relative Eigen links, allows moving safely Grid's directory

.gitignore

@@ -88,7 +88,6 @@ Thumbs.db
 # build directory #
 ###################
 build*/*
-Documentation/_build
 
 # IDE related files #
 #####################

.travis.yml

@@ -0,0 +1,61 @@
+language: cpp
+
+cache:
+  directories:
+    - clang
+
+matrix:
+  include:
+    - os:        osx
+      osx_image: xcode8.3
+      compiler: clang
+      env: PREC=single
+    - os:        osx
+      osx_image: xcode8.3
+      compiler: clang
+      env: PREC=double
+      
+before_install:
+    - export GRIDDIR=`pwd`
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]] && [ ! -e clang/bin ]; then wget $CLANG_LINK; tar -xf `basename $CLANG_LINK`; mkdir clang; mv clang+*/* clang/; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export PATH="${GRIDDIR}/clang/bin:${PATH}"; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc openssl; fi
+    
+install:
+    - export CWD=`pwd`
+    - echo $CWD
+    - export CC=$CC$VERSION
+    - export CXX=$CXX$VERSION
+    - echo $PATH
+    - which autoconf
+    - autoconf  --version
+    - which automake
+    - automake  --version
+    - which $CC
+    - $CC  --version
+    - which $CXX
+    - $CXX --version
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export EXTRACONF='--with-openssl=/usr/local/opt/openssl'; fi
+    
+script:
+    - ./bootstrap.sh
+    - mkdir build
+    - cd build
+    - mkdir lime
+    - cd lime
+    - mkdir build
+    - cd build
+    - wget http://usqcd-software.github.io/downloads/c-lime/lime-1.3.2.tar.gz
+    - tar xf lime-1.3.2.tar.gz
+    - cd lime-1.3.2
+    - ./configure --prefix=$CWD/build/lime/install
+    - make -j4
+    - make install
+    - cd $CWD/build
+    - ../configure --enable-precision=$PREC --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
+    - make -j4 
+    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
+    - make check

Grid/DisableWarnings.h

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

Grid/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/Allocator.h>
+#include <Grid/allocator/AlignedAllocator.h>
 #include <Grid/simd/Simd.h>
-#include <Grid/threads/ThreadReduction.h>
+#include <Grid/threads/Threads.h>
 #include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 

Grid/GridStd.h

@@ -6,7 +6,6 @@
 ///////////////////
 #include <cassert>
 #include <complex>
-#include <memory>
 #include <vector>
 #include <array>
 #include <string>
@@ -28,7 +27,4 @@
 ///////////////////
 #include "Config.h"
 
-#ifdef TOFU
-#undef GRID_COMMS_THREADS
-#endif
 #endif /* GRID_STD_H */

Grid/Grid_Eigen_Dense.h

@@ -18,28 +18,21 @@
 #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 GRID_SYCL
+#ifdef __SYCL_DEVICE_ONLY__  
 #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
 
-/* HIP save and restore compile environment*/
-#ifdef GRID_HIP
-#pragma push
-#pragma push_macro("__HIP_DEVICE_COMPILE__")
-#endif
-#define EIGEN_NO_HIP
 
 #include <Grid/Eigen/Dense>
 #include <Grid/Eigen/unsupported/CXX11/Tensor>
@@ -58,12 +51,6 @@
 #pragma pop
 #endif
 
-/*HIP restore*/
-#ifdef __HIP__REDEFINE__
-#pragma pop_macro("__HIP_DEVICE_COMPILE__")
-#pragma pop
-#endif
-
 #if defined __GNUC__
 #pragma GCC diagnostic pop
 #endif

Grid/Makefile.am

@@ -21,8 +21,7 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
-
-all: version-cache Version.h
+all: version-cache
 
 version-cache:
 	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
@@ -43,7 +42,7 @@ version-cache:
 	fi;\
 	rm -f vertmp
 
-Version.h: version-cache
+Version.h:
 	cp version-cache Version.h
 
 .PHONY: version-cache
@@ -54,19 +53,6 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
-extra_sources+=$(WILS_FERMION_FILES)
-extra_sources+=$(STAG_FERMION_FILES)
-if BUILD_ZMOBIUS
-  extra_sources+=$(ZWILS_FERMION_FILES)
-endif
-if BUILD_GPARITY
-  extra_sources+=$(GP_FERMION_FILES)
-endif
-if BUILD_FERMION_REPS
-  extra_sources+=$(ADJ_FERMION_FILES)
-  extra_sources+=$(TWOIND_FERMION_FILES)
-endif
-
 lib_LIBRARIES = libGrid.a
 
 CCFILES += $(extra_sources)

Grid/algorithms/Algorithms.h

@@ -29,11 +29,9 @@ 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>
@@ -43,12 +41,10 @@ NAMESPACE_CHECK(SparseMatrix);
 #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>
@@ -66,9 +62,7 @@ NAMESPACE_CHECK(BiCGSTAB);
 #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/FFT.h

@@ -1,3 +1,4 @@
+
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -36,6 +37,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif
 #endif
 
+
 NAMESPACE_BEGIN(Grid);
 
 template<class scalar> struct FFTW { };
@@ -136,7 +138,7 @@ public:
     flops=0;
     usec =0;
     Coordinate layout(Nd,1);
-    sgrid = new GridCartesian(dimensions,layout,processors,*grid);
+    sgrid = new GridCartesian(dimensions,layout,processors);
   };
     
   ~FFT ( void)  {
@@ -182,14 +184,14 @@ public:
     pencil_gd[dim] = G*processors[dim];
       
     // Pencil global vol LxLxGxLxL per node
-    GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid);
+    GridCartesian pencil_g(pencil_gd,layout,processors);
       
     // Construct pencils
     typedef typename vobj::scalar_object sobj;
     typedef typename sobj::scalar_type   scalar;
       
     Lattice<sobj> pgbuf(&pencil_g);
-    autoView(pgbuf_v , pgbuf, CpuWrite);
+    auto pgbuf_v = pgbuf.View();
 
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -230,18 +232,15 @@ public:
     result = source;
     int pc = processor_coor[dim];
     for(int p=0;p<processors[dim];p++) {
-      {
-	autoView(r_v,result,CpuRead);
-	autoView(p_v,pgbuf,CpuWrite);
-	thread_for(idx, sgrid->lSites(),{
+      thread_for(idx, sgrid->lSites(),{
           Coordinate cbuf(Nd);
           sobj s;
 	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,r_v,cbuf);
+	  peekLocalSite(s,result,cbuf);
 	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  pokeLocalSite(s,p_v,cbuf);
-        });
-      }
+	  //            cbuf[dim]+=p*L;
+	  pokeLocalSite(s,pgbuf,cbuf);
+      });
       if (p != processors[dim] - 1) {
 	result = Cshift(result,dim,L);
       }
@@ -270,19 +269,15 @@ public:
     flops+= flops_call*NN;
       
     // writing out result
-    {
-      autoView(pgbuf_v,pgbuf,CpuRead);
-      autoView(result_v,result,CpuWrite);
-      thread_for(idx,sgrid->lSites(),{
+    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_v,cgbuf);
-	pokeLocalSite(s,result_v,clbuf);
-      });
-    }
+	peekLocalSite(s,pgbuf,cgbuf);
+	pokeLocalSite(s,result,clbuf);
+    });
     result = result*div;
       
     // destroying plan

Grid/algorithms/LinearOperator.h

@@ -52,7 +52,6 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
-  virtual ~LinearOperatorBase(){};
 };
 
 
@@ -508,7 +507,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   virtual  void MpcDag   (const Field &in, Field &out){
     Mpc(in,out);
   }
-  virtual void MpcDagMpc(const Field &in, Field &out) {
+  virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
     assert(0);// Never need with staggered
   }
 };
@@ -531,16 +530,6 @@ public:
 template<class Field> class LinearFunction {
 public:
   virtual void operator() (const Field &in, Field &out) = 0;
-
-  virtual void operator() (const std::vector<Field> &in, std::vector<Field> &out)
-  {
-    assert(in.size() == out.size());
-
-    for (unsigned int i = 0; i < in.size(); ++i)
-    {
-      (*this)(in[i], out[i]);
-    }
-  }
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
@@ -586,7 +575,6 @@ class HermOpOperatorFunction : public OperatorFunction<Field> {
 template<typename Field>
 class PlainHermOp : public LinearFunction<Field> {
 public:
-  using LinearFunction<Field>::operator();
   LinearOperatorBase<Field> &_Linop;
       
   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
@@ -600,7 +588,6 @@ public:
 template<typename Field>
 class FunctionHermOp : public LinearFunction<Field> {
 public:
-  using LinearFunction<Field>::operator(); 
   OperatorFunction<Field>   & _poly;
   LinearOperatorBase<Field> &_Linop;
       

Grid/algorithms/Preconditioner.h

@@ -30,19 +30,13 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Field> using Preconditioner =  LinearFunction<Field> ;
-
-/*
-template<class Field> class Preconditioner :  public LinearFunction<Field> {
-  using LinearFunction<Field>::operator();
+template<class Field> class Preconditioner :  public LinearFunction<Field> { 
   virtual void operator()(const Field &src, Field & psi)=0;
 };
-*/
 
 template<class Field> class TrivialPrecon :  public Preconditioner<Field> { 
 public:
-  using Preconditioner<Field>::operator();
-  virtual void operator()(const Field &src, Field & psi){
+  void operator()(const Field &src, Field & psi){
     psi = src;
   }
   TrivialPrecon(void){};

Grid/algorithms/SparseMatrix.h

@@ -48,7 +48,6 @@ public:
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;
-  virtual ~SparseMatrixBase() {};
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -73,7 +72,7 @@ public:
   virtual  void MeooeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeInvDag (const Field &in, Field &out)=0;
-  virtual ~CheckerBoardedSparseMatrixBase() {};
+
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/BiCGSTAB.h

@@ -122,14 +122,12 @@ class BiCGSTAB : public OperatorFunction<Field>
 
         LinearCombTimer.Start();
         bo = beta * omega;
-	{
-	  autoView( p_v , p, AcceleratorWrite);
-	  autoView( r_v , r, AcceleratorRead);
-	  autoView( v_v , v, AcceleratorRead);
-	  accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
-	      coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
-	    });
-	}
+        auto p_v = p.View();
+        auto r_v = r.View();
+        auto v_v = v.View();
+        accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
+          coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
+        });
         LinearCombTimer.Stop();
         LinalgTimer.Stop();
 
@@ -144,20 +142,16 @@ class BiCGSTAB : public OperatorFunction<Field>
         alpha = rho / Calpha.real();
 
         LinearCombTimer.Start();
-	{
-	  autoView( p_v , p, AcceleratorRead);
-	  autoView( r_v , r, AcceleratorRead);
-	  autoView( v_v , v, AcceleratorRead);
-	  autoView( psi_v,psi, AcceleratorRead);
-	  autoView( h_v  ,  h, AcceleratorWrite);
-	  autoView( s_v  ,  s, AcceleratorWrite);
-	  accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
-	      coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
-	    });
-	  accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
-	      coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
- 	  });
-        }
+        auto h_v = h.View();
+        auto psi_v = psi.View();
+        accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
+          coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
+        });
+        
+        auto s_v = s.View();
+        accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
+          coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
+        });
         LinearCombTimer.Stop();
         LinalgTimer.Stop();
 
@@ -172,19 +166,13 @@ class BiCGSTAB : public OperatorFunction<Field>
         omega = Comega.real() / norm2(t);
 
         LinearCombTimer.Start();
-	{
-	  autoView( psi_v,psi, AcceleratorWrite);
-	  autoView( r_v , r, AcceleratorWrite);
-	  autoView( h_v , h, AcceleratorRead);
-	  autoView( s_v , s, AcceleratorRead);
-	  autoView( t_v , t, AcceleratorRead);
-	  accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
-	      coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
-	      coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
-	    });
-	}
+        auto t_v = t.View();
+        accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
+          coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
+          coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
+        });
         LinearCombTimer.Stop();
-	
+
         cp = norm2(r);
         LinalgTimer.Stop();
 

Grid/algorithms/iterative/BiCGSTABMixedPrec.h

@@ -36,8 +36,7 @@ NAMESPACE_BEGIN(Grid);
 template<class FieldD, class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
 class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD> 
 {
-  public:
-    using LinearFunction<FieldD>::operator();
+  public:                                                
     RealD   Tolerance;
     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/ConjugateGradient.h

@@ -140,15 +140,13 @@ public:
       b = cp / c;
 
       LinearCombTimer.Start();
-      {
-	autoView( psi_v , psi, AcceleratorWrite);
-	autoView( p_v   , p,   AcceleratorWrite);
-	autoView( r_v   , r,   AcceleratorWrite);
-	accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
-	    coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
-	    coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
-	});
-      }
+      auto psi_v = psi.View();
+      auto p_v   = p.View();
+      auto r_v   = r.View();
+      accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+	  coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+	  coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+      });
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -35,8 +35,7 @@ NAMESPACE_BEGIN(Grid);
     typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
-  public:
-    using LinearFunction<FieldD>::operator();
+  public:                                                
     RealD   Tolerance;
     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/Deflation.h

@@ -33,19 +33,16 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
-  using LinearFunction<Field>::operator();
     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
 };
 template<class Field>
 class DoNothingGuesser: public LinearFunction<Field> {
 public:
-  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) {  };
 };
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
-  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) { guess = src; };
 };
 
@@ -57,24 +54,15 @@ class DeflatedGuesser: public LinearFunction<Field> {
 private:
   const std::vector<Field> &evec;
   const std::vector<RealD> &eval;
-  const unsigned int       N;
 
 public:
-  using LinearFunction<Field>::operator();
 
-  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
-  : DeflatedGuesser(_evec, _eval, _evec.size())
-  {}
-
-  DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N)
-  : evec(_evec), eval(_eval), N(_N)
-  {
-    assert(evec.size()==eval.size());
-    assert(N <= evec.size());
-  } 
+  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
 
   virtual void operator()(const Field &src,Field &guess) {
     guess = Zero();
+    assert(evec.size()==eval.size());
+    auto N = evec.size();
     for (int i=0;i<N;i++) {
       const Field& tmp = evec[i];
       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
@@ -91,7 +79,6 @@ private:
   const std::vector<RealD>       &eval_coarse;
 public:
   
-  using LinearFunction<FineField>::operator();
   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
 				const std::vector<CoarseField> &_evec_coarse,
 				const std::vector<RealD>       &_eval_coarse)

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -67,7 +67,6 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
-  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -98,7 +97,6 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
-  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
 template<class Field>
 class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
 public:                                                
-  using LinearFunction<Field>::operator();
+
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -1,242 +0,0 @@
-/*************************************************************************************
-
-    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:                                                
-  using LinearFunction<Field>::operator();
-  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;
-    //    ComplexD 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/algorithms/iterative/SchurRedBlack.h

@@ -132,31 +132,6 @@ namespace Grid {
       (*this)(_Matrix,in,out,guess);
     }
 
-    void RedBlackSource(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &src_o) 
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      Field tmp(grid);
-      int nblock = in.size();
-      for(int b=0;b<nblock;b++){
-	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
-      }
-    }
-    // James can write his own deflated guesser
-    // with optimised code for the inner products
-    //    RedBlackSolveSplitGrid();
-    //    RedBlackSolve(_Matrix,src_o,sol_o); 
-
-    void RedBlackSolution(Matrix &_Matrix, const std::vector<Field> &in, const std::vector<Field> &sol_o, std::vector<Field> &out)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      Field tmp(grid);
-      int nblock = in.size();
-      for(int b=0;b<nblock;b++) {
-	pickCheckerboard(Even,tmp,in[b]);
-	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]);
-      }
-    }
-
     template<class Guesser>
     void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess) 
     {
@@ -175,29 +150,24 @@ namespace Grid {
       ////////////////////////////////////////////////
       // Prepare RedBlack source
       ////////////////////////////////////////////////
-      RedBlackSource(_Matrix,in,src_o);
-	//      for(int b=0;b<nblock;b++){
-	//	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
-	//      }
-      
+      for(int b=0;b<nblock;b++){
+	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+      }
       ////////////////////////////////////////////////
       // Make the guesses
       ////////////////////////////////////////////////
       if ( subGuess ) guess_save.resize(nblock,grid);
 
-      
-      if(useSolnAsInitGuess) {
-        for(int b=0;b<nblock;b++){
+      for(int b=0;b<nblock;b++){
+        if(useSolnAsInitGuess) {
           pickCheckerboard(Odd, sol_o[b], out[b]);
+        } else {
+          guess(src_o[b],sol_o[b]); 
         }
-      } else {
-        guess(src_o, sol_o); 
-      }
 
-	    if ( subGuess ) { 
-        for(int b=0;b<nblock;b++){
-          guess_save[b] = sol_o[b];
-        }
+	if ( subGuess ) { 
+	  guess_save[b] = sol_o[b];
+	}
       }
       //////////////////////////////////////////////////////////////
       // Call the block solver

Grid/allocator/AlignedAllocator.cc

@@ -0,0 +1,154 @@
+#include <Grid/GridCore.h>
+#include <fcntl.h>
+
+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__
+  int fd = open("/proc/self/pagemap", O_RDONLY);
+  assert(fd >= 0);
+  const int page_size = 4096;
+  uint64_t virt_pfn = (uint64_t)Buf / page_size;
+  off_t offset = sizeof(uint64_t) * virt_pfn;
+  uint64_t npages = (BYTES + page_size-1) / page_size;
+  uint64_t pagedata[npages];
+  uint64_t ret = lseek(fd, offset, SEEK_SET);
+  assert(ret == offset);
+  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
+  assert(ret == sizeof(uint64_t) * npages);
+  int nhugepages = npages / 512;
+  int n4ktotal, nnothuge;
+  n4ktotal = 0;
+  nnothuge = 0;
+  for (int i = 0; i < nhugepages; ++i) {
+    uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
+    for (int j = 0; j < 512; ++j) {
+      uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
+      ++n4ktotal;
+      if (pageaddr != baseaddr + j * page_size)
+	++nnothuge;
+    }
+  }
+  int rank = CartesianCommunicator::RankWorld();
+  printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
+#endif
+}
+
+std::string sizeString(const size_t bytes)
+{
+  constexpr unsigned int bufSize = 256;
+  const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
+  char                   buf[256];
+  size_t                 s     = 0;
+  double                 count = bytes;
+  
+  while (count >= 1024 && s < 7)
+    {
+      s++;
+      count /= 1024;
+    }
+  if (count - floor(count) == 0.0)
+    {
+      snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
+    }
+  else
+    {
+      snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
+    }
+  
+  return std::string(buf);
+}
+
+NAMESPACE_END(Grid);
+

Grid/allocator/AlignedAllocator.h

@@ -26,10 +26,129 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#pragma once
+#ifndef GRID_ALIGNED_ALLOCATOR_H
+#define GRID_ALIGNED_ALLOCATOR_H
+
+#ifdef HAVE_MALLOC_MALLOC_H
+#include <malloc/malloc.h>
+#endif
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
+
+#ifdef HAVE_MM_MALLOC_H
+#include <mm_malloc.h>
+#endif
+
+#define POINTER_CACHE
+#define GRID_ALLOC_ALIGN (2*1024*1024)
+#define GRID_ALLOC_SMALL_LIMIT (4096)
 
 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: 
@@ -53,131 +172,89 @@ 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
+    pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
+#else 
+    pointer __freeme = __p;
+#endif
+
+#ifdef GRID_NVCC
+    if ( __freeme ) cudaFree((void *)__freeme);
+#else 
+  #ifdef HAVE_MM_MALLOC_H
+    if ( __freeme ) _mm_free((void *)__freeme); 
+  #else
+    if ( __freeme ) free((void *)__freeme);
+  #endif
+#endif
   }
 
-  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
-  void construct(pointer __p, const _Tp& __val) { assert(0);};
+  // FIXME: hack for the copy constructor, eventually it must be avoided
+  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
+  //void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
 template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
-//////////////////////////////////////////////////////////////////////////////////////
-// Unified virtual memory
-//////////////////////////////////////////////////////////////////////////////////////
-template<typename _Tp>
-class uvmAllocator {
-public: 
-  typedef std::size_t     size_type;
-  typedef std::ptrdiff_t  difference_type;
-  typedef _Tp*       pointer;
-  typedef const _Tp* const_pointer;
-  typedef _Tp&       reference;
-  typedef const _Tp& const_reference;
-  typedef _Tp        value_type;
-
-  template<typename _Tp1>  struct rebind { typedef uvmAllocator<_Tp1> other; };
-  uvmAllocator() throw() { }
-  uvmAllocator(const uvmAllocator&) throw() { }
-  template<typename _Tp1> uvmAllocator(const uvmAllocator<_Tp1>&) throw() { }
-  ~uvmAllocator() throw() { }
-  pointer       address(reference __x)       const { return &__x; }
-  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
-
-  pointer allocate(size_type __n, const void* _p= 0)
-  { 
-    size_type bytes = __n*sizeof(_Tp);
-    profilerAllocate(bytes);
-    _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-    return ptr;
-  }
-
-  void deallocate(pointer __p, size_type __n) 
-  { 
-    size_type bytes = __n * sizeof(_Tp);
-    profilerFree(bytes);
-    MemoryManager::SharedFree((void *)__p,bytes);
-  }
-
-  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
-  void construct(pointer __p) { };
-  void destroy(pointer __p) { };
-};
-template<typename _Tp>  inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
-template<typename _Tp>  inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
-
-////////////////////////////////////////////////////////////////////////////////
-// Device memory
-////////////////////////////////////////////////////////////////////////////////
-template<typename _Tp>
-class devAllocator {
-public: 
-  typedef std::size_t     size_type;
-  typedef std::ptrdiff_t  difference_type;
-  typedef _Tp*       pointer;
-  typedef const _Tp* const_pointer;
-  typedef _Tp&       reference;
-  typedef const _Tp& const_reference;
-  typedef _Tp        value_type;
-
-  template<typename _Tp1>  struct rebind { typedef devAllocator<_Tp1> other; };
-  devAllocator() throw() { }
-  devAllocator(const devAllocator&) throw() { }
-  template<typename _Tp1> devAllocator(const devAllocator<_Tp1>&) throw() { }
-  ~devAllocator() 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::AcceleratorAllocate(bytes);
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-    return ptr;
-  }
-
-  void deallocate(pointer __p, size_type __n) 
-  { 
-    size_type bytes = __n * sizeof(_Tp);
-    profilerFree(bytes);
-    MemoryManager::AcceleratorFree((void *)__p,bytes);
-  }
-  void construct(pointer __p, const _Tp& __val) { };
-  void construct(pointer __p) { };
-  void destroy(pointer __p) { };
-};
-template<typename _Tp>  inline bool operator==(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return true; }
-template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return false; }
-
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-#ifdef ACCELERATOR_CSHIFT
-// Cshift on device
-template<class T> using cshiftAllocator = devAllocator<T>;
-#else
-// Cshift on host
-template<class T> using cshiftAllocator = std::allocator<T>;
-#endif
-
-template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
-template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,devAllocator<T> >;
-template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
+template<class T> using commAllocator = alignedAllocator<T>;
+template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
+template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
+template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
 
 NAMESPACE_END(Grid);
 
