PokeIndex<LorentzIndex>(mom, mommu, mu);
-    auto U_v      = U.View();
-    auto mom_v    = mom.View();
-    auto Uprime_v = Uprime.View();
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
     // fourth order exponential approx
     thread_foreach( i, mom_v,{
       Uprime_v[i](mu) = U_v[i](mu) + mom_v[i](mu)*U_v[i](mu)*dt + mom_v[i](mu) *mom_v[i](mu) *U_v[i](mu)*(dt*dt/2.0)
diff --git a/tests/forces/Test_partfrac_force.cc b/tests/forces/Test_partfrac_force.cc
index 3ea2c6aa..17dce530 100644
--- a/tests/forces/Test_partfrac_force.cc
+++ b/tests/forces/Test_partfrac_force.cc
@@ -101,9 +101,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto U_v = U.View();
-    auto mom_v = mom.View();
-    auto Uprime_v = Uprime.View();
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =  U_v[i](mu)
 	+ mom_v[i](mu)*U_v[i](mu)*dt 
diff --git a/tests/forces/Test_rect_force.cc b/tests/forces/Test_rect_force.cc
index 9a78de24..ed72f2c0 100644
--- a/tests/forces/Test_rect_force.cc
+++ b/tests/forces/Test_rect_force.cc
@@ -87,9 +87,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto Uprime_v = Uprime.View();
-    auto U_v      = U.View();
-    auto mom_v    = mom.View();
+    autoView(Uprime_v, Uprime, CpuWrite);
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
     thread_foreach(i,mom_v,{ // exp(pmu dt) * Umu
       Uprime_v[i](mu) = U_v[i](mu) + mom_v[i](mu)*U_v[i](mu)*dt ;
     });
diff --git a/tests/forces/Test_wilson_force.cc b/tests/forces/Test_wilson_force.cc
index 47f1516a..c8b3a7f4 100644
--- a/tests/forces/Test_wilson_force.cc
+++ b/tests/forces/Test_wilson_force.cc
@@ -105,9 +105,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto U_v = U.View();
-    auto mom_v = mom.View();
-    auto Uprime_v = Uprime.View();
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
     thread_foreach( i,mom_v,{
       Uprime_v[i](mu)  =	  U_v[i](mu);
       Uprime_v[i](mu) += mom_v[i](mu)*U_v[i](mu)*dt ;
diff --git a/tests/forces/Test_wilsonclover_force.cc b/tests/forces/Test_wilsonclover_force.cc
index d9ace23c..f26f0ac9 100644
--- a/tests/forces/Test_wilsonclover_force.cc
+++ b/tests/forces/Test_wilsonclover_force.cc
@@ -105,9 +105,9 @@ int main(int argc, char **argv)
     Hmom -= real(sum(trace(mommu * mommu)));
     PokeIndex<LorentzIndex>(mom, mommu, mu);
 
-    auto Uprime_v = Uprime.View();
-    auto U_v      = U.View();
-    auto mom_v    = mom.View();
+    autoView(Uprime_v, Uprime, CpuWrite);
+    autoView( U_v , U, CpuRead);
+    autoView( mom_v, mom, CpuRead);
     thread_foreach(ss,mom_v,
     {
       Uprime_v[ss]._internal[mu] = ProjectOnGroup(Exponentiate(mom_v[ss]._internal[mu], dt, 12) * U_v[ss]._internal[mu]);
diff --git a/tests/forces/Test_zmobius_force.cc b/tests/forces/Test_zmobius_force.cc
index 2730885f..e24ae601 100644
--- a/tests/forces/Test_zmobius_force.cc
+++ b/tests/forces/Test_zmobius_force.cc
@@ -114,9 +114,9 @@ int main (int argc, char ** argv)
     PokeIndex<LorentzIndex>(mom,mommu,mu);
 
     // fourth order exponential approx
-    auto mom_v    = mom.View();
-    auto U_v      = U.View();
-    auto Uprime_v = Uprime.View();
+    autoView( mom_v, mom, CpuRead);
+    autoView( U_v , U, CpuRead);
+    autoView(Uprime_v, Uprime, CpuWrite);
 
     thread_foreach(i,mom_v,{
       Uprime_v[i](mu) =	  U_v[i](mu)
diff --git a/tests/hmc/Test_hmc_WilsonMixedRepresentationsFermionGauge.cc b/tests/hmc/Test_hmc_WilsonMixedRepresentationsFermionGauge.cc
index 6fa90f32..3b8cdda6 100644
--- a/tests/hmc/Test_hmc_WilsonMixedRepresentationsFermionGauge.cc
+++ b/tests/hmc/Test_hmc_WilsonMixedRepresentationsFermionGauge.cc
@@ -35,7 +35,7 @@ directory
 
 
 int main(int argc, char **argv) {
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
   using namespace Grid;
 
 
diff --git a/tests/hmc/Test_multishift_sqrt.cc b/tests/hmc/Test_multishift_sqrt.cc
index f8477220..31697c12 100644
--- a/tests/hmc/Test_multishift_sqrt.cc
+++ b/tests/hmc/Test_multishift_sqrt.cc
@@ -31,7 +31,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 using namespace std;
 using namespace Grid;
- ;
 
 template<class Field> class DumbOperator  : public LinearOperatorBase<Field> {
 public:
@@ -57,7 +56,7 @@ public:
