expose gauge group in GImpl and generic Nc fix

Hdf5 namespace

.gitignore

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

.travis.yml

@@ -1,61 +0,0 @@
-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

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

Grid/GridCore.h

@@ -47,9 +47,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
-#include <Grid/allocator/AlignedAllocator.h>
+#include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
-#include <Grid/threads/Threads.h>
+#include <Grid/threads/ThreadReduction.h>
 #include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 

Grid/GridStd.h

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

Grid/Grid_Eigen_Dense.h

@@ -18,21 +18,28 @@
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")
+#undef __CUDA_ARCH__
 #undef __NVCC__
 #undef __CUDACC__
-#undef __CUDA_ARCH__
 #define __NVCC__REDEFINE__
 #endif 
 
 /* SYCL save and restore compile environment*/
-#ifdef __SYCL_DEVICE_ONLY__  
+#ifdef GRID_SYCL
 #pragma push
 #pragma push_macro("__SYCL_DEVICE_ONLY__")
 #undef __SYCL_DEVICE_ONLY__
-#undef EIGEN_USE_SYCL
 #define EIGEN_DONT_VECTORIZE
+//#undef EIGEN_USE_SYCL
+#define __SYCL__REDEFINE__
 #endif
 
+/* 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>
@@ -51,6 +58,12 @@
 #pragma pop
 #endif
 
+/*HIP restore*/
+#ifdef __HIP__REDEFINE__
+#pragma pop_macro("__HIP_DEVICE_COMPILE__")
+#pragma pop
+#endif
+
 #if defined __GNUC__
 #pragma GCC diagnostic pop
 #endif

Grid/Makefile.am

@@ -21,7 +21,8 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
-all: version-cache
+
+all: version-cache Version.h
 
 version-cache:
 	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
@@ -42,7 +43,7 @@ version-cache:
 	fi;\
 	rm -f vertmp
 
-Version.h:
+Version.h: version-cache
 	cp version-cache Version.h
 
 .PHONY: version-cache
@@ -53,6 +54,19 @@ Version.h:
 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

@@ -29,9 +29,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_H
 #define GRID_ALGORITHMS_H
 
+NAMESPACE_CHECK(algorithms);
 #include <Grid/algorithms/SparseMatrix.h>
 #include <Grid/algorithms/LinearOperator.h>
 #include <Grid/algorithms/Preconditioner.h>
+NAMESPACE_CHECK(SparseMatrix);
 
 #include <Grid/algorithms/approx/Zolotarev.h>
 #include <Grid/algorithms/approx/Chebyshev.h>
@@ -41,10 +43,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/approx/Forecast.h>
 #include <Grid/algorithms/approx/RemezGeneral.h>
 #include <Grid/algorithms/approx/ZMobius.h>
-
+NAMESPACE_CHECK(approx);
 #include <Grid/algorithms/iterative/Deflation.h>
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
+NAMESPACE_CHECK(ConjGrad);
 #include <Grid/algorithms/iterative/BiCGSTAB.h>
+NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
@@ -62,7 +66,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/iterative/PowerMethod.h>
 
+NAMESPACE_CHECK(PowerMethod);
 #include <Grid/algorithms/CoarsenedMatrix.h>
+NAMESPACE_CHECK(CoarsendMatrix);
 #include <Grid/algorithms/FFT.h>
 
 #endif

Grid/algorithms/FFT.h

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

Grid/algorithms/LinearOperator.h

@@ -530,6 +530,16 @@ 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> {

Grid/algorithms/iterative/BiCGSTAB.h

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

Grid/algorithms/iterative/ConjugateGradient.h

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

Grid/algorithms/iterative/Deflation.h

@@ -54,15 +54,23 @@ class DeflatedGuesser: public LinearFunction<Field> {
 private:
   const std::vector<Field> &evec;
   const std::vector<RealD> &eval;
+  const unsigned int       N;
 
 public:
 
-  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
+  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());
+  } 
 
   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);

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -0,0 +1,241 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/PrecGeneralisedConjugateResidual.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_PREC_GCR_NON_HERM_H
+#define GRID_PREC_GCR_NON_HERM_H
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+//VPGCR Abe and Zhang, 2005.
+//INTERNATIONAL JOURNAL OF NUMERICAL ANALYSIS AND MODELING
+//Computing and Information Volume 2, Number 2, Pages 147-161
+//NB. Likely not original reference since they are focussing on a preconditioner variant.
+//    but VPGCR was nicely written up in their paper
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+
+#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" " 
+
+template<class Field>
+class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
+public:                                                
+
+  RealD   Tolerance;
+  Integer MaxIterations;
+  int verbose;
+  int mmax;
+  int nstep;
+  int steps;
+  int level;
+  GridStopWatch PrecTimer;
+  GridStopWatch MatTimer;
+  GridStopWatch LinalgTimer;
+
+  LinearFunction<Field>     &Preconditioner;
+  LinearOperatorBase<Field> &Linop;
+
+  void Level(int lv) { level=lv; };
+
+  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
+    Tolerance(tol), 
+    MaxIterations(maxit),
+    Linop(_Linop),
+    Preconditioner(Prec),
+    mmax(_mmax),
+    nstep(_nstep)
+  { 
+    level=1;
+    verbose=1;
+  };
+
+  void operator() (const Field &src, Field &psi){
+
+    psi=Zero();
+    RealD cp, ssq,rsq;
+    ssq=norm2(src);
+    rsq=Tolerance*Tolerance*ssq;
+      
+    Field r(src.Grid());
+
+    PrecTimer.Reset();
+    MatTimer.Reset();
+    LinalgTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    steps=0;
+    for(int k=0;k<MaxIterations;k++){
+
+      cp=GCRnStep(src,psi,rsq);
+
+      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
+
+      if(cp<rsq) {
+
+	SolverTimer.Stop();
+
+	Linop.Op(psi,r);
+	axpy(r,-1.0,src,r);
+	RealD tr = norm2(r);
+	GCRLogLevel<<"PGCR: Converged on iteration " <<steps
+		 << " computed residual "<<sqrt(cp/ssq)
+		 << " true residual "    <<sqrt(tr/ssq)
+		 << " target "           <<Tolerance <<std::endl;
+
+	GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
+	return;
+      }
+
+    }
+    GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
+    //    assert(0);
+  }
+
+  RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
+
+    RealD cp;
+    ComplexD a, b, zAz;
+    RealD zAAz;
+    ComplexD rq;
+
+    GridBase *grid = src.Grid();
+
+    Field r(grid);
+    Field z(grid);
+    Field tmp(grid);
+    Field ttmp(grid);
+    Field Az(grid);
+
+    ////////////////////////////////
+    // history for flexible orthog
+    ////////////////////////////////
+    std::vector<Field> q(mmax,grid);
+    std::vector<Field> p(mmax,grid);
+    std::vector<RealD> qq(mmax);
+      
+    GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
+
+    //////////////////////////////////
+    // initial guess x0 is taken as nonzero.
+    // r0=src-A x0 = src
+    //////////////////////////////////
+    MatTimer.Start();
+    Linop.Op(psi,Az);
+    zAz = innerProduct(Az,psi);
+    zAAz= norm2(Az);
+    MatTimer.Stop();
+    
+
+    LinalgTimer.Start();
+    r=src-Az;
+    LinalgTimer.Stop();
+    GCRLogLevel<< "PGCR true residual r = src - A psi   "<<norm2(r) <<std::endl;
+    
+    /////////////////////
+    // p = Prec(r)
+    /////////////////////
+
+    PrecTimer.Start();
+    Preconditioner(r,z);
+    PrecTimer.Stop();
+
+    MatTimer.Start();
+    Linop.Op(z,Az);
+    MatTimer.Stop();
+
+    LinalgTimer.Start();
+
+    zAz = innerProduct(Az,psi);
+    zAAz= norm2(Az);
+
+    //p[0],q[0],qq[0] 
+    p[0]= z;
+    q[0]= Az;
+    qq[0]= zAAz;
+    
+    cp =norm2(r);
+    LinalgTimer.Stop();
+
+    for(int k=0;k<nstep;k++){
+
+      steps++;
+
+      int kp     = k+1;
+      int peri_k = k %mmax;
+      int peri_kp= kp%mmax;
+
+      LinalgTimer.Start();
+      rq= innerProduct(q[peri_k],r); // what if rAr not real?
+      a = rq/qq[peri_k];
+
+      axpy(psi,a,p[peri_k],psi);         
+
+      cp = axpy_norm(r,-a,q[peri_k],r);
+      LinalgTimer.Stop();
+
+      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
+
+      if((k==nstep-1)||(cp<rsq)){
+	return cp;
+      }
+
+
+      PrecTimer.Start();
+      Preconditioner(r,z);// solve Az = r
+      PrecTimer.Stop();
+
+      MatTimer.Start();
+      Linop.Op(z,Az);
+      MatTimer.Stop();
+      zAz = innerProduct(Az,psi);
+      zAAz= norm2(Az);
+
+      LinalgTimer.Start();
+
+      q[peri_kp]=Az;
+      p[peri_kp]=z;
+
+      int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
+      for(int back=0;back<northog;back++){
+
+	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+
+	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
+	p[peri_kp]=p[peri_kp]+b*p[peri_back];
+	q[peri_kp]=q[peri_kp]+b*q[peri_back];
+
+      }
+      qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
+      LinalgTimer.Stop();
+    }
+    assert(0); // never reached
+    return cp;
+  }
+};
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/SchurRedBlack.h

@@ -132,6 +132,31 @@ 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) 
     {
@@ -150,24 +175,29 @@ namespace Grid {
       ////////////////////////////////////////////////
       // Prepare RedBlack source
       ////////////////////////////////////////////////
-      for(int b=0;b<nblock;b++){
-	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
-      }
+      RedBlackSource(_Matrix,in,src_o);
+	//      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);
 
-      for(int b=0;b<nblock;b++){
-        if(useSolnAsInitGuess) {
+      
+      if(useSolnAsInitGuess) {
+        for(int b=0;b<nblock;b++){
           pickCheckerboard(Odd, sol_o[b], out[b]);
-        } else {
-          guess(src_o[b],sol_o[b]); 
         }
+      } else {
+        guess(src_o, sol_o); 
+      }
 
-	if ( subGuess ) { 
-	  guess_save[b] = sol_o[b];
-	}
+	    if ( subGuess ) { 
+        for(int b=0;b<nblock;b++){
+          guess_save[b] = sol_o[b];
+        }
       }
       //////////////////////////////////////////////////////////////
       // Call the block solver

Grid/allocator/AlignedAllocator.cc

@@ -1,154 +0,0 @@
-#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

@@ -26,129 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_ALIGNED_ALLOCATOR_H
-#define GRID_ALIGNED_ALLOCATOR_H
-
-#ifdef HAVE_MALLOC_MALLOC_H
-#include <malloc/malloc.h>
-#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_MM_MALLOC_H
-#include <mm_malloc.h>
-#endif
-
-#define POINTER_CACHE
-#define GRID_ALLOC_ALIGN (2*1024*1024)
-#define GRID_ALLOC_SMALL_LIMIT (4096)
+#pragma once
 
 NAMESPACE_BEGIN(Grid);
 
-// Move control to configure.ac and Config.h?
-
-class PointerCache {
-private:
-/*Pinning pages is costly*/
-/*Could maintain separate large and small allocation caches*/
-/* Could make these configurable, perhaps up to a max size*/
-  static const int NcacheSmallMax=128; 
-  static const int NcacheMax=16;
-  static int NcacheSmall;
-  static int Ncache;
-
-  typedef struct { 
-    void *address;
-    size_t bytes;
-    int valid;
-  } PointerCacheEntry;
-    
-  static PointerCacheEntry Entries[NcacheMax];
-  static int Victim;
-  static PointerCacheEntry EntriesSmall[NcacheSmallMax];
-  static int VictimSmall;
-
-public:
-  static void Init(void);
-  static void *Insert(void *ptr,size_t bytes) ;
-  static void *Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) ;
-  static void *Lookup(size_t bytes) ;
-  static void *Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) ;
-};
-
-std::string sizeString(size_t bytes);
-
-struct MemoryStats
-{
-  size_t totalAllocated{0}, maxAllocated{0}, 
-    currentlyAllocated{0}, totalFreed{0};
-};
-    
-class MemoryProfiler
-{
-public:
-  static MemoryStats *stats;
-  static bool        debug;
-};
-
-#ifdef GRID_NVCC
-#define profilerCudaMeminfo \
-  { size_t f, t ; cudaMemGetInfo ( &f,&t); std::cout << GridLogDebug << "[Memory debug] Cuda free "<<f<<"/"<<t << std::endl;}
-#else
-#define profilerCudaMeminfo
-#endif
-
-#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
-#define profilerDebugPrint						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl; \
-      std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
-		<< std::endl;						\
-    }									\
-  profilerCudaMeminfo;
-
-#define profilerAllocate(bytes)						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      s->totalAllocated     += (bytes);					\
-      s->currentlyAllocated += (bytes);					\
-      s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated); \
-    }									\
-  if (MemoryProfiler::debug)						\
-    {									\
-      std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl; \
-      profilerDebugPrint;						\
-    }
-
-#define profilerFree(bytes)						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      s->totalFreed         += (bytes);					\
-      s->currentlyAllocated -= (bytes);					\
-    }									\
-  if (MemoryProfiler::debug)						\
-    {									\
-      std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl; \
-      profilerDebugPrint;						\
-    }
-
-void check_huge_pages(void *Buf,uint64_t BYTES);
-
-////////////////////////////////////////////////////////////////////
-// A lattice of something, but assume the something is SIMDized.
-////////////////////////////////////////////////////////////////////
-
 template<typename _Tp>
 class alignedAllocator {
 public: 
@@ -172,89 +53,131 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
-
-
-#ifdef POINTER_CACHE
-    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
-#else
-    pointer ptr = nullptr;
-#endif
-
-#ifdef GRID_NVCC
-    ////////////////////////////////////
-    // Unified (managed) memory
-    ////////////////////////////////////
-    if ( ptr == (_Tp *) NULL ) {
-      //      printf(" alignedAllocater cache miss %ld bytes ",bytes);      BACKTRACEFP(stdout);
-      auto err = 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
+    _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
-  void deallocate(pointer __p, size_type __n) { 
+  void deallocate(pointer __p, size_type __n) 
+  { 
     size_type bytes = __n * sizeof(_Tp);
-
     profilerFree(bytes);
-
-#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
+    MemoryManager::CpuFree((void *)__p,bytes);
   }
 
-  // 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) { };
+  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
+  void construct(pointer __p, const _Tp& __val) { assert(0);};
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
 template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
+//////////////////////////////////////////////////////////////////////////////////////
+// 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
 ////////////////////////////////////////////////////////////////////////////////
-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> > >;
+#ifdef ACCELERATOR_CSHIFT
+// Cshift on device
+template<class T> using cshiftAllocator = devAllocator<T>;
+#else
+// Cshift on host
+template<class T> using cshiftAllocator = std::allocator<T>;
+#endif
+
+template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
+template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
+template<class T> using commVector = std::vector<T,devAllocator<T> >;
+template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);
 
-#endif
+

Grid/allocator/Allocator.h

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

Grid/allocator/MemoryManager.cc

@@ -0,0 +1,254 @@
+#include <Grid/GridCore.h>
+
+NAMESPACE_BEGIN(Grid);
+
+/*Allocation types, saying which pointer cache should be used*/
+#define Cpu      (0)
+#define CpuSmall (1)
+#define Acc      (2)
+#define AccSmall (3)
+#define Shared   (4)
+#define SharedSmall (5)
+uint64_t total_shared;
+uint64_t total_device;
+uint64_t total_host;;
+void MemoryManager::PrintBytes(void)
+{
+  std::cout << " MemoryManager : "<<total_shared<<" shared      bytes "<<std::endl;
+  std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
+  std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<std::endl;
+}
+
+//////////////////////////////////////////////////////////////////////
+// Data tables for recently freed pooiniter caches
+//////////////////////////////////////////////////////////////////////
+MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
+int MemoryManager::Victim[MemoryManager::NallocType];
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 8, 32, 8, 32, 8, 32 };
+
+//////////////////////////////////////////////////////////////////////
+// Actual allocation and deallocation utils
+//////////////////////////////////////////////////////////////////////
+void *MemoryManager::AcceleratorAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Acc);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocDevice(bytes);
+    total_device+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes)
+{
+  void *__freeme = Insert(ptr,bytes,Acc);
+  if ( __freeme ) {
+    acceleratorFreeDevice(__freeme);
+    total_device-=bytes;
+    //    PrintBytes();
+  }
+}
+void *MemoryManager::SharedAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Shared);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+    total_shared+=bytes;
+    //    std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
+    //    PrintBytes();
+  }
+  return ptr;
+}
+void  MemoryManager::SharedFree    (void *ptr,size_t bytes)
+{
+  void *__freeme = Insert(ptr,bytes,Shared);
+  if ( __freeme ) {
+    acceleratorFreeShared(__freeme);
+    total_shared-=bytes;
+    //    PrintBytes();
+  }
+}
+#ifdef GRID_UVM
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+    total_host+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeShared(__freeme);
+    total_host-=bytes;
+  }
+}
+#else
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocCpu(bytes);
+    total_host+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeCpu(__freeme);
+    total_host-=bytes;
+  }
+}
+#endif
+
+//////////////////////////////////////////
+// call only once
+//////////////////////////////////////////
+void MemoryManager::Init(void)
+{
+
+  char * str;
+  int Nc;
+  int NcS;
+  
+  str= getenv("GRID_ALLOC_NCACHE_LARGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[Cpu]=Nc;
+      Ncache[Acc]=Nc;
+      Ncache[Shared]=Nc;
+    }
+  }
+
+  str= getenv("GRID_ALLOC_NCACHE_SMALL");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuSmall]=Nc;
+      Ncache[AccSmall]=Nc;
+      Ncache[SharedSmall]=Nc;
+    }
+  }
+
+}
+
+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]);  
+#else
+  return ptr;
+#endif
+}
+
+void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) 
+{
+  assert(ncache>0);
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  void * ret = NULL;
+  int v = -1;
+
+  for(int e=0;e<ncache;e++) {
+    if ( entries[e].valid==0 ) {
+      v=e; 
+      break;
+    }
+  }
+
+  if ( v==-1 ) {
+    v=victim;
+    victim = (victim+1)%ncache;
+  }
+
+  if ( entries[v].valid ) {
+    ret = entries[v].address;
+    entries[v].valid = 0;
+    entries[v].address = NULL;
+    entries[v].bytes = 0;
+  }
+
+  entries[v].address=ptr;
+  entries[v].bytes  =bytes;
+  entries[v].valid  =1;
+
+  return ret;
+}
+
+void *MemoryManager::Lookup(size_t bytes,int type)
+{
+#ifdef ALLOCATION_CACHE
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type+small;
+  return Lookup(bytes,Entries[cache],Ncache[cache]);
+#else
+  return NULL;
+#endif
+}
+
+void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) 
+{
+  assert(ncache>0);
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+  for(int e=0;e<ncache;e++){
+    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
+      entries[e].valid = 0;
+      return entries[e].address;
+    }
+  }
+  return NULL;
+}
+
+
+NAMESPACE_END(Grid);
+

Grid/allocator/MemoryManager.h

@@ -0,0 +1,180 @@
+/*************************************************************************************
+
+    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];
+
+  /////////////////////////////////////////////////
+  // Free pool
+  /////////////////////////////////////////////////
+  static void *Insert(void *ptr,size_t bytes,int type) ;
+  static void *Lookup(size_t bytes,int type) ;
+  static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
+  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
+
+  static void 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 int   isOpen   (void* CpuPtr);
+  static void  ViewClose(void* CpuPtr,ViewMode mode);
+  static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+
+};
+
+NAMESPACE_END(Grid);
+
+

Grid/allocator/MemoryManagerCache.cc

@@ -0,0 +1,478 @@
+#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)
+{
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+	      << "\t" << AccCache.cpuLock
+	      << "\t" << AccCache.accLock
+	      << "\t" << AccCache.LRU_valid<<std::endl;
+  }
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+
+};
+int   MemoryManager::isOpen   (void* _CpuPtr) 
+{ 
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    return AccCache.cpuLock+AccCache.accLock;
+  } else { 
+    return 0;
+  }
+}
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/allocator/MemoryManagerShared.cc

