Zero() change

Conflicts:
	Grid/qcd/action/fermion/WilsonKernelsAsm.cc
	Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h
	Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
	benchmarks/Benchmark_comms.cc

.gitignore

@@ -114,3 +114,4 @@ gh-pages/
 #####################
 Grid/qcd/spin/gamma-gen/*.h
 Grid/qcd/spin/gamma-gen/*.cc
+Grid/util/Version.h

ChangeLog

@@ -0,0 +1,5 @@
+Version : 0.8.0
+
+- Clang 3.5 and above, ICPC v16 and above, GCC 6.3 and above recommended
+- MPI and MPI3 comms optimisations for KNL and OPA finished
+- Half precision comms

Grid/DisableWarnings.h

@@ -30,8 +30,34 @@ directory
 #ifndef DISABLE_WARNINGS_H
 #define DISABLE_WARNINGS_H
 
+
+
+#if defined __GNUC__ && __GNUC__>=6
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
  //disables and intel compiler specific warning (in json.hpp)
 #pragma warning disable 488  
 
+#ifdef __NVCC__
+ //disables nvcc specific warning in json.hpp
+#pragma clang diagnostic ignored "-Wdeprecated-register"
+#pragma diag_suppress unsigned_compare_with_zero
+#pragma diag_suppress cast_to_qualified_type
+
+ //disables nvcc specific warning in many files
+#pragma diag_suppress esa_on_defaulted_function_ignored
+#pragma diag_suppress extra_semicolon
+
+//Eigen only
+#endif
+
+// Disable vectorisation in Eigen on the Power8/9 and PowerPC
+#ifdef  __ALTIVEC__
+#define  EIGEN_DONT_VECTORIZE
+#endif
+#ifdef  __VSX__
+#define  EIGEN_DONT_VECTORIZE
+#endif
 
 #endif

Grid/Grid.h

@@ -42,6 +42,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridQCDcore.h>
 #include <Grid/qcd/action/Action.h>
 #include <Grid/qcd/utils/GaugeFix.h>
+#include <Grid/qcd/utils/CovariantSmearing.h>
 #include <Grid/qcd/smearing/Smearing.h>
 #include <Grid/parallelIO/MetaData.h>
 #include <Grid/qcd/hmc/HMC_aggregate.h>

Grid/GridCore.h

@@ -38,16 +38,19 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_BASE_H
 #define GRID_BASE_H
 
-#include <Grid/GridStd.h>
 
+#include <Grid/DisableWarnings.h>
+#include <Grid/Namespace.h>
+#include <Grid/GridStd.h>
+#include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
 #include <Grid/perfmon/PerfCount.h>
+#include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
 #include <Grid/allocator/AlignedAllocator.h>
 #include <Grid/simd/Simd.h>
-#include <Grid/serialisation/Serialisation.h>
 #include <Grid/threads/Threads.h>
-#include <Grid/util/Util.h>
+#include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 
 #include <Grid/cartesian/Cartesian.h>    
@@ -57,5 +60,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/stencil/Stencil.h>      
 #include <Grid/parallelIO/BinaryIO.h>
 #include <Grid/algorithms/Algorithms.h>   
+NAMESPACE_CHECK(GridCore)
 
 #endif

Grid/GridQCDcore.h

@@ -38,5 +38,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/qcd/spin/Spin.h>
 #include <Grid/qcd/utils/Utils.h>
 #include <Grid/qcd/representations/Representations.h>
+NAMESPACE_CHECK(GridQCDCore);
 
 #endif

Grid/GridStd.h

@@ -7,6 +7,7 @@
 #include <cassert>
 #include <complex>
 #include <vector>
+#include <array>
 #include <string>
 #include <iostream>
 #include <iomanip>

Grid/Grid_Eigen_Dense.h

@@ -1,14 +1,41 @@
+#include <Grid/GridCore.h>
 #pragma once
 // Force Eigen to use MKL if Grid has been configured with --enable-mkl
 #ifdef USE_MKL
 #define EIGEN_USE_MKL_ALL
 #endif
 
+
 #if defined __GNUC__
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
+
+/* NVCC save and restore compile environment*/
+#ifdef __NVCC__
+#pragma push
+#pragma diag_suppress code_is_unreachable
+#pragma push_macro("__CUDA_ARCH__")
+#pragma push_macro("__NVCC__")
+#pragma push_macro("__CUDACC__")
+#undef __NVCC__
+#undef __CUDACC__
+#undef __CUDA_ARCH__
+#define __NVCC__REDEFINE__
+#endif
+
 #include <Grid/Eigen/Dense>
+#include <Grid/Eigen/unsupported/CXX11/Tensor>
+
+/* NVCC restore */
+#ifdef __NVCC__REDEFINE__
+#pragma pop_macro("__CUDACC__")
+#pragma pop_macro("__NVCC__")
+#pragma pop_macro("__CUDA_ARCH__")
+#pragma pop
+#endif
+
 #if defined __GNUC__
 #pragma GCC diagnostic pop
 #endif
+

Grid/Grid_Eigen_Tensor.h

@@ -0,0 +1 @@
+#include <Grid/Grid_Eigen_Dense.h>

Grid/Namespace.h

@@ -0,0 +1,38 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/Namespace.h
+
+Copyright (C) 2016
+
+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 <type_traits>
+#include <cassert>
+
+#define NAMESPACE_BEGIN(A) namespace A {
+#define NAMESPACE_END(A)   }
+#define GRID_NAMESPACE_BEGIN NAMESPACE_BEGIN(Grid)
+#define GRID_NAMESPACE_END   NAMESPACE_END(Grid)
+#define NAMESPACE_CHECK(x) struct namespaceTEST##x {};  static_assert(std::is_same<namespaceTEST##x, ::namespaceTEST##x>::value,"Not in :: at"  ); 

Grid/algorithms/Algorithms.h

@@ -48,15 +48,16 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
+#include <Grid/algorithms/iterative/MinimalResidual.h>
+#include <Grid/algorithms/iterative/GeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
+#include <Grid/algorithms/iterative/PowerMethod.h>
+
 #include <Grid/algorithms/CoarsenedMatrix.h>
 #include <Grid/algorithms/FFT.h>
 
-
-// EigCg
-// Pcg
-// Hdcg
-// GCR
-// etc..
-
 #endif

Grid/algorithms/CoarsenedMatrix.h

@@ -32,7 +32,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define  GRID_ALGORITHM_COARSENED_MATRIX_H
 
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 class Geometry {
   //    int dimension;
@@ -104,7 +104,7 @@ namespace Grid {
   GridBase *FineGrid;
   std::vector<Lattice<Fobj> > subspace;
   int checkerboard;
-
+  int Checkerboard(void){return checkerboard;}
   Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
     CoarseGrid(_CoarseGrid),
     FineGrid(_FineGrid),
@@ -127,10 +127,10 @@ namespace Grid {
     CoarseVector eProj(CoarseGrid); 
     for(int i=0;i<nbasis;i++){
       blockProject(iProj,subspace[i],subspace);
-	eProj=zero; 
-	parallel_for(int ss=0;ss<CoarseGrid->oSites();ss++){
-	  eProj._odata[ss](i)=CComplex(1.0);
-	}
+      eProj=Zero(); 
+      thread_for(ss, CoarseGrid->oSites(),{
+	eProj[ss](i)=CComplex(1.0);
+      });
       eProj=eProj - iProj;
       std::cout<<GridLogMessage<<"Orthog check error "<<i<<" " << norm2(eProj)<<std::endl;
     }
@@ -140,7 +140,7 @@ namespace Grid {
     blockProject(CoarseVec,FineVec,subspace);
   }
   void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
-      FineVec.checkerboard = subspace[0].checkerboard;
+    FineVec.Checkerboard() = subspace[0].Checkerboard();
     blockPromote(CoarseVec,FineVec,subspace);
   }
   void CreateSubspaceRandom(GridParallelRNG &RNG){
@@ -211,6 +211,7 @@ namespace Grid {
 
     for(int b=0;b<nn;b++){
 	
+      subspace[b] = Zero();
       gaussian(RNG,noise);
       scale = std::pow(norm2(noise),-0.5); 
       noise=noise*scale;
@@ -254,7 +255,8 @@ namespace Grid {
   ////////////////////
   Geometry         geom;
   GridBase *       _grid; 
-    CartesianStencil<siteVector,siteVector> Stencil; 
+
+  CartesianStencil<siteVector,siteVector,int> Stencil; 
 
   std::vector<CoarseMatrix> A;
 
@@ -266,14 +268,15 @@ namespace Grid {
 
   RealD M (const CoarseVector &in, CoarseVector &out){
 
-      conformable(_grid,in._grid);
-      conformable(in._grid,out._grid);
+    conformable(_grid,in.Grid());
+    conformable(in.Grid(),out.Grid());
 
     SimpleCompressor<siteVector> compressor;
     Stencil.HaloExchange(in,compressor);
-
-      parallel_for(int ss=0;ss<Grid()->oSites();ss++){
-        siteVector res = zero;
+    auto in_v = in.View();
+    auto out_v = in.View();
+    thread_for(ss,Grid()->oSites(),{
+      siteVector res = Zero();
       siteVector nbr;
       int ptype;
       StencilEntry *SE;
@@ -282,32 +285,82 @@ namespace Grid {
 	SE=Stencil.GetEntry(ptype,point,ss);
 	  
 	if(SE->_is_local&&SE->_permute) { 
-	    permute(nbr,in._odata[SE->_offset],ptype);
+	  permute(nbr,in_v[SE->_offset],ptype);
 	} else if(SE->_is_local) { 
-	    nbr = in._odata[SE->_offset];
+	  nbr = in_v[SE->_offset];
 	} else {
 	  nbr = Stencil.CommBuf()[SE->_offset];
 	}
-	  res = res + A[point]._odata[ss]*nbr;
-	}
-	vstream(out._odata[ss],res);
+	auto A_point = A[point].View();
+	res = res + A_point[ss]*nbr;
       }
+      vstream(out_v[ss],res);
+    });
     return norm2(out);
   };
 
   RealD Mdag (const CoarseVector &in, CoarseVector &out){
-      return M(in,out);
+    // // corresponds to Petrov-Galerkin coarsening
+    // return M(in,out);
+    
+    // corresponds to Galerkin coarsening
+    CoarseVector tmp(Grid());
+    G5C(tmp, in);
+    M(tmp, out);
+    G5C(out, out);
+    return norm2(out);
+  };
+
+  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
+    
+    conformable(_grid,in.Grid());
+    conformable(in.Grid(),out.Grid());
+    
+    SimpleCompressor<siteVector> compressor;
+    Stencil.HaloExchange(in,compressor);
+    
+    auto point = [dir, disp](){
+      if(dir == 0 and disp == 0)
+	return 8;
+      else
+	return (4 * dir + 1 - disp) / 2;
+    }();
+
+    auto out_v = out.View();
+    auto in_v  = in.View();
+    thread_for(ss,Grid()->oSites(),{
+      siteVector res = Zero();
+      siteVector nbr;
+      int ptype;
+      StencilEntry *SE;
+      
+      SE=Stencil.GetEntry(ptype,point,ss);
+      
+      if(SE->_is_local&&SE->_permute) {
+	permute(nbr,in_v[SE->_offset],ptype);
+      } else if(SE->_is_local) {
+	nbr = in_v[SE->_offset];
+      } else {
+	nbr = Stencil.CommBuf()[SE->_offset];
+      }
+
+      auto A_point = A[point].View();
+      res = res + A_point[ss]*nbr;
+      
+      vstream(out_v[ss],res);
+    });
+  };
+
+  void Mdiag(const CoarseVector &in, CoarseVector &out){
+    Mdir(in, out, 0, 0); // use the self coupling (= last) point of the stencil
   };
 
-    // Defer support for further coarsening for now
-    void Mdiag    (const CoarseVector &in,  CoarseVector &out){};
-    void Mdir     (const CoarseVector &in,  CoarseVector &out,int dir, int disp){};
   
  CoarsenedMatrix(GridCartesian &CoarseGrid) 	: 
 
     _grid(&CoarseGrid),
     geom(CoarseGrid._ndimension),
-      Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
     A(geom.npoint,&CoarseGrid)
   {
   };
@@ -320,7 +373,7 @@ namespace Grid {
 
     FineField     phi(FineGrid);
     FineField     tmp(FineGrid);
-      FineField     zz(FineGrid); zz=zero;
+    FineField     zz(FineGrid); zz=Zero();
     FineField    Mphi(FineGrid);
 
     Lattice<iScalar<vInteger> > coor(FineGrid);
@@ -336,7 +389,7 @@ namespace Grid {
     // set of vectors.
     int self_stencil=-1;
     for(int p=0;p<geom.npoint;p++){ 
-	A[p]=zero;
+      A[p]=Zero();
       if( geom.displacements[p]==0){
 	self_stencil=p;
       }
@@ -369,7 +422,7 @@ namespace Grid {
 	////////////////////////////////////////////////////////////////////////
 	if ( disp==0 ) {
 	  iblock = Mphi;
-	    oblock = zero;
+	  oblock = Zero();
 	} else if ( disp==1 ) {
 	  oblock = where(mod(coor,block)==(block-1),Mphi,zz);
 	  iblock = where(mod(coor,block)!=(block-1),Mphi,zz);
@@ -384,14 +437,18 @@ namespace Grid {
 	Subspace.ProjectToSubspace(oProj,oblock);
 	//	  blockProject(iProj,iblock,Subspace.subspace);
 	//	  blockProject(oProj,oblock,Subspace.subspace);
-	  parallel_for(int ss=0;ss<Grid()->oSites();ss++){
+	auto iProj_v = iProj.View() ;
+	auto oProj_v = oProj.View() ;
+	auto A_p     =  A[p].View();
+	auto A_self  = A[self_stencil].View();
+	thread_for(ss, Grid()->oSites(),{
 	  for(int j=0;j<nbasis;j++){
 	    if( disp!= 0 ) {
-		A[p]._odata[ss](j,i) = oProj._odata[ss](j);
-	      }
-	      A[self_stencil]._odata[ss](j,i) =	A[self_stencil]._odata[ss](j,i) + iProj._odata[ss](j);
+	      A_p[ss](j,i) = oProj_v[ss](j);
 	    }
+	    A_self[ss](j,i) =	A_self[ss](j,i) + iProj_v[ss](j);
 	  }
+	});
       }
     }
 
@@ -417,35 +474,10 @@ namespace Grid {
     std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
 #endif
       //      ForceHermitian();
-      AssertHermitian();
+      // AssertHermitian();
       // ForceDiagonal();
   }
-    void ForceDiagonal(void) {
 
-
-      std::cout<<GridLogMessage<<"**************************************************"<<std::endl;
-      std::cout<<GridLogMessage<<"****   Forcing coarse operator to be diagonal ****"<<std::endl;
-      std::cout<<GridLogMessage<<"**************************************************"<<std::endl;
-      for(int p=0;p<8;p++){
-	A[p]=zero;
-      }
-
-      GridParallelRNG  RNG(Grid()); RNG.SeedFixedIntegers(std::vector<int>({55,72,19,17,34}));
-      Lattice<iScalar<CComplex> > val(Grid()); random(RNG,val);
-
-      Complex one(1.0);
-
-      iMatrix<CComplex,nbasis> ident;  ident=one;
-
-      val = val*adj(val);
-      val = val + 1.0;
-
-      A[8] = val*ident;
-
-      //      for(int s=0;s<Grid()->oSites();s++) {
-      //	A[8]._odata[s]=val._odata[s];
-      //      }
-    }
   void ForceHermitian(void) {
     for(int d=0;d<4;d++){
       int dd=d+1;
@@ -476,5 +508,5 @@ namespace Grid {
     
 };
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/FFT.h

@@ -38,7 +38,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif
 
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class scalar> struct FFTW { };
 
@@ -115,9 +115,9 @@ namespace Grid {
   double flops_call;
   uint64_t usec;
     
-    std::vector<int> dimensions;
-    std::vector<int> processors;
-    std::vector<int> processor_coor;
+  Coordinate dimensions;
+  Coordinate processors;
+  Coordinate processor_coor;
     
 public:
     
@@ -137,7 +137,7 @@ namespace Grid {
   {
     flops=0;
     usec =0;
-      std::vector<int> layout(Nd,1);
+    Coordinate layout(Nd,1);
     sgrid = new GridCartesian(dimensions,layout,processors);
   };
     
@@ -146,10 +146,10 @@ namespace Grid {
   }
     
   template<class vobj>
-    void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,std::vector<int> mask,int sign){
+  void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
 
-      conformable(result._grid,vgrid);
-      conformable(source._grid,vgrid);
+    conformable(result.Grid(),vgrid);
+    conformable(source.Grid(),vgrid);
     Lattice<vobj> tmp(vgrid);
     tmp = source;
     for(int d=0;d<Nd;d++){
@@ -162,7 +162,7 @@ namespace Grid {
 
   template<class vobj>
   void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
-      std::vector<int> mask(Nd,1);
+    Coordinate mask(Nd,1);
     FFT_dim_mask(result,source,mask,sign);
   }
 
@@ -172,14 +172,14 @@ namespace Grid {
 #ifndef HAVE_FFTW
     assert(0);
 #else
-      conformable(result._grid,vgrid);
-      conformable(source._grid,vgrid);
+    conformable(result.Grid(),vgrid);
+    conformable(source.Grid(),vgrid);
 
     int L = vgrid->_ldimensions[dim];
     int G = vgrid->_fdimensions[dim];
       
-      std::vector<int> layout(Nd,1);
-      std::vector<int> pencil_gd(vgrid->_fdimensions);
+    Coordinate layout(Nd,1);
+    Coordinate pencil_gd(vgrid->_fdimensions);
       
     pencil_gd[dim] = G*processors[dim];
       
@@ -191,7 +191,7 @@ namespace Grid {
     typedef typename sobj::scalar_type   scalar;
       
     Lattice<sobj> pgbuf(&pencil_g);
-      
+    auto pgbuf_v = pgbuf.View();
 
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -217,8 +217,8 @@ namespace Grid {
       
     FFTW_plan p;
     {
-        FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[0];
-        FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[0];
+      FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
+      FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[0];
       p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
 					   in,inembed,
 					   istride,idist,
@@ -228,26 +228,20 @@ namespace Grid {
     }
       
     // Barrel shift and collect global pencil
-      std::vector<int> lcoor(Nd), gcoor(Nd);
+    Coordinate lcoor(Nd), gcoor(Nd);
     result = source;
     int pc = processor_coor[dim];
     for(int p=0;p<processors[dim];p++) {
-        PARALLEL_REGION
-        {
-          std::vector<int> cbuf(Nd);
+      thread_for(idx, sgrid->lSites(),{
+          Coordinate cbuf(Nd);
           sobj s;
-          
-          PARALLEL_FOR_LOOP_INTERN
-          for(int idx=0;idx<sgrid->lSites();idx++) {
 	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
 	  peekLocalSite(s,result,cbuf);
 	  cbuf[dim]+=((pc+p) % processors[dim])*L;
 	  //            cbuf[dim]+=p*L;
 	  pokeLocalSite(s,pgbuf,cbuf);
-          }
-        }
-        if (p != processors[dim] - 1)
-        {
+      });
+      if (p != processors[dim] - 1) {
 	result = Cshift(result,dim,L);
       }
     }
@@ -256,20 +250,15 @@ namespace Grid {
     int NN=pencil_g.lSites();
     GridStopWatch timer;
     timer.Start();
-      PARALLEL_REGION
-      {
-        std::vector<int> cbuf(Nd);
-        
-        PARALLEL_FOR_LOOP_INTERN
-        for(int idx=0;idx<NN;idx++) {
+    thread_for( idx,NN,{
+        Coordinate cbuf(Nd);
 	pencil_g.LocalIndexToLocalCoor(idx, cbuf);
 	if ( cbuf[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
-            FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[idx];
-            FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[idx];
+	  FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[idx];
+	  FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[idx];
 	  FFTW<scalar>::fftw_execute_dft(p,in,out);
 	}
-        }
-      }
+    });
     timer.Stop();
       
     // performance counting
@@ -280,20 +269,15 @@ namespace Grid {
     flops+= flops_call*NN;
       
     // writing out result
-      PARALLEL_REGION
-      {
-        std::vector<int> clbuf(Nd), cgbuf(Nd);
+    thread_for(idx,sgrid->lSites(),{
+	Coordinate clbuf(Nd), cgbuf(Nd);
 	sobj s;
-        
-        PARALLEL_FOR_LOOP_INTERN
-        for(int idx=0;idx<sgrid->lSites();idx++) {
 	sgrid->LocalIndexToLocalCoor(idx,clbuf);
 	cgbuf = clbuf;
 	cgbuf[dim] = clbuf[dim]+L*pc;
 	peekLocalSite(s,pgbuf,cgbuf);
 	pokeLocalSite(s,result,clbuf);
-        }
-      }
+    });
     result = result*div;
       
     // destroying plan
@@ -301,6 +285,7 @@ namespace Grid {
 #endif
   }
 };
-}
+
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/LinearOperator.h

@@ -26,16 +26,15 @@ 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_ALGORITHM_LINEAR_OP_H
-#define  GRID_ALGORITHM_LINEAR_OP_H
+#pragma once 
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////////////////////////////////////
 // LinearOperators Take a something and return a something.
 /////////////////////////////////////////////////////////////////////////////////////////////
 //
-  // Hopefully linearity is satisfied and the AdjOp is indeed the Hermitian conjugateugate (transpose if real):
+// Hopefully linearity is satisfied and the AdjOp is indeed the Hermitian Conjugateugate (transpose if real):
 //SBase
 //   i)  F(a x + b y) = aF(x) + b F(y).
 //  ii)  <x|Op|y> = <y|AdjOp|x>^\ast
@@ -183,13 +182,13 @@ namespace Grid {
       virtual  RealD Mpc      (const Field &in, Field &out) =0;
       virtual  RealD MpcDag   (const Field &in, Field &out) =0;
       virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
-      Field tmp(in._grid);
-      tmp.checkerboard = in.checkerboard;
+      Field tmp(in.Grid());
+      tmp.Checkerboard() = in.Checkerboard();
 	ni=Mpc(in,tmp);
 	no=MpcDag(tmp,out);
       }
       virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-      out.checkerboard = in.checkerboard;
+      out.Checkerboard() = in.Checkerboard();
 	MpcDagMpc(in,out,n1,n2);
       }
       virtual void HermOp(const Field &in, Field &out){
@@ -212,25 +211,24 @@ namespace Grid {
     };
     template<class Matrix,class Field>
     class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
-    protected:
-      Matrix &_Mat;
     public:
+      Matrix &_Mat;
       SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
       virtual  RealD Mpc      (const Field &in, Field &out) {
-      Field tmp(in._grid);
-      tmp.checkerboard = !in.checkerboard;
+      Field tmp(in.Grid());
+      tmp.Checkerboard() = !in.Checkerboard();
 	//std::cout <<"grid pointers: in._grid="<< in._grid << " out._grid=" << out._grid << "  _Mat.Grid=" << _Mat.Grid() << " _Mat.RedBlackGrid=" << _Mat.RedBlackGrid() << std::endl;
 
 	_Mat.Meooe(in,tmp);
 	_Mat.MooeeInv(tmp,out);
 	_Mat.Meooe(out,tmp);
 
-      //std::cout << "cb in " << in.checkerboard << "  cb out " << out.checkerboard << std::endl;
+      //std::cout << "cb in " << in.Checkerboard() << "  cb out " << out.Checkerboard() << std::endl;
 	_Mat.Mooee(in,out);
 	return axpy_norm(out,-1.0,tmp,out);
       }
       virtual  RealD MpcDag   (const Field &in, Field &out){
-	Field tmp(in._grid);
+	Field tmp(in.Grid());
 
 	_Mat.MeooeDag(in,tmp);
         _Mat.MooeeInvDag(tmp,out);
@@ -248,7 +246,7 @@ namespace Grid {
       SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
 
       virtual  RealD Mpc      (const Field &in, Field &out) {
-	Field tmp(in._grid);
+	Field tmp(in.Grid());
 
 	_Mat.Meooe(in,out);
 	_Mat.MooeeInv(out,tmp);
@@ -258,7 +256,7 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
       virtual  RealD MpcDag   (const Field &in, Field &out){
-	Field tmp(in._grid);
+	Field tmp(in.Grid());
 
 	_Mat.MooeeInvDag(in,out);
 	_Mat.MeooeDag(out,tmp);
@@ -276,7 +274,7 @@ namespace Grid {
       SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
 
       virtual  RealD Mpc      (const Field &in, Field &out) {
-	Field tmp(in._grid);
+	Field tmp(in.Grid());
 
 	_Mat.MooeeInv(in,out);
 	_Mat.Meooe(out,tmp);
@@ -286,7 +284,7 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
       virtual  RealD MpcDag   (const Field &in, Field &out){
-	Field tmp(in._grid);
+	Field tmp(in.Grid());
 
 	_Mat.MeooeDag(in,out);
 	_Mat.MooeeInvDag(out,tmp);
@@ -354,7 +352,17 @@ namespace Grid {
 	axpby(out,-1.0,mass*mass,tmp,in);
 	taxpby_norm+=usecond();
   }
-      virtual  RealD Mpc      (const Field &in, Field &out) {
+  virtual  RealD Mpc      (const Field &in, Field &out) 
+  {
+
+    Field tmp(in.Grid());
+    Field tmp2(in.Grid());
+
+    //    std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
+    _Mat.Mooee(in,out);
+    _Mat.Mooee(out,tmp);
+    //    std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
+
     tMeo-=usecond();
     _Mat.Meooe(in,out);
     _Mat.Meooe(out,tmp);
@@ -380,6 +388,12 @@ namespace Grid {
 template<class Field> class OperatorFunction {
 public:
   virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
+  virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
+    assert(in.size()==out.size());
+    for(int k=0;k<in.size();k++){
+      (*this)(Linop,in[k],out[k]);
+    }
+  };
 };
 
 template<class Field> class LinearFunction {
@@ -459,13 +473,15 @@ namespace Grid {
 private:
   std::vector<RealD> Coeffs;
 public:
+  using OperatorFunction<Field>::operator();
+
   Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
 
   // Implement the required interface
   void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 
-      Field AtoN(in._grid);
-      Field Mtmp(in._grid);
+    Field AtoN(in.Grid());
+    Field Mtmp(in.Grid());
     AtoN = in;
     out = AtoN*Coeffs[0];
     for(int n=1;n<Coeffs.size();n++){
@@ -476,6 +492,4 @@ namespace Grid {
   };
 };
 
-}
-
-#endif
+NAMESPACE_END(Grid);

Grid/algorithms/Preconditioner.h

@@ -28,7 +28,7 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 #ifndef GRID_PRECONDITIONER_H
 #define GRID_PRECONDITIONER_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class Field> class Preconditioner :  public LinearFunction<Field> { 
   virtual void operator()(const Field &src, Field & psi)=0;
@@ -42,5 +42,5 @@ namespace Grid {
   TrivialPrecon(void){};
 };
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/SparseMatrix.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #define  GRID_ALGORITHM_SPARSE_MATRIX_H
 
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////////////////////////////////////
 // Interface defining what I expect of a general sparse matrix, such as a Fermion action
@@ -41,7 +41,7 @@ namespace Grid {
   virtual RealD M    (const Field &in, Field &out)=0;
   virtual RealD Mdag (const Field &in, Field &out)=0;
   virtual void  MdagM(const Field &in, Field &out,RealD &ni,RealD &no) {
-	Field tmp (in._grid);
+    Field tmp (in.Grid());
     ni=M(in,tmp);
     no=Mdag(tmp,out);
   }
@@ -55,6 +55,14 @@ namespace Grid {
 template<class Field> class CheckerBoardedSparseMatrixBase : public SparseMatrixBase<Field> {
 public:
   virtual GridBase *RedBlackGrid(void)=0;
+
+      //////////////////////////////////////////////////////////////////////
+      // Query the even even properties to make algorithmic decisions
+      //////////////////////////////////////////////////////////////////////
+      virtual RealD  Mass(void)        { return 0.0; };
+      virtual int    ConstEE(void)     { return 1; }; // Disable assumptions unless overridden
+      virtual int    isTrivialEE(void) { return 0; }; // by a derived class that knows better
+
   // half checkerboard operaions
   virtual  void Meooe    (const Field &in, Field &out)=0;
   virtual  void Mooee    (const Field &in, Field &out)=0;
@@ -66,6 +74,6 @@ namespace Grid {
 
 };
 
-}
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/approx/Chebyshev.h

@@ -32,7 +32,7 @@ Author: Christoph Lehner <clehner@bnl.gov>
 
 #include <Grid/algorithms/LinearOperator.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 struct ChebyParams : Serializable {
   GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
@@ -47,6 +47,8 @@ struct ChebyParams : Serializable {
 template<class Field>
 class Chebyshev : public OperatorFunction<Field> {
 private:
+  using OperatorFunction<Field>::operator();
+
   std::vector<RealD> Coeffs;
   int order;
   RealD hi;
@@ -55,7 +57,7 @@ struct ChebyParams : Serializable {
 public:
   void csv(std::ostream &out){
     RealD diff = hi-lo;
-      RealD delta = (hi-lo)*1.0e-9;
+    RealD delta = diff*1.0e-9;
     for (RealD x=lo; x<hi; x+=delta) {
       delta*=1.1;
       RealD f = approx(x);
@@ -212,9 +214,9 @@ struct ChebyParams : Serializable {
   // Implement the required interface
   void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 
-      GridBase *grid=in._grid;
+    GridBase *grid=in.Grid();
 
-      // std::cout << "Chevyshef(): in._grid="<<in._grid<<std::endl;
+    // std::cout << "Chevyshef(): in.Grid()="<<in.Grid()<<std::endl;
     //std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
 
     int vol=grid->gSites();
@@ -321,7 +323,7 @@ struct ChebyParams : Serializable {
   // shift_Multiply in Rudy's code
   void AminusMuSq(LinearOperatorBase<Field> &Linop, const Field &in, Field &out) 
   {
-      GridBase *grid=in._grid;
+    GridBase *grid=in.Grid();
     Field tmp(grid);
 
     RealD aa= alpha*alpha;
@@ -338,7 +340,7 @@ struct ChebyParams : Serializable {
   // Implement the required interface
   void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 
-      GridBase *grid=in._grid;
+    GridBase *grid=in.Grid();
 
     int vol=grid->gSites();
 
@@ -373,5 +375,5 @@ struct ChebyParams : Serializable {
     }
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/approx/Forecast.h

@@ -31,7 +31,7 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #ifndef INCLUDED_FORECAST_H
 #define INCLUDED_FORECAST_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 // Abstract base class.
 // Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
@@ -57,10 +57,10 @@ namespace Grid {
     Field chi(phi); // forecasted solution
 
     // Trivial cases
-        if(degree == 0){ chi = zero; return chi; }
+    if(degree == 0){ chi = Zero(); return chi; }
     else if(degree == 1){ return prev_solns[0]; }
 
-        RealD dot;
+    //    RealD dot;
     ComplexD xp;
     Field r(phi); // residual
     Field Mv(phi);
@@ -92,7 +92,7 @@ namespace Grid {
     for(int j=0; j<degree; j++){
       for(int k=j+1; k<degree; k++){
 	G[j][k] = innerProduct(v[j],MdagMv[k]);
-          G[k][j] = std::conj(G[j][k]);
+	G[k][j] = conjugate(G[j][k]);
       }}
 
     // Gauss-Jordan elimination with partial pivoting
@@ -100,7 +100,7 @@ namespace Grid {
 
       // Perform partial pivoting
       int k = i;
-          for(int j=i+1; j<degree; j++){ if(std::abs(G[j][j]) > std::abs(G[k][k])){ k = j; } }
+      for(int j=i+1; j<degree; j++){ if(abs(G[j][j]) > abs(G[k][k])){ k = j; } }
       if(k != i){
 	xp = b[k];
 	b[k] = b[i];
@@ -121,7 +121,7 @@ namespace Grid {
     }
 
     // Use Gaussian elimination to solve equations and calculate initial guess
-        chi = zero;
+    chi = Zero();
     r = phi;
     for(int i=degree-1; i>=0; i--){
       a[i] = 0.0;
@@ -136,7 +136,7 @@ namespace Grid {
     for(int i=0; i<degree; i++){
       tmp = -b[i];
       for(int j=0; j<degree; j++){ tmp += G[i][j]*a[j]; }
-          tmp = std::conj(tmp)*tmp;
+      tmp = conjugate(tmp)*tmp;
       true_r += std::sqrt(tmp.real());
     }
 
@@ -147,6 +147,6 @@ namespace Grid {
   };
 };
 
-}
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/approx/MultiShiftFunction.cc

@@ -27,7 +27,8 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
     /*  END LEGAL */
 #include <Grid/GridCore.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
+
 double MultiShiftFunction::approx(double x)
 {
   double a = norm;
@@ -53,4 +54,4 @@ void MultiShiftFunction::csv(std::ostream &out)
   }
   return;
 }
-}
+NAMESPACE_END(Grid);

Grid/algorithms/approx/MultiShiftFunction.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef MULTI_SHIFT_FUNCTION
 #define MULTI_SHIFT_FUNCTION
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 class MultiShiftFunction {
 public:
@@ -63,5 +63,5 @@ public:
   }
 
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/approx/Remez.cc

@@ -298,7 +298,7 @@ void AlgRemez::stpini(bigfloat *step) {
 // Search for error maxima and minima
 void AlgRemez::search(bigfloat *step) {
   bigfloat a, q, xm, ym, xn, yn, xx0, xx1;
-  int i, j, meq, emsign, ensign, steps;
+  int i, meq, emsign, ensign, steps;
 
   meq = neq + 1;
   bigfloat *yy = new bigfloat[meq];
@@ -306,7 +306,6 @@ void AlgRemez::search(bigfloat *step) {
   bigfloat eclose = 1.0e30;
   bigfloat farther = 0l;
 
-  j = 1;
   xx0 = apstrt;
 
   for (i = 0; i < meq; i++) {

Grid/algorithms/approx/Zolotarev.cc

@@ -58,8 +58,8 @@
 
 /* Compute the partial fraction expansion coefficients (alpha) from the
  * factored form */
-namespace Grid {
-namespace Approx {
+NAMESPACE_BEGIN(Grid);
+NAMESPACE_BEGIN(Approx);
 
 static void construct_partfrac(izd *z) {
   int dn = z -> dn, dd = z -> dd, type = z -> type;
@@ -516,7 +516,9 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
   free(d);
   return zd;
 }
-}}
+
+NAMESPACE_END(Approx);
+NAMESPACE_END(Grid);
 
 #ifdef TEST
 
@@ -585,6 +587,7 @@ static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) {
   return (ONE - T) / (ONE + T);
 }
 
+
 /* Test program. Apart from printing out the parameters for R(x) it produces
  * the following data files for plotting (unless NPLOT is defined):
  *
@@ -723,5 +726,5 @@ int main(int argc, char** argv) {
   return EXIT_SUCCESS;
 }
 
-
 #endif /* TEST */
+

Grid/algorithms/approx/Zolotarev.h

@@ -1,13 +1,13 @@
 /* -*- Mode: C; comment-column: 22; fill-column: 79; -*- */
 
 #ifdef __cplusplus
-namespace Grid {
-namespace Approx {
+#include <Grid/Namespace.h>
+NAMESPACE_BEGIN(Grid);
+NAMESPACE_BEGIN(Approx);
 #endif
 
 #define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY>
 
-
 #ifndef ZOLOTAREV_INTERNAL
 #ifndef PRECISION
 #define PRECISION double
@@ -83,5 +83,6 @@ void zolotarev_free(zolotarev_data *zdata);
 #endif
 
 #ifdef __cplusplus
-}}
+NAMESPACE_END(Approx);
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/approx/bigfloat.h

@@ -10,10 +10,12 @@
 #ifndef INCLUDED_BIGFLOAT_H
 #define INCLUDED_BIGFLOAT_H
 
-
+#define __GMP_WITHIN_CONFIGURE
 #include <gmp.h>
 #include <mpf2mpfr.h>
 #include <mpfr.h>
+#undef  __GMP_WITHIN_CONFIGURE
+
 class bigfloat {
 
 private:

Grid/algorithms/iterative/AdefGeneric.h

@@ -90,8 +90,8 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
   void operator() (const Field &src, Field &psi){
   void operator() (const Field &src, Field &psi){
 
-    psi.checkerboard = src.checkerboard;
-    grid             = src._grid;
+    psi.Checkerboard() = src.Checkerboard();
+    grid             = src.Grid();
 
     RealD f;
     RealD rtzp,rtz,a,d,b;

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -27,13 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_BLOCK_CONJUGATE_GRADIENT_H
-#define GRID_BLOCK_CONJUGATE_GRADIENT_H
+#pragma once
 
+NAMESPACE_BEGIN(Grid);
 
-namespace Grid {
-
-enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS };
+enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
 
 //////////////////////////////////////////////////////////////////////////
 // Block conjugate gradient. Dimension zero should be the block direction
@@ -42,7 +40,6 @@ template <class Field>
 class BlockConjugateGradient : public OperatorFunction<Field> {
  public:
 
-
   typedef typename Field::scalar_type scomplex;
 
   int blockDim ;
@@ -54,21 +51,15 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
   RealD Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  Integer PrintInterval; //GridLogMessages or Iterative
   
   BlockConjugateGradient(BlockCGtype cgtype,int _Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : Tolerance(tol), CGtype(cgtype),   blockDim(_Orthog),  MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv)
+    : Tolerance(tol), CGtype(cgtype),   blockDim(_Orthog),  MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv),PrintInterval(100)
   {};
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Thin QR factorisation (google it)
 ////////////////////////////////////////////////////////////////////////////////////////////////////
-void ThinQRfact (Eigen::MatrixXcd &m_rr,
-		 Eigen::MatrixXcd &C,
-		 Eigen::MatrixXcd &Cinv,
-		 Field & Q,
-		 const Field & R)
-{
-  int Orthog = blockDim; // First dimension is block dim; this is an assumption
   ////////////////////////////////////////////////////////////////////////////////////////////////////
   //Dimensions
   // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
@@ -85,22 +76,20 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   // Cdag C = Rdag R ; passes.
   // QdagQ  = 1      ; passes
   ////////////////////////////////////////////////////////////////////////////////////////////////////
+void ThinQRfact (Eigen::MatrixXcd &m_rr,
+		 Eigen::MatrixXcd &C,
+		 Eigen::MatrixXcd &Cinv,
+		 Field & Q,
+		 const Field & R)
+{
+  int Orthog = blockDim; // First dimension is block dim; this is an assumption
   sliceInnerProductMatrix(m_rr,R,R,Orthog);
 
   // Force manifest hermitian to avoid rounding related
   m_rr = 0.5*(m_rr+m_rr.adjoint());
 
-#if 0
-  std::cout << " Calling Cholesky  ldlt on m_rr "  << m_rr <<std::endl;
-  Eigen::MatrixXcd L_ldlt = m_rr.ldlt().matrixL(); 
-  std::cout << " Called Cholesky  ldlt on m_rr "  << L_ldlt <<std::endl;
-  auto  D_ldlt = m_rr.ldlt().vectorD(); 
-  std::cout << " Called Cholesky  ldlt on m_rr "  << D_ldlt <<std::endl;
-#endif
-
-  //  std::cout << " Calling Cholesky  llt on m_rr "  <<std::endl;
   Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
-  //  std::cout << " Called Cholesky  llt on m_rr "  << L <<std::endl;
+
   C    = L.adjoint();
   Cinv = C.inverse();
   ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -112,6 +101,25 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   ////////////////////////////////////////////////////////////////////////////////////////////////////
   sliceMulMatrix(Q,Cinv,R,Orthog);
 }
+// see comments above
+void ThinQRfact (Eigen::MatrixXcd &m_rr,
+		 Eigen::MatrixXcd &C,
+		 Eigen::MatrixXcd &Cinv,
+		 std::vector<Field> & Q,
+		 const std::vector<Field> & R)
+{
+  InnerProductMatrix(m_rr,R,R);
+
+  m_rr = 0.5*(m_rr+m_rr.adjoint());
+
+  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
+
+  C    = L.adjoint();
+  Cinv = C.inverse();
+
+  MulMatrix(Q,Cinv,R);
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Call one of several implementations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -119,14 +127,20 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
 {
   if ( CGtype == BlockCGrQ ) {
     BlockCGrQsolve(Linop,Src,Psi);
-  } else if (CGtype == BlockCG ) {
-    BlockCGsolve(Linop,Src,Psi);
   } else if (CGtype == CGmultiRHS ) {
     CGmultiRHSsolve(Linop,Src,Psi);
   } else {
     assert(0);
   }
 }
+virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Field> &Src, std::vector<Field> &Psi) 
+{
+  if ( CGtype == BlockCGrQVec ) {
+    BlockCGrQsolveVec(Linop,Src,Psi);
+  } else {
+    assert(0);
+  }
+}
 
 ////////////////////////////////////////////////////////////////////////////
 // BlockCGrQ implementation:
@@ -138,11 +152,12 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
 void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X) 
 {
   int Orthog = blockDim; // First dimension is block dim; this is an assumption
-  Nblock = B._grid->_fdimensions[Orthog];
-
+  Nblock = B.Grid()->_fdimensions[Orthog];
+/* FAKE */
+  Nblock=8;
   std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
 
-  X.checkerboard = B.checkerboard;
+  X.Checkerboard() = B.Checkerboard();
   conformable(X, B);
 
   Field tmp(B);
@@ -202,15 +217,10 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   std::cout << GridLogMessage<<"BlockCGrQ algorithm initialisation " <<std::endl;
 
   //1.  QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
-
   Linop.HermOp(X, AD);
   tmp = B - AD;  
-  //std::cout << GridLogMessage << " initial tmp " << norm2(tmp)<< std::endl;
+
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
-  //std::cout << GridLogMessage << " initial Q " << norm2(Q)<< std::endl;
-  //std::cout << GridLogMessage << " m_rr " << m_rr<<std::endl;
-  //std::cout << GridLogMessage << " m_C " << m_C<<std::endl;
-  //std::cout << GridLogMessage << " m_Cinv " << m_Cinv<<std::endl;
   D=Q;
 
   std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
@@ -232,14 +242,12 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     MatrixTimer.Start();
     Linop.HermOp(D, Z);      
     MatrixTimer.Stop();
-    //std::cout << GridLogMessage << " norm2 Z " <<norm2(Z)<<std::endl;
 
