Merge branch 'develop' into feature/baryon

Grid/GridCore.h

@@ -47,9 +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/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> 

Grid/GridStd.h

@@ -6,6 +6,7 @@
 ///////////////////
 #include <cassert>
 #include <complex>
+#include <memory>
 #include <vector>
 #include <array>
 #include <string>

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
 

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

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

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 
@@ -91,35 +80,8 @@ 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;
-  };
-  */    
-
 };
   
 template<class Fobj,class CComplex,int nbasis>
@@ -149,24 +111,6 @@ 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) {
 
@@ -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();
+	autoView( y_v , y, AcceleratorWrite);
-	auto Tn_v = Tn->View();
+	autoView( Tn_v , (*Tn), AcceleratorWrite);
-	auto Tnp_v = Tnp->View();
+	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
-	auto Tnm_v = Tnm->View();
+	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
-
 
 };
 
@@ -549,13 +293,13 @@ public:
     SimpleCompressor<siteVector> compressor;
 
     Stencil.HaloExchange(in,compressor);
-
+    autoView( in_v , in, AcceleratorRead);
-    auto in_v = in.View();
+    autoView( out_v , out, AcceleratorWrite);
-    auto out_v = out.View();
     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();
+    autoView( out_v , out, AcceleratorWrite);
-    auto in_v  = in.View();
+    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
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
-    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
   }
   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() ;
+	    autoView( iZProj_v , iZProj, AcceleratorRead) ;
-	    auto oZProj_v = oZProj.View() ;
+	    autoView( oZProj_v , oZProj, AcceleratorRead) ;
-	    auto A_p     =  A[p].View();
+	    autoView( A_p     ,  A[p], AcceleratorWrite);
-	    auto A_self  = A[self_stencil].View();
+	    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();
+	  autoView( tmp_      , tmp, AcceleratorWrite);
-	  auto evenmask_ = evenmask.View();
+	  autoView( evenmask_ , evenmask, AcceleratorRead);
-	  auto oddmask_  =  oddmask.View();
+	  autoView( oddmask_  ,  oddmask, AcceleratorRead);
-	  auto Mphie_    =  Mphie.View();
+	  autoView( Mphie_    ,  Mphie, AcceleratorRead);
-	  auto Mphio_    =  Mphio.View();
+	  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();
+	autoView( SelfProj_ , SelfProj, AcceleratorRead);
-	auto A_self  = A[self_stencil].View();
+	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);

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,p_v,cbuf);
-	  pokeLocalSite(s,pgbuf,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);
+	peekLocalSite(s,pgbuf_v,cgbuf);
-	pokeLocalSite(s,result,clbuf);
+	pokeLocalSite(s,result_v,clbuf);
-    });
+      });
+    }
     result = result*div;
       
     // destroying plan

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();
+	  autoView( p_v , p, AcceleratorWrite);
-        auto v_v = v.View();
+	  autoView( r_v , r, AcceleratorRead);
-        accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
+	  autoView( v_v , v, AcceleratorRead);
-          coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
+	  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();
+	  autoView( p_v , p, AcceleratorRead);
-        accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
+	  autoView( r_v , r, AcceleratorRead);
-          coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
+	  autoView( v_v , v, AcceleratorRead);
-        });
+	  autoView( psi_v,psi, AcceleratorRead);
-        
+	  autoView( h_v  ,  h, AcceleratorWrite);
-        auto s_v = s.View();
+	  autoView( s_v  ,  s, AcceleratorWrite);
-        accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
+	  accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
+	      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();
 
@@ -166,11 +172,17 @@ 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(),{
+	  autoView( psi_v,psi, AcceleratorWrite);
-          coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
+	  autoView( r_v , r, AcceleratorWrite);
-          coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
+	  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);

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();
+	autoView( psi_v , psi, AcceleratorWrite);
-      auto r_v   = r.View();
+	autoView( p_v   , p,   AcceleratorWrite);
-      accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+	autoView( r_v   , r,   AcceleratorWrite);
-	  coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+	accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
-	  coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+	    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();
 

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

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__

Grid/allocator/AlignedAllocator.h

@@ -26,129 +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
+#pragma once
-#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)
 
 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;
-};
-
-#ifdef GRID_NVCC
-#define profilerCudaMeminfo \
-  { size_t f, t ; cudaMemGetInfo ( &f,&t); std::cout << GridLogDebug << "[Memory debug] Cuda free "<<f<<"/"<<t << std::endl;}
-#else
-#define profilerCudaMeminfo
-#endif
-
-#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;						\
-    }									\
-  profilerCudaMeminfo;
-
-#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: 
@@ -172,70 +53,60 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
-
+    _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
-
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-#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 = cudaMallocManaged((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
     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
+  // FIXME: hack for the copy constructor, eventually it must be avoided
-    pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
+  //void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
-#else 
+  void construct(pointer __p, const _Tp& __val) { assert(0);};
-    pointer __freeme = __p;
+  void construct(pointer __p) { };
-#endif
+  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
+template<typename _Tp>
-    if ( __freeme ) cudaFree((void *)__freeme);
+class uvmAllocator {
-#else 
+public: 
-  #ifdef HAVE_MM_MALLOC_H
+  typedef std::size_t     size_type;
-    if ( __freeme ) _mm_free((void *)__freeme); 
+  typedef std::ptrdiff_t  difference_type;
-  #else
+  typedef _Tp*       pointer;
-    if ( __freeme ) free((void *)__freeme);
+  typedef const _Tp* const_pointer;
-  #endif
+  typedef _Tp&       reference;
-#endif
+  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
@@ -244,17 +115,17 @@ public:
   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 uvmAllocator<_Tp>&, const uvmAllocator<_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 false; }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using commAllocator = alignedAllocator<T>;
+template<class T> using commAllocator = uvmAllocator<T>;
-template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
+template<class T> using Vector     = std::vector<T,uvmAllocator<T> >;           
-template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
+template<class T> using commVector = std::vector<T,uvmAllocator<T> >;
-template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
+//template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
 
 NAMESPACE_END(Grid);
 
-#endif
+

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>

Grid/allocator/MemoryManager.cc

@@ -0,0 +1,244 @@
+#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;
+    }
+  }
+  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);
+

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);
+
+

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

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

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);
+

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);
+

Grid/cartesian/Cartesian_base.h

@@ -81,6 +81,7 @@ public:
 
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
+  Coordinate _checker_dim_mask;
 
 public:
 

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];

Grid/cartesian/Cartesian_red_black.h

@@ -36,11 +36,27 @@ 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;
 