@@ -0,0 +1,23 @@
+#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::Print(void){};
+void  MemoryManager::NotifyDeletion(void *ptr){};
+
+NAMESPACE_END(Grid);
+#endif

Grid/allocator/MemoryStats.cc

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

Grid/allocator/MemoryStats.h

@@ -0,0 +1,95 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/MemoryStats.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+
+NAMESPACE_BEGIN(Grid);
+
+std::string sizeString(size_t bytes);
+
+struct MemoryStats
+{
+  size_t totalAllocated{0}, maxAllocated{0}, 
+    currentlyAllocated{0}, totalFreed{0};
+};
+    
+class MemoryProfiler
+{
+public:
+  static MemoryStats *stats;
+  static bool        debug;
+};
+
+#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
+#define profilerDebugPrint						\
+  if (MemoryProfiler::stats)						\
+    {									\
+      auto s = MemoryProfiler::stats;					\
+      std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl; \
+      std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
+		<< std::endl;						\
+      std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
+		<< std::endl;						\
+      std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
+		<< std::endl;						\
+      std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
+		<< std::endl;						\
+    }
+
+#define profilerAllocate(bytes)						\
+  if (MemoryProfiler::stats)						\
+    {									\
+      auto s = MemoryProfiler::stats;					\
+      s->totalAllocated     += (bytes);					\
+      s->currentlyAllocated += (bytes);					\
+      s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated); \
+    }									\
+  if (MemoryProfiler::debug)						\
+    {									\
+      std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl; \
+      profilerDebugPrint;						\
+    }
+
+#define profilerFree(bytes)						\
+  if (MemoryProfiler::stats)						\
+    {									\
+      auto s = MemoryProfiler::stats;					\
+      s->totalFreed         += (bytes);					\
+      s->currentlyAllocated -= (bytes);					\
+    }									\
+  if (MemoryProfiler::debug)						\
+    {									\
+      std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl; \
+      profilerDebugPrint;						\
+    }
+
+void check_huge_pages(void *Buf,uint64_t BYTES);
+
+NAMESPACE_END(Grid);
+

Grid/cartesian/Cartesian_base.h

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

Grid/cartesian/Cartesian_full.h

@@ -38,6 +38,7 @@ class GridCartesian: public GridBase {
 
 public:
   int dummy;
+  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -104,6 +105,7 @@ public:
     _ldimensions.resize(_ndimension);
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
+    _checker_dim_mask.resize(_ndimension);;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -114,6 +116,8 @@ public:
 
     for (int d = 0; d < _ndimension; d++)
       {
+	_checker_dim_mask[d]=0;
+
         _fdimensions[d] = dimensions[d];   // Global dimensions
         _gdimensions[d] = _fdimensions[d]; // Global dimensions
         _simd_layout[d] = simd_layout[d];

Grid/cartesian/Cartesian_red_black.h

@@ -35,12 +35,28 @@ static const int CbRed  =0;
 static const int CbBlack=1;
 static const int Even   =CbRed;
 static const int Odd    =CbBlack;
+
+accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex,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

@@ -33,6 +33,8 @@ 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

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -36,6 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+extern bool Stencil_force_mpi ;
+
 class CartesianCommunicator : public SharedMemory {
 
 public:    
@@ -108,6 +109,8 @@ public:
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
+  void GlobalMax(RealD &);
+  void GlobalMax(RealF &);
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
   void GlobalSum(RealD &);
@@ -138,21 +141,6 @@ 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;
@@ -43,8 +43,16 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
 
+#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);
@@ -91,7 +99,7 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Initialises from communicator_world
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   MPI_Comm optimal_comm;
   ////////////////////////////////////////////////////
@@ -110,7 +118,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)    
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
 {
   _ndimension = processors.size();  assert(_ndimension>=1);
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
@@ -127,7 +135,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  //  int Nparent = parent._processors ; 
+  //  int Nparent = parent._processors ;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
 
@@ -149,13 +157,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
-  int crank;  
+  int crank;
   // Mpi uses the reverse Lexico convention to us; so reversed routines called
   Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
   Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
 
   MPI_Comm comm_split;
-  if ( Nchild > 1 ) { 
+  if ( Nchild > 1 ) {
 
     ////////////////////////////////////////////////////////////////
     // Split the communicator
@@ -180,11 +188,11 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   SetCommunicator(comm_split);
 
   ///////////////////////////////////////////////
-  // Free the temp communicator 
+  // Free the temp communicator
   ///////////////////////////////////////////////
   MPI_Comm_free(&comm_split);
 
-  if(0){ 
+  if(0){
     std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
     for(int d=0;d<processors.size();d++){
       std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
@@ -245,7 +253,7 @@ CartesianCommunicator::~CartesianCommunicator()
     for(int i=0;i<communicator_halo.size();i++){
       MPI_Comm_free(&communicator_halo[i]);
     }
-  }  
+  }
 }
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
@@ -267,6 +275,16 @@ void CartesianCommunicator::GlobalXOR(uint64_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalMax(float &f)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSum(float &f){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
@@ -294,60 +312,28 @@ 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);
-  //    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)
-{
+  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+
   int myrank = _processor;
   int ierr;
 
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    MPI_Request xrq;
-    MPI_Request rrq;
+  // Enforce no UVM in comms, device or host OK
+  assert(acceleratorIsCommunicable(xmit));
+  assert(acceleratorIsCommunicable(recv));
 
-    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);
-  }
+  // 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
 }
-
+// Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int dest,
 						     void *recv,
@@ -367,7 +353,7 @@ 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;
@@ -382,36 +368,37 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
+  int tag;
 
-  if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
-  if ( gdest == MPI_UNDEFINED ) {
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
+  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+_processor*32;
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
+  } else {
+    // TODO : make a OMP loop on CPU, call threaded bcopy
+    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+    assert(shm!=NULL);
+    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+    acceleratorCopySynchronise(); // MPI prob slower
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);
   }
 
   return off_node_bytes;
 }
-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)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
   int nreq=list.size();
 
@@ -422,6 +409,13 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
   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);
@@ -436,8 +430,8 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-int CartesianCommunicator::RankWorld(void){ 
-  int r; 
+int CartesianCommunicator::RankWorld(void){
+  int r;
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
@@ -470,7 +464,7 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
   // (Turns up on 32^3 x 64 Gparity too)
   MPI_Datatype object;
-  int iwords; 
+  int iwords;
   int ibytes;
   iwords = words;
   ibytes = bytes;
@@ -483,5 +477,3 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 }
 
 NAMESPACE_END(Grid);
-
-

Grid/communicator/Communicator_none.cc

@@ -67,6 +67,8 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 
 CartesianCommunicator::~CartesianCommunicator(){}
 
+void CartesianCommunicator::GlobalMax(float &){}
+void CartesianCommunicator::GlobalMax(double &){}
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}
@@ -77,15 +79,6 @@ 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,
@@ -96,20 +89,6 @@ 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);
@@ -137,10 +116,6 @@ 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,
@@ -150,13 +125,10 @@ 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,const void *src,size_t bytes);
+  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
   static void SharedMemoryZero(void *dest,size_t bytes);
 
 };

Grid/communicator/SharedMemoryMPI.cc

@@ -7,6 +7,7 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christoph Lehner <christoph@lhnr.de>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -29,8 +30,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
+#endif
+#ifdef GRID_HIP
+#include <hip/hip_runtime_api.h>
+#endif
+#ifdef GRID_SYCl
+
 #endif
 
 NAMESPACE_BEGIN(Grid); 
@@ -47,7 +54,12 @@ 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);
 
@@ -61,6 +73,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   WorldNodes = WorldSize/WorldShmSize;
   assert( (WorldNodes * WorldShmSize) == WorldSize );
 
+
   // FIXME: Check all WorldShmSize are the same ?
 
   /////////////////////////////////////////////////////////////////////
@@ -161,6 +174,23 @@ static inline int divides(int a,int b)
 }
 void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
 {
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
   ////////////////////////////////////////////////////////////////
   // Powers of 2,3,5 only in prime decomposition for now
   ////////////////////////////////////////////////////////////////
@@ -170,17 +200,24 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
   std::vector<int> primes({2,3,5});
 
   int dim = 0;
+  int last_dim = ndimension - 1;
   int AutoShmSize = 1;
   while(AutoShmSize != WorldShmSize) {
-    for(int p=0;p<primes.size();p++) {
+    int p;
+    for(p=0;p<primes.size();p++) {
       int prime=primes[p];
       if ( divides(prime,WorldDims[dim]/ShmDims[dim])
         && divides(prime,WorldShmSize/AutoShmSize)  ) {
 	AutoShmSize*=prime;
 	ShmDims[dim]*=prime;
+	last_dim = dim;
 	break;
       }
     }
+    if (p == primes.size() && last_dim == dim) {
+      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
+      exit(EXIT_FAILURE);
+    }
     dim=(dim+1) %ndimension;
   }
 }
@@ -413,7 +450,47 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_NVCC
+#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)  
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -433,47 +510,73 @@ 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
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  auto err =  cudaMalloc(&ShmCommBuf, bytes);
-  if ( err !=  cudaSuccess) {
-    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
-    exit(EXIT_FAILURE);  
-  }
+  ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< 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
     //////////////////////////////////////////////////
-    cudaIpcMemHandle_t handle;
-    
+    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 ) { 
-      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      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);
       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
     //////////////////////////////////////////////////
@@ -489,23 +592,68 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     ///////////////////////////////////////////////////////////////
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
-    void * thisBuf = ShmCommBuf;
+
+#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
     if ( r!=WorldShmRank ) { 
-      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      auto 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)
@@ -633,7 +781,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -677,16 +824,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 /////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
-#ifdef GRID_NVCC
-  cudaMemset(dest,0,bytes);
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  acceleratorMemSet(dest,0,bytes);
 #else
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
 {
-#ifdef GRID_NVCC
-  cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  acceleratorCopyToDevice(src,dest,bytes);
 #else   
   bcopy(src,dest,bytes);
 #endif
@@ -705,7 +852,11 @@ 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);
@@ -735,25 +886,18 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
 
-#ifdef GRID_IBM_SUMMIT
-  // Hide the shared memory path between sockets 
-  // if even number of nodes
-  if ( (ShmSize & 0x1)==0 ) {
-    int SocketSize = ShmSize/2;
-    int mySocket = ShmRank/SocketSize; 
+#ifdef GRID_SHM_FORCE_MPI
+  // Hide the shared memory path between ranks
+  {
     for(int r=0;r<size;r++){
-      int hisRank=ShmRanks[r];
-      if ( hisRank!= MPI_UNDEFINED ) {
-	int hisSocket=hisRank/SocketSize;
-	if ( hisSocket != mySocket ) {
-	  ShmRanks[r] = MPI_UNDEFINED;
-	}
+      if ( r!=rank ) {
+	ShmRanks[r] = MPI_UNDEFINED;
       }
     }
   }
 #endif
 
-  SharedMemoryTest();
+  //SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier

Grid/communicator/SharedMemoryNone.cc

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

Grid/cshift/Cshift.h

@@ -52,23 +52,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-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)))> 
+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)) 
 {
   return Cshift(closure(expr),dim,shift);
 }

Grid/cshift/Cshift_common.h

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

Grid/cshift/Cshift_mpi.h

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

Grid/cshift/Cshift_table.cc

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

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_ET.h

@@ -42,9 +42,24 @@ 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) {
+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) 
+{
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
@@ -68,6 +83,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftru
   merge(ret, falsevals);
   return ret;
 }
+#endif
 
 /////////////////////////////////////////////////////
 //Specialization of getVectorType for lattices
@@ -81,26 +97,62 @@ struct getVectorType<Lattice<T> >{
 //--  recursive evaluation of expressions; --
 // handle leaves of syntax tree
 ///////////////////////////////////////////////////
-template<class sobj> accelerator_inline 
+template<class sobj,
+  typename std::enable_if<!is_lattice<sobj>::value&&!is_lattice_expr<sobj>::value,sobj>::type * = nullptr> 
+accelerator_inline 
 sobj eval(const uint64_t ss, const sobj &arg)
 {
   return arg;
 }
-
 template <class lobj> accelerator_inline 
-const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg) 
+auto eval(const uint64_t ss, const LatticeView<lobj> &arg) -> decltype(arg(ss))
 {
-  return 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 & eval(const uint64_t ss, const Lattice<lobj> &arg) 
+const lobj & vecEval(const uint64_t ss, const LatticeView<lobj> &arg) 
 {
-  auto view = arg.AcceleratorView(ViewRead);
-  return view[ss];
+  return arg[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)  
@@ -108,23 +160,41 @@ 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)))
 {
-  return 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
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -180,16 +250,12 @@ inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
   cb = lat.Checkerboard();
 }
 template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-inline void CBFromExpression(int &cb, const T1 &notlat)  // non-lattice leaf
-{
-}
-
+inline void CBFromExpression(int &cb, const T1 &notlat) {} // non-lattice leaf
 template <typename Op, typename T1> inline 
 void CBFromExpression(int &cb,const LatticeUnaryExpression<Op, T1> &expr) 
 {
   CBFromExpression(cb, expr.arg1);  // recurse AST
 }
-
 template <typename Op, typename T1, typename T2> inline 
 void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr) 
 {
@@ -204,32 +270,86 @@ 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 {								\
-    static auto accelerator_inline func(const arg a) -> decltype(ret) { return ret; } \
+    template<class _arg> 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));
@@ -241,10 +361,10 @@ 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;					\
@@ -264,10 +384,10 @@ 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;						\
@@ -275,17 +395,17 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
   };
 
 GridTrinOpClass(TrinaryWhere,
-		(predicatedWhere<predicate, 
-		 typename std::remove_reference<left>::type,
-		 typename std::remove_reference<right>::type>(pred, lhs,rhs)));
+		(predicatedWhere<
+		 typename std::remove_reference<_predicate>::type, 
+		 typename std::remove_reference<_left>::type,
+		 typename std::remove_reference<_right>::type>(pred, lhs,rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
 ////////////////////////////////////////////
-
-#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_UNOP(name)   name
+#define GRID_BINOP(name)  name
+#define GRID_TRINOP(name) name
 
 #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> \
@@ -331,22 +451,17 @@ GridTrinOpClass(TrinaryWhere,
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
-GRID_DEF_UNOP(adj, UnaryAdj);
-GRID_DEF_UNOP(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);
@@ -371,29 +486,36 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-  -> Lattice<decltype(expr.op.func(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 >
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<decltype(expr.op.func(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 >
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1),
-				eval(0, expr.arg2),
-				eval(0, expr.arg3)))>  ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+				vecEval(0, expr.arg2),
+			        vecEval(0, expr.arg3)))>::type >  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();
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   conformable(ret,rhs);
   conformable(lhs,rhs);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
@@ -56,13 +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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -89,9 +89,9 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -108,8 +108,8 @@ template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     mult(&tmp,&lhs_v(ss),&rhs);
@@ -121,10 +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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -148,8 +148,8 @@ template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -165,8 +165,8 @@ template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -179,10 +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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -206,8 +206,8 @@ template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto rhs_v = lhs.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -221,11 +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);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto x_v = x.AcceleratorView(ViewRead);
-  auto y_v = y.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( x_v , x, AcceleratorRead);
+  autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-    auto tmp = a*x_v(ss)+y_v(ss);
+    auto tmp = a*coalescedRead(x_v[ss])+coalescedRead(y_v[ss]);
     coalescedWrite(ret_v[ss],tmp);
   });
 }
@@ -234,9 +234,9 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-  auto ret_v = ret.AcceleratorView(ViewWrite);
-  auto x_v = x.AcceleratorView(ViewRead);
-  auto y_v = y.AcceleratorView(ViewRead);
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( x_v , x, AcceleratorRead);
+  autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
     auto tmp = a*x_v(ss)+b*y_v(ss);
     coalescedWrite(ret_v[ss],tmp);

Grid/lattice/Lattice_base.h

@@ -29,6 +29,7 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
+
 #pragma once 
 
 #define STREAMING_STORES
@@ -37,180 +38,6 @@ NAMESPACE_BEGIN(Grid);
 
 extern int GridCshiftPermuteMap[4][16];
 
-///////////////////////////////////////////////////////////////////
-// Base class which can be used by traits to pick up behaviour
-///////////////////////////////////////////////////////////////////
-class LatticeBase {};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Conformable checks; same instance of Grid required
-/////////////////////////////////////////////////////////////////////////////////////////
-void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
-{
-  assert(lhs == rhs);
-}
-
-////////////////////////////////////////////////////////////////////////////
-// Advise the LatticeAccelerator class
-////////////////////////////////////////////////////////////////////////////
-enum LatticeAcceleratorAdvise {
-  AdviseInfrequentUse = 0x1,    // Advise that the data is used infrequently.  This can
-                                // significantly influence performance of bulk storage.
-  AdviseReadMostly = 0x2,       // Data will mostly be read.  On some architectures
-                                // enables read-only copies of memory to be kept on
-                                // host and device.
-};
-
-////////////////////////////////////////////////////////////////////////////
-// View Access Mode
-////////////////////////////////////////////////////////////////////////////
-enum ViewMode {
-  ViewRead = 0x1,
-  ViewWrite = 0x2,
-  ViewReadWrite = 0x3
-};
-
-////////////////////////////////////////////////////////////////////////////
-// Minimal base class containing only data valid to access from accelerator
-// _odata will be a managed pointer in CUDA
-////////////////////////////////////////////////////////////////////////////
-// Force access to lattice through a view object.
-// prevents writing of code that will not offload to GPU, but perhaps annoyingly
-// strict since host could could in principle direct access through the lattice object
-// Need to decide programming model.
-#define LATTICE_VIEW_STRICT
-template<class vobj> class LatticeAccelerator : public LatticeBase
-{
-protected:
-  GridBase *_grid;
-  int checkerboard;
-  vobj     *_odata;    // A managed pointer
-  uint64_t _odata_size;    
-public:
-  accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr) { }; 
-  accelerator_inline uint64_t oSites(void) const { return _odata_size; };
-  accelerator_inline int  Checkerboard(void) const { return checkerboard; };
-  accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
-  accelerator_inline void Conformable(GridBase * &grid) const
-  { 
-    if (grid) conformable(grid, _grid);
-    else      grid = _grid;
-  };
-
-  accelerator_inline void Advise(int advise) {
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    if (advise & AdviseInfrequentUse) {
-      cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetPreferredLocation,cudaCpuDeviceId);
-    }
-    if (advise & AdviseReadMostly) {
-      cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetReadMostly,-1);
-    }
-#endif
-#endif
-  };
-
-  accelerator_inline void AcceleratorPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    int target;
-    cudaGetDevice(&target);
-    cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),target);
-#endif
-#endif
-  };
-
-  accelerator_inline void HostPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
-#ifdef GRID_NVCC
-#ifndef __CUDA_ARCH__ // only on host
-    cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),cudaCpuDeviceId);
-#endif
-#endif
-  };
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// A View class which provides accessor to the data.
-// This will be safe to call from accelerator_for and is trivially copy constructible
-// The copy constructor for this will need to be used by device lambda functions
-/////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> 
-class LatticeView : public LatticeAccelerator<vobj>
-{
-public:
-
-
-  // Rvalue
-#ifdef __CUDA_ARCH__
-  accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { return coalescedRead(this->_odata[i]); }
-#else 
-  accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
-#endif
-
-  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
-
-  accelerator_inline uint64_t begin(void) const { return 0;};
-  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
-  accelerator_inline uint64_t size(void)  const { return this->_odata_size; };
-
-  LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me)
-  {
-  }
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Lattice expression types used by ET to assemble the AST
-// 
-// Need to be able to detect code paths according to the whether a lattice object or not
-// so introduce some trait type things
-/////////////////////////////////////////////////////////////////////////////////////////
-
-class LatticeExpressionBase {};
-
-template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
-template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
-
-template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
-template<class T>                 struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
-template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
-
-template <typename Op, typename _T1>                           
-class LatticeUnaryExpression : public  LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  Op op;
-  T1 arg1;
-  LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
-};
-
-template <typename Op, typename _T1, typename _T2>              
-class LatticeBinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
-};
-
-template <typename Op, typename _T1, typename _T2, typename _T3> 
-class LatticeTrinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  typedef typename ViewMap<_T3>::Type T3;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  T3 arg3;
-  LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
-};
-
 /////////////////////////////////////////////////////////////////////////////////////////
 // The real lattice class, with normal copy and assignment semantics.
 // This contains extra (host resident) grid pointer data that may be accessed by host code
@@ -246,38 +73,33 @@ 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();
+  }
   /////////////////////////////////////////////////////////////////////////////////
   // Return a view object that may be dereferenced in site loops.
   // The view is trivially copy constructible and may be copied to an accelerator device
   // in device lambdas
   /////////////////////////////////////////////////////////////////////////////////
-  LatticeView<vobj> View (void) const // deprecated, should pick AcceleratorView for accelerator_for
-  {                                   //                     and HostView        for thread_for
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
+
+  LatticeView<vobj> View (ViewMode mode) const 
+  {
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
     return accessor;
   }
 