     //4. M  = [D^dag Z]^{-1}
     sliceInnerTimer.Start();
     sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
     sliceInnerTimer.Stop();
     m_M       = m_DZ.inverse();
-    //std::cout << GridLogMessage << " m_DZ " <<m_DZ<<std::endl;
     
     //5. X  = X + D MC
     m_tmp     = m_M * m_C;
@@ -257,6 +265,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     
     //7. D  = Q + D S^dag
     m_tmp = m_S.adjoint();
+
     sliceMaddTimer.Start();
     sliceMaddMatrix(D,m_tmp,D,Q,Orthog);
     sliceMaddTimer.Stop();
@@ -317,163 +326,17 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   IterationsToComplete = k;
 }
 //////////////////////////////////////////////////////////////////////////
-// Block conjugate gradient; Original O'Leary Dimension zero should be the block direction
-//////////////////////////////////////////////////////////////////////////
-void BlockCGsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
-{
-  int Orthog = blockDim; // First dimension is block dim; this is an assumption
-  Nblock = Src._grid->_fdimensions[Orthog];
-
-  std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
-
-  Psi.checkerboard = Src.checkerboard;
-  conformable(Psi, Src);
-
-  Field P(Src);
-  Field AP(Src);
-  Field R(Src);
-  
-  Eigen::MatrixXcd m_pAp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_pAp_inv= Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_rr_inv = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  Eigen::MatrixXcd m_alpha      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_beta   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  // Initial residual computation & set up
-  std::vector<RealD> residuals(Nblock);
-  std::vector<RealD> ssq(Nblock);
-
-  sliceNorm(ssq,Src,Orthog);
-  RealD sssum=0;
-  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
-
-  sliceNorm(residuals,Src,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  sliceNorm(residuals,Psi,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  // Initial search dir is guess
-  Linop.HermOp(Psi, AP);
-  
-
-  /************************************************************************
-   * Block conjugate gradient (Stephen Pickles, thesis 1995, pp 71, O Leary 1980)
-   ************************************************************************
-   * O'Leary : R = B - A X
-   * O'Leary : P = M R ; preconditioner M = 1
-   * O'Leary : alpha = PAP^{-1} RMR
-   * O'Leary : beta  = RMR^{-1}_old RMR_new
-   * O'Leary : X=X+Palpha
-   * O'Leary : R_new=R_old-AP alpha
-   * O'Leary : P=MR_new+P beta
-   */
-
-  R = Src - AP;  
-  P = R;
-  sliceInnerProductMatrix(m_rr,R,R,Orthog);
-
-  GridStopWatch sliceInnerTimer;
-  GridStopWatch sliceMaddTimer;
-  GridStopWatch MatrixTimer;
-  GridStopWatch SolverTimer;
-  SolverTimer.Start();
-
-  int k;
-  for (k = 1; k <= MaxIterations; k++) {
-
-    RealD rrsum=0;
-    for(int b=0;b<Nblock;b++) rrsum+=real(m_rr(b,b));
-
-    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
-	      <<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
-
-    MatrixTimer.Start();
-    Linop.HermOp(P, AP);
-    MatrixTimer.Stop();
-
-    // Alpha
-    sliceInnerTimer.Start();
-    sliceInnerProductMatrix(m_pAp,P,AP,Orthog);
-    sliceInnerTimer.Stop();
-    m_pAp_inv = m_pAp.inverse();
-    m_alpha   = m_pAp_inv * m_rr ;
-
-    // Psi, R update
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(Psi,m_alpha, P,Psi,Orthog);     // add alpha *  P to psi
-    sliceMaddMatrix(R  ,m_alpha,AP,  R,Orthog,-1.0);// sub alpha * AP to resid
-    sliceMaddTimer.Stop();
-
-    // Beta
-    m_rr_inv = m_rr.inverse();
-    sliceInnerTimer.Start();
-    sliceInnerProductMatrix(m_rr,R,R,Orthog);
-    sliceInnerTimer.Stop();
-    m_beta = m_rr_inv *m_rr;
-
-    // Search update
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(AP,m_beta,P,R,Orthog);
-    sliceMaddTimer.Stop();
-    P= AP;
-
-    /*********************
-     * convergence monitor
-     *********************
-     */
-    RealD max_resid=0;
-    RealD rr;
-    for(int b=0;b<Nblock;b++){
-      rr = real(m_rr(b,b))/ssq[b];
-      if ( rr > max_resid ) max_resid = rr;
-    }
-    
-    if ( max_resid < Tolerance*Tolerance ) { 
-
-      SolverTimer.Stop();
-
-      std::cout << GridLogMessage<<"BlockCG converged in "<<k<<" iterations"<<std::endl;
-      for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "
-		  << std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
-      }
-      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
-
-      Linop.HermOp(Psi, AP);
-      AP = AP-Src;
-      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
-
-      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
-      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
-      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
-	    
-      IterationsToComplete = k;
-      return;
-    }
-
-  }
-  std::cout << GridLogMessage << "BlockConjugateGradient did NOT converge" << std::endl;
-
-  if (ErrorOnNoConverge) assert(0);
-  IterationsToComplete = k;
-}
-//////////////////////////////////////////////////////////////////////////
 // multiRHS conjugate gradient. Dimension zero should be the block direction
 // Use this for spread out across nodes
 //////////////////////////////////////////////////////////////////////////
 void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
 {
   int Orthog = blockDim; // First dimension is block dim
-  Nblock = Src._grid->_fdimensions[Orthog];
+  Nblock = Src.Grid()->_fdimensions[Orthog];
 
   std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
 
-  Psi.checkerboard = Src.checkerboard;
+  Psi.Checkerboard() = Src.Checkerboard();
   conformable(Psi, Src);
 
   Field P(Src);
@@ -600,7 +463,232 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   IterationsToComplete = k;
 }
 
-};
+void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
+  for(int b=0;b<Nblock;b++){
+  for(int bp=0;bp<Nblock;bp++) {
+    m(b,bp) = innerProduct(X[b],Y[bp]);  
+  }}
+}
+void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
+  // Should make this cache friendly with site outermost, parallel_for
+  // Deal with case AP aliases with either Y or X
+  std::vector<Field> tmp(Nblock,X[0]);
+  for(int b=0;b<Nblock;b++){
+    tmp[b]   = Y[b];
+    for(int bp=0;bp<Nblock;bp++) {
+      tmp[b] = tmp[b] + scomplex(scale*m(bp,b))*X[bp]; 
+    }
+  }
+  for(int b=0;b<Nblock;b++){
+    AP[b] = tmp[b];
+  }
+}
+void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
+  // Should make this cache friendly with site outermost, parallel_for
+  for(int b=0;b<Nblock;b++){
+    AP[b] = Zero();
+    for(int bp=0;bp<Nblock;bp++) {
+      AP[b] += scomplex(m(bp,b))*X[bp]; 
+    }
+  }
+}
+double normv(const std::vector<Field> &P){
+  double nn = 0.0;
+  for(int b=0;b<Nblock;b++) {
+    nn+=norm2(P[b]);
+  }
+  return nn;
+}
+
+////////////////////////////////////////////////////////////////////////////
+// BlockCGrQvec implementation:
+//--------------------------
+// X is guess/Solution
+// B is RHS
+// Solve A X_i = B_i    ;        i refers to Nblock index
+////////////////////////////////////////////////////////////////////////////
+void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field> &B, std::vector<Field> &X) 
+{
+  Nblock = B.size();
+  assert(Nblock == X.size());
+
+  std::cout<<GridLogMessage<<" Block Conjugate Gradient Vec rQ : Nblock "<<Nblock<<std::endl;
+
+  for(int b=0;b<Nblock;b++){ 
+    X[b].Checkerboard() = B[b].Checkerboard();
+    conformable(X[b], B[b]);
+    conformable(X[b], X[0]); 
+  }
+
+  Field Fake(B[0]);
+
+  std::vector<Field> tmp(Nblock,Fake);
+  std::vector<Field>   Q(Nblock,Fake);
+  std::vector<Field>   D(Nblock,Fake);
+  std::vector<Field>   Z(Nblock,Fake);
+  std::vector<Field>  AD(Nblock,Fake);
+
+  Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+
+  // Initial residual computation & set up
+  std::vector<RealD> residuals(Nblock);
+  std::vector<RealD> ssq(Nblock);
+
+  RealD sssum=0;
+  for(int b=0;b<Nblock;b++){ ssq[b] = norm2(B[b]);}
+  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+
+  for(int b=0;b<Nblock;b++){ residuals[b] = norm2(B[b]);}
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  for(int b=0;b<Nblock;b++){ residuals[b] = norm2(X[b]);}
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  /************************************************************************
+   * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
+   ************************************************************************
+   * Dimensions:
+   *
+   *   X,B==(Nferm x Nblock)
+   *   A==(Nferm x Nferm)
+   *  
+   * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
+   * 
+   * QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
+   * for k: 
+   *   Z  = AD
+   *   M  = [D^dag Z]^{-1}
+   *   X  = X + D MC
+   *   QS = Q - ZM
+   *   D  = Q + D S^dag
+   *   C  = S C
+   */
+  ///////////////////////////////////////
+  // Initial block: initial search dir is guess
+  ///////////////////////////////////////
+  std::cout << GridLogMessage<<"BlockCGrQvec algorithm initialisation " <<std::endl;
+
+  //1.  QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
+  for(int b=0;b<Nblock;b++) {
+    Linop.HermOp(X[b], AD[b]);
+    tmp[b] = B[b] - AD[b];  
+  }
+
+  ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
+
+  for(int b=0;b<Nblock;b++) D[b]=Q[b];
+
+  std::cout << GridLogMessage<<"BlockCGrQ vec computed initial residual and QR fact " <<std::endl;
+
+  ///////////////////////////////////////
+  // Timers
+  ///////////////////////////////////////
+  GridStopWatch sliceInnerTimer;
+  GridStopWatch sliceMaddTimer;
+  GridStopWatch QRTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch SolverTimer;
+  SolverTimer.Start();
+
+  int k;
+  for (k = 1; k <= MaxIterations; k++) {
+
+    //3. Z  = AD
+    MatrixTimer.Start();
+    for(int b=0;b<Nblock;b++) Linop.HermOp(D[b], Z[b]);      
+    MatrixTimer.Stop();
+
+    //4. M  = [D^dag Z]^{-1}
+    sliceInnerTimer.Start();
+    InnerProductMatrix(m_DZ,D,Z);
+    sliceInnerTimer.Stop();
+    m_M       = m_DZ.inverse();
+    
+    //5. X  = X + D MC
+    m_tmp     = m_M * m_C;
+    sliceMaddTimer.Start();
+    MaddMatrix(X,m_tmp, D,X);     
+    sliceMaddTimer.Stop();
+
+    //6. QS = Q - ZM
+    sliceMaddTimer.Start();
+    MaddMatrix(tmp,m_M,Z,Q,-1.0);
+    sliceMaddTimer.Stop();
+    QRTimer.Start();
+    ThinQRfact (m_rr, m_S, m_Sinv, Q, tmp);
+    QRTimer.Stop();
+    
+    //7. D  = Q + D S^dag
+    m_tmp = m_S.adjoint();
+    sliceMaddTimer.Start();
+    MaddMatrix(D,m_tmp,D,Q);
+    sliceMaddTimer.Stop();
+
+    //8. C  = S C
+    m_C = m_S*m_C;
+    
+    /*********************
+     * convergence monitor
+     *********************
+     */
+    m_rr = m_C.adjoint() * m_C;
+
+    RealD max_resid=0;
+    RealD rrsum=0;
+    RealD rr;
+
+    for(int b=0;b<Nblock;b++) {
+      rrsum+=real(m_rr(b,b));
+      rr = real(m_rr(b,b))/ssq[b];
+      if ( rr > max_resid ) max_resid = rr;
+    }
+
+    std::cout << GridLogIterative << "\t Block Iteration "<<k<<" ave resid "<< sqrt(rrsum/sssum) << " max "<< sqrt(max_resid) <<std::endl;
+
+    if ( max_resid < Tolerance*Tolerance ) { 
+
+      SolverTimer.Stop();
+
+      std::cout << GridLogMessage<<"BlockCGrQ converged in "<<k<<" iterations"<<std::endl;
+
+      for(int b=0;b<Nblock;b++){
+	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
+      }
+      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
+
+      for(int b=0;b<Nblock;b++) Linop.HermOp(X[b], AD[b]);
+      for(int b=0;b<Nblock;b++) AD[b] = AD[b]-B[b];
+      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(normv(AD)/normv(B)) <<std::endl;
+
+      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
+      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
+      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
+      std::cout << GridLogMessage << "\tThinQRfact " << QRTimer.Elapsed()  <<std::endl;
+	    
+      IterationsToComplete = k;
+      return;
+    }
 
   }
-#endif
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
+
+  if (ErrorOnNoConverge) assert(0);
+  IterationsToComplete = k;
+}
+
+};
+
+NAMESPACE_END(Grid);
+

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,248 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  using OperatorFunction<Field>::operator();
+
+  bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the CAGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<ComplexD> y;
+  std::vector<ComplexD> gamma;
+  std::vector<ComplexD> c;
+  std::vector<ComplexD> s;
+
+  CommunicationAvoidingGeneralisedMinimalResidual(RealD   tol,
+                                                  Integer maxit,
+                                                  Integer restart_length,
+                                                  bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular GMRES" << std::endl;
+
+    psi.Checkerboard() = src.Checkerboard();
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src.Grid());
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "CommunicationAvoidingGeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src.Grid());
+    Field r(src.Grid());
+
+    // this should probably be made a class member so that it is only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    ComplexD scale = 1.0/gamma[0];
+    v[0] = scale * r;
+
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, w, i);
+
+      qrUpdate(i);
+
+      cp = norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(v, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, Field &w, int iter) {
+
+    MatrixTimer.Start();
+    LinOp.Op(v[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - ComplexD(H(iter, i)) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
+      H(iter, i)     = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    auto nu     = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = conjugate(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &v, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - ComplexD(H(k, i)) * y[k];
+      y[i] = y[i] / ComplexD(H(i, i));
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + v[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/ConjugateGradient.h

@@ -31,7 +31,7 @@ directory
 #ifndef GRID_CONJUGATE_GRADIENT_H
 #define GRID_CONJUGATE_GRADIENT_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -41,6 +41,9 @@ namespace Grid {
 template <class Field>
 class ConjugateGradient : public OperatorFunction<Field> {
 public:
+
+  using OperatorFunction<Field>::operator();
+
   bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
@@ -54,11 +57,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
+    psi.Checkerboard() = src.Checkerboard();
 
-    psi.checkerboard = src.checkerboard;
     conformable(psi, src);
 
-    RealD cp, c, a, d, b, ssq, qq, b_pred;
+    RealD cp, c, a, d, b, ssq, qq;
+    //RealD b_pred;
 
     Field p(src);
     Field mmp(src);
@@ -89,6 +93,8 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
     // Check if guess is really REALLY good :)
     if (cp <= rsq) {
+      std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
+      IterationsToComplete = 0;	
       return;
     }
 
@@ -104,7 +110,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
     SolverTimer.Start();
     int k;
-    for (k = 1; k <= MaxIterations*1000; k++) {
+    for (k = 1; k <= MaxIterations; k++) {
       c = cp;
 
       MatrixTimer.Start();
@@ -125,15 +131,18 @@ class ConjugateGradient : public OperatorFunction<Field> {
       b = cp / c;
 
       LinearCombTimer.Start();
-      parallel_for(int ss=0;ss<src._grid->oSites();ss++){
-	vstream(psi[ss], a      *  p[ss] + psi[ss]);
-	vstream(p  [ss], b      *  p[ss] + r[ss]);
-      }
+      auto psi_v = psi.View();
+      auto p_v   = p.View();
+      auto r_v   = r.View();
+      accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+	  coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+	  coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+      });
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
       std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
-                << " residual " << cp << " target " << rsq << std::endl;
+                << " residual^2 " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -141,12 +150,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
 
-        RealD srcnorm = sqrt(norm2(src));
-        RealD resnorm = sqrt(norm2(p));
+        RealD srcnorm = std::sqrt(norm2(src));
+        RealD resnorm = std::sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
 
         std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k << std::endl;
-        std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+        std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
 	std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
 	std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
 
@@ -165,13 +174,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
         return;
       }
     }
-    std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
-              << std::endl;
+    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
 
     if (ErrorOnNoConverge) assert(0);
     IterationsToComplete = k;
 
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -28,10 +28,12 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
 #ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
 #define GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
   //Mixed precision restarted defect correction CG
-  template<class FieldD,class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+  template<class FieldD,class FieldF, 
+    typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+    typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
   public:                                                
     RealD   Tolerance;
@@ -50,7 +52,12 @@ namespace Grid {
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
     
-    MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) :
+    MixedPrecisionConjugateGradient(RealD tol, 
+				    Integer maxinnerit, 
+				    Integer maxouterit, 
+				    GridBase* _sp_grid, 
+				    LinearOperatorBase<FieldF> &_Linop_f, 
+				    LinearOperatorBase<FieldD> &_Linop_d) :
       Linop_f(_Linop_f), Linop_d(_Linop_d),
       Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
       OuterLoopNormMult(100.), guesser(NULL){ };
@@ -65,18 +72,18 @@ namespace Grid {
     GridStopWatch TotalTimer;
     TotalTimer.Start();
     
-      int cb = src_d_in.checkerboard;
-      sol_d.checkerboard = cb;
+    int cb = src_d_in.Checkerboard();
+    sol_d.Checkerboard() = cb;
     
     RealD src_norm = norm2(src_d_in);
     RealD stop = src_norm * Tolerance*Tolerance;
 
-      GridBase* DoublePrecGrid = src_d_in._grid;
+    GridBase* DoublePrecGrid = src_d_in.Grid();
     FieldD tmp_d(DoublePrecGrid);
-      tmp_d.checkerboard = cb;
+    tmp_d.Checkerboard() = cb;
     
     FieldD tmp2_d(DoublePrecGrid);
-      tmp2_d.checkerboard = cb;
+    tmp2_d.Checkerboard() = cb;
     
     FieldD src_d(DoublePrecGrid);
     src_d = src_d_in; //source for next inner iteration, computed from residual during operation
@@ -84,10 +91,10 @@ namespace Grid {
     RealD inner_tol = InnerTolerance;
     
     FieldF src_f(SinglePrecGrid);
-      src_f.checkerboard = cb;
+    src_f.Checkerboard() = cb;
     
     FieldF sol_f(SinglePrecGrid);
-      sol_f.checkerboard = cb;
+    sol_f.Checkerboard() = cb;
     
     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
     CG_f.ErrorOnNoConverge = false;
@@ -115,7 +122,7 @@ namespace Grid {
       precisionChange(src_f, src_d);
       PrecChangeTimer.Stop();
       
-	zeroit(sol_f);
+      sol_f = Zero();
 
       //Optionally improve inner solver guess (eg using known eigenvectors)
       if(guesser != NULL)
@@ -149,6 +156,6 @@ namespace Grid {
   }
 };
 
-}
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
 #define GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -41,6 +41,9 @@ namespace Grid {
 				    public OperatorFunction<Field>
 {
 public:                                                
+
+  using OperatorFunction<Field>::operator();
+
   RealD   Tolerance;
   Integer MaxIterations;
     Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
@@ -56,7 +59,7 @@ public:
 
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
   {
-  GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
     int nshift = shifts.order;
     std::vector<Field> results(nshift,grid);
     (*this)(Linop,src,results,psi);
@@ -78,7 +81,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
   
-  GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
   
     ////////////////////////////////////////////////////////////////////////
     // Convenience references to the info stored in "MultiShiftFunction"
@@ -318,5 +321,5 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   }
 
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -28,9 +28,11 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
 #ifndef GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
 #define GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
-  template<class FieldD,class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+template<class FieldD,class FieldF, 
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
 class ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
 public:
   bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
@@ -74,7 +76,7 @@ namespace Grid {
     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
     bool using_fallback = false;
       
-      psi.checkerboard = src.checkerboard;
+    psi.Checkerboard() = src.Checkerboard();
     conformable(psi, src);
 
     RealD cp, c, a, d, b, ssq, qq, b_pred;
@@ -108,17 +110,17 @@ namespace Grid {
     // Check if guess is really REALLY good :)
     if (cp <= rsq) {
       std::cout << GridLogMessage << "ConjugateGradientReliableUpdate guess was REALLY good\n";
-	std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+      std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
       return;
     }
 
     //Single prec initialization
     FieldF r_f(SinglePrecGrid);
-      r_f.checkerboard = r.checkerboard;
+    r_f.Checkerboard() = r.Checkerboard();
     precisionChange(r_f, r);
 
     FieldF psi_f(r_f);
-      psi_f = zero;
+    psi_f = Zero();
 
     FieldF p_f(r_f);
     FieldF mmp_f(r_f);
@@ -178,12 +180,12 @@ namespace Grid {
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
 	p = mmp - src;
 
-	  RealD srcnorm = sqrt(norm2(src));
-	  RealD resnorm = sqrt(norm2(p));
+	RealD srcnorm = std::sqrt(norm2(src));
+	RealD resnorm = std::sqrt(norm2(p));
 	RealD true_residual = resnorm / srcnorm;
 
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate Converged on iteration " << k << " after " << l << " reliable updates" << std::endl;
-	  std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+	std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
 	std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
 	std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
 
@@ -217,7 +219,7 @@ namespace Grid {
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
 	r = src - mmp;
 
-	  psi_f = zero;
+	psi_f = Zero();
 	precisionChange(r_f, r);
 	cp = norm2(r);
 	MaxResidSinceLastRelUp = cp;
@@ -249,7 +251,7 @@ namespace Grid {
 };
 
 
-};
+NAMESPACE_END(Grid);
 
 
 

Grid/algorithms/iterative/ConjugateResidual.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CONJUGATE_RESIDUAL_H
 #define GRID_CONJUGATE_RESIDUAL_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -39,6 +39,8 @@ namespace Grid {
 template<class Field> 
 class ConjugateResidual : public OperatorFunction<Field> {
 public:                                                
+  using OperatorFunction<Field>::operator();
+
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;
@@ -49,14 +51,14 @@ namespace Grid {
 
   void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
 
-      RealD a, b, c, d;
+    RealD a, b; // c, d;
     RealD cp, ssq,rsq;
       
     RealD rAr, rAAr, rArp;
     RealD pAp, pAAp;
 
-      GridBase *grid = src._grid;
-      psi=zero;
+    GridBase *grid = src.Grid();
+    psi=Zero();
     Field r(grid),  p(grid), Ap(grid), Ar(grid);
       
     r=src;
@@ -95,8 +97,8 @@ namespace Grid {
 	axpy(r,-1.0,src,Ap);
 	RealD true_resid = norm2(r)/ssq;
 	std::cout<<GridLogMessage<<"ConjugateResidual: Converged on iteration " <<k
-		   << " computed residual "<<sqrt(cp/ssq)
-	           << " true residual "<<sqrt(true_resid)
+		 << " computed residual "<<std::sqrt(cp/ssq)
+		 << " true residual "<<std::sqrt(true_resid)
 		 << " target "       <<Tolerance <<std::endl;
 	return;
       }
@@ -107,5 +109,5 @@ namespace Grid {
     assert(0);
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/Deflation.h

@@ -33,9 +33,13 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
-  virtual void operator()(const Field &src, Field &guess) { guess = zero; };
+    virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
+};
+template<class Field>
+class DoNothingGuesser: public LinearFunction<Field> {
+public:
+  virtual void operator()(const Field &src, Field &guess) {  };
 };
-
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
@@ -56,14 +60,14 @@ public:
   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
 
   virtual void operator()(const Field &src,Field &guess) {
-    guess = zero;
+    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);
     }
-    guess.checkerboard = src.checkerboard;
+    guess.Checkerboard() = src.Checkerboard();
   }
 };
 
@@ -86,15 +90,15 @@ public:
   
   void operator()(const FineField &src,FineField &guess) { 
     int N = (int)evec_coarse.size();
-    CoarseField src_coarse(evec_coarse[0]._grid);
-    CoarseField guess_coarse(evec_coarse[0]._grid);    guess_coarse = zero;
+    CoarseField src_coarse(evec_coarse[0].Grid());
+    CoarseField guess_coarse(evec_coarse[0].Grid());    guess_coarse = Zero();
     blockProject(src_coarse,src,subspace);    
     for (int i=0;i<N;i++) {
       const CoarseField & tmp = evec_coarse[i];
       axpy(guess_coarse,TensorRemove(innerProduct(tmp,src_coarse)) / eval_coarse[i],tmp,guess_coarse);
     }
     blockPromote(guess_coarse,guess,subspace);
-    guess.checkerboard = src.checkerboard;
+    guess.Checkerboard() = src.Checkerboard();
   };
 };
 

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,258 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_FLEXIBLE_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_FLEXIBLE_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  using OperatorFunction<Field>::operator();
+
+  bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the FCAGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch PrecTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<ComplexD> y;
+  std::vector<ComplexD> gamma;
+  std::vector<ComplexD> c;
+  std::vector<ComplexD> s;
+
+  LinearFunction<Field> &Preconditioner;
+
+  FlexibleCommunicationAvoidingGeneralisedMinimalResidual(RealD   tol,
+                                                          Integer maxit,
+                                                          LinearFunction<Field> &Prec,
+                                                          Integer restart_length,
+                                                          bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.)
+      , Preconditioner(Prec) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular FGMRES" << std::endl;
+
+    psi.Checkerboard() = src.Checkerboard();
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src.Grid());
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    PrecTimer.Reset();
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Precon  " <<         PrecTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src.Grid());
+    Field r(src.Grid());
+
+    // these should probably be made class members so that they are only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
+    std::vector<Field> z(RestartLength + 1, src.Grid()); for (auto &elem : z) elem = Zero();
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, z, w, i);
+
+      qrUpdate(i);
+
+      cp = norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(z, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, std::vector<Field> &z, Field &w, int iter) {
+
+    PrecTimer.Start();
+    Preconditioner(v[iter], z[iter]);
+    PrecTimer.Stop();
+
+    MatrixTimer.Start();
+    LinOp.Op(z[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - ComplexD(H(iter, i)) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
+      H(iter, i)     = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = conjugate(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &z, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - ComplexD(H(k, i)) * y[k];
+      y[i] = y[i] / ComplexD(H(i, i));
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h

@@ -0,0 +1,256 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  using OperatorFunction<Field>::operator();
+
+  bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the FGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch PrecTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<ComplexD> y;
+  std::vector<ComplexD> gamma;
+  std::vector<ComplexD> c;
+  std::vector<ComplexD> s;
+
+  LinearFunction<Field> &Preconditioner;
+
+  FlexibleGeneralisedMinimalResidual(RealD   tol,
+                                     Integer maxit,
+                                     LinearFunction<Field> &Prec,
+                                     Integer restart_length,
+                                     bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.)
+      , Preconditioner(Prec) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    psi.Checkerboard() = src.Checkerboard();
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src.Grid());
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    PrecTimer.Reset();
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "FlexibleGeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Precon  " <<         PrecTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src.Grid());
+    Field r(src.Grid());
+
+    // these should probably be made class members so that they are only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
+    std::vector<Field> z(RestartLength + 1, src.Grid()); for (auto &elem : z) elem = Zero();
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, z, w, i);
+
+      qrUpdate(i);
+
+      cp = norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(z, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, std::vector<Field> &z, Field &w, int iter) {
+
+    PrecTimer.Start();
+    Preconditioner(v[iter], z[iter]);
+    PrecTimer.Stop();
+
+    MatrixTimer.Start();
+    LinOp.Op(z[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - ComplexD(H(iter, i)) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
+      H(iter, i)     = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = conjugate(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &z, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - ComplexD(H(k, i)) * y[k];
+      y[i] = y[i] / ComplexD(H(i, i));
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/GeneralisedMinimalResidual.h

@@ -0,0 +1,244 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/GeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class GeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  using OperatorFunction<Field>::operator();
+
+  bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the GMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<ComplexD> y;
+  std::vector<ComplexD> gamma;
+  std::vector<ComplexD> c;
+  std::vector<ComplexD> s;
+
+  GeneralisedMinimalResidual(RealD   tol,
+                             Integer maxit,
+                             Integer restart_length,
+                             bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    psi.Checkerboard() = src.Checkerboard();
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src.Grid());
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "GeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "GeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "GeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "GMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "GeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src.Grid());
+    Field r(src.Grid());
+
+    // this should probably be made a class member so that it is only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, w, i);
+
+      qrUpdate(i);
+
+      cp = norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "GeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(v, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, Field &w, int iter) {
+
+    MatrixTimer.Start();
+    LinOp.Op(v[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - ComplexD(H(iter, i)) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
+      H(iter, i)     = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = conjugate(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &v, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - ComplexD(H(k, i)) * y[k];
+      y[i] = y[i] / ComplexD(H(i, i));
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + v[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -35,7 +35,7 @@ Author: Christoph Lehner <clehner@bnl.gov>
 //#include <zlib.h>
 #include <sys/stat.h>
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
 
   ////////////////////////////////////////////////////////
   // Move following 100 LOC to lattice/Lattice_basis.h
@@ -52,26 +52,31 @@ void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
 template<class Field>
 void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm) 
 {
+  typedef decltype(basis[0].View()) View;
+  auto tmp_v = basis[0].View();
+  std::vector<View> basis_v(basis.size(),tmp_v);
   typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0]._grid;
+  GridBase* grid = basis[0].Grid();
       
-  parallel_region
+  for(int k=0;k<basis.size();k++){
+    basis_v[k] = basis[k].View();
+  }
+
+  thread_region
   {
-
     std::vector < vobj , commAllocator<vobj> > B(Nm); // Thread private
-       
-    parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
+    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[k]._odata[ss];
+	  B[j] +=Qt(j,k) * basis_v[k][ss];
 	}
       }
       for(int j=j0; j<j1; ++j){
-	  basis[j]._odata[ss] = B[j];
-      }
+	basis_v[j][ss] = B[j];
       }
+    });
   }
 }
 
@@ -80,16 +85,18 @@ template<class Field>
 void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
 {
   typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0]._grid;
+  GridBase* grid = basis[0].Grid();
 
-  result.checkerboard = basis[0].checkerboard;
-  parallel_for(int ss=0;ss < grid->oSites();ss++){
-    vobj B = zero;
+  result.Checkerboard() = basis[0].Checkerboard();
+  auto result_v=result.View();
+  thread_for(ss, grid->oSites(),{
+    vobj B = Zero();
     for(int k=k0; k<k1; ++k){
-      B +=Qt(j,k) * basis[k]._odata[ss];
-    }
-    result._odata[ss] = B;
+      auto basis_k = basis[k].View();
+      B +=Qt(j,k) * basis_k[ss];
     }
+    result_v[ss] = B;
+  });
 }
 
 template<class Field>
@@ -119,7 +126,7 @@ void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, s
 
       assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
 
-      std::swap(_v[i]._odata,_v[idx[i]]._odata); // should use vector move constructor, no data copy
+      swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
       std::swap(sort_vals[i],sort_vals[idx[i]]);
 
       idx[j] = idx[i];
@@ -150,6 +157,19 @@ void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, boo
   basisReorderInPlace(_v,sort_vals,idx);
 }
 
+// PAB: faster to compute the inner products first then fuse loops.
+// If performance critical can improve.
+template<class Field>
+void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
+  result = Zero();
+  assert(_v.size()==eval.size());
+  int N = (int)_v.size();
+  for (int i=0;i<N;i++) {
+    Field& tmp = _v[i];
+    axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
+  }
+}
+
 /////////////////////////////////////////////////////////////
 // Implicitly restarted lanczos
 /////////////////////////////////////////////////////////////
@@ -289,7 +309,7 @@ public:
   template<typename T>  static RealD normalise(T& v) 
   {
     RealD nn = norm2(v);
-    nn = sqrt(nn);
+    nn = std::sqrt(nn);
     v = v * (1.0/nn);
     return nn;
   }
@@ -321,8 +341,8 @@ until convergence
 */
   void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
   {
-    GridBase *grid = src._grid;
-    assert(grid == evec[0]._grid);
+    GridBase *grid = src.Grid();
+    assert(grid == evec[0].Grid());
     
     GridLogIRL.TimingMode(1);
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
@@ -446,7 +466,7 @@ until convergence
       assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
 
       basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
-      std::cout<<GridLogIRL <<"basisRotated  by Qt"<<std::endl;
+      std::cout<<GridLogIRL <<"basisRotated  by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
       
       ////////////////////////////////////////////////////
       // Compressed vector f and beta(k2)
@@ -454,7 +474,7 @@ until convergence
       f *= Qt(k2-1,Nm-1);
       f += lme[k2-1] * evec[k2];
       beta_k = norm2(f);
-      beta_k = sqrt(beta_k);
+      beta_k = std::sqrt(beta_k);
       std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
 	  
       RealD betar = 1.0/beta_k;
@@ -477,7 +497,7 @@ until convergence
 
 	std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
 
-	Field B(grid); B.checkerboard = evec[0].checkerboard;
+	Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
 
 	//  power of two search pattern;  not every evalue in eval2 is assessed.
 	int allconv =1;
@@ -515,7 +535,7 @@ until convergence
 	
   converged:
     {
-      Field B(grid); B.checkerboard = evec[0].checkerboard;
+      Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
       basisRotate(evec,Qt,0,Nk,0,Nk,Nm);	    
       std::cout << GridLogIRL << " Rotated basis"<<std::endl;
       Nconv=0;
@@ -807,7 +827,7 @@ void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
     
     // determination of 2x2 leading submatrix
     RealD dsub = lmd[kmax-1]-lmd[kmax-2];
-    RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
+    RealD dd = std::sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
     RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
     // (Dsh: shift)
     
@@ -838,5 +858,6 @@ void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
   abort();
 }
 };
-}
+
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -29,8 +29,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LOCAL_COHERENCE_IRL_H
 #define GRID_LOCAL_COHERENCE_IRL_H
 
-namespace Grid { 
-
+NAMESPACE_BEGIN(Grid); 
 
 struct LanczosParams : Serializable {
 public:
@@ -59,7 +58,7 @@ struct LocalCoherenceLanczosParams : Serializable {
 				  RealD        , coarse_relax_tol,
 				  std::vector<int>, blockSize,
 				  std::string, config,
-				  std::vector < std::complex<double>  >, omega,
+				  std::vector < ComplexD  >, omega,
 				  RealD, mass,
 				  RealD, M5);
 };
@@ -83,11 +82,11 @@ public:
   };
 
   void operator()(const CoarseField& in, CoarseField& out) {
-    GridBase *FineGrid = subspace[0]._grid;    
-    int   checkerboard = subspace[0].checkerboard;
+    GridBase *FineGrid = subspace[0].Grid();    
+    int   checkerboard = subspace[0].Checkerboard();
 
-    FineField fin (FineGrid);     fin.checkerboard= checkerboard;
-    FineField fout(FineGrid);   fout.checkerboard = checkerboard;
+    FineField fin (FineGrid);     fin.Checkerboard()= checkerboard;
+    FineField fout(FineGrid);   fout.Checkerboard() = checkerboard;
 
     blockPromote(in,fin,subspace);       std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
     _Linop.HermOp(fin,fout);                   std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
@@ -118,11 +117,11 @@ public:
 
   void operator()(const CoarseField& in, CoarseField& out) {
 
-    GridBase *FineGrid = subspace[0]._grid;    
-    int   checkerboard = subspace[0].checkerboard;
+    GridBase *FineGrid = subspace[0].Grid();    
+    int   checkerboard = subspace[0].Checkerboard();
 
-    FineField fin (FineGrid); fin.checkerboard =checkerboard;
-    FineField fout(FineGrid);fout.checkerboard =checkerboard;
+    FineField fin (FineGrid); fin.Checkerboard() =checkerboard;
+    FineField fout(FineGrid);fout.Checkerboard() =checkerboard;
     
     blockPromote(in,fin,subspace);             std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
     _poly(_Linop,fin,fout);                    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
@@ -182,10 +181,10 @@ class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanc
   }
   int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
-    GridBase *FineGrid = _subspace[0]._grid;    
-    int checkerboard   = _subspace[0].checkerboard;
-    FineField fB(FineGrid);fB.checkerboard =checkerboard;
-    FineField fv(FineGrid);fv.checkerboard =checkerboard;
+    GridBase *FineGrid = _subspace[0].Grid();    
+    int checkerboard   = _subspace[0].Checkerboard();
+    FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
+    FineField fv(FineGrid);fv.Checkerboard() =checkerboard;
 
     blockPromote(B,fv,_subspace);  
     
@@ -305,11 +304,11 @@ public:
     int Nk = nbasis;
     subspace.resize(Nk,_FineGrid);
     subspace[0]=1.0;
-    subspace[0].checkerboard=_checkerboard;
+    subspace[0].Checkerboard()=_checkerboard;
     normalise(subspace[0]);
     PlainHermOp<FineField>    Op(_FineOp);
     for(int k=1;k<Nk;k++){
-      subspace[k].checkerboard=_checkerboard;
+      subspace[k].Checkerboard()=_checkerboard;
       Op(subspace[k-1],subspace[k]);
       normalise(subspace[k]);
     }
@@ -360,7 +359,11 @@ public:
 
     ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
 
-    FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
+    FineField src(_FineGrid); 
+    typedef typename FineField::scalar_type Scalar;
+    // src=1.0; 
+    src=Scalar(1.0); 
+    src.Checkerboard() = _checkerboard;
 
     int Nconv;
     IRL.calc(evals_fine,subspace,src,Nconv,false);
@@ -402,5 +405,5 @@ public:
   }
 };
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/MinimalResidual.h

@@ -0,0 +1,157 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/MinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_MINIMAL_RESIDUAL_H
+#define GRID_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field> class MinimalResidual : public OperatorFunction<Field> {
+ public:
+  using OperatorFunction<Field>::operator();
+
+  bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
+                          // Defaults true.
+  RealD   Tolerance;
+  Integer MaxIterations;
+  RealD   overRelaxParam;
+  Integer IterationsToComplete; // Number of iterations the MR took to finish.
+                                // Filled in upon completion
+
+  MinimalResidual(RealD tol, Integer maxit, Real ovrelparam = 1.0, bool err_on_no_conv = true)
+    : Tolerance(tol), MaxIterations(maxit), overRelaxParam(ovrelparam), ErrorOnNoConverge(err_on_no_conv){};
+
+  void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
+
+    psi.Checkerboard() = src.Checkerboard();
+    conformable(psi, src);
+
+    ComplexD a, c;
+    RealD    d;
+
+    Field Mr(src);
+    Field r(src);
+
+    // Initial residual computation & set up
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Linop.Op(psi, Mr);
+
+    r = src - Mr;
+
+    RealD cp = norm2(r);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "MinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "MinimalResidual:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << "MinimalResidual:  cp,r " << cp << std::endl;
+
+    if (cp <= rsq) {
+      return;
+    }
+
+    std::cout << GridLogIterative << "MinimalResidual: k=0 residual " << cp << " target " << rsq << std::endl;
+
+    GridStopWatch LinalgTimer;
+    GridStopWatch MatrixTimer;
+    GridStopWatch SolverTimer;
+
+    SolverTimer.Start();
+    int k;
+    for (k = 1; k <= MaxIterations; k++) {
+
+      MatrixTimer.Start();
+      Linop.Op(r, Mr);
+      MatrixTimer.Stop();
+
+      LinalgTimer.Start();
+
+      c = innerProduct(Mr, r);
+
+      d = norm2(Mr);
+
+      a = c / d;
+
+      a = a * overRelaxParam;
+
+      psi = psi + r * a;
+
+      r = r - Mr * a;
+
+      cp = norm2(r);
+
+      LinalgTimer.Stop();
+
+      std::cout << GridLogIterative << "MinimalResidual: Iteration " << k
+                << " residual " << cp << " target " << rsq << std::endl;
+      std::cout << GridLogDebug << "a = " << a << " c = " << c << " d = " << d << std::endl;
+
+      // Stopping condition
+      if (cp <= rsq) {
+        SolverTimer.Stop();
+
+        Linop.Op(psi, Mr);
+        r = src - Mr;
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "MinimalResidual Converged on iteration " << k
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "MR Time elapsed: Total   " << SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MR Time elapsed: Matrix  " << MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MR Time elapsed: Linalg  " << LinalgTimer.Elapsed() << std::endl;
+
+        if (ErrorOnNoConverge)
+          assert(true_residual / Tolerance < 10000.0);
+
+        IterationsToComplete = k;
+
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "MinimalResidual did NOT converge"
+              << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+
+    IterationsToComplete = k;
+  }
+};
+} // namespace Grid
+#endif

Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h

@@ -0,0 +1,276 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_MIXED_PRECISION_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_MIXED_PRECISION_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class FieldD, class FieldF, typename std::enable_if<getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
+class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction<FieldD> {
+ public:
+
+  using OperatorFunction<FieldD>::operator();
+
+  bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the MPFGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch PrecTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+  GridStopWatch ChangePrecTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<ComplexD> y;
+  std::vector<ComplexD> gamma;
+  std::vector<ComplexD> c;
+  std::vector<ComplexD> s;
+
+  GridBase* SinglePrecGrid;
+
+  LinearFunction<FieldF> &Preconditioner;
+
+  MixedPrecisionFlexibleGeneralisedMinimalResidual(RealD   tol,
+                                                   Integer maxit,
+                                                   GridBase * sp_grid,
+                                                   LinearFunction<FieldF> &Prec,
+                                                   Integer restart_length,
+                                                   bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.)
+      , SinglePrecGrid(sp_grid)
+      , Preconditioner(Prec) {};
+
+  void operator()(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi) {
+
+    psi.Checkerboard() = src.Checkerboard();
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    FieldD r(src.Grid());
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "MPFGMRES: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "MPFGMRES:   src " << ssq   << std::endl;
+
+    PrecTimer.Reset();
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+    ChangePrecTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "MPFGMRES: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Total      " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Precon     " <<         PrecTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Matrix     " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Linalg     " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: QR         " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: CompSol    " << CompSolutionTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: PrecChange " <<   ChangePrecTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "MPFGMRES did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    FieldD w(src.Grid());
+    FieldD r(src.Grid());
+
+    // these should probably be made class members so that they are only allocated once, not in every restart
+    std::vector<FieldD> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
+    std::vector<FieldD> z(RestartLength + 1, src.Grid()); for (auto &elem : z) elem = Zero();
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, z, w, i);
+
+      qrUpdate(i);
+
+      cp = norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "MPFGMRES: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(z, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<FieldD> &LinOp, std::vector<FieldD> &v, std::vector<FieldD> &z, FieldD &w, int iter) {
+
+    FieldF v_f(SinglePrecGrid);
+    FieldF z_f(SinglePrecGrid);
+
+    ChangePrecTimer.Start();
+    precisionChange(v_f, v[iter]);
+    precisionChange(z_f, z[iter]);
+    ChangePrecTimer.Stop();
+
+    PrecTimer.Start();
+    Preconditioner(v_f, z_f);
+    PrecTimer.Stop();
+
+    ChangePrecTimer.Start();
+    precisionChange(z[iter], z_f);
+    ChangePrecTimer.Stop();
+
+    MatrixTimer.Start();
+    LinOp.Op(z[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - ComplexD(H(iter, i)) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
+      H(iter, i)     = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = conjugate(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<FieldD> const &z, FieldD &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - ComplexD(H(k, i)) * y[k];
+      y[i] = y[i] / ComplexD(H(i, i));
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/NormalEquations.h