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
 
@@ -170,17 +170,24 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
   std::vector<int> primes({2,3,5});
 
   int dim = 0;
+  int last_dim = ndimension - 1;
   int AutoShmSize = 1;
   while(AutoShmSize != WorldShmSize) {
-    for(int p=0;p<primes.size();p++) {
+    int p;
+    for(p=0;p<primes.size();p++) {
       int prime=primes[p];
       if ( divides(prime,WorldDims[dim]/ShmDims[dim])
         && divides(prime,WorldShmSize/AutoShmSize)  ) {
 	AutoShmSize*=prime;
 	ShmDims[dim]*=prime;
+	last_dim = dim;
 	break;
       }
     }
+    if (p == primes.size() && last_dim == dim) {
+      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
+      exit(EXIT_FAILURE);
+    }
     dim=(dim+1) %ndimension;
   }
 }
@@ -413,7 +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 ; 
@@ -433,13 +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
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -677,7 +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);
@@ -685,7 +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);

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,{
+    autoView(rhs_v , rhs, AcceleratorRead);
-      for(int b=0;b<e2;b++){
+    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 
+    Coordinate rdim=rhs.Grid()->_rdimensions;
-    // Test_cshift_red_black code.
+    Coordinate cdm =rhs.Grid()->_checker_dim_mask;
-    std::cout << " Dense packed buffer WARNING " <<std::endl;
+    std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-    thread_for_collapse(2,n,e1,{
+    accelerator_for2d(n,e1,b,e2,1,{
-      for(int b=0;b<e2;b++){
+
+	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,{
+    autoView( rhs_v, rhs, AcceleratorWrite);
-    rhs_v[table[i].first]=buffer[table[i].second];
+    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();
+    autoView( rhs_v , rhs, AcceleratorWrite);
-    thread_for_collapse(2,n,e1,{
+    accelerator_for2d(n,e1,b,e2,1,{
-      for(int b=0;b<e2;b++){
 	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();
+    autoView(rhs_v , rhs, AcceleratorRead);
-  thread_for(i,ent,{
+    autoView(lhs_v , lhs, AcceleratorWrite);
-    lhs_v[table[i].first]=rhs_v[table[i].second];
+    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();
+    autoView( rhs_v, rhs, AcceleratorRead);
-  thread_for(i,ent,{
+    autoView( lhs_v, lhs, AcceleratorWrite);
-    permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+    auto table = &Cshift_table[0];
-  });
+    accelerator_for(i,ent,1,{
+      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+    });
+  }
 }
 
 //////////////////////////////////////////////////////

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);

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>

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
 ////////////////////////////////////////////

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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto x_v = x.AcceleratorView(ViewRead);
+  autoView( x_v , x, AcceleratorRead);
-  auto y_v = y.AcceleratorView(ViewRead);
+  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);
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto x_v = x.AcceleratorView(ViewRead);
+  autoView( x_v , x, AcceleratorRead);
-  auto y_v = y.AcceleratorView(ViewRead);
+  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);

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
@@ -37,180 +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 Advise(int advise) {
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    if (advise & AdviseInfrequentUse) {
-      cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetPreferredLocation,cudaCpuDeviceId);
-    }
-    if (advise & AdviseReadMostly) {
-      cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetReadMostly,-1);
-    }
-#endif
-#endif
-  };
-
-  accelerator_inline void AcceleratorPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    int target;
-    cudaGetDevice(&target);
-    cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),target);
-#endif
-#endif
-  };
-
-  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
@@ -253,28 +80,23 @@ private:
     }
   }
 public:
+
+  /////////////////////////////////////////////////////////////////////////////////
+  // 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));
-    return accessor;
-  }
 
-  LatticeView<vobj> AcceleratorView(int mode = ViewReadWrite) const 
+  LatticeView<vobj> View (ViewMode mode) const 
   {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
-    accessor.AcceleratorPrefetch(mode);
-    return accessor;
-  }
-
-  LatticeView<vobj> HostView(int mode = ViewReadWrite) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
-    accessor.HostPrefetch(mode);
     return accessor;
   }
 
@@ -298,11 +120,15 @@ 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> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
@@ -317,11 +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)
@@ -335,11 +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
@@ -390,10 +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;
   }
 
@@ -403,11 +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;
@@ -445,11 +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 me =   View(AcceleratorWriteDiscard);
-    auto him= r.AcceleratorView(ViewRead);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
+    me.ViewClose();    him.ViewClose();
     return *this;
   }
 
@@ -459,11 +296,12 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   AcceleratorView(ViewWrite);
+    auto me =   View(AcceleratorWriteDiscard);
-    auto him= r.AcceleratorView(ViewRead);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
+    me.ViewClose();    him.ViewClose();
     return *this;
   }
   ///////////////////////////////////////////

Grid/lattice/Lattice_basis.h

@@ -51,34 +51,18 @@ 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;
+  typedef decltype(basis[0].View(AcceleratorRead)) View;
-  auto tmp_v = basis[0].AcceleratorView(ViewReadWrite);
+
-  Vector<View> basis_v(basis.size(),tmp_v);
+  Vector<View> basis_v; basis_v.reserve(basis.size());
-  typedef typename std::remove_reference<decltype(tmp_v[0])>::type vobj;
   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
-  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.;
 
-	for(int j=j0; j<j1; ++j){
+  View *basis_vp = &basis_v[0];
-	  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];
-	}
-      });
-  }
-#else
   int nrot = j1-j0;
   if (!nrot) // edge case not handled gracefully by Cuda
     return;
@@ -86,6 +70,8 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   uint64_t oSites   =grid->oSites();
   uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
 
+  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
+
   Vector <vobj> Bt(siteBlock * nrot); 
   auto Bp=&Bt[0];
 
@@ -96,7 +82,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int 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){
@@ -132,27 +118,30 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j]));
       });
   }
-#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 +149,7 @@ 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>

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();
+  autoView( lhs_v, lhs, CpuRead);
-  auto rhs_v = rhs.View();
+  autoView( rhs_v, rhs, CpuRead);
-  auto ret_v = ret.View();
+  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();
+  autoView( lhs_v, lhs, CpuRead);
-  auto ret_v = ret.View();
+  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();
+  autoView( rhs_v, rhs, CpuRead);
-  auto ret_v = ret.View();
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
     ret_v[ss]=op(lhs,rhs_v[ss]);
   });

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);
 

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();
+  autoView( rhs_v , rhs, AcceleratorRead);
-  auto ret_v = ret.View();
+  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();
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.View();
+  autoView( rhs_v , rhs, AcceleratorRead);
-  auto ret_v = ret.View();
+  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();
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.View();
+  autoView( rhs_v , rhs, AcceleratorRead);
-  auto ret_v = ret.View();
+  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

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();
+  autoView( X_v , X, CpuRead);
-  auto Y_v = Y.View();
+  autoView( Y_v , Y, CpuRead);
-  auto R_v = R.View();
+  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();
+  autoView( X_v , X, CpuRead);
-  auto R_v = R.View();
+  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();
+  autoView( lhs_v , lhs, CpuRead);
-  auto rhs_v = rhs.View();
+  autoView( rhs_v , rhs, CpuRead);
   thread_region {
     std::vector<vobj> Left(Nblock);
     std::vector<vobj> Right(Nblock);