-  LatticeView<vobj> AcceleratorView(int mode = ViewReadWrite) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
-    accessor.AcceleratorPrefetch(mode);
-    return accessor;
-  }
-
-  LatticeView<vobj> HostView(int mode = ViewReadWrite) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
-    accessor.HostPrefetch(mode);
-    return accessor;
-  }
-  
   ~Lattice() { 
     if ( this->_odata_size ) {
       dealloc();
@@ -297,12 +119,16 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-
-    auto me  = AcceleratorView(ViewWrite);
-    accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
-      vstream(me[ss],tmp);
+    
+    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);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
@@ -317,11 +143,15 @@ public:
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
-    auto me  = AcceleratorView(ViewWrite);
-    accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
-      vstream(me[ss],tmp);
+    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);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
@@ -335,11 +165,15 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-    auto me  = AcceleratorView(ViewWrite);
-    accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
-      vstream(me[ss],tmp);
+    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);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   //GridFromExpression is tricky to do
@@ -390,10 +224,11 @@ public:
   }
 
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
-    auto me  = View();
+    auto me  = View(CpuWrite);
     thread_for(ss,me.size(),{
-      me[ss] = r;
+	me[ss]= r;
     });
+    me.ViewClose();
     return *this;
   }
 
@@ -403,11 +238,12 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid) { 
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
     this->checkerboard=0;
+    SetViewMode(mode);
   }
   
   //  virtual ~Lattice(void) = default;
@@ -445,11 +281,12 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   AcceleratorView(ViewWrite);
-    auto him= r.AcceleratorView(ViewRead);
+    auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
+    me.ViewClose();    him.ViewClose();
     return *this;
   }
 
@@ -459,11 +296,12 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   AcceleratorView(ViewWrite);
-    auto him= r.AcceleratorView(ViewRead);
+    auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
+    me.ViewClose();    him.ViewClose();
     return *this;
   }
   ///////////////////////////////////////////

Grid/lattice/Lattice_basis.h

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

Grid/lattice/Lattice_comparison.h

@@ -42,34 +42,6 @@ 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
 //////////////////////////////////////////////////////////////////////////
@@ -78,9 +50,9 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
 {
   Lattice<vPredicate> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
       ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
   });
@@ -93,8 +65,8 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
 {
   Lattice<vPredicate> ret(lhs.Grid());
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, lhs_v.size(), {
     ret_v[ss]=op(lhs_v[ss],rhs);
   });
@@ -107,8 +79,8 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
 {
   Lattice<vPredicate> ret(rhs.Grid());
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( rhs_v, rhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
     ret_v[ss]=op(lhs,rhs_v[ss]);
   });

Grid/lattice/Lattice_coordinate.h

@@ -37,7 +37,7 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
   GridBase *grid = l.Grid();
   int Nsimd = grid->iSites();
 
-  auto l_v = l.View();
+  autoView(l_v, l, CpuWrite);
   thread_for( o, grid->oSites(), {
     vector_type vI;
     Coordinate gcoor;
@@ -51,23 +51,5 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
   });
 };
 
-// LatticeCoordinate();
-// FIXME for debug; deprecate this; made obscelete by 
-template<class vobj> void lex_sites(Lattice<vobj> &l){
-  auto l_v = l.View();
-  Real *v_ptr = (Real *)&l_v[0];
-  size_t o_len = l.Grid()->oSites();
-  size_t v_len = sizeof(vobj)/sizeof(vRealF);
-  size_t vec_len = vRealF::Nsimd();
-
-  for(int i=0;i<o_len;i++){
-    for(int j=0;j<v_len;j++){
-      for(int vv=0;vv<vec_len;vv+=2){
-	v_ptr[i*v_len*vec_len+j*vec_len+vv  ]= i+vv*500;
-	v_ptr[i*v_len*vec_len+j*vec_len+vv+1]= i+vv*500;
-      }
-    }}
-}
-
 NAMESPACE_END(Grid);
 

Grid/lattice/Lattice_local.h

@@ -43,8 +43,8 @@ template<class vobj>
 inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
 {
   Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
   });
@@ -56,9 +56,9 @@ template<class vobj>
 inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
 {
   Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
   });
@@ -73,9 +73,9 @@ inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Latt
   typedef decltype(coalescedRead(ll())) sll;
   typedef decltype(coalescedRead(rr())) srr;
   Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),1,{
     // FIXME had issues with scalar version of outer 
     // Use vector [] operator and don't read coalesce this loop

Grid/lattice/Lattice_matrix_reduction.h

@@ -51,9 +51,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int block =FullGrid->_slice_block [Orthog];
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
-  auto X_v = X.View();
-  auto Y_v = Y.View();
-  auto R_v = R.View();
+  autoView( X_v , X, CpuRead);
+  autoView( Y_v , Y, CpuRead);
+  autoView( R_v , R, CpuWrite);
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -97,8 +97,8 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  auto X_v = X.View();
-  auto R_v = R.View();
+  autoView( X_v , X, CpuRead);
+  autoView( R_v , R, CpuWrite);
 
   thread_region
   {
@@ -156,8 +156,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   int ostride=FullGrid->_ostride[Orthog];
 
   typedef typename vobj::vector_typeD vector_typeD;
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( lhs_v , lhs, CpuRead);
+  autoView( rhs_v , rhs, CpuRead);
   thread_region {
     std::vector<vobj> Left(Nblock);
     std::vector<vobj> Right(Nblock);

Grid/lattice/Lattice_peekpoke.h

@@ -46,9 +46,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Ind
 {
   Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
   ret.Checkerboard()=lhs.Checkerboard();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
+  accelerator_for( ss, lhs_v.size(), 1, {
     ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
   });
   return ret;
@@ -58,9 +58,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
 {
   Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
   ret.Checkerboard()=lhs.Checkerboard();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
+  accelerator_for( ss, lhs_v.size(), 1, {
     ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
   });
   return ret;
@@ -72,18 +72,18 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
 template<int Index,class vobj>  
 void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
 {
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( rhs_v, rhs, AcceleratorRead);
+  autoView( lhs_v, lhs, AcceleratorWrite);
+  accelerator_for( ss, lhs_v.size(), 1, {
     pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
   });
 }
 template<int Index,class vobj> 
 void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
 {
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( rhs_v, rhs, AcceleratorRead);
+  autoView( lhs_v, lhs, AcceleratorWrite);
+  accelerator_for( ss, lhs_v.size(), 1, {
     pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
   });
 }
@@ -111,7 +111,7 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   // extract-modify-merge cycle is easiest way and this is not perf critical
   ExtractBuffer<sobj> buf(Nsimd);
-  auto l_v = l.View();
+  autoView( l_v , l, CpuWrite);
   if ( rank == grid->ThisRank() ) {
     extract(l_v[odx],buf);
     buf[idx] = s;
@@ -141,7 +141,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
 
   ExtractBuffer<sobj> buf(Nsimd);
-  auto l_v = l.View();
+  autoView( l_v , l, CpuWrite);
   extract(l_v[odx],buf);
 
   s = buf[idx];
@@ -151,21 +151,21 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
   return;
 };
 
-
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+// Must be CPU read view
 template<class vobj,class sobj>
-inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
-        
-  GridBase *grid = l.Grid();
-
+inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
+{
+  GridBase *grid = l.getGrid();
+  assert(l.mode==CpuRead);
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
+  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -173,8 +173,7 @@ inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  auto l_v = l.View();
-  scalar_type * vp = (scalar_type *)&l_v[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
@@ -183,18 +182,27 @@ inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
       
   return;
 };
-
 template<class vobj,class sobj>
-inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
+inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
+{
+  autoView(lv,l,CpuRead);
+  peekLocalSite(s,lv,site);
+  return;
+};
 
-  GridBase *grid=l.Grid();
+// 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);
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
+  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -202,13 +210,19 @@ inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  auto l_v = l.View();
-  scalar_type * vp = (scalar_type *)&l_v[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
   for(int w=0;w<words;w++){
     vp[idx+w*Nsimd] = pt[w];
   }
+  return;
+};
 
+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

@@ -0,0 +1,79 @@
+/*************************************************************************************
+
+    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

@@ -40,26 +40,77 @@ NAMESPACE_BEGIN(Grid);
 
 template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
+
+  autoView( lhs_v, lhs, AcceleratorRead);
+  autoView( ret_v, ret, AcceleratorWrite);
+
   ret.Checkerboard()=lhs.Checkerboard();
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
-  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
-    coalescedWrite(ret_v[ss], adj(lhs_v(ss)));
+  accelerator_for( ss, lhs_v.size(), 1, {
+     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>
 
 
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
 
@@ -39,7 +39,36 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
 {
   typedef typename vobj::scalar_object  sobj;
 
-  const int Nsimd = vobj::Nsimd();
+  //  const int Nsimd = vobj::Nsimd();
+  const int nthread = GridThread::GetThreads();
+
+  Vector<sobj> sumarray(nthread);
+  for(int i=0;i<nthread;i++){
+    sumarray[i]=Zero();
+  }
+  
+  thread_for(thr,nthread, {
+    int nwork, mywork, myoff;
+    nwork = osites;
+    GridThread::GetWork(nwork,thr,mywork,myoff);
+    vobj vvsum=Zero();
+    for(int ss=myoff;ss<mywork+myoff; ss++){
+      vvsum = vvsum + arg[ss];
+    }
+    sumarray[thr]=Reduce(vvsum);
+  });
+  
+  sobj ssum=Zero();  // sum across threads
+  for(int i=0;i<nthread;i++){
+    ssum = ssum+sumarray[i];
+  } 
+  return ssum;
+}
+template<class vobj>
+inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
+{
+  typedef typename vobj::scalar_objectD  sobj;
+
   const int nthread = GridThread::GetThreads();
 
   Vector<sobj> sumarray(nthread);
@@ -63,23 +92,69 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
     ssum = ssum+sumarray[i];
   } 
   
-  return ssum;
+  typedef typename vobj::scalar_object ssobj;
+  ssobj ret = ssum;
+  return ret;
 }
+/*
+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)
 {
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
 #endif  
 }
+template<class vobj>
+inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  return sumD_gpu(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
+
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-  auto arg_v = arg.View();
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum(&arg_v[0],osites);
+  auto ssum= sum_gpu(&arg_v[0],osites);
+#else
+  autoView(arg_v, arg, CpuRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif  
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -92,6 +167,32 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
   return real(nrm); 
 }
 
+//The global maximum of the site norm2
+template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
+{
+  typedef typename vobj::tensor_reduced vscalar;  //iScalar<iScalar<.... <vPODtype> > >
+  typedef typename vscalar::scalar_object  scalar;   //iScalar<iScalar<.... <PODtype> > >
+
+  Lattice<vscalar> inner = localNorm2(arg);
+
+  auto grid = arg.Grid();
+
+  RealD max;
+  for(int l=0;l<grid->lSites();l++){
+    Coordinate coor;
+    scalar val;
+    RealD r;
+    grid->LocalIndexToLocalCoor(l,coor);
+    peekLocalSite(val,inner,coor);
+    r=real(TensorRemove(val));
+    if( (l==0) || (r>max)){
+      max=r;
+    }
+  }
+  grid->GlobalMax(max);
+  return max;
+}
+
 // Double inner product
 template<class vobj>
 inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
@@ -101,43 +202,30 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   ComplexD  nrm;
   
   GridBase *grid = left.Grid();
-  
-  // Might make all code paths go this way.
-  auto left_v = left.AcceleratorView(ViewRead);
-  auto right_v=right.AcceleratorView(ViewRead);
 
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-#ifdef GRID_NVCC
-  // GPU - SIMT lane compliance...
-  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
+  // Might make all code paths go this way.
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  
+    
+  {
+    autoView( left_v , left, AcceleratorRead);
+    autoView( right_v,right, AcceleratorRead);
 
-  accelerator_for( ss, sites, nsimd,{
-      auto x_l = left_v(ss);
-      auto y_l = right_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
-  })
+    // GPU - SIMT lane compliance...
+    accelerator_for( ss, sites, 1,{
+	auto x_l = left_v[ss];
+	auto y_l = right_v[ss];
+	inner_tmp_v[ss]=innerProductD(x_l,y_l);
+    });
+  }
 
   // This is in single precision and fails some tests
-  // Need a sumD that sums in double
-  nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));  
-#else
-  // CPU 
-  typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  
-  accelerator_for( ss, sites, nsimd,{
-      auto x_l = left_v[ss];
-      auto y_l = right_v[ss];
-      inner_tmp_v[ss]=innerProductD(x_l,y_l);
-  })
-  nrm = TensorRemove(sum(inner_tmp_v,sites));
-#endif
+  auto anrm = sum(inner_tmp_v,sites);  
+  nrm = anrm;
   return nrm;
 }
 
@@ -175,40 +263,24 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   
   GridBase *grid = x.Grid();
 
-  auto x_v=x.AcceleratorView(ViewRead);
-  auto y_v=y.AcceleratorView(ViewRead);
-  auto z_v=z.AcceleratorView(ViewWrite);
-
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-#ifdef GRID_NVCC
   // GPU
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
 
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-
-  nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
-#else
-  // CPU 
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
+
+  accelerator_for( ss, sites, 1,{
+      auto tmp = a*x_v[ss]+b*y_v[ss];
       inner_tmp_v[ss]=innerProductD(tmp,tmp);
       z_v[ss]=tmp;
   });
-  // Already promoted to double
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -224,47 +296,29 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
 
   GridBase *grid = left.Grid();
 
-  auto left_v=left.AcceleratorView(ViewRead);
-  auto right_v=right.AcceleratorView(ViewRead);
-
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
 
-#ifdef GRID_NVCC
   // GPU
-  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
-  typedef decltype(innerProduct(left_v[0],left_v[0])) norm_t;
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
+  typedef decltype(innerProductD(vobj(),vobj())) norm_t;
   Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t> norm_tmp(sites);
+  Vector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
+  {
+    autoView(left_v,left, AcceleratorRead);
+    autoView(right_v,right,AcceleratorRead);
+    accelerator_for( ss, sites, 1,{
+	auto left_tmp = left_v[ss];
+	inner_tmp_v[ss]=innerProductD(left_tmp,right_v[ss]);
+        norm_tmp_v [ss]=innerProductD(left_tmp,left_tmp);
+      });
+  }
 
-  accelerator_for( ss, sites, nsimd,{
-      auto left_tmp = left_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(left_tmp,right_v(ss)));
-      coalescedWrite(norm_tmp_v[ss],innerProduct(left_tmp,left_tmp));
-  });
-
-  tmp[0] = TensorRemove(sumD_gpu(inner_tmp_v,sites));
-  tmp[1] = TensorRemove(sumD_gpu(norm_tmp_v,sites));
-#else
-  // CPU
-  typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
-  typedef decltype(innerProductD(left_v[0],left_v[0])) norm_t;
-  Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t> norm_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  auto norm_tmp_v = &norm_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto left_tmp = left_v(ss);
-      inner_tmp_v[ss] = innerProductD(left_tmp,right_v(ss));
-      norm_tmp_v[ss] = innerProductD(left_tmp,left_tmp);
-  });
-  // Already promoted to double
   tmp[0] = TensorRemove(sum(inner_tmp_v,sites));
   tmp[1] = TensorRemove(sum(norm_tmp_v,sites));
-#endif
+
   grid->GlobalSumVector(&tmp[0],2); // keep norm Complex -> can use GlobalSumVector
   ip = tmp[0];
   nrm = real(tmp[1]);
@@ -307,6 +361,7 @@ 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);
 
@@ -335,7 +390,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
 
   // sum over reduced dimension planes, breaking out orthog dir
   // Parallel over orthog direction
-  auto Data_v=Data.View();
+  autoView( Data_v, Data, CpuRead);
   thread_for( r,rd, {
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
     for(int n=0;n<e1;n++){
@@ -365,20 +420,19 @@ 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] ) {
-      gsum=lsSum[lt];
+      result[t]=lsSum[lt];
     } else {
-      gsum=Zero();
+      result[t]=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>
@@ -413,8 +467,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
 
-  auto lhv=lhs.View();
-  auto rhv=rhs.View();
+  autoView( lhv, lhs, CpuRead);
+  autoView( rhv, rhs, CpuRead);
   thread_for( r,rd,{
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
@@ -521,14 +575,12 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
 
     tensor_reduced at; at=av;
 
-    auto Rv=R.View();
-    auto Xv=X.View();
-    auto Yv=Y.View();
-    thread_for_collapse(2, n, e1, {
-      for(int b=0;b<e2;b++){
+    autoView( Rv, R, CpuWrite);
+    autoView( Xv, X, CpuRead);
+    autoView( Yv, Y, CpuRead);
+    thread_for2d( n, e1, b,e2, {
 	int ss= so+n*stride+b;
 	Rv[ss] = at*Xv[ss]+Yv[ss];
-      }
     });
   }
 };
@@ -581,9 +633,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  auto X_v=X.View();
-  auto Y_v=Y.View();
-  auto R_v=R.View();
+  autoView( X_v, X, CpuRead);
+  autoView( Y_v, Y, CpuRead);
+  autoView( R_v, R, CpuWrite);
   thread_region
   {
     Vector<vobj> s_x(Nblock);
@@ -628,13 +680,14 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   //  int nl=1;
 
   //FIXME package in a convenient iterator
+  // thread_for2d_in_region
   //Should loop over a plane orthogonal to direction "Orthog"
   int stride=FullGrid->_slice_stride[Orthog];
   int block =FullGrid->_slice_block [Orthog];
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
-  auto R_v = R.View();
-  auto X_v = X.View();
+  autoView( R_v, R, CpuWrite);
+  autoView( X_v, X, CpuRead);
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -692,8 +745,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
 
   typedef typename vobj::vector_typeD vector_typeD;
 
-  auto lhs_v=lhs.View();
-  auto rhs_v=rhs.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
   thread_region
   {
     std::vector<vobj> Left(Nblock);

Grid/lattice/Lattice_reduction_gpu.h

@@ -1,7 +1,14 @@
 NAMESPACE_BEGIN(Grid);
 
-#define WARP_SIZE 32
+#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
+
 __device__ unsigned int retirementCount = 0;
 
 template <class Iterator>
@@ -19,7 +26,12 @@ template <class Iterator>
 void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
   
   int device;
+#ifdef GRID_CUDA
   cudaGetDevice(&device);
+#endif
+#ifdef GRID_HIP
+  hipGetDevice(&device);
+#endif
   
   Iterator warpSize            = gpu_props[device].warpSize;
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
@@ -53,7 +65,7 @@ __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));
-  __syncwarp();
+  acceleratorSynchronise();
   
   const Iterator VEC = WARP_SIZE;
   const Iterator vid = tid & (VEC-1);
@@ -67,9 +79,9 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
       beta += temp;
       memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
     }
-    __syncwarp();
+    acceleratorSynchronise();
   }
-  __syncthreads();
+  acceleratorSynchroniseAll();
   
   if (threadIdx.x == 0) {
     beta  = Zero();
@@ -79,7 +91,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
     }
     memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
   }
-  __syncthreads();
+  acceleratorSynchroniseAll();
 }
 
 
@@ -147,7 +159,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     sobj *smem = (sobj *)shmem_pointer;
     
     // wait until all outstanding memory instructions in this thread are finished
-    __threadfence();
+    acceleratorFence();
     
     if (tid==0) {
       unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
@@ -156,8 +168,8 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     }
     