-
+#endif

Grid/allocator/Allocator.h

@@ -1,4 +0,0 @@
-#pragma once
-#include <Grid/allocator/MemoryStats.h>
-#include <Grid/allocator/MemoryManager.h>
-#include <Grid/allocator/AlignedAllocator.h>

Grid/allocator/MemoryManager.cc

@@ -1,300 +0,0 @@
-#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)
-#undef GRID_MM_VERBOSE 
-uint64_t total_shared;
-uint64_t total_device;
-uint64_t total_host;;
-void MemoryManager::PrintBytes(void)
-{
-  std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
-  std::cout << " MemoryManager : PrintBytes "<<std::endl;
-  std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
-  std::cout << " MemoryManager : "<<(total_shared>>20)<<" shared      Mbytes "<<std::endl;
-  std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
-  std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl;
-  uint64_t cacheBytes;
-  cacheBytes = CacheBytes[Cpu];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Acc];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Shared];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
-  
-#ifdef GRID_CUDA
-  cuda_mem();
-#endif
-  
-}
-
-//////////////////////////////////////////////////////////////////////
-// 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] = { 2, 8, 2, 8, 2, 8 };
-uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
-//////////////////////////////////////////////////////////////////////
-// Actual allocation and deallocation utils
-//////////////////////////////////////////////////////////////////////
-void *MemoryManager::AcceleratorAllocate(size_t bytes)
-{
-  total_device+=bytes;
-  void *ptr = (void *) Lookup(bytes,Acc);
-  if ( ptr == (void *) NULL ) {
-    ptr = (void *) acceleratorAllocDevice(bytes);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"AcceleratorAllocate "<<std::endl;
-  PrintBytes();
-#endif
-  return ptr;
-}
-void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes)
-{
-  total_device-=bytes;
-  void *__freeme = Insert(ptr,bytes,Acc);
-  if ( __freeme ) {
-    acceleratorFreeDevice(__freeme);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"AcceleratorFree "<<std::endl;
-  PrintBytes();
-#endif
-}
-void *MemoryManager::SharedAllocate(size_t bytes)
-{
-  total_shared+=bytes;
-  void *ptr = (void *) Lookup(bytes,Shared);
-  if ( ptr == (void *) NULL ) {
-    ptr = (void *) acceleratorAllocShared(bytes);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"SharedAllocate "<<std::endl;
-  PrintBytes();
-#endif
-  return ptr;
-}
-void  MemoryManager::SharedFree    (void *ptr,size_t bytes)
-{
-  total_shared-=bytes;
-  void *__freeme = Insert(ptr,bytes,Shared);
-  if ( __freeme ) {
-    acceleratorFreeShared(__freeme);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"SharedFree "<<std::endl;
-  PrintBytes();
-#endif
-}
-#ifdef GRID_UVM
-void *MemoryManager::CpuAllocate(size_t bytes)
-{
-  total_host+=bytes;
-  void *ptr = (void *) Lookup(bytes,Cpu);
-  if ( ptr == (void *) NULL ) {
-    ptr = (void *) acceleratorAllocShared(bytes);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"CpuAllocate "<<std::endl;
-  PrintBytes();
-#endif
-  return ptr;
-}
-void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
-{
-  total_host-=bytes;
-  NotifyDeletion(_ptr);
-  void *__freeme = Insert(_ptr,bytes,Cpu);
-  if ( __freeme ) { 
-    acceleratorFreeShared(__freeme);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"CpuFree "<<std::endl;
-  PrintBytes();
-#endif
-}
-#else
-void *MemoryManager::CpuAllocate(size_t bytes)
-{
-  total_host+=bytes;
-  void *ptr = (void *) Lookup(bytes,Cpu);
-  if ( ptr == (void *) NULL ) {
-    ptr = (void *) acceleratorAllocCpu(bytes);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"CpuAllocate "<<std::endl;
-  PrintBytes();
-#endif
-  return ptr;
-}
-void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
-{
-  total_host-=bytes;
-  NotifyDeletion(_ptr);
-  void *__freeme = Insert(_ptr,bytes,Cpu);
-  if ( __freeme ) { 
-    acceleratorFreeCpu(__freeme);
-  }
-#ifdef GRID_MM_VERBOSE
-  std::cout <<"CpuFree "<<std::endl;
-  PrintBytes();
-#endif
-}
-#endif
-
-//////////////////////////////////////////
-// call only once
-//////////////////////////////////////////
-void MemoryManager::Init(void)
-{
-
-  char * str;
-  int Nc;
-  
-  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;
-    }
-  }
-
-}
-
-void MemoryManager::InitMessage(void) {
-
-#ifndef GRID_UVM
-  std::cout << GridLogMessage << "MemoryManager Cache "<< MemoryManager::DeviceMaxBytes <<" bytes "<<std::endl;
-#endif
-  
-  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],CacheBytes[cache]);  
-#else
-  return ptr;
-#endif
-}
-
-void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
-{
-  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;
-    cacheBytes -= entries[v].bytes;
-    entries[v].valid = 0;
-    entries[v].address = NULL;
-    entries[v].bytes = 0;
-  }
-
-  entries[v].address=ptr;
-  entries[v].bytes  =bytes;
-  entries[v].valid  =1;
-  cacheBytes += bytes;
-
-  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],CacheBytes[cache]);
-#else
-  return NULL;
-#endif
-}
-
-void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
-{
-  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;
-      cacheBytes -= entries[e].bytes;
-      return entries[e].address;
-    }
-  }
-  return NULL;
-}
-
-
-NAMESPACE_END(Grid);
-

Grid/allocator/MemoryManager.h

@@ -1,182 +0,0 @@
-/*************************************************************************************
-
-    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 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];
-  static uint64_t CacheBytes[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,uint64_t &cbytes) ;
-  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
-
-  static void PrintBytes(void);
- public:
-  static void Init(void);
-  static void InitMessage(void);
-  static void *AcceleratorAllocate(size_t bytes);
-  static void  AcceleratorFree    (void *ptr,size_t bytes);
-  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 void PrintState( void* CpuPtr);
-  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

@@ -1,505 +0,0 @@
-#include <Grid/GridCore.h>
-#ifndef GRID_UVM
-
-#warning "Using explicit device memory copies"
-NAMESPACE_BEGIN(Grid);
-//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define dprintf(...)
-
-
-////////////////////////////////////////////////////////////
-// For caching copies of data on device
-////////////////////////////////////////////////////////////
-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 ){
-    EntryCreate(CpuPtr,bytes,mode,hint);
-  }
-
-  auto AccCacheIterator = EntryLookup(CpuPtr);
-  auto & AccCache = AccCacheIterator->second;
-  if (!AccCache.AccPtr) {
-    EvictVictims(bytes); 
-  } 
-  assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
-
-  assert(AccCache.cpuLock==0);  // Programming error
-
-  if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
-		    (uint64_t)AccCache.CpuPtr,
-		    (uint64_t)CpuPtr,
-		    (uint64_t)AccCache.bytes,
-		    (uint64_t)bytes);
-    assert(AccCache.CpuPtr == CpuPtr);
-    assert(AccCache.bytes  ==bytes);
-  }
-/*
- *  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++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
-  } else if(AccCache.state==Consistent) {
-    if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
-      AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
-    else
-      AccCache.state  = Consistent; // Consistent + AccRead => Consistent
-    AccCache.accLock++;
-    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
-  } else if(AccCache.state==AccDirty) {
-    if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
-      AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
-    else
-      AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
-    AccCache.accLock++;
-    dprintf("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 ){
-    EntryCreate(CpuPtr,bytes,mode,transient);
-  }
-
-  auto AccCacheIterator = EntryLookup(CpuPtr);
-  auto & AccCache = AccCacheIterator->second;
-
-  if (!AccCache.AccPtr) {
-     EvictVictims(bytes);
-  }
-
-  assert((mode==CpuRead)||(mode==CpuWrite));
-  assert(AccCache.accLock==0);  // Programming error
-
-  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)
-{
-  PrintBytes();
-  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;
-  }
-}
-
-void MemoryManager::PrintState(void* _CpuPtr)
-{
-  uint64_t CpuPtr = (uint64_t)_CpuPtr;
-
-  if ( EntryPresent(CpuPtr) ){
-    auto AccCacheIterator = EntryLookup(CpuPtr);
-    auto & AccCache = AccCacheIterator->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");
-    if ( AccCache.state==EvictNext) str = std::string("EvictNext");
-
-    std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "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;
-
-  } else {
-    std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl; 
-  }
-}
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/allocator/MemoryManagerShared.cc

@@ -1,27 +0,0 @@
-#include <Grid/GridCore.h>
-#ifdef GRID_UVM
-
-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::PrintState(void* CpuPtr)
-{
-std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
-};
-void  MemoryManager::Print(void){};
-void  MemoryManager::NotifyDeletion(void *ptr){};
-
-NAMESPACE_END(Grid);
-#endif

Grid/allocator/MemoryStats.cc

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

Grid/allocator/MemoryStats.h

@@ -1,95 +0,0 @@
-/*************************************************************************************
-
-    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

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

Grid/cartesian/Cartesian_full.h

@@ -38,7 +38,6 @@ class GridCartesian: public GridBase {
 
 public:
   int dummy;
-  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -105,7 +104,6 @@ public:
     _ldimensions.resize(_ndimension);
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
-    _checker_dim_mask.resize(_ndimension);;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -116,8 +114,6 @@ 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

@@ -35,28 +35,12 @@ static const int CbRed  =0;
 static const int CbBlack=1;
 static const int Even   =CbRed;
 static const int Odd    =CbBlack;
-
-accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex,const Coordinate &rdim,const 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/Communicator_base.cc

@@ -33,8 +33,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-bool Stencil_force_mpi = true;
-
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////

Grid/communicator/Communicator_base.h

@@ -1,3 +1,4 @@
+
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -35,8 +36,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-extern bool Stencil_force_mpi ;
-
 class CartesianCommunicator : public SharedMemory {
 
 public:    
@@ -109,8 +108,6 @@ public:
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
-  void GlobalMax(RealD &);
-  void GlobalMax(RealF &);
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
   void GlobalSum(RealD &);
@@ -141,6 +138,21 @@ public:
 		      int recv_from_rank,
 		      int bytes);
   
+  void SendRecvPacket(void *xmit,
+		      void *recv,
+		      int xmit_to_rank,
+		      int recv_from_rank,
+		      int bytes);
+  
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+			   void *xmit,
+			   int xmit_to_rank,
+			   void *recv,
+			   int recv_from_rank,
+			   int bytes);
+  
+  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,
 			       void *recv,

Grid/communicator/Communicator_mpi3.cc

@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid
+    Grid physics library, www.github.com/paboyle/Grid 
 
     Source file: ./lib/communicator/Communicator_mpi.cc
 
@@ -35,7 +35,7 @@ Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 ////////////////////////////////////////////
 // First initialise of comms system
 ////////////////////////////////////////////
-void CartesianCommunicator::Init(int *argc, char ***argv)
+void CartesianCommunicator::Init(int *argc, char ***argv) 
 {
 
   int flag;
@@ -43,16 +43,8 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-
-#ifndef GRID_COMMS_THREADS
-    nCommThreads=1;
-    // wrong results here too
-    // For now: comms-overlap leads to wrong results in Benchmark_wilson even on single node MPI runs
-    // other comms schemes are ok
-    MPI_Init_thread(argc,argv,MPI_THREAD_SERIALIZED,&provided);
-#else
     MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
-#endif
+
     //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
     if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
       assert(0);
@@ -99,7 +91,7 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Initialises from communicator_world
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) 
 {
   MPI_Comm optimal_comm;
   ////////////////////////////////////////////////////
@@ -118,7 +110,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)    
 {
   _ndimension = processors.size();  assert(_ndimension>=1);
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
@@ -135,7 +127,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  //  int Nparent = parent._processors ;
+  //  int Nparent = parent._processors ; 
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
 
@@ -157,13 +149,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
-  int crank;
+  int crank;  
   // Mpi uses the reverse Lexico convention to us; so reversed routines called
   Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
   Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
 
   MPI_Comm comm_split;
-  if ( Nchild > 1 ) {
+  if ( Nchild > 1 ) { 
 
     ////////////////////////////////////////////////////////////////
     // Split the communicator
@@ -188,11 +180,11 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   SetCommunicator(comm_split);
 
   ///////////////////////////////////////////////
-  // Free the temp communicator
+  // Free the temp communicator 
   ///////////////////////////////////////////////
   MPI_Comm_free(&comm_split);
 
-  if(0){
+  if(0){ 
     std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
     for(int d=0;d<processors.size();d++){
       std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
@@ -253,7 +245,7 @@ CartesianCommunicator::~CartesianCommunicator()
     for(int i=0;i<communicator_halo.size();i++){
       MPI_Comm_free(&communicator_halo[i]);
     }
-  }
+  }  
 }
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
@@ -275,16 +267,6 @@ void CartesianCommunicator::GlobalXOR(uint64_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
   assert(ierr==0);
 }
-void CartesianCommunicator::GlobalMax(float &f)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
-  assert(ierr==0);
-}
 void CartesianCommunicator::GlobalSum(float &f){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
@@ -312,28 +294,60 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int bytes)
 {
   std::vector<CommsRequest_t> reqs(0);
-  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
-
+  //    unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  //    unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  //    xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
+  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
+  SendToRecvFromComplete(reqs);
+  //    rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+  //    printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
+}
+void CartesianCommunicator::SendRecvPacket(void *xmit,
+					   void *recv,
+					   int sender,
+					   int receiver,
+					   int bytes)
+{
+  MPI_Status stat;
+  assert(sender != receiver);
+  int tag = sender;
+  if ( _processor == sender ) {
+    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
+  }
+  if ( _processor == receiver ) { 
+    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
+  }
+}
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes)
+{
   int myrank = _processor;
   int ierr;
 
-  // Enforce no UVM in comms, device or host OK
-  assert(acceleratorIsCommunicable(xmit));
-  assert(acceleratorIsCommunicable(recv));
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
 
-  // Give the CPU to MPI immediately; can use threads to overlap optionally
-  //  printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
-  ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		    recv,bytes,MPI_CHAR,from, from,
-		    communicator,MPI_STATUS_IGNORE);
-  assert(ierr==0);
-
-  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    
+    assert(ierr==0);
+    list.push_back(xrq);
+    list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
+  }
 }
-// Basic Halo comms primitive
+
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int dest,
 						     void *recv,
@@ -353,7 +367,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int from,
 							 int bytes,int dir)
 {
-  int ncomm  =communicator_halo.size();
+  int ncomm  =communicator_halo.size(); 
   int commdir=dir%ncomm;
 
   MPI_Request xrq;
@@ -368,41 +382,37 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
-  int tag;
 
-  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+from*32;
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+  if ( gfrom ==MPI_UNDEFINED) {
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
-  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+_processor*32;
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+  if ( gdest == MPI_UNDEFINED ) {
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
-  } else {
-    // TODO : make a OMP loop on CPU, call threaded bcopy
-    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-    assert(shm!=NULL);
-    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
-    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
     this->StencilSendToRecvFromComplete(list,dir);
   }
 
   return off_node_bytes;
 }
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
+{
+  SendToRecvFromComplete(waitall);
+}
+void CartesianCommunicator::StencilBarrier(void)
+{
+  MPI_Barrier  (ShmComm);
+}
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
-  //   std::cout << "Copy Synchronised\n"<<std::endl;
-  acceleratorCopySynchronise();
-
   int nreq=list.size();
 
   if (nreq==0) return;
@@ -412,13 +422,6 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
   assert(ierr==0);
   list.resize(0);
 }
-void CartesianCommunicator::StencilBarrier(void)
-{
-  MPI_Barrier  (ShmComm);
-}
-//void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-//{
-//}
 void CartesianCommunicator::Barrier(void)
 {
   int ierr = MPI_Barrier(communicator);
@@ -433,8 +436,8 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-int CartesianCommunicator::RankWorld(void){
-  int r;
+int CartesianCommunicator::RankWorld(void){ 
+  int r; 
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
@@ -467,7 +470,7 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
   // (Turns up on 32^3 x 64 Gparity too)
   MPI_Datatype object;
-  int iwords;
+  int iwords; 
   int ibytes;
   iwords = words;
   ibytes = bytes;
@@ -480,3 +483,5 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 }
 
 NAMESPACE_END(Grid);
+
+

Grid/communicator/Communicator_none.cc

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

Grid/communicator/SharedMemory.h

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

Grid/communicator/SharedMemoryMPI.cc

@@ -7,7 +7,6 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Christoph Lehner <christoph@lhnr.de>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -30,14 +29,8 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
-#ifdef GRID_CUDA
+#ifdef GRID_NVCC
 #include <cuda_runtime_api.h>
-#endif
-#ifdef GRID_HIP
-#include <hip/hip_runtime_api.h>
-#endif
-#ifdef GRID_SYCl
-
 #endif
 
 NAMESPACE_BEGIN(Grid); 
@@ -54,12 +47,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
-#ifndef GRID_MPI3_SHM_NONE
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
-#else
-  MPI_Comm_split(comm, WorldRank, 0, &WorldShmComm);
-#endif
-
   MPI_Comm_rank(WorldShmComm     ,&WorldShmRank);
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
@@ -73,7 +61,6 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   WorldNodes = WorldSize/WorldShmSize;
   assert( (WorldNodes * WorldShmSize) == WorldSize );
 
-
   // FIXME: Check all WorldShmSize are the same ?
 
   /////////////////////////////////////////////////////////////////////
@@ -174,23 +161,6 @@ static inline int divides(int a,int b)
 }
 void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
 {
-  ////////////////////////////////////////////////////////////////
-  // Allow user to configure through environment variable
-  ////////////////////////////////////////////////////////////////
-  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
-  if ( str ) {
-    std::vector<int> IntShmDims;
-    GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
-    long ShmSize = 1;
-    for (int dim=0;dim<WorldDims.size();dim++) {
-      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
-    }
-    assert(ShmSize == WorldShmSize);
-    return;
-  }
-  
   ////////////////////////////////////////////////////////////////
   // Powers of 2,3,5 only in prime decomposition for now
   ////////////////////////////////////////////////////////////////
@@ -200,24 +170,17 @@ 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) {
-    int p;
-    for(p=0;p<primes.size();p++) {
+    for(int 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;
   }
 }
@@ -450,47 +413,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
-
-//if defined(GRID_SYCL)
-#if 0
-void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
-{
-  void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // allocate the pointer array for shared windows for our group
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  MPI_Barrier(WorldShmComm);
-  WorldShmCommBufs.resize(WorldShmSize);
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Each MPI rank should allocate our own buffer
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
-
-  if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
-    exit(EXIT_FAILURE);  
-  }
-
-  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
-
-  SharedMemoryZero(ShmCommBuf,bytes);
-
-  assert(WorldShmSize == 1);
-  for(int r=0;r<WorldShmSize;r++){
-    WorldShmCommBufs[r] = ShmCommBuf;
-  }
-  _ShmAllocBytes=bytes;
-  _ShmAlloc=1;
-}
-#endif
-
-#if defined(GRID_CUDA) ||defined(GRID_HIP) ||defined(GRID_SYCL)  
+#ifdef GRID_NVCC
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -510,73 +433,47 @@ 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
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
+  auto err =  cudaMalloc(&ShmCommBuf, bytes);
+  if ( err !=  cudaSuccess) {
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
+    exit(EXIT_FAILURE);  
+  }
   if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
-  std::cout<< "Setting up IPC"<<std::endl;
+
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   for(int r=0;r<WorldShmSize;r++){
-
-#ifndef GRID_MPI3_SHM_NONE
+    
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
     //////////////////////////////////////////////////
-    void * thisBuf = ShmCommBuf;
-    if(!Stencil_force_mpi) {
-#ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
-
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
-      
-    ze_ipc_mem_handle_t ihandle;
-    clone_mem_t handle;
-
-    if ( r==WorldShmRank ) { 
-      auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
-      if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
-	exit(EXIT_FAILURE);
-      } else {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
-      }
-      memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
-      handle.pid = getpid();
-    }
-#endif
-#ifdef GRID_CUDA
     cudaIpcMemHandle_t handle;
+    
     if ( r==WorldShmRank ) { 
-      auto err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
-#endif
-#ifdef GRID_HIP
-    hipIpcMemHandle_t handle;    
-    if ( r==WorldShmRank ) { 
-      auto err = hipIpcGetMemHandle(&handle,ShmCommBuf);
-      if ( err !=  hipSuccess) {
-	std::cerr << " SharedMemoryMPI.cc hipIpcGetMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
-	exit(EXIT_FAILURE);
-      }
-    }
-#endif
-
     //////////////////////////////////////////////////
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
@@ -592,68 +489,23 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     ///////////////////////////////////////////////////////////////
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
-
-#ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    if ( r!=WorldShmRank ) {
-      thisBuf = nullptr;
-      std::cout<<"mapping seeking remote pid/fd "
-	       <<handle.pid<<"/"
-	       <<handle.fd<<std::endl;
-
-      int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
-      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
-      //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
-      int myfd  = syscall(438,pidfd,handle.fd,0);
-
-      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-      
-      memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
-
-      auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
-      if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
-	exit(EXIT_FAILURE);
-      } else {
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl;
-      }
-      assert(thisBuf!=nullptr);
-    }
-#endif
-#ifdef GRID_CUDA
+    void * thisBuf = ShmCommBuf;
     if ( r!=WorldShmRank ) { 
-      auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
       if ( err !=  cudaSuccess) {
 	std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
 	exit(EXIT_FAILURE);
       }
     }
-#endif
-#ifdef GRID_HIP
-    if ( r!=WorldShmRank ) { 
-      auto err = hipIpcOpenMemHandle(&thisBuf,handle,hipIpcMemLazyEnablePeerAccess);
-      if ( err !=  hipSuccess) {
-	std::cerr << " SharedMemoryMPI.cc hipIpcOpenMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
-	exit(EXIT_FAILURE);
-      }
-    }
-#endif
     ///////////////////////////////////////////////////////////////
     // Save a copy of the device buffers
     ///////////////////////////////////////////////////////////////
-    }
     WorldShmCommBufs[r] = thisBuf;
-#else
-    WorldShmCommBufs[r] = ShmCommBuf;
-#endif
   }
 
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
-#endif
-
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
@@ -781,6 +633,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
+      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -824,16 +677,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 /////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
-#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  acceleratorMemSet(dest,0,bytes);
+#ifdef GRID_NVCC
+  cudaMemset(dest,0,bytes);
 #else
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
 {
-#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  acceleratorCopyToDevice(src,dest,bytes);
+#ifdef GRID_NVCC
+  cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
 #else   
   bcopy(src,dest,bytes);
 #endif
@@ -852,11 +705,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
-#ifndef GRID_MPI3_SHM_NONE
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
-#else
-  MPI_Comm_split(comm, rank, 0, &ShmComm);
-#endif
   MPI_Comm_rank(ShmComm     ,&ShmRank);
   MPI_Comm_size(ShmComm     ,&ShmSize);
   ShmCommBufs.resize(ShmSize);
@@ -886,18 +735,25 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
 