@@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_NORMAL_EQUATIONS_H
 #define GRID_NORMAL_EQUATIONS_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
 // Take a matrix and form an NE solver calling a Herm solver
@@ -48,7 +48,7 @@ namespace Grid {
 
   void operator() (const Field &in, Field &out){
  
-      Field src(in._grid);
+    Field src(in.Grid());
 
     _Matrix.Mdag(in,src);
     _HermitianSolver(src,out);  // Mdag M out = Mdag in
@@ -56,5 +56,5 @@ namespace Grid {
   }     
 };
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/PowerMethod.h

@@ -0,0 +1,45 @@
+#pragma once
+namespace Grid {
+template<class Field> class PowerMethod  
+{ 
+ public: 
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  RealD operator()(LinearOperatorBase<Field> &HermOp, const Field &src) 
+  { 
+    GridBase *grid = src.Grid(); 
+    
+    // quickly get an idea of the largest eigenvalue to more properly normalize the residuum 
+    RealD evalMaxApprox = 0.0; 
+    auto src_n = src; 
+    auto tmp = src; 
+    const int _MAX_ITER_EST_ = 50; 
+
+    for (int i=0;i<_MAX_ITER_EST_;i++) { 
+      
+      normalise(src_n); 
+      HermOp.HermOp(src_n,tmp); 
+      RealD vnum = real(innerProduct(src_n,tmp)); // HermOp. 
+      RealD vden = norm2(src_n); 
+      RealD na = vnum/vden; 
+      
+      if ( (fabs(evalMaxApprox/na - 1.0) < 0.01) || (i==_MAX_ITER_EST_-1) ) { 
+ 	evalMaxApprox = na; 
+ 	return evalMaxApprox; 
+      } 
+      evalMaxApprox = na; 
+      std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+      src_n = tmp;
+    }
+    assert(0);
+    return 0;
+  }
+};
+}

Grid/algorithms/iterative/PrecConjugateResidual.h

@@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PREC_CONJUGATE_RESIDUAL_H
 #define GRID_PREC_CONJUGATE_RESIDUAL_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -56,7 +56,7 @@ namespace Grid {
     RealD rAr, rAAr, rArp;
     RealD pAp, pAAp;
 
-      GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
     Field r(grid),  p(grid), Ap(grid), Ar(grid), z(grid);
       
     psi=zero;
@@ -115,5 +115,5 @@ namespace Grid {
     assert(0);
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -36,11 +36,13 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 //NB. Likely not original reference since they are focussing on a preconditioner variant.
 //    but VPGCR was nicely written up in their paper
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class Field>
 class PrecGeneralisedConjugateResidual : public OperatorFunction<Field> {
 public:                                                
+  using OperatorFunction<Field>::operator();
+
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;
@@ -65,12 +67,12 @@ namespace Grid {
 
   void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
 
-      psi=zero;
+    psi=Zero();
     RealD cp, ssq,rsq;
     ssq=norm2(src);
     rsq=Tolerance*Tolerance*ssq;
       
-      Field r(src._grid);
+    Field r(src.Grid());
 
     PrecTimer.Reset();
     MatTimer.Reset();
@@ -113,11 +115,11 @@ namespace Grid {
   RealD GCRnStep(LinearOperatorBase<Field> &Linop,const Field &src, Field &psi,RealD rsq){
 
     RealD cp;
-      RealD a, b, c, d;
+    RealD a, b;
     RealD zAz, zAAz;
-      RealD rAq, rq;
+    RealD rq;
 
-      GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
 
     Field r(grid);
     Field z(grid);
@@ -132,6 +134,7 @@ namespace Grid {
     std::vector<Field> p(mmax,grid);
     std::vector<RealD> qq(mmax);
       
+
     //////////////////////////////////
     // initial guess x0 is taken as nonzero.
     // r0=src-A x0 = src
@@ -139,7 +142,10 @@ namespace Grid {
     MatTimer.Start();
     Linop.HermOpAndNorm(psi,Az,zAz,zAAz); 
     MatTimer.Stop();
+    
+    LinalgTimer.Start();
     r=src-Az;
+    LinalgTimer.Stop();
     
     /////////////////////
     // p = Prec(r)
@@ -152,8 +158,10 @@ namespace Grid {
     Linop.HermOp(z,tmp); 
     MatTimer.Stop();
 
+    LinalgTimer.Start();
     ttmp=tmp;
     tmp=tmp-r;
+    LinalgTimer.Stop();
 
     /*
       std::cout<<GridLogMessage<<r<<std::endl;
@@ -166,12 +174,14 @@ namespace Grid {
     Linop.HermOpAndNorm(z,Az,zAz,zAAz); 
     MatTimer.Stop();
 
+    LinalgTimer.Start();
     //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++){
 
@@ -181,12 +191,14 @@ namespace Grid {
       int peri_k = k %mmax;
       int peri_kp= kp%mmax;
 
+      LinalgTimer.Start();
       rq= real(innerProduct(r,q[peri_k])); // 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();
 
       if((k==nstep-1)||(cp<rsq)){
 	return cp;
@@ -202,6 +214,8 @@ namespace Grid {
       Linop.HermOpAndNorm(z,Az,zAz,zAAz);
       Linop.HermOp(z,tmp);
       MatTimer.Stop();
+
+      LinalgTimer.Start();
       tmp=tmp-r;
       std::cout<<GridLogMessage<< " Preconditioner resid " <<sqrt(norm2(tmp)/norm2(r))<<std::endl; 
 
@@ -219,12 +233,11 @@ namespace Grid {
 
       }
       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

@@ -86,23 +86,26 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    */
 namespace Grid {
 
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Use base class to share code
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
   // Take a matrix and form a Red Black solver calling a Herm solver
   // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Now make the norm reflect extra factor of Mee
-  template<class Field> class SchurRedBlackStaggeredSolve {
-  private:
+  template<class Field> class SchurRedBlackBase {
+  protected:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
     OperatorFunction<Field> & _HermitianRBSolver;
     int CBfactorise;
     bool subGuess;
+    bool useSolnAsInitGuess; // if true user-supplied solution vector is used as initial guess for solver
   public:
 
-    /////////////////////////////////////////////////////
-    // Wrap the usual normal equations Schur trick
-    /////////////////////////////////////////////////////
-  SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false)  :
-     _HermitianRBSolver(HermitianRBSolver) 
+    SchurRedBlackBase(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+        const bool _solnAsInitGuess = false)  :
+    _HermitianRBSolver(HermitianRBSolver),
+    useSolnAsInitGuess(_solnAsInitGuess)
     { 
       CBfactorise = 0;
       subtractGuess(initSubGuess);
@@ -116,12 +119,90 @@ namespace Grid {
       return subGuess;
     }
 
-    template<class Matrix>
+    /////////////////////////////////////////////////////////////
+    // Shared code
+    /////////////////////////////////////////////////////////////
     void operator() (Matrix & _Matrix,const Field &in, Field &out){
       ZeroGuesser<Field> guess;
       (*this)(_Matrix,in,out,guess);
     }
-    template<class Matrix, class Guesser>
+    void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out) 
+    {
+      ZeroGuesser<Field> guess;
+      (*this)(_Matrix,in,out,guess);
+    }
+
+    template<class Guesser>
+    void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess) 
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+      int nblock = in.size();
+
+      std::vector<Field> src_o(nblock,grid);
+      std::vector<Field> sol_o(nblock,grid);
+      
+      std::vector<Field> guess_save;
+
+      Field resid(fgrid);
+      Field tmp(grid);
+
+      ////////////////////////////////////////////////
+      // Prepare RedBlack source
+      ////////////////////////////////////////////////
+      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) {
+          pickCheckerboard(Odd, sol_o[b], out[b]);
+        } else {
+          guess(src_o[b],sol_o[b]); 
+        }
+
+	if ( subGuess ) { 
+	  guess_save[b] = sol_o[b];
+	}
+      }
+      //////////////////////////////////////////////////////////////
+      // Call the block solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlackBase calling the solver for "<<nblock<<" RHS" <<std::endl;
+      RedBlackSolve(_Matrix,src_o,sol_o);
+
+      ////////////////////////////////////////////////
+      // A2A boolean behavioural control & reconstruct other checkerboard
+      ////////////////////////////////////////////////
+      for(int b=0;b<nblock;b++) {
+
+	if (subGuess)   sol_o[b] = sol_o[b] - guess_save[b];
+
+	///////// Needs even source //////////////
+	pickCheckerboard(Even,tmp,in[b]);
+	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]);
+
+	/////////////////////////////////////////////////
+	// Check unprec residual if possible
+	/////////////////////////////////////////////////
+	if ( ! subGuess ) {
+	  _Matrix.M(out[b],resid); 
+	  resid = resid-in[b];
+	  RealD ns = norm2(in[b]);
+	  RealD nr = norm2(resid);
+	
+	  std::cout<<GridLogMessage<< "SchurRedBlackBase solver true unprec resid["<<b<<"] "<<std::sqrt(nr/ns) << std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "SchurRedBlackBase Guess subtracted after solve["<<b<<"] " << std::endl;
+	}
+
+      }
+    }
+    template<class Guesser>
     void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
 
       // FIXME CGdiagonalMee not implemented virtual function
@@ -129,55 +210,42 @@ namespace Grid {
       GridBase *grid = _Matrix.RedBlackGrid();
       GridBase *fgrid= _Matrix.Grid();
 
-      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
- 
-      Field src_e(grid);
-      Field src_o(grid);
-      Field sol_e(grid);
-      Field sol_o(grid);
-      Field   tmp(grid);
-      Field  Mtmp(grid);
       Field resid(fgrid);
+      Field src_o(grid);
+      Field src_e(grid);
+      Field sol_o(grid);
 
-      std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve " <<std::endl;
-      pickCheckerboard(Even,src_e,in);
-      pickCheckerboard(Odd ,src_o,in);
-      pickCheckerboard(Even,sol_e,out);
+      ////////////////////////////////////////////////
+      // RedBlack source
+      ////////////////////////////////////////////////
+      RedBlackSource(_Matrix,in,src_e,src_o);
+
+      ////////////////////////////////
+      // Construct the guess
+      ////////////////////////////////
+      if(useSolnAsInitGuess) {
         pickCheckerboard(Odd, sol_o, out);
-      std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve checkerboards picked" <<std::endl;
+      } else {
+        guess(src_o,sol_o);
+      }
 
-      /////////////////////////////////////////////////////
-      // src_o = (source_o - Moe MeeInv source_e)
-      /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
-
-      //src_o = tmp;     assert(src_o.checkerboard ==Odd);
-      _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
+      Field  guess_save(grid);
+      guess_save = sol_o;
 
       //////////////////////////////////////////////////////////////
       // Call the red-black solver
       //////////////////////////////////////////////////////////////
-      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver calling the Mpc solver" <<std::endl;
-      guess(src_o, sol_o);
-      Mtmp = sol_o;
-      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
-      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver called  the Mpc solver" <<std::endl;
+      RedBlackSolve(_Matrix,src_o,sol_o);
+
+      ////////////////////////////////////////////////
       // Fionn A2A boolean behavioural control
-      if (subGuess)        sol_o = sol_o-Mtmp;
+      ////////////////////////////////////////////////
+      if (subGuess)      sol_o= sol_o-guess_save;
 
       ///////////////////////////////////////////////////
-      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      // RedBlack solution needs the even source
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
-     
-      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver reconstructed other CB" <<std::endl;
-      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
-      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
-      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver inserted solution" <<std::endl;
+      RedBlackSolution(_Matrix,sol_o,src_e,out);
 
       // Verify the unprec residual
       if ( ! subGuess ) {
@@ -185,319 +253,234 @@ namespace Grid {
         resid = resid-in;
         RealD ns = norm2(in);
         RealD nr = norm2(resid);
-        std::cout<<GridLogMessage << "SchurRedBlackStaggered solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+
+        std::cout<<GridLogMessage << "SchurRedBlackBase solver true unprec resid "<< std::sqrt(nr/ns) << std::endl;
       } else {
-        std::cout << GridLogMessage << "Guess subtracted after solve." << std::endl;
+        std::cout << GridLogMessage << "SchurRedBlackBase Guess subtracted after solve." << std::endl;
       }
     }     
+    
+    /////////////////////////////////////////////////////////////
+    // Override in derived. 
+    /////////////////////////////////////////////////////////////
+    virtual void RedBlackSource  (Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)                =0;
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)          =0;
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)                           =0;
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)=0;
+
+  };
+
+  template<class Field> class SchurRedBlackStaggeredSolve : public SchurRedBlackBase<Field> {
+  public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
+
+    SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+        const bool _solnAsInitGuess = false) 
+      :    SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) 
+    {
+    }
+
+    //////////////////////////////////////////////////////
+    // Override RedBlack specialisation
+    //////////////////////////////////////////////////////
+    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd ,src_o,src);
+
+      /////////////////////////////////////////////////////
+      // src_o = (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+
+      _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
+    }
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e_c,Field &sol)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field   tmp(grid);
+      Field   sol_e(grid);
+      Field   src_e(grid);
+
+      src_e = src_e_c; // Const correctness
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.Checkerboard()   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+     
+      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+    }
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
+    {
+      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+    };
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
+    {
+      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
+    }
   };
   template<class Field> using SchurRedBlackStagSolve = SchurRedBlackStaggeredSolve<Field>;
 
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Take a matrix and form a Red Black solver calling a Herm solver
-  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  // Site diagonal has Mooee on it.
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class SchurRedBlackDiagMooeeSolve {
-  private:
-    OperatorFunction<Field> & _HermitianRBSolver;
-    int CBfactorise;
-    bool subGuess;
+  template<class Field> class SchurRedBlackDiagMooeeSolve : public SchurRedBlackBase<Field> {
   public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
 
-    /////////////////////////////////////////////////////
-    // Wrap the usual normal equations Schur trick
-    /////////////////////////////////////////////////////
-  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver,int cb=0, const bool initSubGuess = false)  :  _HermitianRBSolver(HermitianRBSolver) 
-  { 
-    CBfactorise=cb;
-    subtractGuess(initSubGuess);
-  };
-    void subtractGuess(const bool initSubGuess)
-    {
-      subGuess = initSubGuess;
-    }
-    bool isSubtractGuess(void)
-    {
-      return subGuess;
-    }
-    template<class Matrix>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out){
-      ZeroGuesser<Field> guess;
-      (*this)(_Matrix,in,out,guess);
-    }
-    template<class Matrix, class Guesser>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
+    SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+        const bool _solnAsInitGuess = false)  
+      : SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
 
-      // FIXME CGdiagonalMee not implemented virtual function
-      // FIXME use CBfactorise to control schur decomp
+
+    //////////////////////////////////////////////////////
+    // Override RedBlack specialisation
+    //////////////////////////////////////////////////////
+    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
+    {
       GridBase *grid = _Matrix.RedBlackGrid();
       GridBase *fgrid= _Matrix.Grid();
 
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd ,src_o,src);
+
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+
+      // get the right MpcDag
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+
+    }
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
 
-      Field src_e(grid);
-      Field src_o(grid);
-      Field sol_e(grid);
-      Field sol_o(grid);
       Field   tmp(grid);
-      Field  Mtmp(grid);
-      Field resid(fgrid);
-
-      pickCheckerboard(Even,src_e,in);
-      pickCheckerboard(Odd ,src_o,in);
-      pickCheckerboard(Even,sol_e,out);
-      pickCheckerboard(Odd ,sol_o,out);
-    
-      /////////////////////////////////////////////////////
-      // src_o = Mdag * (source_o - Moe MeeInv source_e)
-      /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
-
-      // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
-
-      //////////////////////////////////////////////////////////////
-      // Call the red-black solver
-      //////////////////////////////////////////////////////////////
-      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
-      guess(src_o,sol_o);
-      Mtmp = sol_o;
-      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
-      // Fionn A2A boolean behavioural control
-      if (subGuess)        sol_o = sol_o-Mtmp;
-
+      Field  sol_e(grid);
+      Field  src_e_i(grid);
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.Checkerboard()   ==Even);
+      src_e_i = src_e-tmp;               assert(  src_e_i.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
-      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
-
-      // Verify the unprec residual
-      if ( ! subGuess ) {
-        _Matrix.M(out,resid); 
-        resid = resid-in;
-        RealD ns = norm2(in);
-        RealD nr = norm2(resid);
-
-        std::cout<<GridLogMessage << "SchurRedBlackDiagMooee solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
-      } else {
-        std::cout << GridLogMessage << "Guess subtracted after solve." << std::endl;
+      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
     }
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
+    {
+      SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+    };
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
+    {
+      SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
     }
   };
 
-
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Take a matrix and form a Red Black solver calling a Herm solver
-  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  // Site diagonal is identity, right preconditioned by Mee^inv
+  // ( 1 - Meo Moo^inv Moe Mee^inv  ) phi =( 1 - Meo Moo^inv Moe Mee^inv  ) Mee psi =  = eta  = eta
+  //=> psi = MeeInv phi
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class SchurRedBlackDiagTwoSolve {
-  private:
-    OperatorFunction<Field> & _HermitianRBSolver;
-    int CBfactorise;
-    bool subGuess;
+  template<class Field> class SchurRedBlackDiagTwoSolve : public SchurRedBlackBase<Field> {
   public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
 
     /////////////////////////////////////////////////////
     // Wrap the usual normal equations Schur trick
     /////////////////////////////////////////////////////
-  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false)  :
-     _HermitianRBSolver(HermitianRBSolver) 
-    { 
-      CBfactorise = 0;
-      subtractGuess(initSubGuess);
-    };
-    void subtractGuess(const bool initSubGuess)
-    {
-      subGuess = initSubGuess;
-    }
-    bool isSubtractGuess(void)
-    {
-      return subGuess;
-    }
+  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+      const bool _solnAsInitGuess = false)  
+    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
 
-    template<class Matrix>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out){
-      ZeroGuesser<Field> guess;
-      (*this)(_Matrix,in,out,guess);
-    }
-    template<class Matrix,class Guesser>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
-
-      // FIXME CGdiagonalMee not implemented virtual function
-      // FIXME use CBfactorise to control schur decomp
+    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
+    {
       GridBase *grid = _Matrix.RedBlackGrid();
       GridBase *fgrid= _Matrix.Grid();
 
       SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
       
-      Field src_e(grid);
-      Field src_o(grid);
-      Field sol_e(grid);
-      Field sol_o(grid);
       Field   tmp(grid);
       Field  Mtmp(grid);
-      Field resid(fgrid);
 
-      pickCheckerboard(Even,src_e,in);
-      pickCheckerboard(Odd ,src_o,in);
-      pickCheckerboard(Even,sol_e,out);
-      pickCheckerboard(Odd ,sol_o,out);
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd ,src_o,src);
     
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
-
-      //////////////////////////////////////////////////////////////
-      // Call the red-black solver
-      //////////////////////////////////////////////////////////////
-      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
-//      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
-      guess(src_o,tmp);
-      Mtmp = tmp;
-      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
-      // Fionn A2A boolean behavioural control
-      if (subGuess)      tmp = tmp-Mtmp;
-      _Matrix.MooeeInv(tmp,sol_o);       assert(  sol_o.checkerboard   ==Odd);
-
-      ///////////////////////////////////////////////////
-      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-      ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
-     
-      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
-      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
-
-      // Verify the unprec residual
-      if ( ! subGuess ) {
-        _Matrix.M(out,resid); 
-        resid = resid-in;
-        RealD ns = norm2(in);
-        RealD nr = norm2(resid);
-
-        std::cout<<GridLogMessage << "SchurRedBlackDiagTwo solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
-      } else {
-        std::cout << GridLogMessage << "Guess subtracted after solve." << std::endl;
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
     }
-    }     
-  };
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Take a matrix and form a Red Black solver calling a Herm solver
-  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class SchurRedBlackDiagTwoMixed {
-  private:
-    LinearFunction<Field> & _HermitianRBSolver;
-    int CBfactorise;
-    bool subGuess;
-  public:
 
-    /////////////////////////////////////////////////////
-    // Wrap the usual normal equations Schur trick
-    /////////////////////////////////////////////////////
-  SchurRedBlackDiagTwoMixed(LinearFunction<Field> &HermitianRBSolver, const bool initSubGuess = false)  :
-     _HermitianRBSolver(HermitianRBSolver) 
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
     {
-      CBfactorise=0;
-      subtractGuess(initSubGuess);
-    };
-    void subtractGuess(const bool initSubGuess)
-    {
-      subGuess = initSubGuess;
-    }
-    bool isSubtractGuess(void)
-    {
-      return subGuess;
-    }
-
-    template<class Matrix>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out){
-      ZeroGuesser<Field> guess;
-      (*this)(_Matrix,in,out,guess);
-    }
-    template<class Matrix, class Guesser>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
-
-      // FIXME CGdiagonalMee not implemented virtual function
-      // FIXME use CBfactorise to control schur decomp
       GridBase *grid = _Matrix.RedBlackGrid();
       GridBase *fgrid= _Matrix.Grid();
 
-      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
- 
-      Field src_e(grid);
-      Field src_o(grid);
-      Field sol_e(grid);
-      Field sol_o(grid);
+      Field   sol_o_i(grid);
       Field   tmp(grid);
-      Field  Mtmp(grid);
-      Field resid(fgrid);
+      Field   sol_e(grid);
 
-      pickCheckerboard(Even,src_e,in);
-      pickCheckerboard(Odd ,src_o,in);
-      pickCheckerboard(Even,sol_e,out);
-      pickCheckerboard(Odd ,sol_o,out);
-    
-      /////////////////////////////////////////////////////
-      // src_o = Mdag * (source_o - Moe MeeInv source_e)
-      /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
-
-      // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
-
-      //////////////////////////////////////////////////////////////
-      // Call the red-black solver
-      //////////////////////////////////////////////////////////////
-      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
-//      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
-//      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
-      guess(src_o,tmp);
-      Mtmp = tmp;
-      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
-      // Fionn A2A boolean behavioural control
-      if (subGuess)      tmp = tmp-Mtmp;
-      _Matrix.MooeeInv(tmp,sol_o);        assert(  sol_o.checkerboard   ==Odd);
+      ////////////////////////////////////////////////
+      // MooeeInv due to pecond
+      ////////////////////////////////////////////////
+      _Matrix.MooeeInv(sol_o,tmp);
+      sol_o_i = tmp;
 
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.Checkerboard()   ==Even);
+      tmp = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e);   assert(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
-      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
-
-      // Verify the unprec residual
-      if ( ! subGuess ) {
-        _Matrix.M(out,resid); 
-        resid = resid-in;
-        RealD ns = norm2(in);
-        RealD nr = norm2(resid);
-
-        std::cout << GridLogMessage << "SchurRedBlackDiagTwo solver true unprec resid " << std::sqrt(nr / ns) << " nr " << nr << " ns " << ns << std::endl;
-      } else {
-        std::cout << GridLogMessage << "Guess subtracted after solve." << std::endl;
-      }
-    }     
+      setCheckerboard(sol,sol_e);    assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.Checkerboard() ==Odd );
     };
 
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
+    {
+      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);
+    };
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
+    {
+      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
+    }
+  };
 }
 #endif

Grid/allocator/AlignedAllocator.cc

@@ -1,11 +1,12 @@
 #include <Grid/GridCore.h>
 #include <fcntl.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 MemoryStats *MemoryProfiler::stats = nullptr;
 bool         MemoryProfiler::debug = false;
 
+#ifdef POINTER_CACHE
 int PointerCache::victim;
 
 PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
@@ -51,7 +52,7 @@ void *PointerCache::Lookup(size_t bytes) {
 
   if (bytes < 4096 ) return NULL;
 
-#ifdef _OPENMP
+#ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
 
@@ -63,7 +64,7 @@ void *PointerCache::Lookup(size_t bytes) {
   }
   return NULL;
 }
-
+#endif
 
 void check_huge_pages(void *Buf,uint64_t BYTES)
 {
@@ -122,4 +123,5 @@ std::string sizeString(const size_t bytes)
   return std::string(buf);
 }
 
-}
+NAMESPACE_END(Grid);
+

Grid/allocator/AlignedAllocator.h

@@ -40,8 +40,13 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <mm_malloc.h>
 #endif
 
-namespace Grid {
+#define POINTER_CACHE
+#define GRID_ALLOC_ALIGN (2*1024*1024)
 
+NAMESPACE_BEGIN(Grid);
+
+// Move control to configure.ac and Config.h?
+#ifdef POINTER_CACHE
 class PointerCache {
 private:
 
@@ -63,6 +68,7 @@ namespace Grid {
   static void *Lookup(size_t bytes) ;
 
 };
+#endif  
 
 std::string sizeString(size_t bytes);
 
@@ -152,29 +158,45 @@ public:
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
 
-    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
-    //    if ( ptr != NULL ) 
-    //      std::cout << "alignedAllocator "<<__n << " cache hit "<< std::hex << ptr <<std::dec <<std::endl;
 
-    //////////////////
-    // Hack 2MB align; could make option probably doesn't need configurability
-    //////////////////
-//define GRID_ALLOC_ALIGN (128)
-#define GRID_ALLOC_ALIGN (2*1024*1024)
+#ifdef POINTER_CACHE
+    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
+#else
+    pointer ptr = nullptr;
+#endif
+
+#ifdef GRID_NVCC
+    ////////////////////////////////////
+    // Unified (managed) memory
+    ////////////////////////////////////
+    if ( ptr == (_Tp *) NULL ) {
+      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
-    //    std::cout << "alignedAllocator " << std::hex << ptr <<std::dec <<std::endl;
+    assert( ptr != (_Tp *)NULL);
+
+    //////////////////////////////////////////////////
     // First touch optimise in threaded loop 
-    uint8_t *cp = (uint8_t *)ptr;
-#ifdef GRID_OMP
-#pragma omp parallel for
-#endif
-    for(size_type n=0;n<bytes;n+=4096){
+    //////////////////////////////////////////////////
+    uint64_t *cp = (uint64_t *)ptr;
+    thread_for(n,bytes/sizeof(uint64_t), { // need only one touch per page
       cp[n]=0;
-    }
+    });
+#endif
     return ptr;
   }
 
@@ -183,133 +205,40 @@ public:
 
     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
   }
-  void construct(pointer __p, const _Tp& __val) { };
+
+  // FIXME: hack for the copy constructor, eventually it must be avoided
+  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
+  //void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
 template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
-//////////////////////////////////////////////////////////////////////////////////////////
-// MPI3 : comms must use shm region
-// SHMEM: comms must use symmetric heap
-//////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_COMMS_SHMEM
-extern "C" { 
-#include <mpp/shmem.h>
-extern void * shmem_align(size_t, size_t);
-extern void  shmem_free(void *);
-}
-#define PARANOID_SYMMETRIC_HEAP
-#endif
-
-template<typename _Tp>
-class commAllocator {
-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 commAllocator<_Tp1> other; };
-  commAllocator() throw() { }
-  commAllocator(const commAllocator&) throw() { }
-  template<typename _Tp1> commAllocator(const commAllocator<_Tp1>&) throw() { }
-  ~commAllocator() throw() { }
-  pointer       address(reference __x)       const { return &__x; }
-  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
-
-#ifdef GRID_COMMS_SHMEM
-  pointer allocate(size_type __n, const void* _p= 0)
-  {
-    size_type bytes = __n*sizeof(_Tp);
-
-    profilerAllocate(bytes);
-#ifdef CRAY
-    _Tp *ptr = (_Tp *) shmem_align(bytes,64);
-#else
-    _Tp *ptr = (_Tp *) shmem_align(64,bytes);
-#endif
-#ifdef PARANOID_SYMMETRIC_HEAP
-    static void * bcast;
-    static long  psync[_SHMEM_REDUCE_SYNC_SIZE];
-
-    bcast = (void *) ptr;
-    shmem_broadcast32((void *)&bcast,(void *)&bcast,sizeof(void *)/4,0,0,0,shmem_n_pes(),psync);
-
-    if ( bcast != ptr ) {
-      std::printf("inconsistent alloc pe %d %lx %lx \n",shmem_my_pe(),bcast,ptr);std::fflush(stdout);
-      //      BACKTRACEFILE();
-      exit(0);
-    }
-    assert( bcast == (void *) ptr);
-#endif 
-    return ptr;
-  }
-  void deallocate(pointer __p, size_type __n) { 
-    size_type bytes = __n*sizeof(_Tp);
-
-    profilerFree(bytes);
-    shmem_free((void *)__p);
-  }
-#else
-  pointer allocate(size_type __n, const void* _p= 0) 
-  {
-    size_type bytes = __n*sizeof(_Tp);
-    
-    profilerAllocate(bytes);
-#ifdef HAVE_MM_MALLOC_H
-    _Tp * ptr = (_Tp *) _mm_malloc(bytes, GRID_ALLOC_ALIGN);
-#else
-    _Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN, bytes);
-#endif
-    uint8_t *cp = (uint8_t *)ptr;
-    if ( ptr ) { 
-    // One touch per 4k page, static OMP loop to catch same loop order
-#ifdef GRID_OMP
-#pragma omp parallel for schedule(static)
-#endif
-      for(size_type n=0;n<bytes;n+=4096){
-	cp[n]=0;
-      }
-    }
-    return ptr;
-  }
-  void deallocate(pointer __p, size_type __n) {
-    size_type bytes = __n*sizeof(_Tp);
-
-    profilerFree(bytes);
-#ifdef HAVE_MM_MALLOC_H
-    _mm_free((void *)__p); 
-#else
-    free((void *)__p);
-#endif
-  }
-#endif
-  void construct(pointer __p, const _Tp& __val) { };
-  void construct(pointer __p) { };
-  void destroy(pointer __p) { };
-};
-template<typename _Tp>  inline bool operator==(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return true; }
-template<typename _Tp>  inline bool operator!=(const commAllocator<_Tp>&, const commAllocator<_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,commAllocator<T> >;              
+template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
 template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
 
-}; // namespace Grid
+NAMESPACE_END(Grid);
+
 #endif

Grid/cartesian/Cartesian_base.h

@@ -30,16 +30,15 @@
 #ifndef GRID_CARTESIAN_BASE_H
 #define GRID_CARTESIAN_BASE_H
 
-
-namespace Grid{
+NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////
 // Commicator provides information on the processor grid
 //////////////////////////////////////////////////////////////////////
 //    unsigned long _ndimension;
-  //    std::vector<int> _processors; // processor grid
+//    Coordinate _processors; // processor grid
 //    int              _processor;  // linear processor rank
-  //    std::vector<int> _processor_coor;  // linear processor rank
+//    Coordinate _processor_coor;  // linear processor rank
 //////////////////////////////////////////////////////////////////////
 class GridBase : public CartesianCommunicator , public GridThread {
 
@@ -48,12 +47,14 @@ public:
   // Give Lattice access
   template<class object> friend class Lattice;
 
-    GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
-    GridBase(const std::vector<int> & processor_grid,
+  GridBase(const Coordinate & processor_grid) : CartesianCommunicator(processor_grid) {}; 
+
+  GridBase(const Coordinate & processor_grid,
 	   const CartesianCommunicator &parent,
 	   int &split_rank) 
     : CartesianCommunicator(processor_grid,parent,split_rank) {};
-    GridBase(const std::vector<int> & processor_grid,
+
+  GridBase(const Coordinate & processor_grid,
 	   const CartesianCommunicator &parent) 
     : CartesianCommunicator(processor_grid,parent,dummy) {};
 
@@ -61,23 +62,23 @@ public:
 
 
   // Physics Grid information.
-    std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
-    std::vector<int> _fdimensions;// (full) Global dimensions of array prior to cb removal
-    std::vector<int> _gdimensions;// Global dimensions of array after cb removal
-    std::vector<int> _ldimensions;// local dimensions of array with processor images removed
-    std::vector<int> _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
-    std::vector<int> _ostride;    // Outer stride for each dimension
-    std::vector<int> _istride;    // Inner stride i.e. within simd lane
+  Coordinate _simd_layout;// Which dimensions get relayed out over simd lanes.
+  Coordinate _fdimensions;// (full) Global dimensions of array prior to cb removal
+  Coordinate _gdimensions;// Global dimensions of array after cb removal
+  Coordinate _ldimensions;// local dimensions of array with processor images removed
+  Coordinate _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
+  Coordinate _ostride;    // Outer stride for each dimension
+  Coordinate _istride;    // Inner stride i.e. within simd lane
   int _osites;                  // _isites*_osites = product(dimensions).
   int _isites;
   int _fsites;                  // _isites*_osites = product(dimensions).
   int _gsites;
-    std::vector<int> _slice_block;// subslice information
-    std::vector<int> _slice_stride;
-    std::vector<int> _slice_nblock;
+  Coordinate _slice_block;// subslice information
+  Coordinate _slice_stride;
+  Coordinate _slice_nblock;
 
-    std::vector<int> _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
-    std::vector<int> _lend  ;     // local end of array in gcoors   _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
+  Coordinate _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
+  Coordinate _lend  ;     // local end of array in gcoors   _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
 
     bool _isCheckerBoarded; 
 
@@ -88,7 +89,7 @@ public:
   // GridCartesian / GridRedBlackCartesian
   ////////////////////////////////////////////////////////////////
   virtual int CheckerBoarded(int dim)=0;
-    virtual int CheckerBoard(const std::vector<int> &site)=0;
+  virtual int CheckerBoard(const Coordinate &site)=0;
   virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
   virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
   virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;
@@ -107,20 +108,20 @@ public:
   // coordinate. Note, however, for data parallel operations the "inner" indexing cost is not paid and all
   // lanes are operated upon simultaneously.
   
-    virtual int oIndex(std::vector<int> &coor)
+  virtual int oIndex(Coordinate &coor)
   {
     int idx=0;
     // Works with either global or local coordinates
     for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
     return idx;
   }
-    virtual int iIndex(std::vector<int> &lcoor)
+  virtual int iIndex(Coordinate &lcoor)
   {
     int idx=0;
     for(int d=0;d<_ndimension;d++) idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
     return idx;
   }
-    inline int oIndexReduced(std::vector<int> &ocoor)
+  inline int oIndexReduced(Coordinate &ocoor)
   {
     int idx=0; 
     // ocoor is already reduced so can eliminate the modulo operation
@@ -128,11 +129,11 @@ public:
     for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*ocoor[d];
     return idx;
   }
-    inline void oCoorFromOindex (std::vector<int>& coor,int Oindex){
+  inline void oCoorFromOindex (Coordinate& coor,int Oindex){
     Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
   }
 
-    inline void InOutCoorToLocalCoor (std::vector<int> &ocoor, std::vector<int> &icoor, std::vector<int> &lcoor) {
+  inline void InOutCoorToLocalCoor (Coordinate &ocoor, Coordinate &icoor, Coordinate &lcoor) {
     lcoor.resize(_ndimension);
     for (int d = 0; d < _ndimension; d++)
       lcoor[d] = ocoor[d] + _rdimensions[d] * icoor[d];
@@ -141,7 +142,7 @@ public:
   //////////////////////////////////////////////////////////
   // SIMD lane addressing
   //////////////////////////////////////////////////////////
-    inline void iCoorFromIindex(std::vector<int> &coor,int lane)
+  inline void iCoorFromIindex(Coordinate &coor,int lane)
   {
     Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
   }
@@ -152,8 +153,6 @@ public:
   inline int PermuteType(int dimension){
     int permute_type=0;
     //
-      // FIXME:
-      //
     // Best way to encode this would be to present a mask 
     // for which simd dimensions are rotated, and the rotation
     // size. If there is only one simd dimension rotated, this is just 
@@ -186,11 +185,11 @@ public:
   inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
-    inline const std::vector<int> LocalStarts(void)             { return _lstart;    };
-    inline const std::vector<int> &FullDimensions(void)         { return _fdimensions;};
-    inline const std::vector<int> &GlobalDimensions(void)       { return _gdimensions;};
-    inline const std::vector<int> &LocalDimensions(void)        { return _ldimensions;};
-    inline const std::vector<int> &VirtualLocalDimensions(void) { return _ldimensions;};
+  inline const Coordinate LocalStarts(void)             { return _lstart;    };
+  inline const Coordinate &FullDimensions(void)         { return _fdimensions;};
+  inline const Coordinate &GlobalDimensions(void)       { return _gdimensions;};
+  inline const Coordinate &LocalDimensions(void)        { return _ldimensions;};
+  inline const Coordinate &VirtualLocalDimensions(void) { return _ldimensions;};
 
   ////////////////////////////////////////////////////////////////
   // Utility to print the full decomposition details 
@@ -214,15 +213,15 @@ public:
   ////////////////////////////////////////////////////////////////
   // Global addressing
   ////////////////////////////////////////////////////////////////
-    void GlobalIndexToGlobalCoor(int gidx,std::vector<int> &gcoor){
+  void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
     assert(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
-    void LocalIndexToLocalCoor(int lidx,std::vector<int> &lcoor){
+  void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
     assert(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
-    void GlobalCoorToGlobalIndex(const std::vector<int> & gcoor,int & gidx){
+  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
     gidx=0;
     int mult=1;
     for(int mu=0;mu<_ndimension;mu++) {
@@ -230,7 +229,7 @@ public:
       mult*=_gdimensions[mu];
     }
   }
-    void GlobalCoorToProcessorCoorLocalCoor(std::vector<int> &pcoor,std::vector<int> &lcoor,const std::vector<int> &gcoor)
+  void GlobalCoorToProcessorCoorLocalCoor(Coordinate &pcoor,Coordinate &lcoor,const Coordinate &gcoor)
   {
     pcoor.resize(_ndimension);
     lcoor.resize(_ndimension);
@@ -240,14 +239,14 @@ public:
       lcoor[mu] = gcoor[mu]%_fld;
     }
   }
-    void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const std::vector<int> &gcoor)
+  void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const Coordinate &gcoor)
   {
-      std::vector<int> pcoor;
-      std::vector<int> lcoor;
+    Coordinate pcoor;
+    Coordinate lcoor;
     GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
     rank = RankFromProcessorCoor(pcoor);
     /*
-      std::vector<int> cblcoor(lcoor);
+      Coordinate cblcoor(lcoor);
       for(int d=0;d<cblcoor.size();d++){
       if( this->CheckerBoarded(d) ) {
       cblcoor[d] = lcoor[d]/2;
@@ -258,10 +257,10 @@ public:
     o_idx= oIndex(lcoor);
   }
 
-    void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , std::vector<int> &gcoor)
+  void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , Coordinate &gcoor)
   {
     gcoor.resize(_ndimension);
-      std::vector<int> coor(_ndimension);
+    Coordinate coor(_ndimension);
 
     ProcessorCoorFromRank(rank,coor);
     for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = _ldimensions[mu]*coor[mu];
@@ -273,20 +272,19 @@ public:
     for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += coor[mu];
       
   }
-    void RankIndexCbToFullGlobalCoor(int rank, int o_idx, int i_idx, int cb,std::vector<int> &fcoor)
+  void RankIndexCbToFullGlobalCoor(int rank, int o_idx, int i_idx, int cb,Coordinate &fcoor)
   {
     RankIndexToGlobalCoor(rank,o_idx,i_idx ,fcoor);
     if(CheckerBoarded(0)){
       fcoor[0] = fcoor[0]*2+cb;
     }
   }
-    void ProcessorCoorLocalCoorToGlobalCoor(std::vector<int> &Pcoor,std::vector<int> &Lcoor,std::vector<int> &gcoor)
+  void ProcessorCoorLocalCoorToGlobalCoor(Coordinate &Pcoor,Coordinate &Lcoor,Coordinate &gcoor)
   {
     gcoor.resize(_ndimension);
     for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = Pcoor[mu]*_ldimensions[mu]+Lcoor[mu];
   }
 };
 
-
-}
+NAMESPACE_END(Grid);
 #endif

Grid/cartesian/Cartesian_full.h

@@ -28,13 +28,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CARTESIAN_FULL_H
 #define GRID_CARTESIAN_FULL_H
 
-namespace Grid{
+NAMESPACE_BEGIN(Grid);
     
 /////////////////////////////////////////////////////////////////////////////////////////
 // Grid Support.
 /////////////////////////////////////////////////////////////////////////////////////////
 