     // each thread must read the correct value of amLast
-    __syncthreads();
-    
+    acceleratorSynchroniseAll();
+
     if (amLast) {
       // reduce buffer[0], ..., buffer[gridDim.x-1]
       Iterator i = tid;
@@ -199,13 +211,7 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   sobj *buffer_v = &buffer[0];
   
   reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
-  cudaDeviceSynchronize();
-  
-  cudaError err = cudaGetLastError();
-  if ( cudaSuccess != err ) {
-    printf("Cuda error %s\n",cudaGetErrorString( err ));
-    exit(0);
-  }
+  accelerator_barrier();
   auto result = buffer_v[0];
   return result;
 }

Grid/lattice/Lattice_rng.h

@@ -375,7 +375,7 @@ public:
     int osites = _grid->oSites();  // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
     int words  = sizeof(scalar_object) / sizeof(scalar_type);
 
-    auto l_v = l.View();
+    autoView(l_v, l, CpuWrite);
     thread_for( ss, osites, {
       ExtractBuffer<scalar_object> buf(Nsimd);
       for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
@@ -461,8 +461,8 @@ public:
     }
 
     {
-      // Obtain one reseeded generator per thread
-      int Nthread = GridThread::GetThreads();
+      // Obtain one reseeded generator per thread      
+      int Nthread = 32; // Hardwire a good level or parallelism
       std::vector<RngEngine> seeders(Nthread);
       for(int t=0;t<Nthread;t++){
 	seeders[t] = Reseed(master_engine);

Grid/lattice/Lattice_trace.h

@@ -42,8 +42,8 @@ template<class vobj>
 inline auto trace(const Lattice<vobj> &lhs)  -> Lattice<decltype(trace(vobj()))>
 {
   Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView(ret_v , ret, AcceleratorWrite);
+  autoView(lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
   });
@@ -58,8 +58,8 @@ template<int Index,class vobj>
 inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
 {
   Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
   });

Grid/lattice/Lattice_transfer.h

@@ -47,11 +47,12 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // remove and insert a half checkerboard
 ////////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
+template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full)
+{
   half.Checkerboard() = cb;
 
-  auto half_v = half.View();
-  auto full_v = full.View();
+  autoView( half_v, half, CpuWrite);
+  autoView( full_v, full, CpuRead);
   thread_for(ss, full.Grid()->oSites(),{
     int cbos;
     Coordinate coor;
@@ -64,11 +65,11 @@ template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,con
     }
   });
 }
-
-template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
+template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half)
+{
   int cb = half.Checkerboard();
-  auto half_v = half.View();
-  auto full_v = full.View();
+  autoView( half_v , half, CpuRead);
+  autoView( full_v , full, CpuWrite);
   thread_for(ss,full.Grid()->oSites(),{
 
     Coordinate coor;
@@ -96,15 +97,29 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> &
   out = in;
 }
 
-#ifdef __CUDA_ARCH__
+template<typename T>
+accelerator_inline EnableIf<isGridFundamental<T>> convertType(T & out, const T & in) {
+  out = in;
+}
+
+// This would allow for conversions between GridFundamental types, but is not strictly needed as yet
+/*template<typename T1, typename T2>
+accelerator_inline typename std::enable_if<isGridFundamental<T1>::value && isGridFundamental<T2>::value>::type
+// Or to make this very broad, conversions between anything that's not a GridTensor could be allowed
+//accelerator_inline typename std::enable_if<!isGridTensor<T1>::value && !isGridTensor<T2>::value>::type
+convertType(T1 & out, const T2 & in) {
+  out = in;
+}*/
+
+#ifdef GRID_SIMT
 accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
-  ((ComplexF*)&out)[SIMTlane(vComplexF::Nsimd())] = in;
+  ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
 }
 accelerator_inline void convertType(vComplexD & out, const ComplexD & in) {
-  ((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd())] = in;
+  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd())] = in;
 }
 accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
-  ((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd()*2)] = in;
+  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd()*2)] = in;
 }
 #endif
 
@@ -116,18 +131,18 @@ accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
   Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
 }
 
-template<typename T1,typename T2,int N>
-  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 void convertType(iScalar<T1> & out, const iScalar<T2> & in) {
+  convertType(out._internal,in._internal);
+}
 
-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) {
+template<typename T1,typename T2>
+accelerator_inline NotEnableIf<isGridScalar<T1>> convertType(T1 & out, const iScalar<T2> & in) {
   convertType(out,in._internal);
 }
 
 template<typename T1,typename T2>
-accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
+accelerator_inline NotEnableIf<isGridScalar<T2>> convertType(iScalar<T1> & out, const T2 & in) {
   convertType(out._internal,in);
 }
 
@@ -144,19 +159,13 @@ 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) {
-  auto out_v = out.AcceleratorView(ViewWrite);
-  auto in_v  = in.AcceleratorView(ViewRead);
-
+  autoView( out_v , out,AcceleratorWrite);
+  autoView( in_v  , in ,AcceleratorRead);
   accelerator_for(ss,out_v.size(),T1::Nsimd(),{
       convertType(out_v[ss],in_v(ss));
-    });
+  });
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////
@@ -164,19 +173,20 @@ accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
 ////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj>
 inline auto localInnerProductD(const Lattice<vobj> &lhs,const Lattice<vobj> &rhs)
--> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View()[0],rhs.View()[0])))>>
+-> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View(CpuRead)[0],rhs.View(CpuRead)[0])))>>
 {
-  auto lhs_v = lhs.AcceleratorView(ViewRead);
-  auto rhs_v = rhs.AcceleratorView(ViewRead);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
 
   typedef decltype(TensorRemove(innerProductD2(lhs_v[0],rhs_v[0]))) t_inner;
   Lattice<iScalar<t_inner>> ret(lhs.Grid());
-  auto ret_v = ret.AcceleratorView(ViewWrite);
 
-  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
+  {
+    autoView(ret_v, ret,AcceleratorWrite);
+    accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
       convertType(ret_v[ss],innerProductD2(lhs_v(ss),rhs_v(ss)));
     });
-
+  }
   return ret;
 }
 
@@ -194,14 +204,13 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
   Lattice<iScalar<CComplex>> ip(coarse);
   Lattice<vobj>     fineDataRed = fineData;
 
-  //  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.AcceleratorView(ViewWrite);
-  auto ip_         = ip.AcceleratorView(ViewReadWrite);
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( ip_         , ip,         AcceleratorWrite);
   for(int v=0;v<nbasis;v++) {
     blockInnerProductD(ip,Basis[v],fineDataRed); // ip = <basis|fine>
     accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
 	convertType(coarseData_[sc](v),ip_[sc]);
-      });
+    });
 
     // improve numerical stability of projection
     // |fine> = |fine> - <basis|fine> |basis>
@@ -210,68 +219,6 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
   }
 }
 
-template<class vobj,class CComplex,int nbasis>
-inline void blockProject1(Lattice<iVector<CComplex,nbasis > > &coarseData,
-			 const             Lattice<vobj>   &fineData,
-			 const std::vector<Lattice<vobj> > &Basis)
-{
-  typedef iVector<CComplex,nbasis > coarseSiteData;
-  coarseSiteData elide;
-  typedef decltype(coalescedRead(elide)) ScalarComplex;
-  GridBase * fine  = fineData.Grid();
-  GridBase * coarse= coarseData.Grid();
-  int  _ndimension = coarse->_ndimension;
-
-  // checks
-  assert( nbasis == Basis.size() );
-  subdivides(coarse,fine); 
-  for(int i=0;i<nbasis;i++){
-    conformable(Basis[i],fineData);
-  }
-
-  Coordinate block_r      (_ndimension);
-  
-  for(int d=0 ; d<_ndimension;d++){
-    block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
-    assert(block_r[d]*coarse->_rdimensions[d] == fine->_rdimensions[d]);
-  }
-  int blockVol = fine->oSites()/coarse->oSites();
-
-  coarseData=Zero();
-
-  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
-  ////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // To make this lock free, loop over coars parallel, and then loop over fine associated with coarse.
-  // Otherwise do fine inner product per site, and make the update atomic
-  ////////////////////////////////////////////////////////////////////////////////////////////////////////
-  accelerator_for( sci, nbasis*coarse->oSites(), vobj::Nsimd(), {
-
-    auto sc=sci/nbasis;
-    auto i=sci%nbasis;
-    auto Basis_      = Basis[i].View();
-
-    Coordinate coor_c(_ndimension);
-    Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions);  // Block coordinate
-
-    int sf;
-    decltype(innerProduct(Basis_(sf),fineData_(sf))) reduce=Zero();
-
-    for(int sb=0;sb<blockVol;sb++){
-
-      Coordinate coor_b(_ndimension);
-      Coordinate coor_f(_ndimension);
-
-      Lexicographic::CoorFromIndex(coor_b,sb,block_r);
-      for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d]+coor_b[d];
-      Lexicographic::IndexFromCoor(coor_f,sf,fine->_rdimensions);
-      
-      reduce=reduce+innerProduct(Basis_(sf),fineData_(sf));
-    }
-    coalescedWrite(coarseData_[sc](i),reduce);
-  });
-  return;
-}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -298,10 +245,12 @@ template<class vobj,class vobj2,class CComplex>
     assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
-  auto fineZ_  = fineZ.AcceleratorView(ViewWrite);
-  auto fineX_  = fineX.AcceleratorView(ViewRead);
-  auto fineY_  = fineY.AcceleratorView(ViewRead);
-  auto coarseA_= coarseA.AcceleratorView(ViewRead);
+  autoView( fineZ_  , fineZ, AcceleratorWrite);
+  autoView( fineX_  , fineX, AcceleratorRead);
+  autoView( fineY_  , fineY, AcceleratorRead);
+  autoView( coarseA_, coarseA, AcceleratorRead);
+  Coordinate fine_rdimensions = fine->_rdimensions;
+  Coordinate coarse_rdimensions = coarse->_rdimensions;
 
   accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
 
@@ -309,12 +258,12 @@ 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 __CUDA_ARCH__
+#ifdef GRID_SIMT
       typename vobj2::tensor_reduced::scalar_object cA;
       typename vobj::scalar_object cAx;
 #else
@@ -344,15 +293,16 @@ template<class vobj,class CComplex>
   Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
   Lattice<dotp> coarse_inner(coarse);
 
-  auto CoarseInner_  = CoarseInner.AcceleratorView(ViewWrite);
-  auto coarse_inner_ = coarse_inner.AcceleratorView(ViewReadWrite);
-
   // Precision promotion
-  fine_inner = localInnerProductD(fineX,fineY);
+  fine_inner = localInnerProductD<vobj>(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  accelerator_for(ss, coarse->oSites(), 1, {
+  {
+    autoView( CoarseInner_  , CoarseInner,AcceleratorWrite);
+    autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
+    accelerator_for(ss, coarse->oSites(), 1, {
       convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
     });
+  }
  
 }
 
@@ -370,14 +320,15 @@ inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
   Lattice<dotp> coarse_inner(coarse);
 
   // Precision promotion?
-  auto CoarseInner_  = CoarseInner.AcceleratorView(ViewWrite);
-  auto coarse_inner_ = coarse_inner.AcceleratorView(ViewReadWrite);
-
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  accelerator_for(ss, coarse->oSites(), 1, {
-    CoarseInner_[ss] = coarse_inner_[ss];
-  });
+  {
+    autoView( CoarseInner_  , CoarseInner, AcceleratorWrite);
+    autoView( coarse_inner_ , coarse_inner, AcceleratorRead);
+    accelerator_for(ss, coarse->oSites(), 1, {
+	CoarseInner_[ss] = coarse_inner_[ss];
+    });
+  }
 }
 
 template<class vobj,class CComplex>
@@ -408,16 +359,27 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   }
   int blockVol = fine->oSites()/coarse->oSites();
 
-  auto coarseData_ = coarseData.AcceleratorView(ViewReadWrite);
-  auto fineData_   = fineData.AcceleratorView(ViewRead);
+  // Turn this around to loop threaded over sc and interior loop 
+  // over sf would thread better
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( fineData_   , fineData, AcceleratorRead);
 
+  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
-      coarseData_[sc]=Zero();
+      Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
 
+      vobj cd = zz;
+      
       for(int sb=0;sb<blockVol;sb++){
 
 	int sf;
@@ -425,11 +387,13 @@ 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);
 
-	coarseData_[sc]=coarseData_[sc]+fineData_[sf];
+	cd=cd+fineData_p[sf];
       }
 
+      coarseData_p[sc] = cd;
+
     });
   return;
 }
@@ -510,8 +474,8 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
-  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
+  autoView( fineData_   , fineData, AcceleratorWrite);
+  autoView( coarseData_ , coarseData, AcceleratorRead);
 
   // Loop with a cache friendly loop ordering
   accelerator_for(sf,fine->oSites(),1,{
@@ -524,7 +488,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
     Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
     for(int i=0;i<nbasis;i++) {
-      auto basis_ = Basis[i].View();
+      /*      auto basis_ = Basis[i],  );*/
       if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
       else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
     }
@@ -543,7 +507,14 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   fineData=Zero();
   for(int i=0;i<nbasis;i++) {
     Lattice<iScalar<CComplex> > ip = PeekIndex<0>(coarseData,i);
-    auto  ip_ =  ip.AcceleratorView(ViewRead);
+
+    //Lattice<CComplex> cip(coarse);
+    //autoView( cip_ , cip, AcceleratorWrite);
+    //autoView(  ip_ ,  ip, AcceleratorRead);
+    //accelerator_forNB(sc,coarse->oSites(),CComplex::Nsimd(),{
+    //	coalescedWrite(cip_[sc], ip_(sc)());
+    //  });
+    //blockZAXPY<vobj,CComplex >(fineData,cip,Basis[i],fineData);
     blockZAXPY(fineData,ip,Basis[i],fineData);
   }
 }
@@ -571,15 +542,17 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
     assert(ig->lSites() == og->lSites());
   }
 
+  autoView(in_v,in,CpuRead);
+  autoView(out_v,out,CpuWrite);
   thread_for(idx, ig->lSites(),{
     sobj s;
     ssobj ss;
 
     Coordinate lcoor(ni);
     ig->LocalIndexToLocalCoor(idx,lcoor);
-    peekLocalSite(s,in,lcoor);
+    peekLocalSite(s,in_v,lcoor);
     ss=s;
-    pokeLocalSite(ss,out,lcoor);
+    pokeLocalSite(ss,out_v,lcoor);
   });
 }
 
@@ -614,8 +587,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate rdt = Tg->_rdimensions;
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
-  auto t_v = To.AcceleratorView(ViewWrite);
-  auto f_v = From.AcceleratorView(ViewRead);
+
+  autoView( t_v , To, AcceleratorWrite);
+  autoView( f_v , From, AcceleratorRead);
   accelerator_for(idx,Fg->lSites(),1,{
     sobj s;
     Coordinate Fcoor(nd);
@@ -638,8 +612,6 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       for(int w=0;w<words;w++){
 	tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd];  // FIXME IF RRII layout, type pun no worke
       }
-      //      peekLocalSite(s,From,Fcoor);
-      //      pokeLocalSite(s,To  ,Tcoor);
     }
   });
 }
@@ -670,6 +642,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuRead);
+  autoView(higherDimv,higherDim,CpuWrite);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -682,8 +656,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,lowDim,lcoor);
-    pokeLocalSite(s,higherDim,hcoor);
+    peekLocalSite(s,lowDimv,lcoor);
+    pokeLocalSite(s,higherDimv,hcoor);
   });
 }
 
@@ -711,6 +685,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     }
   }
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuWrite);
+  autoView(higherDimv,higherDim,CpuRead);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -723,8 +699,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,higherDim,hcoor);
-    pokeLocalSite(s,lowDim,lcoor);
+    peekLocalSite(s,higherDimv,hcoor);
+    pokeLocalSite(s,lowDimv,lcoor);
   });
 
 }
@@ -752,6 +728,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuRead);
+  autoView(higherDimv,higherDim,CpuWrite);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -760,8 +738,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,lowDim,lcoor);
-      pokeLocalSite(s,higherDim,hcoor);
+      peekLocalSite(s,lowDimv,lcoor);
+      pokeLocalSite(s,higherDimv,hcoor);
     }
   });
 }
@@ -789,6 +767,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuWrite);
+  autoView(higherDimv,higherDim,CpuRead);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -797,8 +777,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,higherDim,hcoor);
-      pokeLocalSite(s,lowDim,lcoor);
+      peekLocalSite(s,higherDimv,hcoor);
+      pokeLocalSite(s,lowDimv,lcoor);
     }
   });
 }
@@ -862,7 +842,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
   }
 
   //loop over outer index
-  auto in_v  = in.View();
+  autoView( in_v  , in, CpuRead);
   thread_for(in_oidx,in_grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> out_ptrs(in_nsimd);
@@ -955,7 +935,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
     icoor[lane].resize(ndim);
     grid->iCoorFromIindex(icoor[lane],lane);
   }
-  auto out_v = out.View();
+  autoView( out_v , out, CpuWrite);
   thread_for(oidx, grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> ptrs(nsimd);
@@ -1058,7 +1038,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   std::vector<SobjOut> in_slex_conv(in_grid->lSites());
   unvectorizeToLexOrdArray(in_slex_conv, in);
     
-  auto out_v = out.View();
+  autoView( out_v , out, CpuWrite);
   thread_for(out_oidx,out_grid->oSites(),{
     Coordinate out_ocoor(ndim);
     out_grid->oCoorFromOindex(out_ocoor, out_oidx);

Grid/lattice/Lattice_transpose.h

@@ -42,8 +42,8 @@ NAMESPACE_BEGIN(Grid);
 template<class vobj>
 inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
   });
@@ -58,8 +58,8 @@ template<int Index,class vobj>
 inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
 {
   Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
   });

Grid/lattice/Lattice_unary.h

@@ -35,8 +35,8 @@ NAMESPACE_BEGIN(Grid);
 
 template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs, rhs_i, AcceleratorRead);
+  autoView( ret, ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),1,{
       ret[ss]=pow(rhs[ss],y);
@@ -45,8 +45,8 @@ template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
 }
 template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs , rhs_i, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],mod(rhs(ss),y));
@@ -56,8 +56,8 @@ template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
 
 template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto ret = ret_i.View();
-  auto rhs = rhs_i.View();
+  autoView( ret , ret_i, AcceleratorWrite);
+  autoView( rhs , rhs_i, AcceleratorRead);
   ret.Checkerboard() = rhs_i.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],div(rhs(ss),y));
@@ -67,8 +67,8 @@ template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
 
 template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs , rhs_i, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));

Grid/lattice/Lattice_view.h

@@ -0,0 +1,173 @@
+#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;
@@ -52,22 +52,23 @@ inline void whereWolf(Lattice<vobj> &ret,const Lattice<iobj> &predicate,Lattice<
 
   const int Nsimd = grid->Nsimd();
 
-  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];
+  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);
     }
-
-    merge(ret._odata[ss],falsevals);
-  }
+  });
 }
 
 template<class vobj,class iobj>
@@ -76,9 +77,9 @@ inline Lattice<vobj> whereWolf(const Lattice<iobj> &predicate,Lattice<vobj> &ift
   conformable(iftrue,iffalse);
   conformable(iftrue,predicate);
 
-  Lattice<vobj> ret(iftrue._grid);
+  Lattice<vobj> ret(iftrue.Grid());
 
-  where(ret,predicate,iftrue,iffalse);
+  whereWolf(ret,predicate,iftrue,iffalse);
 
   return ret;
 }

Grid/log/Log.h

@@ -130,6 +130,8 @@ 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)
       {
@@ -152,6 +154,8 @@ public:
 	       << now	       << log.background() << " : " ;
       }
       stream << log.colour();
+      stream.flags(f);
+
       return stream;
     } else { 
       return devnull;

Grid/parallelIO/BinaryIO.cc

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

Grid/parallelIO/BinaryIO.h

@@ -79,6 +79,13 @@ 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;
 