-#ifdef GRID_SHM_FORCE_MPI
-  // Hide the shared memory path between ranks
-  {
+#ifdef GRID_IBM_SUMMIT
+  // Hide the shared memory path between sockets 
+  // if even number of nodes
+  if ( (ShmSize & 0x1)==0 ) {
+    int SocketSize = ShmSize/2;
+    int mySocket = ShmRank/SocketSize; 
     for(int r=0;r<size;r++){
-      if ( r!=rank ) {
-	ShmRanks[r] = MPI_UNDEFINED;
+      int hisRank=ShmRanks[r];
+      if ( hisRank!= MPI_UNDEFINED ) {
+	int hisSocket=hisRank/SocketSize;
+	if ( hisSocket != mySocket ) {
+	  ShmRanks[r] = MPI_UNDEFINED;
+	}
       }
     }
   }
 #endif
 
-  //SharedMemoryTest();
+  SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier

Grid/communicator/SharedMemoryNone.cc

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

Grid/cshift/Cshift.h

@@ -52,8 +52,23 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto Cshift(const Expression &expr,int dim,int shift)  -> decltype(closure(expr)) 
+template<typename Op, typename T1> 
+auto Cshift(const LatticeUnaryExpression<Op,T1> &expr,int dim,int shift)
+    -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
+{
+  return Cshift(closure(expr),dim,shift);
+}
+template <class Op, class T1, class T2>
+auto Cshift(const LatticeBinaryExpression<Op,T1,T2> &expr,int dim,int shift)
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> 
+{
+  return Cshift(closure(expr),dim,shift);
+}
+template <class Op, class T1, class T2, class T3>
+auto Cshift(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr,int dim,int shift)
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
+				   eval(0, expr.arg2),
+				   eval(0, expr.arg3)))> 
 {
   return Cshift(closure(expr),dim,shift);
 }

Grid/cshift/Cshift_common.h

@@ -29,13 +29,11 @@ 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 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -48,16 +46,16 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
   int e2=rhs.Grid()->_slice_block[dimension];
   int ent = 0;
 
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
+  static Vector<std::pair<int,int> > table; table.resize(e1*e2);
   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;
-	Cshift_table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
+	table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
       }
     }
   } else { 
@@ -67,26 +65,14 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
 	 int o  = n*stride;
 	 int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	 if ( ocb &cbmask ) {
-	   Cshift_table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
+	   table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
 	 }
        }
      }
   }
-  {
-    auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for(i,ent,vobj::Nsimd(),{
-	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
-    });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for(i,ent,{
-      buffer_p[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
-  }
+  thread_for(i,ent,{
+    buffer[table[i].first]=rhs_v[table[i].second];
+  });
 }
 
 ///////////////////////////////////////////////////////////////////
@@ -109,80 +95,43 @@ 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){
-#ifdef ACCELERATOR_CSHIFT
-    autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for(nn,e1*e2,1,{
-	int n = nn%e1;
-	int b = nn/e1;
+    thread_for_collapse(2,n,e1,{
+      for(int b=0;b<e2;b++){
+
 	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, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-	int o      =   n*n1;
-	int offset = b+n*e2;
-	
-	vobj temp =rhs_v[so+o+b];
-	extract<vobj>(temp,pointers,offset);
-      });
-#endif
+      }
+    });
   } else { 
-    Coordinate rdim=rhs.Grid()->_rdimensions;
-    Coordinate cdm =rhs.Grid()->_checker_dim_mask;
-    std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for(nn,e1*e2,1,{
-	int n = nn%e1;
-	int b = nn/e1;
 
-	Coordinate coor;
+    // Case of SIMD split AND checker dim cannot currently be hit, except in 
+    // Test_cshift_red_black code.
+    std::cout << " Dense packed buffer WARNING " <<std::endl;
+    thread_for_collapse(2,n,e1,{
+      for(int b=0;b<e2;b++){
 
 	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
+	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	int offset = b+n*e2;
 
 	if ( ocb & cbmask ) {
 	  vobj temp =rhs_v[so+o+b];
 	  extract<vobj>(temp,pointers,offset);
 	}
-      });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-
-	Coordinate coor;
-
-	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
-	int offset = b+n*e2;
-
-	if ( ocb & cbmask ) {
-	  vobj temp =rhs_v[so+o+b];
-	  extract<vobj>(temp,pointers,offset);
-	}
-      });
-#endif
+      }
+    });
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -196,8 +145,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   int e2=rhs.Grid()->_slice_block[dimension];
   int stride=rhs.Grid()->_slice_stride[dimension];
 
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
+  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
   int ent    =0;
 
   if ( cbmask ==0x3 ) {
@@ -206,7 +154,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
       for(int b=0;b<e2;b++){
 	int o   =n*rhs.Grid()->_slice_stride[dimension];
 	int bo  =n*rhs.Grid()->_slice_block[dimension];
-	Cshift_table[ent++] = std::pair<int,int>(so+o+b,bo+b);
+	table[ent++] = std::pair<int,int>(so+o+b,bo+b);
       }
     }
 
@@ -217,27 +165,16 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
 	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 ) {
-	  Cshift_table[ent++]=std::pair<int,int> (so+o+b,bo++);
+	  table[ent++]=std::pair<int,int> (so+o+b,bo++);
 	}
       }
     }
   }
   
-  {
-    auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v, rhs, AcceleratorWrite);
-    accelerator_for(i,ent,vobj::Nsimd(),{
-	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
-    });
-#else
-    autoView( rhs_v, rhs, CpuWrite);
-    thread_for(i,ent,{
-      rhs_v[table[i].first]=buffer_p[table[i].second];
-    });
-#endif
-  }
+  auto rhs_v = rhs.View();
+  thread_for(i,ent,{
+    rhs_v[table[i].first]=buffer[table[i].second];
+  });
 }
 
 //////////////////////////////////////////////////////
@@ -257,33 +194,21 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
-    int _slice_stride = rhs.Grid()->_slice_stride[dimension];
-    int _slice_block = rhs.Grid()->_slice_block[dimension];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v , rhs, AcceleratorWrite);
-    accelerator_for(nn,e1*e2,1,{
-	int n = nn%e1;
-	int b = nn/e1;
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
-	merge(rhs_v[so+o+b],pointers,offset);
-      });
-#else
-    autoView( rhs_v , rhs, CpuWrite);
-    thread_for2d(n,e1,b,e2,{
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
+    auto rhs_v = rhs.View();
+    thread_for_collapse(2,n,e1,{
+      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);
+      }
     });
-#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
     // Test_cshift_red_black code.
-    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
+    //    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
-    assert(0); // This will fail if hit on GPU
-    autoView( rhs_v, rhs, CpuWrite);
+    auto rhs_v = rhs.View();
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o      = n*rhs.Grid()->_slice_stride[dimension];
@@ -300,7 +225,6 @@ 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];
@@ -315,16 +239,14 @@ 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];
-
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
+  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
   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;
-	Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
+	table[ent++] = std::pair<int,int>(lo+o,ro+o);
       }
     }
   } else { 
@@ -333,32 +255,23 @@ 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 ) {
-	  Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
+	  table[ent++] = std::pair<int,int>(lo+o,ro+o);
 	}
       }
     }
   }
 
-  {
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
-    autoView(lhs_v , lhs, AcceleratorWrite);
-    accelerator_for(i,ent,vobj::Nsimd(),{
-      coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
-    });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    autoView(lhs_v , lhs, CpuWrite);
-    thread_for(i,ent,{
-      lhs_v[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
-  }
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for(i,ent,{
+    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) ) {
@@ -372,41 +285,29 @@ 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];
 
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
+  static std::vector<std::pair<int,int> > table;  table.resize(e1*e2);
   int ent=0;
 
   if ( cbmask == 0x3 ) {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
-      Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+      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 ) Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+      if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
   }
 
-  {
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v, rhs, AcceleratorRead);
-    autoView( lhs_v, lhs, AcceleratorWrite);
-    accelerator_for(i,ent,1,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#else
-    autoView( rhs_v, rhs, CpuRead);
-    autoView( lhs_v, lhs, CpuWrite);
-    thread_for(i,ent,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#endif
-  }
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for(i,ent,{
+    permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+  });
 }
 
 //////////////////////////////////////////////////////

Grid/cshift/Cshift_mpi.h

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

Grid/cshift/Cshift_table.cc

@@ -1,4 +0,0 @@
-#include <Grid/GridCore.h>       
-NAMESPACE_BEGIN(Grid);
-Vector<std::pair<int,int> > Cshift_table; 
-NAMESPACE_END(Grid);

Grid/lattice/Lattice.h

@@ -26,7 +26,6 @@ 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>
@@ -36,8 +35,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_local.h>
 #include <Grid/lattice/Lattice_reduction.h>
 #include <Grid/lattice/Lattice_peekpoke.h>
-#include <Grid/lattice/Lattice_reality.h>
-#include <Grid/lattice/Lattice_real_imag.h>
+//#include <Grid/lattice/Lattice_reality.h>
 #include <Grid/lattice/Lattice_comparison_utils.h>
 #include <Grid/lattice/Lattice_comparison.h>
 #include <Grid/lattice/Lattice_coordinate.h>

Grid/lattice/Lattice_ET.h

@@ -42,24 +42,9 @@ NAMESPACE_BEGIN(Grid);
 ////////////////////////////////////////////////////
 // Predicated where support
 ////////////////////////////////////////////////////
-#ifdef GRID_SIMT
-// drop to scalar in SIMT; cleaner in fact
 template <class iobj, class vobj, class robj>
-accelerator_inline vobj predicatedWhere(const iobj &predicate, 
-					const vobj &iftrue, 
-					const robj &iffalse) 
-{
-  Integer mask = TensorRemove(predicate);
-  typename std::remove_const<vobj>::type ret= iffalse;
-  if (mask) ret=iftrue;
-  return ret;
-}
-#else
-template <class iobj, class vobj, class robj>
-accelerator_inline vobj predicatedWhere(const iobj &predicate, 
-					const vobj &iftrue, 
-					const robj &iffalse) 
-{
+accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
+                            const robj &iffalse) {
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
@@ -83,7 +68,6 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
   merge(ret, falsevals);
   return ret;
 }
-#endif
 
 /////////////////////////////////////////////////////
 //Specialization of getVectorType for lattices
@@ -97,62 +81,26 @@ struct getVectorType<Lattice<T> >{
 //--  recursive evaluation of expressions; --
 // handle leaves of syntax tree
 ///////////////////////////////////////////////////
-template<class sobj,
-  typename std::enable_if<!is_lattice<sobj>::value&&!is_lattice_expr<sobj>::value,sobj>::type * = nullptr> 
-accelerator_inline 
+template<class sobj> accelerator_inline 
 sobj eval(const uint64_t ss, const sobj &arg)
 {
   return arg;
 }
-template <class lobj> accelerator_inline 
-auto eval(const uint64_t ss, const LatticeView<lobj> &arg) -> decltype(arg(ss))
-{
-  return arg(ss);
-}
 
-////////////////////////////////////////////
-//--  recursive evaluation of expressions; --
-// whole vector return, used only for expression return type inference
-///////////////////////////////////////////////////
-template<class sobj> accelerator_inline 
-sobj vecEval(const uint64_t ss, const sobj &arg)
-{
-  return arg;
-}
 template <class lobj> accelerator_inline 
-const lobj & vecEval(const uint64_t ss, const LatticeView<lobj> &arg) 
+const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg) 
 {
   return arg[ss];
 }
+template <class lobj> accelerator_inline 
+const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg) 
+{
+  auto view = arg.AcceleratorView(ViewRead);
+  return view[ss];
+}
 
 ///////////////////////////////////////////////////
 // handle nodes in syntax tree- eval one operand
-// vecEval needed (but never called as all expressions offloaded) to infer the return type
-// in SIMT contexts of closure.
-///////////////////////////////////////////////////
-template <typename Op, typename T1> accelerator_inline 
-auto vecEval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
-  -> decltype(expr.op.func( vecEval(ss, expr.arg1)))
-{
-  return expr.op.func( vecEval(ss, expr.arg1) );
-}
-// vecEval two operands
-template <typename Op, typename T1, typename T2> accelerator_inline
-auto vecEval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
-  -> decltype(expr.op.func( vecEval(ss,expr.arg1),vecEval(ss,expr.arg2)))
-{
-  return expr.op.func( vecEval(ss,expr.arg1), vecEval(ss,expr.arg2) );
-}
-// vecEval three operands
-template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
-auto vecEval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
-  -> decltype(expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3)))
-{
-  return expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3));
-}
-
-///////////////////////////////////////////////////
-// handle nodes in syntax tree- eval one operand coalesced
 ///////////////////////////////////////////////////
 template <typename Op, typename T1> accelerator_inline 
 auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
@@ -160,41 +108,23 @@ auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)
 {
   return expr.op.func( eval(ss, expr.arg1) );
 }
+///////////////////////
 // eval two operands
+///////////////////////
 template <typename Op, typename T1, typename T2> accelerator_inline
 auto eval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
   -> decltype(expr.op.func( eval(ss,expr.arg1),eval(ss,expr.arg2)))
 {
   return expr.op.func( eval(ss,expr.arg1), eval(ss,expr.arg2) );
 }
+///////////////////////
 // eval three operands
+///////////////////////
 template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
 auto eval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
-  -> decltype(expr.op.func(eval(ss, expr.arg1), 
-			   eval(ss, expr.arg2), 
-			   eval(ss, expr.arg3)))
+  -> decltype(expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3)))
 {
-#ifdef GRID_SIMT
-  // Handles Nsimd (vInteger) != Nsimd(ComplexD)
-  typedef decltype(vecEval(ss, expr.arg2)) rvobj;
-  typedef typename std::remove_reference<rvobj>::type vobj;
-
-  const int Nsimd = vobj::vector_type::Nsimd();
-
-  auto vpred = vecEval(ss,expr.arg1);
-
-  ExtractBuffer<Integer> mask(Nsimd);
-  extract<vInteger, Integer>(TensorRemove(vpred), mask);
-
-  int s = acceleratorSIMTlane(Nsimd);
-  return expr.op.func(mask[s],
-		      eval(ss, expr.arg2), 
-		      eval(ss, expr.arg3));
-#else
-  return expr.op.func(eval(ss, expr.arg1),
-		      eval(ss, expr.arg2), 
-		      eval(ss, expr.arg3));
-#endif
+  return expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3));
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -250,12 +180,16 @@ 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) 
 {
@@ -270,86 +204,32 @@ 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);  // rrecurse 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
 ////////////////////////////////////////////
 #define GridUnopClass(name, ret)					\
+  template <class arg>							\
   struct name {								\
-    template<class _arg> static auto accelerator_inline func(const _arg a) -> decltype(ret) { return ret; } \
+    static auto accelerator_inline func(const arg a) -> decltype(ret) { return ret; } \
   };
 
 GridUnopClass(UnarySub, -a);
 GridUnopClass(UnaryNot, Not(a));
+GridUnopClass(UnaryAdj, adj(a));
+GridUnopClass(UnaryConj, conjugate(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
+GridUnopClass(UnaryReal, real(a));
+GridUnopClass(UnaryImag, imag(a));
+GridUnopClass(UnaryToReal, toReal(a));
+GridUnopClass(UnaryToComplex, toComplex(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
 GridUnopClass(UnarySqrt, sqrt(a));
+GridUnopClass(UnaryRsqrt, rsqrt(a));
 GridUnopClass(UnarySin, sin(a));
 GridUnopClass(UnaryCos, cos(a));
 GridUnopClass(UnaryAsin, asin(a));
@@ -361,10 +241,10 @@ GridUnopClass(UnaryExp, exp(a));
 // Binary operators
 ////////////////////////////////////////////
 #define GridBinOpClass(name, combination)			\
+  template <class left, class right>				\
   struct name {							\
-    template <class _left, class _right>			\
     static auto accelerator_inline				\
-    func(const _left &lhs, const _right &rhs)			\
+    func(const left &lhs, const right &rhs)			\
       -> decltype(combination) const				\
     {								\
       return combination;					\
@@ -384,10 +264,10 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
 // Trinary conditional op
 ////////////////////////////////////////////////////
 #define GridTrinOpClass(name, combination)				\
+  template <class predicate, class left, class right>			\
   struct name {								\
-    template <class _predicate,class _left, class _right>		\
     static auto accelerator_inline					\
-    func(const _predicate &pred, const _left &lhs, const _right &rhs)	\
+    func(const predicate &pred, const left &lhs, const right &rhs)	\
       -> decltype(combination) const					\
     {									\
       return combination;						\
@@ -395,17 +275,17 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
   };
 
 GridTrinOpClass(TrinaryWhere,
-		(predicatedWhere<
-		 typename std::remove_reference<_predicate>::type, 
-		 typename std::remove_reference<_left>::type,
-		 typename std::remove_reference<_right>::type>(pred, lhs,rhs)));
+		(predicatedWhere<predicate, 
+		 typename std::remove_reference<left>::type,
+		 typename std::remove_reference<right>::type>(pred, lhs,rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
 ////////////////////////////////////////////
-#define GRID_UNOP(name)   name
-#define GRID_BINOP(name)  name
-#define GRID_TRINOP(name) name
+
+#define GRID_UNOP(name)   name<decltype(eval(0, arg))>
+#define GRID_BINOP(name)  name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
+#define GRID_TRINOP(name) name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
 
 #define GRID_DEF_UNOP(op, name)						\
   template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
@@ -451,17 +331,22 @@ GridTrinOpClass(TrinaryWhere,
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
-//GRID_DEF_UNOP(adj, UnaryAdj);
-//GRID_DEF_UNOP(conjugate, UnaryConj);
+GRID_DEF_UNOP(adj, UnaryAdj);
+GRID_DEF_UNOP(conjugate, UnaryConj);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
+GRID_DEF_UNOP(real, UnaryReal);
+GRID_DEF_UNOP(imag, UnaryImag);
+GRID_DEF_UNOP(toReal, UnaryToReal);
+GRID_DEF_UNOP(toComplex, UnaryToComplex);
 GRID_DEF_UNOP(timesI, UnaryTimesI);
 GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
 GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the
                                // abs-fabs-dabs-labs thing
 GRID_DEF_UNOP(sqrt, UnarySqrt);
+GRID_DEF_UNOP(rsqrt, UnaryRsqrt);
 GRID_DEF_UNOP(sin, UnarySin);
 GRID_DEF_UNOP(cos, UnaryCos);
 GRID_DEF_UNOP(asin, UnaryAsin);
@@ -486,36 +371,29 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
 {
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > ret(expr);
+  Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type >
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> 
 {
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
+  Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
-				   vecEval(0, expr.arg2),
-				   vecEval(0, expr.arg3)))>::type >
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
+				   eval(0, expr.arg2),
+				   eval(0, expr.arg3)))> 
 {
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
-				vecEval(0, expr.arg2),
-			        vecEval(0, expr.arg3)))>::type >  ret(expr);
+  Lattice<decltype(expr.op.func(eval(0, expr.arg1),
+				eval(0, expr.arg2),
+				eval(0, expr.arg3)))>  ret(expr);
   return ret;
 }
-#define EXPRESSION_CLOSURE(function)					\
-  template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> \
-    auto function(Expression &expr) -> decltype(function(closure(expr))) \
-  {									\
-    return function(closure(expr));					\
-  }
-
 
 #undef GRID_UNOP
 #undef GRID_BINOP

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();
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  auto rhs_v = rhs.AcceleratorView(ViewRead);
   conformable(ret,rhs);
   conformable(lhs,rhs);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
@@ -56,13 +56,13 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  auto rhs_v = rhs.AcceleratorView(ViewRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     auto rhs_t=rhs_v(ss);
-    auto tmp  =ret_v(ss);
     mac(&tmp,&lhs_t,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
   });
@@ -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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  auto rhs_v = rhs.AcceleratorView(ViewRead);
   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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  auto rhs_v = rhs.AcceleratorView(ViewRead);
   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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     mult(&tmp,&lhs_v(ss),&rhs);
@@ -121,10 +121,10 @@ 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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
+    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     mac(&tmp,&lhs_t,&rhs);
     coalescedWrite(ret_v[ss],tmp);
@@ -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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
   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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
   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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto rhs_v = lhs.AcceleratorView(ViewRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -179,10 +179,10 @@ 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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto rhs_v = lhs.AcceleratorView(ViewRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
+    decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
     mac(&tmp,&lhs,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
@@ -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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto rhs_v = lhs.AcceleratorView(ViewRead);
   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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto rhs_v = lhs.AcceleratorView(ViewRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -221,11 +221,11 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( x_v , x, AcceleratorRead);
-  autoView( y_v , y, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto x_v = x.AcceleratorView(ViewRead);
+  auto y_v = y.AcceleratorView(ViewRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-    auto tmp = a*coalescedRead(x_v[ss])+coalescedRead(y_v[ss]);
+    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);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( x_v , x, AcceleratorRead);
-  autoView( y_v , y, AcceleratorRead);
+  auto ret_v = ret.AcceleratorView(ViewWrite);
+  auto x_v = x.AcceleratorView(ViewRead);
+  auto y_v = y.AcceleratorView(ViewRead);
   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

@@ -29,7 +29,6 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
-
 #pragma once 
 
 #define STREAMING_STORES
@@ -38,6 +37,180 @@ 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
@@ -73,40 +246,38 @@ private:
       dealloc();
       
       this->_odata_size = size;
-      if ( size )
+      if ( size ) 
 	this->_odata      = alloc.allocate(this->_odata_size);
       else 
 	this->_odata      = nullptr;
     }
   }
 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();
-  }
-
-  // Helper function to print the state of this object in the AccCache
-  void PrintCacheState(void)
-  {
-    MemoryManager::PrintState(this->_odata);
-  }
-
   /////////////////////////////////////////////////////////////////////////////////
   // 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 (ViewMode mode) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
+  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> accessor(*( (LatticeAccelerator<vobj> *) this));
+    accessor.AcceleratorPrefetch(mode);
+    return accessor;
+  }
+
+  LatticeView<vobj> HostView(int mode = ViewReadWrite) const 
+  {
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
+    accessor.HostPrefetch(mode);
+    return accessor;
+  }
+  
   ~Lattice() { 
     if ( this->_odata_size ) {
       dealloc();
@@ -126,16 +297,12 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-    
-    auto exprCopy = expr;
-    ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      auto tmp = eval(ss,exprCopy);
-      coalescedWrite(me[ss],tmp);
+
+    auto me  = AcceleratorView(ViewWrite);
+    accelerator_for(ss,me.size(),1,{
+      auto tmp = eval(ss,expr);
+      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)
@@ -150,15 +317,11 @@ public:
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
-    auto exprCopy = expr;
-    ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      auto tmp = eval(ss,exprCopy);
-      coalescedWrite(me[ss],tmp);
+    auto me  = AcceleratorView(ViewWrite);
+    accelerator_for(ss,me.size(),1,{
+      auto tmp = eval(ss,expr);
+      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)
@@ -172,15 +335,11 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-    auto exprCopy = expr;
-    ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      auto tmp = eval(ss,exprCopy);
-      coalescedWrite(me[ss],tmp);
+    auto me  = AcceleratorView(ViewWrite);
+    accelerator_for(ss,me.size(),1,{
+      auto tmp = eval(ss,expr);
+      vstream(me[ss],tmp);
     });
-    me.ViewClose();
-    ExpressionViewClose(exprCopy);
     return *this;
   }
   //GridFromExpression is tricky to do
@@ -231,11 +390,10 @@ public:
   }
 
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
-    auto me  = View(CpuWrite);
+    auto me  = View();
     thread_for(ss,me.size(),{
-	me[ss]= r;
+      me[ss] = r;
     });
-    me.ViewClose();
     return *this;
   }
 