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {};
   void OpDir  (const Field &in, Field &out,int dir,int disp){};
-  void OpDirAll  (const Field &in, std::vector<Field> &out) {}; // Abstract base
+  void OpDirAll  (const Field &in, std::vector<Field> &out) {};
 
   void Op     (const Field &in, Field &out){
     out = scale * in;
@@ -105,7 +104,7 @@ int main (int argc, char ** argv)
 						       GridDefaultMpi());
 
   double     lo=0.001;
-  double     hi=1.0;
+  double     hi=20.0;
   int precision=64;
   int    degree=10;
   AlgRemez remez(lo,hi,precision);
diff --git a/tests/solver/Test_dwf_hdcr.cc b/tests/solver/Test_dwf_hdcr.cc
index 873530ff..8e083231 100644
--- a/tests/solver/Test_dwf_hdcr.cc
+++ b/tests/solver/Test_dwf_hdcr.cc
@@ -1,5 +1,3 @@
-
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -302,8 +300,8 @@ int main (int argc, char ** argv)
     int nb=nbasisc/2;
     CoarseAggregates.CreateSubspaceChebyshev(CRNG,PosdefLdop,nb,12.0,0.02,500,100,100,0.0);
     for(int n=0;n<nb;n++){
-      auto subspace    = CoarseAggregates.subspace[n].View();
-      auto subspace_g5 = CoarseAggregates.subspace[n+nb].View();
+      autoView( subspace,   CoarseAggregates.subspace[n]   ,CpuRead);
+      autoView( subspace_g5,CoarseAggregates.subspace[n+nb],CpuWrite);
       for(int nn=0;nn<nb;nn++){
 	for(int site=0;site<Coarse5d->oSites();site++){
 	  subspace_g5[site](nn)   = subspace[site](nn);
diff --git a/tests/solver/Test_dwf_multigrid.cc b/tests/solver/Test_dwf_multigrid.cc
new file mode 100644
index 00000000..9e11c160
--- /dev/null
+++ b/tests/solver/Test_dwf_multigrid.cc
@@ -0,0 +1,594 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_hdcr.cc
+
+    Copyright (C) 2015
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+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>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+using namespace std;
+using namespace Grid;
+/* Params
+ * Grid: 
+ * block1(4)
+ * block2(4)
+ * 
+ * Subspace
+ * * Fine  : Subspace(nbasis,hi,lo,order,first,step) -- 32, 60,0.02,500,100,100
+ * * Coarse: Subspace(nbasis,hi,lo,order,first,step) -- 32, 18,0.02,500,100,100
+
+ * Smoother:
+ * * Fine: Cheby(hi, lo, order)            --  60,0.5,10
+ * * Coarse: Cheby(hi, lo, order)          --  12,0.1,4
+
+ * Lanczos:
+ * CoarseCoarse IRL( Nk, Nm, Nstop, poly(lo,hi,order))   24,36,24,0.002,4.0,61 
+ */
+
+template<class Field> class SolverWrapper : public LinearFunction<Field> {
+private:
+  LinearOperatorBase<Field> & _Matrix;
+  OperatorFunction<Field> & _Solver;
+  LinearFunction<Field>   & _Guess;
+public:
+
+  /////////////////////////////////////////////////////
+  // Wrap the usual normal equations trick
+  /////////////////////////////////////////////////////
+  SolverWrapper(LinearOperatorBase<Field> &Matrix,
+	      OperatorFunction<Field> &Solver,
+	      LinearFunction<Field> &Guess) 
+   :  _Matrix(Matrix), _Solver(Solver), _Guess(Guess) {}; 
+
+  void operator() (const Field &in, Field &out){
+ 
+    _Guess(in,out);
+    _Solver(_Matrix,in,out);  // Mdag M out = Mdag in
+
+  }     
+};
+
+
+// Must use a non-hermitian solver
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {
+    assert(0);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    assert(0);
+  };
+  void Op     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    assert(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    assert(0);
+  }
+};
+
+
+RealD InverseApproximation(RealD x){
+  return 1.0/x;
+}
+
+template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
+{
+public:
+  typedef LinearOperatorBase<Field>                            FineOperator;
+  Matrix         & _SmootherMatrix;
+  FineOperator   & _SmootherOperator;
+  
+  Chebyshev<Field> Cheby;
+
+  ChebyshevSmoother(RealD _lo,RealD _hi,int _ord, FineOperator &SmootherOperator,Matrix &SmootherMatrix) :
+    _SmootherOperator(SmootherOperator),
+    _SmootherMatrix(SmootherMatrix),
+    Cheby(_lo,_hi,_ord,InverseApproximation)
+  {};
+
+  void operator() (const Field &in, Field &out) 
+  {
+    Field tmp(in.Grid());
+    MdagMLinearOperator<Matrix,Field>   MdagMOp(_SmootherMatrix); 
+    _SmootherOperator.AdjOp(in,tmp);