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();
+  autoView( ret_v, ret, AcceleratorWrite);
-  auto lhs_v = lhs.View();
+  autoView( lhs_v, lhs, AcceleratorRead);
-  thread_for( ss, lhs_v.size(), {
+  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();
+  autoView( ret_v, ret, AcceleratorWrite);
-  auto lhs_v = lhs.View();
+  autoView( lhs_v, lhs, AcceleratorRead);
-  thread_for( ss, lhs_v.size(), {
+  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();
+  autoView( rhs_v, rhs, AcceleratorRead);
-  auto lhs_v = lhs.View();
+  autoView( lhs_v, lhs, AcceleratorWrite);
-  thread_for( ss, lhs_v.size(), {
+  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();
+  autoView( rhs_v, rhs, AcceleratorRead);
-  auto lhs_v = lhs.View();
+  autoView( lhs_v, lhs, AcceleratorWrite);
-  thread_for( ss, lhs_v.size(), {
+  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){
+inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
-        
+{
-  GridBase *grid = l.Grid();
+  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[odx];
-  scalar_type * vp = (scalar_type *)&l_v[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){
+inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
-
+{
-  GridBase *grid=l.Grid();
+  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[odx];
-  scalar_type * vp = (scalar_type *)&l_v[odx];
   scalar_type * pt = (scalar_type *)&s;
   for(int w=0;w<words;w++){
     vp[idx+w*Nsimd] = pt[w];
   }
-
   return;
 };
 

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)));
   });

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;
 }
@@ -102,42 +151,29 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   
   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
+  // Might make all code paths go this way.
-  // GPU - SIMT lane compliance...
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
-  typedef decltype(innerProduct(left_v[0],right_v[0])) 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,{
+    // GPU - SIMT lane compliance...
-      auto x_l = left_v(ss);
+    accelerator_for( ss, sites, 1,{
-      auto y_l = right_v(ss);
+	auto x_l = left_v[ss];
-      coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
+	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
+  auto anrm = sum(inner_tmp_v,sites);  
-  nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));  
+  nrm = anrm;
-#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
   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;
+  autoView( x_v, x, AcceleratorRead);
-  Vector<inner_t> inner_tmp(sites);
+  autoView( y_v, y, AcceleratorRead);
-  auto inner_tmp_v = &inner_tmp[0];
+  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,{
+  accelerator_for( ss, sites, 1,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
+      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(innerProductD(vobj(),vobj())) inner_t;
-  typedef decltype(innerProduct(left_v[0],left_v[0])) norm_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();
+  autoView( lhv, lhs, CpuRead);
-  auto rhv=rhs.View();
+  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();
+    autoView( Rv, R, CpuWrite);
-    auto Xv=X.View();
+    autoView( Xv, X, CpuRead);
-    auto Yv=Y.View();
+    autoView( Yv, Y, CpuRead);
-    thread_for_collapse(2, n, e1, {
+    thread_for2d( n, e1, b,e2, {
-      for(int b=0;b<e2;b++){
 	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();
+  autoView( X_v, X, CpuRead);
-  auto Y_v=Y.View();
+  autoView( Y_v, Y, CpuRead);
-  auto R_v=R.View();
+  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();
+  autoView( R_v, R, CpuWrite);
-  auto X_v = X.View();
+  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();
+  autoView( lhs_v, lhs, CpuRead);
-  auto rhs_v=rhs.View();
+  autoView( rhs_v, rhs, CpuRead);
   thread_region
   {
     std::vector<vobj> Left(Nblock);

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);
@@ -156,7 +167,7 @@ __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]
@@ -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();
+  accelerator_barrier();
-  
-  cudaError err = cudaGetLastError();
-  if ( cudaSuccess != err ) {
-    printf("Cuda error %s\n",cudaGetErrorString( err ));
-    exit(0);
-  }
   auto result = buffer_v[0];
   return result;
 }

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
@@ -462,7 +462,7 @@ public:
 
     {
       // Obtain one reseeded generator per thread      
-      int Nthread = GridThread::GetThreads();
+      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);

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();
+  autoView(ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.View();
+  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();
+  autoView( ret_v , ret, AcceleratorWrite);
-  auto lhs_v = lhs.View();
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
   });

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();
+  autoView( half_v, half, CpuWrite);
-  auto full_v = full.View();
+  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();
+  autoView( half_v , half, CpuRead);
-  auto full_v = full.View();
+  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 __CUDA_ARCH__
+#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);
+  autoView( out_v , out,AcceleratorWrite);
-  auto in_v  = in.AcceleratorView(ViewRead);
+  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);
+  autoView( lhs_v , lhs, AcceleratorRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  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();
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
-  auto coarseData_ = coarseData.AcceleratorView(ViewWrite);
+  autoView( ip_         , ip,         AcceleratorWrite);
-  auto ip_         = ip.AcceleratorView(ViewReadWrite);
   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);
+  autoView( fineZ_  , fineZ, AcceleratorWrite);
-  auto fineX_  = fineX.AcceleratorView(ViewRead);
+  autoView( fineX_  , fineX, AcceleratorRead);
-  auto fineY_  = fineY.AcceleratorView(ViewRead);
+  autoView( fineY_  , fineY, AcceleratorRead);
-  auto coarseA_= coarseA.AcceleratorView(ViewRead);
+  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, {
+  {
-    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);
+  // Turn this around to loop threaded over sc and interior loop 
-  auto fineData_   = fineData.AcceleratorView(ViewRead);
+  // 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();
+  autoView( fineData_   , fineData, AcceleratorWrite);
-  auto coarseData_ = coarseData.View();
+  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);
+    peekLocalSite(s,lowDimv,lcoor);
-    pokeLocalSite(s,higherDim,hcoor);
+    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);
+    peekLocalSite(s,higherDimv,hcoor);
-    pokeLocalSite(s,lowDim,lcoor);
+    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);
+      peekLocalSite(s,lowDimv,lcoor);
-      pokeLocalSite(s,higherDim,hcoor);
+      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);
+      peekLocalSite(s,higherDimv,hcoor);
-      pokeLocalSite(s,lowDim,lcoor);
+      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);

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();
+  autoView( ret_v, ret, AcceleratorWrite);
-  auto lhs_v = lhs.View();
+  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();
+  autoView( ret_v, ret, AcceleratorWrite);
-  auto lhs_v = lhs.View();
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
   });

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();
+  autoView( rhs, rhs_i, AcceleratorRead);
-  auto ret = ret_i.View();
+  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();
+  autoView( rhs , rhs_i, AcceleratorRead);
-  auto ret = ret_i.View();
+  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();
+  autoView( ret , ret_i, AcceleratorWrite);
-  auto rhs = rhs_i.View();
+  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();
+  autoView( rhs , rhs_i, AcceleratorRead);
-  auto ret = ret_i.View();
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));

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);

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;

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
 

Grid/qcd/action/fermion/DomainWallVec5dImpl.h

@@ -115,18 +115,21 @@ public:
       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) 

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;