@@ -245,12 +403,11 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
+  Lattice(GridBase *grid) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
     this->checkerboard=0;
-    SetViewMode(mode);
   }
   
   //  virtual ~Lattice(void) = default;
@@ -288,12 +445,11 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWriteDiscard);
-    auto him= r.View(AcceleratorRead);
+    auto me =   AcceleratorView(ViewWrite);
+    auto him= r.AcceleratorView(ViewRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
-    me.ViewClose();    him.ViewClose();
     return *this;
   }
 
@@ -303,12 +459,11 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
-    auto him= r.View(AcceleratorRead);
+    auto me =   AcceleratorView(ViewWrite);
+    auto him= r.AcceleratorView(ViewRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
-    me.ViewClose();    him.ViewClose();
     return *this;
   }
   ///////////////////////////////////////////

Grid/lattice/Lattice_basis.h

@@ -51,39 +51,34 @@ 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(AcceleratorRead)) View;
-
-  Vector<View> basis_v; basis_v.reserve(basis.size());
-  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
-  typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
+  typedef decltype(basis[0].View()) View;
+  auto tmp_v = basis[0].AcceleratorView(ViewReadWrite);
+  Vector<View> basis_v(basis.size(),tmp_v);
+  typedef typename std::remove_reference<decltype(tmp_v[0])>::type vobj;
   GridBase* grid = basis[0].Grid();
       
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorWrite));
+    basis_v[k] = basis[k].AcceleratorView(ViewReadWrite);
   }
 
-#if ( (!defined(GRID_CUDA)) )
-  int max_threads = thread_max();
-  Vector < vobj > Bt(Nm * max_threads);
+#ifndef GRID_NVCC
   thread_region
-    {
-      vobj* B = &Bt[Nm * thread_num()];
-      thread_for_in_region(ss, grid->oSites(),{
-	  for(int j=j0; j<j1; ++j) B[j]=0.;
+  {
+    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){
-	    for(int k=k0; k<k1; ++k){
-	      B[j] +=Qt(j,k) * basis_v[k][ss];
-	    }
+	for(int j=j0; j<j1; ++j){
+	  for(int k=k0; k<k1; ++k){
+	    B[j] +=Qt(j,k) * basis_v[k][ss];
 	  }
-	  for(int j=j0; j<j1; ++j){
-	    basis_v[j][ss] = B[j];
-	  }
-	});
-    }
+	}
+	for(int j=j0; j<j1; ++j){
+	  basis_v[j][ss] = B[j];
+	}
+      });
+  }
 #else
-  View *basis_vp = &basis_v[0];
-
   int nrot = j1-j0;
   if (!nrot) // edge case not handled gracefully by Cuda
     return;
@@ -95,13 +90,13 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  Vector<Coeff_t> Qt_jv(Nm*Nm);
-  Coeff_t *Qt_p = & Qt_jv[0];
+  Vector<double> Qt_jv(Nm*Nm);
+  double *Qt_p = & Qt_jv[0];
   thread_for(i,Nm*Nm,{
       int j = i/Nm;
       int k = i%Nm;
       Qt_p[i]=Qt(j,k);
-  });
+    });
 
   // Block the loop to keep storage footprint down
   for(uint64_t s=0;s<oSites;s+=siteBlock){
@@ -125,7 +120,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 
 	for(int k=k0; k<k1; ++k){
 	  auto tmp = coalescedRead(Bp[ss*nrot+j]);
-	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_vp[k][sss]));
+	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss]));
 	}
       });
 
@@ -134,44 +129,37 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	int jj  =j0+j;
 	int ss =sj/nrot;
 	int sss=ss+s;
-	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
+	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].View(AcceleratorRead)) View;
+  typedef decltype(basis[0].AcceleratorView()) View;
   typedef typename Field::vector_object vobj;
   GridBase* grid = basis[0].Grid();
 
   result.Checkerboard() = basis[0].Checkerboard();
-
-  Vector<View> basis_v; basis_v.reserve(basis.size());
+  auto result_v=result.AcceleratorView(ViewWrite);
+  Vector<View> basis_v(basis.size(),result_v);
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorRead));
+    basis_v[k] = basis[k].AcceleratorView(ViewRead);
   }
   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);
-
-  auto basis_vp=& basis_v[0];
-  autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    vobj zzz=Zero();
-    auto B=coalescedRead(zzz);
+    auto B=coalescedRead(zz);
     for(int k=k0; k<k1; ++k){
-      B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]);
+      B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
     }
     coalescedWrite(result_v[ss], B);
   });
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
 }
 
 template<class Field>

Grid/lattice/Lattice_comparison.h

@@ -42,6 +42,34 @@ NAMESPACE_BEGIN(Grid);
 
 typedef iScalar<vInteger> vPredicate ;
 
+/*
+template <class iobj, class vobj, class robj> accelerator_inline 
+vobj predicatedWhere(const iobj &predicate, const vobj &iftrue, const robj &iffalse) 
+{
+  typename std::remove_const<vobj>::type ret;
+
+  typedef typename vobj::scalar_object scalar_object;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  const int Nsimd = vobj::vector_type::Nsimd();
+
+  ExtractBuffer<Integer> mask(Nsimd);
+  ExtractBuffer<scalar_object> truevals(Nsimd);
+  ExtractBuffer<scalar_object> falsevals(Nsimd);
+
+  extract(iftrue, truevals);
+  extract(iffalse, falsevals);
+  extract<vInteger, Integer>(TensorRemove(predicate), mask);
+
+  for (int s = 0; s < Nsimd; s++) {
+    if (mask[s]) falsevals[s] = truevals[s];
+  }
+
+  merge(ret, falsevals);
+  return ret;
+}
+*/
 //////////////////////////////////////////////////////////////////////////
 // compare lattice to lattice
 //////////////////////////////////////////////////////////////////////////
@@ -50,9 +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());
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( rhs_v, rhs, CpuRead);
-  autoView( ret_v, ret, CpuWrite);
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
   thread_for( ss, rhs_v.size(), {
       ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
   });
@@ -65,8 +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());
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( ret_v, ret, CpuWrite);
+  auto lhs_v = lhs.View();
+  auto ret_v = ret.View();
   thread_for( ss, lhs_v.size(), {
     ret_v[ss]=op(lhs_v[ss],rhs);
   });
@@ -79,8 +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());
-  autoView( rhs_v, rhs, CpuRead);
-  autoView( ret_v, ret, CpuWrite);
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
   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();
 
-  autoView(l_v, l, CpuWrite);
+  auto l_v = l.View();
   thread_for( o, grid->oSites(), {
     vector_type vI;
     Coordinate gcoor;
@@ -51,5 +51,23 @@ 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());
-  autoView( rhs_v , rhs, AcceleratorRead);
-  autoView( ret_v , ret, AcceleratorWrite);
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
   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());
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  autoView( ret_v , ret, AcceleratorWrite);
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
   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());
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  autoView( ret_v , ret, AcceleratorWrite);
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
   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];
-  autoView( X_v , X, CpuRead);
-  autoView( Y_v , Y, CpuRead);
-  autoView( R_v , R, CpuWrite);
+  auto X_v = X.View();
+  auto Y_v = Y.View();
+  auto R_v = R.View();
   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];
 
-  autoView( X_v , X, CpuRead);
-  autoView( R_v , R, CpuWrite);
+  auto X_v = X.View();
+  auto R_v = R.View();
 
   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;
-  autoView( lhs_v , lhs, CpuRead);
-  autoView( rhs_v , rhs, CpuRead);
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
   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();
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
-  accelerator_for( ss, lhs_v.size(), 1, {
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  thread_for( ss, lhs_v.size(), {
     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();
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
-  accelerator_for( ss, lhs_v.size(), 1, {
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  thread_for( ss, lhs_v.size(), {
     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)
 {
-  autoView( rhs_v, rhs, AcceleratorRead);
-  autoView( lhs_v, lhs, AcceleratorWrite);
-  accelerator_for( ss, lhs_v.size(), 1, {
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for( ss, lhs_v.size(), {
     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)
 {
-  autoView( rhs_v, rhs, AcceleratorRead);
-  autoView( lhs_v, lhs, AcceleratorWrite);
-  accelerator_for( ss, lhs_v.size(), 1, {
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for( ss, lhs_v.size(), {
     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);
-  autoView( l_v , l, CpuWrite);
+  auto l_v = l.View();
   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);
-  autoView( l_v , l, CpuWrite);
+  auto l_v = l.View();
   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 LatticeView<vobj> &l,Coordinate &site)
-{
-  GridBase *grid = l.getGrid();
-  assert(l.mode==CpuRead);
+inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
+        
+  GridBase *grid = l.Grid();
+
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
+  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -173,7 +173,8 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  scalar_type * vp = (scalar_type *)&l[odx];
+  auto l_v = l.View();
+  scalar_type * vp = (scalar_type *)&l_v[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
@@ -182,27 +183,18 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
       
   return;
 };
-template<class vobj,class sobj>
-inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
-{
-  autoView(lv,l,CpuRead);
-  peekLocalSite(s,lv,site);
-  return;
-};
 
-// Must be CPU write view
 template<class vobj,class sobj>
-inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
-{
-  GridBase *grid=l.getGrid();
-  assert(l.mode==CpuWrite);
+inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
+
+  GridBase *grid=l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
+  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -210,19 +202,13 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  scalar_type * vp = (scalar_type *)&l[odx];
+  auto l_v = l.View();
+  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;
-};
 
-template<class vobj,class sobj>
-inline void pokeLocalSite(const sobj &s, Lattice<vobj> &l,Coordinate &site)
-{
-  autoView(lv,l,CpuWrite);
-  pokeLocalSite(s,lv,site);
   return;
 };
 

Grid/lattice/Lattice_real_imag.h

@@ -1,79 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_reality.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-
-    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_LATTICE_REAL_IMAG_H
-#define GRID_LATTICE_REAL_IMAG_H
-
-
-// FIXME .. this is the sector of the code 
-// I am most worried about the directions
-// The choice of burying complex in the SIMD
-// is making the use of "real" and "imag" very cumbersome
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj> inline Lattice<vobj> real(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard()=lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] =real(lhs_v[ss]);
-  });
-  return ret;
-};
-template<class vobj> inline Lattice<vobj> imag(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard()=lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] =imag(lhs_v[ss]);
-  });
-  return ret;
-};
-
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-  auto real(const Expression &expr) -> decltype(real(closure(expr)))		
-{									
-  return real(closure(expr));					
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-  auto imag(const Expression &expr) -> decltype(imag(closure(expr)))		
-{									
-  return imag(closure(expr));					
-}
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/lattice/Lattice_reality.h

@@ -40,77 +40,26 @@ 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();
-  accelerator_for( ss, lhs_v.size(), 1, {
-     ret_v[ss] = adj(lhs_v[ss]);
+  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)));
   });
   return ret;
 };
 
 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)));
   });
   return ret;
 };
 
-template<class vobj> inline Lattice<typename vobj::Complexified> toComplex(const Lattice<vobj> &lhs){
-  Lattice<typename vobj::Complexified> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard() = lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] = toComplex(lhs_v[ss]);
-  });
-  return ret;
-};
-template<class vobj> inline Lattice<typename vobj::Realified> toReal(const Lattice<vobj> &lhs){
-  Lattice<typename vobj::Realified> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard() = lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] = toReal(lhs_v[ss]);
-  });
-  return ret;
-};
-
-
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto toComplex(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return toComplex(closure(expr));
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto toReal(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return toReal(closure(expr));
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto adj(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return adj(closure(expr));
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto conjugate(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return conjugate(closure(expr));
-}
-
 NAMESPACE_END(Grid);
 
 #endif

Grid/lattice/Lattice_reduction.h

@@ -25,7 +25,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #include <Grid/Grid_Eigen_Dense.h>
 
 
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#ifdef GRID_NVCC
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
 
@@ -39,7 +39,7 @@ 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);
@@ -62,124 +62,24 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
   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);
-  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];
-  } 
-  
-  typedef typename vobj::scalar_object ssobj;
-  ssobj ret = ssum;
-  return ret;
-}
-/*
-Threaded max, don't use for now
-template<class Double>
-inline Double max(const Double *arg, Integer osites)
-{
-  //  const int Nsimd = vobj::Nsimd();
-  const int nthread = GridThread::GetThreads();
-
-  std::vector<Double> maxarray(nthread);
-  
-  thread_for(thr,nthread, {
-    int nwork, mywork, myoff;
-    nwork = osites;
-    GridThread::GetWork(nwork,thr,mywork,myoff);
-    Double max=arg[0];
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      if( arg[ss] > max ) max = arg[ss];
-    }
-    maxarray[thr]=max;
-  });
-  
-  Double tmax=maxarray[0];
-  for(int i=0;i<nthread;i++){
-    if (maxarray[i]>tmax) tmax = maxarray[i];
-  } 
-  return tmax;
-}
-*/
-template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#ifdef GRID_NVCC
   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_objectD sumD_large(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  return sumD_gpu_large(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
-
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
+  auto arg_v = arg.View();
   Integer osites = arg.Grid()->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;
-}
-
-template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu_large(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
-#endif
+  auto ssum= sum(&arg_v[0],osites);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -192,32 +92,6 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
   return real(nrm); 
 }
 
-//The global maximum of the site norm2
-template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
-{
-  typedef typename vobj::tensor_reduced vscalar;  //iScalar<iScalar<.... <vPODtype> > >
-  typedef typename vscalar::scalar_object  scalar;   //iScalar<iScalar<.... <PODtype> > >
-
-  Lattice<vscalar> inner = localNorm2(arg);
-
-  auto grid = arg.Grid();
-
-  RealD max;
-  for(int l=0;l<grid->lSites();l++){
-    Coordinate coor;
-    scalar val;
-    RealD r;
-    grid->LocalIndexToLocalCoor(l,coor);
-    peekLocalSite(val,inner,coor);
-    r=real(TensorRemove(val));
-    if( (l==0) || (r>max)){
-      max=r;
-    }
-  }
-  grid->GlobalMax(max);
-  return max;
-}
-
 // Double inner product
 template<class vobj>
 inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
@@ -227,30 +101,43 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   ComplexD  nrm;
   
   GridBase *grid = left.Grid();
+  
+  // Might make all code paths go this way.
+  auto left_v = left.AcceleratorView(ViewRead);
+  auto right_v=right.AcceleratorView(ViewRead);
 
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-  // Might make all code paths go this way.
-  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
+#ifdef GRID_NVCC
+  // GPU - SIMT lane compliance...
+  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);
+  
 
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, 1,{
-	auto x_l = left_v[ss];
-	auto y_l = right_v[ss];
-	inner_tmp_v[ss]=innerProductD(x_l,y_l);
-    });
-  }
+  accelerator_for( ss, sites, nsimd,{
+      auto x_l = left_v(ss);
+      auto y_l = right_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
+  })
 
   // This is in single precision and fails some tests
-  auto anrm = sum(inner_tmp_v,sites);  
-  nrm = anrm;
+  // Need a sumD that sums in double
+  nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));  
+#else
+  // CPU 
+  typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+  
+  accelerator_for( ss, sites, nsimd,{
+      auto x_l = left_v[ss];
+      auto y_l = right_v[ss];
+      inner_tmp_v[ss]=innerProductD(x_l,y_l);
+  })
+  nrm = TensorRemove(sum(inner_tmp_v,sites));
+#endif
   return nrm;
 }
 
@@ -288,24 +175,40 @@ 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
-  autoView( x_v, x, AcceleratorRead);
-  autoView( y_v, y, AcceleratorRead);
-  autoView( z_v, z, AcceleratorWrite);
-
-  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
+  typedef decltype(innerProduct(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, 1,{
-      auto tmp = a*x_v[ss]+b*y_v[ss];
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
+  });
+
+  nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
+#else
+  // CPU 
+  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+  
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
       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; 
 }
@@ -321,29 +224,47 @@ 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(innerProductD(vobj(),vobj())) inner_t;
-  typedef decltype(innerProductD(vobj(),vobj())) norm_t;
+  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
+  typedef decltype(innerProduct(left_v[0],left_v[0])) 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]);
@@ -386,7 +307,6 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   // But easily avoided by using double precision fields
   ///////////////////////////////////////////////////////
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_object::scalar_type scalar_type;
   GridBase  *grid = Data.Grid();
   assert(grid!=NULL);
 
@@ -415,7 +335,7 @@ 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
-  autoView( Data_v, Data, CpuRead);
+  auto Data_v=Data.View();
   thread_for( r,rd, {
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
     for(int n=0;n<e1;n++){
@@ -445,19 +365,20 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   }
   
   // sum over nodes.
+  sobj gsum;
   for(int t=0;t<fd;t++){
     int pt = t/ld; // processor plane
     int lt = t%ld;
     if ( pt == grid->_processor_coor[orthogdim] ) {
-      result[t]=lsSum[lt];
+      gsum=lsSum[lt];
     } else {
-      result[t]=Zero();
+      gsum=Zero();
     }
 
+    grid->GlobalSum(gsum);
+
+    result[t]=gsum;
   }
-  scalar_type * ptr = (scalar_type *) &result[0];
-  int words = fd*sizeof(sobj)/sizeof(scalar_type);
-  grid->GlobalSumVector(ptr, words);
 }
 
 template<class vobj>
@@ -492,8 +413,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
 
-  autoView( lhv, lhs, CpuRead);
-  autoView( rhv, rhs, CpuRead);
+  auto lhv=lhs.View();
+  auto rhv=rhs.View();
   thread_for( r,rd,{
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
@@ -600,12 +521,14 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
 
     tensor_reduced at; at=av;
 
-    autoView( Rv, R, CpuWrite);
-    autoView( Xv, X, CpuRead);
-    autoView( Yv, Y, CpuRead);
-    thread_for2d( n, e1, b,e2, {
+    auto Rv=R.View();
+    auto Xv=X.View();
+    auto Yv=Y.View();
+    thread_for_collapse(2, n, e1, {
+      for(int b=0;b<e2;b++){
 	int ss= so+n*stride+b;
 	Rv[ss] = at*Xv[ss]+Yv[ss];
+      }
     });
   }
 };
@@ -658,9 +581,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  autoView( X_v, X, CpuRead);
-  autoView( Y_v, Y, CpuRead);
-  autoView( R_v, R, CpuWrite);
+  auto X_v=X.View();
+  auto Y_v=Y.View();
+  auto R_v=R.View();
   thread_region
   {
     Vector<vobj> s_x(Nblock);
@@ -705,14 +628,13 @@ 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];
-  autoView( R_v, R, CpuWrite);
-  autoView( X_v, X, CpuRead);
+  auto R_v = R.View();
+  auto X_v = X.View();
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -770,8 +692,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
 
   typedef typename vobj::vector_typeD vector_typeD;
 
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( rhs_v, rhs, CpuRead);
+  auto lhs_v=lhs.View();
+  auto rhs_v=rhs.View();
   thread_region
   {
     std::vector<vobj> Left(Nblock);

Grid/lattice/Lattice_reduction_gpu.h

@@ -1,14 +1,7 @@
 NAMESPACE_BEGIN(Grid);
 
-#ifdef GRID_HIP
-extern hipDeviceProp_t *gpu_props;
-#define WARP_SIZE 64
-#endif
-#ifdef GRID_CUDA
-extern cudaDeviceProp *gpu_props;
 #define WARP_SIZE 32
-#endif
-
+extern cudaDeviceProp *gpu_props;
 __device__ unsigned int retirementCount = 0;
 
 template <class Iterator>
@@ -23,27 +16,23 @@ unsigned int nextPow2(Iterator x) {
 }
 
 template <class Iterator>
-int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+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;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  /*  
+  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
+  std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  */  
+  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -51,14 +40,10 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   
   // let the number of threads in a block be a multiple of 2, starting from warpSize
   threads = warpSize;
-  if ( threads*sizeofsobj > sharedMemPerBlock ) {
-    std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
-    return 0;
-  }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  return 1;
+  
 }
 
 template <class sobj, class Iterator>
@@ -68,7 +53,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
   
   // cannot use overloaded operators for sobj as they are not volatile-qualified
   memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
-  acceleratorSynchronise();
+  __syncwarp();
   
   const Iterator VEC = WARP_SIZE;
   const Iterator vid = tid & (VEC-1);
@@ -82,9 +67,9 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
       beta += temp;
       memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
     }
-    acceleratorSynchronise();
+    __syncwarp();
   }
-  acceleratorSynchroniseAll();
+  __syncthreads();
   
   if (threadIdx.x == 0) {
     beta  = Zero();
@@ -94,7 +79,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
     }
     memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
   }
-  acceleratorSynchroniseAll();
+  __syncthreads();
 }
 
 
@@ -162,7 +147,7 @@ __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
-    acceleratorFence();
+    __threadfence();
     
     if (tid==0) {
       unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
@@ -171,8 +156,8 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     }
     
     // each thread must read the correct value of amLast
-    acceleratorSynchroniseAll();
-
+    __syncthreads();
+    
     if (amLast) {
       // reduce buffer[0], ..., buffer[gridDim.x-1]
       Iterator i = tid;
@@ -198,7 +183,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -207,67 +192,23 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   Integer size = osites*nsimd;
 
   Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
-
+  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
   Integer smemSize = numThreads * sizeof(sobj);
 
   Vector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
   
   reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
-  accelerator_barrier();
+  cudaDeviceSynchronize();
+  
+  cudaError err = cudaGetLastError();
+  if ( cudaSuccess != err ) {
+    printf("Cuda error %s\n",cudaGetErrorString( err ));
+    exit(0);
+  }
   auto result = buffer_v[0];
   return result;
 }
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<vector> buffer(osites);
-  vector *dat = (vector *)lat;
-  vector *buf = &buffer[0];
-  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,1,{
-	buf[ss] = dat[ss*words+w];
-      });
-      
-    ret_p[w] = sumD_gpu_small(tbuf,osites);
-  }
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  
-  Integer nsimd= vobj::Nsimd();
-  Integer size = osites*nsimd;
-  Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  
-  if ( ok ) {
-    ret = sumD_gpu_small(lat,osites);
-  } else {
-    ret = sumD_gpu_large(lat,osites);
-  }
-  return ret;
-}
-
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Return as same precision as input performing reduction in double precision though
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -280,13 +221,6 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
+
 
 NAMESPACE_END(Grid);

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);
 
-    autoView(l_v, l, CpuWrite);
+    auto l_v = l.View();
     thread_for( ss, osites, {
       ExtractBuffer<scalar_object> buf(Nsimd);
       for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
@@ -461,8 +461,8 @@ public:
     }
 
     {
-      // Obtain one reseeded generator per thread      
-      int Nthread = 32; // Hardwire a good level or parallelism
+      // Obtain one reseeded generator per thread
+      int Nthread = GridThread::GetThreads();
       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());
-  autoView(ret_v , ret, AcceleratorWrite);
-  autoView(lhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
   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());
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
   });

Grid/lattice/Lattice_transfer.h

@@ -47,12 +47,11 @@ 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;
 