+    Cheby(MdagMOp,tmp,out);         
+  }
+};
+
+template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
+{
+public:
+  typedef LinearOperatorBase<Field>                            FineOperator;
+  Matrix         & SmootherMatrix;
+  FineOperator   & SmootherOperator;
+  RealD tol;
+  RealD shift;
+  int   maxit;
+
+  MirsSmoother(RealD _shift,RealD _tol,int _maxit,FineOperator &_SmootherOperator,Matrix &_SmootherMatrix) :
+    shift(_shift),tol(_tol),maxit(_maxit),
+    SmootherOperator(_SmootherOperator),
+    SmootherMatrix(_SmootherMatrix)
+  {};
+
+  void operator() (const Field &in, Field &out) 
+  {
+    ZeroGuesser<Field> Guess;
+    ConjugateGradient<Field>  CG(tol,maxit,false);
+ 
+    Field src(in.Grid());
+
+    ShiftedMdagMLinearOperator<SparseMatrixBase<Field>,Field> MdagMOp(SmootherMatrix,shift);
+    SmootherOperator.AdjOp(in,src);
+    Guess(src,out);
+    CG(MdagMOp,src,out); 
+  }
+};
+
+#define GridLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level <<" "
+
+template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
+class HDCRPreconditioner : public LinearFunction< Lattice<Fobj> > {
+public:
+
+  typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
+  typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::FineField    FineField;
+  typedef LinearOperatorBase<FineField>                            FineOperator;
+  typedef LinearFunction    <FineField>                            FineSmoother;
+
+  Aggregates     & _Aggregates;
+  FineOperator   & _FineOperator;
+  FineSmoother   & _Smoother;
+  CoarseSolver   & _CoarseSolve;
+
+  int    level;  void Level(int lv) {level = lv; };
+
+
+  HDCRPreconditioner(Aggregates &Agg,
+		     FineOperator &Fine,
+		     FineSmoother &Smoother,
+		     CoarseSolver &CoarseSolve_)
+    : _Aggregates(Agg),
+      _FineOperator(Fine),
+      _Smoother(Smoother),
+      _CoarseSolve(CoarseSolve_),
+      level(1)  {  }
+
+  virtual void operator()(const FineField &in, FineField & out) 
+  {
+    auto CoarseGrid = _Aggregates.CoarseGrid;
+    CoarseVector Csrc(CoarseGrid);
+    CoarseVector Csol(CoarseGrid);
+    FineField vec1(in.Grid());
+    FineField vec2(in.Grid());
+
+    double t;
+    // Fine Smoother
+    t=-usecond();
+    _Smoother(in,out);
+    t+=usecond();
+    GridLogLevel << "Smoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Update the residual
+    _FineOperator.Op(out,vec1);  sub(vec1, in ,vec1);   
+
+    // Fine to Coarse 
+    t=-usecond();
+    _Aggregates.ProjectToSubspace  (Csrc,vec1);
+    t+=usecond();
+    GridLogLevel << "Project to coarse took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse correction
+    t=-usecond();
+    _CoarseSolve(Csrc,Csol);
+    t+=usecond();
+    GridLogLevel << "Coarse solve took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse to Fine
+    t=-usecond();
+    _Aggregates.PromoteFromSubspace(Csol,vec1); 
+    add(out,out,vec1);
+    t+=usecond();
+    GridLogLevel << "Promote to this level took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Residual
+    _FineOperator.Op(out,vec1);  sub(vec1 ,in , vec1);  
+
+    // Fine Smoother
+    t=-usecond();
+    _Smoother(vec1,vec2);
+    t+=usecond();
+    GridLogLevel << "Smoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    add( out,out,vec2);
+  }
+};
+
+/*
+template<class Fobj,class CComplex,int nbasis, class Guesser, class CoarseSolver>
+class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
+public:
+
+  typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
+  typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::FineField    FineField;
+  typedef LinearOperatorBase<FineField>                            FineOperator;
+  typedef LinearFunction    <FineField>                            FineSmoother;
+
+  Aggregates     & _Aggregates;
+  CoarseOperator & _CoarseOperator;
+  FineOperator   & _FineOperator;
+  Guesser        & _Guess;
+  FineSmoother   & _Smoother;
+  CoarseSolver   & _CoarseSolve;
+
+  int    level;  void Level(int lv) {level = lv; };
+
+
+  MultiGridPreconditioner(Aggregates &Agg, CoarseOperator &Coarse, 
+			  FineOperator &Fine,
+			  FineSmoother &Smoother,