-
 class GridCartesian: public GridBase {
 
 public:
@@ -49,7 +48,7 @@ public:
   virtual int CheckerBoarded(int dim){
     return 0;
   }
-    virtual int CheckerBoard(const std::vector<int> &site){
+  virtual int CheckerBoard(const Coordinate &site){
     return 0;
   }
   virtual int CheckerBoardDestination(int cb,int shift,int dim){
@@ -64,16 +63,16 @@ public:
   /////////////////////////////////////////////////////////////////////////
   // Constructor takes a parent grid and possibly subdivides communicator.
   /////////////////////////////////////////////////////////////////////////
-    GridCartesian(const std::vector<int> &dimensions,
-		  const std::vector<int> &simd_layout,
-		  const std::vector<int> &processor_grid,
+  GridCartesian(const Coordinate &dimensions,
+		const Coordinate &simd_layout,
+		const Coordinate &processor_grid,
 		const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
   {
     Init(dimensions,simd_layout,processor_grid);
   }
-    GridCartesian(const std::vector<int> &dimensions,
-		  const std::vector<int> &simd_layout,
-		  const std::vector<int> &processor_grid,
+  GridCartesian(const Coordinate &dimensions,
+		const Coordinate &simd_layout,
+		const Coordinate &processor_grid,
 		const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
   {
     Init(dimensions,simd_layout,processor_grid);
@@ -81,18 +80,18 @@ public:
   /////////////////////////////////////////////////////////////////////////
   // Construct from comm world
   /////////////////////////////////////////////////////////////////////////
-    GridCartesian(const std::vector<int> &dimensions,
-		  const std::vector<int> &simd_layout,
-		  const std::vector<int> &processor_grid) : GridBase(processor_grid)
+  GridCartesian(const Coordinate &dimensions,
+		const Coordinate &simd_layout,
+		const Coordinate &processor_grid) : GridBase(processor_grid)
   {
     Init(dimensions,simd_layout,processor_grid);
   }
 
   virtual ~GridCartesian() = default;
 
-    void Init(const std::vector<int> &dimensions,
-	      const std::vector<int> &simd_layout,
-	      const std::vector<int> &processor_grid)
+  void Init(const Coordinate &dimensions,
+	    const Coordinate &simd_layout,
+	    const Coordinate &processor_grid)
   {
     ///////////////////////
     // Grid information
@@ -170,5 +169,6 @@ public:
   };
 
 };
-}
+
+NAMESPACE_END(Grid);
 #endif

Grid/cartesian/Cartesian_red_black.h

@@ -29,8 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CARTESIAN_RED_BLACK_H
 #define GRID_CARTESIAN_RED_BLACK_H
 
-
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 static const int CbRed  =0;
 static const int CbBlack=1;
@@ -41,7 +40,7 @@ namespace Grid {
 class GridRedBlackCartesian : public GridBase
 {
 public:
-    std::vector<int> _checker_dim_mask;
+  Coordinate _checker_dim_mask;
   int              _checker_dim;
   std::vector<int> _checker_board;
 
@@ -49,7 +48,7 @@ public:
     if( dim==_checker_dim) return 1;
     else return 0;
   }
-    virtual int CheckerBoard(const std::vector<int> &site){
+  virtual int CheckerBoard(const Coordinate &site){
     int linear=0;
     assert(site.size()==_ndimension);
     for(int d=0;d<_ndimension;d++){ 
@@ -59,7 +58,6 @@ public:
     return (linear&0x1);
   }
 
-
   // Depending on the cb of site, we toggle source cb.
   // for block #b, element #e = (b, e)
   // we need 
@@ -83,7 +81,7 @@ public:
   }
   virtual int  CheckerBoardFromOindex (int Oindex)
   {
-      std::vector<int> ocoor;
+    Coordinate ocoor;
     oCoorFromOindex(ocoor,Oindex);
     return CheckerBoard(ocoor);
   }
@@ -118,7 +116,7 @@ public:
   GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
   {
     int dims = base->_ndimension;
-      std::vector<int> checker_dim_mask(dims,1);
+    Coordinate checker_dim_mask(dims,1);
     int checker_dim = 0;
     Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
   };
@@ -127,7 +125,7 @@ public:
   // Create redblack from original grid, with non-trivial checker dim mask
   ////////////////////////////////////////////////////////////
   GridRedBlackCartesian(const GridBase *base,
-			  const std::vector<int> &checker_dim_mask,
+			const Coordinate &checker_dim_mask,
 			int checker_dim
 			) :  GridBase(base->_processors,*base) 
   {
@@ -135,40 +133,11 @@ public:
   }
 
   virtual ~GridRedBlackCartesian() = default;
-#if 0
-    ////////////////////////////////////////////////////////////
-    // Create redblack grid ;; deprecate these. Should not
-    // need direct creation of redblack without a full grid to base on
-    ////////////////////////////////////////////////////////////
-    GridRedBlackCartesian(const GridBase *base,
-			  const std::vector<int> &dimensions,
-			  const std::vector<int> &simd_layout,
-			  const std::vector<int> &processor_grid,
-			  const std::vector<int> &checker_dim_mask,
-			  int checker_dim
-			  ) :  GridBase(processor_grid,*base) 
-    {
-      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
-    }
 
-    ////////////////////////////////////////////////////////////
-    // Create redblack grid
-    ////////////////////////////////////////////////////////////
-    GridRedBlackCartesian(const GridBase *base,
-			  const std::vector<int> &dimensions,
-			  const std::vector<int> &simd_layout,
-			  const std::vector<int> &processor_grid) : GridBase(processor_grid,*base) 
-    {
-      std::vector<int> checker_dim_mask(dimensions.size(),1);
-      int checker_dim = 0;
-      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
-    }
-#endif
-
-    void Init(const std::vector<int> &dimensions,
-              const std::vector<int> &simd_layout,
-              const std::vector<int> &processor_grid,
-              const std::vector<int> &checker_dim_mask,
+  void Init(const Coordinate &dimensions,
+	    const Coordinate &simd_layout,
+	    const Coordinate &processor_grid,
+	    const Coordinate &checker_dim_mask,
 	    int checker_dim)
   {
 
@@ -282,7 +251,7 @@ public:
   };
 
 protected:
-    virtual int oIndex(std::vector<int> &coor)
+  virtual int oIndex(Coordinate &coor)
   {
     int idx = 0;
     for (int d = 0; d < _ndimension; d++)
@@ -299,7 +268,7 @@ public:
     return idx;
   };
 
-    virtual int iIndex(std::vector<int> &lcoor)
+  virtual int iIndex(Coordinate &lcoor)
   {
     int idx = 0;
     for (int d = 0; d < _ndimension; d++)
@@ -316,5 +285,5 @@ public:
     return idx;
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/communicator/Communicator.h

@@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_COMMUNICATOR_H
 #define GRID_COMMUNICATOR_H
 
+#include <Grid/util/Coordinate.h>
 #include <Grid/communicator/SharedMemory.h>
 #include <Grid/communicator/Communicator_base.h>
 

Grid/communicator/Communicator_base.cc

@@ -31,7 +31,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <limits.h>
 #include <sys/mman.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
@@ -47,8 +47,8 @@ int                      CartesianCommunicator::Dimensions(void)        { return
 int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
 int                      CartesianCommunicator::BossRank(void)          { return 0; };
 int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
-const std::vector<int> & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
-const std::vector<int> & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
+const Coordinate & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
+const Coordinate & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
 int                      CartesianCommunicator::ProcessorCount(void)    { return _Nprocessors; };
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -72,5 +72,6 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
   GlobalSumVector((double *)c,2*N);
 }
   
-}
+NAMESPACE_END(Grid);
+
 

Grid/communicator/Communicator_base.h

@@ -34,7 +34,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 class CartesianCommunicator : public SharedMemory {
 
@@ -52,9 +52,9 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  std::vector<int> _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  std::vector<int> _processor_coor;  // linear processor coordinate
+  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
@@ -69,8 +69,8 @@ public:
   // Constructors to sub-divide a parent communicator
   // and default to comm world
   ////////////////////////////////////////////////
-  CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank);
-  CartesianCommunicator(const std::vector<int> &pdimensions_in);
+  CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank);
+  CartesianCommunicator(const Coordinate &pdimensions_in);
   virtual ~CartesianCommunicator();
 
 private:
@@ -79,7 +79,7 @@ public:
   // Private initialise from an MPI communicator
   // Can use after an MPI_Comm_split, but hidden from user so private
   ////////////////////////////////////////////////
-  void InitFromMPICommunicator(const std::vector<int> &processors, Grid_MPI_Comm communicator_base);
+  void InitFromMPICommunicator(const Coordinate &processors, Grid_MPI_Comm communicator_base);
 
 public:
   
@@ -88,15 +88,15 @@ public:
   // Wraps MPI_Cart routines, or implements equivalent on other impls
   ////////////////////////////////////////////////////////////////////////////////////////
   void ShiftedRanks(int dim,int shift,int & source, int & dest);
-  int  RankFromProcessorCoor(std::vector<int> &coor);
-  void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
+  int  RankFromProcessorCoor(Coordinate &coor);
+  void ProcessorCoorFromRank(int rank,Coordinate &coor);
   
   int                      Dimensions(void)        ;
   int                      IsBoss(void)            ;
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
-  const std::vector<int> & ThisProcessorCoor(void) ;
-  const std::vector<int> & ProcessorGrid(void)     ;
+  const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ProcessorGrid(void)     ;
   int                      ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -199,9 +199,10 @@ public:
   template<class obj> void Broadcast(int root,obj &data)
   {
     Broadcast(root,(void *)&data,sizeof(data));
-    };
-
-}; 
   }
 
+}; 
+
+NAMESPACE_END(Grid);
+
 #endif

Grid/communicator/Communicator_mpi3.cc

@@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/communicator/SharedMemory.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
@@ -44,21 +44,26 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
     MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+
     //If 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) ||
-        (nCommThreads > 1 && provided != MPI_THREAD_MULTIPLE) )
+    if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
       assert(0);
     }
 
-  Grid_quiesce_nodes();
+    if( (nCommThreads > 1) && (provided != MPI_THREAD_MULTIPLE) ) {
+      assert(0);
+    }
+  }
 
   // Never clean up as done once.
   MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
 
+  Grid_quiesce_nodes();
   GlobalSharedMemory::Init(communicator_world);
   GlobalSharedMemory::SharedMemoryAllocate(
 		   GlobalSharedMemory::MAX_MPI_SHM_BYTES,
 		   GlobalSharedMemory::Hugepages);
+  Grid_unquiesce_nodes();
 }
 
 ///////////////////////////////////////////////////////////////////////////
@@ -69,14 +74,14 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
   assert(ierr==0);
 }
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
+int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor)
 {
   int rank;
   int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
   assert(ierr==0);
   return rank;
 }
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 {
   coor.resize(_ndimension);
   int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
@@ -86,7 +91,7 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Initialises from communicator_world
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) 
 {
   MPI_Comm optimal_comm;
   ////////////////////////////////////////////////////
@@ -105,13 +110,12 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)    
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)    
 {
-  _ndimension = processors.size();
-
+  _ndimension = processors.size();  assert(_ndimension>=1);
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
-  std::vector<int> parent_processor_coor(_ndimension,0);
-  std::vector<int> parent_processors    (_ndimension,1);
+  Coordinate parent_processor_coor(_ndimension,0);
+  Coordinate parent_processors    (_ndimension,1);
 
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
@@ -124,10 +128,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   //  int Nparent = parent._processors ; 
-  //  std::cout << " splitting from communicator "<<parent.communicator <<std::endl;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
-  //  std::cout << " Parent size  "<<Nparent <<std::endl;
 
   int childsize=1;
   for(int d=0;d<processors.size();d++) {
@@ -136,11 +138,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   int Nchild = Nparent/childsize;
   assert (childsize * Nchild == Nparent);
 
-  //  std::cout << " child size  "<<childsize <<std::endl;
-
-  std::vector<int> ccoor(_ndimension); // coor within subcommunicator
-  std::vector<int> scoor(_ndimension); // coor of split within parent
-  std::vector<int> ssize(_ndimension); // coor of split within parent
+  Coordinate ccoor(_ndimension); // coor within subcommunicator
+  Coordinate scoor(_ndimension); // coor of split within parent
+  Coordinate ssize(_ndimension); // coor of split within parent
 
   for(int d=0;d<_ndimension;d++){
     ccoor[d] = parent_processor_coor[d] % processors[d];
@@ -157,36 +157,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   MPI_Comm comm_split;
   if ( Nchild > 1 ) { 
 
-    if(0){
-      std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
-      std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"]    ";
-      for(int d=0;d<parent._ndimension;d++)  std::cout << parent._processors[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" child grid["<< _ndimension <<"]    ";
-      for(int d=0;d<processors.size();d++)  std::cout << processors[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"]    ";
-      for(int d=0;d<parent._ndimension;d++)  std::cout << parent._processor_coor[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"]    ";
-      for(int d=0;d<processors.size();d++)  std::cout << scoor[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"]    ";
-      for(int d=0;d<processors.size();d++)  std::cout << ccoor[d] << " ";
-      std::cout<<std::endl;
-
-      //////////////////////////////////////////////////////////////////////////////////////////////////////
-      // Declare victory
-      //////////////////////////////////////////////////////////////////////////////////////////////////////
-      std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
-		<< Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
-      std::cout << " Split communicator " <<comm_split <<std::endl;
-    }
-
     ////////////////////////////////////////////////////////////////
     // Split the communicator
     ////////////////////////////////////////////////////////////////
@@ -225,7 +195,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   }
 }
 
-void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
+void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors, MPI_Comm communicator_base)
 {
   ////////////////////////////////////////////////////
   // Creates communicator, and the communicator_halo
@@ -242,7 +212,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &proc
     _Nprocessors*=_processors[i];
   }
 
-  std::vector<int> periodic(_ndimension,1);
+  Coordinate periodic(_ndimension,1);
   MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
   MPI_Comm_rank(communicator,&_processor);
   MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
@@ -479,7 +449,7 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
-  std::vector<int> row(_ndimension,1);
+  Coordinate row(_ndimension,1);
   assert(dim>=0 && dim<_ndimension);
 
   //  Split the communicator
@@ -508,7 +478,6 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   MPI_Type_free(&object);
 }
 
+NAMESPACE_END(Grid);
 
 
-}
-

Grid/communicator/Communicator_none.cc

@@ -27,7 +27,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #include <Grid/GridCore.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
@@ -42,17 +42,17 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 					   GlobalSharedMemory::Hugepages);
 }
 
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
   srank=0;
   SetCommunicator(communicator_world);
 }
 
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   _processors = processors;
-  _ndimension = processors.size();
+  _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
   
   // Require 1^N processor grid for fake
@@ -122,8 +122,8 @@ int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
-int  CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) {  return 0;}
-void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){  coor = _processor_coor; }
+int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
+void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   source =0;
@@ -160,6 +160,6 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
 
 void CartesianCommunicator::StencilBarrier(void){};
 
+NAMESPACE_END(Grid);
 
-}
 

Grid/communicator/SharedMemory.cc

@@ -28,10 +28,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 #include <Grid/GridCore.h>
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
 
 // static data
 
+int                 GlobalSharedMemory::HPEhypercube = 1;
 uint64_t            GlobalSharedMemory::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
 int                 GlobalSharedMemory::Hugepages = 0;
 int                 GlobalSharedMemory::_ShmSetup;
@@ -76,6 +77,7 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
     std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
     assert(heap_bytes<heap_size);
   }
+  //std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;
   return ptr;
 }
 void SharedMemory::ShmBufferFreeAll(void) { 
@@ -84,9 +86,9 @@ void SharedMemory::ShmBufferFreeAll(void) {
 }
 void *SharedMemory::ShmBufferSelf(void)
 {
+  //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
   return ShmCommBufs[ShmRank];
 }
 
+NAMESPACE_END(Grid); 
 
-
-}

Grid/communicator/SharedMemory.h

@@ -25,18 +25,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-
-
-// TODO
-// 1) move includes into SharedMemory.cc
-//
-// 2) split shared memory into a) optimal communicator creation from comm world
-// 
-//                             b) shared memory buffers container
-//                                -- static globally shared; init once
-//                                -- per instance set of buffers.
-//                                   
-
 #pragma once 
 
 #include <Grid/GridCore.h>
@@ -57,7 +45,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <numaif.h>
 #endif
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 #if defined (GRID_COMMS_MPI3) 
 typedef MPI_Comm    Grid_MPI_Comm;
@@ -71,12 +59,18 @@ class GlobalSharedMemory {
 private:
   static const int     MAXLOG2RANKSPERNODE = 16;            
 
+
   // Init once lock on the buffer allocation
   static int      _ShmSetup;
   static int      _ShmAlloc;
   static uint64_t _ShmAllocBytes;
 
 public:
+  ///////////////////////////////////////
+  // HPE 8600 hypercube optimisation
+  ///////////////////////////////////////
+  static int HPEhypercube;
+
   static int      ShmSetup(void)      { return _ShmSetup; }
   static int      ShmAlloc(void)      { return _ShmAlloc; }
   static uint64_t ShmAllocBytes(void) { return _ShmAllocBytes; }
@@ -102,12 +96,16 @@ class GlobalSharedMemory {
   // Create an optimal reordered communicator that makes MPI_Cart_create get it right
   //////////////////////////////////////////////////////////////////////////////////////
   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
   ///////////////////////////////////////////////////
   // Provide shared memory facilities off comm world
   ///////////////////////////////////////////////////
   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 SharedMemoryZero(void *dest,size_t bytes);
 
 };
 
@@ -148,6 +146,7 @@ class SharedMemory
   // Call on any instance
   ///////////////////////////////////////////////////
   void SharedMemoryTest(void);
+  
   void *ShmBufferSelf(void);
   void *ShmBuffer    (int rank);
   void *ShmBufferTranslate(int rank,void * local_p);
@@ -162,4 +161,5 @@ class SharedMemory
 
 };
 
-}
+NAMESPACE_END(Grid);
+

Grid/communicator/SharedMemoryMPI.cc

@@ -29,8 +29,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
-namespace Grid { 
+#ifdef GRID_NVCC
+#include <cuda_runtime_api.h>
+#endif
 
+NAMESPACE_BEGIN(Grid); 
+#define header "SharedMemoryMpi: "
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -46,6 +50,11 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
   MPI_Comm_rank(WorldShmComm     ,&WorldShmRank);
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
+
+  if ( WorldRank == 0) {
+    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+  }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
   // WorldNodes
@@ -130,9 +139,24 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+{
+  //////////////////////////////////////////////////////////////////////////////
+  // Look and see if it looks like an HPE 8600 based on hostname conventions
+  //////////////////////////////////////////////////////////////////////////////
+  const int namelen = _POSIX_HOST_NAME_MAX;
+  char name[namelen];
+  int R;
+  int I;
+  int N;
+  gethostname(name,namelen);
+  int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
+
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
+}
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
-#ifdef HYPERCUBE
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
   ////////////////////////////////////////////////////////////////
@@ -173,9 +197,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   }
 
   std::string hname(name);
-  std::cout << "hostname "<<hname<<std::endl;
-  std::cout << "R " << R << " I " << I << " N "<< N
-            << " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
+  //  std::cout << "hostname "<<hname<<std::endl;
+  //  std::cout << "R " << R << " I " << I << " N "<< N
+  //            << " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
 
   //////////////////////////////////////////////////////////////////
   // broadcast node 0's base coordinate for this partition.
@@ -198,7 +222,8 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   ////////////////////////////////////////////////////////////////
   int ndimension              = processors.size();
   std::vector<int> processor_coor(ndimension);
-  std::vector<int> WorldDims = processors;   std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
+  std::vector<int> WorldDims = processors.toVector();
+  std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
   std::vector<int> ShmCoor  (ndimension);    std::vector<int> NodeCoor (ndimension);    std::vector<int> WorldCoor(ndimension);
   std::vector<int> HyperCoor(ndimension);
   int dim = 0;
@@ -253,7 +278,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
-#else 
+}
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+{
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
   ////////////////////////////////////////////////////////////////
@@ -265,9 +292,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   // in a maximally symmetrical way
   ////////////////////////////////////////////////////////////////
   int ndimension              = processors.size();
-  std::vector<int> processor_coor(ndimension);
-  std::vector<int> WorldDims = processors;   std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
-  std::vector<int> ShmCoor  (ndimension);    std::vector<int> NodeCoor (ndimension);    std::vector<int> WorldCoor(ndimension);
+  Coordinate processor_coor(ndimension);
+  Coordinate WorldDims = processors; Coordinate ShmDims(ndimension,1);  Coordinate NodeDims (ndimension);
+  Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
   int dim = 0;
   for(int l2=0;l2<log2size;l2++){
     while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
@@ -306,7 +333,6 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
-#endif
 }
 ////////////////////////////////////////////////////////////////////////////////////////////
 // SHMGET
@@ -314,7 +340,7 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -337,7 +363,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
         int errsv = errno;
         printf("Errno %d\n",errsv);
         printf("key   %d\n",key);
-        printf("size  %lld\n",size);
+        printf("size  %ld\n",size);
         printf("flags %d\n",flags);
         perror("shmget");
         exit(1);
@@ -373,10 +399,97 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_NVCC
+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);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // TODO/FIXME : NOT ALL NVLINK BOARDS have full Peer to peer connectivity.
+  // The annoyance is that they have partial peer 2 peer. This occurs on the 8 GPU blades.
+  // e.g. DGX1, supermicro board, 
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  //  cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
+  cudaSetDevice(WorldShmRank);
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // 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);  
+  }
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+  if ( WorldRank == 0 ){
+    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+  }
+  SharedMemoryZero(ShmCommBuf,bytes);
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Loop over ranks/gpu's on our node
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  for(int r=0;r<WorldShmSize;r++){
+    
+    //////////////////////////////////////////////////
+    // If it is me, pass around the IPC access key
+    //////////////////////////////////////////////////
+    cudaIpcMemHandle_t handle;
+
+    if ( r==WorldShmRank ) { 
+      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      if ( err !=  cudaSuccess) {
+	std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+    //////////////////////////////////////////////////
+    // Share this IPC handle across the Shm Comm
+    //////////////////////////////////////////////////
+    { 
+      int ierr=MPI_Bcast(&handle,
+			 sizeof(handle),
+			 MPI_BYTE,
+			 r,
+			 WorldShmComm);
+      assert(ierr==0);
+    }
+    
+    ///////////////////////////////////////////////////////////////
+    // If I am not the source, overwrite thisBuf with remote buffer
+    ///////////////////////////////////////////////////////////////
+    void * thisBuf = ShmCommBuf;
+    if ( r!=WorldShmRank ) { 
+      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      if ( err !=  cudaSuccess) {
+	std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+    ///////////////////////////////////////////////////////////////
+    // Save a copy of the device buffers
+    ///////////////////////////////////////////////////////////////
+    WorldShmCommBufs[r] = thisBuf;
+  }
+
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+}
+#else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -413,7 +526,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -423,7 +536,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -455,7 +568,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -470,7 +583,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
@@ -499,7 +612,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
+      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -536,10 +649,27 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _ShmAllocBytes = bytes;
 }
 #endif
+#endif // End NVCC case for GPU device buffers
 
-
-
-
+/////////////////////////////////////////////////////////////////////////
+// Routines accessing shared memory should route through for GPU safety
+/////////////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
+{
+#ifdef GRID_NVCC
+  cudaMemset(dest,0,bytes);
+#else
+  bzero(dest,bytes);
+#endif
+}
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
+{
+#ifdef GRID_NVCC
+  cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
+#else   
+  bcopy(src,dest,bytes);
+#endif
+}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
@@ -571,7 +701,6 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
     MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
 
     ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
-    //    std::cout << "SetCommunicator ShmCommBufs ["<< r<< "] = "<< ShmCommBufs[r]<< "  wsr = "<<wsr<<std::endl;
   }
   ShmBufferFreeAll();
 
@@ -584,6 +713,8 @@ 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]); 
+
+  SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier
@@ -598,25 +729,27 @@ void SharedMemory::ShmBarrier(void)
 void SharedMemory::SharedMemoryTest(void)
 {
   ShmBarrier();
+  uint64_t check[3];
+  uint64_t magic = 0x5A5A5A;
   if ( ShmRank == 0 ) {
-    for(int r=0;r<ShmSize;r++){
-      uint64_t * check = (uint64_t *) ShmCommBufs[r];
+    for(uint64_t r=0;r<ShmSize;r++){
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
-      check[2] = 0x5A5A5A;
+       check[2]=magic;
+       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
-  for(int r=0;r<ShmSize;r++){
-    uint64_t * check = (uint64_t *) ShmCommBufs[r];
-    
+  for(uint64_t r=0;r<ShmSize;r++){
+    ShmBarrier();
+    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
+    ShmBarrier();
     assert(check[0]==GlobalSharedMemory::WorldNode);
     assert(check[1]==r);
-    assert(check[2]==0x5A5A5A);
-    
-  }
+    assert(check[2]==magic);
     ShmBarrier();
   }
+}
 
 void *SharedMemory::ShmBuffer(int rank)
 {
@@ -629,7 +762,6 @@ void *SharedMemory::ShmBuffer(int rank)
 }
 void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 {
-  static int count =0;
   int gpeer = ShmRanks[rank];
   assert(gpeer!=ShmRank); // never send to self
   if (gpeer == MPI_UNDEFINED){
@@ -648,4 +780,5 @@ SharedMemory::~SharedMemory()
   }
 };
 
-}
+NAMESPACE_END(Grid); 
+

Grid/communicator/SharedMemoryNone.cc

@@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 #include <Grid/GridCore.h>
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
 
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@@ -47,7 +47,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   _ShmSetup=1;
 }
 
-void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   optimal_comm = WorldComm;
 }
@@ -125,4 +125,5 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 SharedMemory::~SharedMemory()
 {};
 
-}
+NAMESPACE_END(Grid); 
+

Grid/cshift/Cshift_common.h

@@ -25,10 +25,9 @@ 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_CSHIFT_COMMON_H_
-#define _GRID_CSHIFT_COMMON_H_
+#pragma once
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
@@ -36,20 +35,21 @@ namespace Grid {
 template<class vobj> void 
 Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
-  int rd = rhs._grid->_rdimensions[dimension];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask = 0x3;
   }
   
-  int so=plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
-  int e1=rhs._grid->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
+  int so=plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
+  int e1=rhs.Grid()->_slice_nblock[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
   int ent = 0;
 
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+  static Vector<std::pair<int,int> > table; table.resize(e1*e2);
+  int stride=rhs.Grid()->_slice_stride[dimension];
 
-  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++){
@@ -63,66 +63,68 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
      for(int n=0;n<e1;n++){
        for(int b=0;b<e2;b++){
 	 int o  = n*stride;
-	 int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
+	 int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	 if ( ocb &cbmask ) {
 	   table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
 	 }
        }
      }
   }
-  parallel_for(int i=0;i<ent;i++){
-    buffer[table[i].first]=rhs._odata[table[i].second];
-  }
+  thread_for(i,ent,{
+    buffer[table[i].first]=rhs_v[table[i].second];
+  });
 }
 
 ///////////////////////////////////////////////////////////////////
 // Gather for when there *is* need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
+Gather_plane_extract(const Lattice<vobj> &rhs,
+		     ExtractPointerArray<typename vobj::scalar_object> pointers,
+		     int dimension,int plane,int cbmask)
 {
-  int rd = rhs._grid->_rdimensions[dimension];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask = 0x3;
   }
 
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
 
-  int e1=rhs._grid->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
-  int n1=rhs._grid->_slice_stride[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
+  int n1=rhs.Grid()->_slice_stride[dimension];
 
+  auto rhs_v = rhs.View();
   if ( cbmask ==0x3){
-    parallel_for_nest2(int n=0;n<e1;n++){
+    thread_for_collapse(2,n,e1,{
       for(int b=0;b<e2;b++){
 
 	int o      =   n*n1;
 	int offset = b+n*e2;
 	
-	vobj temp =rhs._odata[so+o+b];
+	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
-
-      }
       }
+    });
   } else { 
 
     // 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;
-    parallel_for_nest2(int n=0;n<e1;n++){
+    thread_for_collapse(2,n,e1,{
       for(int b=0;b<e2;b++){
 
 	int o=n*n1;
-	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
+	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	int offset = b+n*e2;
 
 	if ( ocb & cbmask ) {
-	  vobj temp =rhs._odata[so+o+b];
+	  vobj temp =rhs_v[so+o+b];
 	  extract<vobj>(temp,pointers,offset);
 	}
       }
-    }
+    });
   }
 }
 
@@ -131,17 +133,17 @@ Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_
 //////////////////////////////////////////////////////
 template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
-  int rd = rhs._grid->_rdimensions[dimension];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
     
-  int e1=rhs._grid->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
-  int stride=rhs._grid->_slice_stride[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
+  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);
   int ent    =0;
@@ -150,8 +152,8 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
-	int o   =n*rhs._grid->_slice_stride[dimension];
-	int bo  =n*rhs._grid->_slice_block[dimension];
+	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);
       }
     }
@@ -160,8 +162,8 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
     int bo=0;
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
-	int o   =n*rhs._grid->_slice_stride[dimension];
-	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
+	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++);
 	}
@@ -169,48 +171,51 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
     }
   }
   
-  parallel_for(int i=0;i<ent;i++){
-    rhs._odata[table[i].first]=buffer[table[i].second];
-  }
+  auto rhs_v = rhs.View();
+  thread_for(i,ent,{
+    rhs_v[table[i].first]=buffer[table[i].second];
+  });
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there *is* need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerArray<typename vobj::scalar_object> pointers,int dimension,int plane,int cbmask)
 {
-  int rd = rhs._grid->_rdimensions[dimension];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
     
-  int e1=rhs._grid->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
-    parallel_for_nest2(int n=0;n<e1;n++){
+    auto rhs_v = rhs.View();
+    thread_for_collapse(2,n,e1,{
       for(int b=0;b<e2;b++){
-	int o      = n*rhs._grid->_slice_stride[dimension];
-	int offset = b+n*rhs._grid->_slice_block[dimension];
-	merge(rhs._odata[so+o+b],pointers,offset);
-      }
+	int o      = n*rhs.Grid()->_slice_stride[dimension];
+	int offset = b+n*rhs.Grid()->_slice_block[dimension];
+	merge(rhs_v[so+o+b],pointers,offset);
       }
+    });
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
     // Test_cshift_red_black code.
     //    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
+    auto rhs_v = rhs.View();
     for(int n=0;n<e1;n++){
       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 ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
+	int o      = n*rhs.Grid()->_slice_stride[dimension];
+	int offset = b+n*rhs.Grid()->_slice_block[dimension];
+	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	if ( ocb&cbmask ) {
-	  merge(rhs._odata[so+o+b],pointers,offset);
+	  merge(rhs_v[so+o+b],pointers,offset);
 	}
       }
     }
@@ -222,18 +227,18 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<typ
 //////////////////////////////////////////////////////
 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];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int ro  = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
-  int lo  = lplane*lhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int ro  = rplane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
+  int lo  = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane 
 
-  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];
+  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);
   int ent=0;
 
@@ -248,7 +253,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int o =n*stride+b;
-        int ocb=1<<lhs._grid->CheckerBoardFromOindex(o);
+        int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o);
         if ( ocb&cbmask ) {
 	  table[ent++] = std::pair<int,int>(lo+o,ro+o);
 	}
@@ -256,32 +261,33 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
     }
   }
 
-  parallel_for(int i=0;i<ent;i++){
-    lhs._odata[table[i].first]=rhs._odata[table[i].second];
-  }
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for(i,ent,{
+    lhs_v[table[i].first]=rhs_v[table[i].second];
+  });
 
 }
 
 template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask,int permute_type)
 {
  
-  int rd = rhs._grid->_rdimensions[dimension];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int ro  = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
-  int lo  = lplane*lhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int ro  = rplane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
+  int lo  = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane 
 
-  int e1=rhs._grid->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block [dimension];
-  int stride = rhs._grid->_slice_stride[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
+  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);
   int ent=0;
 
-  double t_tab,t_perm;
   if ( cbmask == 0x3 ) {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
@@ -292,14 +298,16 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
     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);
+      int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o+b);
       if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
   }
 
-  parallel_for(int i=0;i<ent;i++){
-    permute(lhs._odata[table[i].first],rhs._odata[table[i].second],permute_type);
-  }
+  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);
+  });
 }
 
 //////////////////////////////////////////////////////
@@ -309,10 +317,8 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
 {
   int sshift[2];
 
-  sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
-  sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
-
-  double t_local;
+  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
+  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
   if ( sshift[0] == sshift[1] ) {
     Cshift_local(ret,rhs,dimension,shift,0x3);
@@ -324,7 +330,7 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
 
 template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
-  GridBase *grid = rhs._grid;
+  GridBase *grid = rhs.Grid();
   int fd = grid->_fdimensions[dimension];
   int rd = grid->_rdimensions[dimension];
   int ld = grid->_ldimensions[dimension];
@@ -335,18 +341,18 @@ template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &r
   shift = (shift+fd)%fd;
 
   // the permute type
-  ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
+  ret.Checkerboard() = grid->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
   int permute_dim =grid->PermuteDim(dimension);
   int permute_type=grid->PermuteType(dimension);
   int permute_type_dist;
 
   for(int x=0;x<rd;x++){       
 
-    int o   = 0;
+    //    int o   = 0;
     int bo  = x * grid->_ostride[dimension];
     int cb= (cbmask==0x2)? Odd : Even;
 
-    int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
+    int sshift = grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
     int sx     = (x+sshift)%rd;
     
     // wrap is whether sshift > rd.
@@ -387,5 +393,5 @@ template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &r
   
   }
 }
-}
-#endif
+NAMESPACE_END(Grid);
+

Grid/cshift/Cshift_mpi.h

@@ -30,27 +30,27 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define _GRID_CSHIFT_MPI_H_
 
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
 
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
   typedef typename vobj::vector_type vector_type;
   typedef typename vobj::scalar_type scalar_type;
 
-  Lattice<vobj> ret(rhs._grid); 
+  Lattice<vobj> ret(rhs.Grid()); 
   
-  int fd = rhs._grid->_fdimensions[dimension];
-  int rd = rhs._grid->_rdimensions[dimension];
+  int fd = rhs.Grid()->_fdimensions[dimension];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
   // Map to always positive shift modulo global full dimension.
   shift = (shift+fd)%fd;
 
-  ret.checkerboard = rhs._grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
+  ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
         
   // the permute type
-  int simd_layout     = rhs._grid->_simd_layout[dimension];
-  int comm_dim        = rhs._grid->_processors[dimension] >1 ;
-  int splice_dim      = rhs._grid->_simd_layout[dimension]>1 && (comm_dim);
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
+  int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
 
 
   if ( !comm_dim ) {
@@ -70,10 +70,10 @@ template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &r
 {
   int sshift[2];
 
-  sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
-  sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
+  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
+  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
-  //  std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
+  //  std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
   if ( sshift[0] == sshift[1] ) {
     //    std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift,0x3);
@@ -88,8 +88,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
 {
   int sshift[2];
 
-  sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
-  sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
+  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
+  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
   //std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
   if ( sshift[0] == sshift[1] ) {
@@ -107,25 +107,25 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   typedef typename vobj::vector_type vector_type;
   typedef typename vobj::scalar_type scalar_type;
 
-  GridBase *grid=rhs._grid;
-  Lattice<vobj> temp(rhs._grid);
+  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 ;
+  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];
+  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);
 
   int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
+  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
 
   for(int x=0;x<rd;x++){       
 
@@ -145,7 +145,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
 
-      int rank           = grid->_processor;
+      //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
@@ -165,7 +165,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
-  GridBase *grid=rhs._grid;
+  GridBase *grid=rhs.Grid();
   const int Nsimd = grid->Nsimd();
   typedef typename vobj::vector_type vector_type;
   typedef typename vobj::scalar_object scalar_object;
@@ -193,21 +193,21 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   // 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);
+  //  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) );
 
   int bytes = buffer_size*sizeof(scalar_object);
 
-  std::vector<scalar_object *>  pointers(Nsimd); // 
-  std::vector<scalar_object *> rpointers(Nsimd); // received pointers
+  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);
+  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
 
   // loop over outer coord planes orthog to dim
   for(int x=0;x<rd;x++){       
@@ -258,5 +258,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   }
 
 }
-}
+
+NAMESPACE_END(Grid); 
+
 #endif

Grid/cshift/Cshift_none.h

@@ -27,13 +27,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #ifndef _GRID_CSHIFT_NONE_H_
 #define _GRID_CSHIFT_NONE_H_
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
-  Lattice<vobj> ret(rhs._grid);
-  ret.checkerboard = rhs._grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
+  Lattice<vobj> ret(rhs.Grid());
+  ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
   Cshift_local(ret,rhs,dimension,shift);
   return ret;
 }
-}
+NAMESPACE_END(Grid);
+
 #endif

Grid/json/json.hpp

@@ -1,3 +1,4 @@
+#ifndef __NVCC__
 /*
     __ _____ _____ _____
  __|  |   __|     |   | |  JSON for Modern C++
@@ -18918,3 +18919,4 @@ inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std
 
 
 #endif
+#endif

Grid/lattice/Lattice.h

@@ -25,9 +25,22 @@ 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_LATTICE_H
-#define GRID_LATTICE_H
-
+#pragma once
 #include <Grid/lattice/Lattice_base.h>
+#include <Grid/lattice/Lattice_conformable.h>
+#include <Grid/lattice/Lattice_ET.h>
+#include <Grid/lattice/Lattice_arith.h>
+#include <Grid/lattice/Lattice_trace.h>
+#include <Grid/lattice/Lattice_transpose.h>
+#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_comparison_utils.h>
+#include <Grid/lattice/Lattice_comparison.h>
+#include <Grid/lattice/Lattice_coordinate.h>
+//#include <Grid/lattice/Lattice_where.h>
+#include <Grid/lattice/Lattice_rng.h>
+#include <Grid/lattice/Lattice_unary.h>
+#include <Grid/lattice/Lattice_transfer.h>
 
-#endif

Grid/lattice/Lattice_ET.h

@@ -36,13 +36,13 @@ directory
 #include <typeinfo>
 #include <vector>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ////////////////////////////////////////////////////
 // Predicated where support
 ////////////////////////////////////////////////////
 template <class iobj, class vobj, class robj>
-inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
+accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
                             const robj &iffalse) {
   typename std::remove_const<vobj>::type ret;
 
@@ -51,11 +51,10 @@ inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
   typedef typename vobj::vector_type vector_type;
 
   const int Nsimd = vobj::vector_type::Nsimd();
-  const int words = sizeof(vobj) / sizeof(vector_type);
 
-  std::vector<Integer> mask(Nsimd);
-  std::vector<scalar_object> truevals(Nsimd);
-  std::vector<scalar_object> falsevals(Nsimd);
+  ExtractBuffer<Integer> mask(Nsimd);
+  ExtractBuffer<scalar_object> truevals(Nsimd);
+  ExtractBuffer<scalar_object> falsevals(Nsimd);
 
   extract(iftrue, truevals);
   extract(iffalse, falsevals);
@@ -69,149 +68,139 @@ inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
   return ret;
 }
 
-////////////////////////////////////////////
-// recursive evaluation of expressions; Could
-// switch to generic approach with variadics, a la
-// Antonin's Lat Sim but the repack to variadic with popped
-// from tuple is hideous; C++14 introduces std::make_index_sequence for this
-////////////////////////////////////////////
-
-// leaf eval of lattice ; should enable if protect using traits
-
-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 <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
-
+/////////////////////////////////////////////////////
 //Specialization of getVectorType for lattices
+/////////////////////////////////////////////////////
 template<typename T>
 struct getVectorType<Lattice<T> >{
   typedef typename Lattice<T>::vector_object type;
 };
 
-template<class sobj>
-inline sobj eval(const unsigned int ss, const sobj &arg)
+////////////////////////////////////////////
+//--  recursive evaluation of expressions; --
+// handle leaves of syntax tree
+///////////////////////////////////////////////////
+template<class sobj> accelerator_inline 
+sobj eval(const uint64_t ss, const sobj &arg)
 {
   return arg;
 }
-template <class lobj>
-inline const lobj &eval(const unsigned int ss, const Lattice<lobj> &arg) {
-  return arg._odata[ss];
+
+template <class lobj> accelerator_inline 
+const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg) 
+{
+  return arg[ss];
+}
+template <class lobj> accelerator_inline 
+const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg) 
+{
+  auto view = arg.View();
+  return view[ss];
 }
 
-// handle nodes in syntax tree
-template <typename Op, typename T1>
-auto inline eval(
-    const unsigned int ss,
-    const LatticeUnaryExpression<Op, T1> &expr)  // eval one operand
-    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)))) {
-  return expr.first.func(eval(ss, std::get<0>(expr.second)));
+///////////////////////////////////////////////////
+// handle nodes in syntax tree- eval one operand
+///////////////////////////////////////////////////
+template <typename Op, typename T1> accelerator_inline 
+auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
+  -> decltype(expr.op.func( eval(ss, expr.arg1)))
+{
+  return expr.op.func( eval(ss, expr.arg1) );
 }
-
-template <typename Op, typename T1, typename T2>
-auto inline eval(
-    const unsigned int ss,
-    const LatticeBinaryExpression<Op, T1, T2> &expr)  // eval two operands
-    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
-                                eval(ss, std::get<1>(expr.second)))) {
-  return expr.first.func(eval(ss, std::get<0>(expr.second)),
-                         eval(ss, std::get<1>(expr.second)));
+///////////////////////
+// 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) );
 }
-
-template <typename Op, typename T1, typename T2, typename T3>
-auto inline eval(const unsigned int ss,
-                 const LatticeTrinaryExpression<Op, T1, T2, T3>
-                     &expr)  // eval three operands
-    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
-                                eval(ss, std::get<1>(expr.second)),
-                                eval(ss, std::get<2>(expr.second)))) {
-  return expr.first.func(eval(ss, std::get<0>(expr.second)),
-                         eval(ss, std::get<1>(expr.second)),
-                         eval(ss, std::get<2>(expr.second)));
+///////////////////////
+// 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)))
+{
+  return expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3));
 }
 