-  autoView( half_v, half, CpuWrite);
-  autoView( full_v, full, CpuRead);
+  auto half_v = half.View();
+  auto full_v = full.View();
   thread_for(ss, full.Grid()->oSites(),{
     int cbos;
     Coordinate coor;
@@ -65,11 +64,11 @@ 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();
-  autoView( half_v , half, CpuRead);
-  autoView( full_v , full, CpuWrite);
+  auto half_v = half.View();
+  auto full_v = full.View();
   thread_for(ss,full.Grid()->oSites(),{
 
     Coordinate coor;
@@ -85,76 +84,6 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
   });
 }
 
-template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
-{
-  half.Checkerboard() = cb;
-  autoView(half_v, half, AcceleratorWrite);
-  autoView(full_v, full, AcceleratorRead);
-  Coordinate rdim_full             = full.Grid()->_rdimensions;
-  Coordinate rdim_half             = half.Grid()->_rdimensions;
-  unsigned long ndim_half          = half.Grid()->_ndimension;
-  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
-  Coordinate ostride_half          = half.Grid()->_ostride;
-  accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{
-    
-    Coordinate coor;
-    int cbos;
-    int linear=0;
-
-    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
-    assert(coor.size()==ndim_half);
-
-    for(int d=0;d<ndim_half;d++){ 
-      if(checker_dim_mask_half[d]) linear += coor[d];
-    }
-    cbos = (linear&0x1);
-
-    if (cbos==cb) {
-      int ssh=0;
-      for(int d=0;d<ndim_half;d++) {
-        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
-        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
-      }
-      coalescedWrite(half_v[ssh],full_v(ss));
-    }
-  });
-}
-template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
-{
-  int cb = half.Checkerboard();
-  autoView(half_v , half, AcceleratorRead);
-  autoView(full_v , full, AcceleratorWrite);
-  Coordinate rdim_full             = full.Grid()->_rdimensions;
-  Coordinate rdim_half             = half.Grid()->_rdimensions;
-  unsigned long ndim_half          = half.Grid()->_ndimension;
-  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
-  Coordinate ostride_half          = half.Grid()->_ostride;
-  accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{
-
-    Coordinate coor;
-    int cbos;
-    int linear=0;
-  
-    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
-    assert(coor.size()==ndim_half);
-
-    for(int d=0;d<ndim_half;d++){ 
-      if(checker_dim_mask_half[d]) linear += coor[d];
-    }
-    cbos = (linear&0x1);
-
-    if (cbos==cb) {
-      int ssh=0;
-      for(int d=0;d<ndim_half;d++){
-        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
-        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
-      }
-      coalescedWrite(full_v[ss],half_v(ssh));
-    }
-
-  });
-}
-
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Flexible Type Conversion for internal promotion to double as well as graceful
 // treatment of scalar-compatible types
@@ -167,29 +96,15 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> &
   out = in;
 }
 
-template<typename T>
-accelerator_inline EnableIf<isGridFundamental<T>> convertType(T & out, const T & in) {
-  out = in;
-}
-
-// This would allow for conversions between GridFundamental types, but is not strictly needed as yet
-/*template<typename T1, typename T2>
-accelerator_inline typename std::enable_if<isGridFundamental<T1>::value && isGridFundamental<T2>::value>::type
-// Or to make this very broad, conversions between anything that's not a GridTensor could be allowed
-//accelerator_inline typename std::enable_if<!isGridTensor<T1>::value && !isGridTensor<T2>::value>::type
-convertType(T1 & out, const T2 & in) {
-  out = in;
-}*/
-
-#ifdef GRID_SIMT
+#ifdef __CUDA_ARCH__
 accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
-  ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
+  ((ComplexF*)&out)[SIMTlane(vComplexF::Nsimd())] = in;
 }
 accelerator_inline void convertType(vComplexD & out, const ComplexD & in) {
-  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd())] = in;
+  ((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd())] = in;
 }
 accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
-  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd()*2)] = in;
+  ((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd()*2)] = in;
 }
 #endif
 
@@ -201,18 +116,18 @@ accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
   Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
 }
 
-template<typename T1,typename T2>
-accelerator_inline void convertType(iScalar<T1> & out, const iScalar<T2> & in) {
-  convertType(out._internal,in._internal);
-}
+template<typename T1,typename T2,int N>
+  accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in);
+template<typename T1,typename T2,int N>
+  accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in);
 
-template<typename T1,typename T2>
-accelerator_inline NotEnableIf<isGridScalar<T1>> convertType(T1 & out, const iScalar<T2> & in) {
+template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
+accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
   convertType(out,in._internal);
 }
 
 template<typename T1,typename T2>
-accelerator_inline NotEnableIf<isGridScalar<T2>> convertType(iScalar<T1> & out, const T2 & in) {
+accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
   convertType(out._internal,in);
 }
 
@@ -229,13 +144,19 @@ accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & i
     convertType(out._internal[i],in._internal[i]);
 }
 
+template<typename T, typename std::enable_if<isGridFundamental<T>::value, T>::type* = nullptr>
+accelerator_inline void convertType(T & out, const T & in) {
+  out = in;
+}
+
 template<typename T1,typename T2>
 accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
-  autoView( out_v , out,AcceleratorWrite);
-  autoView( in_v  , in ,AcceleratorRead);
+  auto out_v = out.AcceleratorView(ViewWrite);
+  auto in_v  = in.AcceleratorView(ViewRead);
+
   accelerator_for(ss,out_v.size(),T1::Nsimd(),{
       convertType(out_v[ss],in_v(ss));
-  });
+    });
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////
@@ -243,20 +164,19 @@ 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(CpuRead)[0],rhs.View(CpuRead)[0])))>>
+-> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View()[0],rhs.View()[0])))>>
 {
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
+  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  auto rhs_v = rhs.AcceleratorView(ViewRead);
 
   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);
 
-  {
-    autoView(ret_v, ret,AcceleratorWrite);
-    accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
+  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
       convertType(ret_v[ss],innerProductD2(lhs_v(ss),rhs_v(ss)));
     });
-  }
+
   return ret;
 }
 
@@ -274,13 +194,14 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
   Lattice<iScalar<CComplex>> ip(coarse);
   Lattice<vobj>     fineDataRed = fineData;
 
-  autoView( coarseData_ , coarseData, AcceleratorWrite);
-  autoView( ip_         , ip,         AcceleratorWrite);
+  //  auto fineData_   = fineData.View();
+  auto coarseData_ = coarseData.AcceleratorView(ViewWrite);
+  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>
@@ -289,6 +210,68 @@ 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,
@@ -315,12 +298,10 @@ template<class vobj,class vobj2,class CComplex>
     assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
-  autoView( fineZ_  , fineZ, AcceleratorWrite);
-  autoView( fineX_  , fineX, AcceleratorRead);
-  autoView( fineY_  , fineY, AcceleratorRead);
-  autoView( coarseA_, coarseA, AcceleratorRead);
-  Coordinate fine_rdimensions = fine->_rdimensions;
-  Coordinate coarse_rdimensions = coarse->_rdimensions;
+  auto fineZ_  = fineZ.AcceleratorView(ViewWrite);
+  auto fineX_  = fineX.AcceleratorView(ViewRead);
+  auto fineY_  = fineY.AcceleratorView(ViewRead);
+  auto coarseA_= coarseA.AcceleratorView(ViewRead);
 
   accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
 
@@ -328,12 +309,12 @@ template<class vobj,class vobj2,class CComplex>
       Coordinate coor_c(_ndimension);
       Coordinate coor_f(_ndimension);
 
-      Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+      Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
       for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-      Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
+      Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
       // z = A x + y
-#ifdef GRID_SIMT
+#ifdef __CUDA_ARCH__
       typename vobj2::tensor_reduced::scalar_object cA;
       typename vobj::scalar_object cAx;
 #else
@@ -363,16 +344,15 @@ 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<vobj>(fineX,fineY);
+  fine_inner = localInnerProductD(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  {
-    autoView( CoarseInner_  , CoarseInner,AcceleratorWrite);
-    autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
-    accelerator_for(ss, coarse->oSites(), 1, {
+  accelerator_for(ss, coarse->oSites(), 1, {
       convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
     });
-  }
  
 }
 
@@ -390,15 +370,14 @@ 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);
-  {
-    autoView( CoarseInner_  , CoarseInner, AcceleratorWrite);
-    autoView( coarse_inner_ , coarse_inner, AcceleratorRead);
-    accelerator_for(ss, coarse->oSites(), 1, {
-	CoarseInner_[ss] = coarse_inner_[ss];
-    });
-  }
+  accelerator_for(ss, coarse->oSites(), 1, {
+    CoarseInner_[ss] = coarse_inner_[ss];
+  });
 }
 
 template<class vobj,class CComplex>
@@ -429,27 +408,16 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   }
   int blockVol = fine->oSites()/coarse->oSites();
 
-  // Turn this around to loop threaded over sc and interior loop 
-  // over sf would thread better
-  autoView( coarseData_ , coarseData, AcceleratorWrite);
-  autoView( fineData_   , fineData, AcceleratorRead);
+  auto coarseData_ = coarseData.AcceleratorView(ViewReadWrite);
+  auto fineData_   = fineData.AcceleratorView(ViewRead);
 
-  auto coarseData_p = &coarseData_[0];
-  auto fineData_p = &fineData_[0];
-  
-  Coordinate fine_rdimensions = fine->_rdimensions;
-  Coordinate coarse_rdimensions = coarse->_rdimensions;
-
-  vobj zz = Zero();
-  
   accelerator_for(sc,coarse->oSites(),1,{
 
       // One thread per sub block
       Coordinate coor_c(_ndimension);
-      Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
+      Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions);  // Block coordinate
+      coarseData_[sc]=Zero();
 
-      vobj cd = zz;
-      
       for(int sb=0;sb<blockVol;sb++){
 
 	int sf;
@@ -457,13 +425,11 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 	Coordinate coor_f(_ndimension);
 	Lexicographic::CoorFromIndex(coor_b,sb,block_r);               // Block sub coordinate
 	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);
+	Lexicographic::IndexFromCoor(coor_f,sf,fine->_rdimensions);
 
-	cd=cd+fineData_p[sf];
+	coarseData_[sc]=coarseData_[sc]+fineData_[sf];
       }
 
-      coarseData_p[sc] = cd;
-
     });
   return;
 }
@@ -544,8 +510,8 @@ 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];
   }
-  autoView( fineData_   , fineData, AcceleratorWrite);
-  autoView( coarseData_ , coarseData, AcceleratorRead);
+  auto fineData_   = fineData.View();
+  auto coarseData_ = coarseData.View();
 
   // Loop with a cache friendly loop ordering
   accelerator_for(sf,fine->oSites(),1,{
@@ -558,7 +524,7 @@ 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],  );*/
+      auto basis_ = Basis[i].View();
       if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
       else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
     }
@@ -577,14 +543,7 @@ 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);
-
-    //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);
+    auto  ip_ =  ip.AcceleratorView(ViewRead);
     blockZAXPY(fineData,ip,Basis[i],fineData);
   }
 }
@@ -612,17 +571,15 @@ 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_v,lcoor);
+    peekLocalSite(s,in,lcoor);
     ss=s;
-    pokeLocalSite(ss,out_v,lcoor);
+    pokeLocalSite(ss,out,lcoor);
   });
 }
 
@@ -657,9 +614,8 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate rdt = Tg->_rdimensions;
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
-
-  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, AcceleratorRead);
+  auto t_v = To.AcceleratorView(ViewWrite);
+  auto f_v = From.AcceleratorView(ViewRead);
   accelerator_for(idx,Fg->lSites(),1,{
     sobj s;
     Coordinate Fcoor(nd);
@@ -682,6 +638,8 @@ 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);
     }
   });
 }
@@ -712,8 +670,6 @@ 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);
@@ -726,8 +682,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,lowDimv,lcoor);
-    pokeLocalSite(s,higherDimv,hcoor);
+    peekLocalSite(s,lowDim,lcoor);
+    pokeLocalSite(s,higherDim,hcoor);
   });
 }
 
@@ -755,8 +711,6 @@ 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);
@@ -769,8 +723,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,higherDimv,hcoor);
-    pokeLocalSite(s,lowDimv,lcoor);
+    peekLocalSite(s,higherDim,hcoor);
+    pokeLocalSite(s,lowDim,lcoor);
   });
 
 }
@@ -798,8 +752,6 @@ 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);
@@ -808,8 +760,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,lowDimv,lcoor);
-      pokeLocalSite(s,higherDimv,hcoor);
+      peekLocalSite(s,lowDim,lcoor);
+      pokeLocalSite(s,higherDim,hcoor);
     }
   });
 }
@@ -837,8 +789,6 @@ 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);
@@ -847,8 +797,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,higherDimv,hcoor);
-      pokeLocalSite(s,lowDimv,lcoor);
+      peekLocalSite(s,higherDim,hcoor);
+      pokeLocalSite(s,lowDim,lcoor);
     }
   });
 }
@@ -912,7 +862,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
   }
 
   //loop over outer index
-  autoView( in_v  , in, CpuRead);
+  auto in_v  = in.View();
   thread_for(in_oidx,in_grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> out_ptrs(in_nsimd);
@@ -1005,7 +955,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
     icoor[lane].resize(ndim);
     grid->iCoorFromIindex(icoor[lane],lane);
   }
-  autoView( out_v , out, CpuWrite);
+  auto out_v = out.View();
   thread_for(oidx, grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> ptrs(nsimd);
@@ -1108,7 +1058,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   std::vector<SobjOut> in_slex_conv(in_grid->lSites());
   unvectorizeToLexOrdArray(in_slex_conv, in);
     
-  autoView( out_v , out, CpuWrite);
+  auto out_v = out.View();
   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());
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
   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());
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
   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());
-  autoView( rhs, rhs_i, AcceleratorRead);
-  autoView( ret, ret_i, AcceleratorWrite);
+  auto rhs = rhs_i.View();
+  auto ret = ret_i.View();
   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());
-  autoView( rhs , rhs_i, AcceleratorRead);
-  autoView( ret , ret_i, AcceleratorWrite);
+  auto rhs = rhs_i.View();
+  auto ret = ret_i.View();
   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());
-  autoView( ret , ret_i, AcceleratorWrite);
-  autoView( rhs , rhs_i, AcceleratorRead);
+  auto ret = ret_i.View();
+  auto rhs = rhs_i.View();
   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());
-  autoView( rhs , rhs_i, AcceleratorRead);
-  autoView( ret , ret_i, AcceleratorWrite);
+  auto rhs = rhs_i.View();
+  auto ret = ret_i.View();
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));

Grid/lattice/Lattice_view.h

@@ -1,173 +0,0 @@
-#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
-
-#if 1
-  //  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i) const { return this->_odata[i]; };
-#else
-  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
-#endif
-  
-  accelerator_inline uint64_t begin(void) const { return 0;};
-  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
-  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/lattice/Lattice_where.h

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

Grid/log/Log.h

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

Grid/parallelIO/BinaryIO.cc

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

Grid/parallelIO/BinaryIO.h

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

Grid/parallelIO/IldgIO.h

@@ -123,7 +123,7 @@ assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
  ////////////////////////////////////////////////////////////
  // Helper to fill out metadata
  ////////////////////////////////////////////////////////////
-template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
+ template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
 					  FieldMetaData &header,
 					  scidacRecord & _scidacRecord,
 					  scidacFile   & _scidacFile) 
@@ -576,8 +576,6 @@ class ScidacReader : public GridLimeReader {
     std::string rec_name(ILDG_BINARY_DATA);
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
       if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
-  // in principle should do the line below, but that breaks backard compatibility with old data
-  // skipPastObjectRecord(std::string(GRID_FIELD_NORM));
 	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
 	return;
       }
@@ -621,12 +619,12 @@ class IldgWriter : public ScidacWriter {
   // Don't require scidac records EXCEPT checksum
   // Use Grid MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
+  template <class vsimd>
+  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<vLorentzColourMatrixD> GaugeField;
-    typedef vLorentzColourMatrixD vobj;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -638,9 +636,6 @@ class IldgWriter : public ScidacWriter {
 
     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
 
-    stats Stats;
-    Stats(Umu,header);
-    
     std::string format = header.floating_point;
     header.ensemble_id    = description;
     header.ensemble_label = description;
@@ -710,10 +705,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
+  template <class vsimd>
+  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
 
-    typedef Lattice<vLorentzColourMatrixD > GaugeField;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
     typedef typename GaugeField::vector_object  vobj;
     typedef typename vobj::scalar_object sobj;
 
@@ -926,8 +921,7 @@ class IldgReader : public GridLimeReader {
 
     if ( found_FieldMetaData || found_usqcdInfo ) {
       FieldMetaData checker;
-      stats Stats;
-      Stats(Umu,checker);
+      GaugeStatistics(Umu,checker);
       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;

Grid/parallelIO/MetaData.h

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

Grid/parallelIO/NerscIO.h

@@ -40,8 +40,6 @@ using namespace Grid;
 class NerscIO : public BinaryIO { 
 public:
 
-  typedef Lattice<vLorentzColourMatrixD> GaugeField;
-
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -131,12 +129,12 @@ public:
   // Now the meat: the object readers
   /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void readConfiguration(GaugeField &Umu,
+  template<class vsimd>
+  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
 				       FieldMetaData& header,
-				       std::string file,
-				       GaugeStats GaugeStatisticsCalculator=GaugeStats())
+				       std::string file)
   {
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
     GridBase *grid = Umu.Grid();
     uint64_t offset = readHeader(file,Umu.Grid(),header);
@@ -155,23 +153,23 @@ public:
     // munger is a function of <floating point, Real, data_type>
     if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3D> 
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
     } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixD>
+	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
@@ -179,7 +177,7 @@ public:
       assert(0);
     }
 
-    GaugeStats Stats; Stats(Umu,clone);
+    GaugeStatistics(Umu,clone);
 
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl;
@@ -205,22 +203,15 @@ public:
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
-  // Preferred interface
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
-					std::string file, 
-					std::string ens_label = std::string("DWF"))
-  {
-    writeConfiguration(Umu,file,0,1,ens_label);
-  }
-  template<class GaugeStats=PeriodicGaugeStatistics>
-  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
+  template<class vsimd>
+  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
 					std::string file, 
 					int two_row,
-					int bits32,
-					std::string ens_label = std::string("DWF"))
+					int bits32)
   {
-    typedef vLorentzColourMatrixD vobj;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+
+    typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
@@ -228,8 +219,8 @@ public:
     // Following should become arguments
     ///////////////////////////////////////////
     header.sequence_number = 1;
-    header.ensemble_id     = std::string("UKQCD");
-    header.ensemble_label  = ens_label;
+    header.ensemble_id     = "UKQCD";
+    header.ensemble_label  = "DWF";
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -238,28 +229,28 @@ public:
 
     GridMetaData(grid,header);
     assert(header.nd==4);
-    GaugeStats Stats; Stats(Umu,header);
+    GaugeStatistics(Umu,header);
     MachineCharacteristics(header);
 
-    uint64_t offset;
+	uint64_t offset;
 
     // Sod it -- always write 3x3 double
     header.floating_point = std::string("IEEE64BIG");
     header.data_type      = std::string("4D_SU3_GAUGE_3x3");
     GaugeSimpleUnmunger<fobj3D,sobj> munge;
-    if ( grid->IsBoss() ) { 
-      truncate(file);
-      offset = writeHeader(header,file);
-    }
-    grid->Broadcast(0,(void *)&offset,sizeof(offset));
+	if ( grid->IsBoss() ) { 
+	  truncate(file);
+    offset = writeHeader(header,file);
+	}
+	grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
 					      nersc_csum,scidac_csuma,scidac_csumb);
     header.checksum = nersc_csum;
-    if ( grid->IsBoss() ) { 
-      writeHeader(header,file);
-    }
+	if ( grid->IsBoss() ) { 
+    writeHeader(header,file);
+	}
 
     std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
 	     <<std::hex<<header.checksum

Grid/parallelIO/OpenQcdIO.h

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

Grid/parallelIO/OpenQcdIOChromaReference.h

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

Grid/perfmon/PerfCount.h

@@ -44,7 +44,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <sys/syscall.h>
 #endif
 #ifdef __x86_64__
-#ifdef GRID_CUDA
+#ifdef GRID_NVCC
 accelerator_inline uint64_t __rdtsc(void) {  return 0; }
 accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
 #else
@@ -112,6 +112,7 @@ 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_CUDA
+#ifdef GRID_NVCC
 #pragma pop
 #endif
 

Grid/qcd/QCD.h

@@ -47,7 +47,7 @@ static constexpr int Ym = 5;
 static constexpr int Zm = 6;
 static constexpr int Tm = 7;
 
-static constexpr int Nc=Config_Nc;
+static constexpr int Nc=3;
 static constexpr int Ns=4;
 static constexpr int Nd=4;
 static constexpr int Nhs=2; // half spinor
@@ -80,13 +80,6 @@ template<typename T> struct isSpinor {
 template <typename T> using IfSpinor    = Invoke<std::enable_if< isSpinor<T>::value,int> > ;
 template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ;
 
-const int CoarseIndex = 4;
-template<typename T> struct isCoarsened {
-   static constexpr bool value = (CoarseIndex<=T::TensorLevel);
-};
-template <typename T> using IfCoarsened    = Invoke<std::enable_if< isCoarsened<T>::value,int> > ;
-template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ;
-
 // ChrisK very keen to add extra space for Gparity doubling.
 //
 // Also add domain wall index, in a way where Wilson operator 