+			  Guesser &Guess_,
+			  CoarseSolver &CoarseSolve_)
+    : _Aggregates(Agg),
+      _CoarseOperator(Coarse),
+      _FineOperator(Fine),
+      _Smoother(Smoother),
+      _Guess(Guess_),
+      _CoarseSolve(CoarseSolve_),
+      level(1)  {  }
+
+  virtual void operator()(const FineField &in, FineField & out) 
+  {
+    CoarseVector Csrc(_CoarseOperator.Grid());
+    CoarseVector Csol(_CoarseOperator.Grid()); 
+    FineField vec1(in.Grid());
+    FineField vec2(in.Grid());
+
+    double t;
+    // Fine Smoother
+    t=-usecond();
+    _Smoother(in,out);
+    t+=usecond();
+    GridLogLevel << "Smoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Update the residual
+    _FineOperator.Op(out,vec1);  sub(vec1, in ,vec1);   
+
+    // Fine to Coarse 
+    t=-usecond();
+    _Aggregates.ProjectToSubspace  (Csrc,vec1);
+    t+=usecond();
+    GridLogLevel << "Project to coarse took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse correction
+    t=-usecond();
+    _CoarseSolve(Csrc,Csol);
+    t+=usecond();
+    GridLogLevel << "Coarse solve took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse to Fine
+    t=-usecond();
+    _Aggregates.PromoteFromSubspace(Csol,vec1); 
+    add(out,out,vec1);
+    t+=usecond();
+    GridLogLevel << "Promote to this level took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Residual
+    _FineOperator.Op(out,vec1);  sub(vec1 ,in , vec1);  
+
+    // Fine Smoother
+    t=-usecond();
+    _Smoother(vec1,vec2);
+    t+=usecond();
+    GridLogLevel << "Smoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    add( out,out,vec2);
+  }
+};
+*/
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=16;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  ///////////////////////////////////////////////////
+  // Construct a coarsened grid; utility for this?
+  ///////////////////////////////////////////////////
+  std::vector<int> block ({2,2,2,2});
+  std::vector<int> blockc ({2,2,2,2});
+  const int nbasis= 32;
+  const int nbasisc= 32;
+  auto clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/block[d];
+  }
+  auto cclatt = clatt;
+  for(int d=0;d<clatt.size();d++){
+    cclatt[d] = clatt[d]/blockc[d];
+  }
+
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+  GridCartesian *CoarseCoarse4d =  SpaceTimeGrid::makeFourDimGrid(cclatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
+  GridCartesian *CoarseCoarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,CoarseCoarse4d);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+  LatticeFermion    src(FGrid); gaussian(RNG5,src);// src=src+g5*src;
+  LatticeFermion result(FGrid); 
+  LatticeGaugeField Umu(UGrid); 
+
+  FieldMetaData header;
+  std::string file("./ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Building g5R5 hermitian DWF operator" <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  RealD mass=0.001;
+  RealD M5=1.8;
+  DomainWallFermionR Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionR Dpv (Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
+
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis>              Subspace;
+  typedef CoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>          CoarseOperator;
+  typedef CoarseOperator::CoarseVector                                 CoarseVector;
+  typedef CoarseOperator::siteVector siteVector;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Calling Aggregation class to build subspace" <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  MdagMLinearOperator<DomainWallFermionR,LatticeFermion> HermDefOp(Ddwf);
+
+  Subspace Aggregates(Coarse5d,FGrid,0);
+
+  assert ( (nbasis & 0x1)==0);
+  {
+    int nb=nbasis/2;
+    Aggregates.CreateSubspaceChebyshev(RNG5,HermDefOp,nb,60.0,0.02,500,100,100,0.0);
+    for(int n=0;n<nb;n++){
+      G5R5(Aggregates.subspace[n+nb],Aggregates.subspace[n]);
+    }
+    LatticeFermion A(FGrid);
+    LatticeFermion B(FGrid);
+    for(int n=0;n<nb;n++){
+      A = Aggregates.subspace[n];
+      B = Aggregates.subspace[n+nb];
+      Aggregates.subspace[n]   = A+B; // 1+G5 // eigen value of G5R5 is +1
+      Aggregates.subspace[n+nb]= A-B; // 1-G5 // eigen value of G5R5 is -1