 //////////////////////////////////////////////////////////////////////////
 // Obtain the grid from an expression, ensuring conformable. This must follow a
-// tree recursion
+// tree recursion; must retain grid pointer in the LatticeView class which sucks
+// Use a different method, and make it void *.
+// Perhaps a conformable method.
 //////////////////////////////////////////////////////////////////////////
-template <class T1,
-          typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-inline void GridFromExpression(GridBase *&grid, const T1 &lat)  // Lattice leaf
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+accelerator_inline void GridFromExpression(GridBase *&grid, const T1 &lat)  // Lattice leaf
 {
-  if (grid) {
-    conformable(grid, lat._grid);
+  lat.Conformable(grid);
 }
-  grid = lat._grid;
-}
-template <class T1,
-          typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-inline void GridFromExpression(GridBase *&grid,
-                               const T1 &notlat)  // non-lattice leaf
+
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+accelerator_inline 
+void GridFromExpression(GridBase *&grid,const T1 &notlat)  // non-lattice leaf
 {}
+
 template <typename Op, typename T1>
-inline void GridFromExpression(GridBase *&grid,
-                               const LatticeUnaryExpression<Op, T1> &expr) {
-  GridFromExpression(grid, std::get<0>(expr.second));  // recurse
+accelerator_inline 
+void GridFromExpression(GridBase *&grid,const LatticeUnaryExpression<Op, T1> &expr) 
+{
+  GridFromExpression(grid, expr.arg1);  // recurse
 }
 
 template <typename Op, typename T1, typename T2>
-inline void GridFromExpression(
-    GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) {
-  GridFromExpression(grid, std::get<0>(expr.second));  // recurse
-  GridFromExpression(grid, std::get<1>(expr.second));
+accelerator_inline 
+void GridFromExpression(GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) 
+{
+  GridFromExpression(grid, expr.arg1);  // recurse
+  GridFromExpression(grid, expr.arg2);
 }
 template <typename Op, typename T1, typename T2, typename T3>
-inline void GridFromExpression(
-    GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
-  GridFromExpression(grid, std::get<0>(expr.second));  // recurse
-  GridFromExpression(grid, std::get<1>(expr.second));
-  GridFromExpression(grid, std::get<2>(expr.second));
+accelerator_inline 
+void GridFromExpression(GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
+{
+  GridFromExpression(grid, expr.arg1);  // recurse
+  GridFromExpression(grid, expr.arg2);  // recurse
+  GridFromExpression(grid, expr.arg3);  // recurse
 }
 
 //////////////////////////////////////////////////////////////////////////
 // Obtain the CB from an expression, ensuring conformable. This must follow a
 // tree recursion
 //////////////////////////////////////////////////////////////////////////
-template <class T1,
-          typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
 inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
 {
   if ((cb == Odd) || (cb == Even)) {
-    assert(cb == lat.checkerboard);
+    assert(cb == lat.Checkerboard());
   }
-  cb = lat.checkerboard;
-  //  std::cout<<GridLogMessage<<"Lattice leaf cb "<<cb<<std::endl;
+  cb = lat.Checkerboard();
 }
-template <class T1,
-          typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+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
 {
-  //  std::cout<<GridLogMessage<<"Non lattice leaf cb"<<cb<<std::endl;
-}
-template <typename Op, typename T1>
-inline void CBFromExpression(int &cb,
-                             const LatticeUnaryExpression<Op, T1> &expr) {
-  CBFromExpression(cb, std::get<0>(expr.second));  // recurse
-  //  std::cout<<GridLogMessage<<"Unary node cb "<<cb<<std::endl;
 }
 
-template <typename Op, typename T1, typename T2>
-inline void CBFromExpression(int &cb,
-                             const LatticeBinaryExpression<Op, T1, T2> &expr) {
-  CBFromExpression(cb, std::get<0>(expr.second));  // recurse
-  CBFromExpression(cb, std::get<1>(expr.second));
-  //  std::cout<<GridLogMessage<<"Binary node cb "<<cb<<std::endl;
+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) 
+{
+  CBFromExpression(cb, expr.arg1);  // recurse AST
+  CBFromExpression(cb, expr.arg2);  // recurse AST
 }
 template <typename Op, typename T1, typename T2, typename T3>
-inline void CBFromExpression(
-    int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
-  CBFromExpression(cb, std::get<0>(expr.second));  // recurse
-  CBFromExpression(cb, std::get<1>(expr.second));
-  CBFromExpression(cb, std::get<2>(expr.second));
-  //  std::cout<<GridLogMessage<<"Trinary node cb "<<cb<<std::endl;
+inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
+{
+  CBFromExpression(cb, expr.arg1);  // recurse AST
+  CBFromExpression(cb, expr.arg2);  // recurse AST
+  CBFromExpression(cb, expr.arg3);  // recurse AST
 }
 
 ////////////////////////////////////////////
@@ -220,7 +209,7 @@ inline void CBFromExpression(
 #define GridUnopClass(name, ret)					\
   template <class arg>							\
   struct name {								\
-    static auto inline func(const arg a) -> decltype(ret) { return ret; } \
+    static auto accelerator_inline func(const arg a) -> decltype(ret) { return ret; } \
   };
 
 GridUnopClass(UnarySub, -a);
@@ -253,16 +242,18 @@ GridUnopClass(UnaryExp, exp(a));
 #define GridBinOpClass(name, combination)			\
   template <class left, class right>				\
   struct name {							\
-    static auto inline func(const left &lhs, const right &rhs) \
-        -> decltype(combination) const {                       \
+    static auto accelerator_inline				\
+    func(const left &lhs, const right &rhs)			\
+      -> decltype(combination) const				\
+    {								\
       return combination;					\
     }								\
-  }
+  };
+
 GridBinOpClass(BinaryAdd, lhs + rhs);
 GridBinOpClass(BinarySub, lhs - rhs);
 GridBinOpClass(BinaryMul, lhs *rhs);
 GridBinOpClass(BinaryDiv, lhs /rhs);
-
 GridBinOpClass(BinaryAnd, lhs &rhs);
 GridBinOpClass(BinaryOr, lhs | rhs);
 GridBinOpClass(BinaryAndAnd, lhs &&rhs);
@@ -274,17 +265,18 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
 #define GridTrinOpClass(name, combination)				\
   template <class predicate, class left, class right>			\
   struct name {								\
-    static auto inline func(const predicate &pred, const left &lhs,            \
-                            const right &rhs) -> decltype(combination) const { \
+    static auto accelerator_inline					\
+    func(const predicate &pred, const left &lhs, const right &rhs)	\
+      -> decltype(combination) const					\
+    {									\
       return combination;						\
     }									\
-  }
+  };
 
-GridTrinOpClass(
-    TrinaryWhere,
-    (predicatedWhere<predicate, typename std::remove_reference<left>::type,
-                     typename std::remove_reference<right>::type>(pred, lhs,
-                                                                  rhs)));
+GridTrinOpClass(TrinaryWhere,
+		(predicatedWhere<predicate, 
+		 typename std::remove_reference<left>::type,
+		 typename std::remove_reference<right>::type>(pred, lhs,rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
@@ -292,50 +284,32 @@ GridTrinOpClass(
 
 #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_TRINOP(name) name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
 
 #define GRID_DEF_UNOP(op, name)						\
-  template <typename T1,                                                    \
-            typename std::enable_if<is_lattice<T1>::value ||                \
-                                        is_lattice_expr<T1>::value,         \
-                                    T1>::type * = nullptr>                  \
-  inline auto op(const T1 &arg)                                             \
-      ->decltype(LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(       \
-          std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)))) { \
-    return LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(             \
-        std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)));     \
+  template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
+  inline auto op(const T1 &arg) ->decltype(LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg)) \
+  {									\
+    return     LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg); \
   }
 
 #define GRID_BINOP_LEFT(op, name)					\
   template <typename T1, typename T2,					\
-            typename std::enable_if<is_lattice<T1>::value ||                  \
-                                        is_lattice_expr<T1>::value,           \
-                                    T1>::type * = nullptr>                    \
+            typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
   inline auto op(const T1 &lhs, const T2 &rhs)				\
-      ->decltype(                                                             \
-          LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(  \
-              std::make_pair(GRID_BINOP(name)(),                              \
-                             std::forward_as_tuple(lhs, rhs)))) {             \
-    return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
-        std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
+    ->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs)) \
+  {									\
+    return     LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs);\
   }
 
 #define GRID_BINOP_RIGHT(op, name)					\
   template <typename T1, typename T2,					\
-            typename std::enable_if<!is_lattice<T1>::value &&                 \
-                                        !is_lattice_expr<T1>::value,          \
-                                    T1>::type * = nullptr,                    \
-            typename std::enable_if<is_lattice<T2>::value ||                  \
-                                        is_lattice_expr<T2>::value,           \
-                                    T2>::type * = nullptr>                    \
+            typename std::enable_if<!is_lattice<T1>::value&&!is_lattice_expr<T1>::value,T1>::type * = nullptr, \
+            typename std::enable_if< is_lattice<T2>::value|| is_lattice_expr<T2>::value,T2>::type * = nullptr> \
   inline auto op(const T1 &lhs, const T2 &rhs)				\
-      ->decltype(                                                             \
-          LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(  \
-              std::make_pair(GRID_BINOP(name)(),                              \
-                             std::forward_as_tuple(lhs, rhs)))) {             \
-    return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
-        std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
+    ->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs)) \
+  {									\
+    return     LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs); \
   }
 
 #define GRID_DEF_BINOP(op, name)		\
@@ -345,18 +319,14 @@ GridTrinOpClass(
 #define GRID_DEF_TRINOP(op, name)					\
   template <typename T1, typename T2, typename T3>			\
   inline auto op(const T1 &pred, const T2 &lhs, const T3 &rhs)		\
-      ->decltype(                                                              \
-          LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &,  \
-                                   const T3 &>(std::make_pair(                 \
-              GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs)))) {  \
-    return LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &, \
-                                    const T3 &>(std::make_pair(                \
-        GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs)));          \
+    ->decltype(LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs)) \
+  {									\
+    return LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs); \
   }
+
 ////////////////////////
 // Operator definitions
 ////////////////////////
-
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
@@ -400,29 +370,27 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-    -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> {
-  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> ret(
-      expr);
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
+{
+  Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-    -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
-                                        eval(0, std::get<1>(expr.second))))> {
-  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
-                                   eval(0, std::get<1>(expr.second))))>
-      ret(expr);
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> 
+{
+  Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-    -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
-                                        eval(0, std::get<1>(expr.second)),
-                                        eval(0, std::get<2>(expr.second))))> {
-  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
-                                   eval(0, std::get<1>(expr.second)),
-                                   eval(0, std::get<2>(expr.second))))>
-      ret(expr);
+  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
+				   eval(0, expr.arg2),
+				   eval(0, expr.arg3)))> 
+{
+  Lattice<decltype(expr.op.func(eval(0, expr.arg1),
+				eval(0, expr.arg2),
+				eval(0, expr.arg3)))>  ret(expr);
   return ret;
 }
 
@@ -433,34 +401,7 @@ auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
 #undef GRID_DEF_UNOP
 #undef GRID_DEF_BINOP
 #undef GRID_DEF_TRINOP
-}
 
-#if 0
-using namespace Grid;
-        
- int main(int argc,char **argv){
-   
-   Lattice<double> v1(16);
-   Lattice<double> v2(16);
-   Lattice<double> v3(16);
-
-   BinaryAdd<double,double> tmp;
-   LatticeBinaryExpression<BinaryAdd<double,double>,Lattice<double> &,Lattice<double> &> 
-     expr(std::make_pair(tmp,
-    std::forward_as_tuple(v1,v2)));
-   tmp.func(eval(0,v1),eval(0,v2));
-
-   auto var = v1+v2;
-   std::cout<<GridLogMessage<<typeid(var).name()<<std::endl;
-
-   v3=v1+v2;
-   v3=v1+v2+v1*v2;
- };
-
-void testit(Lattice<double> &v1,Lattice<double> &v2,Lattice<double> &v3)
-{
-   v3=v1+v2+v1*v2;
-}
-#endif
+NAMESPACE_END(Grid);
 
 #endif

Grid/lattice/Lattice_arith.h

@@ -28,228 +28,230 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_ARITH_H
 #define GRID_LATTICE_ARITH_H
 
-namespace Grid {
-
+NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 //  avoid copy back routines for mult, mac, sub, add
 //////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class obj1,class obj2,class obj3> strong_inline
+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;
+  ret.Checkerboard() = lhs.Checkerboard();
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      mult(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else
-      mult(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
-#endif
-    }
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t = lhs_v(ss);
+    auto rhs_t = rhs_v(ss);
+    mult(&tmp,&lhs_t,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      mac(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else
-      mac(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
-#endif
-    }
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t=lhs_v(ss);
+    auto rhs_t=rhs_v(ss);
+    mac(&tmp,&lhs_t,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      sub(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else
-      sub(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
-#endif
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t=lhs_v(ss);
+    auto rhs_t=rhs_v(ss);
+    sub(&tmp,&lhs_t,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
-  }
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      add(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else
-      add(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
-#endif
-    }
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t=lhs_v(ss);
+    auto rhs_t=rhs_v(ss);
+    add(&tmp,&lhs_t,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 //  avoid copy back routines for mult, mac, sub, add
 //////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-      obj1 tmp;
-      mult(&tmp,&lhs._odata[ss],&rhs);
-      vstream(ret._odata[ss],tmp);
-    }
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    mult(&tmp,&lhs_v(ss),&rhs);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-      obj1 tmp;
-      mac(&tmp,&lhs._odata[ss],&rhs);
-      vstream(ret._odata[ss],tmp);
-    }
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t=lhs_v(ss);
+    mac(&tmp,&lhs_t,&rhs);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      sub(&tmp,&lhs._odata[ss],&rhs);
-      vstream(ret._odata[ss],tmp);
-#else 
-      sub(&ret._odata[ss],&lhs._odata[ss],&rhs);
-#endif
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t=lhs_v(ss);
+    sub(&tmp,&lhs_t,&rhs);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
-  }
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      add(&tmp,&lhs._odata[ss],&rhs);
-      vstream(ret._odata[ss],tmp);
-#else 
-      add(&ret._odata[ss],&lhs._odata[ss],&rhs);
-#endif
-    }
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto lhs_t=lhs_v(ss);
+    add(&tmp,&lhs_t,&rhs);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
 
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 //  avoid copy back routines for mult, mac, sub, add
 //////////////////////////////////////////////////////////////////////////////////////////////////////
-    template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      mult(&tmp,&lhs,&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else 
-      mult(&ret._odata[ss],&lhs,&rhs._odata[ss]);
-#endif
-    }
+  auto ret_v = ret.View();
+  auto rhs_v = lhs.View();
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto rhs_t=rhs_v(ss);
+    mult(&tmp,&lhs,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      mac(&tmp,&lhs,&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else 
-      mac(&ret._odata[ss],&lhs,&rhs._odata[ss]);
-#endif
-    }
+  auto ret_v = ret.View();
+  auto rhs_v = lhs.View();
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto rhs_t=rhs_v(ss);
+    mac(&tmp,&lhs,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      sub(&tmp,&lhs,&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else 
-      sub(&ret._odata[ss],&lhs,&rhs._odata[ss]);
-#endif
+  auto ret_v = ret.View();
+  auto rhs_v = lhs.View();
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto rhs_t=rhs_v(ss);
+    sub(&tmp,&lhs,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
-  }
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      obj1 tmp;
-      add(&tmp,&lhs,&rhs._odata[ss]);
-      vstream(ret._odata[ss],tmp);
-#else 
-      add(&ret._odata[ss],&lhs,&rhs._odata[ss]);
-#endif
-    }
+  auto ret_v = ret.View();
+  auto rhs_v = lhs.View();
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
+    decltype(coalescedRead(obj1())) tmp;
+    auto rhs_t=rhs_v(ss);
+    add(&tmp,&lhs,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class sobj,class vobj> strong_inline
+template<class sobj,class vobj> inline
 void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
+  ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-    parallel_for(int ss=0;ss<x._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      vobj tmp = a*x._odata[ss]+y._odata[ss];
-      vstream(ret._odata[ss],tmp);
-#else
-      ret._odata[ss]=a*x._odata[ss]+y._odata[ss];
-#endif
+  auto ret_v = ret.View();
+  auto x_v = x.View();
+  auto y_v = y.View();
+  accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
+    auto tmp = a*x_v(ss)+y_v(ss);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
-  }
-  template<class sobj,class vobj> strong_inline
+template<class sobj,class vobj> inline
 void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
+  ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-    parallel_for(int ss=0;ss<x._grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      vobj tmp = a*x._odata[ss]+b*y._odata[ss];
-      vstream(ret._odata[ss],tmp);
-#else
-      ret._odata[ss]=a*x._odata[ss]+b*y._odata[ss];
-#endif
-    }
+  auto ret_v = ret.View();
+  auto x_v = x.View();
+  auto y_v = y.View();
+  accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
+    auto tmp = a*x_v(ss)+b*y_v(ss);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
 
-  template<class sobj,class vobj> strong_inline
-  RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
+template<class sobj,class vobj> inline
+RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
+{
     return axpy_norm_fast(ret,a,x,y);
 }
-  template<class sobj,class vobj> strong_inline
-  RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
+template<class sobj,class vobj> inline
+RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
+{
     return axpby_norm_fast(ret,a,b,x,y);
 }
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_base.h

@@ -28,311 +28,428 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
-#ifndef GRID_LATTICE_BASE_H
-#define GRID_LATTICE_BASE_H
+#pragma once 
 
 #define STREAMING_STORES
 
-namespace Grid {
-
-// TODO: 
-//       mac,real,imag
-
-// Functionality:
-//     -=,+=,*=,()
-//     add,+,sub,-,mult,mac,*
-//     adj,conjugate
-//     real,imag
-//     transpose,transposeIndex  
-//     trace,traceIndex
-//     peekIndex
-//     innerProduct,outerProduct,
-//     localNorm2
-//     localInnerProduct
+NAMESPACE_BEGIN(Grid);
 
 extern int GridCshiftPermuteMap[4][16];
 
-////////////////////////////////////////////////
-// Basic expressions used in Expression Template
-////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////
+// Base class which can be used by traits to pick up behaviour
+///////////////////////////////////////////////////////////////////
+class LatticeBase {};
 
-class LatticeBase
-{
-public:
-    virtual ~LatticeBase(void) = default;
-    GridBase *_grid;
-};
-    
-class LatticeExpressionBase {};
-
-template <typename Op, typename T1>                           
-class LatticeUnaryExpression  : public std::pair<Op,std::tuple<T1> > , public LatticeExpressionBase {
- public:
- LatticeUnaryExpression(const std::pair<Op,std::tuple<T1> > &arg): std::pair<Op,std::tuple<T1> >(arg) {};
-};
-
-template <typename Op, typename T1, typename T2>              
-class LatticeBinaryExpression : public std::pair<Op,std::tuple<T1,T2> > , public LatticeExpressionBase {
- public:
- LatticeBinaryExpression(const std::pair<Op,std::tuple<T1,T2> > &arg): std::pair<Op,std::tuple<T1,T2> >(arg) {};
-};
-
-template <typename Op, typename T1, typename T2, typename T3> 
-class LatticeTrinaryExpression :public std::pair<Op,std::tuple<T1,T2,T3> >, public LatticeExpressionBase {
- public:
- LatticeTrinaryExpression(const std::pair<Op,std::tuple<T1,T2,T3> > &arg): std::pair<Op,std::tuple<T1,T2,T3> >(arg) {};
-};
-
-void inline conformable(GridBase *lhs,GridBase *rhs)
+/////////////////////////////////////////////////////////////////////////////////////////
+// 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:
+  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;
+  };
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// 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 Lattice : public LatticeBase
+class LatticeView : public LatticeAccelerator<vobj>
 {
 public:
-    int checkerboard;
-    Vector<vobj> _odata;
 
-    // to pthread need a computable loop where loop induction is not required
-    int begin(void) { return 0;};
-    int end(void)   { return _odata.size(); }
-    vobj & operator[](int i) { return _odata[i]; };
-    const vobj & operator[](int i) const { return _odata[i]; };
 
+  // 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
+/////////////////////////////////////////////////////////////////////////////////////////
+template<class vobj>
+class Lattice : public LatticeAccelerator<vobj>
+{
+public:
+  GridBase *Grid(void) const { return this->_grid; }
+  ///////////////////////////////////////////////////
+  // Member types
+  ///////////////////////////////////////////////////
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   typedef vobj vector_object;
 
+private:
+  void dealloc(void)
+  {
+    alignedAllocator<vobj> alloc;
+    if( this->_odata_size ) {
+      alloc.deallocate(this->_odata,this->_odata_size);
+      this->_odata=nullptr;
+      this->_odata_size=0;
+    }
+  }
+  void resize(uint64_t size)
+  {
+    alignedAllocator<vobj> alloc;
+    if ( this->_odata_size != size ) {
+      dealloc();
+    }
+    this->_odata_size = size;
+    if ( size ) 
+      this->_odata      = alloc.allocate(this->_odata_size);
+    else 
+      this->_odata      = nullptr;
+  }
+public:
+  /////////////////////////////////////////////////////////////////////////////////
+  // 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 
+  {
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
+    return accessor;
+  }
+  
+  ~Lattice() { 
+    if ( this->_odata_size ) {
+      dealloc();
+    }
+   }
   ////////////////////////////////////////////////////////////////////////////////
   // Expression Template closure support
   ////////////////////////////////////////////////////////////////////////////////
-  template <typename Op, typename T1>                         strong_inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
+  template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
   {
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
-    conformable(_grid,egrid);
+    conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,tmp);
-#else
-      _odata[ss]=eval(ss,expr);
-#endif
-    }
+    auto me  = View();
+    accelerator_for(ss,me.size(),1,{
+      auto tmp = eval(ss,expr);
+      vstream(me[ss],tmp);
+    });
     return *this;
   }
-  template <typename Op, typename T1,typename T2> strong_inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
+  template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
   {
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
-    conformable(_grid,egrid);
+    conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,tmp);
-#else
-      _odata[ss]=eval(ss,expr);
-#endif
-    }
+    auto me  = View();
+    accelerator_for(ss,me.size(),1,{
+      auto tmp = eval(ss,expr);
+      vstream(me[ss],tmp);
+    });
     return *this;
   }
-  template <typename Op, typename T1,typename T2,typename T3> strong_inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
+  template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
   {
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
-    conformable(_grid,egrid);
+    conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
-
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      //vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,eval(ss,expr));
-#else
-      _odata[ss] = eval(ss,expr);
-#endif
-    }
+    this->checkerboard=cb;
+    auto me  = View();
+    accelerator_for(ss,me.size(),1,{
+      auto tmp = eval(ss,expr);
+      vstream(me[ss],tmp);
+    });
     return *this;
   }
   //GridFromExpression is tricky to do
   template<class Op,class T1>
   Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
-    _grid = nullptr;
-    GridFromExpression(_grid,expr);
-    assert(_grid!=nullptr);
+    this->_grid = nullptr;
+    GridFromExpression(this->_grid,expr);
+    assert(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    _odata.resize(_grid->oSites());
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,tmp);
-#else
-      _odata[ss]=eval(ss,expr);
-#endif
+    resize(this->_grid->oSites());
+
+    *this = expr;
   }
-  };
   template<class Op,class T1, class T2>
   Lattice(const LatticeBinaryExpression<Op,T1,T2> & expr) {
-    _grid = nullptr;
-    GridFromExpression(_grid,expr);
-    assert(_grid!=nullptr);
+    this->_grid = nullptr;
+    GridFromExpression(this->_grid,expr);
+    assert(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    _odata.resize(_grid->oSites());
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-#ifdef STREAMING_STORES
-      vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,tmp);
-#else
-      _odata[ss]=eval(ss,expr);
-#endif
+    resize(this->_grid->oSites());
+
+    *this = expr;
   }
-  };
   template<class Op,class T1, class T2, class T3>
   Lattice(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr) {
-    _grid = nullptr;
-    GridFromExpression(_grid,expr);
-    assert(_grid!=nullptr);
+    this->_grid = nullptr;
+    GridFromExpression(this->_grid,expr);
+    assert(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    _odata.resize(_grid->oSites());
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      vstream(_odata[ss] ,eval(ss,expr));
+    resize(this->_grid->oSites());
+
+    *this = expr;
+  }
+
+  template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
+    auto me  = View();
+    thread_for(ss,me.size(),{
+      me[ss] = r;
+    });
+    return *this;
   }
-  };
 
   //////////////////////////////////////////////////////////////////
-  // Constructor requires "grid" passed.
-  // what about a default grid?
-  //////////////////////////////////////////////////////////////////
-  Lattice(GridBase *grid) : _odata(grid->oSites()) {
-    _grid = grid;
-    //        _odata.reserve(_grid->oSites());
-    //        _odata.resize(_grid->oSites());
-    //      std::cout << "Constructing lattice object with Grid pointer "<<_grid<<std::endl;
-    assert((((uint64_t)&_odata[0])&0xF) ==0);
-    checkerboard=0;
+  // Follow rule of five, with Constructor requires "grid" passed
+  // to user defined constructor
+  ///////////////////////////////////////////
+  // user defined constructor
+  ///////////////////////////////////////////
+  Lattice(GridBase *grid) { 
+    this->_grid = grid;
+    resize(this->_grid->oSites());
+    assert((((uint64_t)&this->_odata[0])&0xF) ==0);
+    this->checkerboard=0;
   }
   
-  Lattice(const Lattice& r){ // copy constructor
-    _grid = r._grid;
-    checkerboard = r.checkerboard;
-    _odata.resize(_grid->oSites());// essential
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      _odata[ss]=r._odata[ss];
-    }  	
-  }
-
-  Lattice(Lattice&& r){ // move constructor
-    _grid = r._grid;
-    checkerboard = r.checkerboard;
-    _odata=std::move(r._odata);
-  }
-  
-  inline Lattice<vobj> & operator = (Lattice<vobj> && r)
-  {
-    _grid        = r._grid;
-    checkerboard = r.checkerboard;
-    _odata       =std::move(r._odata);
-    return *this;
-  }
-
-  inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
-    _grid        = r._grid;
-    checkerboard = r.checkerboard;
-    _odata.resize(_grid->oSites());// essential
-    
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      _odata[ss]=r._odata[ss];
-    }  	
-    return *this;
-  }
-
-  template<class robj> strong_inline Lattice<vobj> & operator = (const Lattice<robj> & r){
-    this->checkerboard = r.checkerboard;
-    conformable(*this,r);
-    
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      this->_odata[ss]=r._odata[ss];
-    }
-    return *this;
-  }
-
-  virtual ~Lattice(void) = default;
+  //  virtual ~Lattice(void) = default;
     
   void reset(GridBase* grid) {
-    if (_grid != grid) {
-      _grid = grid;
-      _odata.resize(grid->oSites());
-      checkerboard = 0;
+    if (this->_grid != grid) {
+      this->_grid = grid;
+      this->_odata.resize(grid->oSites());
+      this->checkerboard = 0;
     }
   }
-  
-
-  template<class sobj> strong_inline Lattice<vobj> & operator = (const sobj & r){
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      this->_odata[ss]=r;
+  ///////////////////////////////////////////
+  // copy constructor
+  ///////////////////////////////////////////
+  Lattice(const Lattice& r){ 
+    //    std::cout << "Lattice constructor(const Lattice &) "<<this<<std::endl; 
+    this->_grid = r.Grid();
+    resize(this->_grid->oSites());
+    *this = r;
   }
+  ///////////////////////////////////////////
+  // move constructor
+  ///////////////////////////////////////////
+  Lattice(Lattice && r){ 
+    this->_grid = r.Grid();
+    this->_odata      = r._odata;
+    this->_odata_size = r._odata_size;
+    this->checkerboard= r.Checkerboard();
+    r._odata      = nullptr;
+    r._odata_size = 0;
+  }
+  ///////////////////////////////////////////
+  // assignment template
+  ///////////////////////////////////////////
+  template<class robj> inline Lattice<vobj> & operator = (const Lattice<robj> & r){
+    typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
+    conformable(*this,r);
+    this->checkerboard = r.Checkerboard();
+    auto me =   View();
+    auto him= r.View();
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+      coalescedWrite(me[ss],him(ss));
+    });
     return *this;
   }
 
+  ///////////////////////////////////////////
+  // Copy assignment 
+  ///////////////////////////////////////////
+  inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
+    this->checkerboard = r.Checkerboard();
+    conformable(*this,r);
+    auto me =   View();
+    auto him= r.View();
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+      coalescedWrite(me[ss],him(ss));
+    });
+    return *this;
+  }
+  ///////////////////////////////////////////
+  // Move assignment possible if same type
+  ///////////////////////////////////////////
+  inline Lattice<vobj> & operator = (Lattice<vobj> && r){
 
+    resize(0); // deletes if appropriate
+    this->_grid       = r.Grid();
+    this->_odata      = r._odata;
+    this->_odata_size = r._odata_size;
+    this->checkerboard= r.Checkerboard();
+
+    r._odata      = nullptr;
+    r._odata_size = 0;
+    
+    return *this;
+  }
+
+  /////////////////////////////////////////////////////////////////////////////
   // *=,+=,-= operators inherit behvour from correspond */+/- operation
-  template<class T> strong_inline Lattice<vobj> &operator *=(const T &r) {
+  /////////////////////////////////////////////////////////////////////////////
+  template<class T> inline Lattice<vobj> &operator *=(const T &r) {
     *this = (*this)*r;
     return *this;
   }
   
-  template<class T> strong_inline Lattice<vobj> &operator -=(const T &r) {
+  template<class T> inline Lattice<vobj> &operator -=(const T &r) {
     *this = (*this)-r;
     return *this;
   }
-  template<class T> strong_inline Lattice<vobj> &operator +=(const T &r) {
+  template<class T> inline Lattice<vobj> &operator +=(const T &r) {
     *this = (*this)+r;
     return *this;
   }
+
+  friend inline void swap(Lattice &l, Lattice &r) { 
+    conformable(l,r);
+    LatticeAccelerator<vobj> tmp;
+    LatticeAccelerator<vobj> *lp = (LatticeAccelerator<vobj> *)&l;
+    LatticeAccelerator<vobj> *rp = (LatticeAccelerator<vobj> *)&r;
+    tmp = *lp;    *lp=*rp;    *rp=tmp;
+  }
+
 }; // class Lattice
 
 template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
-    std::vector<int> gcoor;
   typedef typename vobj::scalar_object sobj;
+  for(int g=0;g<o.Grid()->_gsites;g++){
+
+    Coordinate gcoor;
+    o.Grid()->GlobalIndexToGlobalCoor(g,gcoor);
+
     sobj ss;
-    for(int g=0;g<o._grid->_gsites;g++){
-      o._grid->GlobalIndexToGlobalCoor(g,gcoor);
     peekSite(ss,o,gcoor);
     stream<<"[";
     for(int d=0;d<gcoor.size();d++){
@@ -345,31 +462,5 @@ public:
   return stream;
 }
   
-}
+NAMESPACE_END(Grid);
 
-
-
-#include "Lattice_conformable.h"
-#define GRID_LATTICE_EXPRESSION_TEMPLATES
-#ifdef  GRID_LATTICE_EXPRESSION_TEMPLATES
-#include "Lattice_ET.h"
-#else 
-#include "Lattice_overload.h"
-#endif
-#include "Lattice_arith.h"
-#include "Lattice_trace.h"
-#include "Lattice_transpose.h"
-#include "Lattice_local.h"
-#include "Lattice_reduction.h"
-#include "Lattice_peekpoke.h"
-#include "Lattice_reality.h"
-#include "Lattice_comparison_utils.h"
-#include "Lattice_comparison.h"
-#include "Lattice_coordinate.h"
-#include "Lattice_where.h"
-#include "Lattice_rng.h"
-#include "Lattice_unary.h"
-#include "Lattice_transfer.h"
-
-
-#endif

Grid/lattice/Lattice_comparison.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_COMPARISON_H
 #define GRID_LATTICE_COMPARISON_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////////
 // relational operators
@@ -40,40 +40,78 @@ namespace Grid {
 //Query supporting logical &&, ||, 
 //////////////////////////////////////////////////////////////////////////
 
+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
 //////////////////////////////////////////////////////////////////////////
+
 template<class vfunctor,class lobj,class robj>  
-    inline Lattice<vInteger> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
+inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
 {
-    Lattice<vInteger> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      ret._odata[ss]=op(lhs._odata[ss],rhs._odata[ss]);
-    }
+  Lattice<vPredicate> ret(rhs.Grid());
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
+  thread_for( ss, rhs_v.size(), {
+      ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
+  });
   return ret;
 }
 //////////////////////////////////////////////////////////////////////////
 // compare lattice to scalar
 //////////////////////////////////////////////////////////////////////////
 template<class vfunctor,class lobj,class robj> 
-    inline Lattice<vInteger> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
+inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
 {
-    Lattice<vInteger> ret(lhs._grid);
-    parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
-      ret._odata[ss]=op(lhs._odata[ss],rhs);
-    }
+  Lattice<vPredicate> ret(lhs.Grid());
+  auto lhs_v = lhs.View();
+  auto ret_v = ret.View();
+  thread_for( ss, lhs_v.size(), {
+    ret_v[ss]=op(lhs_v[ss],rhs);
+  });
   return ret;
 }
 //////////////////////////////////////////////////////////////////////////
 // compare scalar to lattice
 //////////////////////////////////////////////////////////////////////////
 template<class vfunctor,class lobj,class robj> 
-    inline Lattice<vInteger> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
+inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
 {
-    Lattice<vInteger> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      ret._odata[ss]=op(lhs._odata[ss],rhs);
-    }
+  Lattice<vPredicate> ret(rhs.Grid());
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
+  thread_for( ss, rhs_v.size(), {
+    ret_v[ss]=op(lhs,rhs_v[ss]);
+  });
   return ret;
 }
   
@@ -82,88 +120,88 @@ namespace Grid {
 //////////////////////////////////////////////////////////////////////////
 // Less than
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vlt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vlt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vlt<lobj,robj>(),lhs,rhs);
 }
   
 // Less than equal
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vle<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vle<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vle<lobj,robj>(),lhs,rhs);
 }
   
 // Greater than 
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vgt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vgt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vgt<lobj,robj>(),lhs,rhs);
 }
   
   
 // Greater than equal
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vge<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-   inline Lattice<vInteger> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vge<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vge<lobj,robj>(),lhs,rhs);
 }
    
 // equal
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(veq<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(veq<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(veq<lobj,robj>(),lhs,rhs);
 }
    
    
 // not equal
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vne<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vne<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-     inline Lattice<vInteger> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vne<lobj,robj>(),lhs,rhs);
 }
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_comparison_utils.h

@@ -26,10 +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_COMPARISON_H
-#define GRID_COMPARISON_H
 
-namespace Grid {
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
 
   /////////////////////////////////////////
   // This implementation is a bit poor.
@@ -44,42 +44,42 @@ namespace Grid {
   //
   template<class lobj,class robj> class veq {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) == (rhs);
     }
   };
   template<class lobj,class robj> class vne {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) != (rhs);
     }
   };
   template<class lobj,class robj> class vlt {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) < (rhs);
     }
   };
   template<class lobj,class robj> class vle {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) <= (rhs);
     }
   };
   template<class lobj,class robj> class vgt {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) > (rhs);
     }
   };
   template<class lobj,class robj> class vge {
     public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) >= (rhs);
     }
@@ -88,42 +88,42 @@ namespace Grid {
   // Generic list of functors
   template<class lobj,class robj> class seq {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) == (rhs);
     }
   };
   template<class lobj,class robj> class sne {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) != (rhs);
     }
   };
   template<class lobj,class robj> class slt {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) < (rhs);
     }
   };
   template<class lobj,class robj> class sle {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) <= (rhs);
     }
   };
   template<class lobj,class robj> class sgt {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) > (rhs);
     }
   };
   template<class lobj,class robj> class sge {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) >= (rhs);
     }
@@ -133,12 +133,12 @@ namespace Grid {
   // Integer and real get extra relational functions.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
+    accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
-      std::vector<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
-      std::vector<scalar> vrhs(vsimd::Nsimd());
-      std::vector<Integer> vpred(vsimd::Nsimd());
+      ExtractBuffer<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
+      ExtractBuffer<scalar> vrhs(vsimd::Nsimd());
+      ExtractBuffer<Integer> vpred(vsimd::Nsimd());
       vInteger ret;
       extract<vsimd,scalar>(lhs,vlhs);
       extract<vsimd,scalar>(rhs,vrhs);
@@ -150,11 +150,11 @@ namespace Grid {
     }
 
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
+    accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
-      std::vector<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
-      std::vector<Integer> vpred(vsimd::Nsimd());
+      ExtractBuffer<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
+      ExtractBuffer<Integer> vpred(vsimd::Nsimd());
       vInteger ret;
       extract<vsimd,scalar>(lhs,vlhs);
       for(int s=0;s<vsimd::Nsimd();s++){
@@ -165,11 +165,11 @@ namespace Grid {
     }
 
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
+    accelerator_inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
-      std::vector<scalar> vrhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
-      std::vector<Integer> vpred(vsimd::Nsimd());
+      ExtractBuffer<scalar> vrhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
+      ExtractBuffer<Integer> vpred(vsimd::Nsimd());
       vInteger ret;
       extract<vsimd,scalar>(rhs,vrhs);
       for(int s=0;s<vsimd::Nsimd();s++){
@@ -181,30 +181,30 @@ namespace Grid {
 
 #define DECLARE_RELATIONAL_EQ(op,functor) \
   template<class vsimd,IfSimd<vsimd> = 0>\
-    inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
+    accelerator_inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd,IfSimd<vsimd> = 0>\
-    inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
+    accelerator_inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd,IfSimd<vsimd> = 0>\
-    inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
+    accelerator_inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd>\
-    inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
+    accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
     {									\
       return lhs._internal op rhs;					\
     }									\
   template<class vsimd>\
-    inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
+    accelerator_inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
     {									\
       return lhs op rhs._internal;					\
     }									\
@@ -212,7 +212,7 @@ namespace Grid {
 #define DECLARE_RELATIONAL(op,functor) \
   DECLARE_RELATIONAL_EQ(op,functor)    \
   template<class vsimd>\
-    inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
+    accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
     {									\
       return lhs._internal op rhs._internal;				\
     }									
@@ -226,7 +226,7 @@ DECLARE_RELATIONAL(!=,sne);
 
 #undef DECLARE_RELATIONAL
 
-}
+NAMESPACE_END(Grid);
+
 
 
-#endif

Grid/lattice/Lattice_conformable.h

@@ -28,13 +28,13 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_CONFORMABLE_H
 #define GRID_LATTICE_CONFORMABLE_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class obj1,class obj2> void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
 {
-        assert(lhs._grid == rhs._grid);
-        assert(lhs.checkerboard == rhs.checkerboard);
+  assert(lhs.Grid() == rhs.Grid());
+  assert(lhs.Checkerboard() == rhs.Checkerboard());
 }
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_coordinate.h

@@ -25,32 +25,50 @@ 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_LATTICE_COORDINATE_H
-#define GRID_LATTICE_COORDINATE_H
+#pragma once 
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
 {
   typedef typename iobj::scalar_type scalar_type;
   typedef typename iobj::vector_type vector_type;
 
-      GridBase *grid = l._grid;
+  GridBase *grid = l.Grid();
   int Nsimd = grid->iSites();
 
-      std::vector<int> gcoor;
-      std::vector<scalar_type> mergebuf(Nsimd);
+  Coordinate gcoor;
+  ExtractBuffer<scalar_type> mergebuf(Nsimd);
 
   vector_type vI;
+  auto l_v = l.View();
   for(int o=0;o<grid->oSites();o++){
     for(int i=0;i<grid->iSites();i++){
       grid->RankIndexToGlobalCoor(grid->ThisRank(),o,i,gcoor);
       mergebuf[i]=(Integer)gcoor[mu];
     }
     merge<vector_type,scalar_type>(vI,mergebuf);
-	l._odata[o]=vI;
+    l_v[o]=vI;
   }
 };
 
+// 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;
       }
-#endif
+    }}
+}
+
+NAMESPACE_END(Grid);
+

Grid/lattice/Lattice_local.h

@@ -32,7 +32,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 // localInner, localNorm, outerProduct
 ///////////////////////////////////////////////
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////
 // Non site, reduced locally reduced routines
@@ -42,10 +42,12 @@ namespace Grid {
 template<class vobj>
 inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
 {
-      Lattice<typename vobj::tensor_reduced> ret(rhs._grid);
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	ret._odata[ss]=innerProduct(rhs._odata[ss],rhs._odata[ss]);
-      }
+  Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
+  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
+    coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
+  });
   return ret;
 }
   
@@ -53,23 +55,33 @@ namespace Grid {
 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);
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	ret._odata[ss]=innerProduct(lhs._odata[ss],rhs._odata[ss]);
-      }
+  Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  auto ret_v = ret.View();
+  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
+    coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
+  });
   return ret;
 }
   
 // outerProduct Scalar x Scalar -> Scalar
 //              Vector x Vector -> Matrix
 template<class ll,class rr>
-    inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))>
+inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(ll(),rr()))>
 {
-    Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      ret._odata[ss]=outerProduct(lhs._odata[ss],rhs._odata[ss]);
-    }
+  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();
+  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
+    ret_v[ss]=outerProduct(lhs_v[ss],rhs_v[ss]);
+  });
   return ret;
 }
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_matrix_reduction.h