Grid/qcd/action/ActionBase.h

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

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -1,240 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h
-
-    Copyright (C) 2020 - 2022
-
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Nils Meyer <nils.meyer@ur.de>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-/*  END LEGAL */
-
-#pragma once
-
-#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-
-NAMESPACE_BEGIN(Grid);
-
-// see Grid/qcd/action/fermion/WilsonCloverFermion.h for description
-//
-// Modifications done here:
-//
-// Original: clover term = 12x12 matrix per site
-//
-// But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running)
-// Sufficient to store/transfer only the real parts of the diagonal and one triangular part
-// 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered
-//
-// Here: Above but diagonal as complex numbers, i.e., need to store/transfer
-// 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words
-//
-// Words per site and improvement compared to original (combined with the input and output spinors):
-//
-// - Original: 2*12 + 12*12 = 168 words -> 1.00 x less
-// - Minimal:  2*12 + 36    =  60 words -> 2.80 x less
-// - Here:     2*12 + 42    =  66 words -> 2.55 x less
-//
-// These improvements directly translate to wall-clock time
-//
-// Data layout:
-//
-// - diagonal and triangle part as separate lattice fields,
-//   this was faster than as 1 combined field on all tested machines
-// - diagonal: as expected
-// - triangle: store upper right triangle in row major order
-// - graphical:
-//        0  1  2  3  4
-//           5  6  7  8
-//              9 10 11 = upper right triangle indices
-//                12 13
-//                   14
-//     0
-//        1
-//           2
-//              3       = diagonal indices
-//                 4
-//                    5
-//     0
-//     1  5
-//     2  6  9          = lower left triangle indices
-//     3  7 10 12
-//     4  8 11 13 14
-//
-// Impact on total memory consumption:
-// - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site
-// - Here:     (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts =  24 complex words per site
-//           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
-//                                                                 =  84 complex words per site
-
-template<class Impl>
-class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
-                                   public WilsonCloverHelpers<Impl>,
-                                   public CompactWilsonCloverHelpers<Impl> {
-  /////////////////////////////////////////////
-  // Sizes
-  /////////////////////////////////////////////
-
-public:
-
-  INHERIT_COMPACT_CLOVER_SIZES(Impl);
-
-  /////////////////////////////////////////////
-  // Type definitions
-  /////////////////////////////////////////////
-
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  typedef WilsonFermion<Impl>              WilsonBase;
-  typedef WilsonCloverHelpers<Impl>        Helpers;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  /////////////////////////////////////////////
-  // Constructors
-  /////////////////////////////////////////////
-
-public:
-
-  CompactWilsonCloverFermion(GaugeField& _Umu,
-			    GridCartesian& Fgrid,
-			    GridRedBlackCartesian& Hgrid,
-			    const RealD _mass,
-			    const RealD _csw_r = 0.0,
-			    const RealD _csw_t = 0.0,
-			    const RealD _cF = 1.0,
-			    const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
-			    const ImplParams& impl_p = ImplParams());
-
-  /////////////////////////////////////////////
-  // Member functions (implementing interface)
-  /////////////////////////////////////////////
-
-public:
-
-  virtual void Instantiatable() {};
-  int          ConstEE()     override { return 0; };
-  int          isTrivialEE() override { return 0; };
-
-  void Dhop(const FermionField& in, FermionField& out, int dag) override;
-
-  void DhopOE(const FermionField& in, FermionField& out, int dag) override;
-
-  void DhopEO(const FermionField& in, FermionField& out, int dag) override;
-
-  void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
-
-  void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
-
-  void M(const FermionField& in, FermionField& out) override;
-
-  void Mdag(const FermionField& in, FermionField& out) override;
-
-  void Meooe(const FermionField& in, FermionField& out) override;
-
-  void MeooeDag(const FermionField& in, FermionField& out) override;
-
-  void Mooee(const FermionField& in, FermionField& out) override;
-
-  void MooeeDag(const FermionField& in, FermionField& out) override;
-
-  void MooeeInv(const FermionField& in, FermionField& out) override;
-
-  void MooeeInvDag(const FermionField& in, FermionField& out) override;
-
-  void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
-
-  void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
-
-  void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
-
-  void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
-
-  void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
-
-  /////////////////////////////////////////////
-  // Member functions (internals)
-  /////////////////////////////////////////////
-
-  void MooeeInternal(const FermionField&        in,
-                     FermionField&              out,
-                     const CloverDiagonalField& diagonal,
-                     const CloverTriangleField& triangle);
-
-  /////////////////////////////////////////////
-  // Helpers
-  /////////////////////////////////////////////
-
-  void ImportGauge(const GaugeField& _Umu) override;
-
-  /////////////////////////////////////////////
-  // Helpers
-  /////////////////////////////////////////////
-
-private:
-
-  template<class Field>
-  const MaskField* getCorrectMaskField(const Field &in) const {
-    if(in.Grid()->_isCheckerBoarded) {
-      if(in.Checkerboard() == Odd) {
-        return &this->BoundaryMaskOdd;
-      } else {
-        return &this->BoundaryMaskEven;
-      }
-    } else {
-      return &this->BoundaryMask;
-    }
-  }
-
-  template<class Field>
-  void ApplyBoundaryMask(Field& f) {
-    const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
-    assert(m != nullptr);
-    CompactHelpers::ApplyBoundaryMask(f, *m);
-  }
-
-  /////////////////////////////////////////////
-  // Member Data
-  /////////////////////////////////////////////
-
-public:
-
-  RealD csw_r;
-  RealD csw_t;
-  RealD cF;
-
-  bool open_boundaries;
-
-  CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
-  CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
-
-  CloverTriangleField Triangle,    TriangleEven,    TriangleOdd;
-  CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
-
-  FermionField Tmp;
-
-  MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
-};
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/DomainWallVec5dImpl.h

@@ -114,22 +114,19 @@ public:
       U = adj(Cshift(U, mu, -1));
       PokeIndex<LorentzIndex>(Uadj, U, mu);
     }
-
-    autoView(Umu_v,Umu,CpuRead);
-    autoView(Uadj_v,Uadj,CpuRead);
-    autoView(Uds_v,Uds,CpuWrite);
-    thread_for( lidx, GaugeGrid->lSites(), {
+    
+    for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
       Coordinate lcoor;
       GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
       
-      peekLocalSite(ScalarUmu, Umu_v, lcoor);
+      peekLocalSite(ScalarUmu, Umu, lcoor);
       for (int mu = 0; mu < 4; mu++) ScalarUds(mu) = ScalarUmu(mu);
       
-      peekLocalSite(ScalarUmu, Uadj_v, lcoor);
+      peekLocalSite(ScalarUmu, Uadj, lcoor);
       for (int mu = 0; mu < 4; mu++) ScalarUds(mu + 4) = ScalarUmu(mu);
       
-      pokeLocalSite(ScalarUds, Uds_v, lcoor);
-    });
+      pokeLocalSite(ScalarUds, Uds, lcoor);
+    }
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu) 

Grid/qcd/action/fermion/Fermion.h

@@ -53,12 +53,10 @@ NAMESPACE_CHECK(Wilson);
 #include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(WilsonTM);
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
-#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
 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);
@@ -116,9 +114,9 @@ typedef WilsonFermion<WilsonImplR> WilsonFermionR;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
-//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
-//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
-//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
+typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
+typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
+typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
 
 typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
@@ -154,63 +152,46 @@ typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoInd
 typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
 typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
-// Compact Clover fermions
-typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
-typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
-typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
-
-typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
-typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
-typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
-
-typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
-
-typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
-
 // Domain Wall fermions
 typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
 
-//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
-//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
-//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
+typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
+typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
+typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
 
 typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
-//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
-//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
+typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
+typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
+typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
 
 typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
-//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
-//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
+typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
+typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
+typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
 
 typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
-//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
-//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
+typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
+typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
+typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
 
 typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
-//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
-//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
+typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
+typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
+typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
 
 // Ls vectorised
 typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
@@ -253,62 +234,64 @@ typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 
-//typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
-//typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
-//typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
+typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
+typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
+typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
 
 typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
-//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
-//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
+typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
+typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
+typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
 
 typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
-//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
-//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
+typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
+typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
+typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
 
 typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
-//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
-//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
+typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
+typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
+typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
 
 typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
-//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
-//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
+typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
+typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
+typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
 
 typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 
-//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
-//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
-//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
+typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
+typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
+typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
 
 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
+typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR;
+typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF;
+typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD;
+#endif
+
 NAMESPACE_END(Grid);
 
 ////////////////////

Grid/qcd/action/fermion/FermionOperatorImpl.h

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

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -96,31 +96,43 @@ public:
     int sl        = St._simd_layout[direction];
     Coordinate icoor;
 
-#ifdef GRID_SIMT
+#ifdef __CUDA_ARCH__
+    _Spinor tmp;
+
     const int Nsimd =SiteDoubledGaugeField::Nsimd();
-    int s = acceleratorSIMTlane(Nsimd);
+    int s = SIMTlane(Nsimd);
     St.iCoorFromIindex(icoor,s);
 
     int mmu = mu % Nd;
+    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
+      
+      int permute_lane = (sl==1) 
+    	|| ((distance== 1)&&(icoor[direction]==1))
+	|| ((distance==-1)&&(icoor[direction]==0));
 
-    auto UU0=coalescedRead(U(0)(mu));
-    auto UU1=coalescedRead(U(1)(mu));
-    
-    //Decide whether we do a G-parity flavor twist
-    //Note: this assumes (but does not check) that sl==1 || sl==2 i.e. max 2 SIMD lanes in G-parity dir
-    //It also assumes (but does not check) that abs(distance) == 1
-    int permute_lane = (sl==1) 
-    || ((distance== 1)&&(icoor[direction]==1))
-    || ((distance==-1)&&(icoor[direction]==0));
+      if ( permute_lane ) { 
+	tmp(0) = chi(1);
+	tmp(1) = chi(0);
+      } else {
+	tmp(0) = chi(0);
+	tmp(1) = chi(1);
+      }
 
-    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
+      auto UU0=coalescedRead(U(0)(mu));
+      auto UU1=coalescedRead(U(1)(mu));
 
-    //Apply the links
-    int f_upper = permute_lane ? 1 : 0;
-    int f_lower = !f_upper;
+      mult(&phi(0),&UU0,&tmp(0));
+      mult(&phi(1),&UU1,&tmp(1));
 
-    mult(&phi(0),&UU0,&chi(f_upper));
-    mult(&phi(1),&UU1,&chi(f_lower));
+    } else {
+
+      auto UU0=coalescedRead(U(0)(mu));
+      auto UU1=coalescedRead(U(1)(mu));
+
+      mult(&phi(0),&UU0,&chi(0));
+      mult(&phi(1),&UU1,&chi(1));
+
+    }
 
 #else
     typedef _Spinor vobj;
@@ -220,17 +232,15 @@ public:
       if ( Params.twists[mu] ) { 
 	Uconj = where(coor==neglink,-Uconj,Uconj);
       }
-
-      {
-	autoView( U_v , U, CpuRead);
-	autoView( Uconj_v , Uconj, CpuRead);
-	autoView( Uds_v , Uds, CpuWrite);
-	autoView( Utmp_v, Utmp, CpuWrite);
-	thread_foreach(ss,U_v,{
-	    Uds_v[ss](0)(mu) = U_v[ss]();
-	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
-	  });
-      }
+	  
+      auto U_v = U.View();
+      auto Uds_v = Uds.View();
+      auto Uconj_v = Uconj.View();
+      auto Utmp_v= Utmp.View();
+      thread_foreach(ss,U_v,{
+	Uds_v[ss](0)(mu) = U_v[ss]();
+	Uds_v[ss](1)(mu) = Uconj_v[ss]();
+      });
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
       Uconj = adj(Cshift(Uconj,mu,-1));
@@ -240,25 +250,19 @@ public:
 	Utmp = where(coor==0,Uconj,Utmp);
       }
 
-      {
-	autoView( Uds_v , Uds, CpuWrite);
-	autoView( Utmp_v, Utmp, CpuWrite);
-	thread_foreach(ss,Utmp_v,{
-	    Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
-	  });
-      }
+      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);
       }
-
-      {	  
-	autoView( Uds_v , Uds, CpuWrite);
-	autoView( Utmp_v, Utmp, CpuWrite);
-	thread_foreach(ss,Utmp_v,{
-	    Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
-        });
-      }
+	  
+      thread_foreach(ss,Utmp_v,{
+        Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
+      });
+          
     }
   }
       
@@ -268,14 +272,11 @@ public:
     GaugeLinkField link(mat.Grid());
     // use lorentz for flavour as hack.
     auto tmp = TraceIndex<SpinIndex>(outerProduct(Btilde, A));
-
-    {
-      autoView( link_v , link, CpuWrite);
-      autoView( tmp_v , tmp, CpuRead);
-      thread_foreach(ss,tmp_v,{
-	  link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
-	});
-    }
+    auto link_v = link.View();
+    auto tmp_v = tmp.View();
+    thread_foreach(ss,tmp_v,{
+      link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
+    });
     PokeIndex<LorentzIndex>(mat, link, mu);
     return;
   }
@@ -305,18 +306,16 @@ public:
         
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    {
-      autoView( tmp_v , tmp, CpuWrite);
-      autoView( Atilde_v , Atilde, CpuRead);
-      autoView( Btilde_v , Btilde, CpuRead);
-      thread_for(ss,tmp.Grid()->oSites(),{
-	  for (int s = 0; s < Ls; s++) {
-	    int sF = s + Ls * ss;
-	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
-	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
-	  }
-	});
-    }
+    auto tmp_v = tmp.View();
+    auto Atilde_v = Atilde.View();
+    auto Btilde_v = Btilde.View();
+    thread_for(ss,tmp.Grid()->oSites(),{
+      for (int s = 0; s < Ls; s++) {
+	int sF = s + Ls * ss;
+	auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+	tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+      }
+    });
     PokeIndex<LorentzIndex>(mat, tmp, mu);
     return;
   }
@@ -327,8 +326,8 @@ typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffReal> Gparit
 typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffReal> GparityWilsonImplF;  // Float
 typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffReal> GparityWilsonImplD;  // Double
  
-//typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
-//typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
-//typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
+typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
+typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
+typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -61,8 +61,8 @@ public:
   double DhopCalls;
   double DhopCommTime;
   double DhopComputeTime;
-  double DhopComputeTime2;
-  double DhopFaceTime;
+      double DhopComputeTime2;
+      double DhopFaceTime;
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
@@ -208,7 +208,7 @@ public:
   LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
-  //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
+  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
     
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities

Grid/qcd/action/fermion/MADWF.h

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

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -1,194 +0,0 @@
-/*************************************************************************************
-
-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/StaggeredImpl.h

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

Grid/qcd/action/fermion/StaggeredKernels.h

@@ -47,37 +47,23 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
   INHERIT_IMPL_TYPES(Impl);
   typedef FermionOperator<Impl> Base;
    
- public:
-
-  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
-		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
-		    const FermionField &in, FermionField &out, int dag, int interior,int exterior);
-  void DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeField &U,
-		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
-  
-  void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
-		     int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
- protected:    
+public:
+    
+   void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+		      int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Generic Nc kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   template<int Naik> 
-   static accelerator_inline
-   void DhopSiteGeneric(StencilView &st, 
+   void DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
 			DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteGenericInt(StencilView &st, 
+   void DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteGenericExt(StencilView &st, 
+   void DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -85,21 +71,15 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    ///////////////////////////////////////////////////////////////////////////////////////
    // Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteHand(StencilView &st, 
+   void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
 		     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		     SiteSpinor * buf, int LLs, int sU, 
 		     const FermionFieldView &in, FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteHandInt(StencilView &st, 
+   void DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   
-   template<int Naik> static accelerator_inline
-   void DhopSiteHandExt(StencilView &st, 
+   void DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -107,11 +87,27 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    ///////////////////////////////////////////////////////////////////////////////////////
    // Asm Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   
-   void DhopSiteAsm(StencilView &st, 
+   void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
 		    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

@@ -113,7 +113,20 @@ public:
       
   inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
   {
-    assert(0);
+    GridBase *GaugeGrid = U_ds.Grid();
+    thread_for(lidx, GaugeGrid->lSites(),{
+
+	SiteScalarGaugeLink   ScalarU;
+	SiteDoubledGaugeField ScalarUds;
+	
+	Coordinate lcoor;
+	GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
+	peekLocalSite(ScalarUds, U_ds, lcoor);
+	
+	peekLocalSite(ScalarU, U, lcoor);
+	ScalarUds(mu) = ScalarU();
+	
+    });
   }
   inline void DoubleStore(GridBase *GaugeGrid,
 			  DoubledGaugeField &UUUds, // for Naik term

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -4,11 +4,10 @@
 
     Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
 
-    Copyright (C) 2017 - 2022
+    Copyright (C) 2017
 
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: David Preti <>
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -30,8 +29,7 @@
 
 #pragma once
 
-#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -52,15 +50,18 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////
 
 template <class Impl>
-class WilsonCloverFermion : public WilsonFermion<Impl>,
-                            public WilsonCloverHelpers<Impl>
+class WilsonCloverFermion : public WilsonFermion<Impl>
 {
 public:
+  // Types definitions
   INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
+  template <typename vtype>
+  using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+  typedef iImplClover<Simd> SiteCloverType;
+  typedef Lattice<SiteCloverType> CloverFieldType;
 
-  typedef WilsonFermion<Impl>       WilsonBase;
-  typedef WilsonCloverHelpers<Impl> Helpers;
+public:
+  typedef WilsonFermion<Impl> WilsonBase;
 
   virtual int    ConstEE(void)     { return 0; };
   virtual void Instantiatable(void){};
@@ -71,7 +72,42 @@ public:
                       const RealD _csw_r = 0.0,
                       const RealD _csw_t = 0.0,
                       const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
-                      const ImplParams &impl_p = ImplParams());
+                      const ImplParams &impl_p = ImplParams()) : WilsonFermion<Impl>(_Umu,
+                                                                                     Fgrid,
+                                                                                     Hgrid,
+                                                                                     _mass, impl_p, clover_anisotropy),
+                                                                 CloverTerm(&Fgrid),
+                                                                 CloverTermInv(&Fgrid),
+                                                                 CloverTermEven(&Hgrid),
+                                                                 CloverTermOdd(&Hgrid),
+                                                                 CloverTermInvEven(&Hgrid),
+                                                                 CloverTermInvOdd(&Hgrid),
+                                                                 CloverTermDagEven(&Hgrid),
+                                                                 CloverTermDagOdd(&Hgrid),
+                                                                 CloverTermInvDagEven(&Hgrid),
+                                                                 CloverTermInvDagOdd(&Hgrid)
+  {
+    assert(Nd == 4); // require 4 dimensions
+
+    if (clover_anisotropy.isAnisotropic)
+    {
+      csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
+      diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
+    }
+    else
+    {
+      csw_r = _csw_r * 0.5;
+      diag_mass = 4.0 + _mass;
+    }
+    csw_t = _csw_t * 0.5;
+
+    if (csw_r == 0)
+      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
+    if (csw_t == 0)
+      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
+
+    ImportGauge(_Umu);
+  }
 
   virtual void M(const FermionField &in, FermionField &out);
   virtual void Mdag(const FermionField &in, FermionField &out);
@@ -88,21 +124,249 @@ public:
   void ImportGauge(const GaugeField &_Umu);
 
   // Derivative parts unpreconditioned pseudofermions
-  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
+  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
+  {
+    conformable(X.Grid(), Y.Grid());
+    conformable(X.Grid(), force.Grid());
+    GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
+    GaugeField clover_force(force.Grid());
+    PropagatorField Lambda(force.Grid());
 
-public:
+    // Guido: Here we are hitting some performance issues:
+    // need to extract the components of the DoubledGaugeField
+    // for each call
+    // Possible solution
+    // Create a vector object to store them? (cons: wasting space)
+    std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
+
+    Impl::extractLinkField(U, this->Umu);
+
+    force = Zero();
+    // Derivative of the Wilson hopping term
+    this->DhopDeriv(force, X, Y, dag);
+
+    ///////////////////////////////////////////////////////////
+    // Clover term derivative
+    ///////////////////////////////////////////////////////////
+    Impl::outerProductImpl(Lambda, X, Y);
+    //std::cout << "Lambda:" << Lambda << std::endl;
+
+    Gamma::Algebra sigma[] = {
+        Gamma::Algebra::SigmaXY,
+        Gamma::Algebra::SigmaXZ,
+        Gamma::Algebra::SigmaXT,
+        Gamma::Algebra::MinusSigmaXY,
+        Gamma::Algebra::SigmaYZ,
+        Gamma::Algebra::SigmaYT,
+        Gamma::Algebra::MinusSigmaXZ,
+        Gamma::Algebra::MinusSigmaYZ,
+        Gamma::Algebra::SigmaZT,
+        Gamma::Algebra::MinusSigmaXT,
+        Gamma::Algebra::MinusSigmaYT,
+        Gamma::Algebra::MinusSigmaZT};
+
+    /*
+      sigma_{\mu \nu}=
+      | 0         sigma[0]  sigma[1]  sigma[2] |
+      | sigma[3]    0       sigma[4]  sigma[5] |
+      | sigma[6]  sigma[7]     0      sigma[8] |
+      | sigma[9]  sigma[10] sigma[11]   0      |
+    */
+
+    int count = 0;
+    clover_force = Zero();
+    for (int mu = 0; mu < 4; mu++)
+    {
+      force_mu = Zero();
+      for (int nu = 0; nu < 4; nu++)
+      {
+        if (mu == nu)
+        continue;
+        
+        RealD factor;
+        if (nu == 4 || mu == 4)
+        {
+          factor = 2.0 * csw_t;
+        }
+        else
+        {
+          factor = 2.0 * csw_r;
+        }
+        PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
+        Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
+        force_mu -= factor*Cmunu(U, lambda, mu, nu);                   // checked
+        count++;
+      }
+
+      pokeLorentz(clover_force, U[mu] * force_mu, mu);
+    }
+    //clover_force *= csw;
+    force += clover_force;
+  }
+
+  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
+  GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
+  {
+    conformable(lambda.Grid(), U[0].Grid());
+    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
+    // insertion in upper staple
+    // please check redundancy of shift operations
+
+    // C1+
+    tmp = lambda * U[nu];
+    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C2+
+    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C3+
+    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
+
+    // C4+
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
+
+    // insertion in lower staple
+    // C1-
+    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C2-
+    tmp = adj(lambda) * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C3-
+    tmp = lambda * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
+
+    // C4-
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
+
+    return out;
+  }
+
+private:
   // here fixing the 4 dimensions, make it more general?
 