   /////////////////////////////////////////////////////////////////////////////
@@ -502,12 +509,15 @@ 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<< bytes <<" bytes in "<<timer.Elapsed() <<" "
-	     << (double)bytes/ (double)timer.useconds() <<" MB/s "<<std::endl;
+    std::cout<< lastPerf.size <<" bytes in "<< timer.Elapsed() <<" "
+	     << lastPerf.mbytesPerSecond <<" MB/s "<<std::endl;
 
     std::cout<<GridLogMessage<<"IOobject: endian and checksum overhead "<<bstimer.Elapsed()  <<std::endl;
 
@@ -663,10 +673,15 @@ class BinaryIO {
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     timer.Start();
-    thread_for(lidx,lsites,{
+    thread_for(lidx,lsites,{  // FIX ME, suboptimal implementation
       std::vector<RngStateType> tmp(RngStateCount);
       std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
-      parallel_rng.SetState(tmp,lidx);
+      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);
       });
     timer.Stop();
 
@@ -723,7 +738,12 @@ class BinaryIO {
     std::vector<RNGstate> iodata(lsites);
     thread_for(lidx,lsites,{
       std::vector<RngStateType> tmp(RngStateCount);
-      parallel_rng.GetState(tmp,lidx);
+      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);
       std::copy(tmp.begin(),tmp.end(),iodata[lidx].begin());
     });
     timer.Stop();

Grid/parallelIO/IldgIO.h

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

Grid/parallelIO/MetaData.h

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

Grid/parallelIO/NerscIO.h

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

Grid/parallelIO/OpenQcdIOChromaReference.h

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

Grid/perfmon/PerfCount.h

@@ -44,7 +44,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <sys/syscall.h>
 #endif
 #ifdef __x86_64__
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 accelerator_inline uint64_t __rdtsc(void) {  return 0; }
 accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
 #else
@@ -112,7 +112,6 @@ class PerformanceCounter {
 private:
 
   typedef struct { 
-  public:
     uint32_t type;
     uint64_t config;
     const char *name;

Grid/pugixml/pugixml.cc

@@ -12773,7 +12773,7 @@ namespace pugi
 #undef PUGI__THROW_ERROR
 #undef PUGI__CHECK_ERROR
 
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #pragma pop
 #endif
 

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

Grid/qcd/action/ActionBase.h

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

Grid/qcd/action/fermion/DomainWallVec5dImpl.h

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

Grid/qcd/action/fermion/Fermion.h

@@ -57,6 +57,7 @@ NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
 
+#include <Grid/qcd/action/fermion/NaiveStaggeredFermion.h>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion.h>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
 NAMESPACE_CHECK(Staggered);
@@ -114,9 +115,9 @@ 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;
@@ -157,41 +158,41 @@ 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;
@@ -234,64 +235,62 @@ 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 typename ViewMap<FermionField>::Type      FermionFieldView;	\
-  typedef typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
+  typedef const typename ViewMap<FermionField>::Type      FermionFieldView;	\
+  typedef const typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView;
 
 #define INHERIT_IMPL_TYPES(Base)		\
   INHERIT_GIMPL_TYPES(Base)			\
@@ -183,7 +183,8 @@ NAMESPACE_CHECK(ImplStaggered);
 /////////////////////////////////////////////////////////////////////////////
 // Single flavour one component spinors with colour index. 5d vec
 /////////////////////////////////////////////////////////////////////////////
-#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
-NAMESPACE_CHECK(ImplStaggered5dVec);  
+// Deprecate Vec5d
+//#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h> 
+//NAMESPACE_CHECK(ImplStaggered5dVec);  
 
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -96,43 +96,31 @@ public:
     int sl        = St._simd_layout[direction];
     Coordinate icoor;
 
-#ifdef __CUDA_ARCH__
-    _Spinor tmp;
-
+#ifdef GRID_SIMT
     const int Nsimd =SiteDoubledGaugeField::Nsimd();
-    int s = SIMTlane(Nsimd);
+    int s = acceleratorSIMTlane(Nsimd);
     St.iCoorFromIindex(icoor,s);
 
     int mmu = mu % Nd;
-    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
-      
-      int permute_lane = (sl==1) 
-    	|| ((distance== 1)&&(icoor[direction]==1))
-	|| ((distance==-1)&&(icoor[direction]==0));
 
-      if ( permute_lane ) { 
-	tmp(0) = chi(1);
-	tmp(1) = chi(0);
-      } else {
-	tmp(0) = chi(0);
-	tmp(1) = chi(1);
-      }
+    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));
 
-      auto UU0=coalescedRead(U(0)(mu));
-      auto UU1=coalescedRead(U(1)(mu));
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
 
-      mult(&phi(0),&UU0,&tmp(0));
-      mult(&phi(1),&UU1,&tmp(1));
+    //Apply the links
+    int f_upper = permute_lane ? 1 : 0;
+    int f_lower = !f_upper;
 
-    } 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));
-
-    }
+    mult(&phi(0),&UU0,&chi(f_upper));
+    mult(&phi(1),&UU1,&chi(f_lower));
 
 #else
     typedef _Spinor vobj;
@@ -232,15 +220,17 @@ public:
       if ( Params.twists[mu] ) { 
 	Uconj = where(coor==neglink,-Uconj,Uconj);
       }
-	  
-      auto U_v = U.View();
-      auto Uds_v = Uds.View();
-      auto Uconj_v = Uconj.View();
-      auto Utmp_v= Utmp.View();
-      thread_foreach(ss,U_v,{
-	Uds_v[ss](0)(mu) = U_v[ss]();
-	Uds_v[ss](1)(mu) = Uconj_v[ss]();
-      });
+
+      {
+	autoView( U_v , U, CpuRead);
+	autoView( Uconj_v , Uconj, CpuRead);
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,U_v,{
+	    Uds_v[ss](0)(mu) = U_v[ss]();
+	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
+	  });
+      }
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
       Uconj = adj(Cshift(Uconj,mu,-1));
@@ -250,19 +240,25 @@ public:
 	Utmp = where(coor==0,Uconj,Utmp);
       }
 
-      thread_foreach(ss,Utmp_v,{
-	Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
-      });
-          
+      {
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,Utmp_v,{
+	    Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
+	  });
+      }
       Utmp = Uconj;
       if ( Params.twists[mu] ) { 
 	Utmp = where(coor==0,U,Utmp);
       }
-	  
-      thread_foreach(ss,Utmp_v,{
-        Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
-      });
-          
+
+      {	  
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,Utmp_v,{
+	    Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
+        });
+      }
     }
   }
       
@@ -272,11 +268,14 @@ public:
     GaugeLinkField link(mat.Grid());
     // use lorentz for flavour as hack.
     auto tmp = TraceIndex<SpinIndex>(outerProduct(Btilde, A));
-    auto link_v = link.View();
-    auto tmp_v = tmp.View();
-    thread_foreach(ss,tmp_v,{
-      link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
-    });
+
+    {
+      autoView( link_v , link, CpuWrite);
+      autoView( tmp_v , tmp, CpuRead);
+      thread_foreach(ss,tmp_v,{
+	  link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
+	});
+    }
     PokeIndex<LorentzIndex>(mat, link, mu);
     return;
   }
@@ -306,16 +305,18 @@ public:
         
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    auto tmp_v = tmp.View();
-    auto Atilde_v = Atilde.View();
-    auto Btilde_v = Btilde.View();
-    thread_for(ss,tmp.Grid()->oSites(),{
-      for (int s = 0; s < Ls; s++) {
-	int sF = s + Ls * ss;
-	auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
-	tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
-      }
-    });
+    {
+      autoView( tmp_v , tmp, CpuWrite);
+      autoView( Atilde_v , Atilde, CpuRead);
+      autoView( Btilde_v , Btilde, CpuRead);
+      thread_for(ss,tmp.Grid()->oSites(),{
+	  for (int s = 0; s < Ls; s++) {
+	    int sF = s + Ls * ss;
+	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+	  }
+	});
+    }
     PokeIndex<LorentzIndex>(mat, tmp, mu);
     return;
   }
@@ -326,8 +327,8 @@ 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 &src4,FermionFieldo &sol5)
+  void operator() (const FermionFieldo &src,FermionFieldo &sol5)
   {
     std::cout << GridLogMessage<< " ************************************************" << std::endl;
     std::cout << GridLogMessage<< "  MADWF-like algorithm                           " << std::endl;
@@ -114,8 +114,16 @@ class MADWF
     ///////////////////////////////////////
     //Import source, include Dminus factors
     ///////////////////////////////////////
-    Mato.ImportPhysicalFermionSource(src4,b); 
-    std::cout << GridLogMessage << " src4 " <<norm2(src4)<<std::endl;
+    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;
     std::cout << GridLogMessage << " b    " <<norm2(b)<<std::endl;
 
     defect = b;

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -0,0 +1,194 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/ImprovedStaggered.h
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#ifndef GRID_QCD_NAIVE_STAG_FERMION_H
+#define GRID_QCD_NAIVE_STAG_FERMION_H
+
+NAMESPACE_BEGIN(Grid);
+
+class NaiveStaggeredFermionStatic {
+public:
+  static const std::vector<int> directions;
+  static const std::vector<int> displacements;
+  static const int npoint = 8;
+};
+
+template <class Impl>
+class NaiveStaggeredFermion : public StaggeredKernels<Impl>, public NaiveStaggeredFermionStatic {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+  typedef StaggeredKernels<Impl> Kernels;
+
+  FermionField _tmp;
+  FermionField &tmp(void) { return _tmp; }
+
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
+  ///////////////////////////////////////////////////////////////
+  // Implement the abstract base
+  ///////////////////////////////////////////////////////////////
+  GridBase *GaugeGrid(void) { return _grid; }
+  GridBase *GaugeRedBlackGrid(void) { return _cbgrid; }
+  GridBase *FermionGrid(void) { return _grid; }
+  GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
+
+  //////////////////////////////////////////////////////////////////
+  // override multiply; cut number routines if pass dagger argument
+  // and also make interface more uniformly consistent
+  //////////////////////////////////////////////////////////////////
+  void M(const FermionField &in, FermionField &out);
+  void Mdag(const FermionField &in, FermionField &out);
+
+  /////////////////////////////////////////////////////////
+  // half checkerboard operations
+  /////////////////////////////////////////////////////////
+  void Meooe(const FermionField &in, FermionField &out);
+  void MeooeDag(const FermionField &in, FermionField &out);
+  void Mooee(const FermionField &in, FermionField &out);
+  void MooeeDag(const FermionField &in, FermionField &out);
+  void MooeeInv(const FermionField &in, FermionField &out);
+  void MooeeInvDag(const FermionField &in, FermionField &out);
+
+  ////////////////////////
+  // Derivative interface
+  ////////////////////////
+  // Interface calls an internal routine
+  void DhopDeriv  (GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+  void DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+  void DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+
+  ///////////////////////////////////////////////////////////////
+  // non-hermitian hopping term; half cb or both
+  ///////////////////////////////////////////////////////////////
+  void Dhop  (const FermionField &in, FermionField &out, int dag);
+  void DhopOE(const FermionField &in, FermionField &out, int dag);
+  void DhopEO(const FermionField &in, FermionField &out, int dag);
+
+  ///////////////////////////////////////////////////////////////
+  // Multigrid assistance; force term uses too
+  ///////////////////////////////////////////////////////////////
+  void Mdir(const FermionField &in, FermionField &out, int dir, int disp);
+  void MdirAll(const FermionField &in, std::vector<FermionField> &out);
+  void DhopDir(const FermionField &in, FermionField &out, int dir, int disp);
+
+  ///////////////////////////////////////////////////////////////
+  // Extra methods added by derived
+  ///////////////////////////////////////////////////////////////
+  void DerivInternal(StencilImpl &st, 
+		     DoubledGaugeField &U,
+		     GaugeField &mat, 
+		     const FermionField &A, const FermionField &B, int dag);
+
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
+  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+			       const FermionField &in, FermionField &out, int dag);
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+				   const FermionField &in, FermionField &out, int dag);
+
+  //////////////////////////////////////////////////////////////////////////
+  // Grid own interface Constructor
+  //////////////////////////////////////////////////////////////////////////
+  NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
+			GridRedBlackCartesian &Hgrid, RealD _mass,
+			RealD _c1, RealD _u0,
+			const ImplParams &p = ImplParams());
+  NaiveStaggeredFermion(GridCartesian &Fgrid,
+			GridRedBlackCartesian &Hgrid, RealD _mass,
+			RealD _c1, RealD _u0,
+			const ImplParams &p = ImplParams());
+
+  // DoubleStore impl dependent
+  void ImportGauge      (const GaugeField &_U );
+  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  void CopyGaugeCheckerboards(void);
+
+  ///////////////////////////////////////////////////////////////
+  // Data members require to support the functionality
+  ///////////////////////////////////////////////////////////////
+
+  //    protected:
+public:
+  // any other parameters of action ???
+  virtual int   isTrivialEE(void) { return 1; };
+  virtual RealD Mass(void) { return mass; }
+  RealD mass;
+  RealD u0;
+  RealD c1;
+
+  GridBase *_grid;
+  GridBase *_cbgrid;
+
+  // Defines the stencils for even and odd
+  StencilImpl Stencil;
+  StencilImpl StencilEven;
+  StencilImpl StencilOdd;
+
+  // Copy of the gauge field , with even and odd subsets
+  DoubledGaugeField Umu;
+  DoubledGaugeField UmuEven;
+  DoubledGaugeField UmuOdd;
+
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+  
+  ///////////////////////////////////////////////////////////////
+  // Conserved current utilities
+  ///////////////////////////////////////////////////////////////
+  void ContractConservedCurrent(PropagatorField &q_in_1,
+                                PropagatorField &q_in_2,
+                                PropagatorField &q_out,
+                                PropagatorField &src,
+                                Current curr_type,
+                                unsigned int mu);
+  void SeqConservedCurrent(PropagatorField &q_in,
+                           PropagatorField &q_out,
+                           PropagatorField &srct,
+                           Current curr_type,
+                           unsigned int mu, 
+                           unsigned int tmin,
+                           unsigned int tmax,
+			   ComplexField &lattice_cmplx);
+};
+
+typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
+typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/qcd/action/fermion/StaggeredImpl.h

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

Grid/qcd/action/fermion/StaggeredKernels.h

@@ -47,23 +47,37 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
   INHERIT_IMPL_TYPES(Impl);
   typedef FermionOperator<Impl> Base;
    
-public:
-    
-   void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
-		      int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+ public:
+
+  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
+		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
+		    const FermionField &in, FermionField &out, int dag, int interior,int exterior);
+  void DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
+		 DoubledGaugeField &U,
+		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
+  
+  void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+		     int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+ protected:    
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Generic Nc kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+   template<int Naik> 
+   static accelerator_inline
+   void DhopSiteGeneric(StencilView &st, 
 			DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteGenericInt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteGenericExt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -71,15 +85,21 @@ public:
    ///////////////////////////////////////////////////////////////////////////////////////
    // Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteHand(StencilView &st, 
 		     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		     SiteSpinor * buf, int LLs, int sU, 
 		     const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteHandInt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteHandExt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -87,27 +107,11 @@ public:
    ///////////////////////////////////////////////////////////////////////////////////////
    // Asm Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+   
+   void DhopSiteAsm(StencilView &st, 
 		    DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		    SiteSpinor * buf, int LLs, int sU, 
 		    const FermionFieldView &in, FermionFieldView &out,int dag);
-   ///////////////////////////////////////////////////////////////////////////////////////////////////
-   // Generic interface; fan out to right routine
-   ///////////////////////////////////////////////////////////////////////////////////////////////////
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		 SiteSpinor * buf, int LLs, int sU,
-		 const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
-
-   void DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, 
-		    DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		    SiteSpinor * buf, int LLs, int sU,
-		    const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
-
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		 SiteSpinor * buf, int LLs, int sU,
-		 const FermionFieldView &in, FermionFieldView &out, int dag, int interior,int exterior);
   
 public:
 

Grid/qcd/action/fermion/StaggeredVec5dImpl.h

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

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -245,7 +245,7 @@ public:
     return out;
   }
 
-private:
+protected:
   // here fixing the 4 dimensions, make it more general?
 
   RealD csw_r;                                               // Clover coefficient - spatial
@@ -257,15 +257,16 @@ private:
   CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term Dag EO
   CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
 
+ public:
   // eventually these can be compressed into 6x6 blocks instead of the 12x12
   // using the DeGrand-Rossi basis for the gamma matrices
   CloverFieldType fillCloverYZ(const GaugeLinkField &F)
   {
     CloverFieldType T(F.Grid());
     T = Zero();
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
       T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
@@ -281,9 +282,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
     
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v, T,AcceleratorWrite);
+    autoView(F_v, F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = -F_v[i]()();
       T_v[i]()(1, 0) = F_v[i]()();
@@ -299,9 +300,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
       T_v[i]()(1, 1) = timesI(F_v[i]()());
@@ -317,9 +318,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v , T, AcceleratorWrite);
+    autoView( F_v , F, AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = timesI(F_v[i]()());
       T_v[i]()(1, 0) = timesI(F_v[i]()());
@@ -335,9 +336,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
     
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v ,T,AcceleratorWrite);
+    autoView( F_v ,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = -(F_v[i]()());
       T_v[i]()(1, 0) = (F_v[i]()());
@@ -354,9 +355,9 @@ private:
 
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v , T,AcceleratorWrite);
+    autoView( F_v , F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 0) = timesI(F_v[i]()());
       T_v[i]()(1, 1) = timesMinusI(F_v[i]()());

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -61,18 +61,19 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) {
+  accelerator_inline int CommDatumSize(void) const {
     return sizeof(SiteHalfCommSpinor);
   }
 
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  template<class _SiteHalfSpinor, class _SiteSpinor>
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
-    _SiteHalfSpinor tmp;
-    projector::Proj(tmp,in,mu,dag);
-    vstream(buf[o],tmp);
+  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);
   }
 
   /*****************************************************/
@@ -81,19 +82,24 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
 				   const SiteHalfSpinor * __restrict__ vp0,
 				   const SiteHalfSpinor * __restrict__ vp1,
-				   Integer type,Integer o){
+				   Integer type,Integer o) const {
+#ifdef GRID_SIMT
+    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+#else
     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) {    
+				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
     assert(0);
   }
 
@@ -103,8 +109,30 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
 					   SiteHalfSpinor * __restrict__ out1,
 					   const SiteSpinor * __restrict__ in,
-					   Integer j,Integer k, Integer m,Integer type)
+					   Integer j,Integer k, Integer m,Integer type) const
   {
+#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);
@@ -112,15 +140,17 @@ 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) { return false; }
+  accelerator_inline bool DecompressionStep(void) const { 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 >
@@ -142,20 +172,30 @@ public:
   typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
   constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
-  accelerator_inline int CommDatumSize(void) {
+  accelerator_inline int CommDatumSize(void) const {
     return sizeof(SiteHalfCommSpinor);
   }
 
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  template<class _SiteHalfSpinor, class _SiteSpinor>
-  accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) {
-    _SiteHalfSpinor hsp;
+  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
+    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 */
@@ -163,7 +203,7 @@ public:
   accelerator_inline void Exchange(SiteHalfSpinor *mp,
                        SiteHalfSpinor *vp0,
                        SiteHalfSpinor *vp1,
-		       Integer type,Integer o){
+		       Integer type,Integer o) const {
     SiteHalfSpinor vt0,vt1;
     SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
     SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
@@ -175,7 +215,7 @@ public:
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o){
+  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
     SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
     precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
   }
@@ -186,7 +226,7 @@ public:
   accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
 			       SiteHalfSpinor *out1,
 			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type){
+			       Integer j,Integer k, Integer m,Integer type) const {
     SiteHalfSpinor temp1, temp2,temp3,temp4;
     SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
     SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
@@ -200,9 +240,10 @@ public:
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) { return true; }
+  accelerator_inline bool DecompressionStep(void) const { return true; }
 
 };
+#endif
 
 #define DECLARE_PROJ(Projector,Compressor,spProj)			\
   class Projector {							\
@@ -253,33 +294,8 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  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;
-  }
+  void ZeroCountersi(void)  {  }
+  void Reporti(int calls)  {  }
 
   std::vector<int> surface_list;
 
@@ -321,26 +337,18 @@ 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>
@@ -352,12 +360,9 @@ 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;
       