+    }
+  }
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Will coarsen G5R5 M and G5R5 Mpv in G5R5 compatible way " <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  
+  typedef CoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>    Level1Op;
+  typedef CoarsenedMatrix<siteVector,iScalar<vTComplex>,nbasisc> Level2Op;
+
+  Gamma5R5HermitianLinearOperator<DomainWallFermionR,LatticeFermion> HermIndefOp(Ddwf);
+  Gamma5R5HermitianLinearOperator<DomainWallFermionR,LatticeFermion> HermIndefOpPV(Dpv);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Building coarse representation of Indef operator" <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  Level1Op LDOp(*Coarse5d,1);   LDOp.CoarsenOperator(FGrid,HermIndefOp,Aggregates);
+  Level1Op LDOpPV(*Coarse5d,1); LDOpPV.CoarsenOperator(FGrid,HermIndefOpPV,Aggregates);
+
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Testing fine and coarse solvers " <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  CoarseVector c_src(Coarse5d); c_src=1.0;
+  CoarseVector c_res(Coarse5d);
+
+  LatticeFermion f_src(FGrid); f_src=1.0;
+  LatticeFermion f_res(FGrid);
+
+  LatticeFermion f_src_e(FrbGrid); f_src_e=1.0;
+  LatticeFermion f_res_e(FrbGrid);
+
+  RealD tol=1.0e-8;
+  int MaxIt = 10000;
+
+  BiCGSTAB<CoarseVector>                     CoarseBiCGSTAB(tol,MaxIt);
+  ConjugateGradient<CoarseVector>            CoarseCG(tol,MaxIt);
+  //  GeneralisedMinimalResidual<CoarseVector>   CoarseGMRES(tol,MaxIt,20);
+
+  BiCGSTAB<LatticeFermion>                   FineBiCGSTAB(tol,MaxIt);
+  ConjugateGradient<LatticeFermion>          FineCG(tol,MaxIt);
+  //  GeneralisedMinimalResidual<LatticeFermion> FineGMRES(tol,MaxIt,20);
+  
+  MdagMLinearOperator<DomainWallFermionR,LatticeFermion>    FineMdagM(Ddwf);     //  M^\dag M
+  PVdagMLinearOperator<DomainWallFermionR,LatticeFermion>   FinePVdagM(Ddwf,Dpv);//  M_{pv}^\dag M
+  SchurDiagMooeeOperator<DomainWallFermionR,LatticeFermion> FineDiagMooee(Ddwf); //  M_ee - Meo Moo^-1 Moe 
+  SchurDiagOneOperator<DomainWallFermionR,LatticeFermion>   FineDiagOne(Ddwf);   //  1 - M_ee^{-1} Meo Moo^{-1} Moe e
+
+   MdagMLinearOperator<Level1Op,CoarseVector> CoarseMdagM(LDOp);
+  PVdagMLinearOperator<Level1Op,CoarseVector> CoarsePVdagM(LDOp,LDOpPV);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Fine CG unprec "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  f_res=Zero();
+  FineCG(FineMdagM,f_src,f_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Fine CG prec DiagMooee "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  
+  f_res_e=Zero();
+  FineCG(FineDiagMooee,f_src_e,f_res_e);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Fine CG prec DiagOne "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  f_res_e=Zero();
+  FineCG(FineDiagOne,f_src_e,f_res_e);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Fine BiCGSTAB unprec "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  f_res=Zero();
+  FineBiCGSTAB(FinePVdagM,f_src,f_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Coarse BiCGSTAB "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  
+  c_res=Zero();
+  CoarseBiCGSTAB(CoarsePVdagM,c_src,c_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Coarse CG unprec "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  c_res=Zero();
+  CoarseCG(CoarseMdagM,c_src,c_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Running Coarse grid Lanczos "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  Chebyshev<CoarseVector>      IRLCheby(0.03,12.0,71);  // 1 iter
+  FunctionHermOp<CoarseVector> IRLOpCheby(IRLCheby,CoarseMdagM);
+  PlainHermOp<CoarseVector>    IRLOp    (CoarseMdagM);
+  int Nk=64;
+  int Nm=128;
+  int Nstop=Nk;
+  ImplicitlyRestartedLanczos<CoarseVector> IRL(IRLOpCheby,IRLOp,Nstop,Nk,Nm,1.0e-3,20);
+
+  int Nconv;
+  std::vector<RealD>            eval(Nm);
+  std::vector<CoarseVector>     evec(Nm,Coarse5d);
+  IRL.calc(eval,evec,c_src,Nconv);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Running Coarse grid deflated solver              "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  DeflatedGuesser<CoarseVector> DeflCoarseGuesser(evec,eval);