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;
@@ -233,14 +233,16 @@ public:
 	Uconj = where(coor==neglink,-Uconj,Uconj);
       }
 
-      auto U_v = U.View();
+      {
-      auto Uds_v = Uds.View();
+	autoView( U_v , U, CpuRead);
-      auto Uconj_v = Uconj.View();
+	autoView( Uconj_v , Uconj, CpuRead);
-      auto Utmp_v= Utmp.View();
+	autoView( Uds_v , Uds, CpuWrite);
-      thread_foreach(ss,U_v,{
+	autoView( Utmp_v, Utmp, CpuWrite);
-	Uds_v[ss](0)(mu) = U_v[ss]();
+	thread_foreach(ss,U_v,{
-	Uds_v[ss](1)(mu) = Uconj_v[ss]();
+	    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,{
+      autoView( link_v , link, CpuWrite);
-      link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
+      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();
+      autoView( tmp_v , tmp, CpuWrite);
-    auto Btilde_v = Btilde.View();
+      autoView( Atilde_v , Atilde, CpuRead);
-    thread_for(ss,tmp.Grid()->oSites(),{
+      autoView( Btilde_v , Btilde, CpuRead);
-      for (int s = 0; s < Ls; s++) {
+      thread_for(ss,tmp.Grid()->oSites(),{
-	int sF = s + Ls * ss;
+	  for (int s = 0; s < Ls; s++) {
-	auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+	    int sF = s + Ls * ss;
-	tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+	    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;
   }

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -61,8 +61,8 @@ public:
   double DhopCalls;
   double DhopCommTime;
   double DhopComputeTime;
-      double DhopComputeTime2;
+  double DhopComputeTime2;
-      double DhopFaceTime;
+  double DhopFaceTime;
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base

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

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:
+ public:
 
-   void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
-		      int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+		    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:
 

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();
+    assert(0);
-    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();
-	
-    });
   }
   inline void DoubleStore(GridBase *GaugeGrid,
 			  DoubledGaugeField &UUUds, // for Naik term

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();
+    autoView(T_v,T,AcceleratorWrite);
-    auto F_v = F.View();
+    autoView(F_v,F,AcceleratorRead);
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    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();
+    autoView(T_v, T,AcceleratorWrite);
-    auto F_v = F.View();
+    autoView(F_v, F,AcceleratorRead);
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    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();
+    autoView(T_v,T,AcceleratorWrite);
-    auto F_v = F.View();
+    autoView(F_v,F,AcceleratorRead);
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    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();
+    autoView( T_v , T, AcceleratorWrite);
-    auto F_v = F.View();
+    autoView( F_v , F, AcceleratorRead);
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    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();
+    autoView( T_v ,T,AcceleratorWrite);
-    auto F_v = F.View();
+    autoView( F_v ,F,AcceleratorRead);
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    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();
+    autoView( T_v , T,AcceleratorWrite);
-    auto F_v = F.View();
+    autoView( F_v , F,AcceleratorRead);
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    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]()());

Grid/qcd/action/fermion/WilsonImpl.h

@@ -106,10 +106,10 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
-    auto out_v= out.View();
+    autoView( out_v, out, AcceleratorWrite);
-    auto phi_v= phi.View();
+    autoView( phi_v, phi, AcceleratorRead);
-    auto Umu_v= Umu.View();
+    autoView( Umu_v, Umu, AcceleratorRead);
-    thread_for(sss,out.Grid()->oSites(),{
+    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();
+      autoView( tmp_v , tmp, AcceleratorWrite);
-    auto Atilde_v = Atilde.View();
+      autoView( Btilde_v , Btilde, AcceleratorRead);
-    thread_for(sss,tmp.Grid()->oSites(),{
+      autoView( Atilde_v , Atilde, AcceleratorRead);
-      int sU=sss;
+      accelerator_for(sss,tmp.Grid()->oSites(),1,{
-      for(int s=0;s<Ls;s++){
+	  int sU=sss;
-	int sF = s+Ls*sU;
+	  for(int s=0;s<Ls;s++){
-	tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
+	    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);
-      
   }
 };
 

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

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();
+  autoView(psi , psi_i,AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView(phi , phi_i,AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i,AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView(phi , phi_i,AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i,AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i,AcceleratorRead);
-  auto chi = chi_i.View();
+  autoView(chi , chi_i,AcceleratorWrite);
 
   auto plee  = & lee [0];
   auto pdee  = & dee [0];

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();
+  autoView(psi, psi_i,CpuRead);
-  auto phi = phi_i.View();
+  autoView(phi, phi_i,CpuRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi,psi_i,CpuRead);
-  auto phi=phi_i.View();
+  autoView(phi,phi_i,CpuRead);
-  auto chi=chi_i.View();
+  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();
+  autoView(psi , psi_i,CpuRead);
-  auto chi = chi_i.View();
+  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();
+    autoView(psi , psi_i,CpuRead);
-    auto chi = chi_i.View();
+    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();
+    autoView(psi , psi_i,CpuRead);
-    auto chi = chi_i.View();
+    autoView(chi , chi_i,CpuWrite);
     // pointers
     //  MASK_REGS;
 #define Chi_00 %zmm0

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();
+  autoView( phi , phi_i, AcceleratorRead);
-  auto psi = psi_i.View();
+  autoView( psi , psi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView( psi , psi_i, AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView( phi , phi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView( psi, psi_i, AcceleratorRead);
-  auto chi=chi_i.View();
+  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();
+  autoView( psi, psi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto plee  = & this->lee[0];

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();
+  autoView( Umu_v   ,   Umu, CpuRead);
-  auto UUUmu_v = UUUmu.View();
+  autoView( UUUmu_v , UUUmu, CpuRead);
-  auto in_v    = in.View();
+  autoView( in_v    ,  in, CpuRead);
-  auto out_v   = out.View();
+  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 interior=1;
-    int nthreads = omp_get_num_threads();
+    int exterior=0;
-    int ncomms = CartesianCommunicator::nCommThreads;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
-    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;
-    }
   }
-  DhopCommTime += ctime;
+  DhopComputeTime+=usecond();
-  DhopComputeTime+=ptime;
-
-  // First to enter, last to leave timing
-  st.CollateThreads();
 
   DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
   DhopFaceTime+=usecond();
 
-  DhopComputeTime2-=usecond();
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
-  auto   U_v  =   U.View();
+  DhopComputeTime2-=usecond();
-  auto UUU_v  = UUU.View();
+  {
-  auto  in_v  =  in.View();
+    int interior=0;
-  auto out_v  = out.View();
+    int exterior=1;
-  if (dag == DaggerYes) {
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
-    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();
-#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();
+    int interior=1;
-  auto  in_v  =  in.View();
+    int exterior=1;
-  auto out_v  = out.View();
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
-  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);
-    });
   }
   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*/