   RealD csw_r;                                               // Clover coefficient - spatial
   RealD csw_t;                                               // Clover coefficient - temporal
   RealD diag_mass;                                           // Mass term
-  CloverField CloverTerm, CloverTermInv;                     // Clover term
-  CloverField CloverTermEven, CloverTermOdd;                 // Clover term EO
-  CloverField CloverTermInvEven, CloverTermInvOdd;           // Clover term Inv EO
-  CloverField CloverTermDagEven, CloverTermDagOdd;           // Clover term Dag EO
-  CloverField CloverTermInvDagEven, CloverTermInvDagOdd;     // Clover term Inv Dag EO
-};
+  CloverFieldType CloverTerm, CloverTermInv;                 // Clover term
+  CloverFieldType CloverTermEven, CloverTermOdd;             // Clover term EO
+  CloverFieldType CloverTermInvEven, CloverTermInvOdd;       // Clover term Inv EO
+  CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term Dag EO
+  CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
 
+  // eventually these can be compressed into 6x6 blocks instead of the 12x12
+  // using the DeGrand-Rossi basis for the gamma matrices
+  CloverFieldType fillCloverYZ(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+    auto T_v = T.View();
+    auto F_v = F.View();
+    thread_for(i, CloverTerm.Grid()->oSites(),
+    {
+      T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
+      T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
+      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverXZ(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+    
+    auto T_v = T.View();
+    auto F_v = F.View();
+    thread_for(i, CloverTerm.Grid()->oSites(),
+    {
+      T_v[i]()(0, 1) = -F_v[i]()();
+      T_v[i]()(1, 0) = F_v[i]()();
+      T_v[i]()(2, 3) = -F_v[i]()();
+      T_v[i]()(3, 2) = F_v[i]()();
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverXY(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+
+    auto T_v = T.View();
+    auto F_v = F.View();
+    thread_for(i, CloverTerm.Grid()->oSites(),
+    {
+      T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
+      T_v[i]()(1, 1) = timesI(F_v[i]()());
+      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 3) = timesI(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverXT(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+
+    auto T_v = T.View();
+    auto F_v = F.View();
+    thread_for(i, CloverTerm.Grid()->oSites(),
+    {
+      T_v[i]()(0, 1) = timesI(F_v[i]()());
+      T_v[i]()(1, 0) = timesI(F_v[i]()());
+      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverYT(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+    T = Zero();
+    
+    auto T_v = T.View();
+    auto F_v = F.View();
+    thread_for(i, CloverTerm.Grid()->oSites(),
+    {
+      T_v[i]()(0, 1) = -(F_v[i]()());
+      T_v[i]()(1, 0) = (F_v[i]()());
+      T_v[i]()(2, 3) = (F_v[i]()());
+      T_v[i]()(3, 2) = -(F_v[i]()());
+    });
+
+    return T;
+  }
+
+  CloverFieldType fillCloverZT(const GaugeLinkField &F)
+  {
+    CloverFieldType T(F.Grid());
+
+    T = Zero();
+
+    auto T_v = T.View();
+    auto F_v = F.View();
+    thread_for(i, CloverTerm.Grid()->oSites(),
+    {
+      T_v[i]()(0, 0) = timesI(F_v[i]()());
+      T_v[i]()(1, 1) = timesMinusI(F_v[i]()());
+      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
+      T_v[i]()(3, 3) = timesI(F_v[i]()());
+    });
+
+    return T;
+  }
+};
 NAMESPACE_END(Grid);
 
 

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -1,761 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverHelpers.h
-
-    Copyright (C) 2021 - 2022
-
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#pragma once
-
-// Helper routines that implement common clover functionality
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Impl> class WilsonCloverHelpers {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
-  {
-    conformable(lambda.Grid(), U[0].Grid());
-    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
-    // insertion in upper staple
-    // please check redundancy of shift operations
-
-    // C1+
-    tmp = lambda * U[nu];
-    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C2+
-    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C3+
-    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
-
-    // C4+
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
-
-    // insertion in lower staple
-    // C1-
-    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C2-
-    tmp = adj(lambda) * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C3-
-    tmp = lambda * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
-
-    // C4-
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
-
-    return out;
-  }
-
-  static CloverField fillCloverYZ(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-    autoView(T_v,T,AcceleratorWrite);
-    autoView(F_v,F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverXZ(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-    
-    autoView(T_v, T,AcceleratorWrite);
-    autoView(F_v, F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-F_v[i]()()));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead(F_v[i]()()));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead(-F_v[i]()()));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(F_v[i]()()));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverXY(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-
-    autoView(T_v,T,AcceleratorWrite);
-    autoView(F_v,F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverXT(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-
-    autoView( T_v , T, AcceleratorWrite);
-    autoView( F_v , F, AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverYT(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-    T = Zero();
-    
-    autoView( T_v ,T,AcceleratorWrite);
-    autoView( F_v ,F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 0), coalescedRead((F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 3), coalescedRead((F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 2), coalescedRead(-(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  static CloverField fillCloverZT(const GaugeLinkField &F)
-  {
-    CloverField T(F.Grid());
-
-    T = Zero();
-
-    autoView( T_v , T,AcceleratorWrite);
-    autoView( F_v , F,AcceleratorRead);
-    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
-    {
-      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
-      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
-    });
-
-    return T;
-  }
-
-  template<class _Spinor>
-  static accelerator_inline void multClover(_Spinor& phi, const SiteClover& C, const _Spinor& chi) {
-    auto CC = coalescedRead(C);
-    mult(&phi, &CC, &chi);
-  }
-
-  template<class _SpinorField>
-  inline void multCloverField(_SpinorField& out, const CloverField& C, const _SpinorField& phi) {
-    const int Nsimd = SiteSpinor::Nsimd();
-    autoView(out_v, out, AcceleratorWrite);
-    autoView(phi_v, phi, AcceleratorRead);
-    autoView(C_v,   C,   AcceleratorRead);
-    typedef decltype(coalescedRead(out_v[0])) calcSpinor;
-    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
-      calcSpinor tmp;
-      multClover(tmp,C_v[sss],phi_v(sss));
-      coalescedWrite(out_v[sss],tmp);
-    });
-  }
-};
-
-
-template<class Impl> class CompactWilsonCloverHelpers {
-public:
-
-  INHERIT_COMPACT_CLOVER_SIZES(Impl);
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  #if 0
-  static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
-    assert(i != j);
-    if(i < j) {
-      return triangle()(block)(triangle_index(i, j));
-    } else { // i > j
-      return conjugate(triangle()(block)(triangle_index(i, j)));
-    }
-  }
-  #else
-  template<typename vobj>
-  static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
-    assert(i != j);
-    if(i < j) {
-      return triangle()(block)(triangle_index(i, j));
-    } else { // i > j
-      return conjugate(triangle()(block)(triangle_index(i, j)));
-    }
-  }
-  #endif
-
-  static accelerator_inline int triangle_index(int i, int j) {
-    if(i == j)
-      return 0;
-    else if(i < j)
-      return Nred * (Nred - 1) / 2 - (Nred - i) * (Nred - i - 1) / 2 + j - i - 1;
-    else // i > j
-      return Nred * (Nred - 1) / 2 - (Nred - j) * (Nred - j - 1) / 2 + i - j - 1;
-  }
-
-  static void MooeeKernel_gpu(int                        Nsite,
-                              int                        Ls,
-                              const FermionField&        in,
-                              FermionField&              out,
-                              const CloverDiagonalField& diagonal,
-                              const CloverTriangleField& triangle) {
-    autoView(diagonal_v, diagonal, AcceleratorRead);
-    autoView(triangle_v, triangle, AcceleratorRead);
-    autoView(in_v,       in,       AcceleratorRead);
-    autoView(out_v,      out,      AcceleratorWrite);
-
-    typedef decltype(coalescedRead(out_v[0])) CalcSpinor;
-
-    const uint64_t NN = Nsite * Ls;
-
-    accelerator_for(ss, NN, Simd::Nsimd(), {
-      int sF = ss;
-      int sU = ss/Ls;
-      CalcSpinor res;
-      CalcSpinor in_t = in_v(sF);
-      auto diagonal_t = diagonal_v(sU);
-      auto triangle_t = triangle_v(sU);
-      for(int block=0; block<Nhs; block++) {
-        int s_start = block*Nhs;
-        for(int i=0; i<Nred; i++) {
-          int si = s_start + i/Nc, ci = i%Nc;
-          res()(si)(ci) = diagonal_t()(block)(i) * in_t()(si)(ci);
-          for(int j=0; j<Nred; j++) {
-            if (j == i) continue;
-            int sj = s_start + j/Nc, cj = j%Nc;
-            res()(si)(ci) = res()(si)(ci) + triangle_elem(triangle_t, block, i, j) * in_t()(sj)(cj);
-          };
-        };
-      };
-      coalescedWrite(out_v[sF], res);
-    });
-  }
-
-  static void MooeeKernel_cpu(int                        Nsite,
-                              int                        Ls,
-                              const FermionField&        in,
-                              FermionField&              out,
-                              const CloverDiagonalField& diagonal,
-                              const CloverTriangleField& triangle) {
-    autoView(diagonal_v, diagonal, CpuRead);
-    autoView(triangle_v, triangle, CpuRead);
-    autoView(in_v,       in,       CpuRead);
-    autoView(out_v,      out,      CpuWrite);
-
-    typedef SiteSpinor CalcSpinor;
-
-#if defined(A64FX) || defined(A64FXFIXEDSIZE)
-#define PREFETCH_CLOVER(BASE) {                                     \
-    uint64_t base;                                                  \
-    int pf_dist_L1 = 1;                                             \
-    int pf_dist_L2 = -5; /* -> penalty -> disable */                \
-                                                                    \
-    if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {           \
-      base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);               \
-      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL1STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL1STRM); \
-    }                                                               \
-                                                                    \
-    if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {           \
-      base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);               \
-      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL2STRM); \
-      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL2STRM); \
-    }                                                               \
-  }
-// TODO: Implement/generalize this for other architectures
-// I played around a bit on KNL (see below) but didn't bring anything
-// #elif defined(AVX512)
-// #define PREFETCH_CLOVER(BASE) {                              \
-//     uint64_t base;                                           \
-//     int pf_dist_L1 = 1;                                      \
-//     int pf_dist_L2 = +4;                                     \
-//                                                              \
-//     if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {    \
-//       base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);        \
-//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T0); \
-//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T0); \
-//     }                                                        \
-//                                                              \
-//     if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {    \
-//       base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);        \
-//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T1); \
-//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T1); \
-//     }                                                        \
-//   }
-#else
-#define PREFETCH_CLOVER(BASE)
-#endif
-
-    const uint64_t NN = Nsite * Ls;
-
-    thread_for(ss, NN, {
-      int sF = ss;
-      int sU = ss/Ls;
-      CalcSpinor res;
-      CalcSpinor in_t = in_v[sF];
-      auto diag_t     = diagonal_v[sU]; // "diag" instead of "diagonal" here to make code below easier to read
-      auto triangle_t = triangle_v[sU];
-
-      // upper half
-      PREFETCH_CLOVER(0);
-
-      auto in_cc_0_0 = conjugate(in_t()(0)(0)); // Nils: reduces number
-      auto in_cc_0_1 = conjugate(in_t()(0)(1)); // of conjugates from
-      auto in_cc_0_2 = conjugate(in_t()(0)(2)); // 30 to 20
-      auto in_cc_1_0 = conjugate(in_t()(1)(0));
-      auto in_cc_1_1 = conjugate(in_t()(1)(1));
-
-      res()(0)(0) =               diag_t()(0)( 0) * in_t()(0)(0)
-                  +           triangle_t()(0)( 0) * in_t()(0)(1)
-                  +           triangle_t()(0)( 1) * in_t()(0)(2)
-                  +           triangle_t()(0)( 2) * in_t()(1)(0)
-                  +           triangle_t()(0)( 3) * in_t()(1)(1)
-                  +           triangle_t()(0)( 4) * in_t()(1)(2);
-
-      res()(0)(1) =           triangle_t()(0)( 0) * in_cc_0_0;
-      res()(0)(1) =               diag_t()(0)( 1) * in_t()(0)(1)
-                  +           triangle_t()(0)( 5) * in_t()(0)(2)
-                  +           triangle_t()(0)( 6) * in_t()(1)(0)
-                  +           triangle_t()(0)( 7) * in_t()(1)(1)
-                  +           triangle_t()(0)( 8) * in_t()(1)(2)
-                  + conjugate(       res()(0)( 1));
-
-      res()(0)(2) =           triangle_t()(0)( 1) * in_cc_0_0
-                  +           triangle_t()(0)( 5) * in_cc_0_1;
-      res()(0)(2) =               diag_t()(0)( 2) * in_t()(0)(2)
-                  +           triangle_t()(0)( 9) * in_t()(1)(0)
-                  +           triangle_t()(0)(10) * in_t()(1)(1)
-                  +           triangle_t()(0)(11) * in_t()(1)(2)
-                  + conjugate(       res()(0)( 2));
-
-      res()(1)(0) =           triangle_t()(0)( 2) * in_cc_0_0
-                  +           triangle_t()(0)( 6) * in_cc_0_1
-                  +           triangle_t()(0)( 9) * in_cc_0_2;
-      res()(1)(0) =               diag_t()(0)( 3) * in_t()(1)(0)
-                  +           triangle_t()(0)(12) * in_t()(1)(1)
-                  +           triangle_t()(0)(13) * in_t()(1)(2)
-                  + conjugate(       res()(1)( 0));
-
-      res()(1)(1) =           triangle_t()(0)( 3) * in_cc_0_0
-                  +           triangle_t()(0)( 7) * in_cc_0_1
-                  +           triangle_t()(0)(10) * in_cc_0_2
-                  +           triangle_t()(0)(12) * in_cc_1_0;
-      res()(1)(1) =               diag_t()(0)( 4) * in_t()(1)(1)
-                  +           triangle_t()(0)(14) * in_t()(1)(2)
-                  + conjugate(       res()(1)( 1));
-
-      res()(1)(2) =           triangle_t()(0)( 4) * in_cc_0_0
-                  +           triangle_t()(0)( 8) * in_cc_0_1
-                  +           triangle_t()(0)(11) * in_cc_0_2
-                  +           triangle_t()(0)(13) * in_cc_1_0
-                  +           triangle_t()(0)(14) * in_cc_1_1;
-      res()(1)(2) =               diag_t()(0)( 5) * in_t()(1)(2)
-                  + conjugate(       res()(1)( 2));
-
-      vstream(out_v[sF]()(0)(0), res()(0)(0));
-      vstream(out_v[sF]()(0)(1), res()(0)(1));
-      vstream(out_v[sF]()(0)(2), res()(0)(2));
-      vstream(out_v[sF]()(1)(0), res()(1)(0));
-      vstream(out_v[sF]()(1)(1), res()(1)(1));
-      vstream(out_v[sF]()(1)(2), res()(1)(2));
-
-      // lower half
-      PREFETCH_CLOVER(1);
-
-      auto in_cc_2_0 = conjugate(in_t()(2)(0));
-      auto in_cc_2_1 = conjugate(in_t()(2)(1));
-      auto in_cc_2_2 = conjugate(in_t()(2)(2));
-      auto in_cc_3_0 = conjugate(in_t()(3)(0));
-      auto in_cc_3_1 = conjugate(in_t()(3)(1));
-
-      res()(2)(0) =               diag_t()(1)( 0) * in_t()(2)(0)
-                  +           triangle_t()(1)( 0) * in_t()(2)(1)
-                  +           triangle_t()(1)( 1) * in_t()(2)(2)
-                  +           triangle_t()(1)( 2) * in_t()(3)(0)
-                  +           triangle_t()(1)( 3) * in_t()(3)(1)
-                  +           triangle_t()(1)( 4) * in_t()(3)(2);
-
-      res()(2)(1) =           triangle_t()(1)( 0) * in_cc_2_0;
-      res()(2)(1) =               diag_t()(1)( 1) * in_t()(2)(1)
-                  +           triangle_t()(1)( 5) * in_t()(2)(2)
-                  +           triangle_t()(1)( 6) * in_t()(3)(0)
-                  +           triangle_t()(1)( 7) * in_t()(3)(1)
-                  +           triangle_t()(1)( 8) * in_t()(3)(2)
-                  + conjugate(       res()(2)( 1));
-
-      res()(2)(2) =           triangle_t()(1)( 1) * in_cc_2_0
-                  +           triangle_t()(1)( 5) * in_cc_2_1;
-      res()(2)(2) =               diag_t()(1)( 2) * in_t()(2)(2)
-                  +           triangle_t()(1)( 9) * in_t()(3)(0)
-                  +           triangle_t()(1)(10) * in_t()(3)(1)
-                  +           triangle_t()(1)(11) * in_t()(3)(2)
-                  + conjugate(       res()(2)( 2));
-
-      res()(3)(0) =           triangle_t()(1)( 2) * in_cc_2_0
-                  +           triangle_t()(1)( 6) * in_cc_2_1
-                  +           triangle_t()(1)( 9) * in_cc_2_2;
-      res()(3)(0) =               diag_t()(1)( 3) * in_t()(3)(0)
-                  +           triangle_t()(1)(12) * in_t()(3)(1)
-                  +           triangle_t()(1)(13) * in_t()(3)(2)
-                  + conjugate(       res()(3)( 0));
-
-      res()(3)(1) =           triangle_t()(1)( 3) * in_cc_2_0
-                  +           triangle_t()(1)( 7) * in_cc_2_1
-                  +           triangle_t()(1)(10) * in_cc_2_2
-                  +           triangle_t()(1)(12) * in_cc_3_0;
-      res()(3)(1) =               diag_t()(1)( 4) * in_t()(3)(1)
-                  +           triangle_t()(1)(14) * in_t()(3)(2)
-                  + conjugate(       res()(3)( 1));
-
-      res()(3)(2) =           triangle_t()(1)( 4) * in_cc_2_0
-                  +           triangle_t()(1)( 8) * in_cc_2_1
-                  +           triangle_t()(1)(11) * in_cc_2_2
-                  +           triangle_t()(1)(13) * in_cc_3_0
-                  +           triangle_t()(1)(14) * in_cc_3_1;
-      res()(3)(2) =               diag_t()(1)( 5) * in_t()(3)(2)
-                  + conjugate(       res()(3)( 2));
-
-      vstream(out_v[sF]()(2)(0), res()(2)(0));
-      vstream(out_v[sF]()(2)(1), res()(2)(1));
-      vstream(out_v[sF]()(2)(2), res()(2)(2));
-      vstream(out_v[sF]()(3)(0), res()(3)(0));
-      vstream(out_v[sF]()(3)(1), res()(3)(1));
-      vstream(out_v[sF]()(3)(2), res()(3)(2));
-    });
-  }
-
-  static void MooeeKernel(int                        Nsite,
-                          int                        Ls,
-                          const FermionField&        in,
-                          FermionField&              out,
-                          const CloverDiagonalField& diagonal,
-                          const CloverTriangleField& triangle) {
-#if defined(GRID_CUDA) || defined(GRID_HIP)
-    MooeeKernel_gpu(Nsite, Ls, in, out, diagonal, triangle);
-#else
-    MooeeKernel_cpu(Nsite, Ls, in, out, diagonal, triangle);
-#endif
-  }
-
-  static void Invert(const CloverDiagonalField& diagonal,
-                     const CloverTriangleField& triangle,
-                     CloverDiagonalField&       diagonalInv,
-                     CloverTriangleField&       triangleInv) {
-    conformable(diagonal, diagonalInv);
-    conformable(triangle, triangleInv);
-    conformable(diagonal, triangle);
-
-    diagonalInv.Checkerboard() = diagonal.Checkerboard();
-    triangleInv.Checkerboard() = triangle.Checkerboard();
-
-    GridBase* grid = diagonal.Grid();
-
-    long lsites = grid->lSites();
-
-    typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
-    typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-
-    autoView(diagonal_v,  diagonal,  CpuRead);
-    autoView(triangle_v,  triangle,  CpuRead);
-    autoView(diagonalInv_v, diagonalInv, CpuWrite);
-    autoView(triangleInv_v, triangleInv, CpuWrite);
-
-    thread_for(site, lsites, { // NOTE: Not on GPU because of Eigen & (peek/poke)LocalSite
-      Eigen::MatrixXcd clover_inv_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
-      Eigen::MatrixXcd clover_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
-
-      scalar_object_diagonal diagonal_tmp     = Zero();
-      scalar_object_diagonal diagonal_inv_tmp = Zero();
-      scalar_object_triangle triangle_tmp     = Zero();
-      scalar_object_triangle triangle_inv_tmp = Zero();
-
-      Coordinate lcoor;
-      grid->LocalIndexToLocalCoor(site, lcoor);
-
-      peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-      peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-      // TODO: can we save time here by inverting the two 6x6 hermitian matrices separately?
-      for (long s_row=0;s_row<Ns;s_row++) {
-        for (long s_col=0;s_col<Ns;s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for (long c_row=0;c_row<Nc;c_row++) {
-            for (long c_col=0;c_col<Nc;c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-              else
-                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(triangle_elem(triangle_tmp, block, i, j)));
-            }
-          }
-        }
-      }
-
-      clover_inv_eigen = clover_eigen.inverse();
-
-      for (long s_row=0;s_row<Ns;s_row++) {
-        for (long s_col=0;s_col<Ns;s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for (long c_row=0;c_row<Nc;c_row++) {
-            for (long c_col=0;c_col<Nc;c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                diagonal_inv_tmp()(block)(i) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
-              else if(i < j)
-                triangle_inv_tmp()(block)(triangle_index(i, j)) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
-              else
-                continue;
-            }
-          }
-        }
-      }
-
-      pokeLocalSite(diagonal_inv_tmp, diagonalInv_v, lcoor);
-      pokeLocalSite(triangle_inv_tmp, triangleInv_v, lcoor);
-    });
-  }
-
-  static void ConvertLayout(const CloverField&   full,
-                            CloverDiagonalField& diagonal,
-                            CloverTriangleField& triangle) {
-    conformable(full, diagonal);
-    conformable(full, triangle);
-
-    diagonal.Checkerboard() = full.Checkerboard();
-    triangle.Checkerboard() = full.Checkerboard();
-
-    autoView(full_v,     full,     AcceleratorRead);
-    autoView(diagonal_v, diagonal, AcceleratorWrite);
-    autoView(triangle_v, triangle, AcceleratorWrite);
-
-    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
-    accelerator_for(ss, full.Grid()->oSites(), 1, {
-      for(int s_row = 0; s_row < Ns; s_row++) {
-        for(int s_col = 0; s_col < Ns; s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for(int c_row = 0; c_row < Nc; c_row++) {
-            for(int c_col = 0; c_col < Nc; c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                diagonal_v[ss]()(block)(i) = full_v[ss]()(s_row, s_col)(c_row, c_col);
-              else if(i < j)
-                triangle_v[ss]()(block)(triangle_index(i, j)) = full_v[ss]()(s_row, s_col)(c_row, c_col);
-              else
-                continue;
-            }
-          }
-        }
-      }
-    });
-  }
-
-
-  static void ConvertLayout(const CloverDiagonalField& diagonal,
-                            const CloverTriangleField& triangle,
-                            CloverField&               full) {
-    conformable(full, diagonal);
-    conformable(full, triangle);
-
-    full.Checkerboard() = diagonal.Checkerboard();
-
-    full = Zero();
-
-    autoView(diagonal_v, diagonal, AcceleratorRead);
-    autoView(triangle_v, triangle, AcceleratorRead);
-    autoView(full_v,     full,     AcceleratorWrite);
-
-    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
-    accelerator_for(ss, full.Grid()->oSites(), 1, {
-      for(int s_row = 0; s_row < Ns; s_row++) {
-        for(int s_col = 0; s_col < Ns; s_col++) {
-          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
-          int block       = s_row / Nhs;
-          int s_row_block = s_row % Nhs;
-          int s_col_block = s_col % Nhs;
-          for(int c_row = 0; c_row < Nc; c_row++) {
-            for(int c_col = 0; c_col < Nc; c_col++) {
-              int i = s_row_block * Nc + c_row;
-              int j = s_col_block * Nc + c_col;
-              if(i == j)
-                full_v[ss]()(s_row, s_col)(c_row, c_col) = diagonal_v[ss]()(block)(i);
-              else
-                full_v[ss]()(s_row, s_col)(c_row, c_col) = triangle_elem(triangle_v[ss], block, i, j);
-            }
-          }
-        }
-      }
-    });
-  }
-
-  static void ModifyBoundaries(CloverDiagonalField& diagonal, CloverTriangleField& triangle, RealD csw_t, RealD cF, RealD diag_mass) {
-    // Checks/grid
-    double t0 = usecond();
-    conformable(diagonal, triangle);
-    GridBase* grid = diagonal.Grid();
-
-    // Determine the boundary coordinates/sites
-    double t1 = usecond();
-    int t_dir = Nd - 1;
-    Lattice<iScalar<vInteger>> t_coor(grid);
-    LatticeCoordinate(t_coor, t_dir);
-    int T = grid->GlobalDimensions()[t_dir];
-
-    // Set off-diagonal parts at boundary to zero -- OK
-    double t2 = usecond();
-    CloverTriangleField zeroTriangle(grid);
-    zeroTriangle.Checkerboard() = triangle.Checkerboard();
-    zeroTriangle = Zero();
-    triangle = where(t_coor == 0,   zeroTriangle, triangle);
-    triangle = where(t_coor == T-1, zeroTriangle, triangle);
-
-    // Set diagonal to unity (scaled correctly) -- OK
-    double t3 = usecond();
-    CloverDiagonalField tmp(grid);
-    tmp.Checkerboard() = diagonal.Checkerboard();
-    tmp                = -1.0 * csw_t + diag_mass;
-    diagonal           = where(t_coor == 0,   tmp, diagonal);
-    diagonal           = where(t_coor == T-1, tmp, diagonal);
-
-    // Correct values next to boundary
-    double t4 = usecond();
-    if(cF != 1.0) {
-      tmp = cF - 1.0;
-      tmp += diagonal;
-      diagonal = where(t_coor == 1,   tmp, diagonal);
-      diagonal = where(t_coor == T-2, tmp, diagonal);
-    }
-
-    // Report timings
-    double t5 = usecond();
-#if 0
-    std::cout << GridLogMessage << "CompactWilsonCloverHelpers::ModifyBoundaries timings:"
-              << " checks = "          << (t1 - t0) / 1e6
-              << ", coordinate = "     << (t2 - t1) / 1e6
-              << ", off-diag zero = "  << (t3 - t2) / 1e6
-              << ", diagonal unity = " << (t4 - t3) / 1e6
-              << ", near-boundary = "  << (t5 - t4) / 1e6
-              << ", total = "          << (t5 - t0) / 1e6
-              << std::endl;
-#endif
-  }
-
-  template<class Field, class Mask>
-  static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
-    conformable(f, m);
-    auto grid  = f.Grid();
-    const uint32_t Nsite = grid->oSites();
-    const uint32_t Nsimd = grid->Nsimd();
-    autoView(f_v, f, AcceleratorWrite);
-    autoView(m_v, m, AcceleratorRead);
-    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
-    accelerator_for(ss, Nsite, Nsimd, {
-      coalescedWrite(f_v[ss], m_v(ss) * f_v(ss));
-    });
-  }
-
-  template<class MaskField>
-  static void SetupMasks(MaskField& full, MaskField& even, MaskField& odd) {
-    assert(even.Grid()->_isCheckerBoarded && even.Checkerboard() == Even);
-    assert(odd.Grid()->_isCheckerBoarded  && odd.Checkerboard()  == Odd);
-    assert(!full.Grid()->_isCheckerBoarded);
-
-    GridBase* grid = full.Grid();
-    int t_dir = Nd-1;
-    Lattice<iScalar<vInteger>> t_coor(grid);
-    LatticeCoordinate(t_coor, t_dir);
-    int T = grid->GlobalDimensions()[t_dir];
-
-    MaskField zeroMask(grid); zeroMask = Zero();
-    full = 1.0;
-    full = where(t_coor == 0,   zeroMask, full);
-    full = where(t_coor == T-1, zeroMask, full);
-
-    pickCheckerboard(Even, even, full);
-    pickCheckerboard(Odd,  odd,  full);
-  }
-};
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -1,92 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverTypes.h
-
-    Copyright (C) 2021 - 2022
-
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Impl>
-class WilsonCloverTypes {
-public:
-  INHERIT_IMPL_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-
-  typedef iImplClover<Simd> SiteClover;
-
-  typedef Lattice<SiteClover> CloverField;
-};
-
-template<class Impl>
-class CompactWilsonCloverTypes {
-public:
-  INHERIT_IMPL_TYPES(Impl);
-
-  static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
-
-  static constexpr int Nred      = Nc * Nhs;        // 6
-  static constexpr int Nblock    = Nhs;             // 2
-  static constexpr int Ndiagonal = Nred;            // 6
-  static constexpr int Ntriangle = (Nred - 1) * Nc; // 15
-
-  template<typename vtype> using iImplCloverDiagonal = iScalar<iVector<iVector<vtype, Ndiagonal>, Nblock>>;
-  template<typename vtype> using iImplCloverTriangle = iScalar<iVector<iVector<vtype, Ntriangle>, Nblock>>;
-
-  typedef iImplCloverDiagonal<Simd> SiteCloverDiagonal;
-  typedef iImplCloverTriangle<Simd> SiteCloverTriangle;
-  typedef iSinglet<Simd>            SiteMask;
-
-  typedef Lattice<SiteCloverDiagonal> CloverDiagonalField;
-  typedef Lattice<SiteCloverTriangle> CloverTriangleField;
-  typedef Lattice<SiteMask>           MaskField;
-};
-
-#define INHERIT_CLOVER_TYPES(Impl)                                 \
-  typedef typename WilsonCloverTypes<Impl>::SiteClover SiteClover; \
-  typedef typename WilsonCloverTypes<Impl>::CloverField CloverField;
-
-#define INHERIT_COMPACT_CLOVER_TYPES(Impl) \
-  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverDiagonal  SiteCloverDiagonal; \
-  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverTriangle  SiteCloverTriangle; \
-  typedef typename CompactWilsonCloverTypes<Impl>::SiteMask            SiteMask; \
-  typedef typename CompactWilsonCloverTypes<Impl>::CloverDiagonalField CloverDiagonalField; \
-  typedef typename CompactWilsonCloverTypes<Impl>::CloverTriangleField CloverTriangleField; \
-  typedef typename CompactWilsonCloverTypes<Impl>::MaskField           MaskField; \
-  /* ugly duplication but needed inside functionality classes */ \
-  template<typename vtype> using iImplCloverDiagonal = \
-    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ndiagonal>, CompactWilsonCloverTypes<Impl>::Nblock>>; \
-  template<typename vtype> using iImplCloverTriangle = \
-    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ntriangle>, CompactWilsonCloverTypes<Impl>::Nblock>>;
-
-#define INHERIT_COMPACT_CLOVER_SIZES(Impl)                                    \
-  static constexpr int Nred      = CompactWilsonCloverTypes<Impl>::Nred;      \
-  static constexpr int Nblock    = CompactWilsonCloverTypes<Impl>::Nblock;    \
-  static constexpr int Ndiagonal = CompactWilsonCloverTypes<Impl>::Ndiagonal; \
-  static constexpr int Ntriangle = CompactWilsonCloverTypes<Impl>::Ntriangle;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -61,19 +61,18 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) const {
+  accelerator_inline int CommDatumSize(void) {
     return sizeof(SiteHalfCommSpinor);
   }
 