+  NormalEquations<CoarseVector> DeflCoarseCGNE   (LDOp,CoarseCG,DeflCoarseGuesser);
+  c_res=Zero();
+  DeflCoarseCGNE(c_src,c_res);
+
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Running HDCR                                     "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  ConjugateGradient<CoarseVector>  CoarseMgridCG(0.001,1000);     
+  ChebyshevSmoother<LatticeFermion,DomainWallFermionR> FineSmoother(0.5,60.0,10,HermIndefOp,Ddwf);
+
+  typedef HDCRPreconditioner<vSpinColourVector,  vTComplex,nbasis, NormalEquations<CoarseVector> >   TwoLevelHDCR;
+  TwoLevelHDCR TwoLevelPrecon(Aggregates,
+			      HermIndefOp,
+			      FineSmoother,
+			      DeflCoarseCGNE);
+  TwoLevelPrecon.Level(1);
+  //  PrecGeneralisedConjugateResidual<LatticeFermion> l1PGCR(1.0e-8,100,HermIndefOp,TwoLevelPrecon,16,16);
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> l1PGCR(1.0e-8,100,HermIndefOp,TwoLevelPrecon,16,16);
+  l1PGCR.Level(1);
+
+  f_res=Zero();
+
+  CoarseCG.Tolerance=0.02;
+  l1PGCR(f_src,f_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Running Multigrid                                "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  BiCGSTAB<CoarseVector>    CoarseMgridBiCGSTAB(0.01,1000);     
+  BiCGSTAB<LatticeFermion>  FineMgridBiCGSTAB(0.0,24);
+  ZeroGuesser<CoarseVector> CoarseZeroGuesser;
+  ZeroGuesser<LatticeFermion> FineZeroGuesser;
+
+  SolverWrapper<LatticeFermion> FineBiCGSmoother(  FinePVdagM,  FineMgridBiCGSTAB,  FineZeroGuesser);
+  SolverWrapper<CoarseVector> CoarsePVdagMSolver(CoarsePVdagM,CoarseMgridBiCGSTAB,CoarseZeroGuesser);
+  typedef HDCRPreconditioner<vSpinColourVector, vTComplex,nbasis, SolverWrapper<CoarseVector> >  TwoLevelMG;
+
+  TwoLevelMG _TwoLevelMG(Aggregates,
+			 FinePVdagM,
+			 FineBiCGSmoother,
+			 CoarsePVdagMSolver);
+  _TwoLevelMG.Level(1);
+
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> pvPGCR(1.0e-8,100,FinePVdagM,_TwoLevelMG,16,16);
+  pvPGCR.Level(1);
+
+  f_res=Zero();
+  pvPGCR(f_src,f_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  Grid_finalize();
+  
+}
diff --git a/tests/solver/Test_hw_multigrid.cc b/tests/solver/Test_hw_multigrid.cc
new file mode 100644
index 00000000..b728faa7
--- /dev/null
+++ b/tests/solver/Test_hw_multigrid.cc
@@ -0,0 +1,375 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_hdcr.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>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+//#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+using namespace std;
+using namespace Grid;
+/* Params
+ * Grid: 
+ * block1(4)
+ * block2(4)
+ * 
+ * Subspace
+ * * Fine  : Subspace(nbasis,hi,lo,order,first,step) -- 32, 60,0.02,500,100,100
+ * * Coarse: Subspace(nbasis,hi,lo,order,first,step) -- 32, 18,0.02,500,100,100
+
+ * Smoother:
+ * * Fine: Cheby(hi, lo, order)            --  60,0.5,10
+ * * Coarse: Cheby(hi, lo, order)          --  12,0.1,4
+
+ * Lanczos:
+ * CoarseCoarse IRL( Nk, Nm, Nstop, poly(lo,hi,order))   24,36,24,0.002,4.0,61 
+ */
+
+template<class Field> class SolverWrapper : public LinearFunction<Field> {
+private:
+  LinearOperatorBase<Field> & _Matrix;
+  OperatorFunction<Field> & _Solver;
+  LinearFunction<Field>   & _Guess;
+public:
+
+  /////////////////////////////////////////////////////
+  // Wrap the usual normal equations trick
+  /////////////////////////////////////////////////////
+  SolverWrapper(LinearOperatorBase<Field> &Matrix,
+	      OperatorFunction<Field> &Solver,
+	      LinearFunction<Field> &Guess) 
+   :  _Matrix(Matrix), _Solver(Solver), _Guess(Guess) {}; 
+
+  void operator() (const Field &in, Field &out){
+ 
+    _Guess(in,out);
+    _Solver(_Matrix,in,out);  // Mdag M out = Mdag in
+
+  }     
+};
+
+
+// Must use a non-hermitian solver
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {
+    assert(0);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    assert(0);
+  };
+  void Op     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    assert(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    assert(0);
+  }
+};
+
+
+RealD InverseApproximation(RealD x){
+  return 1.0/x;
+}
+
+template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