@@ -393,7 +398,6 @@ 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);
@@ -74,6 +74,20 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
+
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent
@@ -138,7 +152,7 @@ public:
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
                 GridRedBlackCartesian &Hgrid, RealD _mass,
-                const ImplParams &p = ImplParams(), 
+                const ImplParams &p = ImplParams(),
                 const WilsonAnisotropyCoefficients &anis = WilsonAnisotropyCoefficients() );
 
   // DoubleStore impl dependent
@@ -170,9 +184,9 @@ public:
 
   LebesgueOrder Lebesgue;
   LebesgueOrder LebesgueEvenOdd;
-  
+
   WilsonAnisotropyCoefficients anisotropyCoeff;
-  
+
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
   ///////////////////////////////////////////////////////////////
@@ -186,7 +200,7 @@ public:
                            PropagatorField &q_out,
                            PropagatorField &phys_src,
                            Current curr_type,
-                           unsigned int mu, 
+                           unsigned int mu,
                            unsigned int tmin,
 			   unsigned int tmax,
 			   ComplexField &lattice_cmplx);
@@ -196,5 +210,3 @@ typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
 NAMESPACE_END(Grid);
-
-

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 typename StencilImpl::View_type StencilView;
+  typedef const typename StencilImpl::View_type StencilView;
     
   ImplParams Params;
 
@@ -106,11 +106,15 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
-    auto out_v= out.View();
-    auto phi_v= phi.View();
-    auto Umu_v= Umu.View();
-    thread_for(sss,out.Grid()->oSites(),{
-	multLink(out_v[sss],Umu_v[sss],phi_v[sss],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);
     });
   }
 					   
@@ -180,29 +184,57 @@ public:
       mat = TraceIndex<SpinIndex>(P); 
     }
       
-    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
+    inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds)
+    {
       for (int mu = 0; mu < Nd; mu++)
       mat[mu] = PeekIndex<LorentzIndex>(Uds, mu);
     }
 
-
-  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
-      
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu)
+  {
+#undef USE_OLD_INSERT_FORCE    
     int Ls=Btilde.Grid()->_fdimensions[0];
+    autoView( mat_v , mat, AcceleratorWrite);
+#ifdef USE_OLD_INSERT_FORCE    
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    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
-      }
-    });
+    {
+      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
+	  }
+	});
+    }
     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    
   }
 };
 
@@ -211,17 +243,17 @@ 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

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

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

@@ -180,7 +180,7 @@ template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
     std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
     std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls     : " << MooeeInvCalls   << std::endl;
     std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls            : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
     RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
     std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
     std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
@@ -642,7 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-#ifndef GRID_NVCC
+#if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -799,7 +799,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
-  LatticeInteger zz (UGrid);   zz=0.0;
+  PropagatorField zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
   for (int s=0;s<Ls;s++) {
 
@@ -826,7 +826,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
-#ifndef GRID_NVCC
+#if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
@@ -850,7 +850,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   PropagatorField tmp(UGrid);
   PropagatorField Utmp(UGrid);
 
-  LatticeInteger zz (UGrid);   zz=0.0;
+  PropagatorField  zz (UGrid);   zz=0.0;
   LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
 
   for(int s=0;s<Ls;s++){
@@ -880,17 +880,29 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
+  RealD sign = 1; // sign flip for vector/tadpole
   if ( curr_type == Current::Axial ) {
     for(int s=0;s<Ls/2;s++){
       G_s[s] = -1.0;
     }
   }
+  else if ( curr_type == Current::Tadpole ) {
+    auto b=this->_b;
+    auto c=this->_c;
+    if ( b == 1 && c == 0 ) {
+      sign = -1;    
+    }
+    else {
+      std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
+      assert(b==1 && c==0);
+    }
+  }
 
   for(int s=0;s<Ls;s++){
 
     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];
@@ -907,7 +919,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     tmp    = Cshift(tmp,mu,1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu);
-    tmp    = G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
+    tmp    = sign*G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
     tmp    = where((lcoor>=tmin),tmp,zz); // Mask the time 
     L_Q    = where((lcoor<=tmax),tmp,zz); // Position of current complicated
 

Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h

@@ -50,9 +50,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(phi , phi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
   auto pdiag = &diag[0];
@@ -93,9 +93,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
 {
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(phi , phi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
   auto pdiag = &diag[0];
@@ -131,8 +131,8 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
 
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
 
   int Ls=this->Ls;
 
@@ -193,8 +193,8 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
   GridBase *grid=psi_i.Grid();
   int Ls=this->Ls;
 
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,AcceleratorRead);
+  autoView(chi , chi_i,AcceleratorWrite);
 
   auto plee  = & lee [0];
   auto pdee  = & dee [0];

Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h

@@ -65,9 +65,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi, psi_i,CpuRead);
+  autoView(phi, phi_i,CpuRead);
+  autoView(chi, chi_i,CpuWrite);
   int Ls   = this->Ls;
   int LLs  = grid->_rdimensions[0];
   const int nsimd= Simd::Nsimd();
@@ -213,9 +213,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
-  auto psi=psi_i.View();
-  auto phi=phi_i.View();
-  auto chi=chi_i.View();
+  autoView(psi,psi_i,CpuRead);
+  autoView(phi,phi_i,CpuRead);
+  autoView(chi,chi_i,CpuWrite);
   int Ls   = this->Ls;
   int LLs  = grid->_rdimensions[0];
   int nsimd= Simd::Nsimd();
@@ -357,8 +357,8 @@ CayleyFermion5D<Impl>::MooeeInternalAsm(const FermionField &psi_i, FermionField
 					Vector<iSinglet<Simd> > &Matm)
 {
   EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i,CpuRead);
+  autoView(chi , chi_i,CpuWrite);
 #ifndef AVX512
   {
     SiteHalfSpinor BcastP;
@@ -535,8 +535,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
   EnableIf<Impl::LsVectorised,int> sfinae=0;
 #ifndef AVX512
   {
-    auto psi = psi_i.View();
-    auto chi = chi_i.View();
+    autoView(psi , psi_i,CpuRead);
+    autoView(chi , chi_i,CpuWrite);
 
     SiteHalfSpinor BcastP;
     SiteHalfSpinor BcastM;
@@ -586,8 +586,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
   }
 #else
   {
-    auto psi = psi_i.View();
-    auto chi = chi_i.View();
+    autoView(psi , psi_i,CpuRead);
+    autoView(chi , chi_i,CpuWrite);
     // pointers
     //  MASK_REGS;
 #define Chi_00 %zmm0

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

@@ -46,9 +46,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   chi_i.Checkerboard() = psi_i.Checkerboard();
   int Ls = this->Ls;
   GridBase* grid = psi_i.Grid();
-  auto phi = phi_i.View();
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView( phi , phi_i, AcceleratorRead);
+  autoView( psi , psi_i, AcceleratorRead);
+  autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
   auto pdiag = &diag[0];
   auto pupper = &upper[0];
@@ -82,9 +82,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
   GridBase* grid = psi_i.Grid();
   int Ls = this->Ls;
 
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView( psi , psi_i, AcceleratorRead);
+  autoView( phi , phi_i, AcceleratorRead);
+  autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
   auto pdiag = &diag[0];
   auto pupper = &upper[0];
@@ -116,8 +116,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase* grid = psi_i.Grid();
-  auto psi=psi_i.View();
-  auto chi=chi_i.View();
+  autoView( psi, psi_i, AcceleratorRead);
+  autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto plee  = & this->lee[0];
@@ -172,8 +172,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase* grid = psi_i.Grid();
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView( psi, psi_i, AcceleratorRead);
+  autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto plee  = & this->lee[0];

Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h

@@ -221,10 +221,10 @@ void ImprovedStaggeredFermion5D<Impl>::DhopDir(const FermionField &in, FermionFi
 
   Compressor compressor;
   Stencil.HaloExchange(in,compressor);
-  auto Umu_v   = Umu.View();
-  auto UUUmu_v = UUUmu.View();
-  auto in_v    = in.View();
-  auto out_v   = out.View();
+  autoView( Umu_v   ,   Umu, CpuRead);
+  autoView( UUUmu_v , UUUmu, CpuRead);
+  autoView( in_v    ,  in, CpuRead);
+  autoView( out_v   , out, CpuWrite);
   thread_for( ss,Umu.Grid()->oSites(),{
     for(int s=0;s<Ls;s++){
       int sU=ss;
@@ -281,11 +281,9 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
 						    const FermionField &in, FermionField &out,int dag)
 {
-#ifdef GRID_OMP
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-#endif
     DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 
@@ -294,9 +292,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
 								   DoubledGaugeField & U,DoubledGaugeField & UUU,
 								   const FermionField &in, FermionField &out,int dag)
 {
-#ifdef GRID_OMP
   //  assert((dag==DaggerNo) ||(dag==DaggerYes));
-
   Compressor compressor; 
 
   int LLs = in.Grid()->_rdimensions[0];
@@ -305,99 +301,42 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   DhopFaceTime-=usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
+
+  DhopCommTime -=usecond();
+  std::vector<std::vector<CommsRequest_t> > requests;
+  st.CommunicateBegin(requests);
+
   //  st.HaloExchangeOptGather(in,compressor); // Wilson compressor
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
   DhopFaceTime+=usecond();
 
-  double ctime=0;
-  double ptime=0;
-
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Ugly explicit thread mapping introduced for OPA reasons.
+  // Remove explicit thread mapping introduced for OPA reasons.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
+  DhopComputeTime-=usecond();
   {
-    int tid = omp_get_thread_num();
-    int nthreads = omp_get_num_threads();
-    int ncomms = CartesianCommunicator::nCommThreads;
-    if (ncomms == -1) ncomms = 1;
-    assert(nthreads > ncomms);
-    if (tid >= ncomms) {
-      double start = usecond();
-      nthreads -= ncomms;
-      int ttid  = tid - ncomms;
-      int n     = U.Grid()->oSites(); // 4d vol
-      int chunk = n / nthreads;
-      int rem   = n % nthreads;
-      int myblock, myn;
-      if (ttid < rem) {
-        myblock = ttid * chunk + ttid;
-        myn = chunk+1;
-      } else {
-        myblock = ttid*chunk + rem;
-        myn = chunk;
-      }
-
-      // do the compute
-      auto   U_v  =   U.View();
-      auto UUU_v  = UUU.View();
-      auto  in_v  =  in.View();
-      auto out_v  = out.View();
-
-      if (dag == DaggerYes) {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-          int sU = ss;
-	  // Interior = 1; Exterior = 0; must implement for staggered
-          Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<---------
-        }
-      } else {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-	  // Interior = 1; Exterior = 0;
-          int sU = ss;
-          Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<------------
-        }
-      }
-        ptime = usecond() - start;
-    } else {
-      double start = usecond();
-      st.CommunicateThreaded();
-      ctime = usecond() - start;
-    }
+    int interior=1;
+    int exterior=0;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
-  DhopCommTime += ctime;
-  DhopComputeTime+=ptime;
-
-  // First to enter, last to leave timing
-  st.CollateThreads();
+  DhopComputeTime+=usecond();
 
   DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
   DhopFaceTime+=usecond();
 
-  DhopComputeTime2-=usecond();
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
-  auto   U_v  =   U.View();
-  auto UUU_v  = UUU.View();
-  auto  in_v  =  in.View();
-  auto out_v  = out.View();
-  if (dag == DaggerYes) {
-    int sz=st.surface_list.size();
-    thread_for( ss,sz,{
-      int sU = st.surface_list[ss];
-      Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1); //<----------
-    });
-  } else {
-    int sz=st.surface_list.size();
-    thread_for( ss,sz,{
-      int sU = st.surface_list[ss];
-      Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1);//<----------
-    });
+  DhopComputeTime2-=usecond();
+  {
+    int interior=0;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime2+=usecond();
-#else
-  assert(0);
-#endif
-
 }
 
 template<class Impl>
@@ -408,8 +347,6 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   Compressor compressor;
   int LLs = in.Grid()->_rdimensions[0];
 
-
-
  //double t1=usecond();
   DhopTotalTime -= usecond();
   DhopCommTime -= usecond();
@@ -418,28 +355,13 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   
   DhopComputeTime -= usecond();
   // Dhop takes the 4d grid from U, and makes a 5d index for fermion
-  auto   U_v  =   U.View();
-  auto UUU_v  = UUU.View();
-  auto  in_v  =  in.View();
-  auto out_v  = out.View();
-  if (dag == DaggerYes) {
-    thread_for( ss,U.Grid()->oSites(),{
-      int sU=ss;
-      Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), LLs, sU,in_v, out_v);
-    });
-  } else {
-    thread_for( ss,U.Grid()->oSites(),{
-      int sU=ss;
-      Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v);
-    });
+  {
+    int interior=1;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime += usecond();
   DhopTotalTime   += usecond();
- //double t2=usecond();
- //std::cout << __FILE__ << " " << __func__  << " Total Time " << DhopTotalTime << std::endl;
- //std::cout << __FILE__ << " " << __func__  << " Total Time Org " << t2-t1 << std::endl;
- //std::cout << __FILE__ << " " << __func__  << " Comml Time " << DhopCommTime << std::endl;
- //std::cout << __FILE__ << " " << __func__  << " Compute Time " << DhopComputeTime << std::endl;
 
 }
 /*CHANGE END*/

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

@@ -258,10 +258,10 @@ void ImprovedStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGauge
     ////////////////////////
     // Call the single hop
     ////////////////////////
-    auto U_v   = U.View();
-    auto UUU_v = UUU.View();
-    auto B_v   = B.View();
-    auto Btilde_v   = Btilde.View();
+    autoView( U_v   , U, CpuRead);
+    autoView( UUU_v , UUU, CpuRead);
+    autoView( B_v      , B, CpuWrite);
+    autoView( Btilde_v , Btilde, CpuWrite);
     thread_for(sss,B.Grid()->oSites(),{
       Kernels::DhopDirKernel(st, U_v, UUU_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
     });
@@ -386,10 +386,10 @@ void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionFiel
 
   Compressor compressor;
   Stencil.HaloExchange(in, compressor);
-  auto Umu_v   =   Umu.View();
-  auto UUUmu_v = UUUmu.View();
-  auto in_v    =  in.View();
-  auto out_v   = out.View();
+  autoView( Umu_v   ,   Umu, CpuRead);
+  autoView( UUUmu_v , UUUmu, CpuRead);
+  autoView( in_v    ,  in, CpuRead);
+  autoView( out_v   , out, CpuWrite);
   thread_for( sss, in.Grid()->oSites(),{
     Kernels::DhopDirKernel(Stencil, Umu_v, UUUmu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
   });
@@ -403,11 +403,9 @@ void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder
 						  const FermionField &in,
 						  FermionField &out, int dag) 
 {
-#ifdef GRID_OMP
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-#endif
     DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 template <class Impl>
@@ -417,7 +415,6 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
 								 const FermionField &in,
 								 FermionField &out, int dag) 
 {
-#ifdef GRID_OMP
   Compressor compressor; 
   int len =  U.Grid()->oSites();
 
@@ -426,60 +423,30 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   DhopFaceTime    -= usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
-  st.CommsMergeSHM(compressor);
   DhopFaceTime    += usecond();
 
+  DhopCommTime -=usecond();
+  std::vector<std::vector<CommsRequest_t> > requests;
+  st.CommunicateBegin(requests);
+
+  DhopFaceTime-=usecond();
+  st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
+
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Ugly explicit thread mapping introduced for OPA reasons.
+  // Removed explicit thread comms
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   DhopComputeTime    -= usecond();
-#pragma omp parallel 
   {
-    int tid = omp_get_thread_num();
-    int nthreads = omp_get_num_threads();
-    int ncomms = CartesianCommunicator::nCommThreads;
-    if (ncomms == -1) ncomms = 1;
-    assert(nthreads > ncomms);
-
-    if (tid >= ncomms) {
-      nthreads -= ncomms;
-      int ttid  = tid - ncomms;
-      int n     = len;
-      int chunk = n / nthreads;
-      int rem   = n % nthreads;
-      int myblock, myn;
-      if (ttid < rem) {
-        myblock = ttid * chunk + ttid;
-        myn = chunk+1;
-      } else {
-        myblock = ttid*chunk + rem;
-        myn = chunk;
-      }
-
-      // do the compute
-      auto U_v   = U.View();
-      auto UUU_v = UUU.View();
-      auto in_v  = in.View();
-      auto out_v = out.View();
-      if (dag == DaggerYes) {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-          int sU = ss;
-	  // Interior = 1; Exterior = 0; must implement for staggered
-          Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0); 
-        }
-      } else {
-        for (int ss = myblock; ss < myblock+myn; ++ss) {
-	  // Interior = 1; Exterior = 0;
-          int sU = ss;
-          Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0);
-        }
-      }
-    } else {
-      st.CommunicateThreaded();
-    }
+    int interior=1;
+    int exterior=0;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime    += usecond();
 
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
+
   // First to enter, last to leave timing
   DhopFaceTime    -= usecond();
   st.CommsMerge(compressor);
@@ -487,28 +454,11 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
 
   DhopComputeTime2    -= usecond();
   {
-    auto U_v   = U.View();
-    auto UUU_v = UUU.View();
-    auto in_v  = in.View();
-    auto out_v = out.View();
-    if (dag == DaggerYes) {
-      int sz=st.surface_list.size();
-      thread_for(ss,sz,{
-	int sU = st.surface_list[ss];
-	Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
-      });
-    } else {
-      int sz=st.surface_list.size();
-      thread_for(ss,sz,{
-	int sU = st.surface_list[ss];
-	Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
-      });
-    }
+    int interior=0;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime2    += usecond();
-#else
-  assert(0);
-#endif
 }
 
 
@@ -528,19 +478,11 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
   st.HaloExchange(in, compressor);
   DhopCommTime    += usecond();
 
-  auto U_v   =   U.View();
-  auto UUU_v = UUU.View();
-  auto in_v  =  in.View();
-  auto out_v = out.View();
   DhopComputeTime -= usecond();
-  if (dag == DaggerYes) {
-    thread_for(sss, in.Grid()->oSites(),{
-      Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
-    });
-  } else {
-    thread_for(sss, in.Grid()->oSites(),{
-      Kernels::DhopSite(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
-    });
+  {
+    int interior=1;
+    int exterior=1;
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
   DhopComputeTime += usecond();
   DhopTotalTime   += usecond();

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

@@ -44,9 +44,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
@@ -84,9 +84,9 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto pm  = this->pm;
   int shift_s = (pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
@@ -132,9 +132,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
@@ -174,9 +174,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
   int shift_s = (this->pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
-  auto psi = psi_i.View();
-  auto phi = phi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(phi , phi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
@@ -226,8 +226,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto plee = & this->lee [0];
   auto pdee = & this->dee [0];
@@ -286,8 +286,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto pm = this->pm;
   auto plee = & this->lee [0];
@@ -354,8 +354,8 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
   int Ls = this->Ls;
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
 
   auto plee = & this->lee [0];
   auto pdee = & this->dee [0];
@@ -410,8 +410,8 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
-  auto psi = psi_i.View();
-  auto chi = chi_i.View();
+  autoView(psi , psi_i, AcceleratorRead);
+  autoView(chi , chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
   auto pm = this->pm;