@@ -0,0 +1,202 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+    Source file: ./lib/lattice/Lattice_reduction.h
+    Copyright (C) 2015
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once 
+#include <Grid/Grid_Eigen_Dense.h>
+
+#ifdef GRID_WARN_SUBOPTIMAL
+#warning "Optimisation alert all these reduction loops are NOT threaded "
+#endif     
+
+NAMESPACE_BEGIN(Grid);
+
+template<class vobj>
+static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
+{    
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
+
+  GridBase *FullGrid  = X.Grid();
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+
+  //FIXME package in a convenient iterator
+  //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 X_v = X.View();
+  auto Y_v = Y.View();
+  auto R_v = R.View();
+  thread_region
+  {
+    std::vector<vobj> s_x(Nblock);
+
+    thread_loop_collapse2( (int n=0;n<nblock;n++),{
+      for(int b=0;b<block;b++){
+	int o  = n*stride + b;
+
+	for(int i=0;i<Nblock;i++){
+	  s_x[i] = X_v[o+i*ostride];
+	}
+
+	vobj dot;
+	for(int i=0;i<Nblock;i++){
+	  dot = Y_v[o+i*ostride];
+	  for(int j=0;j<Nblock;j++){
+	    dot = dot + s_x[j]*(scale*aa(j,i));
+	  }
+	  R_v[o+i*ostride]=dot;
+	}
+      }});
+  }
+};
+
+template<class vobj>
+static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
+{    
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
+
+  GridBase *FullGrid  = X.Grid();
+  assert( FullGrid->_simd_layout[Orthog]==1);
+
+  //FIXME package in a convenient iterator
+  //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 X_v = X.View();
+  auto R_v = R.View();
+
+  thread_region
+  {
+    std::vector<vobj> s_x(Nblock);
+    
+    thread_loop_collapse2( (int n=0;n<nblock;n++),{
+      for(int b=0;b<block;b++){
+	int o  = n*stride + b;
+
+	for(int i=0;i<Nblock;i++){
+	  s_x[i] = X_v[o+i*ostride];
+	}
+
+	vobj dot;
+	for(int i=0;i<Nblock;i++){
+	  dot = s_x[0]*(scale*aa(0,i));
+	  for(int j=1;j<Nblock;j++){
+	    dot = dot + s_x[j]*(scale*aa(j,i));
+	  }
+	  R_v[o+i*ostride]=dot;
+	}
+    }});
+  }
+
+};
+
+
+template<class vobj>
+static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+  
+  GridBase *FullGrid  = lhs.Grid();
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  
+  int Nblock = FullGrid->GlobalDimensions()[Orthog];
+  
+  //  Lattice<vobj> Lslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+  
+  mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+  //  int nh =  FullGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  //  int nl = nh-1;
+
+  //FIXME package in a convenient iterator
+  //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];
+
+  typedef typename vobj::vector_typeD vector_typeD;
+  auto lhs_v = lhs.View();
+  auto rhs_v = rhs.View();
+  thread_region {
+    std::vector<vobj> Left(Nblock);
+    std::vector<vobj> Right(Nblock);
+    Eigen::MatrixXcd  mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+    thread_loop_collapse2((int n=0;n<nblock;n++),{
+      for(int b=0;b<block;b++){
+
+	int o  = n*stride + b;
+
+	for(int i=0;i<Nblock;i++){
+	  Left [i] = lhs_v[o+i*ostride];
+	  Right[i] = rhs_v[o+i*ostride];
+	}
+
+	for(int i=0;i<Nblock;i++){
+	  for(int j=0;j<Nblock;j++){
+	    auto tmp = innerProduct(Left[i],Right[j]);
+	    auto rtmp = TensorRemove(tmp);
+	    ComplexD z = Reduce(rtmp);
+	    mat_thread(i,j) += std::complex<double>(real(z),imag(z));
+	  }}
+    }});
+    thread_critical {
+      mat += mat_thread;
+    }  
+  }
+
+  for(int i=0;i<Nblock;i++){
+    for(int j=0;j<Nblock;j++){
+      ComplexD sum = mat(i,j);
+      FullGrid->GlobalSum(sum);
+      mat(i,j)=sum;
+    }}
+
+  return;
+}
+
+NAMESPACE_END(Grid);
+
+
+

Grid/lattice/Lattice_overload.h

@@ -1,138 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_overload.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#ifndef GRID_LATTICE_OVERLOAD_H
-#define GRID_LATTICE_OVERLOAD_H
-
-namespace Grid {
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // unary negation
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class vobj>
-  inline Lattice<vobj> operator -(const Lattice<vobj> &r)
-  {
-    Lattice<vobj> ret(r._grid);
-    parallel_for(int ss=0;ss<r._grid->oSites();ss++){
-      vstream(ret._odata[ss], -r._odata[ss]);
-    }
-    return ret;
-  } 
-  /////////////////////////////////////////////////////////////////////////////////////
-  // Lattice BinOp Lattice,
-  //NB mult performs conformable check. Do not reapply here for performance.
-  /////////////////////////////////////////////////////////////////////////////////////
-  template<class left,class right>
-    inline auto operator * (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]*rhs._odata[0])>
-  {
-    Lattice<decltype(lhs._odata[0]*rhs._odata[0])> ret(rhs._grid);
-    mult(ret,lhs,rhs);
-    return ret;
-  }
-  template<class left,class right>
-    inline auto operator + (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]+rhs._odata[0])>
-  {
-    Lattice<decltype(lhs._odata[0]+rhs._odata[0])> ret(rhs._grid);
-    add(ret,lhs,rhs);
-    return ret;
-  }
-  template<class left,class right>
-    inline auto operator - (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]-rhs._odata[0])>
-  {
-    Lattice<decltype(lhs._odata[0]-rhs._odata[0])> ret(rhs._grid);
-    sub(ret,lhs,rhs);
-    return ret;
-  }
-  
-  // Scalar BinOp Lattice ;generate return type
-  template<class left,class right>
-  inline auto operator * (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs*rhs._odata[0])>
-  {
-    Lattice<decltype(lhs*rhs._odata[0])> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      decltype(lhs*rhs._odata[0]) tmp=lhs*rhs._odata[ss]; 
-      vstream(ret._odata[ss],tmp);
-	   //      ret._odata[ss]=lhs*rhs._odata[ss];
-    }
-    return ret;
-  }
-  template<class left,class right>
-    inline auto operator + (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs+rhs._odata[0])>
-    {
-      Lattice<decltype(lhs+rhs._odata[0])> ret(rhs._grid);
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	decltype(lhs+rhs._odata[0]) tmp =lhs-rhs._odata[ss];  
-	vstream(ret._odata[ss],tmp);
-	//	ret._odata[ss]=lhs+rhs._odata[ss];
-      }
-        return ret;
-    }
-  template<class left,class right>
-    inline auto operator - (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs-rhs._odata[0])>
-  {
-    Lattice<decltype(lhs-rhs._odata[0])> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      decltype(lhs-rhs._odata[0]) tmp=lhs-rhs._odata[ss];  
-      vstream(ret._odata[ss],tmp);
-    }
-    return ret;
-  }
-    template<class left,class right>
-      inline auto operator * (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]*rhs)>
-    {
-      Lattice<decltype(lhs._odata[0]*rhs)> ret(lhs._grid);
-      parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
-	decltype(lhs._odata[0]*rhs) tmp =lhs._odata[ss]*rhs;
-	vstream(ret._odata[ss],tmp);
-	//            ret._odata[ss]=lhs._odata[ss]*rhs;
-      }
-      return ret;
-    }
-    template<class left,class right>
-      inline auto operator + (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]+rhs)>
-    {
-        Lattice<decltype(lhs._odata[0]+rhs)> ret(lhs._grid);
-	parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  decltype(lhs._odata[0]+rhs) tmp=lhs._odata[ss]+rhs; 
-	  vstream(ret._odata[ss],tmp);
-	  //	  ret._odata[ss]=lhs._odata[ss]+rhs;
-        }
-        return ret;
-    }
-    template<class left,class right>
-      inline auto operator - (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]-rhs)>
-    {
-      Lattice<decltype(lhs._odata[0]-rhs)> ret(lhs._grid);
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  decltype(lhs._odata[0]-rhs) tmp=lhs._odata[ss]-rhs;
-	  vstream(ret._odata[ss],tmp);
-	  //	ret._odata[ss]=lhs._odata[ss]-rhs;
-      }
-      return ret;
-    }
-}
-#endif

Grid/lattice/Lattice_peekpoke.h

@@ -34,29 +34,35 @@ Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 // Peeking and poking around
 ///////////////////////////////////////////////
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
+
+// FIXME accelerator_loop and accelerator_inline these
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Peek internal indices of a Lattice object
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<int Index,class vobj> 
-       auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Index>(lhs._odata[0],i))>
+auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Index>(vobj(),i))>
 {
-      Lattice<decltype(peekIndex<Index>(lhs._odata[0],i))> ret(lhs._grid);
-      ret.checkerboard=lhs.checkerboard;
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-	ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i);
-      }
+  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(), {
+    ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
+  });
   return ret;
 };
 template<int Index,class vobj> 
-      auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))>
+auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(vobj(),i,j))>
 {
-      Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))> ret(lhs._grid);
-      ret.checkerboard=lhs.checkerboard;
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-	ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i,j);
-      }
+  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(), {
+    ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
+  });
   return ret;
 };
 
@@ -64,34 +70,38 @@ namespace Grid {
 // Poke internal indices of a Lattice object
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<int Index,class vobj>  
-    void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0))> & rhs,int i)
+void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
 {
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-	pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i);
-      }      
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for( ss, lhs_v.size(), {
+    pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
+  });
 }
 template<int Index,class vobj> 
-      void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0,0))> & rhs,int i,int j)
+void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
 {
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-	pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i,j);
-      }      
+  auto rhs_v = rhs.View();
+  auto lhs_v = lhs.View();
+  thread_for( ss, lhs_v.size(), {
+    pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
+  });
 }
 
 //////////////////////////////////////////////////////
 // Poke a scalar object into the SIMD array
 //////////////////////////////////////////////////////
 template<class vobj,class sobj> 
-    void pokeSite(const sobj &s,Lattice<vobj> &l,const std::vector<int> &site){
+void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
-      GridBase *grid=l._grid;
+  GridBase *grid=l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-      assert( l.checkerboard== l._grid->CheckerBoard(site));
+  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   int rank,odx,idx;
@@ -99,13 +109,13 @@ namespace Grid {
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
   grid->Broadcast(grid->BossRank(),s);
 
-      std::vector<sobj> buf(Nsimd);
-
   // extract-modify-merge cycle is easiest way and this is not perf critical
+  ExtractBuffer<sobj> buf(Nsimd);
+  auto l_v = l.View();
   if ( rank == grid->ThisRank() ) {
-	extract(l._odata[odx],buf);
+    extract(l_v[odx],buf);
     buf[idx] = s;
-	merge(l._odata[odx],buf);
+    merge(l_v[odx],buf);
   }
 
   return;
@@ -116,22 +126,23 @@ namespace Grid {
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
 template<class vobj,class sobj>
-      void peekSite(sobj &s,const Lattice<vobj> &l,const std::vector<int> &site){
+void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         
-      GridBase *grid=l._grid;
+  GridBase *grid=l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-      assert( l.checkerboard == l._grid->CheckerBoard(site));
+  assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
 
   int rank,odx,idx;
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
 
-      std::vector<sobj> buf(Nsimd);
-      extract(l._odata[odx],buf);
+  ExtractBuffer<sobj> buf(Nsimd);
+  auto l_v = l.View();
+  extract(l_v[odx],buf);
 
   s = buf[idx];
 
@@ -145,16 +156,16 @@ namespace Grid {
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
 template<class vobj,class sobj>
-    void peekLocalSite(sobj &s,const Lattice<vobj> &l,std::vector<int> &site){
+void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
         
-      GridBase *grid = l._grid;
+  GridBase *grid = l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-      assert( l.checkerboard== l._grid->CheckerBoard(site));
+  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -162,7 +173,8 @@ namespace Grid {
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-      scalar_type * vp = (scalar_type *)&l._odata[odx];
+  auto l_v = l.View();
+  scalar_type * vp = (scalar_type *)&l_v[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
@@ -173,16 +185,16 @@ namespace Grid {
 };
 
 template<class vobj,class sobj>
-    void pokeLocalSite(const sobj &s,Lattice<vobj> &l,std::vector<int> &site){
+void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
 
-      GridBase *grid=l._grid;
+  GridBase *grid=l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-      assert( l.checkerboard== l._grid->CheckerBoard(site));
+  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -190,9 +202,9 @@ namespace Grid {
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-      scalar_type * vp = (scalar_type *)&l._odata[odx];
+  auto l_v = l.View();
+  scalar_type * vp = (scalar_type *)&l_v[odx];
   scalar_type * pt = (scalar_type *)&s;
-      
   for(int w=0;w<words;w++){
     vp[idx+w*Nsimd] = pt[w];
   }
@@ -200,6 +212,6 @@ namespace Grid {
   return;
 };
 
-}
+NAMESPACE_END(Grid);
 #endif
 

Grid/lattice/Lattice_reality.h

@@ -36,22 +36,28 @@ Author: neo <cossu@post.kek.jp>
 // The choice of burying complex in the SIMD
 // is making the use of "real" and "imag" very cumbersome
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
-        Lattice<vobj> ret(lhs._grid);
-	parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-            ret._odata[ss] = adj(lhs._odata[ss]);
-        }
+  Lattice<vobj> ret(lhs.Grid());
+  auto lhs_v = lhs.View();
+  auto ret_v = ret.View();
+  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
+    coalescedWrite(ret_v[ss], adj(lhs_v(ss)));
+  });
   return ret;
 };
 
 template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
-        Lattice<vobj> ret(lhs._grid);
-	parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-	  ret._odata[ss] = conjugate(lhs._odata[ss]);
-        }
+  Lattice<vobj> ret(lhs.Grid());
+  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;
 };
-}
+
+NAMESPACE_END(Grid);
+
 #endif

Grid/lattice/Lattice_reduction.h

@@ -19,22 +19,76 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_LATTICE_REDUCTION_H
-#define GRID_LATTICE_REDUCTION_H
+#pragma once
 
 #include <Grid/Grid_Eigen_Dense.h>
 
-namespace Grid {
-#ifdef GRID_WARN_SUBOPTIMAL
-#warning "Optimisation alert all these reduction loops are NOT threaded "
+
+#ifdef GRID_NVCC
+#include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
 
+NAMESPACE_BEGIN(Grid);
+
+//////////////////////////////////////////////////////
+// FIXME this should promote to double and accumulate
+//////////////////////////////////////////////////////
+template<class vobj>
+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 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_object sum(const vobj *arg, Integer osites)
+{
+#ifdef GRID_NVCC
+  return sum_gpu(arg,osites);
+#else
+  return sum_cpu(arg,osites);
+#endif  
+}
+template<class vobj>
+inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+{
+  auto arg_v = arg.View();
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum(&arg_v[0],osites);
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Deterministic Reduction operations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
-  auto nrm = innerProduct(arg,arg);
-  return std::real(nrm); 
+  ComplexD nrm = innerProduct(arg,arg);
+  return real(nrm); 
 }
 
 // Double inner product
@@ -43,32 +97,49 @@ inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &righ
 {
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
-  GridBase *grid = left._grid;
-  const int pad = 8;
+  ComplexD  nrm;
   
-  ComplexD  inner;
-  Vector<ComplexD> sumarray(grid->SumArraySize()*pad);
+  GridBase *grid = left.Grid();
   
-  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
-    int nwork, mywork, myoff;
-    GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
+  // Might make all code paths go this way.
+  auto left_v = left.View();
+  auto right_v=right.View();
 
-    decltype(innerProductD(left._odata[0],right._odata[0])) vinner=zero; // private to thread; sub summation
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      vinner = vinner + innerProductD(left._odata[ss],right._odata[ss]);
-    }
-    // All threads sum across SIMD; reduce serial work at end
-    // one write per cacheline with streaming store
-    ComplexD tmp = Reduce(TensorRemove(vinner)) ;
-    vstream(sumarray[thr*pad],tmp);
-  }
+  const uint64_t nsimd = grid->Nsimd();
+  const uint64_t sites = grid->oSites();
   
-  inner=0.0;
-  for(int i=0;i<grid->SumArraySize();i++){
-    inner = inner+sumarray[i*pad];
-  } 
-  right._grid->GlobalSum(inner);
-  return inner;
+#ifdef GRID_NVCC
+  // GPU - SIMT lane compliance...
+  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+  
+
+  accelerator_for( ss, sites, nsimd,{
+      auto x_l = left_v(ss);
+      auto y_l = right_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
+  })
+
+  // This is in single precision and fails some tests
+  // 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
+  grid->GlobalSum(nrm);
+
+  return nrm;
 }
 
 /////////////////////////
@@ -86,8 +157,7 @@ axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj
 template<class sobj,class vobj> strong_inline RealD 
 axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) 
 {
-  const int pad = 8;
-  z.checkerboard = x.checkerboard;
+  z.Checkerboard() = x.Checkerboard();
   conformable(z,x);
   conformable(x,y);
 
@@ -95,43 +165,57 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
-  GridBase *grid = x._grid;
+  GridBase *grid = x.Grid();
 
-  Vector<RealD> sumarray(grid->SumArraySize()*pad);
+  auto x_v=x.View();
+  auto y_v=y.View();
+  auto z_v=z.View();
 
-  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
-    int nwork, mywork, myoff;
-    GridThread::GetWork(x._grid->oSites(),thr,mywork,myoff);
+  const uint64_t nsimd = grid->Nsimd();
+  const uint64_t sites = grid->oSites();
   
-    // private to thread; sub summation
-    decltype(innerProductD(z._odata[0],z._odata[0])) vnrm=zero; 
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      vobj tmp = a*x._odata[ss]+b*y._odata[ss];
-      vnrm = vnrm + innerProductD(tmp,tmp);
-      vstream(z._odata[ss],tmp);
-    }
-    vstream(sumarray[thr*pad],real(Reduce(TensorRemove(vnrm)))) ;
-  }
+#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];
 
-  nrm = 0.0; // sum across threads; linear in thread count but fast
-  for(int i=0;i<grid->SumArraySize();i++){
-    nrm = nrm+sumarray[i*pad];
-  } 
-  z._grid->GlobalSum(nrm);
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
+  });
+
+  nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
+#else
+  // CPU 
+  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+  
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      inner_tmp_v[ss]=innerProductD(tmp,tmp);
+      z_v[ss]=tmp;
+  });
+  // Already promoted to double
+  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
+#endif
+  grid->GlobalSum(nrm);
   return nrm; 
 }
 
  
 template<class Op,class T1>
 inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
-  ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
+  ->typename decltype(expr.op.func(eval(0,expr.arg1)))::scalar_object
 {
   return sum(closure(expr));
 }
 
 template<class Op,class T1,class T2>
 inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
-      ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),eval(0,std::get<1>(expr.second))))::scalar_object
+      ->typename decltype(expr.op.func(eval(0,expr.arg1),eval(0,expr.arg2)))::scalar_object
 {
   return sum(closure(expr));
 }
@@ -139,54 +223,14 @@ inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
 
 template<class Op,class T1,class T2,class T3>
 inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
-  ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
-				      eval(0,std::get<1>(expr.second)),
-				      eval(0,std::get<2>(expr.second))
+  ->typename decltype(expr.op.func(eval(0,expr.arg1),
+				      eval(0,expr.arg2),
+				      eval(0,expr.arg3)
 				      ))::scalar_object
 {
   return sum(closure(expr));
 }
 
-template<class vobj>
-inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
-{
-  GridBase *grid=arg._grid;
-  int Nsimd = grid->Nsimd();
-  
-  std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
-  for(int i=0;i<grid->SumArraySize();i++){
-    sumarray[i]=zero;
-  }
-  
-  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
-    int nwork, mywork, myoff;
-    GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
-    
-    vobj vvsum=zero;
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      vvsum = vvsum + arg._odata[ss];
-    }
-    sumarray[thr]=vvsum;
-  }
-  
-  vobj vsum=zero;  // sum across threads
-  for(int i=0;i<grid->SumArraySize();i++){
-    vsum = vsum+sumarray[i];
-  } 
-  
-  typedef typename vobj::scalar_object sobj;
-  sobj ssum=zero;
-  
-  std::vector<sobj>               buf(Nsimd);
-  extract(vsum,buf);
-  
-  for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
-  arg._grid->GlobalSum(ssum);
-  
-  return ssum;
-}
-
-
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -199,7 +243,7 @@ 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;
-  GridBase  *grid = Data._grid;
+  GridBase  *grid = Data.Grid();
   assert(grid!=NULL);
 
   const int    Nd = grid->_ndimension;
@@ -212,13 +256,13 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vobj,alignedAllocator<vobj> > lvSum(rd); // will locally sum vectors first
-  std::vector<sobj> lsSum(ld,zero);                    // sum across these down to scalars
-  std::vector<sobj> extracted(Nsimd);                  // splitting the SIMD
+  Vector<vobj> lvSum(rd); // will locally sum vectors first
+  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
 
   result.resize(fd); // And then global sum to return the same vector to every node 
   for(int r=0;r<rd;r++){
-    lvSum[r]=zero;
+    lvSum[r]=Zero();
   }
 
   int e1=    grid->_slice_nblock[orthogdim];
@@ -227,20 +271,19 @@ 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
-  parallel_for(int r=0;r<rd;r++){
-
+  auto Data_v=Data.View();
+  thread_for( r,rd, {
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int ss= so+n*stride+b;
-	lvSum[r]=lvSum[r]+Data._odata[ss];
-      }
+	lvSum[r]=lvSum[r]+Data_v[ss];
       }
     }
+  });
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
-  std::vector<int> icoor(Nd);
+  Coordinate icoor(Nd);
 
   for(int rt=0;rt<rd;rt++){
 
@@ -265,7 +308,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
     if ( pt == grid->_processor_coor[orthogdim] ) {
       gsum=lsSum[lt];
     } else {
-      gsum=zero;
+      gsum=Zero();
     }
 
     grid->GlobalSum(gsum);
@@ -292,9 +335,9 @@ static void localSliceInnerProductVector(std::vector<ComplexD> &result, const La
   // std::cout << GridLogMessage << "Start prep" << std::endl;
   typedef typename vobj::vector_type   vector_type;
   typedef typename vobj::scalar_type   scalar_type;
-  GridBase  *grid = lhs._grid;
+  GridBase  *grid = lhs.Grid();
   assert(grid!=NULL);
-  conformable(grid,rhs._grid);
+  conformable(grid,rhs.Grid());
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
@@ -307,14 +350,14 @@ static void localSliceInnerProductVector(std::vector<ComplexD> &result, const La
   int rd=grid->_rdimensions[orthogdim];
   // std::cout << GridLogMessage << "Start alloc" << std::endl;
 
-  std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
+  Vector<vector_type> lvSum(rd); // will locally sum vectors first
   lsSum.resize(ld,scalar_type(0.0));                    // sum across these down to scalars
-  std::vector<iScalar<scalar_type>> extracted(Nsimd);   // splitting the SIMD  
+  ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
   // std::cout << GridLogMessage << "End alloc" << std::endl;
 
   result.resize(fd); // And then global sum to return the same vector to every node for IO to file
   for(int r=0;r<rd;r++){
-    lvSum[r]=zero;
+    lvSum[r]=Zero();
   }
 
   int e1=    grid->_slice_nblock[orthogdim];
@@ -323,23 +366,24 @@ static void localSliceInnerProductVector(std::vector<ComplexD> &result, const La
   // std::cout << GridLogMessage << "End prep" << std::endl;
   // std::cout << GridLogMessage << "Start parallel inner product, _rd = " << rd << std::endl;
   vector_type vv;
-  parallel_for(int r=0;r<rd;r++)
-  {
+  auto l_v=lhs.View();
+  auto r_v=rhs.View();
+  thread_for( r,rd,{
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int ss = so + n * stride + b;
-        vv = TensorRemove(innerProduct(lhs._odata[ss], rhs._odata[ss]));
+        vv = TensorRemove(innerProduct(l_v[ss], r_v[ss]));
         lvSum[r] = lvSum[r] + vv;
       }
     }
-  }
+  });
   // std::cout << GridLogMessage << "End parallel inner product" << std::endl;
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
-  std::vector<int> icoor(Nd);
+  Coordinate icoor(Nd);
   for(int rt=0;rt<rd;rt++){
 
     iScalar<vector_type> temp; 
@@ -362,7 +406,7 @@ template <class vobj>
 static void globalSliceInnerProductVector(std::vector<ComplexD> &result, const Lattice<vobj> &lhs, std::vector<typename vobj::scalar_type> &lsSum, int orthogdim)
 {
   typedef typename vobj::scalar_type scalar_type;
-  GridBase *grid = lhs._grid;
+  GridBase *grid = lhs.Grid();
   int fd = result.size();
   int ld = lsSum.size();
   // sum over nodes.
@@ -388,9 +432,9 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
 {
   typedef typename vobj::vector_type   vector_type;
   typedef typename vobj::scalar_type   scalar_type;
-  GridBase  *grid = lhs._grid;
+  GridBase  *grid = lhs.Grid();
   assert(grid!=NULL);
-  conformable(grid,rhs._grid);
+  conformable(grid,rhs.Grid());
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
@@ -402,34 +446,36 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
-  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
-  std::vector<iScalar<scalar_type> > extracted(Nsimd);                  // splitting the SIMD
+  Vector<vector_type> lvSum(rd); // will locally sum vectors first
+  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
 
   result.resize(fd); // And then global sum to return the same vector to every node for IO to file
   for(int r=0;r<rd;r++){
-    lvSum[r]=zero;
+    lvSum[r]=Zero();
   }
 
   int e1=    grid->_slice_nblock[orthogdim];
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
 
-  parallel_for(int r=0;r<rd;r++){
+  auto lhv=lhs.View();
+  auto rhv=rhs.View();
+  thread_for( r,rd,{
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int ss= so+n*stride+b;
-	vector_type vv = TensorRemove(innerProduct(lhs._odata[ss],rhs._odata[ss]));
+	vector_type vv = TensorRemove(innerProduct(lhv[ss],rhv[ss]));
 	lvSum[r]=lvSum[r]+vv;
       }
     }
-  }
+  });
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
-  std::vector<int> icoor(Nd);
+  Coordinate icoor(Nd);
   for(int rt=0;rt<rd;rt++){
 
     iScalar<vector_type> temp; 
@@ -470,7 +516,7 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
-  int Nblock = rhs._grid->GlobalDimensions()[Orthog];
+  int Nblock = rhs.Grid()->GlobalDimensions()[Orthog];
   std::vector<ComplexD> ip(Nblock);
   sn.resize(Nblock);
   
@@ -492,7 +538,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   
   scalar_type zscale(scale);
 
-  GridBase *grid  = X._grid;
+  GridBase *grid  = X.Grid();
 
   int Nsimd  =grid->Nsimd();
   int Nblock =grid->GlobalDimensions()[orthogdim];
@@ -505,8 +551,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   int e2     =grid->_slice_block [orthogdim];
   int stride =grid->_slice_stride[orthogdim];
 
-  std::vector<int> icoor;
-
+  Coordinate icoor;
   for(int r=0;r<rd;r++){
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
@@ -522,12 +567,15 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
 
     tensor_reduced at; at=av;
 
-    parallel_for_nest2(int n=0;n<e1;n++){
+    auto Rv=R.View();
+    auto Xv=X.View();
+    auto Yv=Y.View();
+    thread_for_collapse(2, n, e1, {
       for(int b=0;b<e2;b++){
 	int ss= so+n*stride+b;
-	R._odata[ss] = at*X._odata[ss]+Y._odata[ss];
-      }
+	Rv[ss] = at*Xv[ss]+Yv[ss];
       }
+    });
   }
 };
 
@@ -559,18 +607,18 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
-  int Nblock = X._grid->GlobalDimensions()[Orthog];
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
 
-  GridBase *FullGrid  = X._grid;
+  GridBase *FullGrid  = X.Grid();
   //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
 
   //  Lattice<vobj> Xslice(SliceGrid);
   //  Lattice<vobj> Rslice(SliceGrid);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
-  int nh =  FullGrid->_ndimension;
+  //  int nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
-  int nl = nh-1;
+  //  int nl = nh-1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -578,28 +626,31 @@ 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];
-#pragma omp parallel 
-  {
-    std::vector<vobj> s_x(Nblock);
 
-#pragma omp for collapse(2)
-    for(int n=0;n<nblock;n++){
+  auto X_v=X.View();
+  auto Y_v=Y.View();
+  auto R_v=R.View();
+  thread_region
+  {
+    Vector<vobj> s_x(Nblock);
+
+    thread_for_collapse_in_region(2, n,nblock, {
      for(int b=0;b<block;b++){
       int o  = n*stride + b;
 
       for(int i=0;i<Nblock;i++){
-	s_x[i] = X[o+i*ostride];
+	s_x[i] = X_v[o+i*ostride];
       }
 
       vobj dot;
       for(int i=0;i<Nblock;i++){
-	dot = Y[o+i*ostride];
+	dot = Y_v[o+i*ostride];
 	for(int j=0;j<Nblock;j++){
 	  dot = dot + s_x[j]*(scale*aa(j,i));
 	}
-	R[o+i*ostride]=dot;
+	R_v[o+i*ostride]=dot;
       }
-    }}
+    }});
   }
 };
 
@@ -610,17 +661,17 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
-  int Nblock = X._grid->GlobalDimensions()[Orthog];
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
 
-  GridBase *FullGrid  = X._grid;
+  GridBase *FullGrid  = X.Grid();
   //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
   //  Lattice<vobj> Xslice(SliceGrid);
   //  Lattice<vobj> Rslice(SliceGrid);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
-  int nh =  FullGrid->_ndimension;
+  //  int nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
-  int nl=1;
+  //  int nl=1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -628,17 +679,19 @@ static void sliceMulMatrix (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];
-#pragma omp parallel 
+  auto R_v = R.View();
+  auto X_v = X.View();
+  thread_region
   {
     std::vector<vobj> s_x(Nblock);
 
-#pragma omp for collapse(2)
-    for(int n=0;n<nblock;n++){
+
+    thread_for_collapse_in_region( 2 ,n,nblock,{
     for(int b=0;b<block;b++){
       int o  = n*stride + b;
 
       for(int i=0;i<Nblock;i++){
-	s_x[i] = X[o+i*ostride];
+	s_x[i] = X_v[o+i*ostride];
       }
 
       vobj dot;
@@ -647,11 +700,10 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
 	for(int j=1;j<Nblock;j++){
 	  dot = dot + s_x[j]*(scale*aa(j,i));
 	}
-	R[o+i*ostride]=dot;
+	R_v[o+i*ostride]=dot;
       }
-    }}
+    }});
   }
-
 };
 
 
@@ -662,7 +714,7 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
-  GridBase *FullGrid  = lhs._grid;
+  GridBase *FullGrid  = lhs.Grid();
   //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
   
   int Nblock = FullGrid->GlobalDimensions()[Orthog];
@@ -673,9 +725,9 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
-  int nh =  FullGrid->_ndimension;
+  //  int nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
-  int nl = nh-1;
+  //  int nl = nh-1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -686,31 +738,33 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
 
   typedef typename vobj::vector_typeD vector_typeD;
 
-#pragma omp parallel 
+  auto lhs_v=lhs.View();
+  auto rhs_v=rhs.View();
+  thread_region
   {
     std::vector<vobj> Left(Nblock);
     std::vector<vobj> Right(Nblock);
     Eigen::MatrixXcd  mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
 
-#pragma omp for collapse(2)
-    for(int n=0;n<nblock;n++){
+    thread_for_collapse_in_region( 2, n,nblock,{
     for(int b=0;b<block;b++){
 
       int o  = n*stride + b;
 
       for(int i=0;i<Nblock;i++){
-	Left [i] = lhs[o+i*ostride];
-	Right[i] = rhs[o+i*ostride];
+	Left [i] = lhs_v[o+i*ostride];
+	Right[i] = rhs_v[o+i*ostride];
       }
 
       for(int i=0;i<Nblock;i++){
       for(int j=0;j<Nblock;j++){
 	auto tmp = innerProduct(Left[i],Right[j]);
 	auto rtmp = TensorRemove(tmp);
-	mat_thread(i,j) += Reduce(rtmp);
+	auto red  =  Reduce(rtmp);
+	mat_thread(i,j) += std::complex<double>(real(red),imag(red));
       }}
-    }}
-#pragma omp critical
+    }});
+    thread_critical
     {
       mat += mat_thread;
     }  
@@ -726,8 +780,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   return;
 }
 
-} /*END NAMESPACE GRID*/
-#endif
+NAMESPACE_END(Grid);
+
 
 
 

Grid/lattice/Lattice_reduction_gpu.h

@@ -0,0 +1,226 @@
+NAMESPACE_BEGIN(Grid);
+
+#define WARP_SIZE 32
+extern cudaDeviceProp *gpu_props;
+__device__ unsigned int retirementCount = 0;
+
+template <class Iterator>
+unsigned int nextPow2(Iterator x) {
+  --x;
+  x |= x >> 1;
+  x |= x >> 2;
+  x |= x >> 4;
+  x |= x >> 8;
+  x |= x >> 16;
+  return ++x;
+}
+
+template <class Iterator>
+void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+  
+  int device;
+  cudaGetDevice(&device);
+  
+  Iterator warpSize            = gpu_props[device].warpSize;
+  Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
+  Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
+  Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
+  
+  std::cout << GridLogDebug << "GPU has:" << std::endl;
+  std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
+  std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
+  std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
+  std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << warpSize << std::endl;
+  std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
+  
+  if (warpSize != WARP_SIZE) {
+    std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
+    exit(EXIT_FAILURE);
+  }
+  
+  // let the number of threads in a block be a multiple of 2, starting from warpSize
+  threads = warpSize;
+  while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
+  // keep all the streaming multiprocessors busy
+  blocks = nextPow2(multiProcessorCount);
+  
+}
+
+template <class sobj, class Iterator>
+__device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid) {
+  
+  Iterator blockSize = blockDim.x;
+  
+  // cannot use overloaded operators for sobj as they are not volatile-qualified
+  memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
+  __syncwarp();
+  
+  const Iterator VEC = WARP_SIZE;
+  const Iterator vid = tid & (VEC-1);
+  
+  sobj beta, temp;
+  memcpy((void *)&beta, (void *)&mySum, sizeof(sobj));
+  
+  for (int i = VEC/2; i > 0; i>>=1) {
+    if (vid < i) {
+      memcpy((void *)&temp, (void *)&sdata[tid+i], sizeof(sobj));
+      beta += temp;
+      memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
+    }
+    __syncwarp();
+  }
+  __syncthreads();
+  
+  if (threadIdx.x == 0) {
+    beta  = Zero();
+    for (Iterator i = 0; i < blockSize; i += VEC) {
+      memcpy((void *)&temp, (void *)&sdata[i], sizeof(sobj));
+      beta  += temp;
+    }
+    memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
+  }
+  __syncthreads();
+}
+
+
+template <class vobj, class sobj, class Iterator>
+__device__ void reduceBlocks(const vobj *g_idata, sobj *g_odata, Iterator n) 
+{
+  constexpr Iterator nsimd = vobj::Nsimd();
+  
+  Iterator blockSize = blockDim.x;
+  
+  // force shared memory alignment
+  extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
+  // it's not possible to have two extern __shared__ arrays with same name
+  // but different types in different scopes -- need to cast each time
+  sobj *sdata = (sobj *)shmem_pointer;
+  
+  // first level of reduction,
+  // each thread writes result in mySum
+  Iterator tid = threadIdx.x;
+  Iterator i = blockIdx.x*(blockSize*2) + threadIdx.x;
+  Iterator gridSize = blockSize*2*gridDim.x;
+  sobj mySum = Zero();
+  
+  while (i < n) {
+    Iterator lane = i % nsimd;
+    Iterator ss   = i / nsimd;
+    auto tmp = extractLane(lane,g_idata[ss]);
+    sobj tmpD;
+    tmpD=tmp;
+    mySum   +=tmpD;
+    
+    if (i + blockSize < n) {
+      lane = (i+blockSize) % nsimd;
+      ss   = (i+blockSize) / nsimd;
+      tmp = extractLane(lane,g_idata[ss]);
+      tmpD = tmp;
+      mySum += tmpD;
+    }
+    i += gridSize;
+  }
+  
+  // copy mySum to shared memory and perform
+  // reduction for all threads in this block
+  reduceBlock(sdata, mySum, tid);
+  if (tid == 0) g_odata[blockIdx.x] = sdata[0];
+}
+
+template <class vobj, class sobj,class Iterator>
+__global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
+  
+  Iterator blockSize = blockDim.x;
+  
+  // perform reduction for this block and
+  // write result to global memory buffer
+  reduceBlocks(lat, buffer, n);
+  
+  if (gridDim.x > 1) {
+    
+    const Iterator tid = threadIdx.x;
+    __shared__ bool amLast;
+    // force shared memory alignment
+    extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
+    // it's not possible to have two extern __shared__ arrays with same name
+    // but different types in different scopes -- need to cast each time
+    sobj *smem = (sobj *)shmem_pointer;
+    
+    // wait until all outstanding memory instructions in this thread are finished
+    __threadfence();
+    
+    if (tid==0) {
+      unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
+      // true if this block is the last block to be done
+      amLast = (ticket == gridDim.x-1);
+    }
+    
+    // each thread must read the correct value of amLast
+    __syncthreads();
+    
+    if (amLast) {
+      // reduce buffer[0], ..., buffer[gridDim.x-1]
+      Iterator i = tid;
+      sobj mySum = Zero();
+      
+      while (i < gridDim.x) {
+        mySum += buffer[i];
+        i += blockSize;
+      }
+      
+      reduceBlock(smem, mySum, tid);
+      
+      if (tid==0) {
+        buffer[0] = smem[0];
+        // reset count variable
+        retirementCount = 0;
+      }
+    }
+  }
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Possibly promote to double and sum
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_objectD sobj;
+  typedef decltype(lat) Iterator;
+  
+  Integer nsimd= vobj::Nsimd();
+  Integer size = osites*nsimd;
+
+  Integer numThreads, numBlocks;
+  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  Integer smemSize = numThreads * sizeof(sobj);
+
+  Vector<sobj> buffer(numBlocks);
+  sobj *buffer_v = &buffer[0];
+  
+  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  cudaDeviceSynchronize();
+  
+  cudaError err = cudaGetLastError();
+  if ( cudaSuccess != err ) {
+    printf("Cuda error %s\n",cudaGetErrorString( err ));
+    exit(0);
+  }
+  auto result = buffer_v[0];
+  return result;
+}
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Return as same precision as input performing reduction in double precision though
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu(lat,osites);
+  return result;
+}
+
+
+
+NAMESPACE_END(Grid);

Grid/lattice/Lattice_rng.h

@@ -41,7 +41,7 @@
 #undef  RNG_FAST_DISCARD
 #endif
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////
 // Allow the RNG state to be less dense than the fine grid
@@ -108,12 +108,16 @@ namespace Grid {
 template<class distribution,class generator> 
 void fillScalar(ComplexF &s,distribution &dist, generator &gen)
 {
-    s=ComplexF(dist(gen),dist(gen));
+  //  s=ComplexF(dist(gen),dist(gen));
+  s.real(dist(gen));
+  s.imag(dist(gen));
 }
 template<class distribution,class generator> 
 void fillScalar(ComplexD &s,distribution &dist,generator &gen)
 {
-    s=ComplexD(dist(gen),dist(gen));
+  //  s=ComplexD(dist(gen),dist(gen));
+  s.real(dist(gen));
+  s.imag(dist(gen));
 }
   
 class GridRNGbase {
@@ -165,7 +169,10 @@ namespace Grid {
     //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
     const int shift = 30;
 
-      uint64_t skip = site;
+    ////////////////////////////////////////////////////////////////////
+    // Weird compiler bug in Intel 2018.1 under O3 was generating 32bit and not 64 bit left shift.
+    ////////////////////////////////////////////////////////////////////
+    volatile uint64_t skip = site;
 
     skip = skip<<shift;
 
@@ -256,7 +263,7 @@ namespace Grid {
 
     CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
 
-    };
+  }
 
   template <class distribution>  inline void fill(ComplexF &l,std::vector<distribution> &dist){
     dist[0].reset();
@@ -333,13 +340,13 @@ namespace Grid {
 };
 
 class GridParallelRNG : public GridRNGbase {
-
+private:
   double _time_counter;
-
-  public:
   GridBase *_grid;
   unsigned int _vol;
 
+public:
+  GridBase *Grid(void) const { return _grid; }
   int generator_idx(int os,int is) {
     return is*_grid->oSites()+os;
   }
@@ -363,13 +370,14 @@ namespace Grid {
 
     double inner_time_counter = usecond();
 
-      int multiplicity = RNGfillable_general(_grid, l._grid); // l has finer or same grid
-      int Nsimd  = _grid->Nsimd();  // guaranteed to be the same for l._grid too
-      int osites = _grid->oSites();  // guaranteed to be <= l._grid->oSites() by a factor multiplicity
+    int multiplicity = RNGfillable_general(_grid, l.Grid()); // l has finer or same grid
+    int Nsimd  = _grid->Nsimd();  // guaranteed to be the same for l.Grid() too
+    int osites = _grid->oSites();  // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
     int words  = sizeof(scalar_object) / sizeof(scalar_type);
 
-      parallel_for(int ss=0;ss<osites;ss++){
-        std::vector<scalar_object> buf(Nsimd);
+    auto l_v = l.View();
+    thread_for( ss, osites, {
+      ExtractBuffer<scalar_object> buf(Nsimd);
       for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
 
 	int sm = multiplicity * ss + m;  // Maps the generator site to the fine site
@@ -383,23 +391,20 @@ namespace Grid {
 	    fillScalar(pointer[idx], dist[gdx], _generators[gdx]);
 	}
 	// merge into SIMD lanes, FIXME suboptimal implementation
-          merge(l._odata[sm], buf);
-        }
+	merge(l_v[sm], buf);
       }
+      });
+    //    });
 
     _time_counter += usecond()- inner_time_counter;
-    };
+  }
 
     void SeedUniqueString(const std::string &s){
       std::vector<int> seeds;
-      std::stringstream sha;
       seeds = GridChecksum::sha256_seeds(s);
-      for(int i=0;i<seeds.size();i++) { 
-        sha << std::hex << seeds[i];
-      }
       std::cout << GridLogMessage << "Intialising parallel RNG with unique string '" 
                 << s << "'" << std::endl;
-      std::cout << GridLogMessage << "Seed SHA256: " << sha.str() << std::endl;
+      std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
       SeedFixedIntegers(seeds);
     }
   void SeedFixedIntegers(const std::vector<int> &seeds){
@@ -421,12 +426,13 @@ namespace Grid {
     ////////////////////////////////////////////////
 
     // Everybody loops over global volume.
-      parallel_for(int gidx=0;gidx<_grid->_gsites;gidx++){
-
+    thread_for( gidx, _grid->_gsites, {
 	// Where is it?
-	int rank,o_idx,i_idx;
-	std::vector<int> gcoor;
+	int rank;
+	int o_idx;
+	int i_idx;
 
+	Coordinate gcoor;
 	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
 	
@@ -436,8 +442,7 @@ namespace Grid {
 	  _generators[l_idx] = master_engine;
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
-
-      }
+    });
 #else 
     ////////////////////////////////////////////////////////////////
     // Machine and thread decomposition dependent seeding is efficient
@@ -463,7 +468,7 @@ namespace Grid {
 	seeders[t] = Reseed(master_engine);
       }
 
-	parallel_for(int t=0;t<Nthread;t++) {
+      thread_for( t, Nthread, {
 	// set up one per local site in threaded fashion
 	std::vector<uint32_t> newseeds;
 	std::uniform_int_distribution<uint32_t> uid;	
@@ -472,7 +477,7 @@ namespace Grid {
 	    _generators[l] = Reseed(seeders[t],newseeds,uid);
 	  }
 	}
-	}
+      });
     }
 #endif
   }
@@ -490,8 +495,8 @@ namespace Grid {
 
     uint32_t the_number;
     // who
-      std::vector<int> gcoor;
     int rank,o_idx,i_idx;
+    Coordinate gcoor;
     _grid->GlobalIndexToGlobalCoor(gsite,gcoor);
     _grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
 