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();
+    autoView( U_v   , U, CpuRead);
-    auto UUU_v = UUU.View();
+    autoView( UUU_v , UUU, CpuRead);
-    auto B_v   = B.View();
+    autoView( B_v      , B, CpuWrite);
-    auto Btilde_v   = Btilde.View();
+    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();
+  autoView( Umu_v   ,   Umu, CpuRead);
-  auto UUUmu_v = UUUmu.View();
+  autoView( UUUmu_v , UUUmu, CpuRead);
-  auto in_v    =  in.View();
+  autoView( in_v    ,  in, CpuRead);
-  auto out_v   = out.View();
+  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 interior=1;
-    int nthreads = omp_get_num_threads();
+    int exterior=0;
-    int ncomms = CartesianCommunicator::nCommThreads;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
-    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();
-    }
   }
   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();
+    int interior=0;
-    auto UUU_v = UUU.View();
+    int exterior=1;
-    auto in_v  = in.View();
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
-    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);
-      });
-    }
   }
   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(),{
+    int interior=1;
-      Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
+    int exterior=1;
-    });
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
-  } else {
-    thread_for(sss, in.Grid()->oSites(),{
-      Kernels::DhopSite(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
-    });
   }
   DhopComputeTime += usecond();
   DhopTotalTime   += usecond();

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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView(phi , phi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView(phi , phi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView(phi , phi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto phi = phi_i.View();
+  autoView(phi , phi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  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();
+  autoView(psi , psi_i, AcceleratorRead);
-  auto chi = chi_i.View();
+  autoView(chi , chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto pm = this->pm;

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);

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++){
+  //  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);
@@ -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++){
+  //  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_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++){
+  //  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_CHIa(addr0,addr1);

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++){
+  //  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //    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++){
+  //  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //    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++){
+  //  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //    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);

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

@@ -37,9 +37,9 @@ NAMESPACE_BEGIN(Grid);
   if (SE->_is_local ) {						\
     if (SE->_permute) {						\
       chi_p = χ						\
-      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 = χ						\
-      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, 
+					     SiteSpinor *buf, int sF, int sU, 
-					     const FermionFieldView &in, FermionFieldView &out, int dag) {
+					     const FermionFieldView &in, FermionFieldView &out, int dag) 
+{
   const SiteSpinor *chi_p;
   SiteSpinor chi;
   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++){
+  //  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+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++){
+  //  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+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++){
+  //  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+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,
+void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
-					 SiteSpinor *buf, int LLs, int sU,
+					   int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp)
-					 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) 
 {
   // 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);
 
 

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);
+    autoView(CTv,CloverTerm,CpuRead);
-    EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+    autoView(CTIv,CloverTermInv,CpuWrite);
-    peekLocalSite(Qx, CloverTerm, lcoor);
+    for (int site = 0; site < lvol; site++) {
-    Qxinv = Zero();
+      grid->LocalIndexToLocalCoor(site, lcoor);
-    //if (csw!=0){
+      EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-    for (int j = 0; j < Ns; j++)
+      peekLocalSite(Qx, CTv, lcoor);
-      for (int k = 0; k < Ns; k++)
+      Qxinv = Zero();
-        for (int a = 0; a < DimRep; a++)
+      //if (csw!=0){
-          for (int b = 0; b < DimRep; b++){
+      for (int j = 0; j < Ns; j++)
-	    auto zz =  Qx()(j, k)(a, b);
+	for (int k = 0; k < Ns; k++)
-            EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+	  for (int a = 0; a < DimRep; a++)
-	  }
+	    for (int b = 0; b < DimRep; b++){
-    //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
+	      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();
+      EigenInvCloverOp = EigenCloverOp.inverse();
-    //std::cout << EigenInvCloverOp << std::endl;
+      //std::cout << EigenInvCloverOp << std::endl;
-    for (int j = 0; j < Ns; j++)
+      for (int j = 0; j < Ns; j++)
-      for (int k = 0; k < Ns; k++)
+	for (int k = 0; k < Ns; k++)
-        for (int a = 0; a < DimRep; a++)
+	  for (int a = 0; a < DimRep; a++)
-          for (int b = 0; b < DimRep; b++)
+	    for (int b = 0; b < DimRep; b++)
-            Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+	      Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-    //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
+      //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
-    //  }
+      //  }
-    pokeLocalSite(Qxinv, CloverTermInv, lcoor);
+      pokeLocalSite(Qxinv, CTIv, lcoor);
+    }
   }
 
   // Separate the even and odd parts

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;
+    autoView(cosha_v,cosha,CpuRead);
-    //    RealD sgn;
+    autoView(a_v,a,CpuWrite);
-    _grid->LocalIndexToLocalCoor(idx,lcoor);
+    for(int idx=0;idx<_grid->lSites();idx++){
-    peekLocalSite(cc,cosha,lcoor);
+      Coordinate lcoor(Nd);
-    assert((double)real(cc)>=1.0);
+      Tcomplex cc;
-    assert(fabs((double)imag(cc))<=1.0e-15);
+      //    RealD sgn;
-    cc = ScalComplex(::acosh(real(cc)),0.0);
+      _grid->LocalIndexToLocalCoor(idx,lcoor);
-    pokeLocalSite(cc,a,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)  );

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

@@ -67,7 +67,12 @@ 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>
@@ -483,32 +488,7 @@ 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
+  assert(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
 }
 
 
@@ -524,62 +504,7 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 {
   conformable(_grid, q_in.Grid());
   conformable(_grid, q_out.Grid());
-#if 0
+  assert(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
 }
 
 NAMESPACE_END(Grid);

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

@@ -39,9 +39,10 @@ NAMESPACE_BEGIN(Grid);
 // Generic implementation; move to different file?
 ////////////////////////////////////////////
 
+/*
 accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
 {
-#ifdef __CUDA_ARCH__
+#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;
@@ -51,6 +52,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
 #endif
   return;
 }
+*/
 
 #define GENERIC_STENCIL_LEG(Dir,spProj,Recon)			\
   SE = st.GetEntry(ptype, Dir, sF);				\
@@ -61,7 +63,7 @@ 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);
   
@@ -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 ) {					\
@@ -99,7 +101,7 @@ 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);					
 
@@ -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);
@@ -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);
@@ -203,7 +205,7 @@ void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st,  DoubledGaugeField
   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);
@@ -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);
@@ -300,7 +302,7 @@ 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);				\
@@ -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);
@@ -346,30 +348,30 @@ template <class Impl>
 void WilsonKernels<Impl>::DhopDirAll( StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor *buf, int Ls,
 				      int Nsite, const FermionField &in, std::vector<FermionField> &out) 
 {
-   auto U_v   = U.View();
+   autoView(U_v  ,U,AcceleratorRead);
-   auto in_v  = in.View();
+   autoView(in_v ,in,AcceleratorRead);
-   auto st_v  = st.View();
+   autoView(st_v ,st,AcceleratorRead);
 