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

@@ -0,0 +1,499 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/Grid.h>
+
+#pragma once 
+
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////
+// Constructor and gauge import
+/////////////////////////////////
+
+template <class Impl>
+NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid, 
+						   RealD _mass,
+						   RealD _c1, RealD _u0,
+						   const ImplParams &p)
+  : Kernels(p),
+    _grid(&Fgrid),
+    _cbgrid(&Hgrid),
+    Stencil(&Fgrid, npoint, Even, directions, displacements,p),
+    StencilEven(&Hgrid, npoint, Even, directions, displacements,p),  // source is Even
+    StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p),  // source is Odd
+    mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
+    Umu(&Fgrid),
+    UmuEven(&Hgrid),
+    UmuOdd(&Hgrid),
+    _tmp(&Hgrid)
+{
+  int vol4;
+  int LLs=1;
+  c1=_c1;
+  u0=_u0;
+  vol4= _grid->oSites();
+  Stencil.BuildSurfaceList(LLs,vol4);
+  vol4= _cbgrid->oSites();
+  StencilEven.BuildSurfaceList(LLs,vol4);
+  StencilOdd.BuildSurfaceList(LLs,vol4);
+}
+
+template <class Impl>
+NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
+						   GridRedBlackCartesian &Hgrid, RealD _mass,
+						   RealD _c1, RealD _u0,
+						   const ImplParams &p)
+  : NaiveStaggeredFermion(Fgrid,Hgrid,_mass,_c1,_u0,p)
+{
+  ImportGauge(_U);
+}
+
+////////////////////////////////////////////////////////////
+// Momentum space propagator should be 
+// https://arxiv.org/pdf/hep-lat/9712010.pdf
+//
+// mom space action.
+//   gamma_mu i ( c1 sin pmu + c2 sin 3 pmu ) + m
+//
+// must track through staggered flavour/spin reduction in literature to 
+// turn to free propagator for the one component chi field, a la page 4/5
+// of above link to implmement fourier based solver.
+////////////////////////////////////////////////////////////
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::CopyGaugeCheckerboards(void)
+{
+  pickCheckerboard(Even, UmuEven,  Umu);
+  pickCheckerboard(Odd,  UmuOdd ,  Umu);
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::ImportGauge(const GaugeField &_U) 
+{
+  GaugeLinkField U(GaugeGrid());
+  DoubledGaugeField _UUU(GaugeGrid());
+  ////////////////////////////////////////////////////////
+  // Double Store should take two fields for Naik and one hop separately.
+  // Discard teh Naik as Naive
+  ////////////////////////////////////////////////////////
+  Impl::DoubleStore(GaugeGrid(), _UUU, Umu, _U, _U );
+
+  ////////////////////////////////////////////////////////
+  // Apply scale factors to get the right fermion Kinetic term
+  // Could pass coeffs into the double store to save work.
+  // 0.5 ( U p(x+mu) - Udag(x-mu) p(x-mu) ) 
+  ////////////////////////////////////////////////////////
+  for (int mu = 0; mu < Nd; mu++) {
+
+    U = PeekIndex<LorentzIndex>(Umu, mu);
+    PokeIndex<LorentzIndex>(Umu, U*( 0.5*c1/u0), mu );
+    
+    U = PeekIndex<LorentzIndex>(Umu, mu+4);
+    PokeIndex<LorentzIndex>(Umu, U*(-0.5*c1/u0), mu+4);
+
+  }
+
+  CopyGaugeCheckerboards();
+}
+
+/////////////////////////////
+// Implement the interface
+/////////////////////////////
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  Dhop(in, out, DaggerNo);
+  axpy(out, mass, in, out);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  Dhop(in, out, DaggerYes);
+  axpy(out, mass, in, out);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerNo);
+  } else {
+    DhopOE(in, out, DaggerNo);
+  }
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerYes);
+  } else {
+    DhopOE(in, out, DaggerYes);
+  }
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  typename FermionField::scalar_type scal(mass);
+  out = scal * in;
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  Mooee(in, out);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  out = (1.0 / (mass)) * in;
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) 
+{
+  out.Checkerboard() = in.Checkerboard();
+  MooeeInv(in, out);
+}
+
+///////////////////////////////////
+// Internal
+///////////////////////////////////
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
+						GaugeField & mat,
+						const FermionField &A, const FermionField &B, int dag) 
+{
+  assert((dag == DaggerNo) || (dag == DaggerYes));
+
+  Compressor compressor;
+
+  FermionField Btilde(B.Grid());
+  FermionField Atilde(B.Grid());
+  Atilde = A;
+
+  st.HaloExchange(B, compressor);
+
+  for (int mu = 0; mu < Nd; mu++) {
+
+    ////////////////////////
+    // Call the single hop
+    ////////////////////////
+    autoView( U_v      , U, CpuRead);
+    autoView( B_v      , B, CpuWrite);
+    autoView( Btilde_v , Btilde, CpuWrite);
+    thread_for(sss,B.Grid()->oSites(),{
+      Kernels::DhopDirKernel(st, U_v, U_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
+    });
+
+    assert(0);// need to figure out the force interface with a blasted three link term.
+    
+  }
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
+
+  conformable(U.Grid(), _grid);
+  conformable(U.Grid(), V.Grid());
+  conformable(U.Grid(), mat.Grid());
+
+  mat.Checkerboard() = U.Checkerboard();
+
+  DerivInternal(Stencil, Umu, mat, U, V, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
+
+  conformable(U.Grid(), _cbgrid);
+  conformable(U.Grid(), V.Grid());
+  conformable(U.Grid(), mat.Grid());
+
+  assert(V.Checkerboard() == Even);
+  assert(U.Checkerboard() == Odd);
+  mat.Checkerboard() = Odd;
+
+  DerivInternal(StencilEven, UmuOdd, mat, U, V, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
+
+  conformable(U.Grid(), _cbgrid);
+  conformable(U.Grid(), V.Grid());
+  conformable(U.Grid(), mat.Grid());
+
+  assert(V.Checkerboard() == Odd);
+  assert(U.Checkerboard() == Even);
+  mat.Checkerboard() = Even;
+
+  DerivInternal(StencilOdd, UmuEven, mat, U, V, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) 
+{
+  DhopCalls+=2;
+  conformable(in.Grid(), _grid);  // verifies full grid
+  conformable(in.Grid(), out.Grid());
+
+  out.Checkerboard() = in.Checkerboard();
+
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) 
+{
+  DhopCalls+=1;
+  conformable(in.Grid(), _cbgrid);    // verifies half grid
+  conformable(in.Grid(), out.Grid());  // drops the cb check
+
+  assert(in.Checkerboard() == Even);
+  out.Checkerboard() = Odd;
+
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) 
+{
+  DhopCalls+=1;
+  conformable(in.Grid(), _cbgrid);    // verifies half grid
+  conformable(in.Grid(), out.Grid());  // drops the cb check
+
+  assert(in.Checkerboard() == Odd);
+  out.Checkerboard() = Even;
+
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) 
+{
+  DhopDir(in, out, dir, disp);
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out) 
+{
+  assert(0); // Not implemented yet
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp) 
+{
+
+  Compressor compressor;
+  Stencil.HaloExchange(in, compressor);
+  autoView( Umu_v   ,  Umu, CpuRead);
+  autoView( in_v    ,  in, CpuRead);
+  autoView( out_v   , out, CpuWrite);
+  //  thread_for( sss, in.Grid()->oSites(),{
+  //    Kernels::DhopDirKernel(Stencil, Umu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
+  //  });
+  assert(0);
+};
+
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
+					       DoubledGaugeField &U,
+					       const FermionField &in,
+					       FermionField &out, int dag) 
+{
+  if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
+  else
+    DhopInternalSerialComms(st,lo,U,in,out,dag);
+}
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
+							      DoubledGaugeField &U,
+							      const FermionField &in,
+							      FermionField &out, int dag) 
+{
+  Compressor compressor; 
+  int len =  U.Grid()->oSites();
+
+  DhopTotalTime   -= usecond();
+
+  DhopFaceTime    -= usecond();
+  st.Prepare();
+  st.HaloGather(in,compressor);
+  DhopFaceTime    += usecond();
+
+  DhopCommTime -=usecond();
+  std::vector<std::vector<CommsRequest_t> > requests;
+  st.CommunicateBegin(requests);
+
+  DhopFaceTime-=usecond();
+  st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Removed explicit thread comms
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime    -= usecond();
+  {
+    int interior=1;
+    int exterior=0;
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
+  }
+  DhopComputeTime    += usecond();
+
+  st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
+
+  // First to enter, last to leave timing
+  DhopFaceTime    -= usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime    -= usecond();
+
+  DhopComputeTime2    -= usecond();
+  {
+    int interior=0;
+    int exterior=1;
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
+  }
+  DhopComputeTime2    += usecond();
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
+							  DoubledGaugeField &U,
+							  const FermionField &in,
+							  FermionField &out, int dag) 
+{
+  assert((dag == DaggerNo) || (dag == DaggerYes));
+
+  DhopTotalTime   -= usecond();
+
+  DhopCommTime    -= usecond();
+  Compressor compressor;
+  st.HaloExchange(in, compressor);
+  DhopCommTime    += usecond();
+
+  DhopComputeTime -= usecond();
+  {
+    int interior=1;
+    int exterior=1;
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
+  }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
+};
+
+  ////////////////////////////////////////////////////////////////
+  // Reporting
+  ////////////////////////////////////////////////////////////////
+template<class Impl>
+void NaiveStaggeredFermion<Impl>::Report(void) 
+{
+  Coordinate latt = _grid->GlobalDimensions();
+  RealD volume = 1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "NaiveStaggeredFermion Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "NaiveStaggeredFermion ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _grid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "NaiveStaggeredFermion Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "NaiveStaggeredFermion StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "NaiveStaggeredFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void NaiveStaggeredFermion<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime   = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
+
+//////////////////////////////////////////////////////// 
+// Conserved current - not yet implemented.
+////////////////////////////////////////////////////////
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
+							      PropagatorField &q_in_2,
+							      PropagatorField &q_out,
+							      PropagatorField &src,
+							      Current curr_type,
+							      unsigned int mu)
+{
+  assert(0);
+}
+
+template <class Impl>
+void NaiveStaggeredFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
+                                                         PropagatorField &q_out,
+                                                         PropagatorField &src,
+                                                         Current curr_type,
+                                                         unsigned int mu, 
+                                                         unsigned int tmin,
+                                              unsigned int tmax,
+					      ComplexField &lattice_cmplx)
+{
+  assert(0);
+
+}
+
+NAMESPACE_END(Grid);

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

@@ -618,10 +618,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+void StaggeredKernels<Impl>::DhopSiteAsm(StencilView &st,
 					 DoubledGaugeFieldView &U,
 					 DoubledGaugeFieldView &UUU,
-					 SiteSpinor *buf, int LLs,
+					 SiteSpinor *buf, int sF,
 					 int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   assert(0);
@@ -680,12 +680,14 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
   gauge2 =(uint64_t)&UU[sU]( Z );				\
   gauge3 =(uint64_t)&UU[sU]( T ); 
   
+#undef STAG_VEC5D
+#ifdef STAG_VEC5D
   // This is the single precision 5th direction vectorised kernel
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
 								    DoubledGaugeFieldView &U,
 								    DoubledGaugeFieldView &UUU,
-								    SiteSpinor *buf, int LLs,
+								    SiteSpinor *buf, int sF,
 								    int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
@@ -702,9 +704,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-   for(int s=0;s<LLs;s++){
+  //   for(int s=0;s<LLs;s++){
 
-    int sF=s+LLs*sU;
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHI(addr0,addr1,addr2,addr3);
@@ -736,10 +739,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView &st, 
 								    DoubledGaugeFieldView &U,
 								    DoubledGaugeFieldView &UUU,
-								    SiteSpinor *buf, int LLs,
+								    SiteSpinor *buf, int sF,
 								    int sU, const FermionFieldView &in, FermionFieldView &out, int dag) 
 {
 #ifdef AVX512
@@ -756,8 +759,9 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHI(addr0,addr1,addr2,addr3);
@@ -787,7 +791,7 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
 #endif
 }
    
-   
+#endif   
 
 
 #define PERMUTE_DIR3 __asm__ (	\
@@ -821,10 +825,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
   // This is the single precision 5th direction vectorised kernel
 
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st, 
 							       DoubledGaugeFieldView &U,
 							       DoubledGaugeFieldView &UUU,
-							       SiteSpinor *buf, int LLs,
+							       SiteSpinor *buf, int sF,
 							       int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
@@ -841,9 +845,9 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-  for(int s=0;s<LLs;s++){
-    
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHIa(addr0,addr1);
@@ -890,10 +894,10 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilView &st, 
 							       DoubledGaugeFieldView &U,
 							       DoubledGaugeFieldView &UUU,
-							       SiteSpinor *buf, int LLs,
+							       SiteSpinor *buf, int sF,
 							       int sU, const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
 #ifdef AVX512
@@ -910,9 +914,9 @@ template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st,
   StencilEntry *SE2;
   StencilEntry *SE3;
 
-  for(int s=0;s<LLs;s++){
-    
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
     // Xp, Yp, Zp, Tp
     PREPARE(Xp,Yp,Zp,Tp,0,U);
     LOAD_CHIa(addr0,addr1);

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

@@ -32,25 +32,50 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-#define LOAD_CHI(b)		\
+#ifdef GRID_SIMT
+
+#define LOAD_CHI(ptype,b)			\
+  const SiteSpinor & ref (b[offset]);				\
+  Chi_0=coalescedReadPermute<ptype>(ref()()(0),perm,lane);	\
+  Chi_1=coalescedReadPermute<ptype>(ref()()(1),perm,lane);	\
+  Chi_2=coalescedReadPermute<ptype>(ref()()(2),perm,lane);
+
+#define LOAD_CHI_COMMS(b)		\
   const SiteSpinor & ref (b[offset]);	\
-    Chi_0=ref()()(0);\
-    Chi_1=ref()()(1);\
-    Chi_2=ref()()(2);
+  Chi_0=coalescedRead(ref()()(0),lane);	\
+  Chi_1=coalescedRead(ref()()(1),lane);	\
+  Chi_2=coalescedRead(ref()()(2),lane);
+
+#define PERMUTE_DIR(dir)	;
+#else
+#define LOAD_CHI(ptype,b)      LOAD_CHI_COMMS(b)
+
+#define LOAD_CHI_COMMS(b)		\
+  const SiteSpinor & ref (b[offset]);	\
+  Chi_0=ref()()(0);			\
+  Chi_1=ref()()(1);			\
+  Chi_2=ref()()(2);
+
+#define PERMUTE_DIR(dir)			\
+  permute##dir(Chi_0,Chi_0);			\
+  permute##dir(Chi_1,Chi_1);			\
+  permute##dir(Chi_2,Chi_2);
+
+#endif
 
 
 // To splat or not to splat depends on the implementation
 #define MULT(A,UChi)				\
   auto & ref(U[sU](A));			\
-   Impl::loadLinkElement(U_00,ref()(0,0));      \
-   Impl::loadLinkElement(U_10,ref()(1,0));      \
-   Impl::loadLinkElement(U_20,ref()(2,0));      \
-   Impl::loadLinkElement(U_01,ref()(0,1));      \
-   Impl::loadLinkElement(U_11,ref()(1,1));      \
-   Impl::loadLinkElement(U_21,ref()(2,1));      \
-   Impl::loadLinkElement(U_02,ref()(0,2));     \
-   Impl::loadLinkElement(U_12,ref()(1,2));     \
-   Impl::loadLinkElement(U_22,ref()(2,2));     \
+    U_00=coalescedRead(ref()(0,0),lane);				\
+    U_10=coalescedRead(ref()(1,0),lane);				\
+    U_20=coalescedRead(ref()(2,0),lane);				\
+    U_01=coalescedRead(ref()(0,1),lane);				\
+    U_11=coalescedRead(ref()(1,1),lane);				\
+    U_21=coalescedRead(ref()(2,1),lane);				\
+    U_02=coalescedRead(ref()(0,2),lane);				\
+    U_12=coalescedRead(ref()(1,2),lane);				\
+    U_22=coalescedRead(ref()(2,2),lane);				\
     UChi ## _0  = U_00*Chi_0;	       \
     UChi ## _1  = U_10*Chi_0;\
     UChi ## _2  = U_20*Chi_0;\
@@ -63,15 +88,15 @@ NAMESPACE_BEGIN(Grid);
 
 #define MULT_ADD(U,A,UChi)			\
   auto & ref(U[sU](A));			\
-   Impl::loadLinkElement(U_00,ref()(0,0));      \
-   Impl::loadLinkElement(U_10,ref()(1,0));      \
-   Impl::loadLinkElement(U_20,ref()(2,0));      \
-   Impl::loadLinkElement(U_01,ref()(0,1));      \
-   Impl::loadLinkElement(U_11,ref()(1,1));      \
-   Impl::loadLinkElement(U_21,ref()(2,1));      \
-   Impl::loadLinkElement(U_02,ref()(0,2));     \
-   Impl::loadLinkElement(U_12,ref()(1,2));     \
-   Impl::loadLinkElement(U_22,ref()(2,2));     \
+    U_00=coalescedRead(ref()(0,0),lane);				\
+    U_10=coalescedRead(ref()(1,0),lane);				\
+    U_20=coalescedRead(ref()(2,0),lane);				\
+    U_01=coalescedRead(ref()(0,1),lane);				\
+    U_11=coalescedRead(ref()(1,1),lane);				\
+    U_21=coalescedRead(ref()(2,1),lane);				\
+    U_02=coalescedRead(ref()(0,2),lane);				\
+    U_12=coalescedRead(ref()(1,2),lane);				\
+    U_22=coalescedRead(ref()(2,2),lane);				\
     UChi ## _0 += U_00*Chi_0;	       \
     UChi ## _1 += U_10*Chi_0;\
     UChi ## _2 += U_20*Chi_0;\
@@ -83,24 +108,18 @@ NAMESPACE_BEGIN(Grid);
     UChi ## _2 += U_22*Chi_2;
 
 
-#define PERMUTE_DIR(dir)			\
-  permute##dir(Chi_0,Chi_0);			\
-  permute##dir(Chi_1,Chi_1);			\
-  permute##dir(Chi_2,Chi_2);
-
-
 #define HAND_STENCIL_LEG_BASE(Dir,Perm,skew)	\
   SE=st.GetEntry(ptype,Dir+skew,sF);	\
   offset = SE->_offset;			\
   local  = SE->_is_local;		\
   perm   = SE->_permute;		\
   if ( local ) {						\
-    LOAD_CHI(in);					\
+    LOAD_CHI(Perm,in);						\
     if ( perm) {						\
       PERMUTE_DIR(Perm);					\
     }								\
   } else {							\
-    LOAD_CHI(buf);						\
+    LOAD_CHI_COMMS(buf);					\
   }								
 
 #define HAND_STENCIL_LEG_BEGIN(Dir,Perm,skew,even)		\
@@ -116,19 +135,18 @@ NAMESPACE_BEGIN(Grid);
   }
 
 
-
 #define HAND_STENCIL_LEG_INT(U,Dir,Perm,skew,even)	\
   SE=st.GetEntry(ptype,Dir+skew,sF);			\
   offset = SE->_offset;					\
   local  = SE->_is_local;				\
   perm   = SE->_permute;				\
   if ( local ) {					\
-    LOAD_CHI(in);				\
+    LOAD_CHI(Perm,in);					\
     if ( perm) {					\
       PERMUTE_DIR(Perm);				\
     }							\
   } else if ( st.same_node[Dir] ) {			\
-    LOAD_CHI(buf);					\
+    LOAD_CHI_COMMS(buf);				\
   }							\
   if (local || st.same_node[Dir] ) {		\
     MULT_ADD(U,Dir,even);				\
@@ -140,49 +158,59 @@ NAMESPACE_BEGIN(Grid);
   local  = SE->_is_local;				\
   if ((!local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
-    { LOAD_CHI(buf);	  }					\
+    { LOAD_CHI_COMMS(buf);	  }				\
     { MULT_ADD(U,Dir,even); }					\
   }								
 
+#define HAND_DECLARATIONS(Simd) \
+  Simd even_0;			\
+  Simd even_1;			\
+  Simd even_2;			\
+  Simd odd_0;			\
+  Simd odd_1;			\
+  Simd odd_2;		        \
+		      		\
+  Simd Chi_0;			\
+  Simd Chi_1;			\
+  Simd Chi_2;			\
+				\
+  Simd U_00;			\
+  Simd U_10;			\
+  Simd U_20;			\
+  Simd U_01;			\
+  Simd U_11;			\
+  Simd U_21;			\
+  Simd U_02;			\
+  Simd U_12;			\
+  Simd U_22;			
+  
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik> accelerator_inline
+void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
 					  DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
-					  SiteSpinor *buf, int LLs, int sU, 
+					  SiteSpinor *buf, int sF, int sU, 
 					  const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
 
-  Simd Chi_0;    // two spinor; 6 regs
-  Simd Chi_1;
-  Simd Chi_2;
-  
-  Simd U_00;  // two rows of U matrix
-  Simd U_10;
-  Simd U_20;  
-  Simd U_01;
-  Simd U_11;
-  Simd U_21;  // 2 reg left.
-  Simd U_02;
-  Simd U_12;
-  Simd U_22; 
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  SiteSpinor result;
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
+  calcSiteSpinor result;
   int offset,local,perm, ptype;
 