@@ -516,5 +521,5 @@ namespace Grid {
 template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
 template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_trace.h

@@ -32,19 +32,20 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 // Tracing, transposing, peeking, poking
 ///////////////////////////////////////////////
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Trace
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj>
-    inline auto trace(const Lattice<vobj> &lhs)
-      -> Lattice<decltype(trace(lhs._odata[0]))>
+inline auto trace(const Lattice<vobj> &lhs)  -> Lattice<decltype(trace(vobj()))>
 {
-      Lattice<decltype(trace(lhs._odata[0]))> ret(lhs._grid);
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-            ret._odata[ss] = trace(lhs._odata[ss]);
-        }
+  Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
+    coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
+  });
   return ret;
 };
     
@@ -52,16 +53,17 @@ namespace Grid {
 // Trace Index level dependent operation
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<int Index,class vobj>
-    inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(lhs._odata[0]))>
+inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
 {
-      Lattice<decltype(traceIndex<Index>(lhs._odata[0]))> ret(lhs._grid);
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-	ret._odata[ss] = traceIndex<Index>(lhs._odata[ss]);
-      }
+  Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
+    coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
+  });
   return ret;
 };
 
-
-}
+NAMESPACE_END(Grid);
 #endif
 

Grid/lattice/Lattice_transfer.h

@@ -25,10 +25,9 @@ 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_LATTICE_TRANSFER_H
-#define GRID_LATTICE_TRANSFER_H
+#pragma once
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 inline void subdivides(GridBase *coarse,GridBase *fine)
 {
@@ -49,34 +48,39 @@ 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){
-    half.checkerboard = cb;
+  half.Checkerboard() = cb;
 
-    parallel_for(int ss=0;ss<full._grid->oSites();ss++){
+  auto half_v = half.View();
+  auto full_v = full.View();
+  thread_for(ss, full.Grid()->oSites(),{
     int cbos;
-      std::vector<int> coor;
-      full._grid->oCoorFromOindex(coor,ss);
-      cbos=half._grid->CheckerBoard(coor);
+    Coordinate coor;
+    full.Grid()->oCoorFromOindex(coor,ss);
+    cbos=half.Grid()->CheckerBoard(coor);
 
     if (cbos==cb) {
-	int ssh=half._grid->oIndex(coor);
-	half._odata[ssh] = full._odata[ss];
-      }
+      int ssh=half.Grid()->oIndex(coor);
+      half_v[ssh] = full_v[ss];
     }
+  });
 }
 template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
-    int cb = half.checkerboard;
-    parallel_for(int ss=0;ss<full._grid->oSites();ss++){
-      std::vector<int> coor;
+  int cb = half.Checkerboard();
+  auto half_v = half.View();
+  auto full_v = full.View();
+  thread_for(ss,full.Grid()->oSites(),{
+
+    Coordinate coor;
     int cbos;
 
-      full._grid->oCoorFromOindex(coor,ss);
-      cbos=half._grid->CheckerBoard(coor);
+    full.Grid()->oCoorFromOindex(coor,ss);
+    cbos=half.Grid()->CheckerBoard(coor);
       
     if (cbos==cb) {
-	int ssh=half._grid->oIndex(coor);
-	full._odata[ss]=half._odata[ssh];
-      }
+      int ssh=half.Grid()->oIndex(coor);
+      full_v[ss]=half_v[ssh];
     }
+  });
 }
   
 
@@ -85,8 +89,8 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
 			 const             Lattice<vobj>   &fineData,
 			 const std::vector<Lattice<vobj> > &Basis)
 {
-  GridBase * fine  = fineData._grid;
-  GridBase * coarse= coarseData._grid;
+  GridBase * fine  = fineData.Grid();
+  GridBase * coarse= coarseData.Grid();
   int  _ndimension = coarse->_ndimension;
 
   // checks
@@ -96,33 +100,33 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     conformable(Basis[i],fineData);
   }
 
-  std::vector<int>  block_r      (_ndimension);
+  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]);
   }
 
-  coarseData=zero;
+  coarseData=Zero();
 
+  auto fineData_   = fineData.View();
+  auto coarseData_ = coarseData.View();
   // Loop over coars parallel, and then loop over fine associated with coarse.
-  parallel_for(int sf=0;sf<fine->oSites();sf++){
-
+  thread_for( sf, fine->oSites(), {
     int sc;
-    std::vector<int> coor_c(_ndimension);
-    std::vector<int> coor_f(_ndimension);
+    Coordinate coor_c(_ndimension);
+    Coordinate coor_f(_ndimension);
     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);
 
-PARALLEL_CRITICAL
+    thread_critical {
       for(int i=0;i<nbasis;i++) {
-
-      coarseData._odata[sc](i)=coarseData._odata[sc](i)
-	+ innerProduct(Basis[i]._odata[sf],fineData._odata[sf]);
-
+	auto Basis_      = Basis[i].View();
+	coarseData_[sc](i)=coarseData_[sc](i) + innerProduct(Basis_[sf],fineData_[sf]);
       }
     }
+  });
   return;
 }
 
@@ -132,18 +136,18 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
 		       const Lattice<vobj> &fineX,
 		       const Lattice<vobj> &fineY)
 {
-  GridBase * fine  = fineZ._grid;
-  GridBase * coarse= coarseA._grid;
+  GridBase * fine  = fineZ.Grid();
+  GridBase * coarse= coarseA.Grid();
 
-  fineZ.checkerboard=fineX.checkerboard;
-  assert(fineX.checkerboard==fineY.checkerboard);
+  fineZ.Checkerboard()=fineX.Checkerboard();
+  assert(fineX.Checkerboard()==fineY.Checkerboard());
   subdivides(coarse,fine); // require they map
   conformable(fineX,fineY);
   conformable(fineX,fineZ);
 
   int _ndimension = coarse->_ndimension;
   
-  std::vector<int>  block_r      (_ndimension);
+  Coordinate  block_r      (_ndimension);
 
   // FIXME merge with subdivide checking routine as this is redundant
   for(int d=0 ; d<_ndimension;d++){
@@ -151,20 +155,25 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
     assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
-  parallel_for(int sf=0;sf<fine->oSites();sf++){
+  auto fineZ_  = fineZ.View();
+  auto fineX_  = fineX.View();
+  auto fineY_  = fineY.View();
+  auto coarseA_= coarseA.View();
+
+  thread_for(sf, fine->oSites(), {
     
     int sc;
-    std::vector<int> coor_c(_ndimension);
-    std::vector<int> coor_f(_ndimension);
+    Coordinate coor_c(_ndimension);
+    Coordinate coor_f(_ndimension);
 
     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);
 
     // z = A x + y
-    fineZ._odata[sf]=coarseA._odata[sc]*fineX._odata[sf]+fineY._odata[sf];
+    fineZ_[sf]=coarseA_[sc]*fineX_[sf]+fineY_[sf];
 
-  }
+  });
 
   return;
 }
@@ -173,26 +182,29 @@ template<class vobj,class CComplex>
 			      const Lattice<vobj> &fineX,
 			      const Lattice<vobj> &fineY)
 {
-  typedef decltype(innerProduct(fineX._odata[0],fineY._odata[0])) dotp;
+  typedef decltype(innerProduct(vobj(),vobj())) dotp;
 
-  GridBase *coarse(CoarseInner._grid);
-  GridBase *fine  (fineX._grid);
+  GridBase *coarse(CoarseInner.Grid());
+  GridBase *fine  (fineX.Grid());
 
-  Lattice<dotp> fine_inner(fine); fine_inner.checkerboard = fineX.checkerboard;
+  Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
   Lattice<dotp> coarse_inner(coarse);
 
   // Precision promotion?
+  auto CoarseInner_  = CoarseInner.View();
+  auto coarse_inner_ = coarse_inner.View();
+
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  parallel_for(int ss=0;ss<coarse->oSites();ss++){
-    CoarseInner._odata[ss] = coarse_inner._odata[ss];
-  }
+  thread_for(ss, coarse->oSites(),{
+    CoarseInner_[ss] = coarse_inner_[ss];
+  });
 }
 template<class vobj,class CComplex>
 inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 {
-  GridBase *coarse = ip._grid;
-  Lattice<vobj> zz(fineX._grid); zz=zero; zz.checkerboard=fineX.checkerboard;
+  GridBase *coarse = ip.Grid();
+  Lattice<vobj> zz(fineX.Grid()); zz=Zero(); zz.Checkerboard()=fineX.Checkerboard();
   blockInnerProduct(ip,fineX,fineX);
   ip = pow(ip,-0.5);
   blockZAXPY(fineX,ip,fineX,zz);
@@ -202,14 +214,14 @@ inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 template<class vobj>
 inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 {
-  GridBase * fine  = fineData._grid;
-  GridBase * coarse= coarseData._grid;
+  GridBase * fine  = fineData.Grid();
+  GridBase * coarse= coarseData.Grid();
 
   subdivides(coarse,fine); // require they map
 
   int _ndimension = coarse->_ndimension;
   
-  std::vector<int>  block_r      (_ndimension);
+  Coordinate  block_r      (_ndimension);
   
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
@@ -217,36 +229,36 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 
   // Turn this around to loop threaded over sc and interior loop 
   // over sf would thread better
-  coarseData=zero;
-  parallel_region {
+  coarseData=Zero();
+  auto coarseData_ = coarseData.View();
+  auto fineData_   = fineData.View();
 
+  thread_for(sf,fine->oSites(),{
     int sc;
-    std::vector<int> coor_c(_ndimension);
-    std::vector<int> coor_f(_ndimension);
-
-    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
+    Coordinate coor_c(_ndimension);
+    Coordinate coor_f(_ndimension);
     
     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);
     
-PARALLEL_CRITICAL
-      coarseData._odata[sc]=coarseData._odata[sc]+fineData._odata[sf];
+    thread_critical { 
+      coarseData_[sc]=coarseData_[sc]+fineData_[sf];
+    }
 
-    }
-  }
+  });
   return;
 }
 
 template<class vobj>
-inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vobj> &picked,std::vector<int> coor)
+inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vobj> &picked,Coordinate coor)
 {
-  GridBase * fine = unpicked._grid;
+  GridBase * fine = unpicked.Grid();
 
-  Lattice<vobj> zz(fine); zz.checkerboard = unpicked.checkerboard;
+  Lattice<vobj> zz(fine); zz.Checkerboard() = unpicked.Checkerboard();
   Lattice<iScalar<vInteger> > fcoor(fine);
 
-  zz = zero;
+  zz = Zero();
 
   picked = unpicked;
   for(int d=0;d<fine->_ndimension;d++){
@@ -262,16 +274,15 @@ inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vob
 template<class vobj,class CComplex>
 inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> > &Basis)
 {
-  GridBase *coarse = ip._grid;
-  GridBase *fine   = Basis[0]._grid;
+  GridBase *coarse = ip.Grid();
+  GridBase *fine   = Basis[0].Grid();
 
   int       nbasis = Basis.size() ;
-  int  _ndimension = coarse->_ndimension;
 
   // checks
   subdivides(coarse,fine); 
   for(int i=0;i<nbasis;i++){
-    conformable(Basis[i]._grid,fine);
+    conformable(Basis[i].Grid(),fine);
   }
 
   for(int v=0;v<nbasis;v++) {
@@ -290,41 +301,41 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 			 Lattice<vobj>   &fineData,
 			 const std::vector<Lattice<vobj> > &Basis)
 {
-  GridBase * fine  = fineData._grid;
-  GridBase * coarse= coarseData._grid;
+  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]._grid,fine);
+    conformable(Basis[i].Grid(),fine);
   }
 
-  std::vector<int>  block_r      (_ndimension);
+  Coordinate  block_r      (_ndimension);
   
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
+  auto fineData_   = fineData.View();
+  auto coarseData_ = coarseData.View();
 
   // Loop with a cache friendly loop ordering
-  parallel_region {
+  thread_for(sf,fine->oSites(),{
     int sc;
-    std::vector<int> coor_c(_ndimension);
-    std::vector<int> coor_f(_ndimension);
-
-    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
+    Coordinate coor_c(_ndimension);
+    Coordinate coor_f(_ndimension);
 
     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);
 
     for(int i=0;i<nbasis;i++) {
-	if(i==0) fineData._odata[sf]=coarseData._odata[sc](i) * Basis[i]._odata[sf];
-	else     fineData._odata[sf]=fineData._odata[sf]+coarseData._odata[sc](i)*Basis[i]._odata[sf];
-      }
-    }
+      auto basis_ = Basis[i].View();
+      if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf];
+      else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf];
     }
+  });
   return;
   
 }
@@ -337,8 +348,8 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
   typedef typename vobj::scalar_object sobj;
   typedef typename vvobj::scalar_object ssobj;
 
-  GridBase *ig = in._grid;
-  GridBase *og = out._grid;
+  GridBase *ig = in.Grid();
+  GridBase *og = out.Grid();
 
   int ni = ig->_ndimension;
   int no = og->_ndimension;
@@ -351,16 +362,16 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
     assert(ig->lSites() == og->lSites());
   }
 
-  parallel_for(int idx=0;idx<ig->lSites();idx++){
+  thread_for(idx, ig->lSites(),{
     sobj s;
     ssobj ss;
 
-    std::vector<int> lcoor(ni);
+    Coordinate lcoor(ni);
     ig->LocalIndexToLocalCoor(idx,lcoor);
     peekLocalSite(s,in,lcoor);
     ss=s;
     pokeLocalSite(ss,out,lcoor);
-  }
+  });
 }
 
 
@@ -369,8 +380,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
 {
   typedef typename vobj::scalar_object sobj;
 
-  GridBase *lg = lowDim._grid;
-  GridBase *hg = higherDim._grid;
+  GridBase *lg = lowDim.Grid();
+  GridBase *hg = higherDim.Grid();
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
@@ -389,10 +400,10 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
   }
 
   // the above should guarantee that the operations are local
-  parallel_for(int idx=0;idx<lg->lSites();idx++){
+  thread_for(idx,lg->lSites(),{
     sobj s;
-    std::vector<int> lcoor(nl);
-    std::vector<int> hcoor(nh);
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
     int ddl=0;
     hcoor[orthog] = slice;
@@ -403,7 +414,7 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
     }
     peekLocalSite(s,lowDim,lcoor);
     pokeLocalSite(s,higherDim,hcoor);
-  }
+  });
 }
 
 template<class vobj>
@@ -411,8 +422,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 {
   typedef typename vobj::scalar_object sobj;
 
-  GridBase *lg = lowDim._grid;
-  GridBase *hg = higherDim._grid;
+  GridBase *lg = lowDim.Grid();
+  GridBase *hg = higherDim.Grid();
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
@@ -430,10 +441,10 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     }
   }
   // the above should guarantee that the operations are local
-  parallel_for(int idx=0;idx<lg->lSites();idx++){
+  thread_for(idx,lg->lSites(),{
     sobj s;
-    std::vector<int> lcoor(nl);
-    std::vector<int> hcoor(nh);
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
     int ddl=0;
     hcoor[orthog] = slice;
@@ -444,7 +455,7 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     }
     peekLocalSite(s,higherDim,hcoor);
     pokeLocalSite(s,lowDim,lcoor);
-  }
+  });
 
 }
 
@@ -454,8 +465,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 {
   typedef typename vobj::scalar_object sobj;
 
-  GridBase *lg = lowDim._grid;
-  GridBase *hg = higherDim._grid;
+  GridBase *lg = lowDim.Grid();
+  GridBase *hg = higherDim.Grid();
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
@@ -464,15 +475,17 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
+    if ( d!=orthog ) {
     assert(lg->_processors[d]  == hg->_processors[d]);
     assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
   }
+  }
 
   // the above should guarantee that the operations are local
-  parallel_for(int idx=0;idx<lg->lSites();idx++){
+  thread_for(idx,lg->lSites(),{
     sobj s;
-    std::vector<int> lcoor(nl);
-    std::vector<int> hcoor(nh);
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
@@ -480,17 +493,17 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       peekLocalSite(s,lowDim,lcoor);
       pokeLocalSite(s,higherDim,hcoor);
     }
-  }
+  });
 }
 
 
 template<class vobj>
-void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
+void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
   typedef typename vobj::scalar_object sobj;
 
-  GridBase *lg = lowDim._grid;
-  GridBase *hg = higherDim._grid;
+  GridBase *lg = lowDim.Grid();
+  GridBase *hg = higherDim.Grid();
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
@@ -499,15 +512,17 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slic
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
+    if ( d!=orthog ) {
     assert(lg->_processors[d]  == hg->_processors[d]);
     assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
   }
+  }
 
   // the above should guarantee that the operations are local
-  parallel_for(int idx=0;idx<lg->lSites();idx++){
+  thread_for(idx,lg->lSites(),{
     sobj s;
-    std::vector<int> lcoor(nl);
-    std::vector<int> hcoor(nh);
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
@@ -515,7 +530,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slic
       peekLocalSite(s,higherDim,hcoor);
       pokeLocalSite(s,lowDim,lcoor);
     }
-  }
+  });
 }
 
 
@@ -524,8 +539,8 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
 {
   typedef typename vobj::scalar_object sobj;
 
-  GridBase *cg = coarse._grid;
-  GridBase *fg =   fine._grid;
+  GridBase *cg = coarse.Grid();
+  GridBase *fg =   fine.Grid();
 
   int nd = cg->_ndimension;
 
@@ -533,14 +548,14 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
   assert(cg->_ndimension==fg->_ndimension);
 
-  std::vector<int> ratio(cg->_ndimension);
+  Coordinate ratio(cg->_ndimension);
 
   for(int d=0;d<cg->_ndimension;d++){
     ratio[d] = fg->_fdimensions[d]/cg->_fdimensions[d];
   }
 
-  std::vector<int> fcoor(nd);
-  std::vector<int> ccoor(nd);
+  Coordinate fcoor(nd);
+  Coordinate ccoor(nd);
   for(int g=0;g<fg->gSites();g++){
 
     fg->GlobalIndexToGlobalCoor(g,fcoor);
@@ -563,41 +578,46 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
 
   typedef typename vobj::vector_type vtype;
   
-  GridBase* in_grid = in._grid;
+  GridBase* in_grid = in.Grid();
   out.resize(in_grid->lSites());
   
   int ndim = in_grid->Nd();
   int in_nsimd = vtype::Nsimd();
 
-  std::vector<std::vector<int> > in_icoor(in_nsimd);
+  std::vector<Coordinate > in_icoor(in_nsimd);
       
   for(int lane=0; lane < in_nsimd; lane++){
     in_icoor[lane].resize(ndim);
     in_grid->iCoorFromIindex(in_icoor[lane], lane);
   }
 
-  parallel_for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
+  //loop over outer index
+  auto in_v  = in.View();
+  thread_for(in_oidx,in_grid->oSites(),{
     //Assemble vector of pointers to output elements
-    std::vector<sobj*> out_ptrs(in_nsimd);
+    ExtractPointerArray<sobj> out_ptrs(in_nsimd);
 
-    std::vector<int> in_ocoor(ndim);
+    Coordinate in_ocoor(ndim);
     in_grid->oCoorFromOindex(in_ocoor, in_oidx);
 
-    std::vector<int> lcoor(in_grid->Nd());
+    Coordinate lcoor(in_grid->Nd());
       
     for(int lane=0; lane < in_nsimd; lane++){
-      for(int mu=0;mu<ndim;mu++)
+
+      for(int mu=0;mu<ndim;mu++){
 	lcoor[mu] = in_ocoor[mu] + in_grid->_rdimensions[mu]*in_icoor[lane][mu];
+      }
 
       int lex;
       Lexicographic::IndexFromCoor(lcoor, lex, in_grid->_ldimensions);
+      assert(lex < out.size());
       out_ptrs[lane] = &out[lex];
     }
     
     //Unpack into those ptrs
-    const vobj & in_vobj = in._odata[in_oidx];
-    extract1(in_vobj, out_ptrs, 0);
-  }
+    const vobj & in_vobj = in_v[in_oidx];
+    extract(in_vobj, out_ptrs, 0);
+  });
 }
 
 template<typename vobj, typename sobj>
@@ -613,21 +633,21 @@ unvectorizeToRevLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
   int ndim = in_grid->Nd();
   int in_nsimd = vtype::Nsimd();
 
-  std::vector<std::vector<int> > in_icoor(in_nsimd);
+  std::vector<Coordinate > in_icoor(in_nsimd);
       
   for(int lane=0; lane < in_nsimd; lane++){
     in_icoor[lane].resize(ndim);
     in_grid->iCoorFromIindex(in_icoor[lane], lane);
   }
   
-  parallel_for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
+  thread_for(in_oidx, in_grid->oSites(),{
     //Assemble vector of pointers to output elements
     std::vector<sobj*> out_ptrs(in_nsimd);
 
-    std::vector<int> in_ocoor(ndim);
+    Coordinate in_ocoor(ndim);
     in_grid->oCoorFromOindex(in_ocoor, in_oidx);
 
-    std::vector<int> lcoor(in_grid->Nd());
+    Coordinate lcoor(in_grid->Nd());
       
     for(int lane=0; lane < in_nsimd; lane++){
       for(int mu=0;mu<ndim;mu++)
@@ -641,7 +661,7 @@ unvectorizeToRevLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
     //Unpack into those ptrs
     const vobj & in_vobj = in._odata[in_oidx];
     extract1(in_vobj, out_ptrs, 0);
-  }
+  });
 }
 
 //Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
@@ -653,28 +673,27 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
 
   typedef typename vobj::vector_type vtype;
   
-  GridBase* grid = out._grid;
+  GridBase* grid = out.Grid();
   assert(in.size()==grid->lSites());
   
-  int ndim     = grid->Nd();
-  int nsimd    = vtype::Nsimd();
+  const int ndim     = grid->Nd();
+  constexpr int nsimd    = vtype::Nsimd();
 
-  std::vector<std::vector<int> > icoor(nsimd);
+  std::vector<Coordinate > icoor(nsimd);
       
   for(int lane=0; lane < nsimd; lane++){
     icoor[lane].resize(ndim);
     grid->iCoorFromIindex(icoor[lane],lane);
   }
-  
-  parallel_for(uint64_t oidx = 0; oidx < grid->oSites(); oidx++){ //loop over outer index
+  auto out_v = out.View();
+  thread_for(oidx, grid->oSites(),{
     //Assemble vector of pointers to output elements
-    std::vector<sobj*> ptrs(nsimd);
+    ExtractPointerArray<sobj> ptrs(nsimd);
 
-    std::vector<int> ocoor(ndim);
+    Coordinate ocoor(ndim);
+    Coordinate lcoor(ndim);
     grid->oCoorFromOindex(ocoor, oidx);
       
-    std::vector<int> lcoor(grid->Nd());
-      
     for(int lane=0; lane < nsimd; lane++){
 
       for(int mu=0;mu<ndim;mu++){
@@ -688,9 +707,9 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
     
     //pack from those ptrs
     vobj vecobj;
-    merge1(vecobj, ptrs, 0);
-    out._odata[oidx] = vecobj; 
-  }
+    merge(vecobj, ptrs, 0);
+    out_v[oidx] = vecobj; 
+  });
 }
 
 template<typename vobj, typename sobj>
@@ -707,21 +726,21 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   int ndim     = grid->Nd();
   int nsimd    = vtype::Nsimd();
 
-  std::vector<std::vector<int> > icoor(nsimd);
+  std::vector<Coordinate > icoor(nsimd);
       
   for(int lane=0; lane < nsimd; lane++){
     icoor[lane].resize(ndim);
     grid->iCoorFromIindex(icoor[lane],lane);
   }
   
-  parallel_for(uint64_t oidx = 0; oidx < grid->oSites(); oidx++){ //loop over outer index
+  thread_for(oidx, grid->oSites(), {
     //Assemble vector of pointers to output elements
     std::vector<sobj*> ptrs(nsimd);
 
-    std::vector<int> ocoor(ndim);
+    Coordinate ocoor(ndim);
     grid->oCoorFromOindex(ocoor, oidx);
 
-    std::vector<int> lcoor(grid->Nd());
+    Coordinate lcoor(grid->Nd());
       
     for(int lane=0; lane < nsimd; lane++){
 
@@ -738,25 +757,28 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
     vobj vecobj;
     merge1(vecobj, ptrs, 0);
     out._odata[oidx] = vecobj; 
-  }
+  });
 }
 
 //Convert a Lattice from one precision to another
 template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
-  assert(out._grid->Nd() == in._grid->Nd());
-  assert(out._grid->FullDimensions() == in._grid->FullDimensions());
-  out.checkerboard = in.checkerboard;
-  GridBase *in_grid=in._grid;
-  GridBase *out_grid = out._grid;
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+{
+  assert(out.Grid()->Nd() == in.Grid()->Nd());
+  for(int d=0;d<out.Grid()->Nd();d++){
+    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
+  }
+  out.Checkerboard() = in.Checkerboard();
+  GridBase *in_grid=in.Grid();
+  GridBase *out_grid = out.Grid();
 
   typedef typename VobjOut::scalar_object SobjOut;
   typedef typename VobjIn::scalar_object SobjIn;
 
-  int ndim = out._grid->Nd();
+  int ndim = out.Grid()->Nd();
   int out_nsimd = out_grid->Nsimd();
     
-  std::vector<std::vector<int> > out_icoor(out_nsimd);
+  std::vector<Coordinate > out_icoor(out_nsimd);
       
   for(int lane=0; lane < out_nsimd; lane++){
     out_icoor[lane].resize(ndim);
@@ -766,13 +788,14 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
   std::vector<SobjOut> in_slex_conv(in_grid->lSites());
   unvectorizeToLexOrdArray(in_slex_conv, in);
     
-  parallel_for(uint64_t out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
-    std::vector<int> out_ocoor(ndim);
+  auto out_v = out.View();
+  thread_for(out_oidx,out_grid->oSites(),{
+    Coordinate out_ocoor(ndim);
     out_grid->oCoorFromOindex(out_ocoor, out_oidx);
 
-    std::vector<SobjOut*> ptrs(out_nsimd);      
+    ExtractPointerArray<SobjOut> ptrs(out_nsimd);      
 
-    std::vector<int> lcoor(out_grid->Nd());
+    Coordinate lcoor(out_grid->Nd());
       
     for(int lane=0; lane < out_nsimd; lane++){
       for(int mu=0;mu<ndim;mu++)
@@ -781,8 +804,8 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
       int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
       ptrs[lane] = &in_slex_conv[llex];
     }
-    merge(out._odata[out_oidx], ptrs, 0);
-  }
+    merge(out_v[out_oidx], ptrs, 0);
+  });
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -841,8 +864,8 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   assert(full_vecs>=1);
 
-  GridBase * full_grid = full[0]._grid;
-  GridBase *split_grid = split._grid;
+  GridBase * full_grid = full[0].Grid();
+  GridBase *split_grid = split.Grid();
 
   int       ndim  = full_grid->_ndimension;
   int  full_nproc = full_grid->_Nprocessors;
@@ -851,18 +874,18 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   ////////////////////////////////
   // Checkerboard management
   ////////////////////////////////
-  int cb = full[0].checkerboard;
-  split.checkerboard = cb;
+  int cb = full[0].Checkerboard();
+  split.Checkerboard() = cb;
 
   //////////////////////////////
   // Checks
   //////////////////////////////
   assert(full_grid->_ndimension==split_grid->_ndimension);
   for(int n=0;n<full_vecs;n++){
-    assert(full[n].checkerboard == cb);
+    assert(full[n].Checkerboard() == cb);
     for(int d=0;d<ndim;d++){
-      assert(full[n]._grid->_gdimensions[d]==split._grid->_gdimensions[d]);
-      assert(full[n]._grid->_fdimensions[d]==split._grid->_fdimensions[d]);
+      assert(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
+      assert(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
     }
   }
 
@@ -870,7 +893,7 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   assert(nvector*split_nproc==full_nproc);
   assert(nvector == full_vecs);
 
-  std::vector<int> ratio(ndim);
+  Coordinate ratio(ndim);
   for(int d=0;d<ndim;d++){
     ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
   }
@@ -883,13 +906,13 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   for(int v=0;v<nvector;v++){
     unvectorizeToLexOrdArray(scalardata,full[v]);    
-    parallel_for(int site=0;site<lsites;site++){
+    thread_for(site,lsites,{
       alldata[v*lsites+site] = scalardata[site];
-    }
+    });
   }
 
   int nvec = nvector; // Counts down to 1 as we collapse dims
-  std::vector<int> ldims = full_grid->_ldimensions;
+  Coordinate ldims = full_grid->_ldimensions;
 
   for(int d=ndim-1;d>=0;d--){
 
@@ -915,8 +938,8 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
       int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
 
       // Loop over reordered data post A2A
-      parallel_for(int c=0;c<chunk;c++){
-	std::vector<int> coor(ndim);
+      thread_for(c, chunk, {
+	Coordinate coor(ndim);
 	for(int m=0;m<M;m++){
 	  for(int s=0;s<sP;s++){
 	    
@@ -938,7 +961,7 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
 	  }
 	}
-      }
+      });
       ldims[d]*= ratio[d];
       lsites  *= ratio[d];
 
@@ -950,8 +973,8 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 template<class Vobj>
 void Grid_split(Lattice<Vobj> &full,Lattice<Vobj>   & split)
 {
-  int nvector = full._grid->_Nprocessors / split._grid->_Nprocessors;
-  std::vector<Lattice<Vobj> > full_v(nvector,full._grid);
+  int nvector = full.Grid()->_Nprocessors / split.Grid()->_Nprocessors;
+  std::vector<Lattice<Vobj> > full_v(nvector,full.Grid());
   for(int n=0;n<nvector;n++){
     full_v[n] = full;
   }
@@ -967,8 +990,8 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   assert(full_vecs>=1);
 
-  GridBase * full_grid = full[0]._grid;
-  GridBase *split_grid = split._grid;
+  GridBase * full_grid = full[0].Grid();
+  GridBase *split_grid = split.Grid();
 
   int       ndim  = full_grid->_ndimension;
   int  full_nproc = full_grid->_Nprocessors;
@@ -977,18 +1000,18 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   ////////////////////////////////
   // Checkerboard management
   ////////////////////////////////
-  int cb = full[0].checkerboard;
-  split.checkerboard = cb;
+  int cb = full[0].Checkerboard();
+  split.Checkerboard() = cb;
 
   //////////////////////////////
   // Checks
   //////////////////////////////
   assert(full_grid->_ndimension==split_grid->_ndimension);
   for(int n=0;n<full_vecs;n++){
-    assert(full[n].checkerboard == cb);
+    assert(full[n].Checkerboard() == cb);
     for(int d=0;d<ndim;d++){
-      assert(full[n]._grid->_gdimensions[d]==split._grid->_gdimensions[d]);
-      assert(full[n]._grid->_fdimensions[d]==split._grid->_fdimensions[d]);
+      assert(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
+      assert(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
     }
   }
 
@@ -996,7 +1019,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   assert(nvector*split_nproc==full_nproc);
   assert(nvector == full_vecs);
 
-  std::vector<int> ratio(ndim);
+  Coordinate ratio(ndim);
   for(int d=0;d<ndim;d++){
     ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
   }
@@ -1015,7 +1038,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   int nvec = 1;
   uint64_t rsites        = split_grid->lSites();
-  std::vector<int> rdims = split_grid->_ldimensions;
+  Coordinate rdims = split_grid->_ldimensions;
 
   for(int d=0;d<ndim;d++){
 
@@ -1034,8 +1057,8 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 	
       {
 	// Loop over reordered data post A2A
-	parallel_for(int c=0;c<chunk;c++){
-	  std::vector<int> coor(ndim);
+	thread_for(c, chunk,{
+	  Coordinate coor(ndim);
 	  for(int m=0;m<M;m++){
 	    for(int s=0;s<sP;s++){
 
@@ -1056,7 +1079,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 	      tmpdata[lex_c] = alldata[lex_r];
 	    }
 	  }
-	}
+        });
       }
 
       if ( split_grid->_processors[d] > 1 ) {
@@ -1072,14 +1095,12 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   lsites = full_grid->lSites();
   for(int v=0;v<nvector;v++){
-    //    assert(v<full.size());
-    parallel_for(int site=0;site<lsites;site++){
-      //      assert(v*lsites+site < alldata.size());
+    thread_for(site, lsites,{
       scalardata[site] = alldata[v*lsites+site];
-    }
+    });
     vectorizeFromLexOrdArray(scalardata,full[v]);    
   }
 }
 
-}
-#endif
+NAMESPACE_END(Grid);
+

Grid/lattice/Lattice_transpose.h

@@ -33,17 +33,19 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 // Transpose
 ///////////////////////////////////////////////
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Transpose
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj>
 inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
-    Lattice<vobj> ret(lhs._grid);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-      ret._odata[ss] = transpose(lhs._odata[ss]);
-    }
+  Lattice<vobj> ret(lhs.Grid());
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
+    coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
+  });
   return ret;
 };
     
@@ -51,13 +53,16 @@ namespace Grid {
 // Index level dependent transpose
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<int Index,class vobj>
-    inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(lhs._odata[0]))>
+inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
 {
-    Lattice<decltype(transposeIndex<Index>(lhs._odata[0]))> ret(lhs._grid);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
-      ret._odata[ss] = transposeIndex<Index>(lhs._odata[ss]);
-    }
+  Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
+  auto ret_v = ret.View();
+  auto lhs_v = lhs.View();
+  accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
+    coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
+  });
   return ret;
 };
-}
+
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_unary.h

@@ -31,54 +31,50 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_UNARY_H
 #define GRID_LATTICE_UNARY_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
-  template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs,RealD y){
-    Lattice<obj> ret(rhs._grid);
-    ret.checkerboard = rhs.checkerboard;
-    conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-      ret._odata[ss]=pow(rhs._odata[ss],y);
+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();
+  ret.Checkerboard() = rhs.Checkerboard();
+  accelerator_for(ss,rhs.size(),1,{
+      ret[ss]=pow(rhs[ss],y);
+  });
+  return ret_i;
 }
-    return ret;
-  }
-  template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs,Integer y){
-    Lattice<obj> ret(rhs._grid);
-    ret.checkerboard = rhs.checkerboard;
-    conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-      ret._odata[ss]=mod(rhs._odata[ss],y);
-    }
-    return ret;
+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();
+  ret.Checkerboard() = rhs.Checkerboard();
+  accelerator_for(ss,rhs.size(),obj::Nsimd(),{
+    coalescedWrite(ret[ss],mod(rhs(ss),y));
+  });
+  return ret_i;
 }
 
-  template<class obj> Lattice<obj> div(const Lattice<obj> &rhs,Integer y){
-    Lattice<obj> ret(rhs._grid);
-    ret.checkerboard = rhs.checkerboard;
-    conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-      ret._odata[ss]=div(rhs._odata[ss],y);
-    }
-    return ret;
+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();
+  ret.Checkerboard() = rhs_i.Checkerboard();
+  accelerator_for(ss,rhs.size(),obj::Nsimd(),{
+    coalescedWrite(ret[ss],div(rhs(ss),y));
+  });
+  return ret_i;
 }
 
-  template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
-    Lattice<obj> ret(rhs._grid);
-    ret.checkerboard = rhs.checkerboard;
-    conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
-      ret._odata[ss]=Exponentiate(rhs._odata[ss],alpha, Nexp);
+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();
+  ret.Checkerboard() = rhs.Checkerboard();
+  accelerator_for(ss,rhs.size(),obj::Nsimd(),{
+    coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));
+  });
+  return ret_i;
 }
 
-    return ret;
-
-    
-    
-
-    
-  }
-
-
-
-}
+NAMESPACE_END(Grid);
 #endif

Grid/log/Log.cc

@@ -35,7 +35,7 @@ directory
 #include <cxxabi.h>
 #include <memory>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 std::string demangle(const char* name) {
     
@@ -59,6 +59,7 @@ void GridLogTimestamp(int on){
 }
 
 Colours GridLogColours(0);
+GridLogger GridLogMG     (1, "MG"    , GridLogColours, "NORMAL");
 GridLogger GridLogIRL    (1, "IRL"   , GridLogColours, "NORMAL");
 GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
@@ -76,7 +77,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
-  GridLogIntegrator.Active(0);
+  GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
 
   for (int i = 0; i < logstreams.size(); i++) {
@@ -85,8 +86,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
-    if (logstreams[i] == std::string("Performance"))
-      GridLogPerformance.Active(1);
+    if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
     if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
@@ -109,8 +109,9 @@ void Grid_quiesce_nodes(void) {
 }
 
 void Grid_unquiesce_nodes(void) {
-#ifdef GRID_COMMS_MPI
+#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
   std::cout.clear();
 #endif
 }
-}
+NAMESPACE_END(Grid);
+

Grid/log/Log.h

@@ -37,13 +37,12 @@
 #include <execinfo.h>
 #endif
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////////////////////////////////
 // Dress the output; use std::chrono for time stamping via the StopWatch class
 //////////////////////////////////////////////////////////////////////////////////////////////////
 
-
 class Colours{
 protected:
   bool is_active;
@@ -146,9 +145,11 @@ public:
       if ( log.timestamp ) {
 	log.StopWatch->Stop();
 	GridTime now = log.StopWatch->Elapsed();
+	
 	if ( log.timing_mode==1 ) log.StopWatch->Reset();
 	log.StopWatch->Start();
-	stream << log.evidence()<< std::setw(6)<<now << log.background() << " : " ;
+	stream << log.evidence()
+	       << now	       << log.background() << " : " ;
       }
       stream << log.colour();
       return stream;
@@ -167,6 +168,7 @@ public:
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
 
+extern GridLogger GridLogMG;
 extern GridLogger GridLogIRL;
 extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
@@ -211,6 +213,6 @@ std::fprintf (fp,"BT %d %lx\n",3, __builtin_return_address(3)); std::fflush(fp);
 
 #define BACKTRACE() BACKTRACEFP(stdout) 
 
+NAMESPACE_END(Grid);
 
-}
 #endif

Grid/parallelIO/BinaryIO.cc

@@ -0,0 +1,3 @@
+#include <Grid/GridCore.h>
+
+int Grid::BinaryIO::latticeWriteMaxRetry = -1;

Grid/parallelIO/BinaryIO.h

@@ -26,8 +26,7 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_BINARY_IO_H
-#define GRID_BINARY_IO_H
+#pragma once
 
 #if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT) 
 #define USE_MPI_IO
@@ -42,8 +41,7 @@
 #include <arpa/inet.h>
 #include <algorithm>
 
-namespace Grid { 
-
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////////////////////////
 // Byte reversal garbage
@@ -81,6 +79,7 @@ inline void removeWhitespace(std::string &key)
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 class BinaryIO {
  public:
+  static int latticeWriteMaxRetry;
 
   /////////////////////////////////////////////////////////////////////////////
   // more byte manipulation helpers
@@ -90,7 +89,7 @@ class BinaryIO {
   {
     typedef typename vobj::scalar_object sobj;
 
-    GridBase *grid = lat._grid;
+    GridBase *grid = lat.Grid();
     uint64_t lsites = grid->lSites();
 
     std::vector<sobj> scalardata(lsites); 
@@ -110,21 +109,20 @@ class BinaryIO {
       lsites = 1;
     }
 
-PARALLEL_REGION
+    thread_region
     {
       uint32_t nersc_csum_thr = 0;
 
-PARALLEL_FOR_LOOP_INTERN
-      for (uint64_t local_site = 0; local_site < lsites; local_site++)
+      thread_for_in_region( local_site, lsites, 
       {
         uint32_t *site_buf = (uint32_t *)&fbuf[local_site];
         for (uint64_t j = 0; j < size32; j++)
         {
           nersc_csum_thr = nersc_csum_thr + site_buf[j];
         }
-      }
+      });
 
-PARALLEL_CRITICAL
+      thread_critical
       {
         nersc_csum += nersc_csum_thr;
       }
@@ -133,28 +131,25 @@ PARALLEL_CRITICAL
 
   template<class fobj> static inline void ScidacChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &scidac_csuma,uint32_t &scidac_csumb)
   {
-    const uint64_t size32 = sizeof(fobj)/sizeof(uint32_t);
-
-
     int nd = grid->_ndimension;
 
     uint64_t lsites              =grid->lSites();
     if (fbuf.size()==1) {
       lsites=1;
     }
-    std::vector<int> local_vol   =grid->LocalDimensions();
-    std::vector<int> local_start =grid->LocalStarts();
-    std::vector<int> global_vol  =grid->FullDimensions();
+    Coordinate local_vol   =grid->LocalDimensions();
+    Coordinate local_start =grid->LocalStarts();
+    Coordinate global_vol  =grid->FullDimensions();
 
-PARALLEL_REGION
+    thread_region
     { 
-      std::vector<int> coor(nd);
+      Coordinate coor(nd);
       uint32_t scidac_csuma_thr=0;
       uint32_t scidac_csumb_thr=0;
       uint32_t site_crc=0;
 
-PARALLEL_FOR_LOOP_INTERN
-      for(uint64_t local_site=0;local_site<lsites;local_site++){
+      thread_for_in_region( local_site, lsites, 
+      {
 
 	uint32_t * site_buf = (uint32_t *)&fbuf[local_site];
 
@@ -181,9 +176,9 @@ PARALLEL_FOR_LOOP_INTERN
 	//	std::cout << "Site "<<local_site << std::hex<<site_buf[0] <<site_buf[1]<<std::dec <<std::endl;
 	scidac_csuma_thr ^= site_crc<<gsite29 | site_crc>>(32-gsite29);
 	scidac_csumb_thr ^= site_crc<<gsite31 | site_crc>>(32-gsite31);
-      }
+      });
 
-PARALLEL_CRITICAL
+      thread_critical
       {
 	scidac_csuma^= scidac_csuma_thr;
 	scidac_csumb^= scidac_csumb_thr;
@@ -201,23 +196,23 @@ PARALLEL_CRITICAL
   {
     uint32_t * f = (uint32_t *)file_object;
     uint64_t count = bytes/sizeof(uint32_t);
-    parallel_for(uint64_t i=0;i<count;i++){  
+    thread_for( i, count, {  
       f[i] = ntohl(f[i]);
-    }
+    });
   }
   // LE must Swap and switch to host
   static inline void le32toh_v(void *file_object,uint64_t bytes)
   {
     uint32_t *fp = (uint32_t *)file_object;
-    uint32_t f;
 
     uint64_t count = bytes/sizeof(uint32_t);
-    parallel_for(uint64_t i=0;i<count;i++){  
+    thread_for(i,count,{
+      uint32_t f;
       f = fp[i];
       // got network order and the network to host
       f = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
       fp[i] = ntohl(f);
-    }
+    });
   }
 