-   auto out_Xm = out[0].View();
+   autoView(out_Xm,out[0],AcceleratorWrite);
-   auto out_Ym = out[1].View();
+   autoView(out_Ym,out[1],AcceleratorWrite);
-   auto out_Zm = out[2].View();
+   autoView(out_Zm,out[2],AcceleratorWrite);
-   auto out_Tm = out[3].View();
+   autoView(out_Tm,out[3],AcceleratorWrite);
-   auto out_Xp = out[4].View();
+   autoView(out_Xp,out[4],AcceleratorWrite);
-   auto out_Yp = out[5].View();
+   autoView(out_Yp,out[5],AcceleratorWrite);
-   auto out_Zp = out[6].View();
+   autoView(out_Zp,out[6],AcceleratorWrite);
-   auto out_Tp = out[7].View();
+   autoView(out_Tp,out[7],AcceleratorWrite);
-
+   auto CBp=st.CommBuf();
-   accelerator_forNB(sss,Nsite*Ls,Simd::Nsimd(),{
+   accelerator_for(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);
+      DhopDirXm(st_v,U_v,CBp,sF,sU,in_v,out_Xm,0);
-      DhopDirYm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Ym,1);
+      DhopDirYm(st_v,U_v,CBp,sF,sU,in_v,out_Ym,1);
-      DhopDirZm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Zm,2);
+      DhopDirZm(st_v,U_v,CBp,sF,sU,in_v,out_Zm,2);
-      DhopDirTm(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Tm,3);
+      DhopDirTm(st_v,U_v,CBp,sF,sU,in_v,out_Tm,3);
-      DhopDirXp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Xp,4);
+      DhopDirXp(st_v,U_v,CBp,sF,sU,in_v,out_Xp,4);
-      DhopDirYp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Yp,5);
+      DhopDirYp(st_v,U_v,CBp,sF,sU,in_v,out_Yp,5);
-      DhopDirZp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Zp,6);
+      DhopDirZp(st_v,U_v,CBp,sF,sU,in_v,out_Zp,6);
-      DhopDirTp(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_Tp,7);
+      DhopDirTp(st_v,U_v,CBp,sF,sU,in_v,out_Tp,7);
    });
 }
 
@@ -381,17 +383,18 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
   assert(dirdisp<=7);
   assert(dirdisp>=0);
 
-   auto U_v   = U.View();
+   autoView(U_v  ,U  ,AcceleratorRead);
-   auto in_v  = in.View();
+   autoView(in_v ,in ,AcceleratorRead);
-   auto out_v = out.View();
+   autoView(out_v,out,AcceleratorWrite);
-   auto st_v  = st.View();
+   autoView(st_v ,st ,AcceleratorRead);
+   auto CBp=st.CommBuf();			
 #define LoopBody(Dir)				\
-   case Dir :			\
+   case Dir :					\
-     accelerator_forNB(ss,Nsite,Simd::Nsimd(),{	\
+     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;
@@ -435,26 +438,26 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 				     int Ls, int Nsite, const FermionField &in, FermionField &out,
 				     int interior,int exterior) 
 {
-    auto U_v   =   U.View();
+    autoView(U_v  ,  U,AcceleratorRead);
-    auto in_v  =  in.View();
+    autoView(in_v , in,AcceleratorRead);
-    auto out_v = out.View();
+    autoView(out_v,out,AcceleratorWrite);
-    auto st_v  =  st.View();
+    autoView(st_v , st,AcceleratorRead);
 
    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 ) { 
      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
@@ -466,26 +469,26 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 					  int Ls, int Nsite, const FermionField &in, FermionField &out,
 					  int interior,int exterior) 
   {
-    auto U_v   = U.View();
+    autoView(U_v  ,U,AcceleratorRead);
-    auto in_v  = in.View();
+    autoView(in_v ,in,AcceleratorRead);
-    auto out_v = out.View();
+    autoView(out_v,out,AcceleratorWrite);
-    auto st_v  = st.View();
+    autoView(st_v ,st,AcceleratorRead);
 
    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 ) { 
      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 +496,9 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    assert(0 && " Kernel optimisation case not covered ");
   }
 
+#undef KERNEL_CALLNB
+#undef KERNEL_CALL
+#undef ASM_CALL
+
 NAMESPACE_END(Grid);
 

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);

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);

Grid/qcd/action/fermion/instantiation/StaggeredImplD/NaiveStaggeredFermionInstantiationStaggeredImplD.cc

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

Grid/qcd/action/fermion/instantiation/StaggeredImplF/NaiveStaggeredFermionInstantiationStaggeredImplF.cc

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

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

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();
+    autoView(U_v,U,AcceleratorWrite);
-    auto W_v = W.View();
+    autoView(W_v,W,AcceleratorRead);
-    thread_for( ss, U.Grid()->oSites(), {
+    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();
+    autoView(U_v,U,AcceleratorWrite);
-    auto P_v = P.View();
+    autoView(P_v,P,AcceleratorRead);
-    thread_for(ss, P.Grid()->oSites(),{
+    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";
   }

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();
+    autoView( p_v      , p, CpuRead);
-    auto action_v = action.View();
+    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();
+      autoView( p_v , p, CpuRead);
-      auto force_v = force.View();
+      autoView( force_v , force, CpuWrite);
             
       int permute_type;
       StencilEntry *SE;

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< BiCGSTABModule<WilsonFermionR::FermionField>,   
-//static Registrar< ConjugateResidualModule<WilsonFermionR::FermionField>,   
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __BiCGWFmodXMLInit("BiCGSTAB"); 
-//                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CRWFmodXMLInit("ConjugateResidual"); 
+static Registrar< ConjugateResidualModule<WilsonFermionR::FermionField>,   
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CRWFmodXMLInit("ConjugateResidual"); 
 
 // add the staggered, scalar versions here
 

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_uw(u, w, iQ2, iQ3);
-    StoutSmearing.set_fj(f0, f1, f2, u, w);
+    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);
+    xi0 = StoutSmearing->func_xi0(w);
-    xi1 = StoutSmearing.func_xi1(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)

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();
+  		        autoView(emB0_v,emB0[m],CpuRead);
-  		        auto emB1_v = emB1[m].View();
+		        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();
+	autoView(vs_v,vs[s],CpuRead);
-  auto tmp1 = vs_v[ss];
+	auto tmp1 = vs_v[ss];
-  vobj tmp2 = Zero();
+	vobj tmp2 = Zero();
-  vobj tmp3 = Zero();
+	vobj tmp3 = Zero();
-  for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
+	for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
-    auto vd_v = vd[d].View();
+	  autoView(vd_v,vd[d],CpuRead);
-    Scalar_v coeff = WW_sd(t,s,d);
+	  Scalar_v coeff = WW_sd(t,s,d);
-    tmp3 = conjugate(vd_v[ss]);
+	  tmp3 = conjugate(vd_v[ss]);
-    mac(&tmp2, &coeff, &tmp3);
+	  mac(&tmp2, &coeff, &tmp3);
-  }
+	}
 