   StencilEntry *SE;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
 
     skew = 0;
     HAND_STENCIL_LEG_BEGIN(Xp,3,skew,even);  
@@ -193,6 +221,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG      (U,Ym,2,skew,odd);   
     HAND_STENCIL_LEG      (U,Zm,1,skew,even);  
     HAND_STENCIL_LEG      (U,Tm,0,skew,odd);  
+    if (Naik) {
     skew = 8;
     HAND_STENCIL_LEG(UUU,Xp,3,skew,even);  
     HAND_STENCIL_LEG(UUU,Yp,2,skew,odd);   
@@ -202,7 +231,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG(UUU,Ym,2,skew,odd);   
     HAND_STENCIL_LEG(UUU,Zm,1,skew,even);  
     HAND_STENCIL_LEG(UUU,Tm,0,skew,odd);  
-    
+    }    
     if ( dag ) {
       result()()(0) = - even_0 - odd_0;
       result()()(1) = - even_1 - odd_1;
@@ -212,52 +241,39 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
       result()()(1) = even_1 + odd_1;
       result()()(2) = even_2 + odd_2;
     }
-    vstream(out[sF],result);
+    coalescedWrite(out[sF],result);
   }
 }
 
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik> accelerator_inline
+void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st, 
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					     SiteSpinor *buf, int LLs, int sU, 
+					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  Simd Chi_0;    // two spinor; 6 regs
-  Simd Chi_1;
-  Simd Chi_2;
-  
-  Simd U_00;  // two rows of U matrix
-  Simd U_10;
-  Simd U_20;  
-  Simd U_01;
-  Simd U_11;
-  Simd U_21;  // 2 reg left.
-  Simd U_02;
-  Simd U_12;
-  Simd U_22; 
-
-  SiteSpinor result;
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
+  calcSiteSpinor result;
   int offset, ptype, local, perm;
 
   StencilEntry *SE;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
 
-    even_0 = Zero();    even_1 = Zero();    even_2 = Zero();
-     odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
+    zeroit(even_0);    zeroit(even_1);    zeroit(even_2);
+    zeroit(odd_0);    zeroit(odd_1);    zeroit(odd_2);
 
     skew = 0;
     HAND_STENCIL_LEG_INT(U,Xp,3,skew,even);  
@@ -268,6 +284,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_INT(U,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_INT(U,Zm,1,skew,even);  
     HAND_STENCIL_LEG_INT(U,Tm,0,skew,odd);  
+    if (Naik) {
     skew = 8;
     HAND_STENCIL_LEG_INT(UUU,Xp,3,skew,even);  
     HAND_STENCIL_LEG_INT(UUU,Yp,2,skew,odd);   
@@ -277,7 +294,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_INT(UUU,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_INT(UUU,Zm,1,skew,even);  
     HAND_STENCIL_LEG_INT(UUU,Tm,0,skew,odd);  
-
+    }
     // Assume every site must be connected to at least one interior point. No 1^4 subvols.
     if ( dag ) {
       result()()(0) = - even_0 - odd_0;
@@ -288,52 +305,39 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
       result()()(1) = even_1 + odd_1;
       result()()(2) = even_2 + odd_2;
     }
-    vstream(out[sF],result);
+    coalescedWrite(out[sF],result);
   }
 }
 
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik> accelerator_inline
+void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					     SiteSpinor *buf, int LLs, int sU, 
+					     SiteSpinor *buf, int sF, int sU, 
 					     const FermionFieldView &in, FermionFieldView &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  Simd even_0; // 12 regs on knc
-  Simd even_1;
-  Simd even_2;
-  Simd odd_0; // 12 regs on knc
-  Simd odd_1;
-  Simd odd_2;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  typedef decltype( coalescedRead( in[0]()()(0) )) Simt;
+  HAND_DECLARATIONS(Simt);
 
-  Simd Chi_0;    // two spinor; 6 regs
-  Simd Chi_1;
-  Simd Chi_2;
-  
-  Simd U_00;  // two rows of U matrix
-  Simd U_10;
-  Simd U_20;  
-  Simd U_01;
-  Simd U_11;
-  Simd U_21;  // 2 reg left.
-  Simd U_02;
-  Simd U_12;
-  Simd U_22; 
-
-  SiteSpinor result;
+  typedef decltype( coalescedRead( in[0] )) calcSiteSpinor;
+  calcSiteSpinor result;
   int offset, ptype, local;
 
   StencilEntry *SE;
   int skew;
 
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=s+LLs*sU;
+  {
 
-    even_0 = Zero();    even_1 = Zero();    even_2 = Zero();
-     odd_0 = Zero();     odd_1 = Zero();     odd_2 = Zero();
+    zeroit(even_0);    zeroit(even_1);    zeroit(even_2);
+    zeroit(odd_0);    zeroit(odd_1);    zeroit(odd_2);
     int nmu=0;
     skew = 0;
     HAND_STENCIL_LEG_EXT(U,Xp,3,skew,even);  
@@ -344,6 +348,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_EXT(U,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_EXT(U,Zm,1,skew,even);  
     HAND_STENCIL_LEG_EXT(U,Tm,0,skew,odd);  
+    if (Naik) {
     skew = 8;
     HAND_STENCIL_LEG_EXT(UUU,Xp,3,skew,even);  
     HAND_STENCIL_LEG_EXT(UUU,Yp,2,skew,odd);   
@@ -353,7 +358,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
     HAND_STENCIL_LEG_EXT(UUU,Ym,2,skew,odd);   
     HAND_STENCIL_LEG_EXT(UUU,Zm,1,skew,even);  
     HAND_STENCIL_LEG_EXT(UUU,Tm,0,skew,odd);  
-
+    }
     // Add sum of all exterior connected stencil legs
     if ( nmu ) { 
       if ( dag ) {
@@ -365,11 +370,12 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
 	result()()(1) = even_1 + odd_1;
 	result()()(2) = even_2 + odd_2;
       }
-      out[sF] = out[sF] + result;
+      coalescedWrite(out[sF] , out(sF)+ result);
     }
   }
 }
 
+/*
 #define DHOP_SITE_HAND_INSTANTIATE(IMPL)				\
   template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
 						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
@@ -385,8 +391,9 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
 						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
 						     SiteSpinor *buf, int LLs, int sU, \
 						     const FermionFieldView &in, FermionFieldView &out, int dag); \
-
+*/
 #undef LOAD_CHI
+#undef HAND_DECLARATIONS
 
 NAMESPACE_END(Grid);
 

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

@@ -35,39 +35,32 @@ NAMESPACE_BEGIN(Grid);
 #define GENERIC_STENCIL_LEG(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
-      chi_p = χ						\
-      permute(chi,  in[SE->_offset], ptype);		\
-    } else {							\
-      chi_p = &in[SE->_offset];				\
-    }								\
+    int perm= SE->_permute;						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
   } else {							\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  multLink(Uchi, U[sU], *chi_p, Dir);			
+  acceleratorSynchronise();					\
+  multLink(Uchi, U[sU], chi, Dir);			
 
 #define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if (SE->_is_local ) {						\
-    if (SE->_permute) {						\
-      chi_p = χ						\
-      permute(chi,  in[SE->_offset], ptype);		\
-    } else {							\
-      chi_p = &in[SE->_offset];				\
-    }								\
+    int perm= SE->_permute;						\
+    chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
   } else if ( st.same_node[Dir] ) {				\
-    chi_p = &buf[SE->_offset];					\
+    chi = coalescedRead(buf[SE->_offset],lane);                 \
   }								\
   if (SE->_is_local || st.same_node[Dir] ) {			\
-    multLink(Uchi, U[sU], *chi_p, Dir);			\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 #define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink)		\
   SE = st.GetEntry(ptype, Dir+skew, sF);			\
   if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
     nmu++;							\
-    chi_p = &buf[SE->_offset];					\
-    multLink(Uchi, U[sU], *chi_p, Dir);			\
+    chi = coalescedRead(buf[SE->_offset],lane);			\
+    multLink(Uchi, U[sU], chi, Dir);				\
   }
 
 template <class Impl>
@@ -78,19 +71,25 @@ StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
 // Int, Ext, Int+Ext cases for comms overlap
 ////////////////////////////////////////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik> accelerator_inline
+void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st, 
 					     DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					     SiteSpinor *buf, int LLs, int sU, 
-					     const FermionFieldView &in, FermionFieldView &out, int dag) {
-  const SiteSpinor *chi_p;
-  SiteSpinor chi;
-  SiteSpinor Uchi;
+					     SiteSpinor *buf, int sF, int sU, 
+					     const FermionFieldView &in, FermionFieldView &out, int dag) 
+{
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
-  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+s;
+  //  for(int s=0;s<LLs;s++){
+  //
+  //    int sF=LLs*sU+s;
+  {
     skew = 0;
     GENERIC_STENCIL_LEG(U,Xp,skew,Impl::multLink);
     GENERIC_STENCIL_LEG(U,Yp,skew,Impl::multLinkAdd);
@@ -100,6 +99,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
     GENERIC_STENCIL_LEG(U,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(U,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(U,Tm,skew,Impl::multLinkAdd);
+    if ( Naik ) {
     skew=8;
     GENERIC_STENCIL_LEG(UUU,Xp,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(UUU,Yp,skew,Impl::multLinkAdd);
@@ -109,10 +109,11 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
     GENERIC_STENCIL_LEG(UUU,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG(UUU,Tm,skew,Impl::multLinkAdd);
+    }
     if ( dag ) { 
       Uchi = - Uchi;
     } 
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -120,19 +121,24 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
   // Only contributions from interior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik> accelerator_inline
+void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-						SiteSpinor *buf, int LLs, int sU, 
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
-  const SiteSpinor *chi_p;
-  SiteSpinor chi;
-  SiteSpinor Uchi;
+						SiteSpinor *buf, int sF, int sU, 
+						const FermionFieldView &in, FermionFieldView &out,int dag)
+{
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
-  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+s;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=LLs*sU+s;
+  {
     skew = 0;
     Uchi=Zero();
     GENERIC_STENCIL_LEG_INT(U,Xp,skew,Impl::multLinkAdd);
@@ -143,6 +149,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_INT(U,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(U,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(U,Tm,skew,Impl::multLinkAdd);
+    if ( Naik ) {
     skew=8;
     GENERIC_STENCIL_LEG_INT(UUU,Xp,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(UUU,Yp,skew,Impl::multLinkAdd);
@@ -152,10 +159,11 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_INT(UUU,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_INT(UUU,Tm,skew,Impl::multLinkAdd);
+    }
     if ( dag ) {
       Uchi = - Uchi;
     }
-    vstream(out[sF], Uchi);
+    coalescedWrite(out[sF], Uchi,lane);
   }
 };
 
@@ -164,20 +172,25 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
   // Only contributions from exterior of our node
   ///////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+template <int Naik> accelerator_inline
+void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st, 
 						DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-						SiteSpinor *buf, int LLs, int sU,
-						const FermionFieldView &in, FermionFieldView &out,int dag) {
-  const SiteSpinor *chi_p;
-  //  SiteSpinor chi;
-  SiteSpinor Uchi;
+						SiteSpinor *buf, int sF, int sU,
+						const FermionFieldView &in, FermionFieldView &out,int dag)
+{
+  typedef decltype(coalescedRead(in[0])) calcSpinor;
+  calcSpinor chi;
+  calcSpinor Uchi;
   StencilEntry *SE;
   int ptype;
   int nmu=0;
   int skew ;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
 
-  for(int s=0;s<LLs;s++){
-    int sF=LLs*sU+s;
+  //  for(int s=0;s<LLs;s++){
+  //    int sF=LLs*sU+s;
+  {
     skew = 0;
     Uchi=Zero();
     GENERIC_STENCIL_LEG_EXT(U,Xp,skew,Impl::multLinkAdd);
@@ -188,6 +201,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_EXT(U,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(U,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(U,Tm,skew,Impl::multLinkAdd);
+    if ( Naik ) {
     skew=8;
     GENERIC_STENCIL_LEG_EXT(UUU,Xp,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Yp,skew,Impl::multLinkAdd);
@@ -197,12 +211,13 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
     GENERIC_STENCIL_LEG_EXT(UUU,Ym,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
     GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
-
-    if ( nmu ) { 
-      if ( dag ) { 
-	out[sF] = out[sF] - Uchi;
+    }
+    if ( nmu ) {
+      auto _out = coalescedRead(out[sF],lane);
+      if ( dag ) {
+	coalescedWrite(out[sF], _out-Uchi,lane);
       } else { 
-	out[sF] = out[sF] + Uchi;
+	coalescedWrite(out[sF], _out+Uchi,lane);
       }
     }
   }
@@ -211,72 +226,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
 ////////////////////////////////////////////////////////////////////////////////////
 // Driving / wrapping routine to select right kernel
 ////////////////////////////////////////////////////////////////////////////////////
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-					 SiteSpinor *buf, int LLs, int sU,
-					 const FermionFieldView &in, FermionFieldView &out,
-					 int interior,int exterior)
-{
-  int dag=1;
-  DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
-};
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-				      SiteSpinor *buf, int LLs, int sU,
-				      const FermionFieldView &in, FermionFieldView &out,
-				      int interior,int exterior)
-{
-  int dag=0;
-  DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
-};
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
-				      SiteSpinor *buf, int LLs,
-				      int sU, const FermionFieldView &in, FermionFieldView &out,
-				      int dag,int interior,int exterior) 
-{
-  switch(Opt) {
-#ifdef AVX512
-  case OptInlineAsm:
-    if ( interior && exterior ) {
-      DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else { 
-      std::cout << GridLogError << "Cannot overlap comms and compute with Staggered assembly"<<std::endl;
-      assert(0);
-    }
-    break;
-#endif
-  case OptHandUnroll:
-    if ( interior && exterior ) {
-      DhopSiteHand   (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( interior ) {
-      DhopSiteHandInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( exterior ) {
-      DhopSiteHandExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    }
-    break;
-  case OptGeneric:
-    if ( interior && exterior ) {
-      DhopSiteGeneric   (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( interior ) {
-      DhopSiteGenericInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    } else if ( exterior ) {
-      DhopSiteGenericExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    }
-    break;
-  default:
-    std::cout<<"Oops Opt = "<<Opt<<std::endl;
-    assert(0);
-    break;
-  }
-};
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeFieldView &U,  DoubledGaugeFieldView &UUU, SiteSpinor *buf, int sF,
-					    int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int disp) 
+template <class Impl> 
+void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+					   int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp)
 {
   // Disp should be either +1,-1,+3,-3
   // What about "dag" ?
@@ -285,6 +237,112 @@ void StaggeredKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeFieldVi
   assert(0);
 }
 
+#define KERNEL_CALLNB(A,improved)					\
+  const uint64_t    NN = Nsite*Ls;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      ThisKernel:: template A<improved>(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag); \
+    });
+
+#define KERNEL_CALL(A,improved) KERNEL_CALLNB(A,improved); accelerator_barrier(); 
+
+#define ASM_CALL(A)							\
+  const uint64_t    NN = Nsite*Ls;					\
+  thread_for( ss, NN, {							\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      ThisKernel::A(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag);		\
+  });
+
+template <class Impl> 
+void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
+					  DoubledGaugeField &U, DoubledGaugeField &UUU, 
+					  const FermionField &in, FermionField &out, int dag, int interior,int exterior)
+{
+  GridBase *FGrid=in.Grid();  
+  GridBase *UGrid=U.Grid();  
+  typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  autoView( UUU_v , UUU, AcceleratorRead);
+  autoView( U_v   ,   U, AcceleratorRead);
+  autoView( in_v  ,  in, AcceleratorRead);
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( st_v  ,  st, AcceleratorRead);
+  SiteSpinor * buf = st.CommBuf();
+    
+  int Ls=1;
+  if(FGrid->Nd()==UGrid->Nd()+1){
+    Ls    = FGrid->_rdimensions[0];
+  }
+  int Nsite = UGrid->oSites();
+
+  if( interior && exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGeneric,1); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,1);    return;}
+    if (Opt == OptInlineAsm  ) {  ASM_CALL(DhopSiteAsm);     return;}
+#endif
+  } else if( interior ) {
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericInt,1); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,1);    return;}
+#endif
+  } else if( exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericExt,1); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,1);    return;}
+#endif
+  }
+  assert(0 && " Kernel optimisation case not covered ");
+}
+template <class Impl> 
+void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
+				       DoubledGaugeField &U,
+				       const FermionField &in, FermionField &out, int dag, int interior,int exterior)
+{
+  GridBase *FGrid=in.Grid();  
+  GridBase *UGrid=U.Grid();  
+  typedef StaggeredKernels<Impl> ThisKernel;
+  const int Nsimd = SiteHalfSpinor::Nsimd();
+  const int lane=acceleratorSIMTlane(Nsimd);
+  autoView( UUU_v ,   U, AcceleratorRead);
+  autoView( U_v   ,   U, AcceleratorRead);
+  autoView( in_v  ,  in, AcceleratorRead);
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( st_v  ,  st, AcceleratorRead);
+  SiteSpinor * buf = st.CommBuf();
+
+  int Ls=1;
+  if(FGrid->Nd()==UGrid->Nd()+1){
+    Ls    = FGrid->_rdimensions[0];
+  }
+  int Nsite = UGrid->oSites();
+  
+  if( interior && exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGeneric,0); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,0);    return;}
+#endif
+  } else if( interior ) {
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericInt,0); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,0);    return;}
+#endif
+  } else if( exterior ) { 
+    if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericExt,0); return;}
+#ifndef GRID_CUDA
+    if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,0);    return;}
+#endif
+  }
+}
+
+
+#undef KERNEL_CALLNB
+#undef KERNEL_CALL
+#undef ASM_CALL
+
 NAMESPACE_END(Grid);
 
 

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

@@ -92,52 +92,51 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   int lvol = _Umu.Grid()->lSites();
   int DimRep = Impl::Dimension;
 
-  Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-  Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-
-  Coordinate lcoor;
-  typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
-
-  for (int site = 0; site < lvol; site++)
   {
-    grid->LocalIndexToLocalCoor(site, lcoor);
-    EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-    peekLocalSite(Qx, CloverTerm, lcoor);
-    Qxinv = Zero();
-    //if (csw!=0){
-    for (int j = 0; j < Ns; j++)
-      for (int k = 0; k < Ns; k++)
-        for (int a = 0; a < DimRep; a++)
-          for (int b = 0; b < DimRep; b++){
-	    auto zz =  Qx()(j, k)(a, b);
-            EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
-	  }
-    //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
-
-    EigenInvCloverOp = EigenCloverOp.inverse();
-    //std::cout << EigenInvCloverOp << std::endl;
-    for (int j = 0; j < Ns; j++)
-      for (int k = 0; k < Ns; k++)
-        for (int a = 0; a < DimRep; a++)
-          for (int b = 0; b < DimRep; b++)
-            Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-    //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
-    //  }
-    pokeLocalSite(Qxinv, CloverTermInv, lcoor);
+    autoView(CTv,CloverTerm,CpuRead);
+    autoView(CTIv,CloverTermInv,CpuWrite);
+    thread_for(site, lvol, {
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(site, lcoor);
+      Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
+      peekLocalSite(Qx, CTv, lcoor);
+      //if (csw!=0){
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++){
+	      auto zz =  Qx()(j, k)(a, b);
+	      EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+	    }
+      //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
+      
+      EigenInvCloverOp = EigenCloverOp.inverse();
+      //std::cout << EigenInvCloverOp << std::endl;
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++)
+	      Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+      //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
+      //  }
+      pokeLocalSite(Qxinv, CTIv, lcoor);
+    });
   }
 
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
 
-  pickCheckerboard(Even, CloverTermDagEven, closure(adj(CloverTerm)));
-  pickCheckerboard(Odd, CloverTermDagOdd, closure(adj(CloverTerm)));
+  pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
+  pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
 
   pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
   pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
 
-  pickCheckerboard(Even, CloverTermInvDagEven, closure(adj(CloverTermInv)));
-  pickCheckerboard(Odd, CloverTermInvDagOdd, closure(adj(CloverTermInv)));
+  pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
+  pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
 }
 
 template <class Impl>