   // BE is same as network
@@ -225,19 +220,18 @@ PARALLEL_CRITICAL
   {
     uint64_t * f = (uint64_t *)file_object;
     uint64_t count = bytes/sizeof(uint64_t);
-    parallel_for(uint64_t i=0;i<count;i++){  
+    thread_for( i, count, {
       f[i] = Grid_ntohll(f[i]);
-    }
+    });
   }
   
   // LE must swap and switch;
   static inline void le64toh_v(void *file_object,uint64_t bytes)
   {
     uint64_t *fp = (uint64_t *)file_object;
-    uint64_t f,g;
-    
     uint64_t count = bytes/sizeof(uint64_t);
-    parallel_for(uint64_t i=0;i<count;i++){  
+    thread_for( i, count, {
+      uint64_t f,g;
       f = fp[i];
       // got network order and the network to host
       g = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
@@ -245,7 +239,7 @@ PARALLEL_CRITICAL
       f = f >> 32;
       g|= ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
       fp[i] = Grid_ntohll(g);
-    }
+    });
   }
   /////////////////////////////////////////////////////////////////////////////
   // Real action:
@@ -281,13 +275,13 @@ PARALLEL_CRITICAL
     int nrank                = grid->ProcessorCount();
     int myrank               = grid->ThisRank();
 
-    std::vector<int>  psizes = grid->ProcessorGrid(); 
-    std::vector<int>  pcoor  = grid->ThisProcessorCoor();
-    std::vector<int> gLattice= grid->GlobalDimensions();
-    std::vector<int> lLattice= grid->LocalDimensions();
+    Coordinate  psizes = grid->ProcessorGrid(); 
+    Coordinate  pcoor  = grid->ThisProcessorCoor();
+    Coordinate gLattice= grid->GlobalDimensions();
+    Coordinate lLattice= grid->LocalDimensions();
 
-    std::vector<int> lStart(ndim);
-    std::vector<int> gStart(ndim);
+    Coordinate lStart(ndim);
+    Coordinate gStart(ndim);
 
     // Flatten the file
     uint64_t lsites = grid->lSites();
@@ -348,7 +342,8 @@ PARALLEL_CRITICAL
     int ieee32    = (format == std::string("IEEE32"));
     int ieee64big = (format == std::string("IEEE64BIG"));
     int ieee64    = (format == std::string("IEEE64"));
-
+    assert(ieee64||ieee32|ieee64big||ieee32big);
+    assert((ieee64+ieee32+ieee64big+ieee32big)==1);
     //////////////////////////////////////////////////////////////////////////////
     // Do the I/O
     //////////////////////////////////////////////////////////////////////////////
@@ -370,7 +365,7 @@ PARALLEL_CRITICAL
 #endif
       } else {
 	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
+                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
         std::ifstream fin;
 	fin.open(file, std::ios::binary | std::ios::in);
         if (control & BINARYIO_MASTER_APPEND)
@@ -545,7 +540,7 @@ PARALLEL_CRITICAL
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
 
-    GridBase *grid = Umu._grid;
+    GridBase *grid = Umu.Grid();
     uint64_t lsites = grid->lSites();
 
     std::vector<sobj> scalardata(lsites); 
@@ -557,7 +552,7 @@ PARALLEL_CRITICAL
     GridStopWatch timer; 
     timer.Start();
 
-    parallel_for(uint64_t x=0;x<lsites;x++) munge(iodata[x], scalardata[x]);
+    thread_for(x,lsites, { munge(iodata[x], scalardata[x]); });
 
     vectorizeFromLexOrdArray(scalardata,Umu);    
     grid->Barrier();
@@ -581,8 +576,10 @@ PARALLEL_CRITICAL
   {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
-    GridBase *grid = Umu._grid;
-    uint64_t lsites = grid->lSites();
+    GridBase *grid = Umu.Grid();
+    uint64_t lsites = grid->lSites(), offsetCopy = offset;
+    int attemptsLeft = std::max(0, BinaryIO::latticeWriteMaxRetry);
+    bool checkWrite = (BinaryIO::latticeWriteMaxRetry >= 0);
 
     std::vector<sobj> scalardata(lsites); 
     std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
@@ -593,13 +590,40 @@ PARALLEL_CRITICAL
     GridStopWatch timer; timer.Start();
     unvectorizeToLexOrdArray(scalardata,Umu);    
 
-    parallel_for(uint64_t x=0;x<lsites;x++) munge(scalardata[x],iodata[x]);
+    thread_for(x, lsites, { munge(scalardata[x],iodata[x]); });
 
     grid->Barrier();
     timer.Stop();
-
+    while (attemptsLeft >= 0)
+    {
+      grid->Barrier();
       IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
 	             nersc_csum,scidac_csuma,scidac_csumb);
+      if (checkWrite)
+      {
+        std::vector<fobj> ckiodata(lsites);
+        uint32_t          cknersc_csum, ckscidac_csuma, ckscidac_csumb;
+        uint64_t          ckoffset = offsetCopy;
+
+        std::cout << GridLogMessage << "writeLatticeObject: read back object" << std::endl;
+        grid->Barrier();
+        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+	               cknersc_csum,ckscidac_csuma,ckscidac_csumb);
+        if ((cknersc_csum != nersc_csum) or (ckscidac_csuma != scidac_csuma) or (ckscidac_csumb != scidac_csumb))
+        {
+          std::cout << GridLogMessage << "writeLatticeObject: read test checksum failure, re-writing (" << attemptsLeft << " attempt(s) remaining)" << std::endl;
+          offset = offsetCopy;
+          thread_for(x,lsites, { munge(scalardata[x],iodata[x]); });
+        }
+        else
+        {
+          std::cout << GridLogMessage << "writeLatticeObject: read test checksum correct" << std::endl;
+          break;
+        }
+      }
+      attemptsLeft--;
+    }
+    
 
     std::cout<<GridLogMessage<<"writeLatticeObject: unvectorize overhead "<<timer.Elapsed()  <<std::endl;
   }
@@ -607,8 +631,8 @@ PARALLEL_CRITICAL
   /////////////////////////////////////////////////////////////////////////////
   // Read a RNG;  use IOobject and lexico map to an array of state 
   //////////////////////////////////////////////////////////////////////////////////////
-  static inline void readRNG(GridSerialRNG &serial,
-			     GridParallelRNG &parallel,
+  static inline void readRNG(GridSerialRNG &serial_rng,
+			     GridParallelRNG &parallel_rng,
 			     std::string file,
 			     uint64_t offset,
 			     uint32_t &nersc_csum,
@@ -622,7 +646,7 @@ PARALLEL_CRITICAL
 
     std::string format = "IEEE32BIG";
 
-    GridBase *grid = parallel._grid;
+    GridBase *grid = parallel_rng.Grid();
     uint64_t gsites = grid->gSites();
     uint64_t lsites = grid->lSites();
 
@@ -639,11 +663,11 @@ PARALLEL_CRITICAL
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     timer.Start();
-    parallel_for(uint64_t lidx=0;lidx<lsites;lidx++){
+    thread_for(lidx,lsites,{
       std::vector<RngStateType> tmp(RngStateCount);
       std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
-      parallel.SetState(tmp,lidx);
-    }
+      parallel_rng.SetState(tmp,lidx);
+      });
     timer.Stop();
 
     iodata.resize(1);
@@ -653,7 +677,7 @@ PARALLEL_CRITICAL
     {
       std::vector<RngStateType> tmp(RngStateCount);
       std::copy(iodata[0].begin(),iodata[0].end(),tmp.begin());
-      serial.SetState(tmp,0);
+      serial_rng.SetState(tmp,0);
     }
 
     nersc_csum   = nersc_csum   + nersc_csum_tmp;
@@ -669,8 +693,8 @@ PARALLEL_CRITICAL
   /////////////////////////////////////////////////////////////////////////////
   // Write a RNG; lexico map to an array of state and use IOobject
   //////////////////////////////////////////////////////////////////////////////////////
-  static inline void writeRNG(GridSerialRNG &serial,
-			      GridParallelRNG &parallel,
+  static inline void writeRNG(GridSerialRNG &serial_rng,
+			      GridParallelRNG &parallel_rng,
 			      std::string file,
 			      uint64_t offset,
 			      uint32_t &nersc_csum,
@@ -682,7 +706,7 @@ PARALLEL_CRITICAL
     const int RngStateCount = GridSerialRNG::RngStateCount;
     typedef std::array<RngStateType,RngStateCount> RNGstate;
 
-    GridBase *grid = parallel._grid;
+    GridBase *grid = parallel_rng.Grid();
     uint64_t gsites = grid->gSites();
     uint64_t lsites = grid->lSites();
 
@@ -697,11 +721,11 @@ PARALLEL_CRITICAL
 
     timer.Start();
     std::vector<RNGstate> iodata(lsites);
-    parallel_for(uint64_t lidx=0;lidx<lsites;lidx++){
+    thread_for(lidx,lsites,{
       std::vector<RngStateType> tmp(RngStateCount);
-      parallel.GetState(tmp,lidx);
+      parallel_rng.GetState(tmp,lidx);
       std::copy(tmp.begin(),tmp.end(),iodata[lidx].begin());
-    }
+    });
     timer.Stop();
 
     IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
@@ -709,7 +733,7 @@ PARALLEL_CRITICAL
     iodata.resize(1);
     {
       std::vector<RngStateType> tmp(RngStateCount);
-      serial.GetState(tmp,0);
+      serial_rng.GetState(tmp,0);
       std::copy(tmp.begin(),tmp.end(),iodata[0].begin());
     }
     IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_MASTER_APPEND,
@@ -725,5 +749,5 @@ PARALLEL_CRITICAL
     std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
   }
 };
-}
-#endif
+
+NAMESPACE_END(Grid);

Grid/parallelIO/IldgIO.h

@@ -24,8 +24,7 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_ILDG_IO_H
-#define GRID_ILDG_IO_H
+#pragma once
 
 #ifdef HAVE_LIME
 #include <algorithm>
@@ -43,8 +42,13 @@ extern "C" {
 #include "lime.h"
 }
 
-namespace Grid {
-namespace QCD {
+NAMESPACE_BEGIN(Grid);
+
+#define GRID_FIELD_NORM "FieldNormMetaData"
+#define GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) \
+0.5*fabs(FieldNormMetaData_.norm2 - n2ck)/(FieldNormMetaData_.norm2 + n2ck)
+#define GRID_FIELD_NORM_CHECK(FieldNormMetaData_, n2ck) \
+assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
 
   /////////////////////////////////
   // Encode word types as strings
@@ -134,7 +138,7 @@ namespace QCD {
    /////////////////////////////////////
    // Scidac Private File structure
    /////////////////////////////////////
-   _scidacFile              = scidacFile(field._grid);
+   _scidacFile              = scidacFile(field.Grid());
 
    /////////////////////////////////////
    // Scidac Private Record structure
@@ -205,6 +209,7 @@ class GridLimeReader : public BinaryIO {
   {
     typedef typename vobj::scalar_object sobj;
     scidacChecksum scidacChecksum_;
+    FieldNormMetaData  FieldNormMetaData_;
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
 
     std::string format = getFormatString<vobj>();
@@ -220,10 +225,10 @@ class GridLimeReader : public BinaryIO {
 
 	//	std::cout << GridLogMessage<< " readLimeLatticeBinaryObject matches ! " <<std::endl;
 
-	uint64_t PayloadSize = sizeof(sobj) * field._grid->_gsites;
+	uint64_t PayloadSize = sizeof(sobj) * field.Grid()->_gsites;
 
 	//	std::cout << "R sizeof(sobj)= " <<sizeof(sobj)<<std::endl;
-	//	std::cout << "R Gsites " <<field._grid->_gsites<<std::endl;
+	//	std::cout << "R Gsites " <<field.Grid()->_gsites<<std::endl;
 	//	std::cout << "R Payload expected " <<PayloadSize<<std::endl;
 	//	std::cout << "R file size " <<file_bytes <<std::endl;
 
@@ -233,20 +238,52 @@ class GridLimeReader : public BinaryIO {
 	//	std::cout << " ReadLatticeObject from offset "<<offset << std::endl;
 	BinarySimpleMunger<sobj,sobj> munge;
 	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
-
+	std::cout << GridLogMessage << "SciDAC checksum A " << std::hex << scidac_csuma << std::dec << std::endl;
+	std::cout << GridLogMessage << "SciDAC checksum B " << std::hex << scidac_csumb << std::dec << std::endl;
 	/////////////////////////////////////////////
 	// Insist checksum is next record
 	/////////////////////////////////////////////
-	readLimeObject(scidacChecksum_,std::string("scidacChecksum"),std::string(SCIDAC_CHECKSUM));
-
+	readScidacChecksum(scidacChecksum_,FieldNormMetaData_);
 	/////////////////////////////////////////////
 	// Verify checksums
 	/////////////////////////////////////////////
+	if(FieldNormMetaData_.norm2 != 0.0){ 
+	  RealD n2ck = norm2(field);
+	  std::cout << GridLogMessage << "Field norm: metadata= " << FieldNormMetaData_.norm2 
+              << " / field= " << n2ck << " / rdiff= " << GRID_FIELD_NORM_CALC(FieldNormMetaData_,n2ck) << std::endl;
+	  GRID_FIELD_NORM_CHECK(FieldNormMetaData_,n2ck);
+	}
 	assert(scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb)==1);
+
+	// find out if next field is a GridFieldNorm
 	return;
       }
     }
   }
+  void readScidacChecksum(scidacChecksum     &scidacChecksum_,
+			  FieldNormMetaData  &FieldNormMetaData_)
+  {
+    FieldNormMetaData_.norm2 =0.0;
+    std::string scidac_str(SCIDAC_CHECKSUM);
+    std::string field_norm_str(GRID_FIELD_NORM);
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+      uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
+      std::vector<char> xmlc(nbytes+1,'\0');
+      limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
+      std::string xmlstring = std::string(&xmlc[0]);
+      XmlReader RD(xmlstring, true, "");
+      if ( !strncmp(limeReaderType(LimeR), field_norm_str.c_str(),strlen(field_norm_str.c_str()) )  ) {
+	//	std::cout << "FieldNormMetaData "<<xmlstring<<std::endl;
+	read(RD,field_norm_str,FieldNormMetaData_);
+      }
+      if ( !strncmp(limeReaderType(LimeR), scidac_str.c_str(),strlen(scidac_str.c_str()) )  ) {
+	//	std::cout << SCIDAC_CHECKSUM << " " <<xmlstring<<std::endl;
+	read(RD,std::string("scidacChecksum"),scidacChecksum_);
+	return;
+      }      
+    }
+    assert(0);
+  }
   ////////////////////////////////////////////
   // Read a generic serialisable object
   ////////////////////////////////////////////
@@ -367,7 +404,7 @@ class GridLimeWriter : public BinaryIO
   ////////////////////////////////////////////////////
   // Write a generic lattice field and csum
   // This routine is Collectively called by all nodes
-  // in communicator used by the field._grid
+  // in communicator used by the field.Grid()
   ////////////////////////////////////////////////////
   template<class vobj>
   void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
@@ -386,8 +423,10 @@ class GridLimeWriter : public BinaryIO
     //  v) Continue writing scidac record.
     ////////////////////////////////////////////////////////////////////
     
-    GridBase *grid = field._grid;
-    assert(boss_node == field._grid->IsBoss() );
+    GridBase *grid = field.Grid();
+    assert(boss_node == field.Grid()->IsBoss() );
+
+    FieldNormMetaData FNMD; FNMD.norm2 = norm2(field);
 
     ////////////////////////////////////////////
     // Create record header
@@ -402,7 +441,7 @@ class GridLimeWriter : public BinaryIO
     }
     
     //    std::cout << "W sizeof(sobj)"      <<sizeof(sobj)<<std::endl;
-    //    std::cout << "W Gsites "           <<field._grid->_gsites<<std::endl;
+    //    std::cout << "W Gsites "           <<field.Grid()->_gsites<<std::endl;
     //    std::cout << "W Payload expected " <<PayloadSize<<std::endl;
 
     ////////////////////////////////////////////////
@@ -447,6 +486,7 @@ class GridLimeWriter : public BinaryIO
     checksum.suma= streama.str();
     checksum.sumb= streamb.str();
     if ( boss_node ) { 
+      writeLimeObject(0,0,FNMD,std::string(GRID_FIELD_NORM),std::string(GRID_FIELD_NORM));
       writeLimeObject(0,1,checksum,std::string("scidacChecksum"),std::string(SCIDAC_CHECKSUM));
     }
   }
@@ -473,7 +513,7 @@ class ScidacWriter : public GridLimeWriter {
   void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
                               const unsigned int recordScientificPrec = 0) 
   {
-    GridBase * grid = field._grid;
+    GridBase * grid = field.Grid();
 
     ////////////////////////////////////////
     // fill the Grid header
@@ -515,7 +555,7 @@ class ScidacReader : public GridLimeReader {
   void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord) 
   {
     typedef typename vobj::scalar_object sobj;
-    GridBase * grid = field._grid;
+    GridBase * grid = field.Grid();
 
     ////////////////////////////////////////
     // fill the Grid header
@@ -582,7 +622,7 @@ class IldgWriter : public ScidacWriter {
   template <class vsimd>
   void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
   {
-    GridBase * grid = Umu._grid;
+    GridBase * grid = Umu.Grid();
     typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
     typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
@@ -624,6 +664,12 @@ class IldgWriter : public ScidacWriter {
     assert(header.nd==4);
     assert(header.nd==header.dimension.size());
 
+    //////////////////////////////////////////////////////////////////////////////
+    // Field norm tests
+    //////////////////////////////////////////////////////////////////////////////
+    FieldNormMetaData FieldNormMetaData_;
+    FieldNormMetaData_.norm2 = norm2(Umu);
+
     //////////////////////////////////////////////////////////////////////////////
     // Fill the USQCD info field
     //////////////////////////////////////////////////////////////////////////////
@@ -632,11 +678,12 @@ class IldgWriter : public ScidacWriter {
     info.plaq   = header.plaquette;
     info.linktr = header.link_trace;
 
-    std::cout << GridLogMessage << " Writing config; IldgIO "<<std::endl;
+    //    std::cout << GridLogMessage << " Writing config; IldgIO n2 "<< FieldNormMetaData_.norm2<<std::endl;
     //////////////////////////////////////////////
     // Fill the Lime file record by record
     //////////////////////////////////////////////
     writeLimeObject(1,0,header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
+    writeLimeObject(0,0,FieldNormMetaData_,FieldNormMetaData_.SerialisableClassName(),std::string(GRID_FIELD_NORM));
     writeLimeObject(0,0,_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
     writeLimeObject(0,1,info,info.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
     writeLimeObject(1,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
@@ -668,9 +715,9 @@ class IldgReader : public GridLimeReader {
     typedef LorentzColourMatrixF fobj;
     typedef LorentzColourMatrixD dobj;
 
-    GridBase *grid = Umu._grid;
+    GridBase *grid = Umu.Grid();
 
-    std::vector<int> dims = Umu._grid->FullDimensions();
+    Coordinate dims = Umu.Grid()->FullDimensions();
 
     assert(dims.size()==4);
 
@@ -679,6 +726,7 @@ class IldgReader : public GridLimeReader {
     std::string    ildgLFN_       ;
     scidacChecksum scidacChecksum_; 
     usqcdInfo      usqcdInfo_     ;
+    FieldNormMetaData FieldNormMetaData_;
 
     // track what we read from file
     int found_ildgFormat    =0;
@@ -687,7 +735,7 @@ class IldgReader : public GridLimeReader {
     int found_usqcdInfo     =0;
     int found_ildgBinary =0;
     int found_FieldMetaData =0;
-
+    int found_FieldNormMetaData =0;
     uint32_t nersc_csum;
     uint32_t scidac_csuma;
     uint32_t scidac_csumb;
@@ -774,11 +822,17 @@ class IldgReader : public GridLimeReader {
 	  found_scidacChecksum = 1;
 	}
 
+	if ( !strncmp(limeReaderType(LimeR), GRID_FIELD_NORM,strlen(GRID_FIELD_NORM)) ) { 
+	  XmlReader RD(xmlstring, true, "");
+	  read(RD,GRID_FIELD_NORM,FieldNormMetaData_);
+	  found_FieldNormMetaData = 1;
+	}
+
       } else {  
 	/////////////////////////////////
 	// Binary data
 	/////////////////////////////////
-	std::cout << GridLogMessage << "ILDG Binary record found : "  ILDG_BINARY_DATA << std::endl;
+	//	std::cout << GridLogMessage << "ILDG Binary record found : "  ILDG_BINARY_DATA << std::endl;
 	uint64_t offset= ftello(File);
 	if ( format == std::string("IEEE64BIG") ) {
 	  GaugeSimpleMunger<dobj, sobj> munge;
@@ -797,6 +851,7 @@ class IldgReader : public GridLimeReader {
     // Minimally must find binary segment and checksum
     // Since this is an ILDG reader require ILDG format
     //////////////////////////////////////////////////////
+    assert(found_ildgLFN);
     assert(found_ildgBinary);
     assert(found_ildgFormat);
     assert(found_scidacChecksum);
@@ -845,6 +900,13 @@ class IldgReader : public GridLimeReader {
     ////////////////////////////////////////////////////////////
     // Really really want to mandate a scidac checksum
     ////////////////////////////////////////////////////////////
+    if ( found_FieldNormMetaData ) { 
+      RealD nn = norm2(Umu);
+      GRID_FIELD_NORM_CHECK(FieldNormMetaData_,nn);
+      std::cout << GridLogMessage<<"FieldNormMetaData matches " << std::endl;
+    }  else { 
+      std::cout << GridLogWarning<<"FieldNormMetaData not found. " << std::endl;
+    }
     if ( found_scidacChecksum ) {
       FieldMetaData_.scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
       FieldMetaData_.scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
@@ -867,9 +929,9 @@ class IldgReader : public GridLimeReader {
   }
  };
 
-}}
+NAMESPACE_END(Grid);
+
 
 //HAVE_LIME
 #endif
 
-#endif

Grid/parallelIO/IldgIOtypes.h

@@ -32,7 +32,7 @@ extern "C" { // for linkage
 #include "lime.h"
 }
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////////////////////////
 // Data representation of records that enter ILDG and SciDac formats
@@ -91,7 +91,7 @@ struct scidacFile : Serializable {
     return dimensions;
   }
 
-  void setDimensions(std::vector<int> dimensions) { 
+  void setDimensions(Coordinate dimensions) { 
     char delimiter = ' ';
     std::stringstream stream;
     for(int i=0;i<dimensions.size();i++){ 
@@ -232,6 +232,6 @@ struct usqcdPropInfo : Serializable {
 };
 #endif
 
-}
+NAMESPACE_END(Grid);
 #endif
 #endif

Grid/parallelIO/MetaData.h

@@ -36,7 +36,7 @@
 #include <sys/utsname.h>
 #include <pwd.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////
 // Precision mapping
@@ -52,10 +52,15 @@ namespace Grid {
     format = std::string("IEEE64BIG");
   }
   return format;
-  }
+};
+
   ////////////////////////////////////////////////////////////////////////////////
   // header specification/interpretation
   ////////////////////////////////////////////////////////////////////////////////
+    class FieldNormMetaData : Serializable {
+    public:
+      GRID_SERIALIZABLE_CLASS_MEMBERS(FieldNormMetaData, double, norm2);
+    };
     class FieldMetaData : Serializable {
     public:
 
@@ -91,10 +96,9 @@ namespace Grid {
       {}
   };
 
-  namespace QCD {
-
-    using namespace Grid;
-
+// PB disable using namespace - this is a header and forces namesapce visibility for all 
+// including files
+//using namespace Grid;
 
 //////////////////////////////////////////////////////////////////////
 // Bit and Physical Checksumming and QA of data
@@ -165,7 +169,7 @@ namespace Grid {
 
 template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMetaData &header)
 {
-  GridBase *grid = field._grid;
+  GridBase *grid = field.Grid();
   std::string format = getFormatString<vobj>();
   header.floating_point = format;
   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
@@ -175,19 +179,19 @@ template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMet
 inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
 {
   // How to convert data precision etc...
-   header.link_trace=Grid::QCD::WilsonLoops<PeriodicGimplF>::linkTrace(data);
-   header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
+  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=Grid::QCD::WilsonLoops<PeriodicGimplD>::linkTrace(data);
-   header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
+  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;
+  GridBase *grid = field.Grid();
   std::string format = getFormatString<vLorentzColourMatrixF>();
   header.floating_point = format;
   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
@@ -197,7 +201,7 @@ template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMet
 }
 template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
 {
-   GridBase *grid = field._grid;
+  GridBase *grid = field.Grid();
   std::string format = getFormatString<vLorentzColourMatrixD>();
   header.floating_point = format;
   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
@@ -321,7 +325,6 @@ struct BinarySimpleMunger {
     }
   }
 };
-  }
 
+NAMESPACE_END(Grid);
 
-}

Grid/parallelIO/NerscIO.h

@@ -30,8 +30,7 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
-namespace Grid {
-  namespace QCD {
+NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
 
@@ -57,7 +56,6 @@ namespace Grid {
   // for the header-reader
   static inline int readHeader(std::string file,GridBase *grid,  FieldMetaData &field)
   {
-      uint64_t offset=0;
     std::map<std::string,std::string> header;
     std::string line;
 
@@ -138,8 +136,8 @@ namespace Grid {
   {
     typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
-      GridBase *grid = Umu._grid;
-      uint64_t offset = readHeader(file,Umu._grid,header);
+    GridBase *grid = Umu.Grid();
+    uint64_t offset = readHeader(file,Umu.Grid(),header);
 
     FieldMetaData clone(header);
 
@@ -190,8 +188,6 @@ namespace Grid {
 
     if ( fabs(clone.plaquette -header.plaquette ) >=  1.0e-5 ) { 
       std::cout << " Plaquette mismatch "<<std::endl;
-	std::cout << Umu[0]<<std::endl;
-	std::cout << Umu[1]<<std::endl;
     }
     if ( nersc_csum != header.checksum ) { 
       std::cerr << " checksum mismatch " << std::endl;
@@ -229,7 +225,7 @@ namespace Grid {
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
   
-	GridBase *grid = Umu._grid;
+    GridBase *grid = Umu.Grid();
 
     GridMetaData(grid,header);
     assert(header.nd==4);
@@ -274,7 +270,7 @@ namespace Grid {
     header.ensemble_id     = "UKQCD";
     header.ensemble_label  = "DWF";
 
-	GridBase *grid = parallel._grid;
+    GridBase *grid = parallel.Grid();
 
     GridMetaData(grid,header);
     assert(header.nd==4);
@@ -321,7 +317,7 @@ namespace Grid {
   {
     typedef typename GridParallelRNG::RngStateType RngStateType;
 
-	GridBase *grid = parallel._grid;
+    GridBase *grid = parallel.Grid();
 
 	uint64_t offset = readHeader(file,grid,header);
 
@@ -356,8 +352,8 @@ namespace Grid {
 
     std::cout<<GridLogMessage <<"Read NERSC RNG file "<<file<< " format "<< data_type <<std::endl;
   }
-
 };
 
-  }}
+NAMESPACE_END(QCD);
+
 #endif

Grid/perfmon/PerfCount.cc

@@ -29,7 +29,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/perfmon/PerfCount.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
 #define RawConfig(A,B) (A<<8|B)
@@ -72,4 +72,5 @@ const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::Performan
   //  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_STALLED_CYCLES_FRONTEND, "STALL_CYCLES" },
 #endif
 };
-}
+NAMESPACE_END(Grid);
+

Grid/perfmon/PerfCount.h

@@ -47,7 +47,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <x86intrin.h>
 #endif
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 #ifdef __linux__
 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
@@ -89,6 +89,9 @@ inline uint64_t cyclecount(void){
   return tmp;
 }
 #elif defined __x86_64__
+#ifdef GRID_NVCC
+accelerator_inline uint64_t __rdtsc(void) {  return 0; }
+#endif
 inline uint64_t cyclecount(void){ 
   return __rdtsc();
   //  unsigned int dummy;
@@ -212,7 +215,7 @@ public:
       ::ioctl(cyclefd, PERF_EVENT_IOC_DISABLE, 0);
       ign=::read(fd, &count, sizeof(long long));
       ign+=::read(cyclefd, &cycles, sizeof(long long));
-      assert(ign=2*sizeof(long long));
+      assert(ign==2*sizeof(long long));
     }
     elapsed = cyclecount() - begin;
 #else
@@ -241,5 +244,6 @@ public:
 
 };
 
-}
+NAMESPACE_END(Grid);
+
 #endif

Grid/perfmon/Stat.cc

@@ -2,7 +2,7 @@
 #include <Grid/perfmon/PerfCount.h>
 #include <Grid/perfmon/Stat.h>
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
 
 bool PmuStat::pmu_initialized=false;
 
@@ -242,4 +242,5 @@ void PmuStat::KNLreadctrs(ctrs &c)
 }
 
 #endif
-}
+NAMESPACE_END(Grid); 
+

Grid/perfmon/Stat.h

@@ -5,7 +5,7 @@
 #define _KNIGHTS_LANDING_ROOTONLY
 #endif
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
 
 ///////////////////////////////////////////////////////////////////////////////
 // Extra KNL counters from MCDRAM
@@ -98,7 +98,8 @@ public:
     
 };
 
-}
+NAMESPACE_END(Grid); 
+
 #endif
 
 

Grid/perfmon/Timer.h

@@ -33,11 +33,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <ctime>
 #include <chrono>
 
-namespace Grid {
-
+NAMESPACE_BEGIN(Grid)
 
 // Dress the output; use std::chrono
-
 // C++11 time facilities better?
 inline double usecond(void) {
   struct timeval tv;
@@ -49,20 +47,38 @@ inline double usecond(void) {
 
 typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
-typedef  std::chrono::milliseconds          GridMillisecs;
-typedef  std::chrono::microseconds          GridTime;
-typedef  std::chrono::microseconds          GridUsecs;
 
-inline std::ostream& operator<< (std::ostream & stream, const std::chrono::milliseconds & time)
+typedef  std::chrono::seconds               GridSecs;
+typedef  std::chrono::milliseconds          GridMillisecs;
+typedef  std::chrono::microseconds          GridUsecs;
+typedef  std::chrono::microseconds          GridTime;
+
+inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
-  stream << time.count()<<" ms";
+  stream << time.count()<<" s";
   return stream;
 }
-inline std::ostream& operator<< (std::ostream & stream, const std::chrono::microseconds & time)
+inline std::ostream& operator<< (std::ostream & stream, const GridMillisecs & now)
 {
-  stream << time.count()<<" usec";
+  GridSecs second(1);
+  auto     secs       = now/second ; 
+  auto     subseconds = now%second ;
+  auto     fill       = stream.fill();
+  stream << secs<<"."<<std::setw(3)<<std::setfill('0')<<subseconds.count()<<" s";
+  stream.fill(fill);
   return stream;
 }
+inline std::ostream& operator<< (std::ostream & stream, const GridUsecs & now)
+{
+  GridSecs second(1);
+  auto     seconds    = now/second ; 
+  auto     subseconds = now%second ;
+  auto     fill       = stream.fill();
+  stream << seconds<<"."<<std::setw(6)<<std::setfill('0')<<subseconds.count()<<" s";
+  stream.fill(fill);
+  return stream;
+}
+ 
 
 class GridStopWatch {
 private:
@@ -107,5 +123,6 @@ public:
   }
 };
 
-}
+NAMESPACE_END(Grid)
+
 #endif

Grid/pugixml/pugixml.cc

@@ -14,7 +14,12 @@
 #ifndef SOURCE_PUGIXML_CPP
 #define SOURCE_PUGIXML_CPP
 
-#include <Grid/pugixml/pugixml.h>
+#ifdef __NVCC__
+#pragma push
+#pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
+#endif
+
+#include "pugixml.h"
 
 #include <stdlib.h>
 #include <stdio.h>
@@ -202,7 +207,7 @@ PUGI__NS_BEGIN
 	// Without a template<> we'll get multiple definitions of the same static
 	template <typename T> allocation_function xml_memory_management_function_storage<T>::allocate = default_allocate;
 	template <typename T> deallocation_function xml_memory_management_function_storage<T>::deallocate = default_deallocate;
-
+        template struct xml_memory_management_function_storage<int>;
 	typedef xml_memory_management_function_storage<int> xml_memory;
 PUGI__NS_END
 
@@ -12768,6 +12773,10 @@ namespace pugi
 #undef PUGI__THROW_ERROR
 #undef PUGI__CHECK_ERROR
 
+#ifdef GRID_NVCC
+#pragma pop
+#endif
+
 #endif
 
 /**

Grid/qcd/QCD.h

@@ -29,55 +29,53 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_QCD_BASE_H
-#define GRID_QCD_BASE_H
-namespace Grid{
-namespace QCD {
+#pragma once
 
-    static const int Xdir = 0;
-    static const int Ydir = 1;
-    static const int Zdir = 2;
-    static const int Tdir = 3;
+NAMESPACE_BEGIN(Grid);
 
+static constexpr int Xdir = 0;
+static constexpr int Ydir = 1;
+static constexpr int Zdir = 2;
+static constexpr int Tdir = 3;
 
-    static const int Xp = 0;
-    static const int Yp = 1;
-    static const int Zp = 2;
-    static const int Tp = 3;
-    static const int Xm = 4;
-    static const int Ym = 5;
-    static const int Zm = 6;
-    static const int Tm = 7;
+static constexpr int Xp = 0;
+static constexpr int Yp = 1;
+static constexpr int Zp = 2;
+static constexpr int Tp = 3;
+static constexpr int Xm = 4;
+static constexpr int Ym = 5;
+static constexpr int Zm = 6;
+static constexpr int Tm = 7;
 
-    static const int Nc=3;
-    static const int Ns=4;
-    static const int Nd=4;
-    static const int Nhs=2; // half spinor
-    static const int Nds=8; // double stored gauge field
-    static const int Ngp=2; // gparity index range
+static constexpr int Nc=3;
+static constexpr int Ns=4;
+static constexpr int Nd=4;
+static constexpr int Nhs=2; // half spinor
+static constexpr int Nds=8; // double stored gauge field
+static constexpr int Ngp=2; // gparity index range
 
 //////////////////////////////////////////////////////////////////////////////
 // QCD iMatrix types
 // Index conventions:                            Lorentz x Spin x Colour
-    // note: static const int or constexpr will work for type deductions
+// note: static constexpr int or constexpr will work for type deductions
 //       with the intel compiler (up to version 17)
 //////////////////////////////////////////////////////////////////////////////
-    #define ColourIndex  2
-    #define SpinIndex    1
-    #define LorentzIndex 0
+#define ColourIndex  (2)
+#define SpinIndex    (1)
+#define LorentzIndex (0)
 
 // Also should make these a named enum type
-    static const int DaggerNo=0;
-    static const int DaggerYes=1;
-    static const int InverseNo=0;
-    static const int InverseYes=1;
+static constexpr int DaggerNo=0;
+static constexpr int DaggerYes=1;
+static constexpr int InverseNo=0;
+static constexpr int InverseYes=1;
 
 // Useful traits is this a spin index
 //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
 
 const int SpinorIndex = 2;
 template<typename T> struct isSpinor {
-      static const bool value = (SpinorIndex==T::TensorLevel);
+  static constexpr bool value = (SpinorIndex==T::TensorLevel);
 };
 template <typename T> using IfSpinor    = Invoke<std::enable_if< isSpinor<T>::value,int> > ;
 template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ;
@@ -90,6 +88,7 @@ namespace QCD {
 // That probably makes for GridRedBlack4dCartesian grid.
 
 // s,sp,c,spc,lc
+
 template<typename vtype> using iSinglet                   = iScalar<iScalar<iScalar<vtype> > >;
 template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iScalar<vtype>, Ns> >;
 template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
@@ -101,6 +100,8 @@ namespace QCD {
 template<typename vtype> using iSpinColourVector          = iScalar<iVector<iVector<vtype, Nc>, Ns> >;
 template<typename vtype> using iHalfSpinVector            = iScalar<iVector<iScalar<vtype>, Nhs> >;
 template<typename vtype> using iHalfSpinColourVector      = iScalar<iVector<iVector<vtype, Nc>, Nhs> >;
+    template<typename vtype> using iSpinColourSpinColourMatrix  = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
+
 
 template<typename vtype> using iGparitySpinColourVector       = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
 template<typename vtype> using iGparityHalfSpinColourVector   = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
@@ -132,6 +133,24 @@ namespace QCD {
 typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF;
 typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD;
 
+    // SpinColourSpinColour matrix
+    typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
+    typedef iSpinColourSpinColourMatrix<ComplexF >    SpinColourSpinColourMatrixF;
+    typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
+
+    typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
+    typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
+    typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+
+    // SpinColourSpinColour matrix
+    typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
+    typedef iSpinColourSpinColourMatrix<ComplexF >    SpinColourSpinColourMatrixF;
+    typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
+
+    typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
+    typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
+    typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+
 // LorentzColour
 typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix;
 typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
@@ -229,6 +248,9 @@ namespace QCD {
 typedef Lattice<vSpinColourMatrixF>     LatticeSpinColourMatrixF;
 typedef Lattice<vSpinColourMatrixD>     LatticeSpinColourMatrixD;
 
+typedef Lattice<vSpinColourSpinColourMatrix>      LatticeSpinColourSpinColourMatrix;
+typedef Lattice<vSpinColourSpinColourMatrixF>     LatticeSpinColourSpinColourMatrixF;
+typedef Lattice<vSpinColourSpinColourMatrixD>     LatticeSpinColourSpinColourMatrixD;
 
 typedef Lattice<vLorentzColourMatrix>  LatticeLorentzColourMatrix;
 typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
@@ -358,35 +380,35 @@ namespace QCD {
 //////////////////////////////////////////////
 template<class vobj> 
 void pokeColour(Lattice<vobj> &lhs,
-              const Lattice<decltype(peekIndex<ColourIndex>(lhs._odata[0],0))> & rhs,
+		const Lattice<decltype(peekIndex<ColourIndex>(vobj(),0))> & rhs,
 		int i)
 {
   PokeIndex<ColourIndex>(lhs,rhs,i);
 }
 template<class vobj> 
 void pokeColour(Lattice<vobj> &lhs,
-              const Lattice<decltype(peekIndex<ColourIndex>(lhs._odata[0],0,0))> & rhs,
+		const Lattice<decltype(peekIndex<ColourIndex>(vobj(),0,0))> & rhs,
 		int i,int j)
 {
   PokeIndex<ColourIndex>(lhs,rhs,i,j);
 }
 template<class vobj> 
 void pokeSpin(Lattice<vobj> &lhs,
-              const Lattice<decltype(peekIndex<SpinIndex>(lhs._odata[0],0))> & rhs,
+              const Lattice<decltype(peekIndex<SpinIndex>(vobj(),0))> & rhs,
               int i)
 {
   PokeIndex<SpinIndex>(lhs,rhs,i);
 }
 template<class vobj> 
 void pokeSpin(Lattice<vobj> &lhs,
-              const Lattice<decltype(peekIndex<SpinIndex>(lhs._odata[0],0,0))> & rhs,
+              const Lattice<decltype(peekIndex<SpinIndex>(vobj(),0,0))> & rhs,
               int i,int j)
 {
   PokeIndex<SpinIndex>(lhs,rhs,i,j);
 }
 template<class vobj> 
 void pokeLorentz(Lattice<vobj> &lhs,
-              const Lattice<decltype(peekIndex<LorentzIndex>(lhs._odata[0],0))> & rhs,
+		 const Lattice<decltype(peekIndex<LorentzIndex>(vobj(),0))> & rhs,
 		 int i)
 {
   PokeIndex<LorentzIndex>(lhs,rhs,i);
@@ -475,12 +497,12 @@ namespace QCD {
 // Trace lattice and non-lattice
 //////////////////////////////////////////
 template<int Index,class vobj>
-    inline auto traceSpin(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(lhs._odata[0]))>
+inline auto traceSpin(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(vobj()))>
 {
   return traceIndex<SpinIndex>(lhs);
 }
 template<int Index,class vobj>
-    inline auto traceColour(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(lhs._odata[0]))>
+inline auto traceColour(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(vobj()))>
 {
   return traceIndex<ColourIndex>(lhs);
 }
@@ -503,9 +525,5 @@ namespace QCD {
 		       Axial,   1,
 		       Tadpole, 2);
 
-}   //namespace QCD
-} // Grid
+NAMESPACE_END(Grid);
 
-
-
-#endif

Grid/qcd/action/Action.h

@@ -37,14 +37,18 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 // Abstract base interface
 ////////////////////////////////////////////
 #include <Grid/qcd/action/ActionCore.h>
+NAMESPACE_CHECK(ActionCore);
 ////////////////////////////////////////////////////////////////////////
 // Fermion actions; prevent coupling fermion.cc files to other headers
 ////////////////////////////////////////////////////////////////////////
 #include <Grid/qcd/action/fermion/FermionCore.h>
+NAMESPACE_CHECK(FermionCore);
 #include <Grid/qcd/action/fermion/Fermion.h>
+NAMESPACE_CHECK(Fermion);
 ////////////////////////////////////////
 // Pseudo fermion combinations for HMC
 ////////////////////////////////////////
 #include <Grid/qcd/action/pseudofermion/PseudoFermion.h>
+NAMESPACE_CHECK(PseudoFermion);
 
 #endif

Grid/qcd/action/ActionBase.h

@@ -32,8 +32,7 @@ directory
 #ifndef ACTION_BASE_H
 #define ACTION_BASE_H
 
-namespace Grid {
-namespace QCD {
+NAMESPACE_BEGIN(Grid);
 
 template <class GaugeField >
 class Action 
@@ -50,7 +49,6 @@ class Action
   virtual ~Action(){}
 };
 
-}
-}
+NAMESPACE_END(Grid);
 
 #endif // ACTION_BASE_H