-  //////////////////////////
+	//////////////////////////
-  // Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
+	// Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
-  //////////////////////////
+	//////////////////////////
-  OuterProductWWVV(WWVV[t], tmp1, tmp2, Ns, ss);
+	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();
+	  autoView(vs_v,vs[s],CpuRead);
-    auto tmp1 = vs_v[ss];
+	  auto tmp1 = vs_v[ss];
-    vobj tmp2 = Zero();
+	  vobj tmp2 = Zero();
-    vobj tmp3 = Zero();
+	  vobj tmp3 = Zero();
-    for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
+	  for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
-      auto vd_v = vd[d].View();
+	    autoView(vd_v,vd[d],CpuRead);
-      Scalar_v coeff = buf(s,d);
+	    Scalar_v coeff = buf(s,d);
-      tmp3 = conjugate(vd_v[ss]);
+	    tmp3 = conjugate(vd_v[ss]);
-      mac(&tmp2, &coeff, &tmp3);
+	    mac(&tmp2, &coeff, &tmp3);
-    }
+	  }
-
+	  //////////////////////////
-    //////////////////////////
+	  // Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
-    // Fast outer product of tmp1 with a sum of terms suppressed by d_unroll
+	  //////////////////////////
-    //////////////////////////
+	  OuterProductWWVV(WWVV[t], tmp1, tmp2, Ns, ss);
-    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();
+  autoView(WWVV0_v , WWVV0,CpuRead);
-  auto WWVV1_v = WWVV1.View();
+  autoView(WWVV1_v , WWVV1,CpuRead);
-  auto O_trtr_v= O_trtr.View();
+  autoView(O_trtr_v, O_trtr,CpuWrite);
-  auto O_fig8_v= O_fig8.View();
+  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();
+  autoView( WWVV0_v , WWVV0,CpuRead);
-  auto WWVV1_v = WWVV1.View();
+  autoView( WWVV1_v , WWVV1,CpuRead);
-  auto O_trtr_v= O_trtr.View();
+  autoView( O_trtr_v, O_trtr,CpuWrite);
-  auto O_fig8_v= O_fig8.View();
+  autoView( O_fig8_v, O_fig8,CpuWrite);
 
   thread_for(ss,grid->oSites(),{
 

Grid/qcd/utils/BaryonUtils.h

@@ -351,10 +351,10 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
 
   GridBase *grid = q1_left.Grid();
   
-  auto vbaryon_corr= baryon_corr.View();
+  autoView(vbaryon_corr, baryon_corr,CpuWrite);
-  auto v1 = q1_left.View();
+  autoView( v1 , q1_left, CpuRead);
-  auto v2 = q2_left.View();
+  autoView( v2 , q2_left, CpuRead);
-  auto v3 = q3_left.View();
+  autoView( v3 , q3_left, CpuRead);
 
   Real bytes =0.;
   bytes += grid->oSites() * (432.*sizeof(vComplex) + 126.*sizeof(int) + 36.*sizeof(Real));
@@ -989,10 +989,10 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
 
   GridBase *grid = qs_ti.Grid();
 
-  auto vcorr= stn_corr.View();
+  autoView( vcorr, stn_corr, CpuWrite);
-  auto vq_loop = qq_loop.View();
+  autoView( vq_loop , qq_loop, CpuRead);
-  auto vd_tf = qd_tf.View();
+  autoView( vd_tf , qd_tf, CpuRead);
-  auto vs_ti = qs_ti.View();
+  autoView( vs_ti , qs_ti, CpuRead);
 
   accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
     auto Dq_loop = vq_loop[ss];
@@ -1029,13 +1029,13 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
 
   GridBase *grid = qs_ti.Grid();
 
-  auto vcorr= stn_corr.View();
+  autoView( vcorr , stn_corr, CpuWrite);
-  auto vq_ti = qq_ti.View();
+  autoView( vq_ti , qq_ti, CpuRead);
-  auto vq_tf = qq_tf.View();
+  autoView( vq_tf , qq_tf, CpuRead);
-  auto vd_tf = qd_tf.View();
+  autoView( vd_tf , qd_tf, CpuRead);
-  auto vs_ti = qs_ti.View();
+  autoView( vs_ti , qs_ti, CpuRead);
-
+ // accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
-  accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
+  thread_for(ss,grid->oSites(),{
     auto Dq_ti = vq_ti[ss];
     auto Dq_tf = vq_tf[ss];
     auto Dd_tf = vd_tf[ss];

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();
+  autoView(x_v, x, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto y_v = y.View();
+  autoView( y_v, y, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto y_v = y.View();
+  autoView( y_v, y, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto y_v = y.View();
+  autoView( y_v, y, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto y_v = y.View();
+  autoView( y_v, y, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto y_v = y.View();
+  autoView( y_v, y, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto y_v = y.View();
+  autoView( y_v, y, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( x_v, x, AcceleratorRead);
-  auto z_v = z.View();
+  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();
+  autoView( z_v, z, AcceleratorWrite);
-  auto x_v = x.View();
+  autoView( x_v, x, AcceleratorRead);
   accelerator_for(ss,grid->oSites(),CComplex::Nsimd(),
   {
     for(int n = 0; n < nb; ++n) {

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();
+    autoView( dest_v , dest, AcceleratorWrite);
-    auto subgroup_v = subgroup.View();
+    autoView( subgroup_v, subgroup, AcceleratorRead);
-    thread_for(ss, grid->oSites(),
+    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,
+  static void LieRandomize(GridParallelRNG &pRNG, LatticeMatrixType &out, double scale = 1.0) 
-                           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<
+    typedef typename GridTypeMapper<typename LatticeMatrixType::vector_object>::scalar_object MatrixType;
-      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);
   }

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
 

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);

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;

Grid/stencil/Stencil.h

@@ -67,7 +67,8 @@ void Gather_plane_simple_table (Vector<std::pair<int,int> >& table,const Lattice
 {
   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
 }
 
@@ -94,7 +96,7 @@ void Gather_plane_exchange_table(Vector<std::pair<int,int> >& table,const Lattic
   int num=table.size()/2;
   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];
@@ -104,10 +106,11 @@ void Gather_plane_exchange_table(Vector<std::pair<int,int> >& table,const Lattic
 			      so+tp[2*j+1].second,
 			      type);
   });
+  rhs_v.ViewClose();
 }
 
 struct StencilEntry { 
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
   uint64_t _byte_offset;       // 8 bytes 
   uint32_t _offset;            // 4 bytes 
 #else
@@ -122,7 +125,7 @@ struct StencilEntry {
 // 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           _checkerboard;
-  int                               _npoints; // Move to template param?
+  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;
@@ -175,13 +179,43 @@ class CartesianStencilView {
   {
     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;
@@ -226,8 +260,8 @@ public:
   // Generalise as required later if needed
   ////////////////////////////////////////////////////////////////////////
 
-  View_type View(void) const {
+  View_type View(ViewMode mode) const {
-    View_type accessor(*( (View_type *) this));
+    View_type accessor(*( (View_type *) this),mode);
     return accessor;
   }
   