+{
+public:
+  typedef LinearOperatorBase<Field>                            FineOperator;
+  Matrix         & _SmootherMatrix;
+  FineOperator   & _SmootherOperator;
+  
+  Chebyshev<Field> Cheby;
+
+  ChebyshevSmoother(RealD _lo,RealD _hi,int _ord, FineOperator &SmootherOperator,Matrix &SmootherMatrix) :
+    _SmootherOperator(SmootherOperator),
+    _SmootherMatrix(SmootherMatrix),
+    Cheby(_lo,_hi,_ord,InverseApproximation)
+  {};
+
+  void operator() (const Field &in, Field &out) 
+  {
+    Field tmp(in.Grid());
+    MdagMLinearOperator<Matrix,Field>   MdagMOp(_SmootherMatrix); 
+    _SmootherOperator.AdjOp(in,tmp);
+    Cheby(MdagMOp,tmp,out);         
+  }
+};
+
+template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
+{
+public:
+  typedef LinearOperatorBase<Field>                            FineOperator;
+  Matrix         & SmootherMatrix;
+  FineOperator   & SmootherOperator;
+  RealD tol;
+  RealD shift;
+  int   maxit;
+
+  MirsSmoother(RealD _shift,RealD _tol,int _maxit,FineOperator &_SmootherOperator,Matrix &_SmootherMatrix) :
+    shift(_shift),tol(_tol),maxit(_maxit),
+    SmootherOperator(_SmootherOperator),
+    SmootherMatrix(_SmootherMatrix)
+  {};
+
+  void operator() (const Field &in, Field &out) 
+  {
+    ZeroGuesser<Field> Guess;
+    ConjugateGradient<Field>  CG(tol,maxit,false);
+ 
+    Field src(in.Grid());
+
+    ShiftedMdagMLinearOperator<SparseMatrixBase<Field>,Field> MdagMOp(SmootherMatrix,shift);
+    SmootherOperator.AdjOp(in,src);
+    Guess(src,out);
+    CG(MdagMOp,src,out); 
+  }
+};
+
+#define GridLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level <<" "
+
+template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
+class HDCRPreconditioner : public LinearFunction< Lattice<Fobj> > {
+public:
+
+  typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
+  typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::FineField    FineField;
+  typedef LinearOperatorBase<FineField>                            FineOperator;
+  typedef LinearFunction    <FineField>                            FineSmoother;
+
+  Aggregates     & _Aggregates;
+  FineOperator   & _FineOperator;
+  FineSmoother   & _Smoother;
+  CoarseSolver   & _CoarseSolve;
+
+  int    level;  void Level(int lv) {level = lv; };
+
+
+  HDCRPreconditioner(Aggregates &Agg,
+		     FineOperator &Fine,
+		     FineSmoother &Smoother,
+		     CoarseSolver &CoarseSolve_)
+    : _Aggregates(Agg),
+      _FineOperator(Fine),
+      _Smoother(Smoother),
+      _CoarseSolve(CoarseSolve_),
+      level(1)  {  }
+
+  virtual void operator()(const FineField &in, FineField & out) 
+  {
+    auto CoarseGrid = _Aggregates.CoarseGrid;
+    CoarseVector Csrc(CoarseGrid);
+    CoarseVector Csol(CoarseGrid);
+    FineField vec1(in.Grid());
+    FineField vec2(in.Grid());
+
+    double t;
+    // Fine Smoother
+    t=-usecond();
+    _Smoother(in,out);
+    t+=usecond();
+    GridLogLevel << "Smoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Update the residual
+    _FineOperator.Op(out,vec1);  sub(vec1, in ,vec1);   
+
+    // Fine to Coarse 
+    t=-usecond();
+    _Aggregates.ProjectToSubspace  (Csrc,vec1);
+    t+=usecond();
+    GridLogLevel << "Project to coarse took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse correction
+    t=-usecond();
+    _CoarseSolve(Csrc,Csol);
+    t+=usecond();
+    GridLogLevel << "Coarse solve took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse to Fine
+    t=-usecond();
+    _Aggregates.PromoteFromSubspace(Csol,vec1); 
+    add(out,out,vec1);
+    t+=usecond();
+    GridLogLevel << "Promote to this level took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Residual
+    _FineOperator.Op(out,vec1);  sub(vec1 ,in , vec1);  
+
+    // Fine Smoother
+    t=-usecond();
+    _Smoother(vec1,vec2);
+    t+=usecond();
+    GridLogLevel << "Smoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    add( out,out,vec2);
+  }
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=16;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  ///////////////////////////////////////////////////
+  // Construct a coarsened grid; utility for this?