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-    typedef decltype(coalescedRead(buf)) sobj;
-    sobj sp;
-    auto sin = coalescedRead(in);
-    projector::Proj(sp,sin,mu,dag);
-    coalescedWrite(buf,sp);
+  template<class _SiteHalfSpinor, class _SiteSpinor>
+  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
+    _SiteHalfSpinor tmp;
+    projector::Proj(tmp,in,mu,dag);
+    vstream(buf[o],tmp);
   }
 
   /*****************************************************/
@@ -82,24 +81,19 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
 				   const SiteHalfSpinor * __restrict__ vp0,
 				   const SiteHalfSpinor * __restrict__ vp1,
-				   Integer type,Integer o) const {
-#ifdef GRID_SIMT
-    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
-#else
+				   Integer type,Integer o){
     SiteHalfSpinor tmp1;
     SiteHalfSpinor tmp2;
     exchange(tmp1,tmp2,vp0[o],vp1[o],type);
     vstream(mp[2*o  ],tmp1);
     vstream(mp[2*o+1],tmp2);
-#endif
   }
 
-
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
   accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
-				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
+				     SiteHalfSpinor * __restrict__ in, Integer o) {    
     assert(0);
   }
 
@@ -109,30 +103,8 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
 					   SiteHalfSpinor * __restrict__ out1,
 					   const SiteSpinor * __restrict__ in,
-					   Integer j,Integer k, Integer m,Integer type) const
+					   Integer j,Integer k, Integer m,Integer type)
   {
-#ifdef GRID_SIMT
-    typedef SiteSpinor vobj;
-    typedef SiteHalfSpinor hvobj;
-    typedef decltype(coalescedRead(*in))    sobj;
-    typedef decltype(coalescedRead(*out0)) hsobj;
-
-    unsigned int Nsimd = vobj::Nsimd();
-    unsigned int mask = Nsimd >> (type + 1);
-    int lane = acceleratorSIMTlane(Nsimd);
-    int j0 = lane &(~mask); // inner coor zero
-    int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
-    const vobj *vp1 = &in[m];
-    const vobj *vp = (lane&mask) ? vp1:vp0;
-    auto sa = coalescedRead(*vp,j0);
-    auto sb = coalescedRead(*vp,j1);
-    hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);
-    projector::Proj(psb,sb,mu,dag);
-    coalescedWrite(out0[j],psa);
-    coalescedWrite(out1[j],psb);
-#else
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
     projector::Proj(temp1,in[k],mu,dag);
@@ -140,17 +112,15 @@ public:
     exchange(temp3,temp4,temp1,temp2,type);
     vstream(out0[j],temp3);
     vstream(out1[j],temp4);
-#endif
   }
 
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return false; }
+  accelerator_inline bool DecompressionStep(void) { return false; }
 
 };
 
-#if 0
 template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
 class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
 				typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
@@ -172,30 +142,20 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) const {
+  accelerator_inline int CommDatumSize(void) {
     return sizeof(SiteHalfCommSpinor);
   }
 
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-    SiteHalfSpinor hsp;
+  template<class _SiteHalfSpinor, class _SiteSpinor>
+  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
+    _SiteHalfSpinor hsp;
     SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
     projector::Proj(hsp,in,mu,dag);
     precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
   }
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-#ifdef GRID_SIMT
-    typedef decltype(coalescedRead(buf)) sobj;
-    sobj sp;
-    auto sin = coalescedRead(in);
-    projector::Proj(sp,sin,mu,dag);
-    coalescedWrite(buf,sp);
-#else
-    projector::Proj(buf,in,mu,dag);
-#endif
-  }
 
   /*****************************************************/
   /* Exchange includes precision change if mpi data is not same */
@@ -203,7 +163,7 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
                        SiteHalfSpinor *vp0,
                        SiteHalfSpinor *vp1,
-		       Integer type,Integer o) const {
+		       Integer type,Integer o){
     SiteHalfSpinor vt0,vt1;
     SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
     SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
@@ -215,7 +175,7 @@ public:
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
+  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o){
     SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
     precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
   }
@@ -226,7 +186,7 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
 			       SiteHalfSpinor *out1,
 			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type) const {
+			       Integer j,Integer k, Integer m,Integer type){
     SiteHalfSpinor temp1, temp2,temp3,temp4;
     SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
     SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
@@ -240,10 +200,9 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return true; }
+  accelerator_inline bool DecompressionStep(void) { return true; }
 
 };
-#endif
 
 #define DECLARE_PROJ(Projector,Compressor,spProj)			\
   class Projector {							\
@@ -294,8 +253,33 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  void ZeroCountersi(void)  {  }
-  void Reporti(int calls)  {  }
+  double timer0;
+  double timer1;
+  double timer2;
+  double timer3;
+  double timer4;
+  double timer5;
+  double timer6;
+  uint64_t callsi;
+  void ZeroCountersi(void)
+  {
+    timer0=0;
+    timer1=0;
+    timer2=0;
+    timer3=0;
+    timer4=0;
+    timer5=0;
+    timer6=0;
+    callsi=0;
+  }
+  void Reporti(int calls)
+  {
+    if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
+    if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate)   " <<timer1/calls <<std::endl;
+    if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge )   " <<timer2/calls <<std::endl;
+    if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
+    if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
+  }
 
   std::vector<int> surface_list;
 
@@ -337,18 +321,26 @@ public:
   {
     std::vector<std::vector<CommsRequest_t> > reqs;
     this->HaloExchangeOptGather(source,compress);
+    double t1=usecond();
     // Asynchronous MPI calls multidirectional, Isend etc...
     // Non-overlapped directions within a thread. Asynchronous calls except MPI3, threaded up to comm threads ways.
     this->Communicate();
+    double t2=usecond(); timer1 += t2-t1;
     this->CommsMerge(compress);
+    double t3=usecond(); timer2 += t3-t2;
     this->CommsMergeSHM(compress);
+    double t4=usecond(); timer3 += t4-t3;
   }
   
   template <class compressor>
   void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress) 
   {
     this->Prepare();
+    double t0=usecond();
     this->HaloGatherOpt(source,compress);
+    double t1=usecond();
+    timer0 += t1-t0;
+    callsi++;
   }
 
   template <class compressor>
@@ -360,9 +352,12 @@ public:
     typedef typename compressor::SiteHalfSpinor     SiteHalfSpinor;
     typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
 
+    this->mpi3synctime_g-=usecond();
     this->_grid->StencilBarrier();
+    this->mpi3synctime_g+=usecond();
 
     assert(source.Grid()==this->_grid);
+    this->halogtime-=usecond();
     
     this->u_comm_offset=0;
       
@@ -398,6 +393,7 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
+    this->halogtime+=usecond();
     accelerator_barrier();
   }
 

Grid/qcd/action/fermion/WilsonFermion.h

@@ -50,14 +50,14 @@ public:
   double, nu);
 
   WilsonAnisotropyCoefficients():
-    isAnisotropic(false),
-    t_direction(Nd-1),
-    xi_0(1.0),
+    isAnisotropic(false), 
+    t_direction(Nd-1), 
+    xi_0(1.0), 
     nu(1.0){}
 };
 
 template <class Impl>
-class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic
+class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic 
 {
 public:
   INHERIT_IMPL_TYPES(Impl);
@@ -74,20 +74,6 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
-  void Report(void);
-  void ZeroCounters(void);
-  double DhopCalls;
-  double DhopCommTime;
-  double DhopComputeTime;
-  double DhopComputeTime2;
-  double DhopFaceTime;
-  double DhopTotalTime;
-
-  double DerivCalls;
-  double DerivCommTime;
-  double DerivComputeTime;
-  double DerivDhopComputeTime;
-
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent
@@ -152,7 +138,7 @@ public:
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
                 GridRedBlackCartesian &Hgrid, RealD _mass,
-                const ImplParams &p = ImplParams(),
+                const ImplParams &p = ImplParams(), 
                 const WilsonAnisotropyCoefficients &anis = WilsonAnisotropyCoefficients() );
 
   // DoubleStore impl dependent
@@ -184,9 +170,9 @@ public:
 
   LebesgueOrder Lebesgue;
   LebesgueOrder LebesgueEvenOdd;
-
+  
   WilsonAnisotropyCoefficients anisotropyCoeff;
-
+  
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
   ///////////////////////////////////////////////////////////////
@@ -200,7 +186,7 @@ public:
                            PropagatorField &q_out,
                            PropagatorField &phys_src,
                            Current curr_type,
-                           unsigned int mu,
+                           unsigned int mu, 
                            unsigned int tmin,
 			   unsigned int tmax,
 			   ComplexField &lattice_cmplx);
@@ -210,3 +196,5 @@ typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
 NAMESPACE_END(Grid);
+
+

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -215,7 +215,7 @@ public:
   LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
-  //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
+  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
 
 
 };

Grid/qcd/action/fermion/WilsonImpl.h

@@ -72,7 +72,7 @@ public:
   typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
   typedef WilsonImplParams ImplParams;
   typedef WilsonStencil<SiteSpinor, SiteHalfSpinor,ImplParams> StencilImpl;
-  typedef const typename StencilImpl::View_type StencilView;
+  typedef typename StencilImpl::View_type StencilView;
     
   ImplParams Params;
 
@@ -106,15 +106,11 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
-    const int Nsimd = SiteHalfSpinor::Nsimd();
-    autoView( out_v, out, AcceleratorWrite);
-    autoView( phi_v, phi, AcceleratorRead);
-    autoView( Umu_v, Umu, AcceleratorRead);
-    typedef decltype(coalescedRead(out_v[0]))   calcSpinor;
-    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
-	calcSpinor tmp;
-	multLink(tmp,Umu_v[sss],phi_v(sss),mu);
-	coalescedWrite(out_v[sss],tmp);
+    auto out_v= out.View();
+    auto phi_v= phi.View();
+    auto Umu_v= Umu.View();
+    thread_for(sss,out.Grid()->oSites(),{
+	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
     });
   }
 					   
@@ -184,57 +180,29 @@ public:
       mat = TraceIndex<SpinIndex>(P); 
     }
       
-    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds)
-    {
+    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
       for (int mu = 0; mu < Nd; mu++)
       mat[mu] = PeekIndex<LorentzIndex>(Uds, mu);
     }
 
-  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu)
-  {
-#undef USE_OLD_INSERT_FORCE    
+
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
+      
     int Ls=Btilde.Grid()->_fdimensions[0];
-    autoView( mat_v , mat, AcceleratorWrite);
-#ifdef USE_OLD_INSERT_FORCE    
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    {
-      const int Nsimd = SiteSpinor::Nsimd();
-      autoView( tmp_v , tmp, AcceleratorWrite);
-      autoView( Btilde_v , Btilde, AcceleratorRead);
-      autoView( Atilde_v , Atilde, AcceleratorRead);
-      accelerator_for(sss,tmp.Grid()->oSites(),1,{
-	  int sU=sss;
-	  for(int s=0;s<Ls;s++){
-	    int sF = s+Ls*sU;
-	    tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
-	  }
-	});
-    }
+    auto tmp_v = tmp.View();
+    auto Btilde_v = Btilde.View();
+    auto Atilde_v = Atilde.View();
+    thread_for(sss,tmp.Grid()->oSites(),{
+      int sU=sss;
+      for(int s=0;s<Ls;s++){
+	int sF = s+Ls*sU;
+	tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
+      }
+    });
     PokeIndex<LorentzIndex>(mat,tmp,mu);
-#else
-    {
-      const int Nsimd = SiteSpinor::Nsimd();
-      autoView( Btilde_v , Btilde, AcceleratorRead);
-      autoView( Atilde_v , Atilde, AcceleratorRead);
-      accelerator_for(sss,mat.Grid()->oSites(),Nsimd,{
-	  int sU=sss;
-  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
-  	  ColorMatrixType sum;
-	  zeroit(sum);  
-	  for(int s=0;s<Ls;s++){
-	    int sF = s+Ls*sU;
-  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
-  	      auto bb = coalescedRead(Btilde_v[sF]()(spn) ); //color vector
-  	      auto aa = coalescedRead(Atilde_v[sF]()(spn) );
-	      auto op = outerProduct(bb,aa);
-  	      sum = sum + op;
-	    }
-	  }
-  	  coalescedWrite(mat_v[sU](mu)(), sum);
-      });
-    }
-#endif    
+      
   }
 };
 
@@ -243,17 +211,17 @@ typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
 
-//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
-//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
-//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
+typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
+typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
 
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
 
-//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
-//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
-//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
+typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
+typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
+typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
  
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float

Grid/qcd/action/fermion/WilsonKernels.h

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

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

@@ -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_CUDA
+#ifdef GRID_NVCC
     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)
 {
-#if (!defined(GRID_HIP))
+#ifndef GRID_NVCC
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -799,7 +799,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
-  PropagatorField zz (UGrid);   zz=0.0;
+  LatticeInteger zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
   for (int s=0;s<Ls;s++) {
 
@@ -826,9 +826,8 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
-#if (!defined(GRID_HIP))
+#ifndef GRID_NVCC
   int tshift = (mu == Nd-1) ? 1 : 0;
-  unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
   ////////////////////////////////////////////////
@@ -851,7 +850,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
 
-  PropagatorField  zz (UGrid);   zz=0.0;
+  LatticeInteger zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
 
   for(int s=0;s<Ls;s++){
@@ -881,29 +880,17 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
-  RealD sign = 1.0; // sign flip for vector/tadpole
   if ( curr_type == Current::Axial ) {
     for(int s=0;s<Ls/2;s++){
       G_s[s] = -1.0;
     }
   }
-  else if ( curr_type == Current::Tadpole ) {
-    auto b=this->_b;
-    auto c=this->_c;
-    if ( b == 1 && c == 0 ) {
-      sign = -1.0;    
-    }
-    else {
-      std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
-      assert(b==1 && c==0);
-    }
-  }
 
   for(int s=0;s<Ls;s++){
 
     int sp = (s+1)%Ls;
-    //    int sr = Ls-1-s;
-    //    int srp= (sr+1)%Ls;
+    int sr = Ls-1-s;
+    int srp= (sr+1)%Ls;
 
     // Mobius parameters
     auto b=this->bs[s];
@@ -920,7 +907,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     tmp    = Cshift(tmp,mu,1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu);
-    tmp    = sign*G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
+    tmp    = G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
     tmp    = where((lcoor>=tmin),tmp,zz); // Mask the time 
     L_Q    = where((lcoor<=tmax),tmp,zz); // Position of current complicated
 
@@ -935,13 +922,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
     tmp    = Cshift(tmp,mu,-1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
     tmp = -G_s[s]*( Utmp + gmu*Utmp );
-    // Mask the time
-    if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
-      unsigned int t0 = 0;
-      tmp    = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
-    } else {
-      tmp    = where((lcoor>=tmin+tshift),tmp,zz);
-    }
+    tmp    = where((lcoor>=tmin+tshift),tmp,zz); // Mask the time 
     L_Q   += where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
 
     InsertSlice(L_Q, q_out, s , 0);

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();
-  autoView(psi , psi_i,AcceleratorRead);
-  autoView(phi , phi_i,AcceleratorRead);
-  autoView(chi , chi_i,AcceleratorWrite);
+  auto psi = psi_i.View();
+  auto phi = phi_i.View();
+  auto chi = chi_i.View();
   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();
-  autoView(psi , psi_i,AcceleratorRead);
-  autoView(phi , phi_i,AcceleratorRead);
-  autoView(chi , chi_i,AcceleratorWrite);
+  auto psi = psi_i.View();
+  auto phi = phi_i.View();
+  auto chi = chi_i.View();
   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();
 
-  autoView(psi , psi_i,AcceleratorRead);
-  autoView(chi , chi_i,AcceleratorWrite);
+  auto psi = psi_i.View();
+  auto chi = chi_i.View();
 
   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;
 
-  autoView(psi , psi_i,AcceleratorRead);
-  autoView(chi , chi_i,AcceleratorWrite);
+  auto psi = psi_i.View();
+  auto chi = chi_i.View();
 
   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();
-  autoView(psi, psi_i,CpuRead);
-  autoView(phi, phi_i,CpuRead);
-  autoView(chi, chi_i,CpuWrite);
+  auto psi = psi_i.View();
+  auto phi = phi_i.View();
+  auto chi = chi_i.View();
   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();
-  autoView(psi,psi_i,CpuRead);
-  autoView(phi,phi_i,CpuRead);
-  autoView(chi,chi_i,CpuWrite);
+  auto psi=psi_i.View();
+  auto phi=phi_i.View();
+  auto chi=chi_i.View();
   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;
-  autoView(psi , psi_i,CpuRead);
-  autoView(chi , chi_i,CpuWrite);
+  auto psi = psi_i.View();
+  auto chi = chi_i.View();
 #ifndef AVX512
   {
     SiteHalfSpinor BcastP;
@@ -535,8 +535,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
   EnableIf<Impl::LsVectorised,int> sfinae=0;
 #ifndef AVX512
   {
-    autoView(psi , psi_i,CpuRead);
-    autoView(chi , chi_i,CpuWrite);
+    auto psi = psi_i.View();
+    auto chi = chi_i.View();
 
     SiteHalfSpinor BcastP;
     SiteHalfSpinor BcastM;
@@ -586,8 +586,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
   }
 #else
   {
-    autoView(psi , psi_i,CpuRead);
-    autoView(chi , chi_i,CpuWrite);
+    auto psi = psi_i.View();
+    auto chi = chi_i.View();
     // pointers
     //  MASK_REGS;
 #define Chi_00 %zmm0