@@ -662,9 +696,9 @@ public:
     _unified_buffer_size=0;
     surface_list.resize(0);
 
-    int osites  = _grid->oSites();
+    this->_osites  = _grid->oSites();
     
-    _entries.resize(this->_npoints* osites);
+    _entries.resize(this->_npoints* this->_osites);
     this->_entries_p = &_entries[0];
     for(int ii=0;ii<npoints;ii++){
       

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);
 }

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:

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 )
 {

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);

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);

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
+#include <Grid/threads/Threads.h>
-#define MAX(x,y) ((x)>(y)?(x):(y))
+#include <Grid/threads/Accelerator.h>
-#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

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);
+

Grid/threads/Threads.h

@@ -28,101 +28,47 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once 
 
-
+#ifndef MAX
-// Introduce a class to gain deterministic bit reproducible reduction.
+#define MAX(x,y) ((x)>(y)?(x):(y))
-// make static; perhaps just a namespace is required.
+#define MIN(x,y) ((x)>(y)?(y):(x))
-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) { 
+#define strong_inline     __attribute__((always_inline)) inline
-#ifdef GRID_OMP
+#define UNROLL  _Pragma("unroll")
-    _threads = MIN(thr,omp_get_max_threads()) ;
+
-    omp_set_num_threads(_threads);
+//////////////////////////////////////////////////////////////////////////////////
-#else 
+// New primitives; explicit host thread calls, and accelerator data parallel calls
-    _threads = 1;
+//////////////////////////////////////////////////////////////////////////////////
+
+#ifdef _OPENMP
+#define GRID_OMP
+#include <omp.h>
 #endif
-  };
+
-  static void SetMaxThreads(void) { 
 #ifdef GRID_OMP
-    _threads = omp_get_max_threads();
+#define DO_PRAGMA_(x) _Pragma (#x)
-    omp_set_num_threads(_threads);
+#define DO_PRAGMA(x) DO_PRAGMA_(x)
+#define thread_num(a) omp_get_thread_num()
+#define thread_max(a) omp_get_max_threads()
 #else 
-    _threads = 1;
+#define DO_PRAGMA_(x) 
+#define DO_PRAGMA(x) 
+#define thread_num(a) (0)
+#define thread_max(a) (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){
+#define thread_for( i, num, ... )                           DO_PRAGMA(omp parallel for schedule(static)) for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-    GetWork(nwork,me,mywork,myoff,_threads);
+#define thread_for2d( i1, n1,i2,n2, ... )  \
-  }
+  DO_PRAGMA(omp parallel for collapse(2))  \
-  static void GetWork(int nwork, int me, int & mywork, int & myoff,int units){
+  for ( uint64_t i1=0;i1<n1;i1++) {	   \
-    int basework = nwork/units;
+  for ( uint64_t i2=0;i2<n2;i2++) {	   \
-    int backfill = units-(nwork%units);
+  { __VA_ARGS__ } ;			   \
-    if ( me >= units ) { 
+  }}
-      mywork = myoff = 0;
+#define thread_foreach( i, container, ... )                 DO_PRAGMA(omp parallel for schedule(static)) for ( uint64_t i=container.begin();i<container.end();i++) { __VA_ARGS__ } ;
-    } else { 
+#define thread_for_in_region( i, num, ... )                 DO_PRAGMA(omp for schedule(static))          for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-      mywork = (nwork+me)/units;
+#define thread_for_collapse2( i, num, ... )                 DO_PRAGMA(omp parallel for collapse(2))      for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-      myoff  = basework * me;
+#define thread_for_collapse( N , i, num, ... )              DO_PRAGMA(omp parallel for collapse ( N ) )  for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
-      if ( me > backfill ) 
+#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__ } ;
-	myoff+= (me-backfill);
+#define thread_region                                       DO_PRAGMA(omp parallel)
-    }
+#define thread_critical                                     DO_PRAGMA(omp critical)
-    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);
 

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);
+    resize(sz);
-    assert(sz<=MaxEntries);
-    _size = 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) { 

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
+    arg= GridCmdOptionPayload(argv,argv+argc,"--accelerator-threads");
-    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");
     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
+  acceleratorInit(); // Must come first to set device prior to MPI init due to Omnipath Driver
-
-  PointerCache::Init();
 
   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 +318,11 @@ void Grid_init(int *argc,char ***argv)
     Grid_debug_handler_init();
   }
 
+  //////////////////////////////////////////////////////////
+  // Memory manager
+  //////////////////////////////////////////////////////////
+  MemoryManager::Init();
+
   //////////////////////////////////////////////////////////
   // MPI initialisation
   //////////////////////////////////////////////////////////
@@ -419,11 +361,18 @@ 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;

benchmarks/Benchmark_ITT.cc

@@ -237,9 +237,9 @@ public:
 
       Vec rn ; random(sRNG,rn);
 
-      LatticeVec z(&Grid); z=rn;
+      LatticeVec z(&Grid); z=Zero();
-      LatticeVec x(&Grid); x=rn;
+      LatticeVec x(&Grid); x=Zero();
-      LatticeVec y(&Grid); y=rn;
+      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();
+	autoView( x_v , x, CpuWrite);
-	auto y_v = y.View();
+	autoView( y_v , y, CpuWrite);
-	auto z_v = z.View();
+	autoView( z_v , z, CpuRead);
         x_v[0]=z_v[0]; // force serial dependency to prevent optimise away
         y_v[4]=z_v[4];
       }

benchmarks/Benchmark_dwf.cc

@@ -21,7 +21,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #define CUDA_PROFILE
 #endif
 
@@ -129,8 +129,8 @@ int main (int argc, char ** argv)
   LatticeGaugeField Umu5d(FGrid); 
   std::vector<LatticeColourMatrix> U(4,FGrid);
   {
-    auto Umu5d_v = Umu5d.View();
+    autoView( Umu5d_v, Umu5d, CpuWrite);
-    auto Umu_v = Umu.View();
+    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];
@@ -258,8 +258,8 @@ 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();
+	autoView( ref_v, ref, CpuWrite);
-	auto tmp_v = tmp.View();
+	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 +268,8 @@ int main (int argc, char ** argv)
       tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu+1,-1);
       {
-	auto ref_v = ref.View();
+	autoView( ref_v, ref, CpuWrite);
-	auto tmp_v = tmp.View();
+	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]; ;
 	}

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();
+    autoView(Umu5d_v , Umu5d, CpuWrite);
-  for(int ss=0;ss<Umu.Grid()->oSites();ss++){
+    autoView(  Umu_v , Umu,   CpuRead);
-    for(int s=0;s<Ls;s++){
+    for(int ss=0;ss<Umu.Grid()->oSites();ss++){
-      Umu5d_v[Ls*ss+s] = Umu_v[ss];
+      for(int s=0;s<Ls;s++){
+	Umu5d_v[Ls*ss+s] = Umu_v[ss];
+      }
     }
   }