+  ///////////////////////////////////////////////////
+  std::vector<int> block ({2,2,2,2});
+  const int nbasis= 8;
+
+  auto clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/block[d];
+  }
+
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(Ls,Coarse4d);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(seeds);
+
+  LatticeGaugeField Umu(UGrid); 
+  FieldMetaData header;
+  std::string file("./ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Building g5R5 hermitian DWF operator" <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  RealD mass=0.001;
+  RealD M5=1.8;
+  WilsonFermionR    Dw(Umu,*UGrid,*UrbGrid,-M5);
+  DomainWallFermionR Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionR Dpv (Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
+
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis>              Subspace;
+  typedef CoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>          CoarseOperator;
+  typedef CoarseOperator::CoarseVector                                 CoarseVector;
+  typedef CoarseOperator::siteVector siteVector;
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Calling Aggregation class to build subspace" <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  MdagMLinearOperator<WilsonFermionR,LatticeFermion> SubspaceOp(Dw);
+
+  Subspace Aggregates4D(Coarse4d,UGrid,0);
+  Subspace Aggregates5D(Coarse5d,FGrid,0);
+
+  assert ( (nbasis & 0x1)==0);
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " 4D subspace build                                " <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  int nb=nbasis/2;
+  Gamma g5(Gamma::Algebra::Gamma5);
+  Aggregates4D.CreateSubspaceChebyshev(RNG4,SubspaceOp,nb,60.0,0.02,500,100,100,0.0);
+  for(int n=0;n<nb;n++){
+    Aggregates4D.subspace[n+nb]= Aggregates4D.subspace[n] - g5 * Aggregates4D.subspace[n];
+    Aggregates4D.subspace[n]   = Aggregates4D.subspace[n] + g5 * Aggregates4D.subspace[n];
+  }
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Promote to 5D basis                              " <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  for(int n=0;n<nbasis;n++){
+    for(int s=0;s<Ls;s++){
+      InsertSlice(Aggregates4D.subspace[n],Aggregates5D.subspace[n],s,0);
+    }
+  }
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Coarsen the operator                          " <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  typedef CoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>    Level1Op;
+
+  NonHermitianLinearOperator<DomainWallFermionR,LatticeFermion>  LinOpDwf(Ddwf);
+
+  Level1Op LDOp  (*Coarse5d,0);   
+  
+  std::cout<<GridLogMessage << " Callinig Coarsen the operator                          " <<std::endl;
+  LDOp.CoarsenOperator(FGrid,LinOpDwf,Aggregates5D);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Coarse CG unprec "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+
+  CoarseVector c_src(Coarse5d); c_src=1.0;
+  CoarseVector c_res(Coarse5d);
+
+  LatticeFermion f_src(FGrid); f_src=1.0;
+  LatticeFermion f_res(FGrid);
+
+  RealD tol=1.0e-8;
+  int MaxIt = 10000;
+
+  MdagMLinearOperator<Level1Op,CoarseVector> CoarseMdagM(LDOp);
+  BiCGSTAB<CoarseVector>                     CoarseBiCGSTAB(tol,MaxIt);
+  ConjugateGradient<CoarseVector>            CoarseCG(tol,MaxIt);
+
+  c_res=Zero();
+  CoarseCG(CoarseMdagM,c_src,c_res);
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << " Solve                                            " <<std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  
+  LatticeFermion    src(FGrid); gaussian(RNG5,src);
+  LatticeFermion result(FGrid); 
+
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+  std::cout<<GridLogMessage << "**************************************************"<< std::endl;
+  Grid_finalize();
+  
+}